text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1608.00327v6 [math.CV] 23 Dec 2017
CR Singularities and Generalizations of Moser’s Theorem I
Valentin Burcea
Abstract. Let (z1 , . . . , zN , w) be the coordinates in CN +1 . Let M ⊂ CN +1 be the real-analytic submanifold defined near p = 0 as follows
2
2
2
2
3
w = z1 z 1 + · · · + zN z N + λ1 z1 + z 1 + · · · + λN zN + z N + O |z|
,
where λ1 , . . . , λN ≥ 0. We assume that M is formally equivalent to the following model
2
2
2
2
2
2
w = z1 z 1 + · · · + zN z N + λ1 z1 + z 1 + · · · + λN zN + z N + z1 z N + z 1 zN .
We prove that M is biholomorphicaly equivalent to the model assuming that N > 1.
1. Introduction and Main Result
This note is motivated by the local equivalence problem in complex analysis and Theorem[25] of Moser. Two real-analytic submanifolds
may be only formally biholomorphically equivalent as it has been proven by Moser-Webster[24] and Gong[13] in the CR Singular situation[4],[6],
and respectively by Kossovskiy-Shafikov[19] in the CR situation[2]. Moser[25] considered the local equivalence problem for the real-analytic
surface in C2 defined near p = 0 as follows
(1.1)
w = zz + O |z|3 ,
where (z, w) are the coordinates in C2 . Moser[25] proved that if (1.1) is formally equivalent to the following model
(1.2)
w = zz,
then (1.1) is biholomorphically equivalent to it. This result is known as Theorem of Moser[25].
In this note, the following equivalence result is proven
Theorem 1.1. Let (z1 , . . . , zN , w) be the coordinates in CN+1 . Let M ⊂ CN+1 be a submanifold defined near p = 0 as follows
2
+ z 2N + O |z|3 ,
(1.3)
w = z1 z 1 + · · · + zN z N + λ1 z12 + z 21 + · · · + λN zN
where λ1 , . . . , λN ≥ 0. We assume that that M is real-analytic and formally equivalent to the following model
2
w = z1 z 1 + · · · + zN z N + λ1 z12 + z 21 + · · · + λN zN
+ z 2N + z12 z N + z 21 zN .
(1.4)
Then M is holomorphically equivalent to the model assuming that N > 1.
The proof of this result is based on a careful analysis of formal equivalences in the local defining equations (1.3) of M and (1.4) of
the model. The main role is played by real submanifolds existent near the C.-R. singularity p = 0, which are minimal hypersurfaces in CN .
More precisely, we adapt the methods of Mir[22],[23] applying Approximation Theorem of Artin[1] under the assumption that M is formally
equivalent to the model from (1.4). The presence of the terms of order 3 in (1.4) has a non-degeneracy role for concluding the convergence of
the formal transformation, regardless that the expression of the terms of degree 3 from (1.4) is not invariant.
The CR Singularities[4],[6] in codimension 2 are important for the area of the analysis of several complex variables. Dolbeault[8], [9]
and Dolbeault-Tomassini-Zaitsev[10],[11] observed that near the C.-R. singularity p = 0, the real submanifold M is of strictly positive C.-R.
dimension, when N > 1. This contrasts with the case of Moser-Webster[24] in C2 , where near the C.-R. singularity[4] the real submanifold
is totally real. They[10],[11] used the C.-R. structure existent around the C.-R. singularity and the existence of the CR Singularities[4],[6] in
order to study the problem of the existence of Levi-flat hypersurfaces with prescribed boundary[10],[11], when N > 1. Huang-Yin[16],[17]
impresivelly exploited the C.-R. structure around the C.-R. singularity[4],[6] in order to study the local hull of holomorphy[18], when N > 1.
Gong-Lebl[14] constructed normal forms for real C.-R. singular submanifolds of codimension 2 which are Levi-flat at the C.-R. points, when
N > 1. The existence of CR Singularities has been used by Slapar[27],[26] in order to show the existence of regular Stein neighborhood bases
for compact real surfaces[26] in C2 .
Acknowledgements I thank Prof. Xiaojun Huang for useful discussions regarding the Generalization [15] of Theorem [25] of Moser.
I thank Dr. Jiri Lebl for interesting conversations about [7]. I thank to my supervisor Prof. Dmitri Zaitsev for many long conversations
regarding C.-R. Singularities[11] while I had been working in the School of Mathematics of Trinity College Dublin. I thank for hospitality
to the Departament of Mathematics of the Federal University of Minas Gerais during my short visit there. I would also like to thank Prof.
Nordine Mir for remarks on some preliminary draft of this note. Special Thanks to Science Foundation of Ireland.
Keywords: CR singularity, equivalence problem, Artin Approximation.
This project was supported principally by CAPES, being initiated with Science Foundation Ireland grant 10/RF P/MT H2878.
1
2
VALENTIN BURCEA
2. Preparations
Let
M, M ′
⊂
CN+1
be real-analytic submanifolds defined near p = 0 as follows
M : w = Q (z, z) + ϕ(z, z),
(2.1)
M ′ : w′ = Q z ′ , z ′ ,
where ϕ(z, z) = O |z|3 and Q (z, z) is the model of degree 3 defined by the right-side from (1.4).
We consider the formal equivalence between M and M ′ defined as follows
(2.2)
w ′ , z ′ = (g(z, w), f1 (z, w), . . . , fN (z, w)) = (g(z, w), f (z, w)) .
Following [5],[29], we can eventually compose with an linear automorphism of the quadratic model from (1.4) in order to assume that
X (N)
X (1)
X
X
k
k
k
k
fk (z)w
(2.3)
,
fk (z)w , . . . , zN +
fk (z)w = z1 +
gk (z)w , f (z, w) = z +
g(z, w) = w +
k≥0
k≥0
k≥0
k≥0
where fk (z) and gk (z) are formal holomorphic power series, for all k ∈ N.
The formula of the terms of bidegree 3 in (2.1) may be become different by further changes of coordinates. The presence of these terms
is just a non-degeneracy condition as we shall observe throughout next computations for finding suitable analytic systems. In order to have
simplified computations, we firstly assume that we have formal normalized equivalences. It follows thus by (2.1) and (2.2) that
g(z, w) = Q f (z, w), f (z, w) ,
(2.4)
or by (2.3) equivalently to
X
(2.5)
k≥0
gk (z)w k = Q z +
X
k≥0
fk (z)w k , z +
X
k≥0
fk (z)w k .
Because M and M ′ are formally equivalent, it follows that there exist {ai,j (z, z, w, w)}i,j=1,2 real-formal functions defined near 0 ∈ CN+1
which are not identically vanishing, satisfying by (2.1) the following
X
X
X
k
k
k
fk (z)w
= a11 (z, z, w, w) (Re w − Q (z, z) − Re ϕ(z, z))
fk (z)w , z +
gk (z)w − Q z +
(2.6)
k≥0
k≥0
k≥0
+ a12 (z, z, w, w) (Im w − Im ϕ(z, z)) + i (a21 (z, z, w, w) (Re w − Q (z, z) − Re ϕ(z, z)) + a22 (z, z, w, w) (Im w − Im ϕ(z, z))) ,
Restricting (2.1) at w = x, for each x ∈ (0, ǫ) with given ǫ > 0, we obtain the following real-analytic hypersurfaces
2
+ z 2N + Reϕ(z, z) − x.
(2.7)
ρ1 (z, z, x) = 0, ρ1 (z, z, x) = z1 z 1 + · · · + zN z N + λ1 z12 + z 21 + · · · + λN zN
Dolbeault-Tomassini-Zaitsev [10],[11] observed that near the C.-R. singularity p = 0, the real submanifold M ⊂ CN+1 defined as in (1.3),
may be non-minimal at its C.-R. points. This fact motivates the problem study in CN , because (2.7) defines also a hypersurface in CN . More
precisely, we make the following observation
Lemma 2.1. Let (z1 , . . . , zN ) be the coordinates in CN and let ρ1 (z, z, x) be defined as in (2.7). If px ∈ CN is chosen such that
(2.8)
ρ1 (px , px , x) = 0,
∂ρ1 (px , px , x)
6= 0,
∂zN
then (2.8) is a hypersurface of finite type at px .
Proof. The existence of the point px 6= 0, for each x ∈ (0, ǫ), may be obtained following Dolbeault-Tomassini-Zaitsev [10],[11]. Following
[2], we consider the following base of C.-R. vector fields associated to (2.8) defined as follows
−1
∂
∂ρ1
∂
∂ (Reϕ(z, z))
(2.9)
Lk =
−
(px , px )
zk + 2λk z k +
, k = 1, . . . , N − 1.
∂z k
∂z N
∂z k
∂zN
Making computations, we observe that
−1
−1
∂ρ2
∂ρ2
∂
∂
(px , px )
(1 + O(1)) −
(px , px )
(1 + O(1)) ,
L1 , L1 = −
∂z N
∂zN
∂z N
∂zN
which gives us the missing direction. Here O(1) is a formal power series depending of z, z and x without constant term. It follows that (2.8)
defines a hypersurface of finite type at px .
Following Mir[22],[23], we complexify the real submanifolds defined by w = x ∈ (0, ǫ) in the real part in (2.1). We replace z 1 , . . . , z N with
ζ1 , . . . , ζN ∈ C in (2.1). We conclude the following local defining equation:
C2N ⊃ Mx : ρ1 (z, ζ, x) = 0,
(2.10)
where ρ1 (z, ζ, x) = x − Q (z, ζ) − ϕ (z, ζ) − ϕ (ζ, z).
Next, we consider px ∈ Mx satisfying (2.8). We consider the complexified C.-R. vector fields for Mx defined as follows
(2.11)
L⋆1 =
∂ρ1 (z, ζ, x) ∂
∂ρ1 (z, ζ, x)
∂ρ1 (z, ζ, x) ∂
∂
∂ρ1 (z, ζ, x) ∂
−
, . . . , L⋆N−1 =
−
.
∂ζN
∂ζ1
∂ζ1
∂ζN
∂ζN
∂ζN−1
∂ζN−1 ∂ζN
By Theorem of Implicit Functions, it follows by (2.8) the existence of the real hypersurfaces
(2.12)
Nx : Im zN = ψ (x, z1 , z 1 , . . . , zN−1 , z N−1 , Re zN ) ,
defined near 0 ∈ R. Here ψ : R2N−1 → R depends analytically on x ∈ (0, ǫ). Moreover, we can assume that ψ does not contain pure terms in z
or in z by changing eventually the coordinates.
CR SINGULARITIES AND GENERALIZATIONS OF MOSER’S THEOREM I
3
Continuing the study of (2.6) fixing x ∈ (0, ǫ) in (2.1), it follows by taking the real part in (2.6) that
X
1 X
k
k
gk (z) (x + iIm ϕ(z, ζ)) +
−x−
g k (ζ) (x − iIm ϕ (z, ζ))
2 k≥0
k≥0
(2.13)
X
X
k
k
f k (ζ) (x − iIm ϕ (z, ζ))
fk (z) (x + iIm ϕ(z, ζ) , ζ +
+Q z+
k≥0
k≥0
= a11 (z, ζ, x + iIm ϕ (z, ζ) , x + iIm ϕ (z, ζ)) (x − Q (z, ζ) − ϕ (z, ζ) − ϕ (ζ, z)) .
We observe that by (2.8), it follows by continuity that, for z near px , we have
∂ρ1 (v1 (z), 0, x)
6= 0,
∂zN
(2.14)
where v1 is the first Segre mapping defined by [22],[2] as follows
v1 : CN −→ CN ,
(2.15)
v1 (z) = v1 (z1 , . . . , zN−1 , zN ) = (z1 , . . . , zN−1 , 0) .
Before going ahead, we introduce the following notations
∂ρ1 (z, ζ, x)
, k = 1, . . . , N ;
Hk (z, ζ, x) =
∂zk
(2.16)
∂ (ϕ (z, ζ) − ϕ (ζ, z))
∂ (ϕ (z, ζ) − ϕ (ζ, z))
Vr (z, ζ) = HN (z, ζ, x)
− Hr (z, ζ, x)
,
∂ζr
∂ζr
r = 1, . . . , N − 1.
In order to apply Approximation Theorem of Artin[1], we construct by (2.13), (2.15) convenient analytic systems as follows.
2.1. Analytic Systems. Following the procedure of Mir[22],[23], we apply successively L⋆1 , . . . , L⋆N−1 in (2.13) evaluating everything
using the real-analytic hypersurfaces defined in (2.12). We obtain by (2.16) the following analytic equation
X
Vr (z, ζ) X
k−1
k−1
kgk (ζ) (x − iIm ϕ (z, ζ))
kgk (z) (x + iIm ϕ (z, ζ))
−
− Er (ζ, z, x) +
2
k≥1
k≥1
X
∂gk (ζ)
∂gk (ζ)
−
HN (z, ζ, x)
(ζ) − Hr (z, ζ, x)
(x + iIm ϕ (z, ζ))k = (zr + 2λr ) HN (z, ζ, x) − (zN + 2λN ) Hr (z, ζ, x)
∂ζr
∂ζN
k≥1
N
X
X
Vr (z, ζ) X X (l)
(l)
(l)
kζl fk (z) (x + iIm ϕ (z, ζ))k−1
2λl ζl fk (ζ) (x − iIm ϕ (z, ζ))k−1 +
+
k zl fk (ζ) +
−
2
k≥1
k≥1
l=1
k≥1
−
X
(l)
(l)
(k + p)λl fk (ζ) fp (ζ) (x − iIm ϕ (z, ζ))k+p−1 +
k,p≥1
+
−
(k +
X
(l)
(p)
pfk (z)fk
k+p−1
+ iIm ϕ (z, ζ))
+
k
(ζ) (x + iIm ϕ (z, ζ)) (x − iIm ϕ (z, ζ))
(N)
−Hr (z, ζ, x) fk (z)
+
X
l=1 k≥1
+
λl
HN
N X
X
l=1 k,p≥0
+
X
(p)
(l)
kfk (z)fk
(ζ) (x + iIm ϕ (z, ζ))k−1 (x − iIm ϕ (z, ζ))p
k,p≥0
k,p≥0
N
X
(l)
2kλl zl fk (z) (x + iIm ϕ (z, ζ))k−1
k≥1
(l)
(l)
p) λl fk (z)fp (z) (x
X
k,p≥1
(2.17)
X
N X
X
l=1 k,p≥0
+
N
X
X
k
p−1
X
+
k≥1
(r)
(x + iIm ϕ (z, ζ))k HN (z, ζ, x) fk (z)
(l)
(x − iIm ϕ (z, ζ)) zl HN (z, ζ, x)
(l)
∂fk (ζ)
− Hr (z, ζ, x)
∂ζr
(l)
(l)
∂ ζl fk (ζ)
∂ ζl fk (ζ)
(x − iIm ϕ (z, ζ))k
− Hr (z, ζ, x)
(z, ζ, x)
∂ζr
∂ζN
l=1 k≥1
∂
λl
HN (z, ζ, x)
(l)
(l)
fk (ζ) fp (ζ)
∂ζr
∂
− Hr (z, ζ, x)
p
(l)
∂fk (ζ)
(2.18)
Er (ζ, z, x) = Lr ζ1 +
+ z1 +
X
k≥0
X
k≥0
∂ζr
f k (ζ) (x − iIm ϕ (z, ζ))k zN +
k
fk (z) (x + iIm ϕ (z, ζ))
ζN +
X
k≥0
−
X
k≥0
2
fk (z) (x + iIm ϕ (z, ζ))k
2
k
f k (ζ) (x − iIm ϕ (z, ζ))
(l)
∂f (ζ)
(l)
,
Hr (z, ζ, x) fk (z) k
∂ζN
where Er (ζ, z, x) is defined as follows
∂ζN
(l)
(l)
fk (ζ) fp (ζ)
(x − iIm ϕ (z, ζ))2k
∂ζN
(l)
(x − iIm ϕ (z, ζ)) (x + iIm ϕ (z, ζ)) HN (z, ζ, x) fk (z)
k
∂fk (ζ)
,
for all r = 1, . . . , N − 1,
4
VALENTIN BURCEA
which may be computed and expanded by similar manners as (2.17), for all r = 1, . . . , N − 1. These computations are restricted to the
complexification of (2.12). It is clear that (ζ, z) = (0, v1 (z)) belongs to the complexification of Nx and also that
∂g
∂f k
(0) and
(0) may not vanish, for all k ∈ N and l = 1, . . . , N .
∂ζl
∂ζl
(2.19)
Thus (2.17) gives by (2.19), (2.12) and (2.15) the following analytic system:
2
f1 (v1 (z), x) + (fN (v1 (z), x)) HN (v1 (z), 0, x) + P1 (f1 (v1 (z), x) , . . . , fN (v1 (z), x)) = χ1 (v1 (z), x) ,
f2 (v1 (z), x) HN (v1 (z), 0, x) + P2 (f1 (v1 (z), x) , . . . , fN (v1 (z), x)) = χ2 (v1 (z), x) ,
(2.20)
......................................................................................................
fN−1 (v1 (z), x) HN (v1 (z), 0, x) + PN−1 (f1 (v1 (z), x) , . . . , fN (v1 (z), x)) = χN−1 (v1 (z), x) ,
where P1 , P2 , . . . , PN−1 are polynomials in N variables. The coefficients of these polynomials and χ1 (v1 (z), x) , . . . , χN−1 (v1 (z), x) are power
series holomorphic in z1 , . . . , zN−1 , formal in x ∈ (0, ǫ) and with no constant terms.
The analytic system (2.20) is formed just by N −1 equations, but is defined by N formal power series. Thus (2.20) does not contain sufficient
equations in order to conclude the desired partial convergence. This fact is overcamed by applying L⋆1 L⋆1 in (2.13) evaluating everything using
the complexifications of real hypersurfaces defined in (2.8). This gives by (2.14) and (2.15) another analytic equation
fN (v1 (z), x) (HN (v1 (z), 0, x))2 + P (fN (v1 (z), x)) = χN (v1 (z), x) ,
(2.21)
where P is a polynomial. Its coefficients and χN (v1 (z), x) are power series holomorphic in z1 , . . . , zN−1 , formal in x ∈ (0, ǫ) and with no
constant terms. Identifying the coefficients of x in (2.21), it follows by induction on k ∈ N and (2.14) that
(N)
(2.22)
fk
(z)|z=v
1 (z)
is convergent,
for all k ∈ N .
L⋆1 L⋆1
Moreover, we apply
in (2.13). Then we take derivatives with respect to z1 , . . . , zn in (2.20) of length |I| = n , for any I ∈ NN . Restricting
the obtained identity using the complexifications of real hypersurfaces defined in (2.8), it follows by (2.15) that
I
∂ fN
∂ I fN
(v1 (z), x) (HN (v1 (z), 0, x))2 1 + P ′
(v1 (z), x)
= χ′N (v1 (z), x) .
(2.23)
I
I
∂z
∂z
where P ′ is a polynomial. Its coefficients and χ′N (v1 (z), x) are holomorphic series in z1 , . . . , zN−1 , formal in x ∈ (0, ǫ) and with no constant
term, but depending on
(N)
∂ I fk
(z)
|z=v1 (z) , where k ∈ N, I ∈ NN with |I| < n.
∂z I
Identifying the coefficients of x in (2.23), we conclude by induction on k ∈ N and by (2.14) that
(2.24)
(N)
∂ I fk
(z)
|z=v1 (z) is convergent, where k ∈ N, I ∈ NN with |I| = n.
∂z I
Returning in (2.20), we identitify the coefficients of x in (2.20) by induction on k ∈ N⋆ . Following Proposition 4.2 from Mir[22], we conclude
by (2.20) and (2.14) that
(2.25)
(l)
(2.26)
fk (z)|z=v1 (z) is convergent,
where l = 1, . . . , N and k ∈ N .
Moreover, taking derivatives with respect to z1 , . . . , zn in (2.20) of length |I| = n with I ∈ NN , and then repeating the previous procedure
evaluating using the first Segre mapping (2.15), we conclude the following analytic system
!
I
I
2
I
∂ f1
∂ I fN
∂ f1 (v (z), x) + ∂ fN (v (z), x) H (v (z), 0, x) + P
(v1 (z), x) , . . . ,
(v1 (z), x) = χ′1 (v1 (z), x) ,
1
1
1
1
N
I
I
I
I
∂z
∂z
∂z
∂z
I
∂ I f2
∂ I fN
∂ f1
(v1 (z), x) HN (v1 (z), 0, x) + P2
(v1 (z), x) , . . . ,
(v1 (z), x) = χ′2 (v1 (z), x) ,
(2.27)
∂z I
∂z I
∂z I
.....................................................................................................................
I
If
∂ f1
∂ I fN
∂
N−1
(v
(z),
x)
H
(v
(z),
0,
x)
+
P
(v
(z),
x)
,
.
.
.
,
(v
(z),
x)
= χ′N−1 (v1 (z), x) ,
1
1
1
1
N
N−1
∂z I
∂z I
∂z I
′
are polynomials in N variables. The coefficients of these polynomials and χ′1 (v1 (z), x) , . . . , χ′N−1 (v1 (z), x) are power
where P1′ , P2′ , . . . , PN−1
series holomorphic in z1 , . . . , zN−1 , formal in x ∈ (0, ǫ) and with no constant terms, but depending on
(l)
∂ I fk (z)
(2.28)
By induction on the length of I ∈ NN
|z=v1 (z) , where l = 1, . . . , N , k ∈ N, I ∈ NN with |I| < n.
∂z I
and following Proposition 4.2 from Mir[22], we conclude by (2.20), (2.14) and (2.14) that
(l)
∂ I fk (z)
|z=v1 (z) is convergent, where l = 1, . . . , N , k ∈ N, I ∈ NN with |I| = n.
∂z I
Returning to (2.17), we identify again the coefficients of x by induction on k ∈ N⋆ , concluding by (2.14) and (2.15) that
(2.29)
∂ I gk (z)
|z=v1 (z) is convergent, where k ∈ N, for all I ∈ Nn .
∂z I
We have to observe that the defining equations are considered near certain points px ∈ Mx which is sufficient. These points may be
eventually changed leading to the same conclusion. It remains to show the following
(2.30)
CR SINGULARITIES AND GENERALIZATIONS OF MOSER’S THEOREM I
5
Lemma 2.2. Any holomorphic polynomial change of coordinates of degree 2 leaving the quadratic model invariant in (1.4) produces mixed
terms of bidegree 3 in (1.4).
Proof. It is suficient to consider the following invertible holomorphic transformation of (1.4) of coordinates :
N
N
X
X
(1)
(N)
′
′
(2.31)
z , w = z1 +
aij zi zj + b1 w, . . . , zN +
aij zi zk + bN w, w .
i=1
i,j=1
Replacing everything in the corresponding local defining equations, it follows by (1.4) that
( N
!)
N
X
X
(k)
2
+ z 2N
zk
aij zi zj + bk z1 z 1 + · · · + zN z N + λ1 z12 + z 21 + · · · + λN zN
w = Q(z, z) + Re
k=1
(2.32)
+ 2Re
(
N
X
k=1
λk zk
N
X
i=1
i=1
!)
(k)
2
+ z 2N
aij zi zj + bk z1 z 1 + · · · + zN z N + λ1 z12 + z 21 + · · · + λN zN
+ O(3).
Studying the mixed terms of degree 3 in (2.32), it follows that any holomorphic change of coordinates eliminating the present terms, will generate
other mixed terms of degree 3. Thus the presence of the mixed terms of degree 3 can not be eliminated by changing the coordinates.
The existence of mixed terms of degree 3 is fundamental in order to construct convenient analytic systems for concluding partial convergence
observations. Any change of coordinates eliminating those mixed terms, produces other mixed terms of degree 3 which replace their nondegeneracy role. For instance, if the coefficient of z1 z2 z1 does not vanish, we can apply L⋆1 L⋆2 to the local defining equation in the new
coordinates. We find thus desired analytic systems by similar manners.
3. Proof of Theorem 1.1
The proof is inspired by Lemma 6.1 of Mir[22]. Here the minimality of the hypersurfaces (2.12) is fundamental in order to consider standard
parametrizations using Segre sets[2]. Generally, the minimality is the standard geometrical condition in order to conclude convergence results for
formal holomorphic transformations. Baouendi-Mir-Rothschild[3], Mir[22],[23], Sunyé[28], Merker[20], Meylan-Mir-Zaitsev[21] are suggested
for related studies where the minimality is considered. Kossovskiy-Shafikov[19] showed recently that between two nonminimal real-analytic
formally equivalent submanifolds in the complex space, it may not exist holomorphic equivalences.
The proof relies on Approximation Theorem of Artin[1]. We look at z = (z1 , . . . , zN ), z = ζ = (ζ1 , . . . , ζN ) as independent variables taking
in consideration the complexifiation of (2.12). Moreover, we need to more generally assume that x ∈ C. By a translation, we can consider x
near 0 ∈ C. By (2.2), we rewrite (2.17) more generally as follows
1
R (z, x + iIm ϕ (z, ζ) , A (x, ζ, z)) = − B (x, z, ζ) ,
2
(3.1)
where the following notations are used
1
B (x, z, ζ) = − g (ζ, x − iIm ϕ (ζ, z)) ,
2
Following Mir[22],[23], the following expression is also considered
(3.2)
(3.3)
A (x, ζ, z) = f (ζ, x − iIm ϕ (z, ζ)) ,
T ((λ, µ) , z, x, ζ) = R (z, x + iIm ϕ (z, ζ) , λ) − µ,
R (z, w, λ) = Q (f (z, w), λ) −
1
g(z, w).
2
λ ∈ CN , µ ∈ C.
In order to apply Approximation Theorem of Artin[1], we have to reconsider the situation introducing more variables as follows. We
′
1 − z 2 , where x , x , z 1 , z 2 ∈ C. Defining also z ′ = (z , . . . , z
assume that x = x1 − x2 , zN = zN
1
1
2 N N
N−1 ) and ζ = (ζ1 , . . . , ζN−1 ), we consider the
N
1 , z 2 , ζ, denoted by C z ′ , x , x , z 1 , z 2 , ζ
throughout
the
rest of this note.
ring of the formal power series in z ′ , x1 , x2 , zN
1
2 N N
N
We are ready now to give the proof
Proof. We conclude by (2.29) and (2.30) the following
(3.4)
∂ k+n
2
1
, x1 − x2 , ζ |x1 =x2 ,
− zN
T (λ, µ) , z ′ , zN
∂ k zN ∂xn
1
1 =z 2
zN
N
is convergent, for all n, k ∈ N⋆ .
From Approximation Theorem of Artin[1], we obtain that the following analytic system
∂ k+n
1
2
1
2
T (C, D) z ′ , x1 − x2 , zN
− zN
, ζ z ′ , zN
− zN
, x1 − x2 , ζ |x1 =x2 , z 1 =z 1 ≡ 0, for all n, k ∈ N,
1 n
N
N
∂ k x1 k ∂zN
1
′
1
and
admits an convergent solution (C, D) z , zN , x1 , ζ agreeing in its Taylor expansion up to a certain order with A x1 , ζ, z ′ , zN
′
1
B x1 , z , zN , ζ defined by (3.2). It follows that
′
1
1
2
1 , z2 , ζ .
(3.6)
T (C, D) z ′ , x1 , ζ, zN
, z , x 1 − x 2 , zN
− zN
, ζ ≡ 0 in C z ′ , x1 , x2 , zN
N
(3.5)
2 = 0 in (3.6), we obtain by (3.3) and (3.2) that
Taking x2 = 0 and zN
(3.7)
R (z, x + iIm ϕ (z, ζ) , D (z, x, ζ)) = C (z, x, ζ) in C [[z, x, ζ]].
In order to prove the convergence of f and g, we use Proposition 6.2 from Mir[22] defining a parametrization of (3.7) as follows. By (2.12), we
conclude the existence of the following varieties
n
o
(3.8)
Qz = z ∈ CN ; ζN = Θ z, ζ ′ , x ,
6
VALENTIN BURCEA
where Θ is convergent near 0 ∈ C2N−1 . Since N > 1, it follows by Lemma 2.1 that the hypersurfaces (2.7) are actually of finite type fixing x
near 0 in C. Thus, considering the following mapping
(3.9)
V x, z, ζ ′ = D x, zN , ζ ′ , Θ z, ζ ′ , x ,
we obtain following[22] an application of maximum generic rank:
(3.10)
C × CN × CN−1 ∋ x, z, ζ ′ 7−→ x, z, ζ ′ , Θ z, ζ ′ , x , V x, z, ζ ′ , ,
which parametrizes (3.7). It follows that R is convergent and thus that f (z, x + Im ϕ (z, ζ)) and g (z, x + Im ϕ (z, ζ)) are convergent. Taking
ζ = 0 previously, it follows that f and g are convergent. The proof is finished.
Finally, we mention that new generalizations of Theorem of Moser[25] have been concluded by Huang-Yin[15] using the assumption that
(1.1) is formally equivalent with the quadratic model in (1.1) with λ1 = · · · = λN = 0, and respectively by the author[5] in a more general
situation. In those cases, this approach does not apply because it does not exists sufficient equations in order to obtain partial convergence
observations. Thus the existence of the terms of degree 3 may be seen as a non-degeneracy condition. These terms of bidegree 3 may be
eventually selected differently. Thus there are many other suitable models which may be used in order to generalize Theorem of Moser[25].
References
[1] Artin, M. — On the solutions of analytic equations. Inv. Math. 5 (1968), 277 − 291.
[2] Baouendi, M.S.; Ebenfelt, P.; Rothschild, L.P. — Real Submanifolds in Complex Space and Their Mappings. Princeton Math. Ser. 47, Princeton Univ.
Press, 1999.
[3] Baouendi, M.S.; Mir, N.; Rothschild, L.P.— Reflection ideals and mappings between generic submanifolds in complex space. J. Geom. Anal., 12 (4)
(2002), pp. 543 − 580.
[4] Bishop, E. — Differentiable Manifolds In Complex Euclidian Space. Duke Math. J. 32 (1965), no. 1, 1-21.
[5] Burcea, V. — A normal form for a 2-codimensional submanifold M ⊂ CN +1 near a CR singurity. Adv. in Math. 243 (2013), 262-295.
[6] Burcea, V. — On a family of analytic discs attached to a real submanifold M ⊂ CN +1 , Methods and Applications of Analysis 20, 1, (2013), 69-78.
[7] Burcea, V. — Real Submanifolds in Complex Spaces, Acta Mathematica Sinica-English Series. 33, (2017), 1, 1-20.
[8] Dolbeault, P. — On Levi-flat hypersurfaces with given boundary in Cn . Sci. China Ser. A 51, (2008), no. 4, 541 − 552.
[9] Dolbeault, P. — Boundaries of Levi-flat hypersurfaces: special hyperbolic points. Ann. Polon. Math. 106, (2012), 145 − 170.
[10] Dolbeault, P.; Tomassini, G.; Zaitsev, D. —On Levi-flat hypersurfaces with prescribed boundary. Pure Appl. Math. Q , 6, (2010), no. 3, (Special
Issue: In honor of Joseph J. Kohn. Part 1), 725 − 755.
[11] Dolbeault, P.; Tomassini, G.; Zaitsev, D. —Boundary problem for Levi flat graphs. Indiana Univ. Math. J.60 (2011), no. 1, 161 − 170.
[12] Gong, X. — On the convergence of normalizations of real analytic surfaces near hyperbolic complex tangents. Comment. Math. Helv. 69 (1994), no. 4,
549 − 574.
[13] Gong, X. — Existence of real analytic surfaces with hyperbolic complex tangent that are formally, but not holomorphically equivalent to quadrics.
Indiana Univ. Math. J. 53 (2004), no. 1, 83 − 95.
[14] Gong, X.; Lebl, J.— Normal forms for CR singular codimension-two Levi-flat submanifolds. Pacific J. Math 275 (2015), no. 1, 115 − 165.
[15] Huang, X.; Yin, W. — A codimension two CR singular submanifold that is formally equivalent to a symmetric quadric. Int. Math. Res. Notices (2009),
no. 15, 2789 − 2828.
[16] Huang, X.; Yin, W. — Flattening of CR singular points and the analicity of the local hull of holomorphy I. Math. Ann. 365 (2016), no. 1-2, 381399.
[17] Huang, X.; Yin, W. — Flattening of CR singular points and the analicity of the local hull of holomorphy II. Adv. Math. 308 (2017), 1009-1073.
[18] Huang, X.; Krantz, S. — On a problem of Moser. Duke Math. J., 78, (1995), no. 1, 213 − 228.
[19] Kossovskiy, I. ; Shafikov, R. —Divergent CR-Equivalences and Meromorphic Differential Equations.Journal of European Math. Society,18, (2016),
no.12, 2785 − 2819.
[20] Merker, J.— Convergence of formal invertible CR mappings between minimal holomorphically nondegenerate real analytic hypersurfaces. Int. J. Math.
Math. Sci. 26 (2001), no. 5, 281 − 302.
[21] Meylan, F.; Mir, N.; Zaitsev, D.— Approximation and convergence of formal CR-mappings.Int. Math. Res. Not. 2003, no. 4, 211 − 242.
[22] Mir, N.— Formal biholomorphic maps of real analytic hypersurfaces.Math. Res. Lett. 7 (2000), no. 2-3, 343 − 359.
[23] Mir, N.— On the convergence of formal mappings. Comm. Anal. Geom. 10 (2002), no. 1, 23 − 59.
[24] Moser, J.; Webster, S. — Normal forms for real surfaces in C2 near complex tangents and hyperbolic surface transformations. Acta Math. 150 (1983),
255 − 296.
[25] Moser, J. —Analytic Surfaces in C2 and their local hull of holomorphy. Ann. Acad. Sci. Fenn. Ser. A.I. Math. 10 (1985), 397-410.
[26] Slapar, M. —On Stein Neighborhood Basis of Real Surfaces. Math. Z. 247 (2004), no. 4, 863 − 879.
[27] Slapar, M. —Modelling complex points up to isotopy. J. Geom. Analysis 23 (2013), no. 4, 1932 − 1943.
[28] Sunyé, J.C.— On formal maps between generic submanifolds in complex space. J. Geom. Anal. 19, (2009), no. 4, 944 − 962.
[29] Zaitsev, D. — New Normal Forms for Levi-nondegenerate Hypersurfaces. Several Complex Variables and Connections with PDE Theory and Geometry.
Complex analysis-Trends in Mathematics, Birkhuser/Springer Basel AG, Basel, (Special Issue: In the honour of Linda Preiss Rothschild), pp. 321 − 340,
(2010).
V. Burcea: Departaments of Mathematics, The Federal University of Minas Gerais, Belo Horizonte, Brazil
E-mail address: [email protected]
| 0math.AC
|
Guaranteed Fault Detection and Isolation for Switched Affine Models
arXiv:1704.05947v2 [math.OC] 1 Oct 2017
Farshad Harirchi
Sze Zheng Yong
Abstract— This paper considers the problem of fault detection and isolation (FDI) for switched affine models. We first
study the model invalidation problem and its application to
guaranteed fault detection. Novel and intuitive optimizationbased formulations are proposed for model invalidation and
T -distinguishability problems, which we demonstrate to be
computationally more efficient than an earlier formulation
that required a complicated change of variables. Moreover, we
introduce a distinguishability index as a measure of separation
between the system and fault models, which offers a practical
method for finding the smallest receding time horizon that is
required for fault detection, and for finding potential design
recommendations for ensuring T -distinguishability. Then, we
extend our fault detection guarantees to the problem of fault
isolation with multiple fault models, i.e., the identification of
the type and location of faults, by introducing the concept of
I-isolability. An efficient way to implement the FDI scheme is
also proposed, whose run-time does not grow with the number
of fault models that are considered. Moreover, we derive bounds
on detection and isolation delays and present an adaptive
scheme for reducing isolation delays. Finally, the effectiveness
of the proposed method is illustrated using several examples,
including an HVAC system model with multiple faults.
I. I NTRODUCTION
Cyber-physical systems (CPS), i.e., systems with integrated computation, networking, and physical processes, are
becoming increasingly common in our daily lives. Such systems include critical infrastructures such as traffic, power and
water networks, as well as autonomous vehicles, aircrafts,
home appliances and manufacturing processes. However,
some major incidents involving these critical infrastructure
systems as a result of cyber-attacks and system failures
have taken place in the recent years and are a big source
of concern. Hence, the reliability and security of CPS is
paramount for their successful implementation and operation.
The detection and isolation of faults and anomalies in CPS
play an important role in enhancing the reliability of these
systems, and in understanding the vulnerability of system
components to failures and attacks.
1) Literature Review: The study of fault detection began
with the introduction of the first failure detection filter by
Beard in 1971 [1]. Since then, fault diagnosis has attracted
a great deal of attention and has become an integral part of
most, if not all system designs. Researchers have mainly
approached the fault detection and isolation problem by
This work is supported in part by DARPA grant N66001-14-1-4045 and
an Early Career Faculty grant from NASA’s Space Technology Research
Grants Program.
F. Harirchi and N. Ozay are with the Electrical Engineering and Computer
Science Department, University of Michigan, Ann Arbor, MI, 48109 (email:
{harirchi,necmiye}@umich.edu); S.Z. Yong is with the School for
Engineering in Matter, Transport and Energy, Arizona State University,
Tempe, AZ, 85281 (email: [email protected]).
Necmiye Ozay
employing either data-driven techniques or model-based approaches. These methods can, in general, be further grouped
into active and passive approaches. In active fault detection,
the system is excited by a carefully designed input [2]–[4],
while in passive methods, the behavior of the system is not
controlled or altered by the FDI scheme [5], [6].
In broad strokes, model-based fault detection and isolation
schemes in the literature can be categorized into two classes,
i.e., approaches that are based on residual generation and
on set-membership. The former approach is more common
in the fault diagnosis literature, and in this approach, the
difference between the measurements and the estimates is
defined as a residual or a symptom [7]. Two major trends
in the residual generation techniques are methods based on
observers [5], [8]–[14] and parameter estimation [15], [16].
On the other hand, set-membership based fault detection
and isolation techniques are proposed with the goal of
providing guarantees for the detection of some specific faults.
Most of these methods operate by discarding models that are
not compatible with observed data, rather than identifying
the most likely model. There is an extensive literature on setmembership based methods for active fault detection of linear
models [17]–[19]. In [20], [21], we posed set-membership
based guaranteed passive fault detection approaches for the
class of switched affine models and polynomial state space
models. These approaches are developed by utilizing ideas
from model invalidation [22], [23] and taking advantage
of recent advances in optimization. In addition, a concept
called T -distinguishability has been introduced for finding
conditions under which the fault detection scheme can be
applied in a receding horizon manner without compromising
detection guarantees. T -distinguishability is closely related
to the concept of input-distinguishability of linear systems
[24], [25] and mode discernibility in hybrid systems [26].
2) Main Contributions and Paper Structure: In this paper,
we consider a passive fault detection and isolation scheme
for switched affine systems using an optimization-based
model invalidation framework, that improves and expands
the results of [20] on fault detection. We provide novel formulations of the model invalidation and T -distinguishability
algorithms that we demonstrate to be noticeably faster than
the previous formulation in [20] and that have the added advantage of simplicity as no complicated change of variables
are needed. Furthermore, we introduce a measure of separation between models, called distinguishability index. By
reformulating the T -distinguishability optimization problem
in [20], we can compute the distinguishability index as a
byproduct. This index offers a practical way to find out if
a finite receding time horizon exists, and suggests potential
design options for ensuring T -distinguishability.
We then consider the fault isolation problem using model
invalidation and introduce the concept of I-isolability when
multiple faults are present. Similar to fault detection, we
propose a computationally efficient optimization problem to
check whether a given set of fault models is I-isolable or
not. Moreover, we propose a fault diagnosis scheme that not
only detects the occurrence of a fault, but also outputs a list
of potential faults along with their associated ‘likelihoods’ in
the form of distinguishability indices. Further, a theoretical
analysis for bounds on detection and isolation delays is
provided and an adaptive fault isolation scheme is proposed
to reduce isolation delays. The run-time of our FDI scheme
also does not grow with the number of fault models. Finally,
these results are illustrated using a numerical model of a
heating, ventilating, and air conditioning (HVAC) system.
N
−1
+
Bswa
(G) := {ut , yt }N
t=0 | ut ∈ U and ∃xt ∈ X , σt ∈ Zm ,
[ηη |t ν |t ]| ∈ E, for t = 0, . . . , N − 1 s.t. (3) holds .
Moreover, with a slight abuse of terminology, we will call
N
Bswa
(G) the behavior of the system G for conciseness.
III. M ODEL I NVALIDATION AND T -D ETECTABILITY
II. P RELIMINARIES
In this section, the notation used throughout the paper and
the modeling framework we consider are described.
A. Notation
Let x ∈ Rn denote a vector and M ∈ Rn×m represent a
.
matrix. The infinity norm of a vector x is denoted by kxk =
maxi |xi |, where xi denotes the ith element of vector x. The
set of positive integers up to n is denoted by Z+
n , and the
set of non-negative integers up to n is denoted by Z0n .
B. Modeling Framework
In this paper, we consider systems that can be represented
by discrete-time switched affine (SWA) models.
Definition 1: (SWA Model) A switched affine model is
defined by:
G = (X , E, U, {Gi }m
(1)
i=1 ),
where X ⊂ Rn is the set of states, E ⊂ Rny +np is the set
of measurement and process noise signals, U ⊂ Rnu is the
set of inputs and {Gi }m
i=1 is a collection of m modes. For
th
all i ∈ Z+
mode is an affine model:
m , the i
Gi = {Ai , Bi , Ci , Di , fi , gi }.
Definition 2 (Fault Model): A fault model for a switched
affine system G = (X , E, U, {Gi }m
i=1 ) is another switched
affine model Ḡ = (X̄ , Ē, Ū, {Ḡi }m̄
i=1 ) with the same number
of states, inputs and outputs.
Further, to describe our framework of model invalidation
and T -distinguishability for fault detection and isolation in
the next section, we define the following.
Definition 3 (Length-N behavior): The length-N behavior associated with an SWA system G is the set of all lengthN input-output trajectories compatible with G, given by the
following set:
(2)
The evolution of G is governed by:
xt+1 = Aσt xt + Bσt ut + fσt + ν t ,
(3)
yt = Cσt xt + Dσt ut + gσt + η t ,
where ν ∈ Rnp and η ∈ Rny denote the process and
measurement noise signals, respectively, and σt indicates the
active mode at time t.
Remark 1: We assume X , E, U are convex and compact
sets. In particular, we consider the following form for the
admissible sets:
X = {x | P x ≤ p}, E = {[ηη | ν | ]| | kηη k ≤ η , kνν k ≤ ν },
U = {u | kuk ≤ U },
(4)
where P ∈ Rnp ×n and p ∈ Rnp . Note that our analysis holds
true for any X , E, U that are convex sets, but for simplicity
in notation, we use the above mentioned admissible sets.
We define the fault model as follows:
A. Model Invalidation
In our previous work [20], [21], we established a theoretical framework that can be utilized in order to develop
fault detection schemes based on the achievements in model
invalidation, a framework that we will also consider in this
paper. The model invalidation problem is to check whether
some given data can be represented by a model or not. More
formally, the model invalidation problem is as follows:
Problem 1 (Model Invalidation): Given an SWA model
N −1
G and an input-output sequence ut , yt t=0 , determine
whether or not the input-output sequence is contained in the
behavior of G, i.e., whether or not the following is true:
N −1
N
ut , yt t=0 ∈ Bswa
(G).
(5)
Clearly, if the model is invalidated by data, i.e., (5) does
not hold, and the model is precise, it is equivalent to the
data representing an abnormal behavior. Hence, model invalidation can be used as a fault detection scheme. Moreover,
our previous work [20] has shown that model invalidation
problem for SWA models can be posed as a Mixed-Integer
Linear Program (MILP) feasibility check problem.
In this paper, we propose a new MILP formulation, that we
believe is much more intuitive and computationally efficient.
We obtain this novel formulation by taking advantage of
Special Ordered Set of degree 1 (SOS-1) constraints [27],
that are readily implementable in most off-the-shelf optimization softwares. In brief, an SOS-1 constraint is a set of
variables for which at most one variable in the set may be
non-zero. Our new formulation is cleaner because this type
of constraints allows us to formulate the feasibility check
problem without introducing complicated change of variables
as was previously done in [20]. Moreover, SOS-1 constraints,
which are by nature integral constraints, make the branch
and bound search procedures noticeably faster (see, e.g., [28,
Section 3.3.4] for a discussion). Our new model invalidation
problem using SOS-1 constraints is presented below, which
we will demonstrate to be much faster than an earlier
formulation [20], [29] in Section VI-A.
Proposition 1: Given an SWA model G and an inputN −1
output sequence ut , yt t=0 , the model is invalidated if
and only if the following problem is infeasible.
Find xt , η t , ν t , ai,t , si,t , ri,t for ∀t ∈ Z0N −1 , ∀i ∈ Z+
m
+
+
0
s.t. ∀j ∈ Z+
n , ∀k ∈ Zny , ∀l ∈ Znp , ∀t ∈ ZN −1 , we have:
xt+1 = Ai xt + Bi ut + fi + ν t + si,t ,
(PM I )
yt = Ci xt + Di ut + gi + η t + ri,t ,
P
ai,t = 1, kνν t k ≤ ν ,
P xt ≤ p, ai,t ∈ {0, 1},
i∈Z+
m
kηη t k ≤ η ,
(ai,t , sji,t )
: SOS-1,
(ai,t , rki,t )
: SOS-1,
where si,t and ri,t are slack variables that are free when
ai,t is zero and zero otherwise. We refer to this problem as
−1
F eas({ut , yt }N
t=0 , G).
Intuitively, the infeasibility of (PM I ) indicates that there
are no state, input and noise values that can generate the
input-output sequence from the model, and hence it is impossible that the data is generated by the model. Proposition
1 enables us to solve the model invalidation problem by
checking the feasibility of (PM I ), which is a MILP with
SOS-1 constraints that can be efficiently solved with many
off-the-shelf softwares, e.g., [30], [31].
B. T -Distinguishability
The model invalidation problem can be solved for the
input-output sequence of any given time horizon to detect
faults, but the number of variables and constraints increase
with the size of the time horizon. Thus, a few questions
naturally arise with regards to this time horizon.
First, one may ask if the smallest receding time horizon T
can be found, for which two different models are guaranteed
to be distinguishable. This question leads us to define the notion of T -distinguishability, previously presented in [20]1 . T distinguishability is defined for a pair of system and/or fault
models, which means that the trajectory generated from the
two models cannot be identical for a time horizon of length
T for any initial state and any noise signals. This notion is
very similar to the concept of input-distinguishability, which
is defined for linear time-invariant models in [24], [25]. T distinguishability is formally defined as follows:
Definition 4 (T -distinguishability): A pair of switched
affine models G and Ḡ is called T -distinguishable if
T
T
Bswa
(G) ∩ Bswa
(Ḡ) = ∅, where T is a positive integer.
Thus, given two SWA models and an integer T , the
T -distinguishability problem is to check whether the two
models are T -distinguishable or not. This problem can be addressed using a Satisfiability Modulo Theory approach [20],
or a MILP feasibility check [29]. As with model invalidation
in the previous section, we will also propose an alternative
MILP formulation for checking T -distinguishability, which
employs SOS-1 type constraints and as before, is more
intuitive and computationally superior (cf. Section VI-A).
Note that in the following T -distinguishability test, we have
added a decision variable δ that will be important in a later
discussion, which can be computed with little additional
computational cost.
1 When the pair of models consists of the nominal system model and the
fault model, this is also known as T -detectability [20], [29].
Theorem 1: A pair of switched affine models G and Ḡ is
T -distinguishable if and only if the following is infeasible.
δ̄ =
s. t.
min
η ,η̄
η ,ν
ν ,ν̄
ν ,s,s̄,r,a,δ
x,x̄,u,η
δ
∀t ∈ Z0T −1 , ∀i ∈ Z+
m,
+
∀h ∈ Z+
,
h̄
∈
Z
np
np̄ ,
+
+
∀j ∈ Z+
m̄ , ∀k ∈ Zn , ∀l ∈ Zny ,
xt+1 = Ai xt + Bi ut + fi + ν t + si,t ,
x̄t+1 = Āj x̄t + B̄j ut + f̄j + ν̄ν t + s̄j,t ,
Pxt ≤ p, P̄x̄t ≤ p̄,
(PT )
Ci xt +Di ut +gi +ηη t = C̄j x̄t + D̄j ut + ḡj +η̄η t +ri,j,t ,
P
P
ai,j,t ∈ {0, 1},
a
= 1,
i∈Z+
j∈Z+ i,j,t
m
m̄
kηη t k ≤ η , kη̄η t k ≤ η̄ , kνν t k ≤ ν , kν̄ν t k ≤ ν̄ , kut k ≤ U,
(ai,j,t , ski,t ) : SOS-1, (ai,j,t , s̄kj,t ) : SOS-1,
η̄η
ηt
(ai,j,t , rli,j,t ) : SOS-1,
− ν t ≤ δ.
ν̄ t
νt
We refer to the above-mentioned problem as F easT (G, Ḡ).
Proof: Except for the last constraint, this is an equivalent formulation to the MILP feasibility problem of Theorem
1 in [29]. Clearly, the last constraint does not change the
feasible set, therefore the feasibility of (PT ) is necessary and
sufficient for T -distinguishability.
While Theorem 1 enables us to solve the T distinguishability problem, if the two models are not T distinguishable, i.e., the solution to PT is feasible, it additionally delivers δ̄, which we argue is a good indication and
measure for the separability of two models. In essence, δ̄ can
be interpreted as the noise effort that is required to make the
trajectories of the two models identical. A larger value for
δ̄ indicates a larger separation between the two models that
the noise has to compensate for. Hence, we will refer to the
normalized version of δ̄ as the distinguishability index
δ̄
δ ∗ = max ,
(6)
δ
.
where δ max = min{max{η + η̄ , ν + ν̄ }, max{η , ν } +
max{η̄ , ν̄ }} is an upper bound on δ̄; hence, 0 ≤ δ ∗ ≤ 1.
Moreover, to find the smallest T for which we have T distinguishability, one could iterate with T increasing from 1
until the T -distinguishability problem in Theorem 1 becomes
infeasible. But, if δ ∗ is small for some T , then it may
make sense to consider larger increases in T to speed up
computation. Thus, δ ∗ can be used as a heuristic for choosing
the next T to solve the T -distinguishability problem. In
addition, one may also ask about when the iterations with
increasing T can be terminated with some confidence that a
finite T does not exist. Once again, we can consider the trend
of δ ∗ with increasing T and terminate the iterations when δ ∗
reaches a plateau. This will be demonstrated to be effective in
a simulation example in Section VI-C. In addition, when this
index reaches a (non-zero) plateau and the problem remains
not T -distinguishable, then it would be possible to use any
value that is smaller than the maximum δ ∗ to derive the
maximum allowed uncertainty for a system such that fault
detection is guaranteed. This may suggest possible design
remedies involving the choice of sensors with better precision
or the employment of noise isolation platforms to reduce the
amount of noise, in order to facilitate fault detection.
IV. FAULT I SOLATION
In practice, it is of special interest to identify the source
of a detected fault, as it significantly simplify the process of
fault accommodation. The process of distinguishing among
possible fault models is called fault isolation. In this section,
we utilize the tools we developed for distinguishability and
model invalidation to address the fault isolation problem for
switched affine models, i.e, to derive necessary and sufficient
conditions to guarantee the isolation of a set of possible
faults. In addition, we propose a tractable way to isolate
faults in real-time for many applications by leveraging the
recent advances in the development of mixed-integer linear
programming tools [31]. In order to ensure the isolability
of two fault models, it suffices that the two fault models
are T -distinguishable. In this case, by implementing model
invalidation for each of the two models on a time horizon
of length T , we can simply isolate the faults after detection.
Hence, we turn our attention to the multiple-fault scenario.
A. Multiple-Fault Scenario
The fault isolation problem becomes marginally more
challenging with multiple fault models. Let us assume that
there exist Nf fault models for a specific system. We
consider the following necessary and sufficient assumptions:
A1 (Detectability Assumption): We assume that ∀ j ∈
Z+
Nf , there exists a finite Tj such that the pair of nominal
system G and the fault model Ḡj is Tj -distinguishable.
A2 (Isolability Assumption): We assume that ∀ m, n ∈
Z+
Nf , m 6= n, there exists a finite Im,n such that Ḡm and
Ḡn are Im,n -distinguishable.
Now, we will define I-isolability for multiple faults.
Proposition 2 (I-isolability for multiple faults): Consider
a set of Nf fault models that satisfies Assumption 2. If a fault
occurs, it can be isolated in at most I = maxm,n, m6=n Im,n
steps after the occurrence. Such a set of fault models is
called I-isolable.
Proof: Under Assumption A2, and because I ≥ Im,n
for all possible pairs of fault models, all pairs of faults are
I-distinguishable. Therefore, if any of the faults occur persistently, by observing at most I samples, it will be isolated.
This is because the length-I behavior of the occurred fault
does not have any intersection with the length-I behavior of
any of the other faults.
V. FDI S CHEME
In this section, we propose a two-step FDI scheme:
1) Off-line step: In the off-line step, under Assumptions
A1 and A2, we calculate the following quantities:
Isolability index: I = max Im,n , m, n ∈ Z+
Nf , m 6= n,
m,n
Isolability index for fault i: I˜i = maxj∈Z+
Nf
Detectability index: T = maxj∈Z+ Tj .
Nf
Length of memory: K = max{T, I}
, j6=i
Ii,j ,
2) On-line step: In this step, we leverage Nf + 1 parallel
monitors corresponding to system and fault models. The
monitors are labeled as {M0 , M1 , . . . , MNf }, where M0
corresponds to the system model and Mi corresponds to
the ith fault model. First, only M0 is active for fault
detection. The rest of the monitors will be “off” until a fault
is detected by M0 . The inputs to each monitor at time t
are the input-output sequence of length Ki = max{I˜i , Ti },
{uk , yk }tk=t−Ki +1 , and the corresponding model Ḡi . For
instance, M0 knows G, and at each time step, it solves the
model invalidation problem, F eas({uk , yk }tk=t−T +1 , G). If
the problem is feasible, the monitor outputs 0, otherwise
it outputs 1. In the latter case, the bank of fault monitors
is activated and parallelly solves the model invalidation
problems for all fault models, i.e., to check if Mj solves
F eas({uk , yk }tk=t−Kj +1 , Ḡj ) for each j ∈ Z+
Nf . By Assumptions A1 and A2, it is guaranteed that in this case,
the problem of at most one monitor is feasible. The output
block receives the signal from all the monitors and shows
two elements, the first element is 1, which indicates that a
fault has occurred, and the second element is kf ∈ Z+
Nf if
the fault matches kf th fault model, or 0 if the fault does not
match any of the fault models.
At every time step t, this FDI scheme acts as a function:
N
f
[H, F] = ψ({uk , yk }tk=t−K+1 , G, {Ḡj }j=1
),
(7)
where H is 0 or 1 to indicate healthy or faulty behaviors,
and F either indicates the fault model that is active, or that
none of the fault models caused the faulty behavior.
Remark 2: In some practical examples, Assumptions A1
and A2 may not be satisfied, i.e., the FDI approach is not
guaranteed to detect and isolate the given fault models.
However, the FDI approach can be simply modified such
that F outputs either the set of faults that matches the data
(because some fault models may not be isolable) along with
their corresponding ‘likelihoods’ in terms of their distinguishability indices, or the empty set, if none of the models
matches the data.
A. Detection and Isolation Delays
In this section, we describe the notion of delays in detection and isolation of faults, and provide theoretical bounds
on these delays using detectability and isolability indices.
Definition 5: (Detection/Isolation
Delay)
Detection/isolation delay is the number of time samples it takes
from the occurrence of the fault to its detection/isolation.
We denote detection and isolation delays with τT and τI ,
respectively.
Proposition 3: Given a Ti -distinguishable pair of system
and fault models (G, Ḡi ), the detection delay of the proposed
fault detection scheme is bounded by Ti . In addition, the
isolation delay of a pair of Ii,j -isolable fault models (Ḡi , Ḡj )
is bounded by Ii,j .
Proof: Direct consequence of definitions.
Theorem 2: The detection delay for fault Ḡi using the FDI
scheme proposed in Section V is bounded by Ti , and the
isolation delay is bounded by Ki = max{I˜i , Ti }.
Proof: Assume fault i occurs at time t∗ . The FDI
approach implements model invalidation with a time horizon
size of T ≥ Ti . At time t∗ +Ti −1, the input-output trajectory
that is fed to the model invalidation contains a length Ti
trajectory that is in B Ti (Ḡi ). By Ti -distinguishability of Ḡi ,
this trajectory cannot be generated by G. Thus, the model
will be invalidated by observing at most Ti data points from
fault i. This concludes the proof for the bound on detection
delay. For isolation, the FDI approach requires detection first,
and in the worst case, detection will occur in Ti steps. On the
other hand, if we observe any trajectory from t∗ to t∗ + I˜i −1
˜
that is generated by fault i, it is not in B Ii (Ḡj ), j 6= i. This
˜
is because Ii ≥ Ii,j , j 6= i. Hence, the fault will be isolated
in at most I˜i observations of the fault. Considering that the
fault needs to be detected first, the isolation delay is bounded
by Ki = max{I˜i , Ti }. This concludes the proof.
A. Run-Time Comparison
In this section, we compare the run-time for the formulations proposed in this paper with the one in [29]. Consider
a hidden-mode switched affine model, G, with admissible
sets X = {x | kxk ≤ 11}, U = {u | kuk ≤ 1000} and
E = {ηη | kηη k ≤ 0.1}. We assume there is no process noise.
We also assume B = [1 0 1]| and C = [1 1 1] for all
modes. The system matrices of the modes are:
0.5
A1 = 0.1
−0.4
0.5
A3 = 0.2
−0.9
0.5 0.5
1
0.5
0.5
0.5
0
−0.2 0.5, f1 = 0, A2 = −0.3 −0.2 0.3 , f2 = 1,
0.6 0.2
0
0.1 −0.3 −0.5
0
0.2 0.6
0
−0.2 0.2, f3 = 0.
0.7 0.1
1
In addition, consider a fault model, G f , with:
0.8
0.7
0.6
1
1
f
f
0.1 −0.2 0.3 , B = 0 , f = 1 .
A =
−0.4 0.3 −0.2
0
1
f
The bound on isolation delays represents the worst case
scenario, where the data created by a fault model falls within
the behavior of some other models up until the very last time
step. However, this is not the case in most applications, where
the faults can be isolated much prior to this bound. Here, in
this section, we propose an adaptive fault isolation scheme
that reduces isolation delay, which is based on the idea of
validation of only one of the fault models. Since the data
prior to the time of detection will most probably invalidate
all the fault models, we propose to reduce isolation delays by
using an adaptive receding horizon that considers only the
data starting from the detection time (fixed horizon lower
bound) with increasing horizon until only one fault model
matches or validates the data. In practice, we can achieve
this by considering model invalidation problems for each of
the fault models with the adaptive receding horizon until only
one fault model remains that matches the data.
Since we assumed that the fault is among the predefined
set of models and is persistent, it is guaranteed that the fault
will be isolated with this approach. Such an approach has the
potential to significantly reduce isolation delays, as we have
observed in simulation in Section VI-B (cf. Fig. 3 (bottom)).
The implementation of the T -distinguishability approach
proves that the system and fault model pairs is 12distinguishable. We first randomly generate input-output
trajectories (5 for each time horizon length) from G f . We
then compare the model invalidation approaches that use
the proposed formulation in Prop. 1 and the one in [20],
[29]. The average run-time for each time horizon length
as well as the standard deviation of run-times for both
formulations are illustrated in Fig. 1. Clearly, the results
indicate the superiority of the proposed formulation to the
one in [20], [29]. Similar improvements were also observed
for the proposed T -distinguishability formulation in Thm. 1
when compared to [20], [29] (plots are omitted for brevity).
100
80
Time(s)
B. Adaptive Fault Isolation
SOS-1
Convex hull
60
40
20
0
0
50
100
150
200
Time Horizon(samples)
VI. I LLUSTRATIVE E XAMPLES
First, we demonstrate in Section VI-A that our new
formulations for model invalidation and T -distinguishability
in Prop. 1 and Thm. 1, respectively, are computationally
superior to the previous formulation in [20], [29]. Then,
we illustrate the performance of the proposed FDI scheme
using a numerical model for the Heating, Ventilating, and
Air Conditioning (HVAC) system that is proposed in [32] in
Section VI-B. Moreover, we provide a numerical example in
Section VI-C to illustrate the practical merits of the distinguishability index that was introduced in Section III-B. All
the simulations in this section are implemented on a 3.5 GHz
machine with 32 GB of memory that runs Ubuntu. For the
implementation of the MILP feasibility check problems, we
utilized YALMIP [33] and Gurobi [30]. All the approaches
and examples are implemented in MATLAB.
Fig. 1: Average execution time (with standard deviations) for
invalidating data generated by G f with various time horizons.
B. Fault Diagnosis in HVAC Systems
Fig. 2: Schematic of a single-zone HVAC system.
In [32], a single-zone HVAC system in cooling mode (cf.
schematic in Fig. 2) is considered. This HVAC system is
68
66
Guaranteed
isolation
for Fault 1
Fault 1
0
2
4
6
Guaranteed
isolation for
Fault 2
66
64
62
60
58
Detection
66
Fault 2
2
Detection
67
Detection
0
8
68
Fault 3
4
6
8
0
2
4
Guaranteed
isolation for
Fault 3
6
8
10 -3
10 -3
0.014
10
10
Fault 3
0.012
8
0
2
4
6
0.008
0
8
2
4
6
0
8
Detection
Flag1
Flag2
Flag3
Fault1
Fault1 (Adaptive Isolation)
2
0
2
Detection
Flag1
Flag2
Flag3
Fault 2
Fault 2 (Adaptive Isolation)
2
4
6
0
8
2
4
6
8
4
6
8
4
4
4
0
0.01
8
0
2
Detection
Flag1
Flag2
Flag3
Fault 3
Fault 3 (Adaptive Isolation)
2
0
4
6
8
0
2
Fig. 3: The outputs (top two rows) of 3 fault scenarios; Detection, isolation and adaptive isolation signals for all faults
(bottom row). Flag i is non-zero when the model invalidation problem associated with fault i using the adaptive horizon
length is validated. Adaptive isolation occurs when only one Flag is non-zero.
represented by a non-linear model as follows:
ṪT S
ẆT S
ṪSA
hf g f
f
TT S
Cp Vs
Vs
f
=
0
− Vs
0 WT S +
TSA
0.75 Vf
−0.75 CfphVw
− Vf
he
he
he
h
f
− CfpgVs Ws + Cp4Vs (Qo − hf g Mo )
Mo
f
Vs Ws + ρVs
hw
f hw
gpm
f
(T
−
W
)
+
o
o
4Vhe
Cp
Cp Vhe Ws − 6000 ρCp Vhe
− Vfs
1
0.5
0.5
0
(8)
,
where f , gpm, Mo and Qo are time varying parameters. The
parameters of the model2 are defined in [32].
We leverage an augmented state-space model with additional states Q0 and M0 that is obtained in [32]. To further
simplify the model, we assume that the fan is always turned
on and the flow rate is fixed at 17000 ft3 /min and the chiller
pump is either “off” or “on” with a fixed flow rate of 58
gal/min. These assumptions along with a discretization with
a sampling time of 10 minutes convert the nonlinear system
(8) to a switched affine model parameterized by
0.98 229.63
0.001
0.94
0
0
A1 = A2 = 0.74 −360.61 0.0008
0
0
0
0
0
0
1 0 0 0 0
C1 = C2 =
, f1 = 0.
0 1 0 0 0
1
0 −0.0035
0.3886
0
0
0.0001
0 −0.0030, f2 = −22.576,
1
0
0
0
1
0
The first three states in the SWA model represent the
deviation of TT S , WT S and TSA from their equilibria and
the last two states are Q0 and M0 . In addition, the HVAC
model is represented by GH = (X , E, U, {Gi }2i=1 ), where
X = {x | [−100 − 0.05 − 50 289800 150]| ≤ x ≤
[100 0.05 50 289950 180]| }, E = {ηη | |ηη | ≤ [0.2 0.002]| }
and U = ∅. The last two bounds on the states are for
the augmented states, which are assumed to stay within a
small range of their equilibria. The first mode corresponds
to chiller being “on” and the second mode represents the
model when it is “off”. The controller keeps the temperature
in the comfort zone of 65–75◦ F by turning the chiller on and
off. Control signals are not observed by the FDI scheme.
We consider three fault scenarios:
2h
w = 180, hf g = 1078.25, Wo = 0.018, Ws = 0.007, Cp =
0.24, To = 85, Vs = 58464, Vhe = 60.75, Mo ∈ [150 180], Qo ∈
[289800 289950], ρ = 0.074, f = 17000, gpm ∈ {0, 58}.
5
10
15
0
5
10
15
Fig. 4: Distinguishability index as a function of length of
time horizon. Left: Increase in detectability index for fault
3, T3 and isolability index of faults 2 and 3, I2,3 , for the
HVAC example; Right: Nonlinear increase with a plateau at
around T = 5, for numerical example described by (9).
1) Faulty fan: The fan rotates at half of its nominal speed.
2) Faulty chiller water pump: The pump is stuck and spins
at half of its nominal speed.
3) Faulty humidity sensor: The humidity measurements
are biased by an amount of +0.005.
The proposed approach for T -distinguishability and Iisolability gives us the following results:
TABLE I: Detectability and Isolability Indices
T1 = 4
T2 = 8
T3 = 16
I1,2 = 4
I1,3 = 4
I2,3 = 16
To illustrate the growth in the distinguishability index δ ∗
as the time horizon increases, we plot its trend in Fig. 4 (left)
for T -distinguishability of fault 3 and I-isolability of faults
2 and 3. The plot shows that the distinguishability index we
introduced does indeed deliver a nice measure of how far
two models are from detectability or isolability, and at the
same time, it allows us to estimate the size of time horizon,
T or I, to achieve T -distinguishability or I-Isolability.
Next, we consider 3 scenarios, where for each scenario i
(i ∈ {1, 2, 3}), we generate data from the nominal system for
four hours and from fault i afterwards. The times at which
the faults occur and their detection times, as well as the upper
bounds on isolation delays are indicated in Fig. 3 (top and
middle rows), which show the output trajectories for each
scenario. Furthermore, we plot in Fig. 3 (bottom row) the
detection and isolation signals for all three faults to show that
only the occurred fault is isolated in all scenarios before their
upper bounds are exceeded, and that the proposed adaptive
isolation scheme reduces the isolation delay, as desired.
C. Distinguishability Index
To illustrate the practical use of the distinguishability
index, δ ∗ , we consider two synthetic SWA models G and
Ḡ subject to measurement and process noise, given by
G:
Ḡ :
0.1
0
0.1
0
0
0.15
A1 = 0
0.1 0.2 , A2 = 0.1
0
0 ,
0.2 0.12 0
0.1 0.12 0.1
0.5
1
C1 = C2 = I, f1 = 0.2 , f2 = 0 ,
1
0.5
0.1 0 0.1
0
0 0.1
Ā1 = 0 0.1 0.2 , Ā2 = 0.1 0
0 ,
0.2 0.1 0
0.1 0.1 0.1
0.3
0.8
C̄1 = C̄2 = I, f¯1 = 0 , f¯2 = 0.2 ,
0.9
0.3
(9)
where the rest of the parameters are zero. The bounds
on the process and measurement noise are set to be 0.2
and 0.25, respectively. Fig. 4 (right) depicts the change of
the distinguishability index with increasing T . We observe
that the distinguishability index increases nonlinearly and
reaches a plateau at a value less than one. In this case, the
distinguishability index δ ∗ provides a practical indication that
these two models are very unlikely to be isolable for any
finite I. Moreover, if the noise levels are constrained to be
below the value of the plateau, then we can be sure that these
faults will be isolable. Hence, the distinguishability index can
also be exploited to derive the maximum allowed uncertainty
for a system such that certain faults are guaranteed to
be detectable or isolable. In turn, this suggests possible
measures for ensuring fault detection and isolation through
the reduction of noise levels, either with a better choice of
sensors or with the use of noise isolation platforms.
VII. C ONCLUSION
We considered the FDI problem for switched affine models. For fault detection, we proposed new model invalidation and T -distinguishability formulations using SOS1 constraints, that are demonstrated to be computationally
more efficient and do not require any complicated change of
variables. Further, we introduced the distinguishability index
as a measure of separation between the system and fault
models and showed that this index is also a practical tool for
finding the smallest receding time horizon that is needed for
fault detection and for recommending system design changes
for ensuring fault detection. Finally, a novel approach is
proposed for isolation of a set of faults, with proven isolation
guarantees under certain conditions. The effectiveness of the
proposed approaches is illustrated on an HVAC system.
R EFERENCES
[1] R. Beard. Failure accommodation in linear systems through selfreorganization. PhD thesis, MIT, 1971.
[2] R. Nikoukhah. Guaranteed active failure detection and isolation for
linear dynamical systems. Automatica, 34(11):1345–1358, 1998.
[3] R. Nikoukhah and S. Campbell. Auxiliary signal design for active
failure detection in uncertain linear systems with a priori information.
Automatica, 42(2):219–228, 2006.
[4] F. Harirchi, S.Z. Yong, E. Jacobsen, and N. Ozay. Active model
discrimination with applications to fraud detection in smart buildings.
In IFAC World Congress, Toulouse, France, 2017.
[5] P. Frank and X. Ding. Survey of robust residual generation and
evaluation methods in observer-based fault detection systems. Journal
of Process Control, 7(6):403–424, 1997.
[6] R. Patton and J. Chen. Observer-based fault detection and isolation:
robustness and applications. Control Eng. Prac., 5(5):671–682, 1997.
[7] S. Simani, C. Fantuzzi, and R. J Patton. Model-based fault diagnosis
in dynamic systems using identification techniques. Springer Science
& Business Media, 2003.
[8] P. Frank. Advances in observer-based fault diagnosis. In International
Conference on Fault Diagnosis: TOOLDIAG, 1993.
[9] H. Sneider and P. M. Frank. Observer-based supervision and fault
detection in robots using nonlinear and fuzzy logic residual evaluation.
IEEE Trans. on Control Systems Technology, 4(3):274–282, 1996.
[10] H. Hammouri, M. Kinnaert, and E. H. El Yaagoubi. Observer-based
approach to fault detection and isolation for nonlinear systems. IEEE
Transactions on Automatic Control, 44(10):1879–1884, 1999.
[11] S. Paoletti, A. Garulli, J. Roll, and A. Vicino. A necessary and
sufficient condition for input-output realization of switched affine state
space models. In IEEE CDC, pages 935–940, 2008.
[12] A. Abdo, S. X. Ding, J. Saijai, and W. Damlakhi. Fault detection
for switched systems based on a deterministic method. In IEEE
Conference on Decision and Control (CDC), pages 568–573, 2012.
[13] W. Pan, Y. Yuan, H. Sandberg, J. Gonçalves, and G. Stan. Online fault
diagnosis for nonlinear power systems. Automatica, 55:27–36, 2015.
[14] S. Narasimhan and G. Biswas. Model-based diagnosis of hybrid
systems. IEEE Transactions on Systems, Man, and Cybernetics, Part
A, 37(3):348–361, 2007.
[15] R. Isermann. Fault diagnosis of machines via parameter estimation
and knowledge processing–tutorial paper. Automatica, 29(4):815–835,
1993.
[16] X. Liu, H. Zhang, J. Liu, and J. Yang. Fault detection and diagnosis of
permanent-magnet dc motor based on parameter estimation and neural
network. IEEE Trans. on Ind. Electronics, 47(5):1021–1030, 2000.
[17] S. Campbell and R. Nikoukhah. Auxiliary signal design for failure
detection. Princeton University Press, 2004.
[18] J. K. Scott, R. Findeisen, R. D Braatz, and D. M. Raimondo. Input
design for guaranteed fault diagnosis using zonotopes. Automatica,
50(6):1580–1589, 2014.
[19] P. Rosa, C. Silvestre, J. S. Shamma, and M. Athans. Fault detection
and isolation of ltv systems using set-valued observers. In IEEE
Conference on Decision and Control (CDC), pages 768–773, 2010.
[20] F. Harirchi and N. Ozay. Model invalidation for switched affine
systems with applications to fault and anomaly detection. IFAC ADHS
Conference, 48(27):260–266, 2015.
[21] F. Harirchi, Z. Luo, and N. Ozay. Model (in)validation and fault detection for systems with polynomial state-space models. In American
Control Conference (ACC), pages 1017–1023, July 2016.
[22] J. Anderson and A. Papachristodoulou. On validation and invalidation
of biological models. BMC Bioinformatics, 10(1):1, 2009.
[23] N. Ozay, M. Sznaier, and C. Lagoa. Convex certificates for model
(in)validation of switched affine systems with unknown switches.
IEEE Transactions on Automatic Control, 59(11):2921–2932, 2014.
[24] H. Lou and P. Si. The distinguishability of linear control systems.
Nonlinear Analysis: Hybrid Systems, 3(1):21–38, 2009.
[25] P. Rosa and C. Silvestre. On the distinguishability of discrete linear
time-invariant dynamic systems. In IEEE CDC-ECC, pages 3356–
3361, 2011.
[26] M. Babaali and M. Egerstedt. Observability of switched linear
systems. In International Workshop on Hybrid Systems: Computation
and Control, pages 48–63. Springer, 2004.
[27] E. Beale and J. Forrest. Global optimization using special ordered
sets. Mathematical Programming, 10(1):52–69, 1976.
[28] C. Guéret, C. Prins, and M. Sevaux. Applications of optimization with
xpress-mp. contract, page 00034, 1999.
[29] F. Harirchi and N. Ozay. Guaranteed model-based fault detection in cyber-physical systems: A model invalidation approach.
arXiv:1609.05921 [math.OC], 2016.
[30] Inc. Gurobi Optimization. Gurobi optimizer reference manual, 2015.
[31] IBM ILOG CPLEX. V12. 1: User’s manual for cplex. International
Business Machines Corporation, 46(53):157, 2009.
[32] B. Argüello-Serrano and M. Vélez-Reyes. Nonlinear control of a
heating, ventilating, and air conditioning system with thermal load
estimation. IEEE Trans. on Control Systems Tech., 7(1):56–63, 1999.
[33] J. Löfberg. YALMIP: A toolbox for modeling and optimization in
MATLAB. In CACSD Conference, Taipei, Taiwan, 2004.
| 3cs.SY
|
arXiv:1712.03305v1 [math.ST] 8 Dec 2017
FALSE DISCOVERY CONTROL FOR PAIRWISE
COMPARISONS—AN ASYMPTOTIC SOLUTION TO WILLIAMS,
JONES AND TUKEY’S CONJECTURE
WEIDONG LIU, DENNIS LEUNG, AND QIMAN SHAO
Abstract. Under weak moment and asymptotic conditions, we offer an affirmative answer to whether the BH procedure (Benjamini and Hochberg, 1995)
can control the false discovery rate in testing pairwise comparisons of means
under a one-way ANOVA layout. Specifically, despite the fact that the two
sample t-statistics do not exhibit positive regression dependency (Benjamini
and Yekutieli, 2001), our result shows that the BH procedure can asymptotically control the directional false discovery rate as conjectured by Williams,
Jones, and Tukey (1999). Such a result is useful for most general situations
when the number of variables is moderately large and/or when idealistic assumptions such as normality and a balanced design are violated.
1. Introduction
Suppose we have m independent groups of observations Xi = {Xki , 1 ≤ k ≤ ni },
1 ≤ i ≤ m, where for each i, Xki ’s are independent and identically distributed
random variables with mean µi and variance σi2 . The pairwise comparison problem
(1.1)
H0ij : µi = µj
against H1ij : µi 6= µj , 1 ≤ i < j ≤ m
has been widely studied since Tukey’s early work on multiple comparisons. With the
goal of controlling the family-wise error rate (FWER), Tukey (1953) and Kramer
(1956) independently proposed their famous procedure for testing (1.1) based on
the studentized range distribution. Such developments culminated in the work of
Hayter (1984) who established the conservativeness of the Tukey-Kramer procedure
in the affirmative. In modern applications when m can be large, the number of
hypotheses q := m
2 = m(m−1)/2 to consider is even larger, making any procedure
aiming to control FWER too conservative to be useful. As a more scalable error
measure for modern large scale testing problems, the false discovery rate (FDR),
defined as the expectation of the false discovery proportion (FDP),
#hypotheses incorrectly rejected
,
#hypotheses rejected
was proposed by Benjamini and Hochberg (1995), in which it was demonstrated
that their step-up procedure (widely known as the BH procedure, or “BH” for
short) can control FDR at a pre-specified level 0 < α < 1 when the test statistics
are independent.
2000 Mathematics Subject Classification. 62H05.
1
2
W. LIU, D. LEUNG, AND Q. SHAO
However, the test statistics for the pairwise comparisons problem (1.1), namely,
the two-sample t-statistics,
(1.2)
X̄i − X̄j
, 1≤i<j≤m
Tij = q
σ̂i2 /ni + σ̂j2 /nj
Pni
Pni
where X̄i = n1i k=1
Xki is the sample mean and σ̂i2 = ni1−1 k=1
(Xki − X̄i )2 is
the sample variance, are not independent. To put it into context, let
(1.3)
pij,U = 1 − Fij (Tij ) and pij,L = Fij (Tij ),
be the one-sided upper and lower tailed (approximate) p-values for each Tij , where
Fij can be an exact or approximate cumulative distribution function for Tij under
the null µi = µj . Typically, Fij ’s are taken to be the distribution functions for the
standard normal or some t-distributions. The BH procedure would stipulate that
we first sort all the q two sided p-values, defined as
(1.4)
p|ij| = 2 min(pij,L , pij,U ) for each pair (i, j),
as their order statistics
p(1) ≤ p(2) ≤ . . . ≤ p(q) ,
and reject a hypothesis Hij whenever p|ij| ≤ p(k̂) for k̂ = max{i : p(i) ≤ iα/q}.
Conjecturing that the BH procedure can still control the FDR at level α, Williams,
Jones, and Tukey (1999) conducted an extensive simulation study to verify this
with satisfactory results, to quote their ending remark: “Each of the three authors
believes that the B-H procedure is the best available choice”. Again, the same problem has been mentioned by subsequent work of Benjamini and Yekutieli (2001,
p.1882):
“Another important open question is whether the same procedure
controls the FDR when testing pairwise comparisons of normal
means, either Studentized or not”,
in which they established the validity of the “BH” in controlling FDR under positive
regression dependence of the test statistics. Nevertheless, such a condition is also
not satisfied by the two-sample t-statistics.
In this paper, we pursue an asymptotic treatment, for the regime where m, n −→
∞ with n := maxi ni , and show that “BH” is valid for (1.1). In many applications,
m and q can be large, for example, (m, q) = (10, 45) in Oishi et al. (1998), (m, q) =
(41, 820) in Williams, et al. (1999) and (m, q) = (72, 2556) in Pawluk-Kolc et al.
(2006). Hence, it is important to prove the conjecture when q → ∞. Moreover,
the normality conditions Xki ∼ N (µi , σi2 ), alongside a balanced design, are always
assumed in the ANOVA literature . In real applications any of these assumptions
can be easily violated, and an asymptotic treatment renders us the flexibility to
address more general situations.
Williams et al. (1999) actually took the view that in reality, none of the null
hypotheses in (1.1) can be true, i.e., µi − µj can at best be close to zero to many
decimal points for any given pair (i, j). Thus, a type 1 error of “rejecting a true
null” could never occur. Nevertheless, one can still make sense of the BH procedure:
3
Following the rejection of a hypothesis Hij , depending on which of pij,U and pij,L
is smaller, a declaration of the sign of the quantity µi − µj will be made by the
practitioner. If pij,U is less than pij,L , or equivalently, Tij > 0, then sgn(µi − µj )
will be declared as positive, and vice versa. In this new context, the false discovery
rate can be modified as
#{(i, j) : sgn(µi − µj ) is incorrectly declared}
.
E
#{(i, j) : Hij is rejected}
In numerical studies where none of the µi ’s are set to be equal, Williams et al.
(1999) showed that this modified FDR can be controlled at α/2 by applying “BH”
with the nominal level α followed by sign declarations of µi − µj . This is intuitive
because in an extremely error-prone situation with all differences µi − µj being very
close to zero, every Tij stochastically behaves almost as if µi − µj = 0, in which case
upon rejecting Hij there is approximately 1/2 chance of making false declaration
about sgn(µi − µj ). Such is known as a directional error. In this paper, following
work such as Benjamini and Hochberg (2000) and Benjamini and Yekutieli (2005),
we acknowledge the possibility that some null hypotheses in (1.1) could in fact be
true, and define directional error more generally as follows:
Definition 1.1. A directional error for a pair (i, j) is made, when either sgn(µi −µj )
is incorrectly declared after rejection of Hij when µi −µj 6= 0, or the null hypothesis
Hij is rejected when µi − µj = 0.
With Definition 1.1, we formally define the directional false discovery proportion
(dFDP) as
(1.5)
dF DP =
#{(i, j) : a directional error is made for (i, j)}
#{(i, j) : Hij is rejected}
and the directional false discovery rate (dFDR) is the expectation thereof. Note
that a directional error boils down to the usual type 1 error for the two sided tests
when µi − µj = 0. Under weak conditions, our main results imply that the “BH”
followed by sign declarations can asymptotically control both the dFDP and dFDR
at (i) the usual nominal level α when all the nulls are true, and (ii) at the level α/2
as demonstrated in the simulations of Williams et al. (1999) when none of the nulls
are true.
The rest of the paper is organized as follows. Section 2 give the main results
showing Williams, Jones, and Tukey (1999)’s conjecture holds under some weak
conditions. Section 3 proves a uniform weak law of large numbers that is crucial
for our main result. Section 4 concludes with some numerical studies. Throughout,
C, c are positive constants whose values, unless otherwise specified, are understood
to be independent of (m, n), and may vary from place to place. The dependence of
m = mn on n is implicit. For two sequences of real numbers {an } and {bn }, an . bn
if there exists a constant C such that |an | ≤ C|bn | holds for all sufficiently large n.
The notation Card(A) denotes the cardinality of a set A. Φ(·) is the cumulative
distribution function of the standard normal and Φ̄(·) = 1 − Φ(·). φ is the standard
normal density. logd means taking log for d times and for two functions f and g,
f (x) ∼ g(x), x → ∞ means f (x)/g(x) −→ 1 as x goes to infinity.
4
W. LIU, D. LEUNG, AND Q. SHAO
2. Theoretical results
We formally restate Williams et al. (1999)’s version of the BH procedure with
declaration of signs for the µi − µj ’s after rejections of null hypotheses below.
Definition 2.1 (Williams et al. (1999)’s BH procedure for the testing (1.1) at level
α).
(i) Obtain the two sided p-values p|ij| according to (1.2), (1.3) and (1.4),
and sort them as the order statistics p(1) ≤ p(2) ≤ . . . ≤ p(q) , where
q = m(m − 1)/2.
(ii) Let k̂ = max{k : p(k) ≤ kα/q}, and reject Hij whenver p|ij| ≤ p(k̂) .
(iii) For each rejected Hij , declare sgn(µi − µj ) > 0 if Tij > 0, and declare
sgn(µi − µj ) < 0 if Tij < 0.
Our main result, Theorem 2.1 below, is proved under the following assumptions:
Assumption 1. There exist 0 < cL ≤ 1 ≤ cU such that for all 1 ≤ i 6= j ≤ m,
cL ≤ σi2 /σj2 , ni /nj ≤ cU .
Assumption 2. Suppose that for some constants r, C, ǫ, K > 0,
m ≤ Cnr and max E|(X1i − µi )/σi |4r+2+ǫ ≤ K
1≤i≤m
Assumption 3. For all i < j, Fij are monotone increasing distribution functions,
with the properties that Fij (t) = 1 − Fij (−t) for t ≥ 0 and
max
sup
1≤i<j≤m 0≤t≤2√log m
1 − Fij (t)
− 1 = O(n−c ) as m, n −→ ∞
Φ̄(t)
for some c > 0, where the dependence of the distribution functions Fij on (m, n1 , . . . , nm )
is suppressed for notational simplicity.
Theorem 2.1. Let H0 := {(i, j) : i < j, µi − µj = 0} and q0 = |H0 |. Under
Assumptions 1 to 3 and the signal size condition
p
|µi − µj |
(2.1)
Card (i, j) : 1 ≤ i < j ≤ m, q
≥ 8c2U log m ≥ 1,
σ 2 /n + σ 2 /n
i
i
j
j
we have
(2.2)
lim sup
m,n
and for any ǫ > 0,
(2.3)
as m, n −→ ∞.
P
α
2
dF DR
≤1
1 + qq0
α
q0
dF DP <
1+
+ ǫ −→ 1
2
q
Our theorem is an asymptotic analog to Corollary 3 and 6 in Benjamini and
Yekutieli (2005), although these previous results do not cover the types of dependency exhibited among the two sample t-statistics (1.2). One can see that the term
5
+ qq0 ) becomes α when all hypotheses are true (q0 = q) and α/2 when no hypotheses are true (q0 = 0), validating our claim in Section 1. The theorem is also
proven under very modest assumptions. Assumption 1 only requires our setup to
be approximately balanced, while Assumption 2 allows the growth of m in n to be
of a polynomial order that depends on the variables’ moments. In particular, 4r + ǫ
can in fact be less than 1, despite the commensurate limited growth rate of m in
that case. Assumption 3 says that we only need to form our p-values with reference
distribution functions that are approximately normal in a uniform way. Trivially,
it will hold when Fij ’s are taken to be Φ(·). By the Cramér-type moderate deviation for t-statistics (Jing et al., 2003), it will also hold when the Fij ’s are taken
as t-distribution functions with, say, min(ni , nj ) − 1 degrees of freedom (Scheffé,
1943), under Assumptions 1 and 2. Lastly, (2.1) is a very minimal condition that is
needed for control of the dFDP in (2.3), in particular only one (i, j) pair is required
to give a prominent signal; the reader can compare to condition (2) in Liu and Shao
(2014) and Corollary 2.2 therein to see that such condition is almost necessary for
FDP-type control.
Now we begin to prove Theorem 2.1. In fact, we only need to prove (2.3),
since it implies that lim supm,n E[dF DP ]/ α2 (1 + qq0 ) ≤ 1 + ǫ for any ǫ > 0, and
the arbitrariness of ǫ will give (2.2). These notations will be used throughout this
section and the next: Let H+ := {(i, j) : i < j, µi − µj > 0} and H− := {(i, j) :
i < j, µi − µj < 0}. Define q+ := |H+ | and q− = |H− |, and hence q = q+ + q− + q0 .
P
For notational brevity, we will also use H+ to denote a summation over all pairs
P
P
P
P
(i, j) in the set H+ , and use H− , H0 , H+ ∪H0 and H− ∪H0 similarly. Finally,
for each pair i < j, we define the centered versions of the two-sample t statistics
α
2 (1
T̄ij =
(X̄i − X̄j ) − (µi − µj )
q
.
σ̂i2 /ni + σ̂j2 /nj
Note that the rejection rule in Definition 2.1 (ii) is equivalent to the classical
BH procedure, so by Theorem 2 in Benjamini and Hochberg (1995) it is equivalent
to a procedure that rejects Hij whenever |p|ij| | < α̂, where
αX
1(p|ij| ≤α̃) .
α̂ = sup 0 ≤ α̃ ≤ 1 : α̃ ≤
q i<j
P
Since, with probability one, αq i<j 1(p|ij| ≤α̃) is right continuous as a function in α̃
on the interval [0, 1], by elementary arguments it can be shown that
αX
1(p|ij| ≤ α̂).
(2.4)
α̂ =
q i<j
If, for every ǫ > 0, we can show that, as m, n −→ ∞,
(2.5)
P
sup
√
2Φ̄( 2 log m)≤α̃≤1
P
H+
1(pij,L ≤α̃/2) +
P
H−
1(pij,U ≤α̃/2) +
α̃(q + q0 )/2
P
H0
1(p|ij| ≤α̃)
≤1+ǫ
−→p 1
!
6
W. LIU, D. LEUNG, AND Q. SHAO
and
(2.6)
p
P (2Φ̄( 2 log m) ≤ α̂ ≤ 1) −→p 1,
then (2.3) is proven since, by (2.4), the expression
P
P
P
2q[ H+ 1(pij,L ≤α̂/2) + H− 1(pij,U ≤α̂/2) + H0 1(p|ij| ≤α̂) ]
2q
P
dF DP =
α(q + q0 )
α(q + q0 ) i<j 1(p|ij| ≤α̂)
is less than 1 + ǫ with probability tending to one. (2.5) essentially amounts to
proving a uniform law of large number, which will be done in Section 3. We will
focus on showing the probabilistic bound for the cutoff point α̂ in (2.6) for the rest
of this section.
We claim that, under Assumption 1, there exists a subset Hs ⊂ {(i, j) : 1 ≤ i <
j ≤ m} such that,
p
|µi − µj |
(2.7) Card{Hs } = ⌊m/2⌋ and q
≥ 4 log m for each (i, j) ∈ Hs .
σi2 /ni + σj2 /nj
By (2.1), without loss of generality we can assume
(2.8)
p
|µ1 − µ2 |
2
p
log m.
≥
8c
U
σ12 /n1 + σ22 /n2
|µ1 −µj |
σ12 /n1 +σj2 /nj
There can only be two cases: If Card{j : 3 ≤ j ≤ m, √
√
≥ 4 log m} ≥
m/2, then (2.7) is satisfied; otherwise, we have
p
|µ1 − µj |
(2.9)
Card j : 3 ≤ j ≤ m, q
< 4 log m ≥ m/2.
σ12 /n1 + σj2 /nj
Since |µ2 − µj | ≥ |µ2 − µ1 | − |µ1 − µj |, by Assumption 1,
|µ2 − µ1 |
|µ2 − µj |
cU |µj − µ1 |
q
≥ p 2
.
−q
2
cU σ2 /n2 + σ1 /n1
σ22 /n2 + σj2 /nj
σj2 /nj + σ12 /n1
Recalling that cU ≥ 1, together with (2.8) and (2.9) we have
p
|µ2 − µj |
Card j : 3 ≤ j ≤ m, q
≥ 4 log m ≥ m/2
σ22 /n2 + σj2 /nj
and hence (2.7) also holds. In Appendix A.1 we will also prove that
σ̂i2
−2
− 1 > (log m)
= O(n−r−c ).
(2.10)
P max
1≤i≤m σi2
for some c > 0, which, together with (2.7), gives
p
|µ
−
µ
|
i
j
≥ 3.9 log m −→ 1.
(2.11)
P min q
(i,j)∈Hs
σ̂i2 /ni + σ̂j2 /nj
7
On the other hand, by the Cramér-type moderate deviation for two-sample tstatistics (see Jing et al. (2003), also see Cao (2007)[Theorem 1.2] with a slight
change to the moment conditions), under Assumptions 1 and 2 we have
4r+2+ǫ
1+s
P (T̄ij ≥ s)
= 1 + O(1)
(2.12)
Φ̄(s)
n1/2−1/(4r+2+ǫ)
uniformly in s ∈ [0, o(n1/2−1/(4r+2+ǫ) )) and 1 ≤ i < j ≤ m. A union bound then
implies
p
p
(2.13) P
max T̄ij ≥ 2 log m ≤ ⌊m/2⌋(2 + o(1))Φ̄( 2 log m)
(i,j)∈Hs
= O((log m)−1/2 ) −→ 0
using the fact that Φ̄(x) ≤ (2π)−1/2 x−1 exp(−x2 /2) for all x > 0.
In view of (2.11) and (2.13), we have that
X
P
1(|Tij |≥2√log m) ≥ ⌊m/2⌋ −→ 1.
1≤i<j≤m
√
By Assumption 3, |Tij | ≥ 2 log m implies p|ij| = o(m−2 ), so recognizing q = O(m2 )
we have
X
P
1(p|ij| ≤1/q) ≥ ⌊m/2⌋ → 1,
1≤i<j≤m
which in turn gives
p
α
2Φ̄( 2 log m) ≤
P
P
1≤i<j≤m
1(p|ij| ≤2Φ̄(√2 log m))
q
!
→1
√
since 2Φ̄( 2 log m) = O((log m)−1/2 m−1 ). By the definition of α̂ this implies (2.6).
3. Proof of the uniform law of large numbers
This section proves (2.5). We claim that it suffices to show
(3.1)
P
P
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥t)
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥t) +
− 1 −→p 0,
sup
√
Φ̄(t)(q + q0 )
0≤t≤ 2 log m
where p̄ij,L = Fij (T̄ij ) and p̄ij,U = 1 − Fij (T̄ij ). This can be easily seen to be true
by observing that for each (i, j), the events
{pij,L ≤ α̃/2},
{pij,U ≤ α̃/2} and {p|ij| ≤ α̃}
are identical to
{Φ̄−1 (pij,L ) ≥ Φ̄−1 (α̃/2)}, {Φ̄−1(pij,U ) ≥ Φ̄−1 (α̃/2)} and {Φ̄−1 (p|ij| ) ≥ Φ̄−1 (α̃)}
8
W. LIU, D. LEUNG, AND Q. SHAO
respectively, and that p̄ij,L ≤ pij,L for (i, j) ∈ H+ ∪ H0 , p̄ij,U ≤ pij,U for (i, j) ∈
H− ∪ H0 .
√
Let 0 = t0 < t1 < ... < tg = 2 log m satisfying tl − tl−1 = vm for 1 ≤ l ≤ g − 1
√
and tg − tg−1 ≤ vm , where vm = ( 2 log m log4 m)−1 . By the mean value theorem,
for some c ∈ (tl−1 , tl ),
(3.2)
Φ̄(tl )
Φ̄(tl−1 ) − vm φ(c)
=
= 1 + o(1).
Φ̄(tl−1 )
Φ̄(tl−1 )
The o(1) term in the above display is uniform in all 0 ≤ l ≤ g − 1, and it comes
from the fact that Φ̄(tl−1 ) ∼ t−1
l−1 φ(tl−1 ) as tl−1 −→ ∞ and hence
vm φ(c)/Φ̄(tl−1 ) ∼ vm tl−1 φ(c)/φ(tl−1 )
which goes to 0 uniformly for all l = 0, . . . , g − 1 since vm tl−1 ≤ (log4 m)−1 . For
any tl−1 ≤ t ≤ tl , note that
(3.3)
P
P
P
H+ 1(Φ̄−1 (p̄ij,L )≥t)
H+ 1(Φ̄−1 (p̄ij,L )≥tl−1 ) Φ̄(tl−1 )
H+ 1(Φ̄−1 (p̄ij,L )≥tl ) Φ̄(tl )
≤
≤
Φ̄(tl )(q + q0 )
Φ̄(tl−1 )
Φ̄(t)(q + q0 )
Φ̄(tl−1 )(q + q0 )
Φ̄(tl )
while the preceding display also hold if we replace H+ with H− /H0 , or p̄ij,L with
p̄ij,U . This, together with (3.2), implies that for proving (3.1) it suffices to show
P
P
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥tl )
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥tl ) +
(3.4)
max
− 1 −→p 0.
0≤l≤g
Φ̄(tl )(q + q0 )
By integral comparison (Appendix B.3), for any ǫ > 0,
!
P
P
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥tl ) +
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥tl )
(3.5) P max
−1 >ǫ ≤
0≤l≤g
Φ̄(tl )(q + q0 )
!
P
P
Z tg
1
1
+
−1 (p̄
−1 (p̄
(
Φ̄
)≥t)
(
Φ̄
)≥t)
H
∪H
H
∪H
ij,U
ij,L
−
0
+
0
−1
P
− 1 > ǫ/2 dt+
vm
Φ̄(t)(q + q0 )
0
!
P
P
X
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥tl )
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥tl ) +
P
−1 >ǫ
Φ̄(tl )(q + q0 )
l=0,g
√
In fact for each t ∈ [0, 2 log m], for large enough m the probabilities in the display
above satisfy
!
P
P
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥t)
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥t) +
−1 >ǫ =
(3.6) P
Φ̄(t)(q + q0 )
!
P
P
P
H− 1(T̄ij ≥tij )
H+ 1(−T̄ij ≥tij ) +
H0 1(|T̄ij |≥tij ) +
−1 >ǫ ,
P
Φ̄(t)(q + q0 )
where
tij = Φ̄−1 (Φ̄(t)(1 + ǫ(T̄ij ))),
for some numbers ǫ(T̄ij ) that may depend on T̄ij but have the property that for a
sequence ǫn > 0,
(3.7)
|ǫ(T̄ij )| ≤ ǫn for all (i, j) a.s.
and ǫn = O(n−c ) for some c > 0.
9
The argument leading to (3.6) is a bit delicate and is deferred to Appendix B.1. To
finish the proof we need the following lemma, where we define
Tij∗ =
(X̄i − µi ) − (X̄j − µj )
q
.
σi2 /ni + σj2 /nj
Lemma 3.1. Under our assumptions, we have for large enough n,
(i)
(ii)
P (T̄ij ≥ tij )
= 1 + O(n−c )
Φ̄(t)
√
uniformly in 0 ≤ t ≤ 2 log m and all (i, j), and
P Ti∗1 j1 ≥ (1 − (log m)−2 )ti1 j1 , Ti∗2 j2 ≥ (1 − (log m)−2 )ti2 j2
≤ C(1 + t)−2 exp(−t2 /(1 + δ)) + O(n−c )
√
uniformly in 0 ≤ t ≤ 2 log m and all (i1 , j1 ), (i2 , j2 ) such that |{i1 , j1 } ∩
{i2 , j2 }| = 1, for some constant 1 > δ > 0.
The proof of the lemma is given in Appendix B.2. Continuing from (3.6), we get
√
that for 0 ≤ t ≤ 2 log m and suffciently large n,
!
P
P
H+ ∪H0 1(Φ̄−1 (p̄ij,L )≥t) +
H− ∪H0 1(Φ̄−1 (p̄ij,U )≥t)
P
−1 >ǫ
Φ̄(t)(q + q0 )
(3.8)
.
(
X
(i,j)∈H0 ∪H+
(i′ ,j ′ )∈H0 ∪H−
|{i,j}∩{i′ ,j ′ }|=1
P −Tij∗ ≥ 1 − (log m)−2 tij , Ti∗′ j ′ ≥ 1 − (log m)−2 ti′ j ′
(3.9)
+
X
2
(i,j)∈H0 ∪H+
(i′ ,j ′ )∈H0 ∪H−
|{i,j}∩{i′ ,j ′ }|=0
(i,j)∈H0
P |T̄ij | ≥ tij
)
(ǫΦ̄(t)(q + q0 ))2 + O(m−1 Φ̄(t)−2 n−r−c )
(3.10)
.m
E[(1(−T̄ij ≥tij ,T̄i′ j′ ≥ti′ j′ ) − Φ̄(t)) ] +
X
−2
−1
(Φ̄(t))
+m
−1
−1 2
exp(δ(1 + δ)
t )+O
log m
−c
.
+n
nc
An application of Chebyshev’s inequality to (3.6), together with (2.10), will give
us the first inequality. We obtain the second inequality in the following way: Since
√
Φ̄(t) ∼ t−1 φ(t) as t → ∞, Φ̄( 2 log m) = (4π log m)−1/2 m−1 (1 + o(1)), and hence
the last “big-O” term in (3.9) becomes O((log m)mn−r−c ) = O((log m)n−c ) by
Assumption 2. Application of Lemma 3.1(i) to the two terms on (3.9) give the
terms m−2 (Φ̄(t))−1 and O(n−c ). Lastly, applying Lemma 3.1(ii) to the sum in
(3.8), together with the fact that (1 + t)−1 φ(t2 ) . Φ̄(t) for all t ≥ 0, gives the terms
10
W. LIU, D. LEUNG, AND Q. SHAO
m−1 exp(δ(1 + δ)−1 t2 ) and O(m−1 n−c ), the later of which can be absorbed into
O(n−c ).
√
Putting t = 2 log m in (3.10), up to constants, we can bound (3.5) by
√
δ−1
2 log m
(log m)
−c
(2 log m log4 m)
+
n
+ m 1+δ + O
m
nc
which converges to zero by Assumption 2, and (3.4) is proved.
4. Discussion and Simulation Study
In an asymptotic setting with minimal assumptions, we established, as conjectured in Williams, Jones, and Tukey (1999), that the BH procedure can control
the false discovery proportion (FDP) and the false discovery rate (FDR) for the
pairwise comparison problem, where inferential errors are defined in the directional
sense. While it is widely believed to be true that, for Gaussian test statistics with
arbitrary dependence structure, the BH procedure can control FDR at the nominal
level in the finite sample setting, proving it in the affirmative remains a hard open
problem in mathematical statistics. For most practical purposes, our treatment is
flexible enough that it established the asymptotic conservativeness of “BH” in the
pairwise comparison setting even when the data can be heavy-tailed or the ANOVA
design can be unbalanced. It is conceivable the techniques in this paper can also be
used to prove the conservativeness of “BH” when it is applied for pairwise comparisons in large scale genomic studies (Yekutieli, 2008). It also came to our attention
that finite sample results on F DP control has appeared in recent literature (Li and
Barber, 2017, Theorem 1), and it will be interesting to explore how their proof
techniques will play out when some dependence exists among the test statistics
considered.
Here we conduct a numeric experiment that parallels our results in Theorem 2.1.
For simplicity, we only consider balanced setups where the same number of samples n are observed for each variable. For each experiment we first generate m
means µ1 , . . . , µm that are i.i.d. realizations of a mean-zero normal distribution
N (0, effect size), where effect size is a chosen value for the standard deviation,
given such a name since the larger effect size is, the more dispersed are the values
µ1 , . . . , µm and hence the pairwise differences µi − µj , 1 ≤ i < j ≤ m will generally
be larger. Notice that we are in the scenario described by Williams, Jones, and
Tukey (1999) where none of the differences µi − µj can be exactly zero. Then for
each 1 ≤ i ≤ m, we generate n i.i.d t-distributed error terms ǫki , 1 ≤ k ≤ n with 6
degrees of freedom, and let our observations be Xki = µi + ρi ǫki , where ρ2i are m
numbers generated from N (mean = 1, sd = .1) to give the Xi ’s different variances.
At level α = 0.2, the BH-procedure in Definition 2.1 is applied using Φ to calibrate
the p-values, and each experiment is repeated 500 times for different combinations
of m, n and effect size. The estimated probability of dFDP being less than α/2
and estimated FDR are reported in Tables 1 and 2. Evidently in Tables 1 , the
estimated P (dF DP ≤ α/2) tends to becomes close to one as m and n grow. Moreover, the higher the signal size, the larger are these probabilities, pointing to the
11
fact that condition (2.1) is required to guaranteed control of false discovery proportions. On the other hand, false discovery rate is always controlled for all effect sizes
in Table 2, and when effect size is smaller, i.e. when it is more easy to commit a
directional error, the estimated dFDR’s are closer to the upper bound α/2 = 0.1,
as expected.
Table 1. Estimation of P (dF DP ≤ α/2) based on 500 repetitions
of experiments, at level α = 0.2
m\n
5
15
30
40
40
100 200 400
effect size = 0.01
0.88 0.90 0.91 0.90
0.87 0.91 0.90 0.89
0.88 0.90 0.91 0.90
0.88 0.89 0.91 0.90
40
100 200 400
effect size = 0.05
0.90 0.93 0.93 0.95
0.89 0.93 0.91 0.89
0.87 0.92 0.91 0.90
0.89 0.90 0.92 0.93
5
15
30
40
effect size = 0.15
0.92 0.93 0.95 0.97
0.92 0.96 1.00 1.00
0.88 0.97 1.00 1.00
0.91 0.99 1.00 1.00
effect size = 0.25
0.93 0.94 0.96 0.99
0.98 1.00 1.00 1.00
0.95 1.00 1.00 1.00
0.99 1.00 1.00 1.00
Table 2. Estimation of dF DR based on 500 repetitions of experiments, at level α = 0.2
m\n
5
15
30
40
40
100 200 400
effect size = 0.01
0.08 0.07 0.06 0.06
0.08 0.05 0.06 0.07
0.07 0.06 0.05 0.05
0.08 0.06 0.05 0.05
40
100 200 400
effect size = 0.05
0.05 0.04 0.02 0.02
0.05 0.03 0.03 0.03
0.06 0.03 0.03 0.03
0.05 0.03 0.02 0.02
5
15
20
40
effect size = 0.15
0.03 0.02 0.01 0.00
0.02 0.02 0.01 0.01
0.03 0.02 0.02 0.02
0.03 0.02 0.02 0.01
effect size = 0.25
0.01 0.01 0.00 0.00
0.02 0.01 0.01 0.01
0.02 0.02 0.01 0.01
0.02 0.02 0.01 0.01
12
W. LIU, D. LEUNG, AND Q. SHAO
Appendix A. Supporting lemmas
A.1. Proof of probability bound in (2.10). Note that for all 1 ≤ i ≤ m,
σ̂i2
=
σi2
Pn
− X̄i )2
=
(n − 1)σi2
k=1 (Xki
Pn
− µi )2
n(X̄i − µi )2
−
,
2
(n − 1)σi
(n − 1)σi2
k=1 (Xki
which in turn gives
σ̂i2
−1 ≤
σi2
Pn
− µi )2 − σi2 ]
n
n(X̄i − µi )2
+
.
−
1
+
(n − 1)σi2
n−1
(n − 1)σi2
k=1 [(Xki
Hence for sufficiently large n, given that m ≤ cnr we have
σ̂i2
−2
(A.1) P max 2 − 1 > (log m)
i
σi
!
r
n
X
Xki − µi
n(n − 1)
≤ P max
(log m)−1 +
>
i
σi
2
k=1
!
n
X
(log m)−2 (n − 1)
(Xki − µi )2
.
−1 >
P max
i
σi2
2
k=1
It suffices to show the two terms on the right hand side of (A.1) is of that order.
The first term, for large enough n, under Assumption 2 can be bounded as
P
max
1≤i≤m
n
X
Xki − µi
σi
k=1
(
≤m P
!
n(n − 1)
−1
(log m)
>
2
!)
r
n
X
n(n − 1)
Xki − µi
−1
>
(log m)
σi
2
r
k=1
.m
n1+2r+ǫ/2 (log m)2+4r+ǫ
. n−r−c ,
n2+4r+ǫ
for some c > 0, where the second last inequality follows from Rosenthal’s inequality
(Rosenthal, 1970, Lemma 1). For the second term, we once again have
(A.2) P
max
i
n
X
(Xki − µi )2
k=1
σi2
mP
!
(log m)−2 (n − 1)
≤
−1 >
2
!
n
X
(log m)−2 (n − 1)
(Xki − µi )2
.
−1 >
σi2
2
k=1
If 2r+1+ǫ/2 ≤ 2, by von Bahr and Esseen (1965, Theorem 2), E[|
Pn
k=1 (
(Xki −µi )2
−
σi2
1)|2r+1+ǫ/2 ] . n, and hence with the right hand side of (A.2) can be bounded by
m(log m)4r+2+ǫ n−2r−ǫ/2 = O(n−r−c ). If 2r + 1 + ǫ/2 ≥ 2, then we can apply Liu
13
et al. (2008)[Lemma 6.1, (6.2)] to give
!
n
X
(Xki − µi )2
(log m)−2 (n − 1)
P
−1 >
σi2
2
k=1
n
X
n(n − 1)
c(n − 1)
(Xki − µi )2
+
exp
−
+ Cn− max(2r+ǫ,2)
−
1
>
≤
P
σi2
2(log m)2
c(log m)2
k=1
(log m)2r+1+ǫ/2
n(n − 1)
+ Cn− max(2r+ǫ,2) . n−2r−c1
.
+ exp −
c(log m)2
n2r+ǫ/2
for some small enough c, c1 > 0 and large enough C > 0, which implies the right
hand side of (A.2) can be bounded by n−r−c1 .
Appendix B. Proof for Section 3
We first introduce a lemma that will be used to finish the proofs.
√
Lemma B.1. Under our assumptions, for any constant c > 2 and a sequence
√
bn = o(1), it must be true for all t ∈ [0, 2 log m]
p
Φ̄−1 (Φ̄(t)(1 + bn )) ≤ c log m
for a sufficiently large n.
Proof. Let t̃ = Φ̄−1 (Φ̄(t)(1 + bn )). It must be true that t̃ ≤ am where am =
√
Φ̄−1 (Φ̄( 2 log m)(1 + bn )). Since sΦ̄(s)/φ(s) → 1 as s −→ ∞, it must be that
√
φ(am ) 2 log m
√
= (1 + o(1))(1 + bn ) = 1 + o(1).
am φ( 2 log m)
By taking log on both side of the preceding display it is easy to see
q
p
p
t̃ ≤ am ≤ 2 log m + 2 log( 2 log m) + o(1) ≤ c log m
fo large enough n (and hence m).
B.1. Proof for (3.6). We will first show that for a sufficiently large n, the equivalence of the events
p̄ij,U ≤ Φ̄(t) ⇐⇒ T̄ij ≥ Φ̄−1 (Φ̄(t)(1 + ǫT̄ij ))
with ǫT̄ij having the property in (3.7). To show this we first note the following two
events are identical, by their definitions:
(B.1)
{p̄ij,L ≤ Φ̄(t)} = {1 − Fij (−T̄ij ) ≤ Φ̄(t)}.
√
For a fixed constant 2 < c′ < 2, for a sufficiently large n the event in (B.1) can
√
√
be realized for all 0 ≤ t ≤ 2 log m if −T̄ij ≥ c′ log m since, with Assumption 3,
p
p
p
1 − Fij (−T̄ij ) ≤ 1 − Fij (c′ log m) = Φ̄(c′ log m)(1 + o(1)) ≤ Φ̄( 2 log m) ≤ Φ̄(t)
when m is large enough, using the fact that Φ(s) ∼ s−1 φ(s) as s → ∞. But since
√
the event in (B.1) may also be true for some −T̄ij < c′ log m, using Assumption 3
14
W. LIU, D. LEUNG, AND Q. SHAO
√
again and the identity (B.1) we can conclude, conditioning on −T̄ij < c′ log m,
the equivalence of events
p̄ij,L ≤ Φ̄(t) ⇐⇒
1 − Fij (−T̄ij )
Φ̄(−T̄ij ) ≤ Φ̄(t)
Φ̄(−T̄ij )
⇐⇒ Φ̄(−T̄ij )(1 + ǫT̄ij ) ≤ Φ̄(t) ⇐⇒ −T̄ij ≥ Φ̄−1 (Φ̄(t)(1 + ǫT̄ij ).
for large enough n, where ǫT̄ij has the property in (3.7). However, in light of
√
Lemma B.1, for large enough n, −T̄ij ≥ c′ log m will necessarily imply −T̄ij ≥
Φ̄−1 (Φ̄(t)(1 + ǫT̄ij ), and hence the train of equivalence in the preceding display is
√
also true without conditioning on −T̄ij < c′ log m. By a completely analogous
argument we also have the equivalence of events
p̄ij,U ≤ Φ̄(t) ⇐⇒ T̄ij ≥ Φ̄−1 (Φ̄(t)(1 + ǫT̄ij ))
for large enough n with same property for ǫT̄ij .
B.2. Proof of Lemma 3.1. For notational simplicity we let tU = Φ̄−1 (Φ̄(t)(1 −
ǫn )) and tL = Φ̄−1 (Φ̄(t)(1 + ǫn )), where the subscripts are suggestive of the fact
that tL is always less than tU . We also note that
p
(B.2)
tL ≤ tij ≤ tU for all (i, j) and t ∈ [0, 2 log m].
Proof of Lemma 3.1(i). In fact, it suffices to show that for sufficiently large m,
(B.3)
P (T̄ij ≥ tU )
= 1 + O(n−c )
Φ̄(t)
(B.4)
P (T̄ij ≥ tL )
= 1 + O(n−c )
Φ̄(t)
√
both uniformly in t ∈ [0, 2 log m] and all (i, j), in light of (B.2).
(B.3): By Lemma B.1, we note that
√ for large enough n, it is true that tL , tU ≤
√
c′ log m for a fixed constant c′ > 2. Since it must be that 0 ≤ tU , by Assumption 2 and the Cramér-type moderate deviation (2.12), we have
4r+2+ǫ !
P (T̄ij ≥ tU )
1 + tU
= (1 − ǫn ) 1 + O(1)
Φ̄(t)
n1/2−1/(4r+2+ǫ)
= 1 + O(n−c )
√
uniformly in all (i, j) and 0 ≤ t ≤ 2 log m, for sufficiently large n.
(B.4): Since tL can be a negative number, (2.12) is not directly applicable.
√
However, we can do a separation argument: For any t ∈ [1, 2 log m], for large
enough n it must be that tL ≥ 0, hence with the same argument for proving (B.3),
we have
p
P (T̄ij ≥ tL )
= 1 + O(n−c ) uniformly in t ∈ [1, 2 log m] and all (i, j).
Φ̄(t)
15
For any t ∈ [0, 1), we have
P (T̄ij ≥ tL )
P (T̄ij ≥ t) P (T̄ij ≥ tL ) − P (T̄ij ≥ t)
=
+
Φ̄(t)
Φ̄(t)
Φ̄(t)
P (T̄ij ≥ tL ) − P (T̄ij ≥ t)
= 1 + O(n−c ) +
Φ̄(t)
(B.5)
uniformly in (i, j) and t ∈ [0, 1), where the last equality is based on Assumption 2
and (2.12). But we also note that
(B.6)
|P (T̄ij ≥ tL ) − P (T̄ij ≥ t)|
≤
Φ̄(t)
|P (T̄ij ≥ tL ) − Φ̄(tL )| |P (T̄ij ≥ t) − Φ̄(t)| |Φ̄(t) − Φ̄(tL )|
+
+
.
Φ̄(t)
Φ̄(t)
Φ̄(t)
Given that Φ̄(t) > C > 0 for some C on interval [0, 1), on the right hand side of
(B.6) the first two terms are of order O(n−c ) by (2.12), and the last term is so due
to bounded first derivative of Φ̄−1 on the interval [Φ̄(1), Φ̄(−1)], all uniformly in
(i, j) and t ∈ [0, 1).
Proof of Lemma 3.1(ii). Given (B.2), it suffices to show, for sufficiently large n,
P Ti∗1 j1 ≥ t̃, Ti∗2 j2 ≥ t̃ ≤ C(1 + t)−2 exp(−t2 /(1 + δ))
√
uniformly in 0 ≤ t ≤ 2 log m and |{i1 , j1 } ∩ {i2 , j2 }| = 1 for some constant δ > 0,
where t̃ := (1 − (log m)−2 )tL .
We first recognize that for any fixed pair i 6= j, Tij∗ can be rewritten as the
standardized sum
Pnj
Pni
Pni ∨nj
−1
ηk,ij
k=1 ni nj (Xkj − µj )
k=1 (Xki − µi ) −
∗
q
Tij =
,
= k=1
√
aij
ni σi2 + n2i σj2 /nj
where
ηk,ij :=
(
[(Xki − µi ) − ni nj−1 (Xkj − µj )]/σi
[(Xki − µi )1{ni >nj } −
ni n−1
j (Xkj
− µj )1{ni <nj } ]/σi
if k ≤ min(ni , nj )
if k > min(ni , nj )
,
1 ≤ k ≤ ni ∨ nj , are independent random variables with mean 0, and aij :=
ni + (n2i σj2 )/(nj σi2 ).
Without loss of generality we will prove the lemma by assuming that i1 = i2 ,
and make the identification i1 = i2 = i, j1 = j, j2 = l with three distinct indices
i, j, l from now on. Denote the Euclidean norm on the plane with k · k, and for each
√
√
k = 1, . . . , ni ∨ nj ∨ nl , let ηk = (ηk,ij 1(k≤ni ∨nj ) / aij , ηk,il 1(k≤ni ∨nl ) / ail )′ and
Pni ∨nj ∨nl
Σ = Cov( k=1
η k η ′k ). We suppressed the dependence of η k on the particular
choice of the triple (i, j, l) for brevity. Later we will use the properties that
(B.7)
E[η k ] = (0, 0)′ ,
(B.8)
Σ is a 2 × 2 matrix with 1’s on the diagonal
16
W. LIU, D. LEUNG, AND Q. SHAO
and
(B.9)
(1+c2U )−1 ≤ Σ1,2
v
!−1 s
u
−1
2n
u
σ
σl2 ni
i
j
t
∗
∗
≤ (1+c2L )−1
= Cov(Tij , Til ) =
1+ 2
1+ 2
σi nj
σi nl
under Assumption 1. For 1 ≤ k ≤ ni ∨ nj ∨ nl , by defining the the truncations
η̂ k = (η̂k,ij , η̂k,il )′ := η k 1(kηk k≤(log m)−4 ) − E[η k 1(kηk k≤(log m)−4 ) ]
and
η̃ k = ηk − η̂ k = η k 1(kηk k>(log m)−4 ) − E[η k 1(kηk k>(log m)−4 ) ]
(again the reliance of η̃ k and η̂ k on (i, j, l) is implicit), we get that
!
ni ∨nj ∨nl
X
∗
∗
−2
(B.10) P Tij ≥ t̃, Til ≥ t̃ ≤ P k
η̃k k ≥ (log m)
+
k=1
P
ni ∨nj
X
k=1
−2
η̂k,ij ≥ t̃ − (log m)
,
nX
i ∨nl
k=1
−2
η̂k,il ≥ t̃ − (log m)
!
.
Then, by Assumptions 2 and 1, there exists some constant κ, such that
max E[kη k k4r+2+ǫ ] ≤ κ/n2r+1+ǫ/2,
k,(i,j,l)
and hence
(B.11)
(B.12)
(B.13)
(B.14)
(B.15)
E
≤
≤
"ni ∨nj ∨nl
X
k=1
ni ∨nj ∨nl
X
k=1
ni ∨nj ∨nl
X
η k 1(kηk k>(log m)−4 )
#
E[kη k k1(kηk k>(log m)−4 ) ]
−4
(log m)
k=1
E
"
ηk
(log m)−4
4r+2+ǫ
#
≤ κn−2r−ǫ/2 (log m)16r+4+4ǫ
= o((log m)−2 ).
P
uniformly for all (i, j, l). Hence for a sufficiently large n, in order for k k η̃ k k to
be greater than (log m)−2 , from the definition of η̃ k it can be seen that at least one
of η k 1(kηk k>(log m)−4 ) , k = 1, . . . , ni ∨ nj ∨ nl , must be non-zero, which gives
!
ni ∨nj ∨nl
X
−2
P
η̃k > (log m)
≤ n max P (kη k k > (log m)−4 )
k
k=1
≤ κ(log m)16r+8+4ǫ n−2r−ǫ/2 = O(n−c ),
again uniformly for all (i, j, l). From (B.10) it remains to show
!
ni ∨nj
nX
i ∨nl
X
−2
−2
≤ C(t+1)−2 exp(−t2 /(1+δ))
P
η̂k,ij ≥ t̃ − (log m) ,
η̂k,il ≥ t̃ − (log m)
k=1
k=1
√
uniformly in 0 ≤ t ≤ 2 log m for some δ > 0.
17
Since kη̂k k are bounded by 2(log m)−4 , it satisfies the condition of Bernstein’s
inequality, by applying Zaı̆tsev (1987, Theorem 1.1) (Taking τ = 2(log m)−4 for τ
in that theorem), we have
!
X
X
−2
−2
P
η̂k,ij ≥ t̃ − (log m) ,
η̂k,il ≥ t̃ − (log m)
k
k
(log m)2
≤ P W1 > t̃ − 2(log m)−2 , W2 > t̃ − 2(log m)−2 + c1 exp −
c2
for some absolute constants c1 , c2 > 0, where (W1 , W2 )′ is a bivariate Gaussian
Pn1 ∨nj ∨nl
vector with mean zero
structure Σ̂ := E[ k=1
η̂ k η̂ ′k ]. Note
and covariance
(log m)2
is less than C(1 + t)−2 exp(−t2 /(1 + δ)) for some
that the term c1 exp − c2
√
C, δ > 0 on the range t ∈ [0, 2 log m], and we are left with bounding the first term
on the right hand side of the preceding display.
By (B.7), (B.15) and Assumption 1,
"ni ∨nj ∨nl
#
X
E
η k 1(kηk k≤(log m)−4 ) = o((log m)−2 ),
k=1
hence, by letting k · k∞ denote the matrix max norm,
#
"ni ∨nj ∨nl
X
′
kΣ − Σ̂k∞ ≤ E
ηk η k 1(kηk k>(log m)−4 )
k=1
≤
(B.16)
ni ∨nj ∨nl
X
k=1
+ o((log m)−4 )
∞
E[kηk k2 1(kηk k>(log m)−4 ) ] + o((log m)−4 )
≤ (log m)16r+4ǫ
κ
+ o((log m)−4 ) ≤ o((log m)−2 ),
n2r+ǫ/2
again uniformly in all (i, j, l). By Berman (1962, Lemma 2), for 0 ≤ t ≤
and large enough n,
(B.17) P W1 > t̃ − 2(log m)−2 , W2 > t̃ − 2(log m)−2 ≤
min
C
2 exp −
−2
−2
√ m) , t̃−2(log
√ m)
1 + min t̃−2(log
Σ̂11
Σ̂22
t̃−2(log m)−2
√
Σ̂11
,
√
2 log m
t̃−2(log m)−2
1 + √ Σ̂12
√
Σ̂11 Σ̂22
Σ̂22
2
.
By the bound in (B.16), (B.8), (B.9) and elementary calculations, the right hand
side can further be bounded by
C
exp(−t2 /(1 + δ))
(1 + t)2
for some other constants C > 0 and 1 > δ > 0, uniformly in (i, j, l) and 0 ≤ t ≤
√
2 log m. These elementary calculations are left to the readers, but the fact that
√
tL = Φ̄−1 (Φ̄(t)(1 + ǫn )) converges to t uniformly in t ∈ [0, 2 log m] is helpful: It
is obvious how tL → t uniformly for t ∈ [0, 2], since Φ̄−1 has bounded derivative
18
W. LIU, D. LEUNG, AND Q. SHAO
√
on [Φ̄(2), Φ̄(0)]. On the interval t ∈ [2, 2 log m], by inverse function theorem and
with mean value theorem,
ǫn
|tL − t| ≤ |φ(t)|−1 |Φ(t)|ǫn ≤
→0
t
uniformly since ǫn = O(n−c ).
B.3. Comparison argument. By a union bound,
(B.18)
P
≤
max
0≤l≤g
X
P
H+ ∪H0
I(Φ̄−1 (p̄ij,L ) ≥ tl ) +
P
H− ∪H0
I(Φ̄−1 (p̄ij,U ) ≥ tl )
Φ̄(tl )(q + q0 )
P
P
−1
(p̄ij,L ) ≥ tl ) + H− ∪H0 I(Φ̄−1 (p̄ij,U ) ≥ tl )
H+ ∪H0 I(Φ̄
Φ̄(tl )(q + q0 )
P
P
−1
(p̄ij,L ) ≥ tl ) + H− ∪H0 I(Φ̄−1 (p̄ij,U ) ≥ tl )
H+ ∪H0 I(Φ̄
0≤l≤g
X
P
P
Φ̄(tl )(q + q0 )
0≤l≤g
−1 >ǫ
!
!
>1+ǫ +
!
<1−ǫ .
By (3.3), for large enough n,
(B.19)
X
P
1≤l≤g−1
≤
−1
vm
Z
tg
P
0
P
H+ ∪H0
I(Φ̄−1 (p̄ij,L ) ≥ tl ) +
P
H− ∪H0
I(Φ̄−1 (p̄ij,U ) ≥ tl )
Φ̄(tl )(q + q0 )
P
P
−1
(p̄ij,L ) ≥ t) + H− ∪H0 I(Φ̄−1 (p̄ij,U ) ≥ t)
H+ ∪H0 I(Φ̄
Φ̄(t)(q + q0 )
>1+ǫ
!
!
> 1 + ǫ/2 dt
and similarly,
(B.20)
X
P
1≤l≤g−1
≤
−1
vm
Z
tg
P
0
P
H+ ∪H0
I(Φ̄−1 (p̄ij,L ) ≥ tl ) +
P
H− ∪H0
I(Φ̄−1 (p̄ij,U ) ≥ tl )
Φ̄(tl )(q + q0 )
P
P
−1
(p̄ij,L ) ≥ t) + H− ∪H0 I(Φ̄−1 (p̄ij,U ) ≥ t)
H+ ∪H0 I(Φ̄
Φ̄(t)(q + q0 )
<1−ǫ
!
!
< 1 − ǫ/2 dt.
References
Benjamini, Y. and Hochberg, Y. (1995). “Controlling the false discovery rate: a practical
and powerful approach to multiple testing.” J. Roy. Statist. Soc. Ser. B , 57(1): 289–300.
— (2000). “On the adaptive control of the false discovery rate in multiple testing with
independent statistics.” Journal of educational and Behavioral Statistics, 25(1): 60–83.
Benjamini, Y. and Yekutieli, D. (2001). “The control of the false discovery rate in multiple
testing under dependency.” Ann. Statist., 29(4): 1165–1188.
— (2005). “False discovery rate-adjusted multiple confidence intervals for selected parameters.” J. Amer. Statist. Assoc., 100(469): 71–93. With comments and a rejoinder by
the authors.
Berman, S. M. (1962). “A law of large numbers for the maximum in a stationary Gaussian
sequence.” Ann. Math. Statist., 33: 93–97.
19
Cao, H. (2007). “Moderate deviations for two sample t-statistics.” ESAIM Probab. Stat.,
11: 264–271.
Hayter, A. J. (1984). “A proof of the conjecture that the Tukey-Kramer multiple comparisons procedure is conservative.” Ann. Statist., 12(1): 61–75.
Jing, B.-Y., Shao, Q.-M., and Wang, Q. (2003). “Self-normalized Cramér-type large
deviations for independent random variables.” Ann. Probab., 31(4): 2167–2215.
Kramer, C. Y. (1956). “Extension of multiple range tests to group means with unequal
numbers of replications.” Biometrics, 12: 307–310.
Li, A. and Barber, R. F. (2017). “Accumulation tests for FDR control in ordered hypothesis testing.” J. Amer. Statist. Assoc., 112(518): 837–849.
Liu, W. and Shao, Q.-M. (2014). “Phase transition and regularized bootstrap in large-scale
t-tests with false discovery rate control.” Ann. Statist., 42(5): 2003–2025.
Liu, W.-D., Lin, Z., and Shao, Q.-M. (2008). “The asymptotic distribution and BerryEsseen bound of a new test for independence in high dimension with an application to
stochastic optimization.” Ann. Appl. Probab., 18(6): 2337–2366.
Oishi, S., Schimmack, U., Diener, E., and Suh, E. M. (1998). “The measurement of values and individualism-collectivism.” Personality and social psychology bulletin, 24(11):
1177–1189.
Pawluk-Kolc, M., Zieba-Palus, J., and Parczewski, A. (2006). “Application of false discovery rate procedure to pairwise comparisons of refractive index of glass fragments.”
Forensic science international , 160(1): 53–58.
Rosenthal, H. P. (1970). “On the subspaces of Lp (p > 2) spanned by sequences of
independent random variables.” Israel J. Math., 8: 273–303.
Scheffé, H. (1943). “On solutions of the Behrens-Fisher problem, based on the tdistribution.” Ann. Math. Statistics, 14: 35–44.
von Bahr, B. and Esseen, C.-G. (1965). “Inequalities for the rth absolute moment of a
sum of random variables, 1 ≤ r ≤ 2.” Ann. Math. Statist, 36: 299–303.
Williams, V. S., Jones, L. V., and Tukey, J. W. (1999). “Controlling error in multiple
comparisons, with examples from state-to-state differences in educational achievement.”
Journal of educational and behavioral statistics, 24(1): 42–69.
Yekutieli, D. (2008). “Hierarchical false discovery rate-controlling methodology.” J. Amer.
Statist. Assoc., 103(481): 309–316.
Zaı̆tsev, A. Y. (1987). “On the Gaussian approximation of convolutions under multidimensional analogues of S. N. Bernstein’s inequality conditions.” Probab. Theory Related
Fields, 74(4): 535–566.
Institute of Natural Sciences, Shanghai Jiao Tong University
E-mail address: [email protected]
Department of Data Sciences and Operations, University of Southern California
E-mail address: [email protected]
Department of Statistics, Chinese University of Hong Kong
E-mail address: [email protected]
| 10math.ST
|
Image Crowd Counting Using CNN and MRF
Paper:
Image Crowd Counting Using Convolutional Neural Network
and Markov Random Field
Kang Han, Wanggen Wan, Haiyan Yao, and Li Hou
arXiv:1706.03686v3 [cs.CV] 17 Oct 2017
School of Communication and Information Engineering, Shanghai University, Shanghai, China
Institute of Smart City, Shanghai University, Shanghai, China
E-mail: [email protected]
Abstract.
In this paper, we propose a method
called Convolutional Neural Network-Markov Random Field (CNN-MRF) to estimate the crowd count
in a still image. We first divide the dense crowd visible
image into overlapping patches and then use a deep
convolutional neural network to extract features from
each patch image, followed by a fully connected neural
network to regress the local patch crowd count. Since
the local patches have overlapping portions, the crowd
count of the adjacent patches has a high correlation.
We use this correlation and the Markov random field
to smooth the counting results of the local patches. Experiments show that our approach significantly outperforms the state-of-the-art methods on UCF and
Shanghaitech crowd counting datasets. Code available on GitHub https://github.com/hankong/crowdcounting.
Keywords: crowd counting, convolutional neural network, Markov random field
1. Introduction
In modern society, more and more people gather to live
in the city. This lifestyle provides convenience for people’s lives and improves the utilization rate of urban public resources. Meanwhile, a large number of people living in cities also led to urban congestion problems. This
crowding phenomenon can be observed at traffic junctions, airports, stations, stadiums and other public places.
Many accidents caused by overcrowding have led to many
deaths, such as massive stampede happened in Shanghai
Bund in 2015, where 36 persons died and 49 persons were
injured. Therefore, automatic crowd density estimation
and early warning for overcrowding are important to prevent such tragedies from happening again.
Existing methods for people density estimation and
counting of the crowd can be divided into two main categories: direct and indirect approaches [1]. The main idea
of the direct approach (also called object detection based)
is to detect and segment each individual in crowd scenes
to get the total count, while the indirect approach (also
called feature based) takes an image as a whole and extracts some features, and then get the final count through
the regression analysis of the extracted features. The advantage of the direct method is that people or head detection have been widely studied and applied, these methods can be easily adapted to the crowd of few tens [2–4].
However, a crowd of more than hundreds does not have
a well-defined shape as a single object does, so the direct
detection method is not applicable. In this scene, the indirect methods are generally more reliable and efficient,
since the overall features of a crowd image are easier to
obtain and have a stronger correlation with the number of
people. Some surveys of these methods can be seen in [5].
In this paper, we focus on the indirect method to estimate the crowd count of more than hundreds. Like other
computer vision tasks, features extraction is the first step
and the most important step in the crowd counting problem. Many hand-crafted computer vision features have
been used to represent the density of the crowd, such as
Local Binary Patterns (LBP) [6], Scale Invariant Feature
Transform (SIFT) [7] and Histogram of Oriented Gradient
(HOG) [2]. However, due to the variation of viewpoint,
scene and crowd count, these hand-crafted features can
not represent the crowd density discriminatively. In recent years, with the advancement of deep learning in computer vision, some researchers try to apply deep learning
to crowd density estimation and achieved state-of-the-art
results [8]. Different with the hand-crafted features extraction which follows the certain steps, deep learning can
automatically learn the features from the data. Following
the certain steps means that the features will not be better with the increase of the data, while deep learning can
learn more discriminative features from abundant data.
Some auxiliary methods are also used to improve the
accuracy of the crowd density estimation. For example,
in the crowd of more than hundreds, the density of crowd
is continuously gradient, so MRF can be used to smooth
the counting between adjacent patches [9].
Inspired by the superior feature representation of deep
learning and the smoothness of MRF, we propose a CNN
and MRF based framework for the problem of people
counting in still images. Firstly, we divide the image into
patches with overlaps and use a pre-trained CNN model
to extract deep features from each overlapping patch, followed by a fully connected deep neural network to regress
the patch people count. Finally, we use MRF to smooth
the counts of adjacent patches in order to make the counts
closer to the true value. The reason is that the overlap be-
Journal of Advanced Computational Intelligence
and Intelligent Informatics
1
Han Kang. et al.
tween the adjacent patches leads to the people count of
adjacent patches keep a certain consistency.
The rest of the paper is organized as follows. In Section
2, we briefly review the related work of crowd density estimation and counting. And then the system architecture
and the implement detail of the proposed approach will
be illustrated in Section 3, followed by the experiment in
Section 4. Finally, Section 5 concludes our paper.
2. Related work
Direct method. Li et al [10] proposed a people count
estimation method combining the foreground segmentation and the head-shoulder detection. A Mosaic Image
Difference based foreground segmentation algorithm is
performed first to detect active areas, and then a headshoulder detection algorithm is utilized to detect heads
and count the number from the detected foreground areas.
Cheriyadat et al [11] presented an object detection system
based on coherent motion region detection for counting
and locating objects in the presence of high object density
and inter-object occlusions. They used the locations of
tracked low-level feature points to construct all possible
coherent-motion-regions and chose a good disjoint set of
coherent motion regions representing individual objects
using a greedy algorithm. Brostow et al [12] described an
unsupervised data-driven Bayesian clustering algorithm
to detect individual entities. The method tracked simple
image features and probabilistically group them into clusters representing independently moving entities.
Indirect method. Lempitsky et al [13] used dense
SIFT features and Maximum Excess over SubArrays distance as a loss function to train a regression model on randomly selected patches. In order to adapt to the change
of the crowd density and perspective, Zhang et al [8] proposed a Multi-column CNN architecture to map the image
to its crowd density map. The network structure included
three parallel CNN with different sizes to extract features
from different scales. The features learned by each column CNN were adaptive to variations in people/head size
due to perspective effect or image resolution. Shang et
al [14] proposed an end-to-end CNN architecture that directly maps the whole image to the counting result. A pretrained GoogLeNet model was used to extract high-level
deep features and the long-short time memory (LSTM)
decoders for the local count and fully connected layers
for the final count. A cross-scene crowd counting architecture was proposed by Zhang et al [15]. Firstly, they
trained a deep CNN with two related learning objectives,
crowd density, and crowd count. And then a data-driven
method was introduced to fine-tune the learned CNN to an
unseen target scene, where the training samples similar to
the target scene were retrieved from the training scenes
for fine-tuning.
Some researchers tried to combine two methods to get
a more accurate estimate count. Idrees et al [9] proposed
a multi-source multi-scale counting method to compute
an estimation of the number of individuals presented in
2
an extremely dense crowd visible in a single image. This
method combined the detection approach (low confidence
head detection) and the features regression approach (repetition of texture element and frequency-domain analysis)
to solve the perspective, occlusion, clutter and few pixels
per person problems in a crowd of more than hundreds.
Then the spatial consistency constraints are employed by
MRF to smooth the counting between adjacent patches.
Rodriguez et al. [16] addressed the problem of person
detection and tracking in crowded video scenes. They
explored constraints imposed by the crowd density and
formulate person detection as the optimization of a joint
energy function combining crowd density estimation and
the localization of individual people.
(a)
(b)
Fig. 1. Some example images from UCF crowd dataset.
The crowd density become sparse with the view from far to
near in image (a). The crowd density of the buildings and
mountains is 0 in the image (b).
3. Proposed Method
For dense crowd images, the distribution of crowd density is rarely uniform due to changes in perspective and
scene. Some example images can be seen in Fig. 1.
Therefore, it is unreasonable to count the crowd by taking
the image as an entirety. So our framework adapted the
divide-count-sum strategy. The images are firstly divided
into patches, then a regression model is learned to map the
image patch to the local count. Finally, the global image
count is computed as the total sum over these patches.
The image segmentation has two advantages: Firstly, in
the small image patches, the crowd density is approximately uniform distribution. Secondly, the segmentation
of image increases the number of training data for the regression model. The above advantages allow us to train a
more robust regression model.
Even though the distribution of crowd density is not
uniform, the overall crowd density distribution is continuous. This means that the density of adjacent image
patches should be similar. Furthermore, we divide the
image with overlaps, which enhances the association between image patches. The Markov random field is used to
smooth the estimation count between overlapping image
patches to compensate for the possible estimation errors
of the image patches and to bring the overall result closer
to the true density distribution. The overview of the proposed method can be seen as Fig. 2.
Journal of Advanced Computational Intelligence
and Intelligent Informatics
Image Crowd Counting Using CNN and MRF
ResNet
100 100 50 50
Image
ResNet feature
1
Fully connected regress network
MRF
Global count
Fig. 2. The overview of our proposed CNN-MRF crowd counting method. The proposed method contains three parts, a pre-trained
deep residual network to extract features, a fully connected neural network for count regress and a MRF to smooth the counting
results of the local patches. The final count is obtained by the sum of the patches count.
3.1. Patches counting
3.2. Images counting
We use a pre-trained deep residual network to extract
features from image patches and a fully connected neural network to learn a map from the above features to
the local count. The features learned from deep convolutional networks have been used for many computer vision
tasks such as image recognition, object detection, and image segmentation [17]. This indicates that the features
learned from the deep convolutional network are universal to many computer vision tasks. With the increase in
the number of network layers, the representation ability
of the learned features becomes stronger. However, a
deeper model means that more data is needed for training.
The existing crowd counting datasets are not large enough
to train a very deep convolutional neural network from
scratch. So we use a pre-trained deep residual network
to extract features from image patches. The deep residual
network was proposed by He et al. [17]. Their method
addressed the degradation problem by reformulating the
layers as learning residual functions with reference to the
layer inputs, instead of learning unreferenced functions.
We employ the residual network, which is trained on ImageNet dataset for image classification task, to extract the
deep features to represent the density of the crowd. This
pre-trained CNN network created a residual item for every
three convolution layer to bring the layer of the network
to 152. We resize the image patches to the size of 224 ×
224 as the input of the model and extract the output of the
fc1000 layer to get the 1000 dimensional features.
The features are then used to train 5 layers fully connected neural network. The network’s input is 1000dimensional, and the number of neurons in the network
is given by 100-100-50-50-1. The network’s output is the
local crowd count. The learning task of the fully connected neural network is to minimize the mean squared
error of the training patches, which can be written as:
s
1 M
Loss =
∑ (cg − cr )2 . . . . . . . . (1)
M i=1
Due to the overlapping of the adjacent image patches,
there is a high correlation between adjacent local people
count. This correlation can be used by the Markov random field to smooth the estimation count between adjacent image patches. As previously analyzed, the people
count of adjacent images patches is generally similar and
may change dramatically at some places due to buildings
or other objects in the scene. This characteristic can be
well modeled by the Markov random field. Formally, the
Markov random field framework for the crowd counting
can be defined as follows (we follow the notation in [18]).
Let P be the set of patches in an image and C be a possible set of counts. A counting c assigns a count c p ∈ C to
each patch p ∈ P. The quality of a counting is given by an
energy function:
where M is the number of the training image patches and
cg and cr are the ground truth count and the regression
count of the image patches, respectively.
E(c) =
∑ D p (c p ) + ∑
p∈P
V (c p − cq )
. . . (2)
(p,q)∈N
where N are the (undirected) edges in the four-connected
image patch graph. D p (c p ) is the cost of assigning count
c p to patch p, and is referred to as the data cost. V (c p −cq )
measures the cost of assigning count c p and cq to two
neighboring patch, and is normally referred to as the discontinuity cost.
For the problem of smoothing the adjacent patches
count, D p (c p ) and V (c p − cq ) can take the form of the
following functions:
D p (c p ) = λ min((I(p) − c p )2 , DATA K) . . . (3)
V (c p − cq ) = min((c p − cq )2 , DISC K)
. . . (4)
where λ is a weight of the energy items, I(p) is the ground
truth count of the patch p, DATA K and DISC K are the
truncating item of D p (c p ) and V (c p − cq ), respectively.
The truncating item makes the cost function stop growing
after the difference becomes large, which allows for large
discontinuities. The above energy function minimization
problem can be efficiently solved by belief propagation
algorithm [18]. Fig. 3 shows the smoothing effect of the
MRF on the local counting results. We can see that the
density map is closer to the ground truth after smoothed
by the MRF.
Journal of Advanced Computational Intelligence
and Intelligent Informatics
3
Han Kang. et al.
Images
Ground Truth
Before MRF
After MRF
Fig. 3. Smoothing of adjacent local counts by MRF. The first column are crowd images, the second, third, and final columns are
the ground truth density map, the estimated density map, and the density map smoothed by the MRF, respectively.
We evaluate our method on UCF and Shanghaitech
crowd counting datasets. Each image is divided with
patch size 100 x 100 pixels and stride size 50 pixels. The
final count of the whole image is obtained by calculating
the sum of the count of all non-overlapping patches. If the
image patch is on the edge and its previous image patch
has been summed, then only half count of this image patch
will be summed. The proposed method is implemented in
Matlab, and we utilize MatConvNet [19], a Matlab toolbox implementing CNN for computer vision applications,
which provides many pre-trained CNN model for image
classification, segmentation, face recognition and text detection.
We utilize two evaluation criteria: the mean absolute
error (MAE) and the mean squared error (MSE), which
are defined as follows:
1 N
MAE = ∑ |gi − ei | . . . . . . . . . . (5)
N i=1
s
1 N
MSE =
∑ (gi − ei )2 . . . . . . . . (6)
N i=1
where N is the number of test images and gi and ei are
the ground truth and the estimate count of the i-th image,
respectively.
4.1. UCF dataset
The UCF dataset [9] is a very challenging dataset because the scene of the each image are different and the
crowd count of the image changes dramatically. More
specifically, there are 50 images with counts ranging between 94 and 4543 with an average of 1280 individuals
per image. The ground truth positions of individuals are
marked by the authors and there is a total of 63705 annotations in the 50 images.
4
Table 1. Comparing results of different methods on the
UCF dataset.
Method
Rodriguez et al. [16]
Lempitsky et al. [13]
Idrees et al. [9]
Zhang et al. [15]
MCNN [8]
Shang et al. [14]
Proposed
MAE
655.7
493.4
419.5
467.0
295.1
270.3
254.1
MSE
697.8
487.1
541.6
498.5
490.2
352.5
3500
Ground truth
Estimated count
3000
2500
Absolute Counts
4. Experiments
2000
1500
1000
500
0
0
1
2
3
4
5
6
7
8
9
10
Group ID
Fig. 4. The comparison of the ground truth and the estimated count on UCF dataset. Absolute counts in the vertical
axis is the average crowd number of images in each group.
Followed by Idrees et al. [6], we use 5-fold crossvalidation to test the performance of our algorithm. Since
the image of the UCF dataset is gray, we extend the image
to three channels by copying the data. The counting result
of our method and the comparison with other methods can
be viewed in Table 1. The experimental results of other
methods come from their papers and the same for Shanghaitech[8] dataset. We can see that our proposed CNNMRF outperforms the other methods, including the state
of the art methods MCNN [8] and Shang et al. [14]. In Fig
Journal of Advanced Computational Intelligence
and Intelligent Informatics
Image Crowd Counting Using CNN and MRF
4, we compare the estimated count with the ground truth
in more details. The images are divided into 10 groups
according to crowd counts in an increasing order. The
divide-count-sum strategy narrows the range of the crowd
count so that the proposed method can give an accurate
estimate of the total counts of the images at all ranges.
4.2. Shanghaitech dataset
Shanghaitech dataset [8] is a large-scale crowd counting dataset which contains 330,165 people heads annotation in 1,198 images. The dataset consists of two parts,
Part A is collected from the Internet and Part B is taken
from the busy streets of metropolitan areas in Shanghai.
The average crowd count of the Part A is 501.4, and the
average number of the Part B dataset is 123.6. The crowd
density of Part A is significantly larger than that in Part
B. This dataset has been randomly divided into training
and testing: 300 images of Part A are used for training
and the remaining 182 images for testing, and 400 images
of Part B are for training and 316 images for testing. Table 2 reports the results of different methods in the two
parts. LBP+RR [8] is a regression based method which
uses Local Binary Pattern (LBP) features extracted from
the original image as input and uses ridge regression (RR)
to predict the crowd number for each image. Our method
significantly outperforms state-of-the-art methods.
Table 2. Comparing results of different methods on the
Shanghaitech dataset.
Part A
MAE MSE
303.2 371.0
181.8 277.7
110.2 173.2
79.1 130.1
Method
LBP+RR [8]
Zhang et al. [15]
MCNN [8]
Proposed
Part B
MAE MSE
59.1
81.7
32.0
49.8
26.4
41.3
17.8
26.0
(a) GT: 2391 C: 2362
(b) GT: 1601 C: 1632
(c) GT: 416 C: 414
(d) GT: 91 C: 91
Fig. 6. Sample images with their respective ground truth and
estimated count. GT is ground truth of the image, and C is
the predicted count by the proposed method. Image (a) and
(b) from the UCF dataset, (c) and (d) from the Shanghaitech
dataset.
5. Conclusion
We present a CNN-MRF based approach to counting
the crowd in a still image from different scenes. The
features extracted from the CNN model trained for other
computer vision tasks show a strong ability to represent
crowd density. With the overlapping patches divided
strategies, the adjacent local counts are highly correlated.
This correlation can be used by the MRF to smooth the
adjacent local counts to obtain a more accurate overall
count. Experimental results demonstrate that the proposed approach achieve superior performance compared
with several recent related methods.
Acknowledgements
This research was partially supported by the National Nature Science Foundation of China (No.61373084).
300
1200
Ground truth
Estimated count
Ground truth
Estimated count
250
Absolute Counts
Absolute Counts
1000
800
600
400
150
100
50
200
0
References:
200
0
1
2
3
4
5
6
Group ID
(a) Part A
7
8
9
10
0
0
1
2
3
4
5
6
7
8
9
10
Group ID
(b) Part B
Fig. 5. The comparison of the ground truth and the estimated count on Shanghaitech dataset.
Similar to Fig 4, the comparison of the ground truth
and the estimated count on Shanghaitech dataset can be
seen in Fig 5. We can see that the proposed method can
estimate the crowd count accurately and is robust to the
large variation in crowd density. Some counting examples
of the images with the associated ground truth counts can
be seen in Fig 6.
[1] S. A. M. Saleh, S. A. Suandi, and H. Ibrahim, “Recent survey on
crowd density estimation and counting for visual surveillance”, Engineering Applications of Artificial Intelligence, 41:103–114, 2015.
[2] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection”, In 2005 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR’05), volume 1,
pp. 886–893. IEEE, 2005.
[3] R. Stewart, M. Andriluka, and A. Y. Ng, “End-to-end people detection in crowded scenes”, In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 2325–2333, 2016.
[4] Y. Yasuoka, Y. Shinomiya, and Y. Hoshino, “Simulation of Human
Detection System Using BRIEF and Neural Network”, Journal of
Advanced Computational Intelligence and Intelligent Informatics,
20(7):1159–1164, 2016.
[5] M. Junjie, D. Yaping, and H. Kaoru, “A Survey of Video-Based
Crowd Anomaly Detection in Dense Scenes”, Journal of Advanced
Computational Intelligence and Intelligent Informatics, 21(2):235–
246, 2017.
[6] T. Ojala, M. Pietikainen, and T. Maenpaa, “Multiresolution grayscale and rotation invariant texture classification with local binary
patterns”, IEEE Transactions on pattern analysis and machine intelligence, 24(7):971–987, 2002.
[7] D. G. Lowe, “Distinctive image features from scale-invariant key-
Journal of Advanced Computational Intelligence
and Intelligent Informatics
5
Han Kang. et al.
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
6
points”, International journal of computer vision, 60(2):91–110,
2004.
Y. Zhang, D. Zhou, S. Chen, S. Gao, and Y. Ma, “Single-image
crowd counting via multi-column convolutional neural network”,
In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 589–597, 2016.
H. Idrees, I. Saleemi, C. Seibert, and M. Shah, “Multi-source multiscale counting in extremely dense crowd images”, In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 2547–2554, 2013.
M. Li, Z. Zhang, K. Huang, and T. Tan, “Estimating the number of
people in crowded scenes by mid based foreground segmentation
and head-shoulder detection”, In Pattern Recognition, 2008. ICPR
2008. 19th International Conference on, pp. 1–4. IEEE, 2008.
A. M. Cheriyadat, B. L. Bhaduri, and R. J. Radke, “Detecting multiple moving objects in crowded environments with coherent motion
regions”, In Computer Vision and Pattern Recognition Workshops,
2008. CVPRW’08. IEEE Computer Society Conference on, pp. 1–
8. IEEE, 2008.
G. J. Brostow and R. Cipolla, “Unsupervised bayesian detection
of independent motion in crowds”, In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition
(CVPR’06), volume 1, pp. 594–601. IEEE, 2006.
V. Lempitsky and A. Zisserman, “Learning to count objects in
images”, In Advances in Neural Information Processing Systems,
pp. 1324–1332, 2010.
C. Shang, H. Ai, and B. Bai, “End-to-end crowd counting via joint
learning local and global count”, In Image Processing (ICIP), 2016
IEEE International Conference on, pp. 1215–1219. IEEE, 2016.
C. Zhang, H. Li, X. Wang, and X. Yang, “Cross-scene crowd counting via deep convolutional neural networks”, In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition,
pp. 833–841, 2015.
M. Rodriguez, I. Laptev, J. Sivic, and J.-Y. Audibert, “Densityaware person detection and tracking in crowds”, In 2011 International Conference on Computer Vision, pp. 2423–2430. IEEE,
2011.
K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
image recognition”, arXiv preprint arXiv:1512.03385, 2015.
P. F. Felzenszwalb and D. P. Huttenlocher, “Efficient belief propagation for early vision”, International journal of computer vision,
70(1):41–54, 2006.
A. Vedaldi and K. Lenc, “Matconvnet: Convolutional neural networks for matlab”, In Proceedings of the 23rd ACM international
conference on Multimedia, pp. 689–692. ACM, 2015.
Journal of Advanced Computational Intelligence
and Intelligent Informatics
| 1cs.CV
|
1
A Novel Genetic Algorithm using Helper
Objectives for the 0-1 Knapsack Problem
Jun He, Feidun He and Hongbin Dong
arXiv:1404.0868v1 [cs.NE] 3 Apr 2014
Abstract
The 0-1 knapsack problem is a well-known combinatorial optimisation problem. Approximation algorithms have been designed
for solving it and they return provably good solutions within polynomial time. On the other hand, genetic algorithms are well
suited for solving the knapsack problem and they find reasonably good solutions quickly. A naturally arising question is whether
genetic algorithms are able to find solutions as good as approximation algorithms do. This paper presents a novel multi-objective
optimisation genetic algorithm for solving the 0-1 knapsack problem. Experiment results show that the new algorithm outperforms
its rivals, the greedy algorithm, mixed strategy genetic algorithm, and greedy algorithm + mixed strategy genetic algorithm.
Index Terms
genetic algorithm, knapsack problem, multi-objective optimisation, solution quality
I. I NTRODUCTION
The 0-1 knapsack problem is one of the most important and also most intensively studied combinatorial optimisation problems
[1]. Several approximation algorithms have proposed for solving the 0-1 knapsack problem [1]. These algorithms always can
return provably good solutions, whose values are within a factor of the value of the optimal solution.
In last two decades, evolutionary algorithm, especially genetic algorithms (GAs), have been well adopted for tackling the
knapsack problem [2], [3]. The problem has received a particular interest from the evolutionary computation community for
the following reason. The binary vector representation is a natural encoding of of the candidate solutions to the 0-1 knapsack
problem. Thereby, it provides an ideal setting for the applications of GAs [4, Chapter 4].
Empirical results often assert that GAs produce reasonably good solutions to the knapsack problems [5], [6], [7]. A naturally
arising question is to compare the solution quality (reasonably good versus provably good) between GAs and approximation
algorithms. There are two approaches to answer the question. One approach is to make a theoretical analysis. A GA is proven
that it can produce a solution within a polynomial runtime, the value of which is within a factor of the value of an optimal
solution. This is a standard approach used in the study of approximation algorithms. Another approach is to conduct an empirical
study. A GA is compared with an approximation algorithm via computer experiments. If the GA can produce solutions better
or not worse than an approximative algorithm does in all instances within polynomial time, the GA is able to reach the same
solution quality as the approximation algorithm does.
The current paper is an empirical study of an GA which uses the multi-objectivization technique [8]. In multi-objectivization,
single-objective optimisation problems are transferred into multi-objective optimisation problems by decomposing the original
objective into several components [8] or by adding helper objectives [9]. Multi-objectivization may bring both positive and
negative effects [10], [11], [12]. This approach has been used for solving several combinatorial optimisation problems,
for example, the knapsack problem [13], vertex cover problem [14] and minimum label spanning tree problem [15]. This
paper focusses on the 0-1 knapsack problem. A novel GA using three helper objectives is designed for solving the 0-1
knapsack problem. Then the solution quality of the GA is compared with a well-known approximation algorithm via computer
experiments.
The remainder of the paper is organized as follows. The 0-1 knapsack problem, a greedy algorithm and a GA for it are
introduced in Section II. In Section III we present a novel GA using helper objectives. Section IV is devoted to an empirical
comparison among several algorithms. Section V concludes the article.
II. K NAPSACK P ROBLEM , G REEDY A LGORITHM
AND
G ENETIC A LGORITHM
In an instance of the 0-1 knapsack problem, given a set of n items with weights wi and profits pi , and a knapsack with
capacity C, the task is to maximise the sum of profits of items packed in the knapsack without exceeding the capacity. More
formally the target is to find a binary vector ~x = (x1 · · · xn ) so as to
max f (~x) =
~
x
n
X
i=1
pi xi ,
subject to
n
X
wi xi ≤ C,
i=1
This work was supported by the EPSRC under Grant No. EP/I009809/1.
Jun He is with Department of Computer Science, Aberystwyth University, Aberystwyth, SY23 3DB, UK (email: [email protected]).
Feidun He is with School of Information Science and Technology, Southwest Jiaotong University, Chengdu, China
Hongbin Dong is with College of Computer Science and Technology, Harbin Engineering University, Harbin, China
(1)
2
where xi = 1 if the item i is selected in the knapsack and xi = 0 if the item i is not selected in the knapsack. A feasible
solution is an ~x which satisfies the constraint. An infeasible one is an ~x that violates the constraint.
Several approximation algorithms have been proposed for solving the 0-1 knapsack problem (see [1, Chapter 2] for more
details). Among these, the simplest one is the greedy algorithm described below. The algorithm aims at putting the most
profitable items as many as possible into the knapsack or the items with the highest profit-to-weight ratio as many as possible,
within the knapsack capacity.
1: input an instance of the 0-1 knapsack problem;
p
p
2: resort all the items via the ratio of their profits to their corresponding weights so that w1 ≥ · · · ≥ wn ;
1
n
3: greedily add the items in the above order to the knapsack as long as adding an item to the knapsack does not exceeding
the capacity of the knapsack. Denote the solution by ~y ;
4: resort all the items according to their profits so that p1 ≥ p2 ≥ · · · ≥ pn ;
5: greedily add the items in the above order as long as adding an item to the knapsack does not exceeding the capacity of
the knapsack. Denote the solution by ~z;
6: output the best of ~
y and ~z.
This algorithm is a 1/2-approximation algorithm for the 0-1 knapsack problem [1, Section 2.4], which means it always can
return a solution no worse than 1/2 of the value of the optimal solution.
The greedy algorithm stops after finding an approximation solution, and it has no ability to seek the global optimal solution.
Therefore GAs are often applied for solving the 0-1 knapsack problem.
In order to handle the constraint in the knapsack problem, we use repair methods since they are claimed to be the most
efficient for the knapsack problem [4], [16]. A repair method is explained as follows.
1: input an infeasible solution ~
x;
2: while ~
x is infeasible do
3:
i =: choose an item from the knapsack;
4:
setP
xi = 0;
n
5:
if i=1 xi wi ≤ C then
6:
~x is feasible;
7:
end if
8: end while
9: output a feasible solution ~
x.
There are different methods available for choosing an item in the repair procedure, described as follows.
1) Profit-greedy repair: sort all items according to the decreasing order of their corresponding profits. Then choose the item
with the smallest profit and remove it from the knapsack.
2) Ratio-greedy repair: sort all items according to the decreasing order of the corresponding ratios. Then choose the item
with the smallest ratio and remove it from the knapsack.
3) Random repair: choose an item from the knapsack at random and remove it from the knapsack.
Thanks to the repair method, all of the infeasible solutions are repaired into the feasible ones. The following pseudo-code
is a mixed strategy GA (MSGA) which chooses one of three repair methods in a probabilistic way and then applies the repair
method to generate a feasible solution.
1: input an instance of the 0-1 knapsack problem;
2: initialize population Φ0 consisting of N feasible solutions;
3: for t = 0, 1, · · · , tmax do
4:
generate a random number r in [0, 1];
5:
if r < 0.9 then
6:
children population Φt.c ← bitwise-mutate Φt ;
7:
else
8:
children population Φt.c ← one-point crossover Φt ;
9:
end if
10:
if a child is an infeasible solution then
11:
choose one method from the ratio-greedy repair, random repair and value-greedy repair with probability 1/3 and repair
the child into feasible;
12:
end if
13:
the best individual in the parent and children populations is selected into population Φt+1 ;
14:
N − 1 individuals from the parent and children populations into population Φt+1 by roulette wheel selection;
15: end for
16: output the maximum of the fitness function.
The genetic operators used in the above GA are explained below.
• Bitwise Mutation: Given a binary vector (x1 · · · xn ), flip each bit xi with probability 1/n.
3
•
One-Point Crossover: Given two binary vectors (x1 · · · xn ) and (y1 · · · yn ), randomly choose a crossover point k ∈
{1, · · · , n}, swap their bits at point k. Then generate two new binary vectors as follows,
(x1 · · · xk yk+1 · · · yn ),
(y1 · · · yk xk+1 · · · xn ).
Like most of GAs, the MSGA may find reasonably good solutions but has no guarantee about the solution quality. Thus it
is necessary to design evolutionary approximation algorithms with provably good solution quality. The most straightforward
approach is that we first apply the greedy algorithm for generating approximation solutions and then take these solutions as
the starting point of the MSGA. We call this approach greedy algorithm + MSGA. Since the MSGA starts at local optima, it
becomes hard for the MSGA to leave the absorbing basin of this local optimum for seeking the global optimum. This is the
main drawback of the approach.
III. G ENETIC A LGORITHM USING H ELPER O BJECTIVES FOR THE 0-1 K NAPSACK P ROBLEMS
In this section, we propose a novel multi-objective optimisation GA (MOGA) which can beat the combination of greedy
algorithm + MOGA mentioned in the previous section. The algorithm is based on the multi-objectivization technique. The
original single objective optimization problem (1) is recast into a multi-objective optimization problem using helper objectives.
The design of helper objectives depends on problem-specific knowledge. The first helper objective comes from an observation
on the following instance.
Item
1
2
3
4
5
Profit
Weight
10
10
10
10
10
10
12
10
12
10
Capacity
20
The global optimum in this instance is 00011. In the optimal solution, the average profit of packed items is the largest. Thus
the first helper objective is to maximize the average profit of items in a knapsack. The objective function is
n
1 X
h1 (~x) =
xi pi .
(2)
k ~x k1 i=1
Pn
where k ~x k1 = i=1 xi .
The second objective is inspired from an observation on another instance.
Item
1
2
3
4
5
Profit
Weight
15
10
15
10
20
20
20
20
20
20
Capacity
20
The global optimum in this instance is 11000. In the optimal solution, the average profit-to-weight ratio of packed items is
the largest. However, the average profit of these items is not the largest. Then the second helper objective is to maximize the
average profit-to-weight ratio of items in a knapsack. The objective function is
n
1 X pi
xi .
h2 (~x) =
(3)
k ~x k1 i=1 wi
Finally let’s look at the following instance.
Item
1
2
3
4
5
Profit
Weight
40
30
40
30
40
30
40
30
150
100
Capacity
120
It is not difficult to verify that the global optimum in this instance is 11110. In the optimal solution, neither the average
profit of packed items nor average profit-to-weight ratio is the largest, but the number of packed items is the largest. Thus the
third helper objective is to maximize the number of items in a knapsack. The objective function is
h3 (~x) =k ~x k1 .
(4)
We then come to the following multi-objective optimization problem:
max{f (~x), h1 (~x), h2 (~x), h3 (~x)},
~
x
subject to
n
X
i=1
wi xi ≤ C.
(5)
4
Besides the above three helper objectives, it is possible to add more helper objectives, for example, to minimise the average
weight of packed items.
The multi-objective optimisation problem (5) is solved by a MOGA using bitwise mutation, one-point crossover and multicriteria selection, plus a mixed strategy of three repair methods.
1: input an instance of the 0-1 knapsack problem;
2: initialize Φ0 consisting of N feasible solutions;
3: for t = 0, 1, · · · , tmax do
4:
generate a random number r in [0, 1];
5:
if r < 0.9 then
6:
children population Φt.c ← bitwise mutate Φt ;
7:
else
8:
children population Φt.c ← one-point crossover Φt ;
9:
end if
10:
if any child is an infeasible solution then
11:
choose one repair method from the ratio-greedy repair, random repair and value-greedy repair with probability 1/3;
12:
repair the child into a feasible solution;
13:
end if
14:
population Φt+1 ← multi-criterion select N individuals from Φt and Φt.c ;
15: end for
16: output the maximum of f (~
x) in the final population.
The multi-criteria selection operator, adopted in the above MOGA, is novel and inspired from multi-objective optimisation.
Since the target is to maximise several objectives simultaneously, we select individuals which have higher function values with
respect to each objective function. The pseudo-code of multi-criteria selection is described as follows.
1: input the parent population Φt and child population Φt.c ;
2: merge the parent and children populations into a temporary population which consists of 2N individuals;
(1)
(1)
3: sort all individuals in the temporary populations in the descending order of f (~
x), denote them by ~x1 , · · · , ~x2N ;
(1)
(1)
(1)
(1)
(1)
4: select all individuals from left to right (denote them by ~
xk1 , · · · , ~xkm ) which satisfy h1 (~xki ) < h1 (~xki+1 ) or h2 (~xki ) <
(1)
h2 (~xki+1 ) for any ki .
5: if the number of selected individuals is greater than m N
3 then
6:
truncate them to N3 individuals;
7: end if
8: add these selected individuals into the next population Φt+1 ;
9: resort all individuals in the temporary population in the descending order of h1 (~
x), still denote them by ~x1 , · · · , ~x2N ;
10: select all individuals from left to right (still denote them by ~
xk1 , · · · , ~xkm ) which satisfy h3 (~xki ) < h3 (~xki+1 ) for any ki .
11: if the number of selected individuals is greater than N
3 then
12:
truncate them to N3 individuals;
13: end if
14: add these selected individuals into the next population Φt+1 ;
15: resort all individuals in the temporary populations in the descending order of h2 (~
x), still denote them by ~x1 , · · · , ~x2N ;
16: select all individuals from left to right (still denote them by ~
xk1 , · · · , ~xkm ) which satisfy h3 (~xki ) < h3 (~xki+1 ) for any ki .
17: if the number of selected individuals is greater than N
3 then
18:
truncate them to N3 individuals;
19: end if
20: add these selected individuals into the next population Φt+1 ;
21: while the population size of Φt+1 is less than N do
22:
randomly choose an individual from the parent population and add it into Φt+1 ;
23: end while
24: output a new population Φt+1 .
In the above algorithm, Steps 3-4 are for selecting the individuals with higher values of f (~x). In order to preserve diversity,
we choose these individuals which have different values of h1 (~x) or h2 (~x). Similarly Steps 9-10 are for selecting the individuals
with a higher value of h1 (~x). We choose the individuals which have different values of h3 (~x) for maintaining diversity. Steps
15-16 are for selecting individuals with a higher value of h2 (~x). Again we choose these individuals which have different values
of h3 (~x) for preserving diversity. We don’t explicitly select individuals based on h3 (~x). Instead we implicitly do it during
Steps 9-10, and Steps 15-16.
Steps 5-7, Steps 11-13, Steps 17-19, plus Steps 21-23 are used to maintain an invariant population size N .
5
The benefit of using multi-criterion selection is its ability of making search along different directions f (~x), h1 (~x), h2 (~x)
and implicitly h3 (~x). Hence the MOGA may not get trapped into the absorbing area of a local optimum.
IV. E XPERIMENTS
In this section, we implement computer experiments. According to [1], [4], the instances of the 0-1 knapsack problem are
often classified into two categorises.
1) Restrictive capacity knapsack: the knapsack capacity is so small that only a few items can be packed in the knapsack.
An instance with restrictive capacity knapsack is generated in the following way. Choose a parameter B which is an
upper bound on the weight of each item. In the experiments, set B = n. For item i, its profit pi and weight wi are
generated at uniformly random in [1, B]. Set the capacity of the knapsack C = B.
2) Average capacity knapsack: the knapsack capacity is so large that it is possible to pack half of items into the knapsack.
An instance with average capacity knapsack is generated as follows. Choose a parameter B which is the upper bound
on the weight of each item. In the experiments, set B = n. For item i, its profit pi and weight wi are generated at
uniformly random in [1, B]. Since the average weight of each item is 0.5B, thus the average of the total weight of items
is 0.5nB. So we set the capacity to be the half of the total weight, that is C = 0.25nB.
For each type of the 0-1 knapsack problem, 10 instances are generated at random. For each instance, the number of items
n is 100. The population size is 3n. The number of maximum generations is 30n for the MSGA and 10n for the MOGA. All
individuals in the initial population are generated at random. If an individual is an infeasible solution, it is repaired to feasible
using random repair.
Besides the above randomly generated instances, we also consider two special instances. Special instance I is given Table I.
TABLE I
S PECIAL I NSTANCE I: n = 500 AND α = 0.2
Item i
n
1, · · · , ⌈ 1+α
⌉
Profit pi
Weight wi
1
1
n
⌈ 1+α
⌉+1
n
⌈ 1+α
⌉ + 2, · · · , n
αn
1+α
1
n
1
2n
n
1+α
n
1+α
Capacity
Initialisation
−
α
4+4α
0
1
half bits are 1
Special instance II is given in Table II.
TABLE II
S PECIAL I NSTANCE II: n = 200
Item i
Profit pi
Weight wi
1, · · · , n
4
√
0.25n√n + 2
0.25n n + 1
Capacity
Initialisation
one bit is 1, others 0
n
4
+ 1, · · · ,
√
0.3n√n
0.5n n
√
0.5n n
0
n
2
n
2
+ 1, · · · , n
√
√n
n
half bits are 1
The population size is n for Instances I and II. The number of maximum generations is 15n for the MSGA and 5n for the
MOGA. The initialisation of individuals in both MSGA and MOGA refer to the above tables.
Tables III gives experiment results of comparing the greedy algorithm, MSGA, greedy algorithm + MSGA and MOGA.
From the table, we observe that
• the solution quality of MSGA is better or not worse than the greedy algorithm in 20 random instances. However for
Instance I, the MSGA only finds a solution whose value is about 20% of the optimal value.
• the solution quality of greedy algorithm + MSGA is better or not worse than the greedy algorithm in all instances. However
for Instance II, the algorithm gets trapped into a local optimum, and is worse than the MOGA.
• the MOGA is the winner among 4 algorithms and its the solution quality is better or not worse than the greedy algorithm
and MSGA in all instances.
V. C ONCLUSIONS
A novel MOGA using helper objectives is proposed in this paper for solving the 0-1 knapsack problem. First the original
0-1 knapsack problem is recast into a multi-objective optimization problem (i.e. to maximize the sum of profits packed in the
6
TABLE III
A
COMPARISON AMONG 4 ALGORITHMS IN 20 RANDOMLY GENERATED INSTANCES AND 2 SPECIAL INSTANCES . T HE FIRST 10 INSTANCES BELONG TO
THE RESTRICTIVE CAPACITY KNAPSACK PROBLEM . T HE SECOND 10 INSTANCES BELONG TO THE AVERAGE CAPACITY KNAPSACK PROBLEM . ‘ MAX ’:
THE MAXIMUM VALUE OF f (~
x) PRODUCED DURING 10 RUNS . ‘AVERAGE ’: THE AVERAGE VALUE OF f (~
x) OVER 10 RUNS . ‘ STDEV ’: THE STANDARD
DERIVATION OF f (~
x) IN 10 RUNS .
Greedy
Instance
MSGA
Greedy + MSGA
MOGA
max
average
stdev
max
average
stdev
max
average
stdev
1
2
3
4
5
6
7
8
9
10
674
714
561
631
585
787
736
1042
982
906
683
714
622
631
621
787
773
1076
994
942
683
714
622
631
621
787
773
1076
994
942
0
0
0
0
0
0
0
0
0
0
683
714
622
631
621
787
773
1076
994
942
683
714
622
631
620.7
787
773
1076
993
942
0
0
0
0
0.95
0
0
0
3.16
0
683
714
622
631
621
787
773
1076
994
942
681.2
714
622
631
620.7
787
773
1076
993
942
1.55
0
0
0
0.95
0
0
0
3.16
0
11
12
13
14
15
16
17
18
19
20
4107
4090
4138
3901
3997
3984
3820
3914
4456
4149
4111
4102
4169
3925
4047
3994
3848
3920
4471
4177
4110.9
4100.6
4168.6
3923.9
4047
3993
3848
3919.4
4470.1
4175.3
0.32
1.43
1.26
1.29
0
1.05
0
1.90
0.57
1.34
4111
4102
4169
3925
4047
3994
3848
3920
4470
4177
4111
4097.1
4165.6
3922.2
4043.2
3992.5
3845.1
3915.2
4465.4
4171.9
0
4.41
2.37
2.15
3.71
0.85
1.97
2.53
2.59
2.73
4111
4102
4169
3927
4047
3994
3848
3920
4471
4177
4111
4101.2
4169
3923.7
4047
3993.8
3848
3920
4470.4
4176
0
1.03
0
1.95
0
0.632
0
0
0.97
0.82
I
II
416.2
1402.1
83.4
1402.1
83.4
1402.1
0
0
416.2
1402.1
416.2
1402.1
0
0
416.2
1414.2
416.2
1414.2
0
0
knapsack, to maximize the average profit-to-weight ratio of items, to maximize the average profit of items, and to maximize the
number of packed items). Then a MOGA (using bitwise mutation, one-point crossover and multi-criterion selection plus a mixed
strategy of three repair methods) is designed for the multi-objective optimization problem. Experiment results demonstrate that
the MOGA using helper objectives outperforms its rivals, which are the greedy algorithm, MSGA and greedy algorithm +
MSGA. The results also show that the MSGA can find reasonably good solutions but without a guarantee; and the greedy
algorithm + MSGA sometimes gets trapped into a local optimum.
R EFERENCES
[1] S. Martello and P. Toth, Knapsack Problems. Chichester: John Wiley & Sons, 1990.
[2] Z. Michalewicz and J. Arabas, “Genetic algorithms for the 0/1 knapsack problem,” in Methodologies for Intelligent Systems. Springer, 1994, pp.
134–143.
[3] S. Khuri, T. Bäck, and J. Heitkötter, “The zero/one multiple knapsack problem and genetic algorithms,” in Proceedings of the 1994 ACM Symposium
on Applied Computing. ACM, 1994, pp. 188–193.
[4] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, 3rd ed. New York: Springer Verlag, 1996.
[5] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A comparative case study and the strength pareto approach,” IEEE Transactions on
Evolutionary Computation, vol. 3, no. 4, pp. 257–271, 1999.
[6] A. Jaszkiewicz, “On the performance of multiple-objective genetic local search on the 0/1 knapsack problem-a comparative experiment,” IEEE
Transactions on Evolutionary Computation, vol. 6, no. 4, pp. 402–412, 2002.
[7] M. Eugénia Captivo, J. Clı̀maco, J. Figueira, E. Martins, and J. Luis Santos, “Solving bicriteria 0–1 knapsack problems using a labeling algorithm,”
Computers & Operations Research, vol. 30, no. 12, pp. 1865–1886, 2003.
[8] J. D. Knowles, R. A. Watson, and D. W. Corne, “Reducing local optima in single-objective problems by multi-objectivization,” in Evolutionary MultiCriterion Optimization. Springer, 2001, pp. 269–283.
[9] M. T. Jensen, “Helper-objectives: Using multi-objective evolutionary algorithms for single-objective optimisation,” Journal of Mathematical Modelling
and Algorithms, vol. 3, no. 4, pp. 323–347, 2005.
[10] J. Handl, S. C. Lovell, and J. Knowles, “Multiobjectivization by decomposition of scalar cost functions,” in Parallel Problem Solving from Nature–PPSN
X. Springer, 2008, pp. 31–40.
[11] D. Brockhoff, T. Friedrich, N. Hebbinghaus, C. Klein, F. Neumann, and E. Zitzler, “On the effects of adding objectives to plateau functions,” IEEE
Transactions on Evolutionary Computation, vol. 13, no. 3, pp. 591–603, 2009.
[12] D. F. Lochtefeld and F. W. Ciarallo, “Helper-objective optimization strategies for the job-shop scheduling problem,” Applied Soft Computing, vol. 11,
no. 6, pp. 4161–4174, 2011.
[13] R. Kumar and N. Banerjee, “Analysis of a multiobjective evolutionary algorithm on the 0–1 knapsack problem,” Theoretical Computer Science, vol.
358, no. 1, pp. 104–120, 2006.
[14] T. Friedrich, J. He, N. Hebbinghaus, F. Neumann, and C. Witt, “Approximating covering problems by randomized search heuristics using multi-objective
models,” Evolutionary Computation, vol. 18, no. 4, pp. 617–633, 2010.
[15] X. Lai, Y. Zhou, J. He, and J. Zhang, “Performance analysis of evolutionary algorithms for the minimum label spanning tree problem,” IEEE Transactions
on Evolutionary Computation, 2014, (accpeted, online).
[16] J. He and Y. Zhou, “A comparison of GAs using penalizing infeasible solutions and repairing infeasible solutions II,” in Proceedings of the 2nd
International Symposium on Intelligence Computation and Applications. Wuhan, China: Springer, 2007, C1, pp. 102–110.
| 9cs.NE
|
Prioritized Norms in Formal Argumentation
Beishui Liao1,3 , Nir Oren2 , Leendert van der Torre3 and Serena Villata4
1
2
arXiv:1709.08034v2 [cs.AI] 28 Feb 2018
3
Zhejiang University, China
University of Aberdeen, United Kingdom
University of Luxembourg, Luxembourg
4
CNRS, I3S Laboratory, France
Abstract. To resolve conflicts among norms, various nonmonotonic formalisms
can be used to perform prioritized normative reasoning. Meanwhile, formal argumentation provides a way to represent nonmonotonic logics. In this paper, we
propose a representation of prioritized normative reasoning by argumentation.
Using hierarchical abstract normative systems, we define three kinds of prioritized normative reasoning approaches, called Greedy, Reduction, and Optimization. Then, after formulating an argumentation theory for a hierarchical abstract
normative system, we show that for a totally ordered hierarchical abstract normative system, Greedy and Reduction can be represented in argumentation by
applying the weakest link and the last link principles respectively, and Optimization can be represented by introducing additional defeats capturing the idea that
for each argument that contains a norm not belonging to the maximal obeyable
set then this argument should be rejected.
1
Introduction
Since the work of Alchourrón and Makinson [1] on hierarchies of regulations and their
logic, in which a partial ordering on a code of laws or regulations is used to overcome
logical imperfections in the code itself, reasoning with prioritized norms has been a
central challenge in deontic logic [12,4,23].
The goal of this paper is to study the open issue of reasoning with priorities over
norms through the lens of argumentation theory [10]. More precisely, we focus on reasoning with the abstract normative system proposed by Tosatto et al. [28], which in turn
is based on Makinson and van der Torre’s approach to input/output logic [20]. In this
system, an abstract norm is represented by an ordered pair (a, x), where the body of the
norm a is thought of as an input, representing some kind of condition or situation, and
the head of the norm x is thought of as an output, representing what the norm tells us
to be obligatory in that situation a. As a consequence, an abstract normative system is
a directed graph (L, N ) together with a context C ⊆ L capturing base facts, where L
is a set of nodes, and N ⊆ L × L is the set of abstract norms. When the edge of an
abstract normative system is associated with a number to indicate its priority over the
other norms in the system, we obtain a hierarchical abstract normative system (HANS),
which will be formally defined and studied in the remainder of this paper.
Let us clarify how a hierarchical abstract normative system is defined by considering
the well known Order Puzzle [17] example from the deontic logic literature, which
revolves around three norms.
2
B. Liao N. Oren, L. van der Torre and S. Villata
“Suppose that there is an agent, called Corporal O’Reilly, and that he is subject
to the commands of three superior officers: a Captain, a Major, and a Colonel.
The Captain, who does not like to be cold, issues a standing order that, during
the winter, the heat should be turned on. The Major, who is concerned about
energy conservation, issues an order that, during the winter, the window should
not be opened. And the Colonel, who does not like to be too warm and does
not care about energy conservation, issues an order that, whenever the heat is
on, the window should be opened.”
Hierarchical Abstract
Normative System
1 : (w, h)
2 : (w, ¬o)
3 : (h, o)
Extensions
1
>
w
h
2
3
o
Greedy :{h, ¬o}
Reduction :{h, o}
Optimization :{¬o}
Fig. 1. The Order puzzle example, represented using the graphical notation of Tosatto et al. [29]
with edges annotated by norm strength.
Let w, h and o respectively denote the propositions that it is winter, the heat is
turned on, and the window is open. There are three norms (w, h), (w, ¬o) and (h, o).
These three norms are visualized in Figure 1, extending the graphical notation described in Tosatto et al. [29] by associating edges with numbers denoting priorities of
norms. These priorities are obtained from the rank of the issuer, since Colonels outrank
Majors, and Majors outrank Captains. Within the figure, each circle denotes a proposition; the light part of the circle is the proposition itself, while the dark part denotes
a negated proposition. Dashed lines represent the conditional obligations. Within Figure 1, the line from the light part of w to the dark part of o denotes (w, ¬o). The box on
the left represents the context, in the example containing > and w.
In formalizing examples, one problem in applied logic is that the representation
may be challenged. For example, it may be argued that the Colonel implies that if the
window is closed, then the heating should be turned off. However, in normative systems,
such pragmatic considerations are usually not part of the detachment procedure [16],
with only explicitly given norms and commands being considered. Therefore, any such
additional interpretations or other pragmatic concerns are out of the scope of this paper.
The central notion of inference in normative systems is called detachment. For example, in the Order Puzzle, the fundamental question is whether we can detach o, ¬o, or
both. In the example, the formulas which can be derived from a normative system are
obligations. In general, permissions and institutional facts can also be detached from
normative systems. A detachment procedure therefore defines the way deontic facts are
derived from a normative system. Different detachment procedures have been defined
and studied in deontic logic, as well as in other rule based systems. Moreover, even in
hierarchical normative systems, not all conflicts may be resolved. In such a case, the
Prioritized Norms in Formal Argumentation
3
detachment procedure may derive several so-called extensions, each representing a set
of obligations, permissions and institutional facts.
Abstract normative systems [28] were introduced as a common core representation
for normative systems, which are still expressive enough to define the main detachment
procedures. In particular, analogous to the main input/output logics, they have factual
detachment built in, and have reasoning by cases, deontic detachment and identity as
optional inference patterns [3,16]. Such systems are called ‘abstract’, because negation
is the only logical connective that is defined in the language. Furthermore, Tosatto et
al. [28] considered elements and anti-elements rather than literals and propositions. It is
straightforward to define more connectives within such systems, and it is also possible
to define structured normative systems where the abstract elements are instantiated with
logical formulas, for example with formulas of a propositional or modal logic. The latter
more interesting representation of logical structure is analogous to the use of abstract
arguments in formal argumentation. An advantage of abstract normative systems over
structured ones is that the central inference of detachment can be visualized by walking
paths in the graph. In other words, inference is represented by graph reachability. For
example, in the Order puzzle, node o is reachable from the context, and thus it can be
detached. Moreover, a conflict is represented by a node where both its light and the dark
side are reachable from the context, as in the case of node o in Figure 1.
There are several optional inference patterns for abstract normative systems, because, as is well known, most principles of deontic logic have been criticized for some
examples and applications. However, the absence of the same inference patterns is often
criticized as well due to the lack of explanations and predictions of the detachment procedures resulting from these patterns. Therefore, current approaches to represent and
reason with normative systems, such as input/output logic as well as abstract normative systems, do not restrict themselves to a single logic, but define a family of logics
which can be embedded within them. Deciding which logic to use in a specific context
depends on the requirements of the application. Similarly, with regards to permissions,
there is an even larger diversity of deontic logics [15] which adopt different representations. For example, for each input/output logic, various notions of permission have
been defined, in terms of their relation to obligation. We refer to the work of Parent and
van der Torre [23] for further explanation, discussion and motivation of this topic.
Now let us consider how various detachment procedures might apply norms in hierarchical normative systems in different orders, and result in different outcomes or
extensions. Since detachment operates differently over permissions and obligations,
and since the former introduces significant extra complexity, we only consider regulative norms and obligation detachment procedures in this paper. We examine three
approaches describing well known procedures defined in the literature [32,5,14]. These
procedures all share one important property, namely that the context itself is not necessarily part of the procedure’s output. It is precisely this feature which distinguishes
input/output logics from traditional rule based languages like logic programming or
default logic [26,6]. Such traditional rule based languages where the input is part of
the output are called throughput operators in input/output logic research. Therefore, the
procedures we consider can naturally be captured using input/output logics. We refer to
the three approaches as Greedy, Reduction and Optimization.
4
B. Liao N. Oren, L. van der Torre and S. Villata
Greedy: The context contains propositions that are known to hold. This procedure
always applies the norm with the highest priority that does not introduce inconsistency
to an extension and the context. Here we say that a norm is applicable when its body is
in the context or has been produced by other norms and added to the extension. In the
Order puzzle example, we begin with the context {w}, and (w, ¬o) is first applied. Then
(w, h) is applied. Finally, (h, o) cannot be applied as this would result in a conflict, and
so, by using Greedy, we obtain the extension {h, ¬o}.
Reduction: in this approach, a candidate extension is guessed. All norms that are
applicable according to this candidate extension are selected and transformed into unconditional or body-free norms. For example, a norm (a, b) selected in this way is transformed to a norm (>, b). The modified hierarchical abstract normative system, with the
transformed norms is then evaluated using Greedy. The candidate extension is selected
as an extension by Reduction if it is identified as an extension according to this application of Greedy. In our example, we select a candidate extension {h, o}, obtaining a set
of body-free norms {(>, h), (>, ¬o), (>, o)}. The priorities assigned to these norms
are carried through from the original hierarchical abstract normative system, and are
therefore respectively 1, 2 and 3. After applying Greedy, we get an extension of Reduction: {h, o}. However, if we had selected the candidate extension {h, ¬o}, then this
new extension would not appear in Greedy as (>, ¬o) has a lower priority than (>, o),
and the latter is therefore not an extension of Reduction.
Optimization: In terms of Hansen’s prioritized conditional imperatives, a set of maximally obeyable (i.e., minimally violated) norms is selected by choosing norms in order
of priority which are consistent with the context. Once these norms are selected, Greedy
is applied to identify the extension. In our example, the maximal set of obeyable norms
is {(h, o), (w, ¬o)}. Optimization therefore detaches the unique extension {¬o}.
We can also consider the example in terms of formal argumentation. Given a hierarchical abstract normative system, we may construct an argumentation framework.
The top part of Figure 2 illustrates the argumentation framework obtained for the Order
puzzle which does not consider the priority relation between arguments.
An argumentation framework is a directed graph in which nodes denote arguments,
and edges denote attacks between arguments. In the setting of a hierarchical abstract
normative system, an argument is represented as a path within the directed graph starting from a node in the context. In this example, there are four arguments A0 , A1 , A2
and A3 , represented as [w], [(w, h)], [(w, h), (h, o)] and [(w, ¬o)], respectively. Since
the conclusions of A2 and A3 are inconsistent, A2 attacks A3 and vice-versa. To obtain
correct conclusions, we must take priorities between arguments into account, transforming attacks into defeats. This in turn requires lifting the priorities given for the constituents of an argument to a priority for the argument itself, and two different principles
have been commonly used for such a lifting. The last link principle ranks an argument
based on the strength of its last inference, while the weakest link principle ranks an
argument based on the strength of its weakest inference. In the order example, if the
last link principle is applied, then [(w, h), (h, o)] defeats [(w, ¬o)]. The corresponding
argumentation framework is illustrated in the middle portion of Figure 2. If the weakest
link principle is used instead, then [(w, ¬o)] defeats [(w, h), (h, o)]. The corresponding
argumentation framework is illustrated at the bottom of Figure 2. As a result, the former
Prioritized Norms in Formal Argumentation
5
principle allows us to conclude {h, o}, while the latter concludes {h, ¬o}. In turn, the
first result coincides with that obtained by Reduction, while the second is the same as
that obtained by Greedy.
A0
[w]
A0
[w]
A0
[w]
A1
A2
A3
[(w, h)] [(w, h), (h, o)] [(w, ¬o)]
A1
A2
A3
[(w, h)] [(w, h), (h, o)] [(w, ¬o)]
A1
A2
A3
[(w, h)] [(w, h), (h, o)] [(w, ¬o)]
Fig. 2. Top: the argumentation framework obtained from the order puzzle hierarchical normative
system with the four arguments and attacks between them visualized as directed arrows. Middle: the argumentation framework obtained when last link is applied. Bottom: the argumentation
framework obtained when weakest link is applied.
Inspired by the example above, we wish to investigate the links between the three
detachment procedures for prioritized normative reasoning and argumentation theory.
More specifically, our main research question is as follows.
How can the three detachment procedures (Reduction, Greedy, and Optimization) proposed in the context of abstract normative reasoning be represented in
formal argumentation?
To answer this research question, we propose a formal framework to connect hierarchical normative reasoning with argumentation theory. More precisely, our framework
represents the above-mentioned detachment procedures by lifting priorities from norms
to arguments, with the underlying goal of making as few commitments as possible to
specific argumentation systems. For this reason, we build on a structured argumentation framework which admits undercuts and rebuts between arguments, and allows for
priorities between norms making up arguments. We show that variants of approaches to
lifting priorities from rules to arguments allow us to capture both Greedy and Reduction, while the introduction of additional defeats allows us to obtain Optimization.
Some preliminary results found in this paper were originally presented in DEON
2016 [19]. We keep, but extend, the representation results for Greedy and Reduction (previously called the “Brewka-Eiter construction”). To capture Optimization (previously called “Hansen’s construction”) the DEON 2016 paper adds new arguments
through the introduction of additional permissive norms. In the current paper, this is
6
B. Liao N. Oren, L. van der Torre and S. Villata
replaced with an approach which introduces additional defeats. The reasons for this
replacement are two-fold. First, the preliminary approaches for Optimization from the
earlier work [19] cannot fully capture the idea that each argument containing a norm not
belonging to the maximal obeyable set should be rejected. Second, permissive norms
were used as a means to support the representation, which making the approach more
complicated.
The layout of the remainder of the paper is as follows. Section 2 formalizes the
above-mentioned three detachment procedures of hierarchical normative reasoning (i.e.,
Greedy, Reduction, and Optimization). In Section 3, we introduce an argumentation theory for a hierarchical abstract normative system. Sections 4, 5 and 6 show how Greedy,
Reduction and Optimization can be represented in argumentation. Finally, in Section 7
we discuss open problems and compare the proposed approach with related work, and
in Section 8 we point out possible directions for future work.
2
Hierarchical abstract normative systems
In this section, we formally introduce the notion of hierarchical abstract normative systems and the three detachment procedures to compute their normative conclusions. A
hierarchical abstract normative system captures the context of a system and the norms in
force in such a system. There is an element in the universe called >, contained in every
context. In this paper, we consider only a finite universe. A hierarchical abstract normative system also encodes a ranking function over the norms to allow for the resolution
of conflicts.
Based on the notion of abstract normative system defined by Tosatto et al. [28], a
hierarchical abstract normative system can be defined as follows.
Definition 1 (Hierarchical abstract normative system). A hierarchical abstract normative system is a tuple H = hL, N, C, ri, where
– L = E ∪ {¬e | e ∈ E} ∪ {>} is the universe, a set of literals based on some finite
set E of atomic elements;
– N ⊆ L × L is a finite set of regulative norms;
– C ⊆ L is a subset of the universe, called a context, such that > ∈ C and for all e
in E, {e, ¬e} 6⊆ C;
– r : N → IN is a function from norms to natural numbers.
Regulative (ordinary) norms are of the kind “if you turn on the heat, then you should
open the window”. These norms are conditional norms, requiring some condition to
hold (e.g., turning on the heat) before their conclusions can be drawn.
We write (a, x) for a regulative norm, where a, x ∈ L are the antecedent and conclusion of the norm, respectively. Given (a, x), we use r(a, x) to denote r((a, x)). Let
u, v ∈ N be two norms, we say that v is at least as preferred as u (denoted u ≤ v)
if and only if r(u) is not larger than r(v) (denoted r(u) ≤ r(v)), where r(u) is also
called the rank of u. We write u < v or v > u if and only if u ≤ v and v 6≤ u. Given
a norm u = (a, x), we write ant(u) for a to represent the antecedent of the norm, and
cons(u) for x to represent the consequent of the norm. Given a set of norms S ⊆ N ,
Prioritized Norms in Formal Argumentation
7
we use cons(S) to denote {cons(u) | u ∈ S}. We say that a hierarchical abstract normative system is totally ordered if and only if the ordering ≤ over N is antisymmetric,
transitive and total. Due to the finiteness of universe, the set of norms is finite. Note that
given this assumption, the notion of total ordering here is identical to that of the full
prioritization in Brewka and Eiter’s [7] and Hansen’s [13] work, and of the linearized
ordering of Young et al. [32]. For a ∈ L, we write a = ¬a if and only if a ∈ E, and
a = e for e ∈ E if and only if a = ¬e. For a set S ⊆ L, we say that S is consistent if
and only if there exist no e1 , e2 ∈ S such that e1 = e2 . To exemplify the notions of a
hierarchical abstract normative system, consider the following example.
Example 1 (Order puzzle). In terms of Definition 1, the set of norms and priorities
that are visualized in Figure 1 can be formally represented as a hierarchical abstract
normative system H = hL, N , C, ri, where L = {w, h, o, ¬w, ¬h, ¬o, >}, N =
{(w, h), (h, o), (w, ¬o)}, C = {w, >}, r(w, h) = 1, r(h, o) = 3, r(w, ¬o) = 2.
In the hierarchical abstract normative system setting, the three detachment procedures for prioritized normative reasoning can be defined as follows.
Firstly, Greedy detachment for a hierarchical abstract normative system always applies the norm with the highest priority among those which can be applied, if this does
not bring inconsistency to the extension and the context. To formally define the notion
of Greedy, we first introduce the following notions of paths and consistent paths.
Definition 2 (Path, consistent path). Let H = hL, N, C, ri be a hierarchical abstract
normative system.
– A path in H from x1 to xn is a sequence of norms (x1 , x2 ), (x2 , x3 ), . . . , (xn−1 , xn )
such that {(x1 , x2 ), (x2 , x3 ), . . . , (xn−1 , xn )} ⊆ N , n ≥ 2, and all norms of the
sequence are distinct.
– A path in H from x1 to xn with respect to R ⊆ N is a sequence of norms (x1 , x2 ),
(x2 , x3 ), . . . , (xn−1 , xn ) such that {(x1 , x2 ), (x2 , x3 ), . . . , (xn−1 , xn )} ⊆ R, n ≥
2, and all norms of the sequence are distinct.
– A consistent path in H from x1 to xn (with respect to R) is a path (x1 , x2 ),
(x2 , x3 ), . . . , (xn−1 , xn ) such that {x1 , x2 , . . . , xn } is consistent.
Then, the unique extension of a totally ordered hierarchical abstract normative system by Greedy can be defined as follows by first selecting a set of applicable norms
with the highest priority in each step, and then collecting the final vertex of each path
with respect to the set of selected norms.
Definition 3 (Greedy). Let H = hL, N, C, ri be a totally ordered hierarchical abstract
normative system. For all R ⊆ N , let R(C) = {x | there is a path in H from an element
in C to x with respect to R}, and Appl(N, C, R) := {(a, x) ∈ N \R | a ∈ C ∪cons(R),
{x, x}6⊆ C ∪ cons(R)}. The extension of H by Greedy, written as Greedy(H), is the set
R(C) such that R = ∪∞
i=0 Ri is built inductively as follows.
R0 = ∅
Ri+1 = Ri ∪ max(N, C, Ri , r)
where max(N, C, Ri , r) = {u ∈ Appl(N, C, Ri ) | ∀v ∈ Appl(N, C, Ri ) : r(u) ≥
r(v)}.
8
B. Liao N. Oren, L. van der Torre and S. Villata
In terms of Definition 3, the unique extension of the Order puzzle can be constructed
as follows.
Example 2 (Extension by Greedy). Given H in Example 1, by Greedy, it holds that
R0 = ∅, R1 = {(w, ¬o)}, R2 = {(w, ¬o), (w, h)}, R = {(w, ¬o), (w, h)}. So,
Greedy(H) = {h, ¬o}.
Reduction is defined by first using a candidate extension to get a modified hierarchical abstract normative system, and then applying Greedy to it. If the candidate
extension is an extension according to this application of Greedy, then it is an extension
of the original hierarchical abstract normative system by Reduction. Formally, we have
the following definition.
Definition 4 (Reduction). Given a totally ordered hierarchical abstract normative system H = hL, N , C, ri, and a set X, let HX =hL, N 0 , C, r0 i, where N 0 = {(>, x2 ) |
(x1 , x2 ) ∈ N, x1 ∈ C∪X} is the set of ordinary norms, and r0 (>, x2 ) = max(r(x1 , x2 ) |
(x1 , x2 ) ∈ N, x1 ∈ C ∪ X) for all (x1 , x2 ) ∈ N are priorities over norms. An extension of H by Reduction is a set U such that U is Greedy(HU ). The set of extensions of
H by Reduction is denoted as Reduction(H).
According to Definition 4, it can be the case that different norms have the same
consequent. To avoid the duplication of multiple body-free norms, only a single norm
with the highest priority is used.
To exemplify the notion of Reduction, consider again the Order puzzle in Example 1.
Example 3 (Extensions by Reduction). By using Reduction, given X = {h, o}, we have
HX = hL, N 0 , C, r0 i, where N 0 = {(>, h), (>, o), (>, ¬o)}, r0 (>, h) = 1, r0 (>, o) =
3 and r0 (>, ¬o) = 2. Since X ∈ Greedy(HX ), and no other set can be an extension,
we have that Reduction(H) = {{h, o}}.
Note that a totally ordered hierarchical abstract normative system might have more
than one extension, as illustrated by the following example.
Example 4 (Multiple extensions by Reduction). Given the hierarchical abstract normative system in Figure 3, assume that we have a context C = {a}. We then consider
X1 = {b, c} and X2 = {¬b}. In the first case, we have HX1 = hL, N 0 , C, r0 i where
N 0 = {(>, b), (>, ¬b), (>, c)} and r0 (>, b) = 4, r0 (>, c) = 3, r0 (>, ¬b) = 2. Here,
Greedy(HX1 ) = {b, c}, i.e., X1 .
In the second case, we obtain HX2 = hL, N 0 , C, r0 i where N 0 = {(>, b), (>, ¬b)}
and r0 (>, b) = 1, r0 (>, ¬b) = 2. Now, Greedy(HX2 ) = {¬b}, i.e., X2 . In this case,
we therefore obtain two extensions using Reduction.
Finally, Optimization can be defined by first choosing a set of maximally obeyable
norms, and then applying Greedy to it.
Prioritized Norms in Formal Argumentation
>
4
1
a
2
b
3
c
9
1 : (a, b)
2 : (a, ¬b)
3 : (b, c)
4 : (c, b)
Fig. 3. The hierarchical abstract normative system of Example 3 containing the two Reduction
extensions {b, c} and {¬b}.
Definition 5 (Optimization). Given a totally ordered hierarchical abstract normative
system H = hL, N , C, ri, let T = (u1 , u2 , . . . , un ) be the linear order on N such that
u1 > u2 > · · · > un . We define a sequence of sets of norms R0 . . . Rn as follows:
R0 = ∅
Ri ∪ {ui }, if C ∪ R(C) is consistent where R = Ri ∪ {ui }
Ri+1 =
Ri , else
The unique extension of H by Optimization is Rn (C) and denoted as Optimization(H).
According to Definition 5, the unique extension of the Order puzzle can be constructed as follows.
Example 5 (Extension by Optimization). Regarding H in Example 1, by Optimization,
let u1 = (h, o), u2 = (w, ¬o), and u3 = (w, h), and T = (u1 , u2 , u3 ). Then, it holds
that R0 = ∅, R1 = {u1 }, R2 = {u1 , u2 }, and R = R3 = R2 = {u1 , u2 }. So, we
obtain that Optimization(H) = {¬o}.
3
Argumentation theory for a hierarchical abstract normative
system
In this section, we introduce an argumentation theory on prioritized norms. Given a
hierarchical abstract normative system, we first define arguments and defeats between
them, then compute extensions of arguments in terms of Dung’s theory [10], and from
these, obtain conclusions.
3.1
Arguments
In a hierarchical abstract normative system, an argument is an acyclic path in the graph
starting in an element of the context. We assume minimal arguments—no norm can be
applied twice in an argument and no redundant norm is included in an argument. We
use concl(α) to denote the conclusion of an argument α, and concl(E) = {concl(α) |
α ∈ E} for the conclusions of a set of arguments E.
Definition 6 (Arguments and sub-arguments). Let H = hL, N, C, ri be a hierarchical abstract normative system.
10
B. Liao N. Oren, L. van der Torre and S. Villata
A context argument in H is an element a ∈ C, and its conclusion is concl(a) = a.
An ordinary argument is a consistent path α in H from x ∈ C to some y ∈ L.
Moreover, we have that concl(α) = y.
The sub-arguments of argument [u1 , . . . , un ] are, for 1 ≤ i ≤ n, [u1 , . . . , ui ]. Note
that context arguments do not have sub-arguments.
The set of all arguments constructed from H is denoted as Arg(H). For readability,
[(a1 , a2 ), . . . , (an−1 , an )] may be written as (a1 , a2 , . . . , an−1 , an ). The set of subarguments of an argument α is denoted as sub(α). When α is a sub-argument of β,
we say that β is a super-argument of α. The set of super-arguments of α is denoted
as sup(α). Furthermore, the set of proper sub-arguments of α is defined as psub(α) =
sub(α) \ {α}. For an ordinary argument α = [u1 , . . . , un ], we call un the top norm
of α. Finally, in each ordinary argument, the norm with lowest priority is called the
weakest norm of the argument. Formally, we have the following definition for this latter
concept.
Definition 7 (The weakest norm of an argument). Let H be a totally ordered hierarchical abstract normative system. For an ordinary argument α = [u1 , . . . , un ] in H, if
r(ui ) = min(r(uj ) | 1 ≤ j ≤ n), then ui is called the weakest norm of α.
3.2
Defeat relation between arguments
We follow the tradition in much of preference-based argumentation [2,21], where attack captures a relation among arguments which ignores preferences, and defeat is a
preference-aware relation on which the semantics is based. To define the defeat relation
among prioritized arguments, we assume that only the priorities of the norms are used
to compare arguments. In other words, we assume a lifting of the ordering on norms to
a binary relation on sequences of norms (i.e., arguments), written as α β, where α
and β are two arguments, indicating that α is at least as preferred as β.
There is no common agreement about the best way to lift ≥ to . In argumentation,
there are at least two ways to introduce weights. As an argumentation framework consists of a set of arguments and an attack relation between them, we can either assign
weights to arguments, or we can assign weights to attacks. Traditionally, weights are assigned to arguments. Two common approaches to give the strength of an argument are
the weakest link and the last link principles, combined with the elitist and democratic
ordering [21]. For example, in the weakest link principle the weight of the argument is
the weight of the weakest rule used in the argument. However, Young et al. [32] showed
that elitist weakest link cannot be used to calculate for Greedy, and proposes a disjoint elitist order which ignores shared rules. It is worth noticing that the strength of an
argument may depend on the argument it is attacking, as identified by Young et al. [32].
Based on this idea, we define the orderings between arguments by assigning a strength
to the attacks between the arguments, to reflect the priority of the norms used in the
arguments, following the same insights of the weakest link and last link principles (denoted as w and l respectively). Defining the weakest link ordering in the same way
as Young et al. [32], we have the following definition.
Prioritized Norms in Formal Argumentation
11
Definition 8 (Weakest link and last link). Let H = hL, N , C, ri be a hierarchical
abstract normative system, and α = [u1 , . . . , un ] and β = [v1 , . . . , vm ] be two arguments in Arg(H). Let Φ1 = {u1 , . . . , un } and Φ2 = {v1 , . . . , vm }. By the weakest link
principle, α w β iff ∃v ∈ Φ2 \ Φ1 such that ∀u ∈ Φ1 \ Φ2 , v ≤ u. By the last link
principle, α l β iff un ≥ vm .
When the context is clear, we write for w , or l . We write α β for α β
without β α. The following proposition shows the transitivity of the two relations
w and l .
Proposition 1 (Transitivity). It holds that the relations w and l are transitive.
Proof. Let α = [u1 , . . . , un ], β = [v1 , . . . , vm ] and γ = [w1 , . . . , wk ] be three arguments in Arg(H). For the case of the weakest link, let Φ1 = {u1 , . . . , un }, Φ2 =
{v1 , . . . , vm }, and Φ3 = {w1 , . . . , wk }, n, m, k ≥ 1. Let x12 ∈ Φ1 ∩Φ2 , x13 ∈ Φ1 ∩Φ3 ,
and x23 ∈ Φ2 ∪ Φ3 , x1 ∈ Φ1 \ (Φ2 ∪ Φ3 ), x2 ∈ Φ2 \ (Φ1 ∪ Φ3 ), and x3 ∈ Φ3 \ (Φ1 ∪ Φ2 ).
Assume that α w β and β w γ. There are only the following four possible cases.
Case 1: There exists x23 ∈ Φ2 , for all x13 , x1 ∈ Φ1 , x23 ≤ x13 and x23 ≤ x1 ; and
since x13 ∈ Φ3 , assume that for all x12 , x2 ∈ Φ2 , x13 ≤ x12 , x13 ≤ x2 . It follows that
x23 ≤ x12 and x23 ≤ x1 . Since x23 ∈ Φ3 , it means that there exists x23 ∈ Φ3 such that
x23 ≤ x12 and x23 ≤ x1 where x12 , x1 ∈ Φ1 . Hence, α w γ.
Case 2: There exist x23 ∈ Φ2 and x3 ∈ Φ3 , for all x13 , x1 ∈ Φ1 , and x12 , x2 ∈ Φ2 :
x23 ≤ x13 , x23 ≤ x1 , x3 ≤ x12 , x3 ≤ x2 . In this case, there are in turn only the
following two possible sub-cases: either x23 ≤ x3 or x23 > x3 . If x23 ≤ x3 , since
x3 ≤ x12 , it holds that x23 ≤ x12 . Since x23 ≤ x1 and x23 ≤ x12 , it holds that
α w γ. Second, if x23 > x3 , since x23 ≤ x1 , x3 ≤ x1 . Since x3 ≤ x12 and x3 ≤ x1 ,
α w γ.
Case 3: There exists x2 ∈ Φ2 , for all x13 , x1 ∈ Φ1 , x2 ≤ x13 and x2 ≤ x1 ; and
since x13 ∈ Φ3 , assume that for all x12 , x2 ∈ Φ2 , x13 ≤ x12 , x13 ≤ x2 . In this case,
there are in turn only the following two possible sub-cases: x23 ≤ x13 or x3 ≤ x13 ,
or x23 > x13 and x3 > x13 . If x23 ≤ x13 and x3 ≤ x13 , it holds that α w γ. If
x23 > x13 and x3 > x13 , it holds that x13 < x23 and x13 ≤ x2 , and therefore β w α.
Contradiction.
Case 4: There exist x2 ∈ Φ2 and x3 ∈ Φ3 , for all x13 , x1 ∈ Φ1 , and x12 , x2 ∈ Φ2 :
x2 ≤ x13 , x2 ≤ x1 , x3 ≤ x12 , x3 ≤ x2 . In this case, since x3 ≤ x12 and x3 ≤ x1 , it
holds that α w γ.
For last link, if α l β and β l γ, then un ≥ vm and vm ≥ wk . It follows that
un ≥ wk , and therefore α l γ. So, it holds that l is transitive.
Given a way to lift the ordering on norms to an ordering on arguments, the notion
of defeat can be defined as follows.
Definition 9 (Defeat among arguments). Let H = hL, N , C, ri be a hierarchical
abstract normative system. For all α, β ∈ Arg(H),
α attacks β iff β has a sub-argument β 0 such that
1. concl(α) = concl(β 0 )
12
B. Liao N. Oren, L. van der Torre and S. Villata
α defeats β iff β has a sub-argument β 0 such that
1. concl(α) = concl(β 0 ) and
2. α is a context argument; or α is an ordinary argument and α β.
The set of defeats between the arguments in Arg(H) based on a preference ordering
is denoted as Def(H, ).
In what follows, an argument α = [u1 , . . . , un ] with ranking on norms is denoted
as u1 . . . un : r(α), where r(α) = (r(u1 ), . . . , r(un )).
The notions of arguments and defeat relations between arguments by the weakest
link and the last link principles respectively can be illustrated by the following example.
Example 6 (Order puzzle continued). Consider the hierarchical abstract normative system in Example 1. We have the following arguments (visually presented in the top of
Figure 2):
A0 : w
A1 : (w, h) : (1)
A2 : (w, h)(h, o) : (1, 3)
A3 : (w, ¬o) : (2)
(context argument)
(ordinary argument)
(ordinary argument)
(ordinary argument)
We have that A2 attacks A3 and vice versa, and there are no other attacks among the
arguments. Moreover, A2 defeats A3 by the last link principle (Figure 2, middle), and
A3 defeats A2 by the weakest link principle (Figure 2, bottom).
3.3
Argument extensions and conclusion extensions
Given a set of arguments A = Arg(H) and a set of defeats R = Def(H, ), we get an
argumentation framework (AF) F = (A, R).
Following the notions of abstract argumentation by Dung [10], we say that a set
B ⊆ A is admissible, if and only if it is conflict-free and it can defend each argument
within the set. A set B ⊆ A is conflict-free if and only if there exist no arguments α
and β in B such that (α, β) ∈ R. Argument α ∈ A is defended by a set B ⊆ A (in
such a situation α can also be said to be acceptable with respect to B) if and only if for
all β ∈ A, if (β, α) ∈ R, then there exists γ ∈ B such that (γ, β) ∈ R. Based on the
notion of admissible sets, some other extensions can be defined. Formally, we have the
following.
Definition 10 (Conflict-freeness, defense and extensions). Let F = (A, R) be an
argumentation framework, and B ⊆ A a set of arguments.
– B is conflict-free if and only if @α, β ∈ B, such that (α, β) ∈ R.
– An argument α ∈ A is defended by B (equivalently α is acceptable with respect to
B), if and only if ∀(β, α) ∈ R, ∃γ ∈ B, such that (γ, β) ∈ R.
– B is admissible if and only if B is conflict-free, and each argument in B is defended
by B.
– B is a complete extension if and only if B is admissible and each argument in A
that is defended by B is in B.
Prioritized Norms in Formal Argumentation
13
– B is a preferred extension if and only if B is a maximal (with respect to setinclusion) complete extension.
– B is a grounded extension if and only if B is the minimal (with respect to setinclusion) complete extension.
– B is a stable extension if and only if B is conflict-free, and ∀α ∈ A \ B, ∃β ∈ B
such that (β, α) ∈ R.
A semantics describes the set of extensions one wishes to obtain. We use sem ∈
{cmp, prf, grd, stb} to denote the complete, preferred, grounded, and stable semantics,
respectively. A set of argument extensions of F = (A, R) is denoted as sem(F). We
write Outfamily for the set of conclusions from the extensions of the argumentation
theory [30].
Definition 11 (Conclusion extensions). Given a hierarchical abstract normative system H = hL, N, C, ri, let F = (Arg(H), Def(H, )) be the AF constructed from H.
The conclusion extensions, written as Outfamily(F, sem), are the conclusions of the
ordinary arguments within argument extensions.
Outfamily(F, sem) = {{concl(α) | α ∈ S, α is an ordinary argument} | S ∈ sem(F)}
The following example shows that for a hierarchical abstract normative system,
when adopting different principles for lifting the priorities over norms to those over arguments, the resulting argumentation frameworks as well as their conclusion extensions
may be different.
Example 7 (Order puzzle in argumentation). According to Example 6, let A = {A0 ,
. . . , A3 }. We have F1 = (A, {(A2 , A3 )}) where A2 l A3 , and F2 = (A, {(A3 , A2 )})
where A3 w A2 . For all sem ∈ {cmp, prf, grd, stb}, Outfamily(F1 , sem) = {{h, o}},
and Outfamily(F2 , sem) = {{h, ¬o}}.
We now turn our attention to the properties of the argumentation theory for a hierarchical abstract normative system.
First, according to Definition 9, we have the following proposition, capturing the
relation between the attack/defeat on an argument and on its super-arguments.
Proposition 2 (Super-argument attack and defeat). Let F = (A, R) be an AF constructed from a hierarchical abstract normative system. For all α, β ∈ A, if α attacks β,
then α attacks all super-arguments of β; if α defeats β, α defeats all super-arguments
of β.
Proof. When α attacks β, according to Definition 9, β has a sub-argument β 0 such
that concl(α) = concl(β 0 ). Let γ be an super-argument of β. It follows that β 0 is a
sub-argument of γ. Hence, α attacks γ.
When α defeats β, according to Definition 9, β has a sub-argument β 0 such that
concl(α) = concl(β 0 ) and α is a context argument; or α is an ordinary argument and
α β 0 , Since β 0 is a sub-argument of γ, α defeats γ.
Second, corresponding to properties of sub-argument closure and direct consistency
in ASPIC+ [21], we have the following two properties.
14
B. Liao N. Oren, L. van der Torre and S. Villata
Proposition 3 (Closure under sub-arguments). Let F = (A, R) be an AF constructed from a hierarchical abstract normative system. For all sem ∈ {cmp, prf, grd, stb},
∀E ∈ sem(F), if an argument α ∈ E, then sub(α) ⊆ E.
Proof. For every β ∈ sub(α), since α is acceptable with respect to E, it holds that β
is acceptable with respect to E. This is because for each γ ∈ R, if γ defeats β, then
according to Proposition 2, γ defeats α; since α ∈ E, there exists an η ∈ E such that η
defeats γ. Given that β is acceptable with respect to E and E is a complete extension,
it holds that β ∈ E.
Since all norms in a hierarchical abstract normative system are defeasible, we only
need to discuss direct consistency and contextual consistency.
Proposition 4 (Direct consistency). Let F = (A, R) be an AF constructed from
a hierarchical abstract normative system. For all sem ∈ {cmp, prf, grd, stb}, ∀E ∈
sem(F), {concl(α) | α ∈ E, α is an ordinary argument} is consistent.
Proof. Assume that there exist α, β ∈ E such that concl(α) = concl(β). Since both α
and β are ordinary arguments, α attacks β, and β attacks α. If α β then α defeats β.
Otherwise, β defeats α. In both cases, E is not conflict-free, contradicting the fact that
E is a complete extension.
Proposition 5 (Contextual consistency). Let F = (A, R) be an AF constructed from
a hierarchical abstract normative system H = hL, N, C, ri. For all sem ∈ {cmp, prf, grd,
stb}, ∀E ∈ sem(F), C ∪ {concl(α) | α ∈ E, α is an ordinary argument} is consistent.
Proof. Since C is consistent, we only need to verify that for all a ∈ C, for all α ∈ E,
{a, concl(α)} is consistent. We use proof by contradiction. Assume the contrary, i.e.,
concl(α) = a. It follows that a defeats α, and therefore α ∈
/ E. Contradiction. So
the assumption is false, i.e., C ∪ {concl(α) | α ∈ E, α is an ordinary argument} is
consistent. This completes the proof.
In the next sections, we present representation results for the Greedy, Reduction
and Optimization approaches introduced in Section 2, identifying equivalences between
these approaches and the argument semantics based descriptions of a hierarchical abstract normative system.
4
Representation results for Greedy
Based on the idea introduced in Section 1, for a totally ordered hierarchical abstract
normative system, we have the following lemma and proposition.
Lemma 1 (Unique extension of Greedy). Given a totally ordered hierarchical abstract normative system H = hL, N, C, ri and the corresponding argumentation framework F = (Arg(H), Def(H, w )), it holds that F is acyclic, and therefore has a unique
extension under stable semantics.
Prioritized Norms in Formal Argumentation
15
Proof. Since H is totally ordered, under w , the relation w among arguments is
acyclic. Assume the contrary. Then, there exist three distinct α, β, γ ∈ Arg(H) such that
α w β, β w γ and γ w α. According to Definition 8, when H is totally ordered,
it holds that α w β, β w γ and γ w α. According to Proposition 1, α w γ,
contradicting γ w α. Hence, F is acyclic, and therefore has a unique extension under
stable semantics.
Proposition 6 (Greedy is weakest link). Given a totally ordered hierarchical abstract
normative system H = hL, N, C, ri and the corresponding argumentation framework
F = (Arg(H), Def(H, w )), it holds that {Greedy(H)} = Outfamily(F, stb).
Proof. Our proof is constructive: given the Greedy extension, we show how to construct
a stable extension of the argumentation framework whose conclusions coincide with the
Greedy extension. Since the argumentation framework has only one extension (Lemma
4.2), this completes the proof.
By Definition 2.4., Greedy(H) is the set of elements x such that there is a path in H
from an element in C to x with respect to a set of norms R ⊆ N , inductively defined by
R = ∪∞
i=0 Ri , where R0 = ∅ and Ri+1 = Ri ∪ max(N, C, Ri , r), max(N, C, Ri , r) =
{u ∈ Appl(N, C, Ri ) | ∀v ∈ Appl(N, C, Ri ) : r(u) ≥ r(v)} and Appl(N, C, R) :=
{(a, x) ∈ N \ R | a ∈ C∪ cons(R), {x, x}6⊆ C ∪ cons(R)}.
Let E = {a ∈ Arg (H) | a ∈ C is a context argument} ∪ {[u1 , . . . , un ] ∈ Arg(H) |
n ≥ 1, ant(u1 ) ∈ C, {ant(u2 ) . . . , ant(un ), cons(un )} ⊆ Greedy(H)}. From the consistency of the extensions (Proposition 3.11) and the construction of Greedy we know
that there is a consistent path to any element of Greedy(H), and thus for any argument
of Greedy(H) there is an ordinary argument with that element as its conclusion, and
thus Greedy(H) = {concl(α) | α ∈ E, α is an ordinary argument}. Now we only need
to prove that E is a stable extension under weakest link.
We use proof by contradiction. Assume that E is not a stable extension. Given that
Greedy(H) is consistent, and thus the extension E is conflict free, and given that context
arguments defeat all conflicting ordinary arguments and are thus included in E, there
must be α = [u1 , . . . , un ] ∈ Arg(H) \ E, such that α is not defeated by any argument in
E. Since α ∈
/ E, there exist α0 = [u1 , . . . , uj−1 ] ∈ E and α00 = [u1 , . . . , uj ] 6∈ E. Let
S ⊆ E be the set of arguments conflicting with α00 , thus each of them has a conclusion
cons(uj ). So, α00 defeats each argument β = [v1 , . . . , vm ] in S using weakest link, i.e.
α00 w β. If Φ1 = {u1 , . . . , un } and Φ2 = {v1 , . . . , vm }, then according to Def 3.2,
∃v ∈ Φ2 \ Φ1 such that ∀u ∈ Φ1 \ Φ2 , v ≤ u. Let M ⊆ N be the set of all these norms
v, i.e. all norms that occur in the arguments of S which are smaller than all the norms
in α00 , excluding norms that occur in α00 itself.
We now consider the point in the construction of the greedy extension where an element of M was added to it, we consider Appl(N, C, R) and max(N, C, Ri , r), and
we derive the contradiction. Let i be the lowest index such that Ri does not contain a norm of M . At this moment, Appl(N, C, Ri ) also contains a norm in α00 , and
max(N, C, Ri , r) contains an element of M . However, by definition, all norms of M
are ranked lower than the norms in α00 . Contradiction. So the assumption is false, i.e.
E is a stable extension. This completes the proof.
16
B. Liao N. Oren, L. van der Torre and S. Villata
Note that Proposition 6 corresponds to Theorem 5.3 of Young et al.[32]. This correspondence arises as follows. First, in the argumentation theory for a hierarchical abstract
normative system, we use disjoint elitist order to compare sets of norms, while in the
argumentation theory for prioritized default logic, Young et al. use a new order called a
structure-preference order, which takes into account the structure of how arguments are
constructed. Since in the setting of hierarchical abstract normative systems, arguments
are acyclic paths, it is not necessary to use the structure-preference order to compare arguments. Second, due to the different ways of constructing argumentation frameworks,
the proof of Proposition 6 differs from that of Theorem 5.3 of Young et al. [32]. The
former considers the order of the applicability of norms in the proof, while the latter
uses the mechanism defined in the structure-preference order.
5
Representation result for Reduction
According to Brewka and Eiter [7], Reduction is based on the following two points.
1) The application of a rule with nonmonotonic assumptions means jumping to a conclusion. This conclusion is yet another assumption which has to be used globally in
the program for the issue of deciding whether a rule is applicable or not.
2) The rules must be applied in an order compatible with the priority information.
This global view of deciding whether a rule is applicable coincides with the lastlink principle of lifting a preference relation between rules to a priority relation between
resulting arguments. According to Definition 4 and the argumentation theory for a hierarchical abstract normative system, we have the following representation result.
Proposition 7 (Reduction is last link). Given a totally ordered hierarchical abstract
normative system H = hL, N, C, ri and the corresponding argumentation framework
F = (Arg(H), Def(H, l )), it holds that Reduction(H) = Outfamily(F, stb).
Proof. We prove Reduction(H) = Outfamily(F, stb). by first proving that the left hand
side is a subset of the right hand side, and then by proving that the right hand side is a
subset of the left hand side.
We first prove Reduction(H) ⊆ Outfamily(F, stb).
Assume a set U ∈ Reduction(H). According to Definition 4 we have Greedy(HU )
where HX =hL, N 0 , C, r0 i, N 0 = {(>, x2 ) | (x1 , x2 ) ∈ N, x1 ∈ C ∪ X} is the set of
ordinary norms, and r0 (>, x2 ) = max(r(x1 , x2 ) | (x1 , x2 ) ∈ N, x1 ∈ C ∪ X) for all
(x1 , x2 ) ∈ N are priorities over norms.
We now construct a stable extension of F such that its conclusions are exactly U .
Let E be the set of arguments that can be constructed from elements of U , just like
in the proof of Greedy. So E = {a ∈ Arg (H) | a ∈ C is a context argument} ∪
{[u1 , . . . , un ] ∈ Arg(H) | n ≥ 1, ant(u1 ) ∈ C, {ant(u2 ) . . . , ant(un ), cons(un )} ⊆
U }. From the consistency of the extensions (Proposition 3.11) and the construction of
Reduction we know that there is a consistent path to any element of U , and thus for any
argument of U there is an ordinary argument with that element as its conclusion, and
Prioritized Norms in Formal Argumentation
17
thus U = {concl(α) | α ∈ E, α is an ordinary argument}. Now we only need to prove
that E is a stable extension.
We use proof by contradiction, just like in the proof of Greedy. Assume that E is not
stable. Given that context arguments defeat all conflicting ordinary arguments, there
must be a β = [v1 , . . . , vn ] ∈ Arg(H) \ E, such that β is not defeated by any argument
in E. Since β ∈
/ E, there exists β 0 = [v1 , . . . , vj ] in E and β 00 = [v1 , . . . , vj+1 ]
not in E. Let S ⊆ E be the set of arguments attacking β 00 , thus each of them has a
conclusion cons(vj+1 ). Since β 00 is not defeated by any argument in E, β 00 defeats each
argument in S. Then, the last link of β 00 is higher than the last links of all arguments of
S. Consequently, the rank of vj+1 is higher than the rank of all norms with consequent
cons(vj+1 ). But this means that in N 0 , the rank of (>, cons(vj+1 )) is higher than the
rank of (>, cons(vj+1 )). Then by the construction of Reduction, i.e., Greedy(HU ), we
have that cons(vj+1 ) must be in U , a contradiction. Thus the assumption is false, i.e.,
E is stable, and this completes the proof that Reduction(H) ⊆ Outfamily(F, stb).
To complete our proof, we now show that Reduction(H) ⊇ Outfamily(F, stb). Let
E be a stable extension of F = (Arg(H), Def(H, l )). Thus E is conflict free, closed
under sub-arguments, and for all arguments not in E, there is an argument in E defeating it. In particular, for any β 0 = [v1 , . . . , vj ] in E and β 00 = [v1 , . . . , vj+1 ] not in E,
there is an argument in E defeating β 00 .
Moreover, let U = {concl(α) | α ∈ E, α is an ordinary argument}. From the
contextual consistency of stable extensions (Proposition 3.12), we know that U ∪ C is
consistent. Since defeat is based on last link, it means that for every norm applicable in
U whose conclusion c is not in U , there is a higher ranked norm applicable in U whose
conclusion is c. Consider HU =hL, N 0 , C, r0 i, N 0 = {(>, x2 ) | (x1 , x2 ) ∈ N, x1 ∈
C ∪ U } is the set of ordinary norms, and r0 (>, x2 ) = max(r(x1 , x2 ) | (x1 , x2 ) ∈
N, x1 ∈ C ∪ U ) for all (x1 , x2 ) ∈ N are priorities over norms. Assume there is an
x ∈ U with (>, x), (>, x) ∈ N 0 . Due to the above the rank of (>, x) is higher than the
rank of (>, x), and thus x ∈ Greedy(HU ).
Thus U = Greedy(HU ), i.e. U ∈ Reduction(H), and that completes our proof.
Now, let us consider the following three examples, which show that by Reduction, a
hierarchical abstract normative system and the corresponding argumentation framework
may have a unique extension, multiple extensions, or an empty extension.
Example 8 (Order puzzle, Reduction). Consider Example 7 when the last link principle
is applied, A2 defeats A3 . Then, we have Outfamily(F1 , stb) = {{h, o}}, which is
equal to Reduction(H).
Example 9 (Reduction, Multiple extensions). Consider the hierarchical abstract normative system of Figure 3 when the last link principle is applied. We obtain the argumentation framework shown in Figure 4, written as F3 , yielding Outfamily(F3 , stb) =
{{b, c}, {¬b}}. Note that here, we have two distinct stable extensions.
Since stable extensions do not necessarily exist for all argumentation frameworks,
the Reduction of a hierarchical abstract normative system might not exist.
18
B. Liao N. Oren, L. van der Torre and S. Villata
A0
A1
[a]
[(a, b)]
[(a, b), (b, c), (c, b)]
A4
A3
A2
[(a, ¬b)]
[(a, b), (b, c)]
Fig. 4. The argumentation framework obtained for the hierarchical abstract normative system of
Figure 3.
[b]
[(b, c), (c, d)]
[(p, ¬d)]
A0
A6
A5
[(b, c)]
A1
A3
[r]
A2
[p]
[(r, z)]
A7
[(r, z), (z, ¬c)]
A8
A4
[(p, ¬d), [(¬d, ¬z)]]
Fig. 5. An argumentation framework with no stable extension.
Example 10 (Reduction, Empty extension). Consider the hierarchical abstract normative system H = hL, N, C, ri where N = {(c, d), (p, ¬d), (z, ¬c), (¬d, ¬z), (r, z), (b, c)},
C = {b, r, p} and r(c, d) = 5, r(p, ¬d) = 4, r(z, ¬c) = 6, r(¬d, ¬z) = 2, r(r, z) = 1
and r(b, c) = 0. When the last link principle is applied, this hierarchical abstract normative system yields the argumentation framework shown in Figure 5, which has no
stable extension.
6
Representation result for Optimization
In principle, Optimization is realized by adding norms in order of priority which are
consistent with the context, until no more norms can be added, obtaining a maximal set
of obeyable norms, following which conclusions can be computed. For each norm that
does not belong to the maximal set of obeyable norms, if it is in a consistent path of a
hierarchical abstract normative system H, it is the weakest norm of this path. In terms
of the terminology of argumentation, a consistent path of H is an argument of the corresponding argumentation framework. Hence, for an ordinary argument, if its top norm
Prioritized Norms in Formal Argumentation
19
is the weakest norm that does not belong to the maximal obeyable set, then this argument should not be accepted. In other words, this argument should be defeated by some
accepted argument. Furthermore, when an argument is rejected, all its super-arguments
should be rejected. For instance, as illustrated in Figure 1, the maximal obeyable set of
norms is {(h, o), (w, ¬o)}. The norm (w, h) does no belong to this set. So, the arguments containing (w, h), i.e., [(w, h)] and [(w, h), (h, o)], should be rejected.
In this paper, given an argument α, a sub-argument of α is called the weakest subargument of α, if it is an ordinary argument and its top norm is the weakest norm of
α.
Definition 12 (Weakest sub-argument). Given a totally ordered hierarchical abstract
normative system H = hL, N, C, ri and the corresponding argumentation framework
F = (Arg(H), Def(H, w )), for all α = [u1 , . . . , un ] ∈ Arg(H), if ui is the weakest
norm of α, then the ordinary sub-argument of α whose top norm is ui is called the
weakest sub-argument of α.
Then, the set of super-arguments of α0 —the weakest sub-argument of α— is called
the weakest arguments with respect to α, formally defined as follows.
Definition 13 (Weakest argument). Given a totally ordered hierarchical abstract normative system H = hL, N, C, ri and the corresponding argumentation framework
F = (Arg(H), Def(H, w )), for all α = [u1 , . . . , un ] ∈ Arg(H), let α0 be the weakest
sub-argument of α. Then, the set of weakest arguments with respect to α is defined as
warg(α) = sup(α0 ).
The notion of weakest arguments can be illustrated by the following example.
Example 11 (Order puzzle, weakest arguments). Consider the argumentation framework in Figure 2. When applying the weakest link principle, A3 defeats A2 . According
to Definition 13, warg(A1 ) = {A1 }, warg(A2 ) = {A1 , A2 }, and warg(A3 ) = {A3 }.
Based on the concept of weakest arguments, when an argument α defeats another argument β according to the weakest link principle, there are two different cases, namely
that α is in the set of weakest arguments with respect to β, or that it is not.
First, consider the case where α is in warg(β). In this case, it means that the weakest
norm of β is also the weakest norm of α.
Proposition 8 (Shared weakest norm). Given F = (Arg(H), Def(H, w )), for all
(α, β) ∈ Def(H, w ), if u is the weakest norm of β and α ∈ warg(β), then u is the
weakest norm of α.
Proof. We use proof by contradiction. Assume that u is not the weakest norm of α.
Then, there exists u0 such that u > u0 . Since u0 does not belong to β, it holds that
β w α. This contradicts (α, β) ∈ Def(H, w ). Thus the assumption is false, i.e., u is
the weakest norm of α.
Given that α and β share a weakest norm, according to the definition of Optimization, this weakest norm does not belong to the maximal obeyable set. Hence, both α
20
B. Liao N. Oren, L. van der Torre and S. Villata
and β cannot be accepted. In this paper, we use an auxiliary argument aux to defeat all
arguments in warg(β) that should not be accepted.
The second case occurs when α is not in warg(β). Here, the weakest norm of α is
superior to the weakest norm of β. So, if α is accepted, then all arguments in warg(β)
should be rejected. For this purpose, we may use α to defeat each argument in warg(β).
In addition, since the defeat from α to each super-argument of β is already in Arg(H),
only the arguments in warg(β) ∩ psub(β) should be added to Arg(H).
According to the above observations, an expanded argumentation framework of F
with auxiliary defeats on weakest arguments is defined as follows.
Definition 14 (Expanded argumentation framework with additional defeats on weakest arguments). Let F = (Arg(H),Def(H, w )) be an argumentation framework that
is constructed from a totally ordered hierarchical abstract normative system H = hL,
N, C, ri, and aux be an auxiliary argument such that aux ∈
/ Arg(H). The expanded
argumentation framework of F with auxiliary defeats on weakest arguments is F 0 =
(Arg(H)∪{aux}, Def(H, w )∪Φ1 ∪Φ2 ) where Φ1 = ∪(α,β)∈Def(H,w )∧α∈warg(β)
{(α, γ) |
/
γ ∈ warg(β) ∩ psub(β)}, and Φ2 = ∪(α,β)∈Def(H,w )∧α∈warg(β) {(aux, γ) | γ ∈
warg(β)}.
In Definition 14, Φ1 is the set of defeats from α that is not a weakest argument with
respect to β, and Φ2 is the set of defeats from the auxiliary argument aux when α is a
weakest argument with respect to β.
The following lemma and proposition show that Optimization can be represented in
formal argumentation by using weakest link together with auxiliary defeats.
Lemma 2 (Unique extension of Optimization). Let H = hL, N, C, ri be a totally ordered hierarchical abstract normative system, and F 0 be an argumentation framework
of H with additional defeats on weakest arguments presented in Definition 13. It holds
that F 0 is acyclic, and therefore has a unique extension under stable semantics.
Proof. According to Lemma 1, F is acyclic. The addition of Φ2 does not produce cycles.
Meanwhile, for all (α, γ) ∈ Φ1 , since α defeats β and γ is a weakest argument with
respect to β, α w γ. So, the addition of Φ2 does not produce cycles either. As a result,
F 0 is acyclic, and therefore has a unique extension under stable semantics.
Proposition 9 (Optimization is weakest link plus auxiliary defeats). Let H = hL,
N, C, ri be a totally ordered hierarchical abstract normative system, and F 0 be an
argumentation framework of H with additional defeats on weakest arguments presented in Definition 13. It holds that Optimization(H) = {concl(α) | α ∈ E \
{aux}, α is an ordinary argument} where E is the unique stable extension of F 0 .
Proof. This proof is similar to that of Greedy: given the Optimization extension, we
show how to construct a stable extension of the expanded argumentation framework
whose conclusions coincide with the Optimization extension. Since the argumentation
framework has only one extension (Lemma 2), this completes the proof.
By Definition 5, Optimization(H) is the set of elements x such that there is a path
in H from an element in C to x with respect to a set of norms R ⊆ N . Given T =
Prioritized Norms in Formal Argumentation
21
(u1 , u2 , . . . , un ) the linear order on N such that u1 > u2 > · · · > un , R is inductively
defined by R = Rn , where R0 = ∅; Ri+1 = Ri ∪ {ui } if C ∪ R(C) is consistent where
R = Ri ∪ {ui }, and Ri+1 = Ri otherwise.
Let E = {aux} ∪ {α ∈ Arg (H) | α ∈ C is a context argument} ∪ {[u1 , . . . , un ] ∈
Arg(H) | n ≥ 1, ant(u1 ) ∈ C, {ant(u2 ) . . . , ant(un ), cons(un )} ⊆ Optimization(H)}.
From the consistency of the extensions (Proposition 3.11) and the construction of Optimization we know that there is a consistent path to any element of Optimization(H),
and thus for any argument of Optimization(H) there is an ordinary argument with
that element as its conclusion, and thus Optimization(H) = {concl(α) | α ∈ E \
{aux}, α is an ordinary argument}.
Now we only need to prove that E is a stable extension of F 0 .
We use proof by contradiction. So assume that E is not a stable extension. On the
one hand, given that Optimization(H) is consistent, and thus the extension E is conflict
free, and given that context arguments defeat all conflicting ordinary arguments and are
thus included in E, there must be α = [u1 , . . . , un ] ∈ Arg(H) \ E, such that α is not
defeated by any argument in E. On the other hand, since α ∈
/ E, there exists a weakest
norm uj of α such that uj ∈
/ R. Let i be the index where uj can not be added to Ri
since C ∪ R(C) is consistent where R = Rj+1 = Ri ∪ {uj }. Let α0 = [u1 , . . . , uk ]
where k ≥ j, and β = [v1 , . . . , vm ] such that {v1 , . . . , vm } ⊆ Ri+1 and vm = uk .
Since uj < v for all v ∈ Ri , it holds that β defeats α0 . Then, we have the following two
possible cases. If uj ∈ {v1 , . . . , vm }, then α is defeated by aux ∈ E. Contradiction.
Otherwise, uj ∈
/ {v1 , . . . , vm }. In this case, {v1 , . . . , vm } ⊆ Ri and C ∪ R(C) is
consistent where R = Ri . So, β is in E. Contradiction. So the assumption is false, i.e.,
E is a stable extension. This completes the proof.
Now, let us consider a revised version of the Order puzzle as follows.
Example 12 (Order puzzle, Optimization). Let H0 = hL, N, C, ri be a hierarchical abstract normative system, where L = {w, h, o, ¬w, ¬h, ¬o, >}, N = {(w, h), (w, ¬h),
(h, o), (¬h, o), (w, ¬o)}, C = {w, >}, r(w, h) = 1, r(w, ¬h) = 0, r(h, o) = 3,
r(¬h, o) = 4, r(w, ¬o) = 2. The maximal set of obeyable norms is R = {(¬h, o), (h, o),
(w, ¬o)} and so Optimization(H0 ) = {¬o}. Figure 6 illustrates the argumentation
framework obtained from this hierarchical abstract normative system, by adding an
auxiliary argument and two auxiliary defeats. In this example, since A5 is neither in
warg(A3 ) = {A1 , A3 } nor in warg(A4 ) = {A2 , A4 }, the two auxiliary defeats are
from A5 to A1 and A2 respectively. Then, under stable semantics, the expanded argumentation framework has a unique extension E = {aux, A0 , A5 }. As a result, the set of
conclusions {concl(α) | α ∈ E\{aux}, α is an ordinary argument} = {concl(A5 )} =
{¬o}.
Finally, let us consider an expanded argumentation framework where auxiliary defeats are from the auxiliary argument aux.
Example 13 (Empty Optimization). Let H = hL, N , C, ri be a hierarchical abstract
normative system, where L = {a, b, c, ¬a, ¬b, ¬c, >}, N = {(a, b), (b, c), (b, ¬c)},
C = {a, >}, r(a, b) = 1, r(b, c) = 2, r(b, ¬c) = 3. The maximal set of obeyable
norms is R = {(b, c), (b, ¬c)} and so Optimization(H) = ∅. Figure 7 illustrates
22
B. Liao N. Oren, L. van der Torre and S. Villata
[(w, h)]
[(w, ¬h)]
[w]
A1
A2
A0
aux
[(w, h), (h, o)]
A5
A4
A3
[(w, ¬h), (¬h, o)] [(w, ¬o)]
Fig. 6. Argumentation framework expanded by an auxiliary argument and two auxiliary defeats
(denoted using dashed lines).
the argumentation framework obtained from this hierarchical abstract normative system, by adding an auxiliary argument and three auxiliary defeats. In this example,
since A3 is in warg(A2 ) = {A1 , A2 , A3 }, the three auxiliary defeats are from aux.
Then, under stable semantics, the expanded argumentation framework has a unique extension E = {aux, A0 }. As a result, the set of conclusions {concl(α) | α ∈ E \
{aux}, α is an ordinary argument} = ∅.
[a]
[(a, b)]
[(a, b), (b, c)]
A0
A1
A2
aux
A3
[(a, b), (b, ¬c)]
Fig. 7. Argumentation framework expanded by an auxiliary argument and three auxiliary defeats
(denoted using dashed lines).
Prioritized Norms in Formal Argumentation
7
23
Discussions and related work
In the previous sections, after defining three detachment procedures for totally ordered
hierarchical abstract normative systems motivated by the Order puzzle example, we
formulated an argumentation theory to represent them. The role of examples in the study
of logic has a long and rich history. Traditionally, a logic was proposed to model some
example problem, following which examples were introduced to highlight paradoxes
or inconsistencies in the logic, whereupon a new logic was proposed to address these
problems, and the cycle repeated. While this approach has significantly enriched the
field, it is not without problems. For example, there is still a debate regarding deontic
detachment within the community, as in some cases, deontic detachment intuitively
holds, and in other cases it does not [25]. Given this, we do not seek to claim that
the detachment procedures we present in this paper are in any sense the ‘right’ logics.
Instead, our goal is to answer the following questions.
1. What are the general properties of systems considered relevant to some problem?
2. Given an application, what choices should be made in order to obtain a solution?
In this paper, the systems we considered are the three different detachment procedures, encoded in the general framework of hierarchical abstract normative systems.
The property we considered is then the conclusions that one can draw from each of the
detachment procedures in the context of prioritized norms, which we describe in the
context of an argumentation system.
Our results then characterize the outputs of Greedy, Reduction and Optimization
in terms of argumentation for a totally ordered hierarchical abstract normative system,
allowing one to decide which approach is relevant to their needs by understanding the
effects of each approach through the argumentation literature. The semantics associated
with each approach also sheds light on the complexity of computing conclusions in the
normative context.
Furthermore, it is important to note that our representation results only hold when
the hierarchical abstract normative system is totally ordered. This is illustrated by the
following example, which shows that Greedy does not match the results of weakest link
when one does not have a total order over preferences.
Let H = hL, N, C, ri be a hierarchical abstract normative system, where N =
{(a, b), (a, c), (b, ¬c), (c, ¬b)} C = {a} and r(a, b) = 1, r(a, c) = 1, r(b, ¬c) = 2
and r(c, ¬b) = 2}. On the one hand, by Greedy, there are two extensions {b, ¬c} and
{c, ¬b}. On the other hand, by the weakest link principle, the argumentation framework
constructed from H is illustrated in Figure 8. Under stable semantics, there are three
extensions {A0 , A1 , A2 }, {A0 , A1 , A3 }, and {A0 , A3 , A4 }. So, there are three conclusion extensions {b, ¬c}, {b, c}, {c, ¬b}. As a result, the set of conclusions obtained by
Greedy is not equal to the one obtained by argumentation using the weakest link principle. We leave identifying valid representation results for hierarchical abstract normative
systems containing preference preorders as an avenue for future research.
Regarding related work, Young et al. [32] endowed Brewka’s prioritized default
logic (PDL) with argumentation semantics using the ASPIC+ framework for structured
argumentation [22]. More precisely, their goal is to define a preference ordering over
24
B. Liao N. Oren, L. van der Torre and S. Villata
[(a, b), (b, ¬c)]
[(a, c)]
A2
A3
A1
A4
[(a, b)]
[(a, c), (c, ¬b)]
A0
[a]
Fig. 8. Argumentation framework constructed from a preordered hierarchical abstract normative
system.
arguments %, based on the strict total order over defeasible rules defined to instantiate
ASP IC + to PDL, so as to ensure that an extension within PDL corresponds to the justified conclusions of its ASP IC + instantiation. Several options are investigated, and
they demonstrate that the standard ASP IC + elitist ordering cannot be used to calculate % as there is no correspondence between the argumentation-defined inferences and
PDL, and the same holds for a disjoint elitist preference ordering. The authors come
up with a new argument preference ordering definition which captures both preferences
over arguments and also when defeasible rules become applicable in the arguments’
construction, leading to the definition of a strict total order on defeasible rules and
corresponding non-strict arguments. Their representation theorem shows that a correspondence always exists between the inferences made in PDL and the conclusions of
justified arguments in the ASP IC + instantiation under stable semantics.
Brewka and Eiter [7] consider programs supplied with priority information, which
is given by a supplementary strict partial ordering of the rules. This additional information is used to solve potential conflicts. Moreover, their idea is that conclusions should
be only those literals that are contained in at least one answer set. They propose to
use preferences on rules for selecting a subset of the answer sets, called the preferred
answer sets. In their approach, a rule is applied unless it is defeated via its assumptions by rules of higher priorities. Our definition (Def. 4) and the original formalism of
Brewka and Eiter [7] are different, in the sense that in our definition we do not make
use of default negation to represent the exceptions, i.e., the defeasibility, of a (strict)
rule. Rather, we use defeasible rules and the notion of the applicability of such rules.
This means that the correct translation of the Order Puzzle of Example 1 ends up with
the following logic program5 :
5
Note that in Nrewka and Eiter’s original formulation [7] r0 < r3 means that r0 has higher
priority than r3 .
Prioritized Norms in Formal Argumentation
r0
r1
r2
r3
r0
25
: w.
: h : - not ¬h, w.
: ¬o : - not o, w.
: o : - not ¬o, h.
< r3 < r2 < r1
If preferences are disregarded, then this logic program has two answer sets: {w, h, ¬o}
and {w, h, o}. Thus, considering preferences, the latter is the unique preferred answer
set. After dropping the context w from the answer set, we get an extension {h, o}, which
is identical to the result obtained in Example 3.
Dung [11] presents an approach to deal with contradictory conclusions in defeasible
reasoning with priorities. More precisely, he starts from the observation that often, the
proposed approaches to defeasible reasoning with priorities (e.g., [5,27,21]) sanction
contradictory conclusions, as exemplified by ASP IC + using the weakest link principle
together with the elitist ordering which returns contradictory conclusions with respect
to its other three attack relations, and the conclusions reached with the well known
approach of Brewka and Eiter [7]. Dung shows then that the semantics for any complex
interpretation of default preferences can be characterized by a subset of the set of stable
extensions with respect to the normal attack relation assignments, i.e., a normal form for
ordinary attack relation assignments. In the setting of this paper, the notion of ‘normal
attack relation’ could be defined as follows. Let α = (a1 , . . . , an ) and β = (b1 , . . . , bm )
be arguments constructed from a hierarchical abstract normative system. Since we have
no Pollock style undercutting argument (as in ASPIC+ ) and each norm is assumed to be
defeasible, α is said to normally attack argument β if and only if β has a sub-argument
β 0 such that concl(α) = concl(β 0 ), and r((an−1 , an )) ≥ r((bm−1 , bm )). According to
Definitions 8 and 9, the normal defeat relation is equivalent to the defeat relation using
the last link principle in this paper.
Kakas et al. [18] present a logic of arguments called argumentation logic, where
the foundations of classical logical reasoning are represented from an argumentation
perspective. More precisely, their goal is to integrate into the single argumentative representation framework both classical reasoning, as in propositional logic, and defeasible
reasoning.
You et al. [31] define a prioritized argumentative characterization of non-monotonic
reasoning, by casting default reasoning as a form of prioritized argumentation. They
illustrate how the parameterized formulation of priority may be used to allow various
extensions and modifications to default reasoning.
We, and all these approaches, share the idea that an argumentative characterization
of NMR formalisms, like prioritized default logic in Young’s case and hierarchical abstract normative systems in our approach, contributes to make the inference process
more transparent to humans. However, the targeted NMR formalism is different, leading to different challenges in the representation results. To the best of our knowledge,
no other approach addressed the challenge of an argumentative characterization of prioritized normative reasoning.
The reason we study prioritized normative reasoning in the setting of formal argumentation is twofold. First, formal argumentation has been recognized as a popular
26
B. Liao N. Oren, L. van der Torre and S. Villata
research area in AI, thanks to its ability to make the inference process more intuitive
and provide natural explanations for the reasoning process [8]; its flexibility in dealing
with the dynamics of the system; and its appeal in sometimes being more computationally efficient than competing approaches. Second, while some progress has been made
on the use of priorities within argumentation (e.g., [2,21]), how to represent different
approaches for prioritized normative reasoning in argumentation is still a challenging
issue.
8
Conclusions
In this paper we embedded three approaches to prioritized normative reasoning—namely
the Greedy [32], Reduction [5] and Optimization [13] approaches—within the framework of a hierarchical abstract normative system. Within such a system, conditional
norms are represented by a binary relation over literals, and priorities are represented
by natural numbers. Hierarchical abstract normative systems provide an elegant visualisation of a normative system, with conflicts shown as two paths to a proposition and its
negation. Since both conflicts and exceptions can be encoded, such systems are inherently non-monotonic. In Dung [10], the author pointed out that “many of the major approaches to nonmonotonic reasoning in AI and logic programming are different forms
of argumentation”, and inspired by this, we described how arguments can be instantiated as paths through a hierarchical abstract normative system; demonstrated that this
instantiation satisfies certain desirable properties; and described how attacks and defeats between these arguments can be identified. Defeats in particular are dependent on
the priorities associated with the arguments, and several different techniques have been
proposed to lift priorities from argument components — made up of norms in the context of our work — to the arguments themselves [21]. We demonstrated that for a total
ordering of priorities, lifting priorities to arguments based on the weakest link principle,
evaluated using the stable semantics, is equivalent to Greedy; that lifting priorities to
arguments based on last link and using the stable semantics is equivalent to Reduction;
and that the Optimization approach can be encoded by an argumentation system which
uses weakest link together with the stable semantics, and which introduces additional
defeats capturing implicit conflicts between arguments.
This last result—which requires a relatively complex argumentative representation—
opens up an interesting avenue for future work, namely in determining which nonmonotonic logics can be easily captured through standard formal argumentation techniques, and which require additional rules or axioms in order to be represented. We note
that on the argumentation side, work on bipolar argumentation (e.g., [9]) has considered
introducing additional defeats between arguments based on some notion of support, and
we intend to investigate how the additional defeats we introduced can be categorized in
such frameworks.
Apart from our representation results, the use of argumentation allows us to make
some useful observations, such as that Reduction will sometimes not reach any conclusions. Furthermore, argumentation can be used to provide explanation [8]. When implemented, a system building on our approach can help users understand what norms they
should comply with, and why. For large normative systems, the use of stable semantics
Prioritized Norms in Formal Argumentation
27
to compute Reduction and Optimization results in a high computational overhead, while
Greedy is computationally efficient. Ultimately, selecting the correct reasoning procedure thus requires giving consideration to both reasoning complexity, and the domain
in which the system will be used.
In closing, our main observations can be summarized as follows. First, from a
normative systems perspective, we know there are many many logics of prioritized
rules/norms, and we consider only three here. The choice we make (Greedy, Reduction
and Optimization) may seem arbitrary. However, many other examples, in particular
detachment procedures not satisfying defeasible deontic detachment, are much easer
to characterize, while the three throughput variants of Greedy, Reduction and Optimization can be derived from the existing results. Furthermore, these three alternatives
display quite diverse behavior, and are illustrative of the various kind of approaches
around.
Second, the results we present are interesting and promising, but the work on representing prioritized rules/norms using argumentation has only begun, and there are many
open issues. In particular, the restriction to totally ordered systems must be relaxed in
future work.
Third, given the large number of possibilities and the vast existing literature on
normative rules/norms, a different methodology is needed for dealing with prioritized
rules/norms in formal argumentation.
Finally, one may wonder why our results have not been shown before, given the
long standing discussion on weakest vs last link at least since the work of Pollock [24],
and the central role of prioritized rules in many structured argumentation theories like
ASPIC+ . The reason, we believe, is that it is easier to study these issues on a small
fragment, like hierarchical abstract normative systems, than on a very general theory
like ASPIC+ .
Acknowledgements
The authors are grateful to the anonymous reviewers for their helpful comments. The
research reported in this paper was partially supported by the Fundamental Research
Funds for the Central Universities of China for the project Big Data, Reasoning and
Decision Making, the National Research Fund Luxembourg (FNR) under grant INTER/MOBILITY/14/8813732 for the project FMUAT: Formal Models for Uncertain
Argumentation from Text, and the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 690974 for
the project MIREL: MIning and REasoning with Legal texts.
References
1. Alchourron, C. E. and D. Makinson, Hierarchies of regulations and their logic, in: R.
Hilpinen, editor, New studies in deontic logic, 1981 pp. 125–148.
2. Amgoud, L. and S. Vesic, Handling inconsistency with preference-based argumentation, in:
A. Deshpande and A. Hunter, editors, Scalable Uncertainty Management 4th International
Conference, SUM 2010, Toulouse, France, September 27-29, 2010. Proceedings, Lecture
28
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
B. Liao N. Oren, L. van der Torre and S. Villata
Notes in Computer Science 6379 (2010), pp. 5669. URL http://dx.doi.org/10.1007/978-3642-15951-0 11
Boella, G. and L. W. N. van der Torre, Permissions and obligations in hierarchical normative
systems, in: Proceedings of the 9th International Conference on Artificial Intelligence and
Law, ICAIL 2003, Edinburgh, Scotland, UK, June 24-28, 2003, 2003, pp. 109–118. URL
http://doi.acm.org/10.1145/1047788.1047818
Boella, G., L. W. N. van der Torre and H. Verhagen, Ten challenges for normative multiagent systems, in: Programming Multi-Agent Systems, 31.08. - 05.09.2008, 2008. URL
http://drops.dagstuhl.de/opus/volltexte/2008/1636/
Brewka, G., Preferred subtheories: An extended logical framework for default reasoning, in:
N. S. Sridharan, editor, Proceedings of the 11th International Joint Conference on Artificial
Intelligence. Detroit, MI, USA, August 1989 (1989), pp. 1043–1048.
Brewka, G., Reasoning about priorities in default logic, in: B. Hayes-Roth and R.
E. Korf, editors, Proceedings of the 12th National Conference on Artificial Intelligence,
Seattle, WA, USA, July 31 - August 4, 1994, Volume 2. (1994), pp. 940–945. URL
http://www.aaai.org/Library/AAAI/1994/aaai94-144.php
Brewka, G. and T. Eiter, Preferred answer sets for extended logic programs, Artificial
Intelligence 109 (1999), pp. 297–356.
Caminada, M., R. Kutlak, N. Oren and W. W. Vasconcelos, Scrutable plan enactment via
argumentation and natural language generation, in: Proceedings of the 2014 International
Conference on Autonomous Agents and Multi-agent Systems, 2014, pp. 1625–1626.
Cayrol, C. and M.-C. Lagasquie-Schiex, An axiomatic approach to support in argumentation,
in: E. Black, S. Modgil and N. Oren, editors, Theory and Applications of Formal Argumentation: Third International Workshop (2015), pp. 74–91. URL https://doi.org/10.1007/978-3319-28460-6 5
Dung, P. M., On the acceptability of arguments and its fundamental role in nonmonotonic
reasoning, logic programming and n-person games, Artificial Intelligence 77 (1995), pp.
321–358. URL http://dx.doi.org/10.1016/0004-3702(94)00041-X
Dung, P. M., An axiomatic analysis of structured argumentation with priorities, Artificial
Intelligence 231 (2016), pp. 107–150. URL https://doi.org/10.1016/j.artint.2015.10.005
Hansen, J., Deontic logics for prioritized imperatives, Artificial Intelligence & Law 14
(2006), pp. 1–34.
Hansen, J.,
Prioritized conditional imperatives: problems and a new proposal,
Autonomous Agents and Multi-Agent Systems 17 (2008), pp. 11–35. URL
http://dx.doi.org/10.1007/s10458-007-9016-7
Hansen, J., Reasoning about permission and obligation, in: S. O. Hansson, editor, David
Makinson on Classical Methods for Non-Classical Problems, Springer Netherlands, 2014
pp. 287–333.
Hansson, S. O., The varieties of permission, in: D. Gabbay, J. Horty, X. Parent, R. van der
Meyden and L. van der Torre, editors, Handbook of Deontic Logic and Normative Systems,
College Publications, 2013 pp. 195–240.
Hilpinen, R. and P. Mcnamara, Deontic logic: A historical survey and introduction, in: D.
Gabbay, J. Horty, X. Parent, R. van der Meyden and L. van der Torre, editors, Handbook of
Deontic Logic and Normative Systems, College Publications, 2013 pp. 3–136.
Horty, J., Defaults with priorities, Journal of Philosophical Logic 36 (2007), pp. 367–413.
Kakas, A. C., F. Toni and P. Mancarella, Argumentation for propositional logic and nonmonotonic reasoning, in: L. Giordano, V. Gliozzi and G. L. Pozzato, editors, Proceedings
of the 29th Italian Conference on Computational Logic, Torino, Italy, June 16-18, 2014.,
CEUR Workshop Proceedings 1195 (2014), pp. 272–286. URL http://ceur-ws.org/Vol1195/long18.pdf
Prioritized Norms in Formal Argumentation
29
19. Liao, B., N. Oren, L. van der Torre and S. Villata, Prioritized norms and defaults in formal
argumentation, in: Proceedings of the 13th International Conference on Deontic Logic and
Normative Systems (DEON2016), 2016, pp. 139–154.
20. Makinson, D. and L. W. N. van der Torre, Input/output logics, J. Philosophical Logic 29
(2000), pp. 383–408. URL http://dx.doi.org/10.1023/A:1004748624537
21. Modgil, S. and H. Prakken, A general account of argumentation with preferences, Artificial
Intelligence 195 (2013), pp. 361–397. URL http://dx.doi.org/10.1016/j.artint.2012.10.008
22. Modgil, S. and H. Prakken,
The ASPIC+ framework for structured argumentation: a tutorial,
Argument & Computation 5 (2014), pp. 31–62. URL
http://dx.doi.org/10.1080/19462166.2013.869766
23. Parent, X. and L. van der Torre, Input/output logic, in: D. Gabbay, J. Horty, X. Parent, R.
van der Meyden and L. van der Torre, editors, Handbook of Deontic Logic and Normative
Systems, College Publications, 2013 pp. 499544.
24. Pollock, J. L., Cognitive Carpentry, Bradford/MIT Press, 1995.
25. Nute, D., editor, Defeasible deontic logic, 263, Springer Science & Business Media, 1997.
26. Prakken, H. and G. Sartor, Argument-based extended logic programming with defeasible priorities, Journal of Applied Non-Classical Logics 7 (1997), pp. 25–75. URL
http://dx.doi.org/10.1080/11663081.1997.10510900
27. Schaub, T. and K. Wang, A comparative study of logic programs with preference, in: B.
Nebel, editor, Proceedings of the Seventeenth International Joint Conference on Artificial
Intelligence, IJCAI 2001, Seattle, Washington, USA, August 4-10, 2001 (2001), pp. 597–
602. URL http://ijcai.org/proceedings/2001-1
28. Tosatto, S. C., G. Boella, L. W. N. van der Torre and S. Villata, Abstract normative
systems: Semantics and proof theory, in: G. Brewka, T. Eiter and S. A. McIlraith, editors, Principles of Knowledge Representation and Reasoning: Proceedings of the Thirteenth International Conference, KR 2012, Rome, Italy, June 10-14, 2012 (2012). URL
http://www.aaai.org/ocs/index.php/KR/KR12/paper/view/4553
29. Tosatto, S. C., G. Boella, L. W. N. van der Torre and S. Villata, Visualizing normative
systems: An abstract approach, in: Deontic Logic in Computer Science - 11th International
Conference, DEON 2012, Bergen, Norway, July 16-18, 2012. Proceedings, 2012, pp. 16–30.
URL https://doi.org/10.1007/978-3-642-31570-1 2
30. van der Torre, L. W. N. and S. Villata, An ASPIC-based legal argumentation framework for
deontic reasoning, in: S. Parsons, N. Oren, C. Reed and F. Cerutti, editors, Computational
Models of Argument -Proceedings of COMMA 2014, Atholl Palace Hotel, Scottish Highlands, UK, September 9-12, 2014, Frontiers in Artificial Intelligence and Applications 266
(2014), pp. 421–432. URL https://doi.org/10.3233/978-1-61499-436-7-421
31. You, J., X. Wang and L. Yuan, Nonmonotonic reasoning as prioritized argumentation, IEEE
Trans. Knowl. Data Eng. 13 (2001), pp. 968–979. URL http://dx.doi.org/10.1109/69.971190
32. Young, A. P., S. Modgil and O. Rodrigues, Prioritised default logic as rational argumentation,
in: Proceedings of the 15th International Conference on Autonomous Agents and Multiagent
Systems (AAMAS 2016), 2016, pp. 626–634.
| 2cs.AI
|
A Sub-Quadratic Exact Medoid Algorithm
arXiv:1605.06950v4 [stat.ML] 12 Apr 2017
James Newling
Idiap Research Institute & EPFL
Abstract
We present a new algorithm trimed for obtaining the medoid of a set, that is the element of the set which minimises the mean
distance to all other elements. The algorithm
is shown to have, under certain assumptions,
3
expected run time O(N 2 ) in Rd where N is
the set size, making it the first sub-quadratic
exact medoid algorithm for d > 1. Experiments show that it performs very well on spatial network data, frequently requiring two
orders of magnitude fewer distance calculations than state-of-the-art approximate algorithms. As an application, we show how
trimed can be used as a component in an
accelerated K-medoids algorithm, and then
how it can be relaxed to obtain further computational gains with only a minor loss in
cluster quality.
1
Introduction
A popular measure of the centrality of an element of
a set is its mean distance to all other elements. In
network analysis, this measure is referred to as closeness centrality, we will refer to it as energy. Given
a set S = {x(1), . . . , x(N )} the energy of element
i ∈ {1, . . . , N } is thus given by,
E(i) =
1
N
X
dist(x(i), x(j)).
j∈{1,...,N }
An element in S with minimum energy is referred to
as a 1-median or a medoid. Without loss of generality, we will assume that S contains a unique medoid.
The problem of determining the medoid of a set arises
in the contexts of clustering, operations research, and
Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS) 2017, Fort Lauderdale, Florida, USA. JMLR: W&CP volume 54. Copyright 2017 by the author(s).
François Fleuret
Idiap Research Institute & EPFL
network analysis. In clustering, the Voronoi iteration
K-medoids algorithm (Hastie et al., 2001; Park and
Jun, 2009) requires determining the medoid of each of
K clusters at each iteration. In operations research,
the facility location problem requires placing one or
several facilities so as to minimise the cost of connecting to clients. In network analysis, the medoid may
represent an influential person in a social network, or
the most central station in a rail network.
1.1
Medoid algorithms and our contribution
A simple algorithm for obtaining the medoid of a set
of N elements computes the energy of all elements and
selects the one with minimum energy, requiring Θ(N 2 )
time. In certain settings Θ(N ) algorithms exist, such
as in 1-D where the problem is solved by Quickselect (Hoare, 1961), and more generally on trees. However, no general purpose o(N 2 ) algorithm exists. An
example illustrating the impossibility of such an algorithm is presented in Supplementary Material B (SMA). Related to finding the medoid of a set is finding
the geometric median, which in vector spaces is defined as the point in the vector space with minimum
energy. The relationship between the two problems is
discussed in §2.1.
Much work has been done to develop approximate algorithms in the context of network analysis. The RAND
algorithm of Eppstein and Wang (2004) can be used to
estimate the energy of all nodes in a graph. The accuracy of RAND depends on the diameter of the network,
which motivated Cohen et al. (2014) to use pivoting
to make RAND more effective for large diameter networks. The work most closely related to ours is that
of Okamoto et al. (2008), where RAND is adapted to
the task of finding the k lowest energy nodes, k = 1
corresponding to the medoid problem. The resulting
TOPRANK algorithm of Okamoto et al. (2008) has run
time Õ(N 5/3 ) under certain assumptions, and returns
the medoid with probability 1 − O(1/N ), that is with
high probability (w.h.p.). Note that only their run time
result requires any assumption, obtaining the medoid
w.h.p. is guaranteed. TOPRANK is discussed in §2.2.
In this paper we present an algorithm which has ex-
A Sub-Quadratic Exact Medoid Algorithm
pected run time O(N 3/2 ) under certain assumptions
and always returns the medoid. In other words, we
present an exact medoid algorithm with improved
complexity over the state-of-the-art approximate algorithm, TOPRANK. We show through experiments that
the new algorithm works well for low-dimensional data
in Rd and for spatial network data. Our new medoid
algorithm, which we call trimed, uses the triangle inequality to quickly eliminate elements which cannot be
the medoid. The O(N 3/2 ) run time follows from the
surprising result that all but O(N 1/2 ) elements can be
eliminated in this way.
The complexity bound on expected run time which we
derive contains a term which grows exponentially in
dimension d, and experiments show that in very high
dimensions trimed often ends up computing O(N 2 )
distances.
1.2
K-medoids algorithms and our
contribution
It should be noted that algorithms other than KMEDS
have been proposed for finding approximate solutions
to the K-medoids problem, and have been shown to be
very effective in Newling and Fleuret (2016b). These
include PAM and CLARA of Kaufman and Rousseeuw
(1990), and CLARANS of Ng et al. (2005). In this paper we do not compare cluster qualities of previous
algorithms, but focus on accelerating the lloyd equivalent for K-medoids as a test setting for our medoid
algorithm trimed.
2
2.1
Previous works
A related problem: the geometric median
A problem closely related to the medoid problem is the
geometric median problem. In the vector space Rd the
geometric median, assuming it is unique, is defined as,
X
g(S) = arg min
kv − yk .
(1)
v∈V
The K-medoids problem is to partition a set into K
clusters, so as to minimise the sum over elements of
dissimilarites with their nearest medoids. That is, to
choose M = {m(1), . . . , m(K)} ⊂ {1, . . . , N } to minimise,
L(M) =
N
X
i=1
min
diss(x(i), x(m(k))).
k∈{1,...,K}
We focus on the special case where the dissimilarity
is a distance (diss = dist), which is still more general
than K-means which only applies to vector spaces. Kmedoids is used in bioinformatics where elements are
genetic sequences or gene expression levels (Chipman
et al., 2003) and has been applied to clustering on
graphs (Rattigan et al., 2007). In machine vision, Kmedoids is often preferred, as a medoid is more easily
interpretable than a mean (Frahm et al., 2010).
The K-medoids problem is NP-hard, but there exist
approximation algorithms. The Voronoi iteration algorithm, appearing in Hastie et al. (2001) and later
in Park and Jun (2009), consists of alternating between updating medoids and assignments, much in the
same way as Lloyd’s algorithm works for the K-means
problem. We will refer to it as KMEDS, and to Lloyd’s
K-means algorithm as lloyd.
One significant difference between KMEDS and lloyd is
that the computation of a medoid is quadratic in the
number of elements per cluster whereas the computation of a mean is linear. By incorporating our new
medoid algorithm into KMEDS, we break the quadratic
dependency of KMEDS, bringing it closer in performance
to lloyd. We also show how ideas for accelerating
lloyd presented in Elkan (2003) can be used in KMEDS.
y∈S
While the medoid of a set is defined in any space with a
distance measure, the geometric median is specific to
vector spaces, where addition and scalar multiplication are defined. The convexity of the objective function being minimised in (1) has enabled the development of fast algorithms. In particular, Cohen et al.
(2016) present an algorithm which obtains an estimate
for the geometric median with relative error 1 + O()
with complexity O(nd log3 ( n )). In Rd , one may hope
that such an algorithm can be converted into an exact
medoid algorithm, but it is not clear how to do this.
Thus, while it may be possible that fast geometric median algorithms can provide inspiration in the development of medoid algorithms, they do not work out
of the box. Moreover, geometric median algorithms
cannot be used for network data as they only work
in vector spaces, thus they are useless for the spatial
network datasets which we consider in §5.
2.2
Medoid Algorithms : TOPRANK and
TOPRANK2
In Eppstein and Wang (2004), the RAND algorithm
for estimating the energy of all elements of a set
S = {x(1), . . . , x(N )} is presented. While RAND is presented in the context of graphs, where the N elements
are nodes of an undirected graph and the metric is
shortest path length, it can equally well be applied to
any set endowed with a distance. The simple idea of
RAND is to estimate the energy of each element from a
sample of anchor nodes I, so that for j ∈ {1, . . . , N },
1 X
Ê(j) =
dist(x(j), x(i)).
|I|
i∈I
James Newling, François Fleuret
An elegant feature of RAND in the context of sparse
graphs is that Dijkstra’s algorithm needs only be run
from anchor nodes i ∈ I, and not from every node.
The key result of Eppstein and Wang (2004) is the
following. Suppose that S has diameter ∆, that is
∆=
max
(i,j)∈{1,...,N }2
dist(x(i), x(j)),
and let > 0 be some error tolerance. If I is of size
Ω(log(N )/), then P(|E(j) − Ê(j)| > ∆) is O N12
for all j ∈ {1, . . . , N }. Using the union bound, this
means there is a O N1 probability that at least one
energy estimate is off by more than ∆, and so we say
that with high probability (w.h.p.) all errors are less
than ∆.
RAND forms the basis of the TOPRANK algorithm
of Okamoto et al. (2008). Whereas RAND w.h.p. returns an element which has energy within of the
minimum, TOPRANK is designed to w.h.p. return the
true medoid. In motivating TOPRANK, Okamoto et al.
(2008) observe that the expected difference between
consecutively ranked energies is O(∆/N ), and so if one
wishes to correctly rank all nodes, one needs to distinguish between energies at a scale = ∆/N , for which
the result of Eppstein and Wang (2004) dictates that
Θ(N log N ) anchor elements are required with RAND,
which is more elements than S contains. However, to
obtain just the highest ranked node should require less
information than obtaining a full ranking of nodes, and
it is to this task that TOPRANK is adapted.
The idea behind TOPRANK is to accurately estimate
only the energies of promising elements. The algorithm proceeds in two passes, where in the first pass
promising elements are earmarked. Specifically, the
first pass runs RAND with N 2/3 log1/3 (N ) anchor elements to obtain Ê(i) for i ∈ {1, . . . , N }, and then
discards elements whose Ê(i) lies below threshold τ
given by,
31
ˆ 0 log n
τ = arg min Ê(j) + 2∆α
,
(2)
n
j∈{1,...,N }
ˆ is an upper bound on ∆ obtained from the
where ∆
anchor nodes, and α0 is some constant satisfying α0 >
1. The second pass computes the true energy of the
undiscarded elements, returning the one with lowest
true energy. Note that a smaller α0 value results in a
lower (better) threshold, we discuss this point further
in SM-C.
To obtain run time guarantees, TOPRANK requires that
the distribution of node energies is non-decreasing near
to the minimum, denoted by E ∗ . More precisely, letting fE be the probability distribution of energies, the
algorithms require the existence of > 0 such that,
E ∗ ≤ ẽ < e < E ∗ + =⇒ fE (ẽ) ≤ fE (e).
(3)
5
If assumption 3 holds, then the run time is Õ(N 3 ). A
second algorithm presented in Okamoto et al. (2008)
is TOPRANK2, where the anchor set I is grown incrementally until some heuristic criterion is met. There
is no runtime guarantee for TOPRANK2, although it has
the potential to run much faster than TOPRANK under
favourable conditions. Pseudocode for RAND, TOPRANK
and TOPRANK2 is presented in SM-C.
2.3
K-medoids algorithm : KMEDS
The Voronoi iteration algorithm, which we refer to as
KMEDS, is similar to lloyd, the main difference being
that cluster medoids are computed instead of cluster
means. It has been desribed in the literature at least
twice, once in Hastie et al. (2001) and then in Park
and Jun (2009), where a novel initialisation scheme is
developed. Pseudocode is presented in SM-B.
All N 2 distances are computed and stored upfront with
KMEDS. Then, at each iteration, KN comparisons are
made during assignment and Ω(N 2 /K) additions are
made during medoid update. The initialisation scheme
of KMEDS requires all N 2 distances. Each iteration of
KMEDS requires retrieving at least max KN, N 2 /K
distinct distances, as can be shown by assuming balanced clusters.
As an alternative to computing all distances upfront,
one could store per-cluster distance matrices which get
updated on-the fly when assignments change. Using
such an approach,the best one could hope for would be
max KN, N 2 /K distance calculations and Θ(N 2 /K)
memory. If one were to completely forego storing distances in memory and calculate distances only when
needed, the number of distance calculations would be
at least r(KN + N 2 /K), where r is the number of
iterations.
The initialisation scheme of Park and Jun (2009) selects K well centered elements as initial medoids. This
goes against the general wisdom for K-means initialisation, where centroids are initialised to be well separated (Arthur and Vassilvitskii, 2007). While the new
scheme of Park and Jun (2009) performs well on a
limited number of small 2-D datasets, we show in § 3
that in general uniform initialisation performs as well
or better.
3
Our new medoid algorithm : trimed
We present our new algorithm, trimed, for determining the medoid of set S = {x(1), . . . , x(N )}. Whereas
the approach with TOPRANK is to empirically estimate
E(i) for i ∈ {1, . . . , N }, the approach with trimed,
presented as Alg. 1, is to bound E(i). When trimed
terminates, an index m∗ ∈ {1, . . . , N } has been de-
A Sub-Quadratic Exact Medoid Algorithm
termined, along with lower bounds l(i) for all i ∈
{1, . . . , N }, such that E(m∗ ) ≤ l(i) ≤ E(i), and thus
x(m∗ ) is the medoid. The bounding approach uses the
triangle inequality, as depicted in Figure 1.
Algorithm 1 The trimed algorithm for computing
the medoid of {x(1), . . . , x(N )}.
1: l ← 0N // lower bounds on energies, maintained
such that l(i) ≤ E(i) and initialised as l(i) = 0.
2: mcl , E cl ← −1, ∞ // index of best medoid candidate found so far, and its energy.
3: for i ∈ shuffle ({1, . . . , N }) do
4:
if l(i) < E cl then
5:
for j ∈ {1, . . . , N } do
6:
d(j) ← dist(x(i), x(j))
7:
end for P
N
8:
l(i) ← N1 j=1 d(j) // set l(i) to be tight,
that is l(i) = E(i).
9:
if l(i) < E cl then
10:
mcl , E cl ← i, l(i)
11:
end if
12:
for j ∈ {1, . . . , N } do
13:
l(j) ← max(l(j), |l(i) − d(j)|) // using
E(i) and dist(x(i), x(j)) to possibly improve bound on E(j).
14:
end for
15:
end if
16: end for
17: m∗ , E ∗ ← mcl , E cl
18: return x(m∗ )
The algorithm trimed iterates through the N elements
of S. Each time a new element with energy lower than
the current lowest energy (E cl ) is found, the index
of the current best medoid (mcl ) is updated (line 10).
Lower bounds on energies are used to quickly eliminate
poor medoid candidates (line 4). Specifically, if lower
bound l(i) on the energy of element i is greater than
or equal to E cl , then i is eliminated. If the bound test
fails to eliminate element i, then it is computed, that is,
all distances to element i are computed (line 6). The
computed distances are used to potentially improve
lower bounds for all elements (line 13). Theorem 3.1
states that trimed finds the medoid. The proof relies
on showing that lower bounds remain consistent when
updated (line 13).
The algorithm is very straightforward to implement,
and requires only two additional floating point values
per datapoint: for sample i, one for l(i) and one for
d(i). Computing either all or no distances from a sample makes particularly good sense for network data,
where computing all distances to a single node is efficiently performed using Dijkstra’s algorithm.
Theorem 3.1. trimed returns the medoid of set S.
x(j)
x(i)
x(j)
x(i)
Figure 1: Using the inequality E(j) ≥ |E(i) −
dist(x(i), x(j)) | to eliminate x(j) as a medoid candidate. Computed element x(i) with energy E(i) ≥ E cl
is used as a pivot to lower bound E(j). The two
cases where the inequality is effective are when (case
1, above) dist(x(i), x(j)) − E(i) ≥ E cl and (case 2,
below) E(i) − dist(x(i), x(j)) ≥ E cl , as both lead to
E(j) ≥ E cl which eliminates x(j) as a medoid candidate.
Proof. We need to prove that l(j) ≤ E(j) for all j ∈
{1, . . . , N } at all iterations of the algorithm. Clearly,
as l(j) = 0 at initialisation, we have l(j) ≤ E(j) at initialisation. E(j) does not change, and the only time
that l(j) may change is on line 13, where we need to
check that |l(i) − d(j)| ≤ E(j). At line 13, l(i) = E(i)
from line 8, and d(j) = dist(x(i), x(j)), so at line 13 we
are effectively checking that |E(i) − dist(x(i), x(j))| ≤
E(j). But this is a simple consequence of the triangle inequality,
PNas we now show. Using the definition,
E(j) = N1 l=1 dist(x(l), x(j)), we have on the one
hand,
E(j) ≥
N
1 X
dist(x(l), x(i)) − dist(x(i), x(j))
N
l=1
≥ E(i) − dist(x(i), x(j)),
(4)
and on the other hand,
E(j) ≥
N
1 X
dist(x(i), x(j)) − dist(x(l), x(i))
N
l=1
≥ dist(x(i), x(j)) − E(i).
(5)
Combining (4) and (5) we obtain the required inequality |E(i) − dist(x(i), x(j))| ≤ E(j).
The bound test (line 4) becomes more effective at later
iterations, for two reasons. Firstly, whenever an element is computed, the lower bounds of other samples
may increase. Secondly, E cl will decrease whenever a
better medoid candidate is found. The main result of
this paper, presented as Theorem 3.2, is that in Rd
1
the expected number of computed elements is O(N 2 )
under some weak assumptions. We show in §5 that
James Newling, François Fleuret
1
The shuffle on line 3 is performed to avoid w.h.p.
pathological orderings, such as when elements are ordered in descending order of energy which would result
in all N elements being computed.
Theorem 3.2. Let S = {x(1), . . . , x(N )} be a set of
N elements in Rd , drawn independently from probability distribution function fX . Let the medoid of S be
x(m∗ ), and let E(m∗ ) = E ∗ . Suppose that there exist strictly positive constants ρ, δ0 and δ1 such that for
any set size N with probability 1 − O(1/N )
x ∈ Bd (x(m∗ ), ρ) =⇒ δ0 ≤ fX (x) ≤ δ1 ,
∗
(7)
∗
2
E(i) − E ≥ αkx(i) − x(m )k .
Then, the expected
number of elements
computed by
1
4 d
N 2 , where Vd [1] =
trimed is O Vd [1]δ1 + d α
d
π 2 /(Γ( d2 + 1)) is the volume of Bd (0, 1).
3.1
On the assumptions in Theorem 3.2
The assumption of constants ρ, δ0 and δ1 made in Theorem 3.2 is weak, and only pathological distributions
might fail it, as we now discuss. For the assumptions to
fail requires that fX vanishes or diverges at the distribution medoid. Any reasonably behaved distribution
does not have this behaviour, as illustrated in Figure 2.
The constant α is a strong convexity constant. The
existence of α > 0 is guaranteed by the existence of
ρ, δ0 and δ1 , as the mean of a sum of uniformly spaced
cones converges to a quadratic function. This is illustrated in 1-D in Figure 5 in SM-G, but holds true in
any dimension.
Note that the assumptions made are on the distribution fX , and not on the data itself. This must be so
in order to prove complexity results in N .
3.2
fX
δ0
αkx − x(m∗ )k2
(6)
where Bd (x, r) = {x0 ∈ Rd : kx0 − xk ≤ r}. Let
α > 0 be a constant (independent of N ) such that with
probability 1 − O(1/N ) all i ∈ {1, . . . , N } satisfy,
x(i) ∈ Bd (x(m∗ ), ρ) =⇒
δ1
E − E∗
the O(N 2 ) result holds even in settings where the assumptions are not valid or relevent, such as for network
data.
Sketch of proof of Theorem 3.2
We now sketch the proof of Theorem 3.2, showing
how (6) and (7) are used. A full proof is presented
in SM-G. Firstly, let the index of the first element
after the shuffle on line 3 be i0 . Then, no elements
beyond radius 2E(i0 ) of x(i0 ) will subsequently be
computed, due to type 1 eliminations (see Figure 1).
Therefore, all computed elements are contained within
Bd (x(i0 ), 2E(i0 )).
x(m∗ )
x(m∗ ) + ρ
Figure 2: Illustration in 1-D of the constants used in
Theorem 3.2. Above, δ0 and δ1 bound the probability
density function in a region containing the distribution
medoid. Below, the energy of samples grows quadratically around the medoid x(m∗ ). The energy E is a sum
of cones centered on samples, which is approximately
quadratic unless fX vanishes or explodes, guaranteeing the existence of α > 0 required in Theorem 3.2.
Next, notice that once an element x(i) has been
computed in trimed, no elements in the ball
Bd (x(i), E(i) − E cl ) will subsequently be computed,
due to type 2 eliminations (see Figure 1). We refer to
such a ball as an exclusion ball. By upper bounding the
number of exclusion balls contained in Bd (x(i0 ), 2E(i0 ))
using a volumetric argument, we can obtain a bound
on the number of computed elements, but obtaining
such an upper bound requires that the radii of exclusion ball E(i)−E cl be bounded below by a strictly positive value. However, by using a volumetric argument
only beyond a certain positive radius of the medoid
(a radius N −1/2d ), we have α > 0 in (15) which provides a lower bound on exclusion ball radii, assuming
E cl ≈ E ∗ . Using δ0 we can show that E cl approaches
E ∗ sufficiently fastsufficiently fast to validate the approximation E cl ≈ E ∗ .
It then remains to count the number of computed
elements within radius N −1/2d of the medoid. One
cannot find a strict upper bound here, but using the
boundedness of fX provided by δ1 , we have w.h.p. that
the number of elements computed within N −1/2d is
O(δ1 N 1/2 ), as the volume of a sphere scales as the
d’th power of its radius.
A Sub-Quadratic Exact Medoid Algorithm
4
Our accelerated K-medoids
algorithm : trikmeds
We adapt our new medoid algorithm trimed and borrow ideas from Elkan (2003) to show how KMEDS can
be accelerated. We abandon the initial N 2 distance
calculations, and only compute distances when necessary. The accelerated version of lloyd of Elkan
(2003) maintains KN bounds on distances between
points and centroids, allowing a large proportion of
distance calculations to be eliminated. We use this
approach to accelerate assignment in trikmeds, incurring a memory cost O(KN ). By adopting the
algorithm of Newling and Fleuret (2016a) or that
of Hamerly (2010), the memory overhead can be reduced to O(N ). We accelerate the medoid update step
by adapting trimed, reusing lower bounds between iterations, so that trimed is only run from scratch once
at the start. Details and pseudocode are presented in
SM-H.
One can relax the bound test in trimed so that for
> 0 element i is computed if l(i)(1 + ) < E cl , guaranteeing that an element with energy within a factor
1 + of E ∗ is found. It is also possible to relax the
bound tests in the assignment step of trikmeds, such
that the distance to an assigned cluster’s medoid is always within a factor 1+ of the distance to the nearest
medoid. We denote by trikmeds- the trikmeds algorithm where the update and assignment steps are
relaxed as just discussed, with trikmeds-0 being exactly trikmeds. The motivation behind such a relaxation is that, at all but the final few iterations, it is
probably a waste of computation obtaining medoids
and assignments at high resolution, as in subsequent
iterations they may change.
5
Results
We first compare the performance of the medoid algorithms TOPRANK, TOPRANK2 and trimed. We then compare the K-medoids algorithms, KMEDS and trikmeds.
5.1
Medoid algorithm results
We compare our new exact medoid algorithm trimed
with state-of-the-art approximate algorithms TOPRANK
and TOPRANK2. Recall, Okamoto et al. (2008) prove
that the approximate algorithms return w.h.p. the true
medoid. We confirm that this is the case in all our experiments, where the approximate algorithms return
the same element as trimed, which we know to be
correct by Theorem 3.1. We now focus on comparing computational costs, which are proportional to the
number of computed points.
Results on artificial datasets are presented in Figure 3,
where our two main observations relate to scaling in
N and dimension d. The artificial data are (left)
uniformly drawn from [0, 1]d and (right) drawn from
Bd (0, 1) with probability of lying within radius 1/21/d
of 1/200, as opposed to 1/2 as would be the case under uniform density. Details about sampling from this
distribution can be found in SM-F. Results on a mix of
publicly available real and artificial datasets are presented in Table 1 and discussed in §5.1.2.
5.1.1
Scaling with N and d on artificial
datasets
In Figure 3 we observe that the number of points computed by trimed is O(N 1/2 ), as predicted by Theorem 3.2. This is illustrated (right) by the close fit of
the number of computed points to exact square root
curves at sufficiently large N for d ∈ {2, 6}.
Recall that TOPRANK consists of two passes, a first
where N 2/3 log1/3 N anchor points are computed, and
a second where all sub-threshold points are computed.
We observe that for small N TOPRANK computes all
N points, which corresponds to all points lying below threshold. At sufficiently large N the threshold
becomes low enough for all points to be eliminated after the first pass. The effect is particularly dramatic
in high dimensions (d = 6 on right), where a phase
transition is observed between all and no points being
computed in the second pass.
Dimension d appears in Theorem 3.2 through a factor
d(4/α)d , where α is the strong convexity of the energy
at the medoid. In Figure 3, we observe that the number of computed points increases with d for fixed N ,
corresponding to a relatively small α. The effect of α
on the number of computed elements is considered in
greater detail in SM-F.
In contrast to the above observation that the number
of computed points increases as dimension increases
for trimed, TOPRANK appears to scale favourably with
dimension. This observation can be explained in terms
of the distribution of energies, with energies close to E ∗
being less common in higher dimensions, as discussed
in SM-J.
5.1.2
Results on publicly available real and
simulated datasets
We present the datasets used here in detail in SM-I.
For all datasets, algorithms TOPRANK, TOPRANK2 and
trimed were run 10 times with a distinct seed, and
the mean number of iterations (n̂) over the 10 runs
was computed. We observe that our algorithm trimed
is the best performing algorithm on all datasets, although in high-dimensions (MNIST-0) and on social
James Newling, François Fleuret
number of computed elements
106
106
TOPRANK
trimed
d=2
d=3
d=4
d=5
d=6
105
104
105
d = 2 TOPRANK
d = 2 trimed
d = 6 TOPRANK
d = 6 trimed
104
103
N 1
18N 2
1
2
N 3 1log 3 N
3N 2
103
102 2
10
103
104
N
105
106
102 3
10
104
105
N
106
107
Figure 3: Comparison of TOPRANK and our algorithm trimed on simulated data. On the left, points are drawn
uniformly from [0, 1]d for d ∈ {2, . . . , 6}, and on the right they are drawn from Bd (0, 1) for d ∈ {2, 6}, with an
increased density near the edge of the ball. Fewer points (elements) are computed by trimed than by TOPRANK
in all scenarios. For small N , TOPRANK computes O(N ) points, before transitioning to Õ(N 2/3 ) computed points
for large N . trimed computes O(N 1/2 ) points. Note that trimed performs better in low-d than in high-d, with
the reverse trend being true for TOPRANK. These observations are discussed in further detail in the text.
network data (Gnutella) no algorithm computes significantly fewer than N elements. The failure in
high-dimensions (MNIST-0) of trimed is in agreement with Theorem 3.2, where dimension appears as
the exponent of a constant term. The small world
network data, Gnutella, can be embedded in a highdimensional Euclidean space, and thus the failure on
this dataset can also be considered as being due to
high-dimensions. For low-dimensional real and spatial
network data, trimed consistently computes O(N 1/2 )
elements.
5.1.3
But who needs the exact medoid
anyway?
A valid criticism that could be raised at this stage
would be that for large datasets, finding the exact
medoid is probably overkill, as any point with energy reasonably close to E ∗ suffices for most applications. But consider, the RAND algorithm requires
computing log N/2 elements to confidently return an
element with energy within E ∗ of E ∗ . For N = 105
and = 0.05, this is 4600, already more than trimed
requires to obtain the exact medoid on low-d datasets
of comparable size.
5.2
K-medoids algorithm results
With N elements to cluster, KMEDS is Θ(N 2 ) in memory, rendering it unusable on even moderately large
datasets. To compare the initialisation scheme proposed in Park and Jun (2009) to random initialisation,
we have performed experiments on 14 small datasets,
with K ∈ {10, dN 1/2 e, dN/10e}. For each of these 42
experimental set-ups, we run the deterministic KMEDS
initialisation once, and then uniform random initialisation, 10 times. Comparing the mean final energy
of the two initialisation schemes, in only 9 of 42 cases
does KMEDS initialisation result in a lower mean final
energy. A Table containing all results from these experiments in presented in SM-E.
Having demonstrated that random uniform initialisation performs at least as well as the initialisation
scheme of KMEDS, and noting that trikmeds-0 returns
exactly the same clustering as would KMEDS with uniform random initialisation, we turn our attention to
the computational performance of trikmeds. Table 2
presents results on 4 datasets, each described in SMI. The first numerical column is the relative number
of distance calculations using trikmeds-0 and KMEDS,
where large savings in distance calculations, especially
in low-dimensions, are observed. Columns φc and φE
are the number of distance calculations and energies
respectively, using ∈ {0.01, 0.1}, relative to = 0.
We observe large reductions in the number of distance
computations with only minor increases in energy.
6
Conclusion and future work
We have presented our new trimed algorithm for computing the medoid of a set, and provided strong theoretical guarantees about its performance in Rd . In
low-dimensions, it outperforms the state-of-the-art approximate algorithm on a large selection of datasets.
The algorithm is very simple to implement, and can
easily be extended to the general ranking problem.
In the future, we propose to explore the idea of us-
A Sub-Quadratic Exact Medoid Algorithm
dataset
Birch 1
Birch 2
Europe
U-Sensor Net
D-Sensor Net
Pennsylvania road
Europe rail
Gnutella
MNIST
type
2-d
2-d
2-d
u-graph
d-graph
u-graph
u-graph
d-graph
784-d
N
1.0 × 105
1.0 × 105
1.6 × 105
3.6 × 105
3.6 × 105
1.1 × 106
4.6 × 104
6.3 × 103
6.7 × 103
TOPRANK TOPRANK2 trimed
n̂
n̂
n̂
57944
100180
2180
66062
100180
2208
176095
169535
2862
113838
327216
1593
99896
176967
1372
216390
time-out
2633
35913
47041
518
7043
6407
6328
7472
6799
6514
Table 1: Comparison of TOPRANK, TOPRANK2 and our algorithm trimed on publicly available real and simulated
datasets. Column 2 provides the type of the dataset, where ‘x-d’ denotes x-dimensional vector data, while
‘d-graph’ and ‘u-graph’ denote directed and undirected graphs respectively. Column n̂ gives the mean number
of elements computed over 10 runs. Our proposed trimed algorithm obtains the true medoid with far fewer
computed points in low dimensions and on spatial network data. On the social network dataset (Gnutella) and
the very high-d dataset (MNIST), all algorithms fail to provide speed-up, computing approximately N elements.
Dataset
Europe
Conflong
Colormo
MNIST50
N
1.6 × 105
1.6 × 105
6.8 × 104
6.0 × 104
=0
d Nc /N 2
2 0.067
3 0.042
9 0.163
50 0.280
K = 10
= 0.01
φc
φE
0.33 1.004
0.67 1.001
0.92 1.000
0.99 1.000
= 0.1
=0
φc
φE Nc /N 2
0.01 1.054 0.008
0.08 1.014 0.006
0.35 1.015 0.011
0.95 1.001 0.019
√
K = d Ne
= 0.01
= 0.1
φc
φE
φc
φE
0.68 1.031 0.39 1.090
0.92 1.003 0.61 1.026
0.98 1.000 0.82 1.005
0.99 1.001 0.97 1.001
Table 2: Relative numbers of distance calculations and final energies using trikmeds- for ∈ {0, 0.01, 0.1}. The
number of distance calculations with trikmeds-0 is Nc , presented here relative to the number computed using
KMEDS (N 2 ) in column Nc /N 2 . The number of distance calculations with ∈ {0.01, 0.1} relative to trikmeds-0
are given in columns φc , so φc = 0.33 means 3× fewer calculations than with = 0. The final energies with
∈ {0.01, 0.1} relative to trikmeds-0 are given in columns φE . We see that trikmeds-0 uses significantly fewer
distance calculations than would KMEDS, especially in low-dimensions where a greater than K× reduction is
observed (NC /N 2 < 1/K). For low-d, additional relaxation further increases the saving in distance calculations
with little cost to final energy.
ing more complex triangle inequality bounds involving
several points, with as goal to improve on the O(N 1/2 )
number of computed points.
We have demonstrated how trimed, when combined
with the approach of Elkan (2003), can greatly reduce
the number of distance calculations required by the
Voronoi iteration K-medoids algorithm of Park and
Jun (2009). In the future we would like to replace the
strategy of Elkan (2003) with that of Hamerly (2010),
which will be better adapted to graph clustering as
either all or no distances are computed with it, making
it more amenable to Dijkstra’s algorithm.
Acknowledgements
The authors are grateful to Wei Chen for helpful discussions of the TOPRANK algorithm. James Newling
was funded by the Hasler Foundation under the grant
13018 MASH2.
References
Arthur, D. and Vassilvitskii, S. (2007). K-means++:
The advantages of careful seeding. In Proceedings of
the Eighteenth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA ’07, pages 1027–1035,
Philadelphia, PA, USA. Society for Industrial and
Applied Mathematics.
Chipman, H., Hastie, T., and Tibshirani, R. (2003).
James Newling, François Fleuret
Statistical Analysis of Gene Expression Microarray
Data. Chapman & Hall. Chapter 4.
Cohen, E., Delling, D., Pajor, T., and Werneck, R. F.
(2014). Computing classic closeness centrality, at
scale. In Proceedings of the Second ACM Conference
on Online Social Networks, COSN ’14, pages 37–50,
New York, NY, USA. ACM.
Cohen, M. B., Lee, Y. T., Miller, G. L., Pachocki,
J. W., and Sidford, A. (2016). Geometric median in
nearly linear time. In STOC16. submitted.
Elkan, C. (2003). Using the triangle inequality to accelerate k-means. In Machine Learning, Proceedings
of the Twentieth International Conference (ICML
2003), August 21-24, 2003, Washington, DC, USA,
pages 147–153.
Eppstein, D. and Wang, J. (2004). Fast approximation
of centrality. J. Graph Algorithms Appl., 8(1):39–45.
Frahm, J.-M., Fite-Georgel, P., Gallup, D., Johnson,
T., Raguram, R., Wu, C., Jen, Y.-H., Dunn, E.,
Clipp, B., Lazebnik, S., and Pollefeys, M. (2010).
Building rome on a cloudless day. In Proceedings of
the 11th European Conference on Computer Vision:
Part IV, ECCV’10, pages 368–381, Berlin, Heidelberg. Springer-Verlag.
Hamerly, G. (2010). Making k-means even faster. In
SDM, pages 130–140.
Hastie, T. J., Tibshirani, R. J., and Friedman, J. H.
(2001). The elements of statistical learning : data
mining, inference, and prediction. Springer series in
statistics. Springer, New York.
Hoare, C. A. R. (1961). Algorithm 65: Find. Commun.
ACM, 4(7):321–322.
Kaufman, L. and Rousseeuw, P. J. (1990). Finding
groups in data : an introduction to cluster analysis. Wiley series in probability and mathematical
statistics. Wiley, New York. A Wiley-Interscience
publication.
Newling, J. and Fleuret, F. (2016a). Fast k-means
with accurate bounds. In Proceedings of the International Conference on Machine Learning (ICML),
pages 936–944.
Newling, J. and Fleuret, F. (2016b). K-medoids for
k-means seeding. arXiv:1609.04723. Under review.
Ng, R. T., Han, J., and Society, I. C. (2005). Clarans:
A method for clustering objects for spatial data mining. IEEE Transactions on Knowledge and Data
Engineering, pages 1003–1017.
Okamoto, K., Chen, W., and Li, X.-Y. (2008). Ranking of closeness centrality for large-scale social networks. In Proceedings of the 2Nd Annual International Workshop on Frontiers in Algorithmics, FAW
’08, pages 186–195, Berlin, Heidelberg. SpringerVerlag.
Park, H.-S. and Jun, C.-H. (2009). A simple and
fast algorithm for k-medoids clustering. Expert Syst.
Appl., 36(2):3336–3341.
Rattigan, M. J., Maier, M., and Jensen, D. (2007).
Graph clustering with network structure indices. In
Proceedings of the 24th International Conference on
Machine Learning, ICML ’07, pages 783–790, New
York, NY, USA. ACM.
A Sub-Quadratic Exact Medoid Algorithm
SM-A
On the difficulty of the medoid problem
We construct an example showing that no general purpose algorithm exists to solve the medoid problem in
o(N 2 ). Consider an almost fully connected graph containing N = 2m + 1 nodes, where the graph is exactly m
edges short of being fully connected: one node has 2m edges and the others have 2m − 1 edges. The graph has
2m2 edges. With the shortest path metric, it is easy to see that the node with 2m edges is the medoid, hence the
medoid problem is as difficult as finding the node with 2m edges. But, supposing that the edges are provided as
an unsorted adjacency list, it is clearly an O(m2 ) task to determine which node has 2m edges as one must look
at all edges until a node with 2m edges is found. Thus determining the medoid is O(m2 ) which is O(N 2 ).
SM-B
KMEDS pseudocode
Alg. 2 presents the KMEDS algorithm of Park and Jun (2009), with the novel initialisation of KMEDS on line 1.
KMEDS is essentially lloyd, with medoids instead of means.
Algorithm 2 KMEDS for clustering data {x(1), . . . , x(N )} around K medoids
P
1: Set all distances D(i, j) ← kx(i) − x(j)k and sums S(i) ← j∈{1,...,N } D(i, j)
P
2: Initialise medoid indices as K indices minimising f (i) = j∈{1,...,N } D(i, j)/S(j)
3: while Some convergence criterion has not been met do
4:
Assign each element to the cluster whose medoid is nearest to the element
5:
Update cluster medoids according to assignments made above
6: end while
SM-C
RAND, TOPRANK and TOPRANK2 pseudocode
We present pseudocode for the RAND, TOPRANK and TOPRANK2 algorithms of Okamoto et al. (2008), and discuss
the explicit and implicit constants.
SM-C.1
2
1
On the number of anchor elements in TOPRANK : the constant in Θ(N 3 (log N ) 3 )
Note that the number of anchor points used in TOPRANK does not affect the result that the medoid is w.h.p.
1
returned. However, Okamoto et al. (2008) show that by choosing the size of the anchor set to be q (log N ) 3 for
any q, the run time is guaranteed to be Õ(N 5/3 ). They do not suggest a specific q, the optimal q being dataset
dependant. We choose q = 1.
Consider Figure 3 in Section 5.1 for example, where q = 1. Had q be chosen to be less than 1, the line
ncomputed = N 2/3 log1/3 N to which TOPRANK runs parallel for large N would be shifted up or down by log q,
however the N at which the transition from ncomputed = N 2/3 log1/3 N to ncomputed = N 2/3 log1/3 N takes
place would also change.
SM-C.2
On the parameter α0 in TOPRANK and TOPRANK2
The threshold τ in (2) is proportional to the parameter α0 . In Okamoto et al. (2008), it is stated that α0 should
be some value greater than 1. Note that the smaller α0 is, the lower the threshold is, and hence fewer the number
of computed points is, thus α0 = 1.00001 would be a fair choice. We use α0 = 1 in our experiments, and observe
that the correct medoid is returned in all experiments.
Personal correspondence with the authors of Okamoto et al. (2008) has brought into doubt the proof of the
result that the medoid is w.h.p. returned for any α0 where α0 > 1. In our most recent correspondence,
the
√
authors suggest that the w.h.p. result can be proven with the more conservative bound of α0 > 1.5. Moreover,
0
we show in SM-D that α0 > 1 is good enough to return the medoid with probability N −(α −1) , a probability
which still tends to 0 as N grows large, but not a w.h.p. result. Please refer to SM-D for further details on our
correspondence with the authors.
James Newling, François Fleuret
SM-C.3
On the parameters specific to TOPRANK2
In addition to α0 , TOPRANK2 requires two parameters to be set. The first is l0 , the starting anchor set size, and
the second is q, the amount by which l should be incremented at each iteration. Okamoto et al. (2008) suggest
taking l0 to be the number of top ranked nodes required, which in our case would be l0 = k = 1. However, in
our experience this is too small as all nodes lie well within the threshold and thus when l increases there is no
change to number below threshold, which makes the algorithm break out of the search for the optimal l too early.
Indeed, l0 needs to be chosen so that at least some points
√ have energies greater than the threshold, which in our
experiments is already quite large. We choose l0 = N , as any value larger than N 2/3 would make TOPRANK2
redundant to TOPRANK. The parameter q we take to be log N as suggested by Okamoto et al. (2008).
Algorithm 3 RAND for estimating energies of elements of set S (Eppstein and Wang, 2004).
I ← random uniform sample from {1, . . . , N }
// Compute all distances from anchor elements (I), using Dijkstra’s algorithm on graphs
for i ∈ I do
for j ∈ {1, . . . , N } do
d(i, j) ← kx(i) − x(j)k,
end for
end for
// Estimate energies as mean distances to anchor elements
for j ∈ {1, . . .P
, N } do
1
Ê(j) ← |I|
i∈I d(i, j)
end for
return Ê
Algorithm 4 TOPRANK for obtaining top k ranked elements of S (Okamoto et al., 2008).
2
1
1
l ← N 3 (log N ) 3 // Okamoto et al. (2008) state that l should be Θ((log N ) 3 ), the choice of 1 as the constant
is arbitrary (see comments in the text of Section SM-C.1).
Run RAND with uniform random I of size l to get Ê(i) for i ∈ {1, . . . , N }.
Sort Ê so that Ê[1] ≤ Ê[2] ≤ . . . ≤ Ê[N ]
ˆ ← 2 mini∈I maxj∈{1,...,N } kx(i) − x(j)k // where kx(i) − x(j)k computed in RAND
∆
q
Q ← i ∈ {1, . . . , N } | Ê(i) ≤ Ê[k] + 2α0 ∆ log(n)
.
l
Compute exact energies of all elements in Q and return the element with the lowest energy.
SM-D
On the proof that TOPRANK returns the medoid with high probability
Through correspondence with the authors of Okamoto et al. (2008), we have located a small problem in the
proof that the medoid is returned w.h.p. for α0 > 1, the problem lying in the second inequality of Lemma 1. To
arrive at this inequality, the authors have used the fact that for all i,
P(E(i) ≥ Ê(i) + f (l) · ∆) ≥ 1 −
1
,
2N 2
(8)
which is a simple consequence of the Hoeffding inequality as shown in Eppstein and Wang (2004). Essentially (8)
says that, for a fixed node i, from which the mean distance to other nodes is E(i), if one uniformly samples l
distances to i and computes the mean Ê(i), the probability that Ê(i) is less than E(i) + f (l) is greater than
1 − 2N1 2 .
The inequality (8) is true for a fixed node i. However, it no longer holds if i is selected to be the node with the
lowest Ê(i). To illustrate this, suppose that E(i) = 1 for all i, and compute Ê(i) for all i. Let Ê ∗ = arg mini Ê(i).
Now, we have a strong prior on Ê ∗ being significantly less than 1, and (8) no longer holds as a statement made
about Ê ∗ .
In personal correspondence, the authors show that the problem can be fixed by the use of an additional layer
of union bounding, with a correction to be published (if not already done so at time of writing). However, the
A Sub-Quadratic Exact Medoid Algorithm
Algorithm 5 TOPRANK2 for obtaining top k ranked elements of S (Okamoto et al., 2008).
// In Okamoto et al. (2008), it is suggested that l0 be taken as k, which in the case of the medoid problem is
1. We have experimented with several choices for l0 , as discussed in the text.
l ← l0
Run RAND with uniform random I of size l to get Ê(i) for i ∈ {1, . . . , N }.
ˆ ← 2 mini∈I maxj∈{1,...,N } kx(i) − x(j)k // where kx(i) − x(j)k computed in RAND
∆
Sort Ê
so that Ê[1] ≤ Ê[2] ≤ . . . ≤ Ê[N ] q
log(n)
0
.
Q ← i ∈ {1, . . . , N } | Ê(i) ≤ Ê[k] + 2α ∆
l
g←1
while g is 1 do
p ← |Q|
// The recommendation for q in Okamoto et al. (2008) is log(n), we follow the suggestion
Increment I with q new anchor points
Update Ê for all data according to new anchor points
l ← |I|
ˆ ← 2 mini∈I maxj∈{1,...,N } kx(i) − x(j)k
∆
Sort Ê
so that Ê[1] ≤ Ê[2] ≤ . . . ≤ Ê[N ] q
log(n)
0
Q ← i ∈ {1, . . . , N } | Ê(i) ≤ Ê[k] + 2α ∆
l
p0 ← |Q|
if p − p0 < log (n) then
g←0
end if
end while
Compute exact energies of all elements in Q and return the element with the lowest energy
additional layer of union bound requires a more conservative constraint
on α0 , which is α0 > 2, although the
√
0
authors propose that the w.h.p. result can be proven
√ with α > 1.5 for N sufficiently large. We now present
a small proof proving the w.h.p. result for α0 > 2 for N sufficiently large, with at the same time α0 > 1
0
guaranteeing that the medoid is returned with probability O(N α −1 ).
SM-D.1
That the medoid is returned with high probability holds for α0 >
vanishing probability it is returned for α0 > 1
√
2 and that with
Recall that we have N nodes with energies E(1), . . . , E(n). We wish to find the k lowest energy nodes (the
original setting of Okamoto et al. (2008)). From Hoeffding’s inequality we have,
P(|E(i) − Ê(i)| ≥ ∆) ≤ 2 exp −l2 .
(9)
Set the probability on the right hand side of 9 to be 2/N 1+β , that is,
2 exp −l2 = 2/N 1+β ,
which corresponds to
s
=
Clearly
√
1+β
l
log (N ) := f˜(l).
1 + β corresponds to α0 . With this notation we have,
P(|E(i) − Ê(i)| ≥ f˜(l)∆) ≤
2
N 1+β
.
Applying the union bound to (10) we have,
2
P ¬ ∧i∈{1,...,N } |E(i) − Ê(i)| ≤ f˜(l)∆ ≤ β .
N
(10)
(11)
James Newling, François Fleuret
K = 10
Dataset
gassensor
house16H
S1
S2
S3
S4
A1
A2
A3
thyroid
yeast
wine
breast
spiral
K=
l√ m
N
K=
N
10
N
d µu /µpark σu /µpark µu /µpark σu /µpark µu /µpark σu /µpark
256 128
1.09
0.08
0.90
0.03
0.83
0.01
1927 17
1.01
0.02
0.97
0.01
0.93
0.01
5000 2
1.05
0.05
0.75
0.01
0.32
0.01
5000 2
1.04
0.07
0.68
0.01
0.34
0.00
5000 2
1.03
0.05
0.76
0.01
0.35
0.00
5000 2
1.02
0.03
0.75
0.01
0.41
0.01
3000 2
0.82
0.03
0.43
0.01
0.19
0.00
5250 2
0.98
0.03
0.47
0.01
0.25
0.00
7500 2
0.96
0.02
0.42
0.02
0.22
0.00
215 5
0.95
0.08
0.97
0.04
0.93
0.04
1484 8
1.00
0.02
0.96
0.02
0.91
0.02
178 14
1.01
0.02
1.02
0.01
0.98
0.02
699 9
0.79
0.03
0.77
0.02
0.68
0.02
312 3
1.03
0.03
0.99
0.02
0.82
0.03
Table 3: Comparing the initialisation scheme proposed in Park and Jun (2009) with random uniform initialisation
for the KMEDS algorithm. The final energy using the deterministic scheme proposed in Park and Jun (2009) is
µpark . The mean over 10 random uniform initialisations is µu , and the corresponding standard deviation is σu .
For small K (K = 10), the performances using the two schemes are comparable, while for larger K, it is clear
that uniform initialisation performs much better on the majority of datasets.
Recall that we wish to obtain the k nodes with lowest energy. Denote by r(j) the index of the node with the
j’th lowest energy, so that
E(r(1)) ≤ . . . ≤ E(r(j)) ≤ . . . ≤ E(r(N )).
Denote by r̂(j) the index of the node with the j’th lowest estimated energy, so that
Ê(r̂(1)) ≤ . . . ≤ Ê(r̂(j)) ≤ . . . ≤ Ê(r̂(N )).
Now assume that for all i, it is true that |E(i) − Ê(i)| ≤ f˜(l). Then consider, for j ≤ k,
Ê(r̂(k)) − Ê(r(j)) = Ê(r̂(k)) − E(r(k)) + E(r(k)) − E(r(j)) + E(r(j)) − Ê(r(j)) ,
|
{z
} |
{z
} |
{z
}
≥−f˜(l)∆
≥0
(12)
≥−f˜(l)∆
≥ −2f˜(l)∆.
The first bound in (12) is obtained by considering the most extreme case possible under the assumption, which
is Ê(i) = a(E) − f˜(l) for all i. The second bound follows from j ≤ k, and the third bound follows directly from
the assumption. We thus have that, under the assumption,
Ê(r(j)) ≤ Ê(r̂(k)) + 2f˜(l)∆,
which says that all nodes of rank less than or equal to k have approximate energy less than Ê(r̂(k)) + 2f˜(l)∆.
As the assumption holds with probability greater than 1 − 2/N β by (11), we are done. Take β = 1 if you want
the statement with high probability, that is
r
2 log(n)
,
=
l
but for any β > 0, which corresponds to α0 > 1, the probability of failing to return the k lowest energy nodes
tends to 0 as N grows.
SM-E
On the initialisation of Park and Jun (2009)
In Table 3 we present the full results of the 48 experiments comparing the initialisation proposed in Park and
Jun (2009) with simple uniform initialisation. The 14 datasets are all available from https://cs.joensuu.fi/
sipu/datasets/.
number of points computed
A Sub-Quadratic Exact Medoid Algorithm
2.5 1e4
2.0
1.5
d =2
d =3
d =4
d =5
p
22.3 N
p
15.5 N
1.0
p
10.6p N
7.8 N
0.5
0.0
0.0
0.2
0.4
0.6
N
0.8
1.0
1.2 1e4
1.0
0.8
0.6
0.4
0.2
0.0
1.2 0.0
0.2
1e6
p
11.2 N
p
7.1 N
p
5.0p N
3.7 N
0.4
0.6
N
0.8
1.0
1.2
1e6
Figure 4: Number of points computed on simulated data. Points are drawn from Bd (0, 1), for d ∈ {2, 3, 4, 5}.
On the left, points are drawn uniformly, while on the right, the density in Bd (0, (1/2)1/d ) is 19× lower that
in Ad (0, (1/2)1/2 , 1), where recall that Ad (x, r1 , r2 ) denotes an annulus centred at x√of inner radius r1 and
outer radius r2 . We observe a near perfect fit of the number of computed points to ξ N where the constant
ξ depends on the dimension and the distribution (left and right). The number of computed points increases
with dimension. The strong convexity constant of the distribution on the right is larger, corresponding to fewer
distance calculations as predicted by Theorem 3.2.
SM-F
Scaling with α, N , and dimension d
We perform more experiments to provide further validation of Theorem 3.2. In particular, we check how the
number of computed elements scales with N , d and α. We generate data from a unit ball in various dimensions,
according to two density functions with different strong convexity constants α. The first density function is
uniform, so that the density everywhere in the ball is uniform. To sample from this distribution, we generate
two random variables, X1 ∼ Nd (0, 1) and X2 ∼ U (0, 1) and use
1
X3 = X1 /kX1 k · X2d ,
(13)
as a sample from the unit ball Bd (0, 1) with uniform distribution. The second distribution we consider has a
higher density beyond radius (1/2)1/d . Specifically, within this radius the density is 19× lower than beyond this
radius. To sample from this distribution, we sample X3 according to (13), and then points lying within radius
(1/2)1/d are with probability 1/10 re-sampled uniformly beyond this radius.
The second distribution has a larger strong convexity constant α. To see this, note that the strong convexity
constant at the center of the ball depends only on the density of the ball on its surface, that is at radius 1, as can
be shown using an argument based on cancelling energies of internal points. As the density at the surface under
distribution 2 is approximately twice that of under distribution 1, the change in energy caused by a small shift
in the medoid is twice as large under distribution 2. Thus, according to Theorem 3.2, we expect the number of
computed points to be larger under distribution 1 than under distribution 2. This is what we observe, as shown
in Figure 4, where distribution 1 is on the left and distribution 2 is on the right.
In Figure 4 we observe a near perfect N 1/2 scaling of number of computed points. Dashed curves are exact N 1/2
relationships, while the coloured points are the observed number of computed points.
SM-G
Proof of Theorem 3.2 (See page 5)
Theorem 3.2. Let S = {x(1), . . . , x(N )} be a set of N elements in Rd , drawn independently from probability
distribution function fX . Let the medoid of S be x(m∗ ), and let E(m∗ ) = E ∗ . Suppose that there exist strictly
positive constants ρ, δ0 and δ1 such that for any set size N with probability 1 − O(1/N )
x ∈ Bd (x(m∗ ), ρ) =⇒ δ0 ≤ fX (x) ≤ δ1 ,
(6)
James Newling, François Fleuret
where Bd (x, r) = {x0 ∈ Rd : kx0 −xk ≤ r}. Let α > 0 be a constant (independent of N ) such that with probability
1 − O(1/N ) all i ∈ {1, . . . , N } satisfy,
x(i) ∈ Bd (x(m∗ ), ρ) =⇒
(7)
∗
∗
2
E(i) − E ≥ αkx(i) − x(m )k .
Then, the expected number of elements computed by trimed is O Vd [1]δ1 + d
4 d
α
1
N 2 , where Vd [1] =
d
π 2 /(Γ( d2 + 1)) is the volume of Bd (0, 1).
Proof. We show that the assumptions made in Th. 3.2 validate the assumptions required in Thm SM-G.1.
Firstly, if e(i) > ρ then e(i) ≥ αρ2 e(i) > ρ, which follows from the convexity of the loss function and. Secondly,
the existance of β follows from continuity of the gradient of the distance, combined with the existence of δ1
(non-exploding).
Theorem SM-G.1 (Main Theorem Expanded). Let S = {x(1), . . . , x(N )} ⊂ Rd have medoid x(m∗ ) with minimum energy E(m∗ ) = E ∗ , where elements in S are drawn independently from probability distribution function
fX . Let e(i) = kx(i) − x(m∗ )k. Suppose that for fX there exist strictly positive constants α, β, ρ, δ0 and δ1
satisfying,
x ∈ Bd (x(m∗ ), ρ) =⇒ δ0 ≤ fX (x) ≤ δ1 ,
(14)
where Bd (x, r) = {x0 ∈ Rd : kx0 − xk ≤ r}, and that for any set size N , w.h.p. all i ∈ {1, . . . , N } satisfy,
(
αe(i)2 if e(i) ≤ ρ,
E(i) − E ∗ ≥
αρ2
if e(i) > ρ,
(15)
and,
E(i) − E ∗ ≤ βe(i)2
if
e(i) ≤ ρ.
(16)
Then
which is to say not eliminated on line 4 of trimed, is
the expected number
1 of elements computed,
d
d
O Vd [1]δ1 + d α4
N 2 , where Vd [1] = π 2 /(Γ( d2 + 1)) is the volume of Bd (0, 1).
E − E∗
x(m∗ )
Figure 5: A sum of uniformly distributed cones is approximately quadratic.
1
1
Proof. We first show that the expected number of computed elements in Bd (x(m∗ ), N − 2d ) is O(Vd [1]δ1 N 2 ).
1
When N is sufficiently large, fX (x) ≤ δ1 within Bd (x(m∗ ), N − 2d ). The expected number of samples in
1
Bd (x(m∗ ), N − 2d ) is thus upper bounded by δ1 multiplied by the volume of the ball. But the volume of a
1
1
ball of radius N − 2d in Rd is Vd [1]N − 2 .
In Lemma SM-G.2 we usea packing argument
to show that the number of computed elements in the annulus
1
1
4 d
∗
− 2d
Ad (x(m ), N
, ∞) is O d α N 2 , but we there assume that the medoid index m∗ is the first element in
shuffle({1, . . . , N }) on line 3 of trimed and thus that the medoid energy is known from the first iteration
(E cl = E ∗ ). We now extend Lemma SM-G.2 to the case where the medoid is not the first element processed. We
1
do this by showing that w.h.p. an element with energy very close to E ∗ has been computed after N − 2 iterations
of trimed, and thus that the bounds on numbers of computed elements obtained using the packing arguments
1
underlying Lemma SM-G.2 are all correct to within some small factor after N − 2 iterations.
A Sub-Quadratic Exact Medoid Algorithm
E(i) − E ∗
0.12
0.06
αρ2
0.4
−ρ
ρ
0.0
kx(i) − x(m∗ )k
0.4
Figure 6: Illustrating the parameters α, β and ρ of Theorem 3.2. Here we draw N = 101 samples uniformly from
[−1, 1] and compute their energies, plotted here as the series of points. Theorem 3.2 states that their exists α,
β and ρ such that irrespective of N , all energies (points) will lie in the envelope (non-hatched region).
2
2
The probability of a sample lying within radius N − 3d of x(m∗ ) is Ω(δ0 N − 3 ), and so the probability that none
1
2
2
1
2
1
of the first N 2 samples lies within radius N − 3d is O((1 − δ0 N − 3d )N ) which is O( N1 ). Thus w.h.p. after N 2
4
iterations of trimed, E cl is within βN − 3d of E ∗ , which means that the radii of the balls used in the packing
1
argument are overestimated by at most a factor N − 3d . Thus w.h.p. the upper bounds obtained with the packing
1
argument are correct to within a factor 1 + N − 3 . The remaining O( N1 ) cases do not affect the expectation, as
we know that no more than N elements can be computed.
Lemma SM-G.2 (Packing beyond the vanishing radius). If we assume (15) from Theorem 3.2 and that
the medoid index m∗ is the
element
processed by trimed, then the number of elements computed in
first
1
1
4 d
∗
− 2d
2
Ad (x(m ), N
, ∞) is O d α N .
Proof. Follows from Lemmas SM-G.3 and SM-G.4.
1
Lemma SM-G.3 (Packing from the vanishing radius N − d to ρ). If we assume (15) from Theorem 3.2 and
that the medoid index m∗ is the first element processed in trimed, then the number of computed elements in
d 1
1
A(x(m∗ ), N − 2d , ρ) is O(d α4 N 2 ).
Proof. According to Assumption 15, an element at radius r < ρ has surplus energy at least αr2 . This means
that, assuming that the medoid has already been computed, an element computed at radius r will be surrounded
by an exclusion zone of radius αr2 in which no element will subsequently be computed. We will use this fact
1
to upper bound the number of computed elements in A(x(m∗ ), N − 2d , ρ), firstly by bounding the number in an
annulus of inner radius r and width αr2 , that is the annulus Ad (x(m∗ ), r, r + αr2 ), and then summing over
1
concentric rings of this form which cover A(x(m∗ ), N − 2d , ρ). Recall that the number of computed elements in
Ad (x(m∗ ), r, r + αr2 ) is denoted by Nc (x(m∗ ), r, r + αr2 ).
James Newling, François Fleuret
We use Lemma SM-G.5 to bound Nc (x(m∗ ), r, r + αr2 ),
d 2
4
αr (r + αr2 )d−1
∗
2
2
Nc (x(m ), r, r + αr ) ≤ (d + 1) √
d
3
(αr2 )
d
d−1
4
1
≤ (d + 1)2 √
1+
αr
3
d
d−1
4
2
2
√
≤ (d + 1)
max 2,
αr
3
d−1 !!
d
2
4
d−1
2
max 2 ,
≤ (d + 1) √
αr
3
d
d−1 !
4
2
≤ (d + 1)2 √
2d−1 +
αr
3
d
d d−1
8
1
8
+ (d + 1)2 √
≤ (d + 1)2 √
αr
3
3
1
Let r0 = N − 2d and ri+1 = ri + αri2 , and let T be the smallest index i such that ri ≤ ρ. With this notation in
hand, we have
T
X
1
Nc (x(m∗ ), N − 2d , ρ) ≤
Nc (x(m∗ ), ri , αri + ri2 ).
i=0
The summation on the right-hand side can be upper-bounded by an integral. Using that the difference between
ri and ri+1 is αri2 , we need to divide terms in the sum by αri2 when converting to an integral. Doing this, we
obtain,
Z ρ+αρ2
1
Nc (x(m∗ ), N − 2d , ρ) ≤
Nc (x(m∗ ), r, αr2 )dr
1
N − 2d
d d Z ∞
8
1
≤ const + (d + 1) √
r−(1+d) dr
1
α
3
N − 2d
d
1
4
≤ const + (d + 1)
N 2.
α
d
This completes the proof, and provides the hidden constant of complexity as (d + 1) α4 . Thus larger values
for α should result in fewer computed elements in the annulus Ad (x(m∗ ), r, r + αr2 ), which makes sense given
that large values of α imply larger surplus energies and thus larger elimination zones.
2
Lemma SM-G.4 (Packing beyond ρ). If we assume (15) from Theorem 3.2 and that the medoid index m∗ is
the first element processed by trimed, then the number of computed elements in Ad (x(m∗ ), ρ, ∞) is less than
(1 + 4E ∗ /(αρ2 ))d .
Proof. Recall that we at assuming m∗ = 1, that is that the medoid is the first element processed in trimed. All
elements beyond radius 2E ∗ are eliminated by type 1 eliminations (Figure 1), which provides the first inequality
below. Then, as the excess energy is at least = αρ2 for all elements beyond radius ρ of x(m∗ ), we apply
Lemma SM-G.8 with = αρ2 /2 to obtain the second inequality below,
Nc (m(x), ρ, ∞) ≤ Nc (m(x), ρ, 2E ∗ )
(2E ∗ + 21 αρ2 )d
( 12 αρ2 )d
d
4E ∗
≤ 1+
.
αρ2
≤
A Sub-Quadratic Exact Medoid Algorithm
Lemma SM-G.5 (Annulus packing). For 0 ≤ r and 0 < ≤ w. If
X ⊂ Ad (0, r, r + w),
where
∀x ∈ X , Bd (x, ) ∪ X = {x},
(17)
then,
|X | ≤ (d + 1)
2
4
√
3
d
d−1
w (r + w)
d
.
Proof. The condition (17) implies,
∀x, x0 ∈ X × X , B x,
∪ B x0 ,
= ∅.
2
2
(18)
Using that ∈ (0, w] and Lemma SM-G.6, one can show that for all x ∈ A(0, r, r + w),
d2
hi
3
1
volume B x,
∩ A(0, r, r + w) >
Vd
2
d+1 4
2
(19)
Combining (18) with (19) we have,
[
volume
x∈X
√ !d
3
|X |d .
4
!
Vd [1]
B x,
∩ A(0, r, r + w) >
2
d+1
(20)
Letting Sd [] denote the surface area of a B(0, ), it is easy to see that
volume (A(0, r, r + w)) < Sd [1] w (r + w)
d−1
.
Combining (20) with (21) we get,
Vd [1]
d+1
√ !d
3
d−1
|X |d < Sd [1] w (r + w)
.
4
which combined with the fact that
Sd [1]
=
Vd [1]
dVd
dr
Vd
!
r=1
= d,
provides us with,
|X | ≤ (d + 1)
2
4
√
3
d
d−1
w (r + w)
d
.
Lemma SM-G.6 (Volume of ball intersection). For x0 , x1 ∈ Rd with kx0 − x1 k = 1,
volume (Bd (x0 , 1) ∩ Bd (x1 , 1))
1
≥
volume (Bd (x0 , 1))
d+1
d2
3
.
4
(21)
James Newling, François Fleuret
Proof. Let Vd [r] denote the volume of Bd (0, r). It is easy to see that,
Z
1
2
i
hp
x(2 − x) dx
0
"r #
Z 21
3
≥2
Vd−1
x dx
2
0
Z 21 d−1
2
3
dx
≥ 2Vd−1 [1]
x
2
0
d−1
d+1
3 2
2
1 2
≥ 2Vd−1 [1]
2
d+1
2
d−1
d−1
3 2
2
1 2
≥ Vd−1 [1]
2
d+1
2
d−1
2
3 2
.
≥ Vd−1 [1]
4
d+1
volume (Bd (x0 , 1) ∩ Bd (x1 , 1)) = 2
Using that
Vd−1
Vd−1 [1]
1
> √ , we divide the intersection volume through by Vd [1] to obtain,
Vd [1]
π
d−1
3 2
2
√
4
π(d + 1)
d2
1
3
≥
d+1 4
volume (Bd (x0 , 1) ∩ Bd (x1 , 1))
≥
volume (Bd (x0 , 1))
Lemma SM-G.7 (Packing balls in a ball). The number of non-intersecting balls of radius which can be packed
d
into a ball of radius r in Rd is less than r
Proof. The technique used here is a loose version of that used in proving Lemma SM-G.5. The volume of Bd (0, )
d
is a factor (r/) smaller than that of Bd (0, r). As the balls of radius are non-overlapping, the volume of their
union is simply the sum of their volumes. The result follow from the fact that the union of the balls of radius
is contained within the ball of radius r.
Lemma SM-G.8 (Packing points in a ball). Given X ⊂ Bd (0, r) such that no two elements of X lie within a
d
distance of of each other, |X | < 2r+
.
Proof. As no two elements lie within distance of each other, balls of radius /2 centred at elements are nonintersecting. As each of the balls of radius /2 centred at elements of X lies entirely within Bd (0, r + /2), we
can apply Lemma (SM-G.7), arriving at the result.
SM-H
Pseudocode for trikmeds
In Alg. (6) we present trikmeds. It is decomposed into algorithms for initialisation (7), updating medoids (8),
assigning data to clusters (9) and updating bounds on the trimed derived bounds (10). Table 4 summarised all
of the variables used in trikmeds.
When there are no distance bounds, the location of the bottleneck in terms of distance calculations depends on
N/K 2 . If N/K K, the bottleneck lies in updating medoids, which can be improved through the strategy
used in trimed. If N/K K, the bottleneck lies in assigning elements to clusters, which is effectively handled
through the approach of Elkan (2003).
A Sub-Quadratic Exact Medoid Algorithm
Algorithm 6 trikmeds
initialise()
while not converged do
update-medoids()
assign-to-clusters()
update-sum-bounds()
end while
Algorithm 7 initialise
// Initialise medoid indices, uniform random sample without replacement (or otherwise)
{m(1), . . . , m(K)} ← uniform-no-replacement({1, . . . , N })
for k = 1 : K do
// Initialise medoid and set cluster count to zero
c(k) ← x(m(k))
v(k) ← 0
// Set sum of in-cluster distances to medoid to zero
s(k) ← 0
end for
for i = 1 : N do
for k = 1 : K do
// Tightly initialise lower bounds on data-to-medoid distances
lc (i, k) ← kx(i) − c(k)k
end for
// Set assignments and distances to nearest (assigned) medoid
a(i) ← arg mink∈{1,...,K} lc (i, k)
d(i) ← lc (i, a(i))
// Update cluster count
v(a(i)) ← v(a(i)) + 1
// Update sum of distances to medoid
s(a(i)) ← s(a(i)) + d(i)
// Initialise lower bound on sum of in-cluster distances to x(i) to zero
ls (i) ← 0
end for
V (0) ← 0
for k = 1 : K do
// Set cumulative cluster count
V (k) ← V (k − 1) + v(k)
// Initialise lower bound on in-cluster sum of distances to be tight for medoids
ls (m(k)) ← s(k)
end for
// Make clusters contiguous
contiguate()
James Newling, François Fleuret
Algorithm 8 update-medoids
for k = 1 : K do
for i = V (k − 1) : V (k) − 1 do
// If the bound test cannot exclude i as m(k)
if ls (i) < s(k) then
// Make ls (i) tight by computing and cumulating all in-cluster distances to x(i),
ls (i) ← 0
for i0 = V (k − 1) : V (k) − 1 do
˜ 0 ) ← kx(i) − x(i0 )k
d(i
˜ 0)
ls (i) ← ls (i) + d(i
end for
// Re-perform the test for i as candidate for m(k), now with exact sums. If i is the new best candidate,
update some cluster information
if ls (i) < s(k) then
s(k) ← ls (i)
m(k) ← i
for i0 = V (k − 1) : V (k) − 1 do
d(i0 ) ← kx(i) − x(i0 )k
end for
end if
// Use computed distances to i to improve lower bounds on sums for all samples in cluster k (see Figure
X)
for i0 = V (k − 1) : V (k) − 1 do
˜ 0 )v(k) − ls (i)|)
ls (i0 ) ← max (ls (i0 ), |d(i
end for
end if
end for
// If the medoid of cluster k has changed, update cluster information
if m(k) 6= V (k − 1) then
p(k) ← kc(k) − x(m(k))k
c(k) ← x(m(k))
end if
end for
A Sub-Quadratic Exact Medoid Algorithm
Algorithm 9 assign-to-clusters
// Reset variables monitoring cluster fluxes,
for k = 1 : K do
// the number of arrivals to cluster k,
∆n−in (k) ← 0
// the number of departures from cluster k,
∆n−out (k) ← 0
// the sum of distances to medoid k of samples which leave cluster k
∆s−out (k) ← 0
// the sum of distances to medoid k of samples which arrive in cluster k
∆s−in (k) ← 0
end for
for i = 1 : N do
// Update lower bounds on distances to medoids based on distances moved by medoids
for k = 1 : K do
l(i, k) = l(i, k) − p(k)
end for
// Use the exact distance of current assignment to keep bound tight (might save future calcs)
l(i, a(i)) = d(i)
// Record current assignment and distance
aold = a(i)
dold = d(i)
// Determine nearest medoid, using bounds to eliminate distance calculations
for k = 1 : K do
if l(i, k) < d(i) then
l(i, k) ← kx(i) − c(k)k
if l(i, k) < d(i) then
a(i) = k
d(i) = l(i, k)
end if
end if
end for
// If the assignment has changed, update statistics
if aold 6= a(i) then
v(aold ) = v(aold ) − 1
v(a(i)) = v(a(i)) + 1
ls (i) = 0
∆n−in (a(i)) = ∆n−in (a(i)) + 1
∆n−out (aold ) = ∆n−out (aold ) + 1
∆s−in (a(i)) = ∆s−in (a(i)) + d(i)
∆s−out (aold ) = ∆s−out (aold ) + dold
end if
end for
// Update cumulative cluster counts
for k = 1 : K do
V (k) ← V (k − 1) + v(k)
end for
contiguate()
James Newling, François Fleuret
Table 4: Table Of Notation For trikmeds
N
i
x(i)
K
k
m(k)
c(k)
n1 (i)
a(i)
d(i)
v(k)
V (k)
lc (i, k)
ls (i)
p(k)
s(k)
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
:
number of training samples
index of a sample, i ∈ {1, . . . , N }
sample i
number of clusters
index of a cluster, k ∈ {1, . . . , K}
index of current medoid of cluster k, m(k) ∈ {1, . . . , N }
current medoid of cluster k, that is c(k) = x(m(k))
cluster index of centroid nearest to x(i)
cluster to which x(i) is currently assigned
distance from x(i) to c(a(i))
number of samples assigned to cluster k
number of samples assigned to a cluster of index less than k + 1
lowerbound on P
distance from x(i) to m(k)
lowerbound on i0 :a(i0 )=a(i) kx(i0 ) − x(i)k
distance moved (teleported) by m(k) in last update
sum of distances of samples in cluster k to medoid k
Algorithm 10 update-sum-bounds
for k = 1 : K do
// Obtain absolute and net fluxes of energy and count, for cluster k
Jsabs (k) = ∆s−in (k) + ∆s−out (k)
Jsnet (k) = ∆s−in (k) − ∆s−out (k)
Jnabs (k) = ∆n−in (k) + ∆n−out (k)
Jnnet (k) = ∆n−in (k) − ∆n−out (k)
for i = V (k − 1) : V (k) − 1 do
// Update the lower bound on the sum of distances
ls (i) ← ls (i) − min(Jsabs (k) − Jnnet (k)d(i), Jnabs (k)d(i) − Jsnet (k))
end for
end for
SM-I
Datasets
• Birch1, Birch2 : Synthetic 2-D datasets available from https://cs.joensuu.fi/sipu/datasets/
• Europe : Border map of Europe available from https://cs.joensuu.fi/sipu/datasets/
• U-Sensor Net : Undirected 2-D graph data. Points drawn uniformly from
√ unit square, with an undirected
edge connecting points when the distance between them is less than 1.25 N
• D-Sensor Net : Directed 2-D graph data. Points drawn uniformly√from unit square, with directed edge
connecting points when the distance between them is less than 1.45 N , direction chosen at random.
• Europe rail : The European rail network, the shapefile is available at http://www.mapcruzin.com/
free-europe-arcgis-maps-shapefiles.htm. We extracted edges from the shapefile using networkx available at https://networkx.github.io/.
Algorithm 11 contiguate
// This function performs an in place rearrangement over of variables a, d, l, x and m
// The permutation applied to a, d, l and x has as result a sorting by cluster,
// a(i) = k if i ∈ {V (k − 1), V (k)} for k ∈ {1, . . . , K}
// and moreover that the first element of each cluster is the medoid,
// m(k) = V (k − 1) for k ∈ {1, . . . , K}
A Sub-Quadratic Exact Medoid Algorithm
• Pennsylvania road The road network of Pennsylvania, the edge list is available directly from https://snap.
stanford.edu/data/
• Gnutella Peer-to-peer network data, available from https://snap.stanford.edu/data/
• MNIST (0) The ‘0’s in the MNIST training dataset.
• Conflong The conflongdemo data is available from https://cs.joensuu.fi/sipu/datasets/
• Colormo The colormoments data is available at http://archive.ics.uci.edu/ml/datasets/Corel+
Image+Features
• MNIST50 The MNIST dataset, projected into 50-dimensions using a random projection matrix where each
of the 784 × 50 elements in the matrix is i.i.d. N (0, 1).
• S1, S2, S3, S4, A1, A2, A3 All of these synthetic datasets are available from https://cs.joensuu.fi/
sipu/datasets/.
• thyroid, yeast, wine, breast, spiral All of these real world datasets are available from https://cs.joensuu.
fi/sipu/datasets/.
SM-J
Scaling with dimension of TOPRANK and TOPRANK2
Recall the assumption (3) made for the TOPRANK and TOPRANK2 algorithms. The assumption states that as one
approaches the minimum energy E ∗ from above, the density of elements decreases. In other words, the lowest
energy elements stand out from the rest and are not bunched up with very similar energies.
Consider the case where elements are points in Rd . Suppose that the density fX of points around the medoid
is bounded by 0 < ρ0 ≤ fX ≤ ρ1 , and that the energy grows quadratically in radius about the medoid. Then,
as the number of points at radius is O(d−1 ), the density (by energy) of points at radius is O(d−2 ). Thus
for d = 1 the assumption for TOPRANK and TOPRANK does not hold, which results in poor performance for d = 1.
For d = 2, the assumption holds, as the density (by energy) of points is constant. For d ≥ 2, as d increases the
energy distribution becomes more and more favourable for TOPRANK and TOPRANK2, as the low ranking elements
become more and more distinct with low energies becoming less probable. This explains the observation that
TOPRANK scales well with dimension in Figure 3.
SM-K
Example where geometric median is a poor approximation of medoid
There is no guarantee that the geometric median is close to the set medoid. Moreover, the element in S
which is nearest to g(S) is not necessarily the medoid, as illustrated in the following example. Suppose S =
{x(1), . . . , x(20)} ⊂ R2 , with x(i) = (0, 1) for i ∈ {1, . . . , 9}, x(i) = (0, −1) for i ∈ {10, . . . , 18}, x(19) = (1/2, 0)
and x(20) = (−1/2, 0). The geometric
median is (0, 0) and the nearest points to the geometric
√
√ median, x(19)
and x(20) have energy 1 + 18 3/2 ≈ 16.6. However, points {x(1), . . . , x(18)} have energy 2 3/2 + 9 = 10.7.
Thus by choosing a point in S which is nearest to the geometric median, one is choosing the element with the
highest energy, the opposite of the medoid.
Note the above example appears to violate the assumptions required for O(N 3/2 ) convergence of trimed, as it
requires that the probability density function vanishes at the distribution median. Indeed, in Rd it is the case
that if the O(N 3/2 ) assumptions are satisfied, the set medoid converges to the geometric median, and so the
geomteric median is a good approximation. We stress however that the geometric median is only relevant in
vector spaces.
SM-L
Miscellaneous
Figure 7 illustrates the idea behind algorithm trimed, comments in the caption.
James Newling, François Fleuret
1N=1
N=4
N = 16
N = 64
N = 256
0
medoid
computed
excluded
-1
0
50
100
N
150
200
250
Figure 7: Eliminating samples as potential medoids using only type 1 elimination, where we assume that the
medoid and its energy E ∗ are known, and so the radius of the exclusion ball of an element x is E(x) − E ∗ .
Uniformly sampling from [−1, 1] × [−1, 1], energies are computed only if the sample drawn does not lie in the
exclusion zone (union of balls). If the energy at x is computed, the exclusion zone is augmented by adding
Bd (x, E(x) − E ∗ ). Top left to right: the distribution of samples which are computed and excluded. Bottom: the
1
times at which samples are computed. We prove that probability of computation at time n is O(n− 2 ).
| 8cs.DS
|
arXiv:1709.04596v1 [stat.ML] 14 Sep 2017
A Framework for Generalizing Graph-based Representation Learning Methods
Nesreen K. Ahmed
Ryan A. Rossi
Rong Zhou
John Boaz Lee
Intel Labs
PARC
Google
WPI
Xiangnan Kong
Theodore L. Willke
Hoda Eldardiry
WPI
Intel Labs
PARC
Abstract
Random walks are at the heart of many existing deep learning algorithms for graph data. However, such algorithms have
many limitations that arise from the use of random walks,
e.g., the features resulting from these methods are unable to
transfer to new nodes and graphs as they are tied to node
identity. In this work, we introduce the notion of attributed
random walks which serves as a basis for generalizing existing methods such as DeepWalk, node2vec, and many others that leverage random walks. Our proposed framework enables these methods to be more widely applicable for both
transductive and inductive learning as well as for use on
graphs with attributes (if available). This is achieved by learning functions that generalize to new nodes and graphs. We
show that our proposed framework is effective with an average AUC improvement of 16.1% while requiring on average
853 times less space than existing methods on a variety of
graphs from several domains.
Keywords random walk · representation learning · inductive
learning · deep learning · attributed graphs
1
Introduction
Graphs (networks) are ubiquitous [5, 9] and allow us to
model entities (nodes) and the dependencies (edges) between them. Graph data is often observed directly in the
natural world (e.g., biological or social networks) [53] or
constructed from non-relational data by deriving a metric
space between entities and retaining only the most significant edges [21, 47, 58]. Learning a useful feature representation from graph data lies at the heart and success of many
machine learning tasks such as node classification [33, 35],
anomaly detection [5], link prediction [6], dynamic network
analysis [37], community detection [36, 44], role discovery [46], visualization and sensemaking [1, 41], graph classification [25, 29], and network alignment [26].
Many existing techniques use random walks as a basis
for learning features or estimating the parameters of a graph
model for a downstream prediction task. Examples include
recent node embedding methods such as DeepWalk [40],
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
node2vec [19], as well as graph-based deep learning algorithms. However, the simple random walk used by these
methods is fundamentally tied to the identity of the node.
This has three main disadvantages. First, these approaches
are inherently transductive and do not generalize to unseen
nodes and other graphs. Furthermore, they are unable to be
used for graph-based transfer learning tasks such as acrossnetwork classification [15, 27], graph similarity [17, 57], and
matching [39, 43]. Second, they are not space-efficient as a
feature vector is learned for each node which is impractical
for large graphs. Third, most of these approaches lack support for attributed graphs. This includes graphs with intrinsic (node) attributes such as age or gender as well as structural features such as higher-order subgraph counts, e.g.,
number of 4-cliques each node participates.
To make these methods more generally applicable, we
introduce a reinterpretation of the notion of random walk
that is not tied to node identity and is instead based on a
function Φ : x → w that maps a node attribute vector to
a type. Using this reinterpretation as a basis we propose a
framework that naturally generalizes many existing methods. This framework provides a number of important advantages. First, the learned features generalize to new nodes and
across graphs and therefore can be used for transfer learning tasks such as across-network link prediction and classification. Our proposed approach supports both transductive
and inductive learning. Second, the generalized approach is
inherently space-efficient since embeddings are learned for
types (as opposed to nodes) and therefore requires significantly less space than existing methods. Third, the generalized approach supports learning from attributed graphs.
However, we stress that the approach does not require graphs
with input attributes since these can be derived from the
graph structure. Furthermore, our approach is shown to be
effective with an average improvement of 16.1% in AUC
while requiring on average 853x less space than existing
methods on a variety of graphs.
Contributions: This paper proposes a framework for generalizing many existing algorithms making them more broadly
applicable for attributed graphs and inductive learning. It has
the following key properties:
• Space-efficient: It requires on average 853x less space
than a number of existing methods.
• Accurate: It is accurate with an average improvement of
16.1% across a variety of graphs from several domains.
• Inductive: It is an inductive learning approach that is able
to learn embeddings for new nodes and graphs.
• Attributed: It naturally supports graphs with attributes (if
available) and serves as a foundation for generalizing existing methods for use on attributed graphs.
2
W = Φ(x1 ) ∪ Φ(x2 ) ∪ · · · ∪ Φ(xN )
Framework
In this section, we formally introduce the attributed random
walk framework which can serve as a basis for generalizing
many existing deep graph models and embedding methods.
The framework consists of two general components:
C1. (Function Mapping Nodes to Types): A function Φ
that maps nodes to types based on a N × K matrix X
of attributes. The function Φ can be defined or learned
automatically. Note X may be given as input or computed
based on the structure of the graph. For more details, see
Section 2.1.
C2. (Attributed Random Walks): Generate attributed
random walks based on the function Φ mapping nodes
to types (Section 2.2). Informally, an attributed walk is
simply the node types that occur during a walk.
A summary of the notation is provided in Table 1. To avoid
abuse of notation, this paper presents the framework for
nodes. However, the approach is applicable to both nodes
and edges.
2.1
Function Mapping Nodes to Types
Given an (un)directed graph G = (V, E) and an N × K matrix X of attributes (if available), the first component of the
framework maps the N nodes to a set W = {w1 , ..., wM } of
M types where 1 ≤ M ≤ N and M is often much smaller
than N , i.e., M N . More formally,
Φ : x → w
(1)
Table 1: Summary of notation.
G
N
M
K
D
R
L
Γi
S
Φ(·)
W
X
xi
x̄j
α
Z
zk
where Φ is a function mapping nodes to types based on the
N ×K attribute matrix X. The function Φ can be learned automatically or defined manually by the user. The framework
is general and flexible for use with any arbitrary function Φ
that maps nodes to types based on an N × K attribute matrix X. For graphs that are not attributed, we simply derive
X by extracting a set of graph features based on the structure of the graph (e.g., X may represent the graphlet features
shown in Figure 2).1 The set of types W is defined as:
(un)directed (attributed) graph
number of nodes |V |
number of unique types
number of attributes used for deriving types
number of dimensions to use for embedding
number of walks for each node
length of a random walk
neighbors (adjacent nodes) of vi
set of (attributed) random walks
a function Φ : x → w that maps an attribute vector x ∈ RK to a
corresponding type
set of types where w ∈ W is a type assigned to one or more nodes
an N × K attribute matrix where the rows represent nodes and the
columns represent attributes
a K-dimensional attribute vector for node vi
a N -dimensional vector for attribute j (column of X)
transformation hyperparameter
an M × D matrix of type embeddings
a D-dimensional embedding for type wk
(2)
where Φ is a function that maps the set of N = |V | nodes
to M = |W | types such that 1 ≤ M ≤ N . In node2vec
and other embedding and deep graph algorithms, notice that
each node is mapped to a unique identifier that identifies the
specific node which is then used in the random walk. However, in the attributed random walk framework two or more
nodes may map to the same type.
There are two general classes of functions for mapping
nodes to types. The first class of functions are simple functions taking the form:
Φ(x) = x1 ◦ x2 ◦ · · · ◦ xK
where x is an attribute vector:
x = x1 x2 · · · xK
(3)
(4)
and ◦ is a binary operator such as concatenation, sum,
among others. The second class of functions are learned by
solving an objective function. This includes functions based
on a low-rank factorization of the N × K matrix X having
the form X ≈ f hUVT i with factor matrices U ∈ RN ×F
and V ∈ RK×F where F > 0 is the rank and f is a linear
or non-linear function. More formally,
h
i
arg min D X, f hUVT i + R(U, V)
(5)
U,V∈C
where D is the loss, C is constraints (e.g., non-negativity
constraints U ≥ 0, V ≥ 0), and R(U, V) is a regularization penalty. Let Vk denote the set of nodes mapped to type
wk ∈ W . We partition U ∈ RN ×F into M disjoint sets
of nodes (for each of the M types) V1 , . . . , VM such that
V = V1 ∪ . . . ∪ VM by solving the k-means objective:
P
M X
X
ui ∈Vj ui
2
min
kui − cj k , where cj =
(6)
|Vj |
{Vj }M
j=1 j=1
ui ∈Vj
Alternatively, the nodes can be mapped to types directly by
setting F = M and using U as follows:
Φ(xi ) = arg max Uik ,
∀ i = 1, . . . , N
(7)
wk ∈W
For other possibilities, we refer the readers to a recent survey
by Xu et al. [54]. Notice that the approaches in Eq. (5)-(7)
can be used for inductive learning as follows: Given a new
1
Graphlets (and orbits) can be derived using exact algorithms [4] or accurately estimated using fast parallel approximation
methods with provable error bounds [3, 49].
Step 1
Step 2
Compute and/or
select attributes
Transform each
attribute
0 3
0 2
v1
v2
v4
v3
0 0
v5
1 3
0 3
v6
1 6
1 4
x1 = number of triangles
x2 = number of 2-stars
0 0
v1
v5
1 0
v2
v4
v3
0 0
v6
1 1
1 1
(e.g., log-binning with bin
size of 0.5)
Step 3
Derive types
from attributes
& assign 0 0
0 0
w1
w2
w3
1 1
Derive set of
attributed
random walks
w1
1 0
w3
Step 4
0 0
w1
1 1
w1
w1
w2
w3
w1
w3 e.g., length = 3
w1 = 0 0
w2 = 1 0
w3 = 1 1
w3 w2 w3
w3 w3 w1
Figure 1: A simple illustrative toy example. This toy example shows only one potential instantiation of the general framework.
In particular, Step 1 computes the number of triangles and 2-stars that each node participates whereas Step 2 transforms each
individual attribute using logarithmic binning. Next, Step 3 derives types using a simple function Φ representing a concatenation
of each node’s attribute values resulting in M = 3 types {w1 , w2 , w3 } among the N = 6 nodes. Finally, Step 4 derives a set of
attributed random walks (type sequences) which are then used to learn embeddings. See Section 2.3 for further discussion.
graph G0 = (V 0 , E 0 ) (or node vN +1 ), we compute the attribute matrix X0 (or attribute vector xN +1 ) using the same
K attributes and then use X0 and V from Eq. (5) to estimate
U0 directly. More formally, given X0 and V, we find U0 by
solving:
h
i
argmin D X0 , f hU0 VT i + R(U0 )
(8)
U0 ∈C
Observe that V gives the mapping from the latent feature
space to the input attribute space and represents how each
of the K attributes map to the latent features and therefore V is not tied to particular nodes but more generally
to the (structural) attributes. Many of these methods avoid
the issue of selecting an appropriate subset of attributes to
use apriori and are more robust to the selection of different sets of attributes compared to simple functions such as
Φ(x) = x1 ◦ x2 ◦ · · · ◦ xK . As an aside, it is straightforward to formulate a function learning problem to automate
the choice of function Φ.
The framework naturally supports both attributed and
non-attributed graphs as the attribute matrix X can be derived by extracting a set of graph features based on the structure of the graph (e.g., X may represent the graphlet features
shown in Figure 2) or learned automatically using an inductive feature learning approach such as DeepGL [48].
2.2
Attributed Random Walks
Most existing deep learning models and embedding algorithms for graphs use simple random walks based on node
identity [19, 40]. The features learned from these methods
are fundamentally tied to the identity of a node and therefore are unable to generalize to new nodes as well as for
graph-based transfer learning tasks or across-network learning, among other limitations. In this section, we introduce
the notion of attributed random walks and demonstrate how
they serve as a basis for generalizing many existing deep
graph models and embedding algorithms.
Previous work uses random walks based on the following
traditional definition of a walk:
Definition 1 (Walk) A walk S of length L is defined as a
sequence of indices i1 , i2 , . . . , iL+1 such that (vit , vit+1 ) ∈
E for all 1 ≤ t ≤ L.
Note that nodes in a walk S are not necessarily distinct. We
replace the traditional notion of walk (Definition 1) with the
more appropriate and useful notion of attributed walk defined as:
Definition 2 (Attributed walk) Let xi be a K-dimensional
vector for node vi . An attributed walk S of length L is defined as a sequence of adjacent node types
Φ(xi1 ), Φ(xi2 ), . . . , Φ(xiL+1 )
(9)
associated with a sequence of indices i1 , i2 , . . . , iL+1 such
that (vit , vit+1 ) ∈ E for all 1 ≤ t ≤ L and Φ : x → w
is a function that maps the input vector x of a node to a
corresponding type Φ(x).
The type sequence Φ(xi1 ), Φ(xi2 ), . . . , Φ(xiL+1 ) is simply the node types that occur during a walk. The framework allows both uniform and non-uniform attributed random walks. It is also straightforward to bias the attributed
random walk in an arbitrary fashion. The notion of attributed
walk can be extended in various ways (e.g., for edge types).
Given an arbitrary embedding method or deep graph
model A that uses simple random walks based on node identity (e.g., node2vec, DeepWalk, among others), we generalize A by replacing the traditional notion of random walk
with the proposed notion of attributed random walks. We
will show that existing methods are actually a special case
of the proposed framework when Φ(xi ) uniquely identifies
node vi . Suppose the graph G has N nodes mapped to M
types using an arbitrary function Φ (Section 2.1). If we select
Φ such that M → N then we recover the original method
A as a special case of the framework. More intuitively, each
node vi ∈ V must be assigned to a unique type Φ(xi ) = wi
that uniquely identifies it from any other node vj ∈ V since
|W | = |V | where W = Φ(x1 ) ∪ · · · ∪ Φ(xN ). Observe that
M → N is actually an extreme case of the framework and
corresponds exactly to the original method A.
There are three key advantages that arise when existing
methods are generalized by the proposed framework. First,
the generalized method naturally supports attributed graphs.
Second, the features learned generalize to new nodes and
across graphs and therefore are naturally inductive and able
to be used for transfer learning tasks. Finally, the learned
embeddings are significantly more space-efficient with a
space complexity of only O(M D) compared to previous
methods that require O(N D) space where M N (M is
much smaller than N ) and D is the embedding dimensionality.
2.3
Illustrative Example
We emphasize that the proposed framework is general and
flexible; and the example discussed in this section represents only one such simple instantiation of the framework.
An overview of the key steps involved in this specific instantiation are shown in Figure 1 and summarized succinctly
below as follows:
Step 1. Compute and/or Select Attributes: Given an
(un)directed graph G, the first step is to compute a set
of attributes using the graph structure (e.g., graphlets).
Step 2. Transform Attributes: Next, we transform each
attribute vector. The goal is to map the values of each
individual attribute vector x̄j to a smaller set of values
(via quantization or discretization algorithm). In Figure 1,
each attribute x̄j is transformed using logarithmic binning.2 For convenience, each initial x̄j is replaced by the
transformed attribute.
Step 3. Derive Types from Attributes & Assign: Now,
the set of types are derived as W = Φ(x1 ) ∪ Φ(x2 ) ∪
· · · ∪ Φ(xN ) where |W | ≤ N and Φ : x → w such that
Φ(x) = x1 ◦ x2 ◦ · · · ◦ xK and ◦ is a binary operator
such as concatenation. We also assign each node to its
corresponding type. See Figure 1 for intuition.
Step 4. Attributed Random Walks: Finally, we generate
a set of attributed random walks using the node types (as
opposed to the unique node identifiers used traditionally).
The attribute random walks represent sequences of node
types which are then used to learn embeddings. For instance, the set of attributed random walks are used in Section 3 to learn the M × D type embedding matrix Z.
From the perspective of the general framework, Steps 1-3 in
Figure 1 correspond to the first main component that maps
nodes to types via the function Φ while Step 4 corresponds
to the last component that generates attributed random walks
representing sequences of node types.
3
Node2Vec Generalization
We use the proposed framework in Section 2 to generalize node2vec [19].3 This gives rise to a new generalized
2
Logarithmic binning assigns the first αN nodes with smallest
attribute value to 0 (where 0 < α < 1), then assigns the α fraction
of remaining unassigned nodes with smallest value to 1, and so on.
3
However, the framework could have been used to generalize
any node embedding or deep graph model that leverages traditional
random walks.
Algorithm 1 Generalized Node2Vec
1
procedure G ENERALIZED N ODE 2V EC( G = (V, E) and possibly attributes X, embedding dimensions D, walks per node R, walk length
L, context (window) size ω, return p, and in-out q, function Φ )
2
Initialize set of attributed walks S to ∅
Extract (graphlet) features if needed and append to X
Transform each attribute in X (e.g., using logarithmic binning)
Precompute transition probabilities π using G, p, and q
G0 = (V, E, π)
parallel for j = 1, 2, ..., R do
. walks per node
Set Π to be a random permutation of the nodes V
for each v ∈ Π in order do
S = ATTRIBUTEDWALK(G0 , X, v, Φ, L)
Add the attributed walk S to S
end for
end parallel
Z = S TOCHASTIC G RADIENT D ESCENT(ω, D, S) . parallel
return the learned type embeddings Z
3
4
5
6
7
8
9
10
11
12
13
14
15
Algorithm 2 Attributed Random Walk
1
2
3
4
5
6
7
8
9
10
procedure ATTRIBUTEDWALK(G0 , X, start node s, function Φ, L)
Initialize attributed walk S to Φ(xs )
Set i = s
. current node
for ` = 1 to L do
Γi = Set of the neighbors for node i
j = A LIAS S AMPLE(Γi , π)
. select node j ∈ Γi
Append Φ(xj ) to S
Set i to be the current node j
end for
return attributed walk S of length L rooted at node s
node2vec algorithm with the following advantages: (a) naturally supports attributed graphs, (b) learns features that generalize for new nodes as well as for extraction on another
arbitrary graph and thus able to be used for transfer learning tasks, and (c) space-efficient by learning an embedding
for each type as opposed to each node. The generalized
node2vec algorithm is shown in Alg. 1. In particular, we replace the notion of random walk in node2vec with the more
appropriate notion of attributed random walk (Line 10-11).
Nodes are mapped to types using an arbitrary function Φ(x)
which can be defined or learned; see Section 2.1 for further
details. As an aside, Alg. 1 assumes the function Φ is defined
or learned apriori. Note that if no attribute matrix X is given
as input then we derive features based on the graph structure (Line 3). However, even if (intrinsic/self- or structural)
attributes are given as input we may also choose to derive additional structural features such as graphlets [2] and append
these to X.
The attributed random walk routine for the generalized
node2vec algorithm is shown in Alg. 2. Given a start node
s, we first obtain the type of the start node s given by the
function Φ(xs ) and initialize an attributed walk S (stored as
a list4 ) to be S = Φ(xs ) (Line 2). Next, Line 3 sets the
current node i to be the start node s. The attributed walk S
of length L rooted at the start node s is obtained in Lines 49. In particular, Line 5 gets the neighbors of the current node
i which is used in Line 6 to sample a new node j ∈ Γi . The
type Φ(xj ) of node j is appended to S (e.g., if S is stored
using an ordered list then we append Φ(xj ) to the back of S
to preserve order). Afterwards, Line 8 sets the current node
i to be j and Lines 4-9 are repeated for L steps. Finally,
an attributed random walk S of length L starting at node
s is returned in Line 10. One can also leverage Alg. 2 (or
an adapted variant of it) for generalizing other deep graph
models and node embedding methods. It is straightforward
to see that Alg. 1 is a generalization of node2vec since if
we allow M → N then we recover the original node2vec
algorithm as a special case of the framework. Furthermore,
DeepWalk [40] is also a special case where M → N , p = 1,
and q = 1.
Implementation details: In our implementation, we evaluate Φ(xi ) = x1 ◦ x2 ◦ · · · ◦ xK only once for each
node i ∈ V and store each Φ(xi ) in a hash table giving
o(1) time lookup. This allows us to derive Φ(xs ) and Φ(xj )
in Line 2 and 7 of Alg. 2 in o(1) constant time. To construct the hash table it takes only O(N K) time to obtain
Φ(xi ), ∀i = 1, ..., N given X and only O(N ) space to store
them efficiently using a hash table.
4
handles other categories of attributes including structural attributes, intrinsic/self-attributes (such as age or other nonrelational attributes), and relational features derived using
the graph G along with existing structural or self-attributes.
We evaluate the generalized node2vec approach presented in
Section 3 that leverages the attributed random walk framework (Section 2) against a number of popular methods including: node2vec6 , DeepWalk [40], and LINE [52]. For
our approach and node2vec, we use the same hyperparameters (D = 128, R = 10, L = 80) and grid search over
p, q ∈ {0.25, 0.50, 1, 2, 4} as mentioned in [19]. We use
logistic regression (LR) with an L2 penalty. The model is
selected using 10-fold cross-validation on 10% of the labeled data. Experiments are repeated for 10 random seed
initializations. All results are statistically significant with pvalue < 0.01. Unless otherwise mentioned, we use AUC to
evaluate the models. Data has been made available at NetworkRepository [45].7
Experiments
In this section, we investigate the effectiveness of the proposed framework using a variety of graphs from several domains. In these experiments, we use the node2vec generalization given in Section 3.
2
1
4
3
G2
G1
5
6
7
9
G3
10
8
12
G4
G5
15
13
11
G6
G7
14
G8
G9
Figure 2: Summary of the 9 node graphlets and 15 orbits
(graphlet automorphisms) with 2-4 nodes.
(a) 27 types (Φ only)
Figure 3: Using only the node types are useful for revealing interesting insights into the graph data including nodes
that are structurally similar. Node types alone reveal nodes
that are structurally similar (karate network). Node color encodes type. Only the number of triangles and wedges are
used by Φ. In (a) nodes with identical types are grouped;
(b) uses log binning (LB) with α = 0.5. See text for further
discussion.
4.2
4.1
Experimental Setup
Unless otherwise mentioned, all experiments use logarithmic binning (defined in Section 2.3) with α = 0.5. In these
experiments, we use a simple function Φ(x) that represents
a concatenation of the attribute values in the node attribute
vector x. In particular, Φ(x) is defined as Φ(x) = x1 ◦ · · · ◦
xK where ◦ is the concatenation operator. For these experiments, we searched over 10 subsets of the 9 graphlet attributes shown in Figure 2.5 However, the approach trivially
4
A list or other efficient data structure may be used for storing
a sequence of items (nodes/edges) with potential duplicates
5
Note we also investigated the 15 graphlet orbits (as opposed to
the 9 graphlet attributes) and found similar results.
(b) 7 types (LB, Φ)
Node Mapping Experiments
This section investigates the intermediate representation defined by an arbitrary function Φ that maps nodes to a set of
types. Recall that for simplicity Φ is defined as a concatenation of the attribute values for a given node and the attributes correspond to the graphlet counts in Figure 2. Strikingly, we observe in Figure 3 that even these simple mappings alone are useful and effective for understanding the
important structural and behavioral characteristics of nodes,
that is before even generating attributed random walks and
learning an embedding based on them. In particular, nodes
assigned to the same type in Figure 3(a) obey a strict notion
6
7
https://github.com/aditya-grover/node2vec
http://networkrepository.com/
of node equivalence on a feature representation:
∀j, 1 ≤ j ≤ K : fj (u) = fj (v) ⇒ u ≡ v
110
(10)
web-EPA
fb-nips-ego
web-BerkStan
fb-Yale4
soc-dolphins
inf-USAir97
soc-anybeat
ia-radoslaw-email
eco-everglades
eco-fweb-baydry
ca-CSphd
bn-rat-cerebral
bn-cat
Ours vs. node2vec
Ours vs. DeepWalk
Ours vs. LINE
bn-rat-brain
50
45
40
35
30
25
20
15
10
5
0
-5
biogrid-human
% Gain/Loss in AUC
80
70
60
50
40
30
20
10
Figure 4: AUC gain over other methods for link prediction
bootstrapped using (zi + zj )/2. The average gain over all
methods and across all graphs is 10.5%.
web-EPA
fb-nips-ego
fb-Yale4
soc-dolphins
inf-USAir97
soc-anybeat
ia-radoslaw-email
eco-everglades
eco-fweb-baydry
ca-CSphd
bn-rat-cerebral
bn-cat
bn-rat-brain
0
Comparison
This section compares the proposed approach to other embedding methods for link prediction. Given a partially observed graph G with a fraction of missing edges, the link
prediction task is to predict these missing edges. We generate a labeled dataset of edges as done in [19]. Positive examples are obtained by removing 50% of edges randomly,
whereas negative examples are generated by randomly sampling an equal number of node pairs that are not connected
with an edge, i.e., each node pair (i, j) 6∈ E. For each
method, we learn features using the remaining graph that
consists of only positive examples. Using the feature representations from each method, we then learn a model to
predict whether a given edge in the test set exists in E or
not. Notice that node embedding methods such as DeepWalk
and node2vec require that each node in G appear in at least
one edge in the training graph (i.e., the graph remains connected), otherwise these methods are unable to derive features for such nodes. This is a significant limitation that prohibits their use in many real-world applications.
For comparison, we use the same set of binary operators [19] to construct features for the edges indirectly using
the learned node representations. Moreover, the AUC scores
Ours vs. node2vec
Ours vs. DeepWalk
Ours vs. LINE
90
biogrid-human
4.3
% Gain/Loss in AUC
where fj is a feature. Eq. 10 is strict since two nodes belong to the same type iff they have identical feature vectors. However, Figure 3(b) captures a more relaxed notion of
structural equivalance called structural similarity [46]. This
result is surprising since no embedding has been learned
yet, only the mapping into the intermediate representation
(types). Hence, it validates the intermediate representation
of mapping nodes to types while also demonstrating the effectiveness of it for grouping structurally similar nodes.
100
Figure 5: AUC gain of our approach over the other methods
for link prediction bootstrapped using Hadamard. The average gain over all methods and across all graphs is 21.7%.
from our method are all significantly better than the other
methods at p < 0.01. We summarize the gain/loss in predictive performance over the other methods in Figure 4 and 5. In
all cases, our method achieves better predictive performance
over the other methods across a wide variety of graphs with
different characteristics. Overall, the mean and product binary operators give the best results with an average gain
in predictive performance (over all graphs) of 10.5% and
21.7%, respectively.
4.4
Space-efficient Embeddings
We now investigate the space-efficiency of the learned embeddings from the proposed framework and intermediate
representation. Observe that any embedding method that implements the proposed attributed random walk framework
(and intermediate representation) learns an embedding for
each distinct node type w ∈ W . In the worst case, an embedding is learned for each of the N nodes in the graph and
we recover the original method as a special case. In general,
the best embedding most often lies between such extremes
and therefore the embedding learned from a method implementing the framework is often orders of magnitude smaller
in size since M N where M = |W | and N = |V |.
Given an attribute vector x of graphlet counts (Figure 2)
for an arbitrary node in G, we derive embeddings using each
of the following:
Φ(xi
Φ(xi
Φ(xi
Φ(xi
= [ x2
= [ x2
= [ x2
= [ x1
x3 ]),
x3 x4
x3 x4
x2 x3
for i = 1, ..., N
x6 x9 ]), for i = 1, ..., N
x5 · · · x9 ]), for i = 1, ..., N
x4 x5 · · · x9 ]), for i = 1, ..., N
(11)
(12)
(13)
(14)
where Φ(·) is a function that maps xi to a type w ∈ W . In
these experiments, we use logarithmic binning (applied to
each N -dimensional graphlet feature) with α = 0.5 and use
Φ defined as the concatenation of the logarithmically binned
attribute values. Embeddings are learned using the different
4.5
4
3.5
3
2.5
2
1.5
1
0.5
-0.5
biogrid-human
bn-cat
bn-fly-medulla
bn-rat-brain
bn-rat-cerebral
ca-CSphd
ca-Erdos992
ca-GrQc
ca-HepPh
ca-citeseer
ca-dblp-2010
eco-everglades
eco-florida
eco-fweb-baydry
eco-mangwet
ia-radoslaw-email
inf-USAir97
inf-power
sc-nasasrb
sc-shipsec1
soc-academia
soc-anybeat
soc-brightkite
soc-dolphins
soc-hamsterster
soc-wiki-elec
fb-Brown11
fb-Harvard1
fb-NotreDame57
fb-Princeton12
fb-Vanderbilt48
fb-Yale4
fb-nips-ego
tech-as-caida2007
tech-internet-as
biogrid-human
bn-cat
0
Figure 6: The space requirements of the proposed approach are orders of magnitude less than existing methods. Note
log10 (σmin /σ) (y-axis) is the relative “gain/loss” (in space) over the best baseline method where σ is the size (bytes) of an
embedding and σmin is the embedding with the smallest size among the baseline methods (node2vec, DeepWalk, LINE); and
[ x2 x3 ], · · · , [ x1 · · · x9 ] is the attribute sets in Eq. (11)-(14) used as input to Φ for mapping nodes to types. These represent
variations of our method. The baseline method with the minimum space (σmin ) is shown at 0 and thus methods that use more
space lie below the horizontal line.
subsets of attributes in Eq. (11)-(14). For instance, Eq. (11)
indicates that node types are derived using the (logarithmic
binned) number of 2-stars x2 and triangles x3 that a given
node participates (Figure 2). To evaluate the space-efficiency
of the various methods, we measure the space (in bytes) required to store the embedding learned by each method. In
Figure 6, we summarize the reduction in space from our
approach compared to the other methods. In all cases, the
embeddings learned from our approach require significantly
less space and thus more space-efficient. Specifically, the
embeddings from our approach require on average 853 times
less space than the best method averaged across all graphs.
In addition, Table 2 provides the number of types derived
when using the various attribute combinations.
5
Related Work
Recent embedding techniques for graphs have largely been
based on the popular skip-gram model [11, 34] originally
introduced for learning vector representations of words in
the natural language processing (NLP) domain. In particular, DeepWalk [40] used this approach to embed the nodes
such that the co-occurrence frequencies of pairs in short random walks are preserved. More recently, node2vec [19] introduced hyperparameters to DeepWalk that tune the depth
and breadth of the random walks. These approaches are becoming increasingly popular and have been shown to outperform a number of existing methods. These methods are all
Table 2: Comparing the number of unique types.
M
G RAPH
N
biogrid–human
ca–citeseer
ca–dblp–2010
fb–Harvard1
fb–NotreDame57
sc–shipsec1
soc–academia
soc–brightkite
tech–internet–as
9527
227320
226413
15126
12155
140385
200169
56739
40164
Eq. (11) Eq. (12) Eq. (13) Eq. (14)
9
9
9
7
9
9
9
9
9
73
132
130
64
75
144
129
92
78
290
1049
1018
253
340
866
954
489
259
429
1783
1706
328
454
1346
1634
768
365
based on simple random walks and thus are well-suited for
generalization using the attributed random walk framework.
On the other hand, methods such as SkipGraph [29] make
use of simple random walks to learn embeddings for entire
graphs (as opposed to individual nodes). These methods can
be used for graph-level tasks such as graph classification and
clustering. Since these methods are still based on simple random walks, it is straightforward to generalize them using our
proposed framework.
While most work has focused on transductive (withinnetwork) learning, there has been some recent work on
graph-based inductive approaches [20, 48, 56]. Yang et al.
proposed an inductive approach called Planetoid [56]. How-
ever, Planetoid is an embedding-based approach for semisupervised learning and does not use any structural features.
Rossi et al. proposed an inductive approach for (attributed)
networks called DeepGL that learns (inductive) relational
functions representing compositions of one or more operators applied to an initial set of graph features [48]. More recently, Hamilton et al. [20] proposed a similar approach that
also aggregates features from node neighborhoods. However, these approaches are not based on random-walks.
Heterogeneous networks [51] have also been recently
considered [10, 14] as well as attributed networks [22, 23].
In particular, Huang et al. proposed an approach for attributed networks with labels [23] whereas Yang et al. used
text features to learn node representations [55]. Liang et al.
proposed a semi-supervised approach for networks with outliers [30]. Bojchevski et al. proposed an unsupervised rankbased approach [8]. More recently, Coley et al. introduced a
convolutional approach for attributed molecular graphs that
learns graph embeddings [13] as opposed to node embeddings. However, most of these approaches are not inductive
nor space-efficient.
Our work is also related to uniform and non-uniform random walks on graphs [12, 32]. Random walks are at the heart
of many important applications such as ranking [38], community detection [36, 42], recommendation [7], link prediction [31], influence modeling [24], search engines [28],
image segmentation [18], routing in wireless sensor networks [50], and time-series forecasting [16]. These applications and techniques may also benefit from the notion of
attributed random walks.
6
Conclusion
To make existing methods more generally applicable, this
work proposed a flexible framework based on the notion of
attributed random walks. The framework serves as a basis
for generalizing existing techniques (that are based on random walks) for use with attributed graphs, unseen nodes,
graph-based transfer learning tasks, and allowing significantly larger graphs due to the inherent space-efficiency of
the approach. Instead of learning individual embeddings for
each node, our approach learns embeddings for each type
based on functions that map feature vectors to types. This
allows for both inductive and transductive learning. Furthermore, the framework was shown to have the following
desired properties: space-efficient, accurate, inductive, and
able to support graphs with attributes (if available). Finally,
the approach is guaranteed to perform at least as well as the
original methods since they are recovered as a special case
in the framework.
References
[1] N. K. Ahmed and R. A. Rossi. Interactive visual graph analytics on the web. In ICWSM, 2015.
[2] N. K. Ahmed, J. Neville, R. A. Rossi, and N. Duffield. Efficient graphlet counting for large networks. In ICDM, page 10,
2015.
[3] N. K. Ahmed, T. L. Willke, and R. A. Rossi. Estimation of
Local Subgraph Counts. In BigData, pages 586–595, 2016.
[4] N. K. Ahmed, J. Neville, R. A. Rossi, N. G. Duffield, and T. L.
Willke. Graphlet Decomposition: Framework, Algorithms,
and Applications. Knowledge and Information Systems, 50
(3):689–722, Mar 2017.
[5] L. Akoglu, H. Tong, and D. Koutra. Graph based anomaly
detection and description: a survey. DMKD, 29(3):626–688,
2015.
[6] M. Al Hasan and M. J. Zaki. A survey of link prediction
in social networks. In Social Network Data Analytics, pages
243–275. 2011.
[7] T. Bogers. Movie recommendation using random walks over
the contextual graph. In Context-Aware Recommender Systems, 2010.
[8] A. Bojchevski and S. Günnemann. Deep gaussian embedding of attributed graphs: Unsupervised inductive learning via
ranking. arXiv:1707.03815, 2017.
[9] S. Borgatti and M. Everett. Notions of position in social network analysis. Sociological methodology, 22(1):1–35, 1992.
[10] S. Chang, W. Han, J. Tang, G.-J. Qi, C. C. Aggarwal, and
T. S. Huang. Heterogeneous network embedding via deep
architectures. In SIGKDD, pages 119–128, 2015.
[11] W. Cheng, C. Greaves, and M. Warren. From n-gram to skipgram to concgram. Int. J. of Corp. Linguistics, 11(4):411–
433, 2006.
[12] F. Chung. Random walks and local cuts in graphs. Linear
Algebra and its applications, 423(1):22–32, 2007.
[13] C. W. Coley, R. Barzilay, W. H. Green, T. S. Jaakkola, and
K. F. Jensen. Convolutional embedding of attributed molecular graphs for physical property prediction. J. Chem. Info. &
Mod., 2017.
[14] Y. Dong, N. V. Chawla, and A. Swami. metapath2vec: Scalable representation learning for heterogeneous networks. In
SIGKDD, pages 135–144, 2017.
[15] L. Getoor and B. Taskar, editors. Intro. to SRL. MIT Press,
2007.
[16] D. F. Gleich and R. A. Rossi. A dynamical system for pagerank with time-dependent teleportation. Inter. Math., pages
188–217, 2014.
[17] T. E. Goldsmith and D. M. Davenport. Assessing structural
similarity of graphs. 1990.
[18] L. Grady. Random walks for image segmentation. TPAMI,
28(11):1768–1783, 2006.
[19] A. Grover and J. Leskovec. node2vec: Scalable feature learning for networks. In SIGKDD, pages 855–864, 2016.
[20] W. L. Hamilton, R. Ying, and J. Leskovec. Inductive representation learning on large graphs. arXiv:1706.02216, 2017.
[21] M. Henaff, J. Bruna, and Y. LeCun. Deep convolutional networks on graph-structured data. arXiv:1506.05163, 2015.
[22] X. Huang, J. Li, and X. Hu. Accelerated attributed network
embedding. In SDM, pages 633–641, 2017.
[23] X. Huang, J. Li, and X. Hu. Label informed attributed network embedding. In WSDM, pages 731–739, 2017.
[24] A. Java, P. Kolari, T. Finin, and T. Oates. Modeling the spread
of influence on the blogosphere. In WWW, pages 22–26,
2006.
[25] S. Kearnes, K. McCloskey, M. Berndl, V. Pande, and P. Riley.
Molecular graph convolutions: moving beyond fingerprints.
Journal of computer-aided molecular design, 30(8):595–608,
2016.
works using random walks. J. Graph Alg. Appl., 10(2):191–
218, 2006.
[26] M. Koyutürk, Y. Kim, U. Topkara, S. Subramaniam, W. Szpankowski, and A. Grama. Pairwise alignment of protein interaction networks. JCB, 13(2):182–199, 2006.
[43] N. Pržulj. Biological network comparison using graphlet degree distribution. Bioinformatics, 23(2):e177–e183, 2007.
[27] A. Kuwadekar and J. Neville. Relational active learning for
joint collective classification models. In ICML, pages 385–
392, 2011.
[28] J.-L. Lassez, R. Rossi, and K. Jeev. Ranking links on the web:
Search and surf engines. In IEA/AIE, pages 199–208, 2008.
[29] J. B. Lee and X. Kong. Skip-Graph: Learning graph embeddings with an encoder-decoder model. In ICLR OpenReview,
2017.
[30] J. Liang, P. Jacobs, and S. Parthasarathy. Seano: Semisupervised embedding in attributed networks with outliers.
arXiv:1703.08100, 2017.
[31] W. Liu and L. Lü. Link prediction based on local random
walk. Europhysics Letters, 89:58007, 2010.
[32] L. Lovász. Random walks on graphs. Combinatorics, 2:1–46,
1993.
[33] L. K. McDowell, K. M. Gupta, and D. W. Aha. Cautious
collective classification. JMLR, 10:2777–2836, 2009.
[34] T. Mikolov, K. Chen, G. Corrado, and J. Dean. Efficient estimation of word representations in vector space.
arXiv:1301.3781, 2013.
[35] J. Neville and D. Jensen. Iterative classification in relational
data. In AAAI SRL Workshop, pages 13–20, 2000.
[36] A. Y. Ng, M. I. Jordan, and Y. Weiss. On spectral clustering:
Analysis and an algorithm. In NIPS, pages 849–856, 2002.
[37] V. Nicosia, J. Tang, C. Mascolo, M. Musolesi, G. Russo, and
V. Latora. Graph metrics for temporal networks. In Temporal
Networks, pages 15–40. Springer, 2013.
[38] L. Page, S. Brin, R. Motwani, and T. Winograd. PageRank
citation ranking: Bringing order to the web. Stanford Tech.
Report, 1998.
[39] Y. Park, D. Reeves, V. Mulukutla, and B. Sundaravel. Fast
malware classification by automated behavioral graph matching. In Workshop on Cyber Security and Info. Intel. Res.,
page 45, 2010.
[40] B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk: Online
learning of social representations. In SIGKDD, pages 701–
710, 2014.
[41] R. Pienta, J. Abello, M. Kahng, and D. H. Chau. Scalable
graph exploration and visualization: Sensemaking challenges
and opportunities. In BigComp, pages 271–278, 2015.
[42] P. Pons and M. Latapy. Computing communities in large net-
[44] F. Radicchi, C. Castellano, F. Cecconi, V. Loreto, and
D. Parisi. Defining and identifying communities in networks.
Proceedings of the National Academy of Sciences, 101(9):
2658–2663, 2004.
[45] R. A. Rossi and N. K. Ahmed.
The network data
repository with interactive graph analytics and visualization. In AAAI, pages 4292–4293, 2015. URL http://
networkrepository.com.
[46] R. A. Rossi and N. K. Ahmed. Role discovery in networks.
TKDE, 27(4):1112–1131, 2015.
[47] R. A. Rossi, L. K. McDowell, D. W. Aha, and J. Neville.
Transforming graph data for statistical relational learning.
JAIR, 45:363–441, 2012.
[48] R. A. Rossi, R. Zhou, and N. K. Ahmed. Deep feature learning for graphs. In arXiv:1704.08829, 2017.
[49] R. A. Rossi, R. Zhou, and N. K. Ahmed. Estimation of
Graphlet Statistics. In arXiv preprint, pages 1–14, 2017.
[50] S. D. Servetto and G. Barrenechea. Constrained random
walks on random graphs: routing algorithms for large scale
wireless sensor networks. In Wireless Sensor Networks &
App., pages 12–21, 2002.
[51] C. Shi, X. Kong, Y. Huang, S. Y. Philip, and B. Wu. HeteSim:
A General Framework for Relevance Measure in Heterogeneous Networks. TKDE, 26(10):2479–2492, 2014.
[52] J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei.
LINE: Large-scale Information Network Embedding. In
WWW, pages 1067–1077, 2015.
[53] B. Viswanath, A. Mislove, M. Cha, and K. P. Gummadi. On
the evolution of user interaction in Facebook. In OSN, pages
37–42, 2009.
[54] R. Xu and D. Wunsch. Survey of clustering algorithms. IEEE
Transactions on Neural Networks, 16(3):645–678, 2005.
[55] C. Yang, Z. Liu, D. Zhao, M. Sun, and E. Y. Chang. Network
representation learning with rich text information. In IJCAI,
pages 2111–2117, 2015.
[56] Z. Yang, W. W. Cohen, and R. Salakhutdinov.
Revisiting semi-supervised learning with graph embeddings.
arXiv:1603.08861, 2016.
[57] L. A. Zager and G. C. Verghese. Graph similarity scoring and
matching. Applied mathematics letters, 21(1):86–94, 2008.
[58] X. Zhu, Z. Ghahramani, and J. D. Lafferty. Semi-supervised
learning using gaussian fields and harmonic functions. In
ICML, pages 912–919, 2003.
| 2cs.AI
|
Inversion-Based Output Tracking and Unknown Input
Reconstruction of Square Discrete-Time Linear Systems
E. Naderi and K. Khorasani a
arXiv:1606.06215v2 [cs.SY] 23 Jan 2017
a
Department of Electrical and Computer Engineering, Concordia University, Montreal, Quebec, Canada
Abstract
In this paper, we propose a framework for output tracking control of both minimum phase (MP) and non-minimum phase
(NMP) systems as well as systems with transmission zeros on the unit circle. Towards this end, we first address the problem of
unknown state and input reconstruction of non-minimum phase systems. An unknown input observer (UIO) is designed that
accurately reconstructs the minimum phase states of the system. The reconstructed minimum phase states serve as inputs
to an FIR filter for a delayed non-minimum phase state reconstruction. It is shown that a quantified upper bound of the
reconstruction error exponentially decreases as the estimation delay is increased. Therefore, an almost perfect reconstruction
can be achieved by selecting the delay to be sufficiently large. Our proposed inversion scheme is then applied to solve the
output-tracking control problem. We have also proposed a methodology to handle the output tracking prob! lem of systems
that have transmission zeros on the unit circle in addition to MP and NMP zeros. Simulation case studies are also presented
that demonstrate the merits and capabilities of our proposed methodologies.
Key words: Inversion-based techniques; Unknown input reconstruction; Output tracking; Non-minimum phase systems;
1
Introduction
Output tracking problems arise in many control applications such as in aerospace and robotics. One possible
solution to this problem is an inversion-based approach
in which the control input is considered as the output
of an inverse system which is stimulated by the actual
system desired output. However, this approach is quite
challenging due to presence of unstable transmission zeros of the system. Unstable transmission zeros also challenge a stable reconstruction of the unknown system inputs using the known outputs. This can be considered
as an equivalent problem to the inversion-based output
tracking problem. Both inversion-based output tracking
and unknown input reconstruction problems have received extensive attention from the control community
researchers.
Inversion of linear systems was first systematically
treated by Brocket and Mesarovic [1]. The classic works
are known as structure algorithm [2], Sain & Massey
algorithm [3], and the Moylan algorithm [4]. Gillijns [5]
has proposed a general form of the Sain & Massey algorithm in which certain free parameters are available for
adjustment based on the design requirements. However,
this is accomplished under the assumption that the
original system does not have any unstable transmission
zeros.
The problem of unknown input reconstruction using inversion schemes has been tackled by more sophisticated
methods. Palanthandalam-Madapusi and his colleagues
([6], [7] and [8]) have considered the problem of input
reconstruction in several papers, yet, the provided solutions are only applicable to minimum-phase (MP)
systems. Xiong and Saif [9] have proposed an observer
for input reconstruction that works under limited cases
Email address: [email protected] (E. Naderi and
K. Khorasani).
of non-minimum phase (NMP) systems. Specifically,
to those non-minimum phase systems with zero feedthrough matix (D) and those systems having special
disturbance dynamics. The restrictive condition of requiring zero feed-through matrix appears in most works
that are related to the input reconstruction problem
[10]. Flouquet and his colleagues [11] have proposed
a sliding mode observer for the input recon! struction
which is also only valid for minimum phase systems
with zero feed-through matrix.
A successful solution to stable inversion of minimum
phase and non-minimum phase systems was proposed
in Zou and Devasia [12]. This solution is extended to
discrete-time systems in [13]. However, the solution
requires that the system should have a well-defined relative degree. A geometric solution for stable inversion
of linear systems is proposed by Marro and Zattoni [14].
The algebraic counterpart of the geometric solution is
provided in [15]. Both the above geometric and algebraic solutions do not provide a framework for handling
systems having transmission zeros on the unit circle in
addition to MP and NMP zeros.
The other problem of inversion-based output tracking
has also been the subject of a number of research in the
literature. It is well-known that unbiased inversion-based
output tracking is essentially non-causal since it requires
the information on the entire trajectory in future that
is not a reasonable assumption for many applications.
Zou and Devasia ([12,16,17]) have introduced previewbased stable-inversion method for continuous-time systems. Basically, this method requires access to a finite
window of future data instead of having the entire future
trajectory, although the approach results in a degraded
output tracking error performance. This technique has
been significantly improved by the recent work ([18,19]),
however, these works are also developed for continuoustime LTI systems. However, the method is constrained
by restrictive assumptions, such as requiring a we! lldefined relative degree condition. Several other work using different approaches are available in the literature
that are mostly application of a particular method such
as the Q-learning [20] or filtered basis functions [21].
In this study, we first address the inversion-based unknown state and input reconstruction problem. A general unknown input observer is proposed that accurately
and independently reconstructs the minimum phase
states of the system by using only the available system
measurements. The minimum phase states here refers
to n − p states of the overall system, where n denotes
the order of the system and p denotes the number of
unstable transmission zeros. Next, the estimated minimum phase states are considered as inputs to an FIR
filter to reconstruct the p non-minimum phase states of
the system. The FIR filter estimates the non-minimum
phase system states with a time delay of n + nd steps.
It also yields an estimation error which is a function
of the to be selected parameter nd . We have explicitly derived subsequently the relationship between the
reconstruction error and nd .
Specifically, we have shown that the estimation error
is proportional to inverse of the smallest non-minimum
phase zero to the power of nd . Hence, if the system does
not have any transmission zeros on the unit circle, the
estimation error asymptotically decays to zero as nd is
increased. This can therefore ensure that an unbiased
input and states estimation can be obtained. For most
cases, an nd equal to four or five times n would yield
an almost perfect estimation results for any smooth or
non-smooth unknown input. For a smooth input, an nd
as small as 2 may suffice.
We comprehensively address and discuss the dynamics
of the non-minimum phase states and have derived the
relationships among the system matrices. By invoking
a minor modification, our proposed methodology is
extended to solve the inversion-based output tracking
control problem. As opposed to a delayed reconstruction, our method now requires data corresponding to
n + nd time steps ahead of the desired trajectory. As in
the previous problem, we have quantified the tracking
error characteristics and have shown that an almost
perfect tracking is achievable by properly selecting nd
that yields an unbiased state reconstruction that can be
achieved as in the first problem.
Finally, we have shown that our proposed methodology
for stable inversion of linear systems can be successfully
extended to handle the output tracking problem in systems that have transmission zeros on the unit circle in
addition to MP and NMP zeros. Our proposed solution
introduces a further delay of nc due to implementation
of a controller of order nc . In contrast to unstable transmission zeros, the output tracking error does not exponentially decrease by increasing nc . Instead, the output
tracking error depends on the norm of a transfer function which is parameterized by the system and the selected controller parameters. We have further characterized design criteria and have formulated a minimization
problem for selection of the controller parameters.
(2)
(3)
(4)
(5)
(6)
minimum phase linear discrete-time systems is proposed and developed,
In our proposed methodology, the MP states are
partitioned and estimated by using the system measurements. The MP states are exactly estimated
with a delay of at most equal to the system order;
in contrast to the available works in the literature
where all states are approximated with a delay that
depends on the location of the smallest unstable
transmission zero,
Our proposed solution does not require that the
system should have a well-defined relative degree,
An algorithm and a simple constructive procedure
for designing the inversion-based output tracking
control scheme is proposed,
We have shown that our proposed solution provides
a framework for handling the output tracking problem of systems that have transmission zeros on the
unit circle in addition to MP and NMP zeros for
the first time in the literature, and finally
The accuracy of our proposed input and state estimation scheme as well as the output tracking control performance as a function of the delay parameter are quantified and investigated. For the case
that the system has transmission zeros on the unit
circle, the output tracking error is characterized by
the norm of a transfer matrix that depends on the
system and controller parameters.
The remainder of this paper is organized as follows. The
problem statement and preliminaries are provided in
Section 2. Section 3 is devoted to the problem of developing and designing unknown state and input reconstruction methodologies. The problem of developing an
inversion-based output tracking strategy is addressed in
Section 4. The extension of the solution to the case where
the system has transmission zeros on the unit circle is
developed in Section 5. Finally, several numerical case
studies are presented in Section 6 to demonstrate and illustrate the capabilities of our proposed methodologies.
2 Problem Statement
Consider the following deterministic discrete-time linear
time-invariant (LTI) system S,
(
x(k + 1) = Ax(k) + Bu(k)
S:
(1)
y(k) = Cx(k) + Du(k)
where x ∈ Rn , u ∈ Rm and y ∈ Rl . The quadruple
Σ := (A, B, C, D) is assumed to be known a priori. The
output measurement y(k) is also assumed to be available, however, both the system states x(k) and u(k) are
assumed to be unmeasurable. In this paper, we consider
the following two specific problems.
Problem 1: The system states and the unknown input
reconstruction: The objective of this problem is to estimate the system state x(k) and the unknown input u(k)
from the only available system measurement y(k). The
main assumption that is imposed to solve this problem
is given by Assumption 1 below.
To summarize, the main contributions of this paper can
be stated as follows:
Assumptions: The system S is assumed to satisfy the
following conditions. namely, i) the system is square
(m = l), ii) the system has a minimal realization, and
iii) the system does not have any zeros on the unit circle.
(1) A methodology for estimation of unknown states
and unknown inputs of both minimum and non-
In Section 5 we relax the Assumption (iii). Other conditions that may be required are provided under each
specific statement and result subsequently.
scheme that estimates the remaining set of the system
states.
Problem 2: The output tracking: The objective of this
problem is to estimate the input signal u(k) such that
the output y(k) follows a desired trajectory yd (k). This
problem is in fact another re-statement of the Problem
1 above with the difference that the actual output of the
system is now replaced by yd (k). The main assumption
that is also required here is Assumption 1.
3.1
We now present the notation that is used throughout
the paper. Given the matrix A, then A⊥ , AT and N (A)
denote the orthogonal space, the transpose, and the null
space of A, respectively. We use the concept of pseudo inverse. If A is full column rank, then we denote the pseudo
inverse of A by A† and compute it by (AT A)−1 AT . If
A is rank deficient, then we denote its pseudo inverse
by A+ , where A+ is a matrix that satisfies the following four conditions: 1) AA+ A = A, 2) A+ AA+ = A+ ,
3) (AA+ )T = AA+ , and 4) (A+ A)T = A+ A. If U ΣV T
denotes the SVD decomposition of A, then A+ is given
by V Σ+ U T , where Σ+ is obtained by reciprocating each
non-zero diagonal element of Σ. If A denotes the system
matrix, then A(1) implies transformation of A under a
standard similarity transformation matrix T(1) . If x(k)
denotes a vector, then x̂(k) represent an estimate of x(k).
Also, x(1) (k) denotes the transformation of x(k) under
the similarity matrix T(1) , i.e. x(1) (k) = T(1) x(k). Finally, diag(V) denotes a diagonal matrix with elements
of the vector V on its diagonal. Consider the Rosenbrock
System Matrix defined by,
"
#
zI − A B
MR (z) =
(2)
C
D
if rank(MR (z)) < n + l, then z is called a transmission
zero (or simply the zero) of the system S or the quadruple (A, B, C, D). The abbreviations MP and NMP stand
for minimum phase and non-minimum phase systems,
respectively.
3
State and Unknown input reconstruction
In this section, we consider and develop methodologies
for solving the Problem 1. Let us first set up an unknown
input observer (UIO) that generates the state η(k) as an
estimate of Mx(k) by using only the system measurements y(k), where M ∈ Rq×n is a full row rank matrix
to be specified. If rank(M) = n, then the system states
can be fully reconstructed since x̂(k) = M−1 η(k). However, such an M with rank equal to n does not always
exist. In fact, it turns out that the rank of M is closely
related to the transmission zeros of the system S.
More specifically, we will show that rank(M) = n −
β, where α and β are now representing the number of
finite MP and NMP transmission zeros of the system
S, respectively. Clearly, α + β is not necessarily equal
to n. Our strategy is to first construct an M having
the rank n − β by using two to be designed matrices
M0 and M# that are specified subsequently based on
the system S matrices. Given M, we then introduce a
transformation to partition the system states that can be
exactly estimated from those where their estimation is
obstructed by the NMP transmission zeros of the system.
The estimated states will then serve as inputs to a causal
Partial or full estimation of the system states
We start by stating our first formal definition.
Definition 1 Assume M ∈ Rq×n , where q ≤ n, is a full
row rank matrix. We denote η(k) = Mx(k) as a partial
or full estimate of the system S states if q < n or q = n,
respectively.
Our goal is to design an unknown input observer (UIO)
that estimates Mx, where M ∈ Rq×n , q ≤ n, is a full
row rank matrix. We consider the governing dynamics
of the unknown input observer (UIO) as follows,
η(k − n + 1) = Âη(k − n) + F Y(k − n)
(3)
where,
y(k − n + 1)
y(k − n + 2)
∈ Rnl
Y(k − n) =
..
.
y(k)
(4)
with the matrices  ∈ Rq×q and F ∈ Rq×(nl) to be
specified subsequently. Our objective is to now select
the matrices M, Â and F such that η(k) − Mx(k) → 0
as k → ∞. The output measurement equation of the
system S can be alternatively expressed as,
Y(k − n) = Cn x(k − n) + Dn U(k − n)
where,
C
D
0
D
CA
CB
; Dn =
Cn =
..
..
..
.
.
.
n−1
n−2
n−3
CA
CA
B CA
B
(5)
0
0
..
.
D
(6)
where Cn ∈ R(nl)×n , Dn ∈ R(nl)×(nm) and U(k − n) ∈
Rmn is constructed similar to Y(k − n) from the input
sequence. The state equation of the system S can be
expressed as,
...
...
..
.
...
x(k − n + 1) = Ax(k − n) + BIn U(k − n)
(7)
h
i
where In = Im×m 0m×(n−m) . Using the equations
(3), (5) and (7), the unknown input observer error dynamics is now governed by,
(η − Mx)(k − n + 1) = Â(η(k − n) − Mx(k − n))
+ (ÂM − MA + F Cn )Y(k − n)
+ (F Dn − MBIn )U(k − n) (8)
It now follows that Mx is accurately estimated if and
only if (i) Â is selected to be a Hurwitz matrix, (ii)
0 = ÂM−MA+F Cn , and (iii) 0 = F Dn −MBIn . The
conditions (i)-(iii) above are the well-known unknown
input observer equations that are solvable under certain
conditions [22]. We will show that these conditions have
a solution if and only if the system S is MP. However,
this will be obtained under the restrictive requirement
that M should be full rank square matrix. We will show
subsequently that a solution for NMP systems exists if
a lower rank matrix M is considered.
From the condition (iii) it follows that,
+
F = MBIn D+
n + Kn (I − Dn Dn )
(9)
where Kn ∈ Rnl×nm is an arbitrary matrix. Let us first
denote by Â0 and M0 as solution to  and M that satisfy the conditions (i)-(iii) corresponding to Kn ≡ 0.
Subsequently, we shall return to the general case where
Kn and (I − Dn D+
n ) are nonzero to obtain another solution to M that we will denote by M# . For now for M0 ,
we have,
F0 = M0 BIn D+
(10)
n
If we substitute F0 from equation (10) into the condition
(ii), we obtain,
Â0 M0 = M0 (A − BIn D+
n Cn )
(11)
PROOF. Follows by direct substitution of the solution
above into the conditions (i)-(iii) that verifies the result.
Remark 2 One may suggest to use the Jordan canonical form of Γ to obtain a solution to the conditions
(i)-(iii), especially when the system S has repeated MP
transmission zeros. This may yield an M having higher
rank condition as compared to the solution provided by
Lemma 1 under certain limited cases. However, in general this will not lead to a robust numerical procedure and
in most cases the algorithm could fail numerically due to
ill-conditioning.
Lemma 1 implies that a solution for NMP systems exists
unless the system S has exactly n NMP transmission
zeros (this is highly unusual in real applications). Our
proposed methodology for state estimation problem that
will be subsequently discussed requires that rank(M) =
n−β . However, rank of M0 that is obtained from Lemma
1 is not necessarily equal to n − β, since Γ may have
multiple eigenvectors due to repeated eigenvalues and
the generalized eigenvectors are not a solution to the
equation (11).
Equation (11) - which is in fact the Sylvester equation has M0 = 0 as its trivial solution. The non-trivial solution to (11) is obtained if M0 is considered as the transpose of the left eigenvectors of Γ = (A− BIn D+
n Cn ) and
Â0 as a diagonal matrix of Γ eigenvalues. It now follows
that the full estimation of the system states by the UIO
observer (3) is obstructed by the NMP transmission zeros of the system due to the fact that the eigenvalues of
(A − BIn D+
n Cn ) contain NMP zeros of the square system S as formally stated in the following theorem.
Specifically, the set Z (as defined in Theorem 1) may
have αz elements sharing the same eigenvectors. We now
consider the term Kn (I − Dn D+
n )Cn in order to obtain linearly independent vectors associated with the elements of Z. If the set Z is not empty, then it implies
that Dn is rank deficient, and therefore (I − Dn D+
n ) is
a nonzero matrix.
Theorem 1 Let Assumption 1 hold, and V = {vi |i =
1, .., p} denote the set of the system S invariant zeros, and
Z = {0, . . . , 0} that contains n−p zeros. The eigenvalues
of Γ = (A − BIn D+
n Cn ) are given by V ∪ Z.
+
Â# M# = M# (A − BIn D+
n Cn ) + Kn (I − Dn Dn )Cn
(12)
)C
is
not
identically
zero,
a
non-trivial
Since (I−Dn D+
n
n
solution exists and Â# , M# and Kn can be selected such
that the condition (i) is satisfied. Therefore, we have the
following theorem.
PROOF. Proof is provided in the Appendix A.
Remark 1 It should be noted that Theorem 1 does not
hold for non-square systems. The eigenvalues of Γ may or
may not coincide with the transmission zeros of S. Each
case needs to be then separately investigated, however,
once the eigenvalues of Γ are determined, the remaining
procedure for obtaining a solution to the conditions (i)(iii) is similar to that of a square system.
If the system S has at least one MP transmission zero,
or it has less than n NMP zeros (therefore, the set Z in
Theorem 1 is not empty), then at least one eigenvalue
of Γ is less than 1, which is denoted by a. Let us now
set Â0 = a. If MT0 is chosen to be the left eigenvector
associated with the eigenvalue a, then equation (11),
and consequently conditions (i)-(iii) are satisfied even if
the system S has nonzero NMP transmission zeros. In
general, we can state the following result.
Lemma 1 Let Assumption 1 hold, and V = {vi |i =
1, .., p} denote the set of the system S invariant zeros,
Z = {0, . . . , 0} that contains n − p zeros, and Θα the set
of MP transmission zeros of S. If {Θα ∪ Z} 6= ∅, then
T
F0 = M0 BIn D+
n , Â0 = diag(Θα ∪ Z) and M0 that has
left eigenvectors of Γ associated with diag(Θα ∪ Z) are
solutions to the conditions (i)-(iii).
Let us now construct M# and Â# such that they satisfy
the following Sylvester equation,
Theorem 2 Let Assumption 1 hold and all the MP
transmission zeros of S have an algebraic multiplicity of
1. Then, the complete solution to the conditions (i)-(iii)
is given by,
#
#
"
"
M0
Â0 0
(13)
;M =
 =
M#
0 Â#
where rank(M) = n − β.
PROOF. The proof is provided in the Appendix B.
Note that if the system S has MP transmission zeros with
an algebraic multiplicity that is higher than 1, then the
rank of M# is reduced proportionally by the multiplicity
of the MP transmission zeros. This is due to the fact that
(I − Dn D+
n ) loses its rank. On the other hand, M0 also
loses its rank by such MP transmission zeros. Therefore,
our method fails, since the rank of M will be less than
n − β. However, we will introduce a technique in Section
5 to relax the assumption on simplicity of the MP zeros.
The solution given in equation (12) is closely related to
equation (5). The matrix I − Dn D+
n gives the null space
of Dn . Multiplication of both sides of equation (5) by
this matrix yields,
+
(I − Dn D+
n )Y(k − n) = (I − Dn Dn )Cn x(k − n) (14)
Let us now define P = (I−Dn D+
n )Cn . It follows that the
rank of P depends on the rank of N (Dn ). If the system S
has exactly p = n transmission zeros, then N (Dn ) = 0,
and consequently P ≡ 0. On the other hand, M0 will be
full row rank and will have n − β linearly independent
rows if the MP transmission zeros are simple. As p is
reduced, then the rank of P increases and the rank of M0
decreases. This relationship reveals several important
characteristics of N (Dn ). A more detailed discussion of
these properties is beyond the scope of this paper.
3.2
Partitioning of the states
If the system S has any NMP transmission zeros, then
rank(M) = q < n, and therefore the states cannot be
fully estimated. Let us now perform an LQ decomposition of the matrix M to partition the estimation of the
q states from the estimation of the other n − q states.
Namely, let us set M = LQ.
The unknown input observer (UIO) is described by equation (3), where  and F are selected according to Theorem 2 and equation (9), and where η(k−n) = Mx̂(k−n).
Equivalently, we have η(k − n) = LQx̂(k − n). Let us
now set the similarity transformation
matrix T(1) = Q.
h
i
Therefore, η(k − n) = M(1)
0 x̂(1) (k − n) 1 , where
q
h
i
(1)
(1)
Mq 0 = L. The matrix Mq ∈ Rq×q is a nonsingular matrix, hence the first q states can be independently reconstructed from η(k − n) as follows,
−1
x̂(1) (1 : q)(k − n) = M(1)
η(k − n)
q
(15)
where x(1 : q) denotes the first q elements of the vector
x.
Definition 2 The MP and NMP states correspond to
the first q and the last n − q states of the system S (1) and
(1)
(1)
are denoted by x1 (k) and x2 (k), respectively. In other
iT
h
(1)
T
T
words, x(1) (k) = (x(1)
, where
(k))
(x
(k))
1
2
S
(1)
:
(
x(1) (k + 1) = A(1) x(1) (k) + B (1) u(k)
y(k) = C (1) x(1) (k) + Du(k)
(16)
Equation (18) shows that the MP states can be independently and accurately estimated from the system
(1)
(1)
measurements. In other words, x̂1 (k − n) → x1 (k −
n) as k → ∞. This is due to the fact that according to the error dynamics (8) and conditions (i)-(iii),
η(k − n) − Mx(k − n) → 0 as k → ∞. Therefore,
Lx̂(1) (k − n) − Lx(1) (k − n) → 0 as k → ∞, which yields
the desired result. An important property of the MP
states is now given by the following theorem.
(1)
Theorem 3 Let Assumption 1 hold. Then x1 (k) → 0
as k → ∞ if and only if y(k) = 0 for k = k0 , k0 +
1, . . . , ∞, k0 > 0.
PROOF. It is known from the state equation of the
system (18) that η(k) = 0 if and only if y(k) = 0 (⇒
Y(k) = 0) for k = k0 , k0 +1, . . . , ∞, k0 > 0. On the other
(1) (1)
(1)
hand, η(k) = Mq x̂1 (k). Since Mq is a nonsingular
(1)
matrix, it follows that x̂1 (k) ≡ 0 if and only if η(k) ≡
(1)
(1)
0. Moreover, x1 (k) → x̂1 (k) as k → ∞. Therefore,
(1)
x1 (k) → 0 as k → ∞, if and only if y(k) = 0 for
k = k0 , k0 + 1, . . . , ∞, k0 > 0.
The above partitioning is quite helpful in several ways.
The most important one is that it renders an elegant expression for the NMP states reconstruction estimation
error as discussed in the next section. Furthermore, in
certain applications such as in fault detection and isolation problems, the considered faults may only affect the
MP states of the system. Therefore, it will not be necessary to estimate the NMP system states that can be
computationally costly as well as an error prone process.
3.3
Dynamics of the MP and NMP states
The unknown input estimation problem requires a successful reconstruction of both the MP and the NMP
states. Towards this end, we partition the state space
(1)
model of the system S or S(1) as follows (x1 (k) ∈ Rq
(1)
and x2 (k) ∈ Rn−q ),
S(1)
Considering the Definition 2 and equation (15), we have,
(1)
−1
x̂1 (k − n) = (M(1)
η(k − n)
q )
or in the state space representation,
(
η(k − n + 1) = Âη(k − n) + F Y(k − n)
(1)
(1)
x̂1 (k − n) = (Mq )−1 η(k − n)
1
(17)
where,
A(1) =
(18)
Recall the notation that was defined in Section 2,
(1)
namely, x(1) (k) = T(1) x(k), x1 (k) = T(1) x(1) (k), A(1) =
(1)
(1) −1
(1)
(1)
T A(T ) , B = T B, and C (1) = C(T(1) )−1 .
"
(1)
(1) (1)
x1 (k − n + 1) = A11 x1 (k − 1)
(1)
(1)
(1)
+A12 x2 (k − n) + B1 u(k − n)
(1)
(1) (1)
x2 (k − n + 1) = A21 x1 (k − n)
(1)
(1)
: +A(1)
22 x2 (k − n) + B2 u(k − n)
(1) (1)
y(k − n) ="C1 x1 (k −#n)+
i x(1) (k − n)
h (1)
2
C
D
2
u(k − n)
(1)
(1)
A11 A12
(1)
(1)
A21 A22
#
; B (1) =
"
(1)
B1
(1)
B2
#
(19)
i
h
; C (1) = C1(1) C2(1) .
(20)
It is now straightforward to conclude from Theorem 3
that the following lemmas imply that the NMP states
cannot be algebraically estimated from the MP states
and the system measurement outputs. Specifically, we
have:
Lemma 2 Let "Assumption#1 hold and 0 < q < n. Then
(1)
C2
D
the columns of
are linearly dependent.
(1)
(1)
A12 B1
PROOF. Proof is provided in the Appendix C.
Lemma 3 Let Assumption" 1 hold and#0 < q < n. Then
(1)
(1)
A22 B2
are a subset of
the transmission zeros of
(1)
(1)
A12 B1
the system S transmission zeros.
(1)
If on the other hand B1 is not a full column rank matrix, then let us assume that D is full column rank. In this
case, the unknown input in terms of the system states is
given by the following expression,
(1) (1)
u(k − n) = D† y(k − n) − C1 x1 (k − n)
(1) (1)
−C2 x2 (k − n)
(26)
By substituting equation (26) into the second equation
of (19), it yields,
(1)
(1)
Lemma 4 Let Assumption" 1 hold and#0 < q < n. Then
(1)
(1)
A22 B2
are a subset of
the transmission zeros of
(1)
C2
D
the system S transmission zeros.
where,
(1)
(1)
Let us now assume that B1 is full column rank. Then,
the unknown input u(k) in terms of the system states
is obtained by the first expression of equation (19), according to
(1)
(1)†
x1 (k − n + 1)−
u(k − n) = B1
(1) (1)
(1) (1)
A11 x1 (k − n) − A12 x2 (k − n)
(21)
By substituting the above equation into the second and
third equations of (19) yields,
(
(1)
(1)
(1)
x2 (k − n + 1) = Az x2 (k − n) + Bz X1 (k − n)
(1)
(1)
y(k − n) = Cz2 x2 (k − n) + Cz1 X1 (k − n)
(22)
where,
(1)
(1) (1)† (1)
Az = A22 − B2 B1 A12
(23)
i
h
(1) (1)†
(1)
(1) (1)† (1)
(24)
Bz = B2 B1
A21 − B2 B1 A11
" (1)
#
x1 (k − n + 1)
(1)
X1 (k − n) =
(25)
(1)
x1 (k − n)
(1)†
(1)
Cz2 = C2 −i DB1 A12 and Cz1 =
hand where
(1)† (1)
(1)
(1)†
C1 − DB1 A11 . The quadruple Σz :=
DB1
(Az , Bz , Cz1 , Cz2 ) have interesting properties that are
related to the transmission zeros of the system S. We
are now in a position to state our next result.
(1)
(1)
Azd = A22 − B2 D† C2
h
i
(1) † (1)
(1) †
Bzd = A(1)
−
B
D
C
B
D
21
2
1
2
"
#
(1)
x1 (k − n)
(1)
X1d (k − n) =
y(k − n)
PROOF. Proof is provided in the Appendix E.
(1)
(1)
x2 (k − n + 1) = Azd x2 (k − n) + Bzd X1d (k − n) (27)
PROOF. Proof is provided in the Appendix D.
(28)
(29)
(30)
We can now state the next result of this paper.
Theorem 5 Let Assumption 1 hold, 0 < q < n, and D
be a full column rank matrix. Then, the eigenvalues of
Azd are the NMP zeros of the system S.
PROOF. Proof is provided in the Appendix G.
(1)
It should be noted that if both B1 and D are column
rank deficient matrices, then the NMP states and the
unknown input can no longer be estimated. This is a
slightly stronger assumption
" than
# the input observability
B (1)
that requires the matrix
to be full column rank.
D
3.4
Estimation of the NMP states
The state equation (22) (or similarly the equation (27)
(1)
depending on the rank condition of B1 or D) describes
the dynamics of the NMP states. The eigenvalues of Az
(or Azd ) coincide with the NMP transmission zeros of
the system S. Therefore, the dynamics of equation (22)
or equation (27) is unstable. This unstable dynamics
should be treated in a manner that provides a stable
estimation of the NMP states. Towards this end, let us
now consider the following non-casual structure that is
obtained by re-arranging the state representation (22)
or (27) as follows
(1)
Theorem 4 Let Assumption 1 hold, 0 < q < n and B1
be a full column rank matrix. Then, the eigenvalues of Az
are the NMP zeros of the system S. Moreover, Cz2 = 0.
PROOF. The proof is provided in the Appendix F.
Remark 3 According to Theorem 4 and the definition
(1)
of Cz2 , if D happens to be zero, then, C2 must be zero
(1) (1)
which implies y(k) = C1 x1 (k). This fact seems to be
useful for design of a robust fault detection and isolation
scheme, that is left as a topic of our future research.
(1)
(1)
(1)
x2 (k − n) = Ãz x2 (k − n + 1) − B̃z Θ1 (k − n) (31)
where,
Ãz = (Az )−1 (for (22)) or (Azd )−1 (for (27))
(32)
B̃z = (Az )−1 Bz (for (22)) or (Azd )−1 Bzd (for (27))
(33)
(1)
(1)
(1)
Θ1 (k−n) = X1 (k−n) (for (22)) or X1d (k−n) (for (27))
(34)
(1)
Iterating equation (31) for nd time steps yields,
(1)
(1)
x2 (k − n − nd ) = Ãnz d x2 (k − n)−
nX
d −1
(1)
(Ãz )i B̃z Θ1 (k − n − i − 1) (35)
i=0
where Ãnz d denotes Ãz raised to the power of nd . The
inverse of Az (or Azd ) exists since Az (or Azd ) does not
have a zero eigenvalue. Also, Ãz is Hurwitz due to the
fact that the eigenvalues of the inverse matrix is the
inverse of the matrix eigenvalues. Equation (35) provides
the key to estimation of the NMP states.
Let us now construct the following FIR filter,
(1)
(1)
x̂2 (k − n − nd ) = Ãnz d x̄20 (k − n)−
nX
d −1
(1)
(Ãz )i B̃z Θ̂1 (k − n − i − 1) (36)
i=0
(1)
where x̄20 (k−n) denotes the random initial condition of
(1)
the FIR filter at each time step k − n and Θ̂1 (k − n) =
(1)
(1)
(1)
X̂1 (k − n) or Θ̂1 (k − n) = X̂1d (k − n), depending on
whether B (1) or D is full column rank, respectively.
"
(1)
X̂1 (k
(1)
x̂1 (k − n + 1)
(1)
x̂1 (k − n)
#
− n) =
and
"
#
(1)
x̂1 (k − n)
(1)
X̂1d (k − n) =
. The estimate of the MP
y(k − n)
Moreover,
(1)
(1)
Since x1 (k) − x̂1 (k) → 0 as k → ∞, then Θ1 (k) −
(1)
Θ̂1 (k) → 0 as k → ∞. Therefore, the NMP state
estimation error is now given by ex2 (k − n − nd ) =
(1)
(1)
Ãnz d (x2 (k − n) − x̄20 (k − n)) as k → ∞.
Theorem 6 highlights a number of important trade-off
analysis considerations regarding the nature of the NMP
state estimation error and the selection of the delay nd .
Specifically, the following observations can be made:
• The farther the NMP transmission zeros are from the
unit circle, a smaller NMP state estimation error can
be ensured since the term Ãnz d decays faster to zero,
• The NMP state estimation error for the MP strictly
stable system is zero since these systems have a NMP
zero at infinity that results in Ãnz d ≡ 0, and
• The closer the NMP transmission zeros are to the unit
circle, one can ensure a larger NMP state estimation
error to the point that if the system S has any transmission zeros on the unit circle, then the NMP state
estimation results will be certainly biased regardless
of the choice of nd .
It turns out that one can obtain a conservative upper
bound on the NMP state estimation error by considering
the 2-norm of ex2 (k). We are now in a position to state
our next result.
Theorem 7 Let Assumption 1 hold, 0 < q < n, either
(1)
(1)
B1 or D is a full column rank matrix and x̄20 (k) =
nd
0 for all k. Then sup(kex2 (k)k2 ) = σmax (Ãz )k(zI −
A(1) )−1 B (1) k∞ , where σmax (.) denotes the largest singular value operator.
(1)
states (x̂1 (k)) as previously discussed is given by (18).
(1)
The random initial condition x̄20 (k − n) at each time
step introduces errors in the estimation process, but for
sufficiently large nd , the effects of the initial conditions
(1)
(1)
will vanish and x̂2 (k − n − nd ) − x2 (k − n − nd ) → 0
as k → ∞ (note that Ãnz d → 0 for nd ≫ 1), as shown
subsequently.
Practically, nd must be as small as possible, however an
accurate estimation requires a large nd . Hence, selection
of nd requires a trade-off analysis by quantification of
the estimation error versus nd at each time step. Below,
we provide an explicit expression for the reconstruction
or estimation error as a function of the delay nd and the
initial condition.
Definition 3 The NMP state estimation error is defined
(1)
(1)
according to ex2 (k) = x2 (k) − x̂2 (k).
Theorem 6 Let Assumption 1 hold, 0 < q < n , and
(1)
either B1 or D is a full column rank matrix. Then the
NMP state estimation error at the time step k − n − nd
(1)
(1)
is given by Ãnz d (x2 (k − n) − x̄20 (k − n)).
PROOF. Note that we have,
(1)
(1)
ex2 (k − n − nd ) = x2 (k − n − nd ) − x̂2 (k − n − nd )
(1)
(1)
= Ãnz d (x2 (k − n) − x̄20 (k − n))
nX
d −1
(1)
(Ãz )i B̃z (Θ1 (k − n − i − 1)
−
i=0
(1)
− Θ̂1 (k
− n − i − 1))
PROOF. It follows from Theorem 6 that,
(1)
kex2 (k − n − nd )k2 = kÃnz d x2 (k − n)k2
(37)
(1)
≤ kÃnz d k2 kx2 (k − n)k2
≤ σmax (Ãnz d )k(zI − A(1) )−1 B (1) k∞
The last inequality holds since the L2 input-output gain
is bounded by the ∞-norm of the system S.
The above upper bound can be plotted as a function of
nd to perform a trade-off analysis. Note that σmax (Ãnz d )
is determined by the smallest NMP transmission zero of
the system S due to the fact that the eigenvalues of Ãz
are inverse of the system S NMP transmission zeros. This
is in accordance with the results that are stated in [14].
Note that σmax (Ãnz d ) asymptotically decays to zero as
nd is increased. Therefore, an almost perfect estimation
can be achieved when nd is equal to several times that
of the system order.
Remark 4 If the system S is stimulated by an input
such that u(k + 1) 6= u(k) at finite k’s (such as in a
step function) or ku(k + 1) − u(k)k is sufficiently small
(such as in a harmonic function), then one can choose
(1)
(1)
x̄20 (k − n) = x̂2 (k − n − nd − 1) in the filter (36) which
may yield an almost unbiased state estimate by selecting
the smallest possible choice of nd = 2. This is due to the
(1)
fact that in these cases x̂2 (k − n − nd − 1) is a close
(1)
(1)
approximation to x2 (k − n − nd − 1) and x2 (k − n)
(for small nd ), and therefore it may yield a sufficiently
small NMP state estimation error, i.e., ex2 (k −n−nd ) =
(1)
(1)
(1)
(1)
Ãnz d (x2 (k−n)−x̄20 (k−n)) = Ãnz d (x2 (k−n)−x̂2 (k−
n − nd − 1)) ≈ 0 even if nd is selected to be sufficiently
small.
We will illustrate the above statement in our simulation
case study section. Once both the MP and NMP states
are estimated, the unknown input can now be easily estimated by using equation (21) (or (26)). Specifically, if
(1)
B1 is full column rank, then û(k) is given by,
(1)†
û(k − n) = B1
(1)
x̂1 (k − n + 1)−
(1) (1)
(1) (1)
A11 x̂1 (k − n) − A12 x̂2 (k − n)
(38)
and if D is full column rank, it is given by,
(1) (1)
û(k − n) = D† y(k − n) − C1 x̂1 (k − n)
(1) (1)
−C2 x̂2 (k − n)
(39)
Definition 4 The unknown input estimation error is
defined according to eu (k) = û(k) − u(k).
Proposition 1 Let Assumption 1 hold, 0 < q < n, and
(1)
B1 be a full column rank matrix. Then,
(1)†
eu (k) → −B1
(1)
A12 ex2 (k) as k → ∞.
(40)
PROOF. The result follows readily from equations (21) and (38) by noting that u(k) − û(k) =
(1)†
(1)
(1) (1)
(1) (1)
(x1 (k + 1) − A11 x1 (k) − A12 x2 (k)) −
(1)
(1) (1)
(1) (1)
(1)†
B1 (x̂1 (k + 1) − A11 x̂1 (k) − A12 x̂2 (k)) →
(1)† (1)
−B1 A12 ex2 (k) as k → ∞. This follows due to
(1)
(1)
the fact that x̂1 (k) → x1 (k) as k → ∞ and
(1)
(1)
ex2 (k) = x2 (k) − x̂2 (k) (Definition 3).
B1
The Proposition 1 links the unknown input estimation
error to the state estimation error. This may serve as
a means for conducting a trade-off analysis. The above
implies that the state estimation error is propagated
(1)† (1)
through the gain −B1 A12 to the unknown input estimation error. One can interestingly conclude that if
(1)† (1)
−B1 A12 happens to be zero, then the unknown input estimation process will be unbiased regardless of the
NMP states estimation error. Therefore, it can immedi(1)† (1)
ately be concluded that −B1 A12 = 0 if and only if the
NMP zero of the system S is at infinity. In other words,
the system S is strictly stable and MP. The proposition
1 provides an explicit unknown input estimation error
(1)
expression if B1 is full column rank. In case that D is
a full column rank matrix, we arrive at the ! following
result.
Proposition 2 Let Assumption 1 hold, 0 < q < n , and
D be a full column rank matrix. Then,
(1)
eu (k) → −D† C2 ex2 (k) as k → ∞.
(1) (1)(k)
(1) (1)
(1) (1)
C2 x2
) − D† (y(k) − C1 x̂1 (k) − C2 x̂2 (k)) →
(1)
−D† C2 ex2 (k) as k → ∞. This follows due to the fact
(1)
(1)
(1)
that x̂1 (k) → x1 (k) as k → ∞ and ex2 (k) = x2 (k)−
(1)
x̂2 (k) (Definition 3).
An immediate conclusion from the Propositions 1 and 2
(1)
is that if the system S is NMP and both B1 and D are
(1)†
(1)
(1)
full column rank matrices, then B1 A12 = D† C2 ,
which we have already derived through a different
(1)
method in Theorem 4 (Cz2 = 0). This completes our solution to the Problem 1. In the next section, we discuss
a solution to the Problem 2.
It is worth pointing that our proposed methodology is
not suitable for systems when all the zeros are NMP and
the system has the same number of poles and zeros. In
fact, this scenario for a square system implies that the
matrix D is full rank, therefore, other methods such as
the one in [10] are available to handle this particular
case.
4 Inversion-based output tracking
We have shown earlier that in presence of NMP states,
accurate estimation of the MP states as well as bounded
error estimation of the NMP states are possible under
certain conditions. In this section, by utilizing the previous results we will introduce and develop an inversionbased output tracking control methodology as a solution to Problem 2. Specifically, we will obtain relationship between the resulting tracking error performance
and the unknown input and state estimation errors. We
also demonstrate that almost perfect tracking of an arbitrary desired output trajectory can be achieved.
For the output tracking problem a delayed state and
input estimation may not be useful or practical given
that the controller should issue the command at a given
present time. This challenge can be resolved if we assume that the desired output trajectory from yd (k) to
yd (k + n + nd ) is known a priori at a given time step
k, which is known as the preview time (window) in the
literature [12]. This is actually a reasonable and acceptable assumption given that the desired trajectory is typically planned in advance and at least it can be assumed
practically to be known for n + nd time steps ahead. Our
proposed estimation scheme is now slightly modified to
incorporate this conditional change. A summary of the
procedure for implementation of our proposed scheme is
presented in Table 1.
h
iT
Let us now define Yd (k) as yd (k)T . . . yd (k + n)T ,
where Yd (k) is assumed to be a known signal. It is now
utilized to derive the unknown input observer following
(1)
equation (18) to yield x̂1 (k) as follows,
(
η(k + 1) = Âη(k) + F Yd (k)
(42)
(1)
(1)−1
x̂1 (k) = Mq
η(k)
(1)
(41)
An estimate of x2 (k) is now given by,
(1)
PROOF. It follows readily from equations (26) and
(1) (1)
(39) that we have u(k) − û(k) = D† (y(k) − C1 x1 (k) −
(1)
x̂2 (k) = Ãnz d x̄20 (k+nd )−
nX
d −1
(1)
(Ãz )i B̃z Θ̂1 (k+nd −i−1)
i=0
(43)
(1)
where x̄20 (k + nd ) is a random initial condition of the
(1)
(1)
FIR filter at each time step k+nd , and Θ̂1 (k) = X̂1 (k)
(1)
(1)
or Θ̂1 (k) = X̂1d (k), if B (1) or D is" full column#rank,
(1)
x̂1 (k + 1)
(1)
respectively. Moreover, X̂1 (k) =
and
(1)
x̂1 (k)
#
"
(1)
x̂1 (k)
(1)
(1)
X̂1d (k) =
. If B1 is full column rank, then
y(k)
û(k) is given by,
(1)
(1) (1)
(1) (1)
(1)†
x̂1 (k + 1) − A11 x̂1 (k) − A12 x̂2 (k)
û(k) = B1
Table 1
Inversion-based output tracking algorithm.
(1) Calculate Â, F and M from Theorem 2.
(1)
(2) Calculate T(1) and Mq from equations given
in Section 3.2.
(3) Calculate A(1) , B (1) and C (1) by applying the
similarity transformation to the system S using
the matrix T(1) (x(1) = T(1) x).
(4) Partition A(1) , B (1) and C (1) according to equation (20).
(1)
(5) If B1 is full column rank, then obtain Az and
(1)
Bz from equations (23) and (24). If B1 is not
full column rank and D is full column rank, then
obtain Azd and Bzd from equations (28) and
(29).
(6) Calculate Ãz and B̃z from equations (32) and
(33).
(7) Select nd according to Theorem 8 to meet the
desired estimation error specifications .
(8) At each time step k,
(1)
(a) Reconstruct x̂1 (k) from equation (42).
(1)
(b) Reconstruct x̂2 (k) using equation (43).
(1)
(c) If B1 is full column rank, then reconstruct
(1)
û(k) from equation (44). If B1 is not full
column rank and D is full column rank, then
reconstruct û(k) from equation (45).
(44)
and if D is full column rank, then û(k) is given by,
(1) (1)
(1) (1)
(45)
û(k) = D† y(k) − C1 x̂1 (k) − C2 x̂2 (k)
Since the NMP state estimation scheme is subject to errors, if the computed û(k) is fed to the system, it will then
generate y(k) that is different from the desired yd (k). In
other words, y(k) is the real output of the system subjected to and stimulated by û(k), that is (in view of the
state space representation (16))
(
x̃(1) (k + 1) = A(1) x̃(1) (k) + B (1) û(k)
(46)
y(k) = C (1) x̃(1) (k) + Dû(k)
where x̃(1) (k) denotes the state response of the system
to the input û(k). If the exact u(k) is known, then we
would have obtained,
(
x(1) (k + 1) = A(1) x(1) (k) + B (1) u(k)
(47)
yd (k) = C (1) x(1) (k) + Du(k)
We are now in a position to define the output tracking
error as follows.
Definition 5 The output tracking error is defined as
ey (k) = y(k) − yd (k).
It now follows from equations (46) and (47) that,
(
ẽx (k + 1) = A(1) ẽx (k) + B (1) eu (k)
(48)
(1)
ey (k) = C (1) ẽx (k) + Deu (k)
where ẽx (k) = x̃(1) (k) − x(1) (k). It is straightforward to
conclude from equation (48) that ey (k) → 0 as k → ∞
if eu (k) → 0 as k → ∞. However, eu (k) is given by the
NMP state estimation error (ex2 (k)) that is multiplied
by a gain as formally stated in Propositions 1 or 2. We
have shown in Theorem 7 that the NMP state estimation error (ex2 (k)) decays asymptotically as nd increases.
Hence, an almost perfect output tracking for any desired
trajectory can be achieved by selecting nd to be sufficiently large by as much as few times of the system order in most cases. The following theorem formally establishes the above statement and provide an upper bound
on the output tracking error versus the delay parameter
nd .
Theorem 8 Let Assumption 1 hold and 0 < q <
(1)
n. If B1 is full column rank, then key (k)k2 ≤
(1)†
(1)
σmax (Ãnz d )kC(zI − A)−1 B + Dk∞ kB1 A12 k2 k(zI −
A(1) )−1 B (1) k∞ . On the other hand, if D is full column
rank, then key (k)k2 ≤ σmax (Ãnz d )kC(zI − A)−1 B +
(1)
Dk∞ kD† C2 k2 k(zI − A(1) )−1 B (1) k∞ .
PROOF. According to equation (48), key (k)k2 =
(1)
kC(zI − A)−1 B + Dk∞ keu (k)k2 . If B1 is full column rank, then from Proposition 1, keu (k)k2 =
(1)†
(1)
(1)†
(1)
kB1 A12 ex2 (k)k2 ≤ kB1 A12 k2 kex2 (k)k2 . Our desired result is now obtained if we substitute kex2 (k)k2 ,
by using Theorem 7, into the above expression as
σmax (Ãnz d )k(zI − A(1) )−1 B (1) k∞ . Following along the
same procedure yields our other desired result for the
case when D is full column rank.
As expected, Theorem 8 implies that the upper bound
of the output tracking error has the same functionality
in terms of the delay parameter nd as that of the upper
bound of the NMP states estimation error. Theorem 8 is
quite useful for performing a trade-off analysis between
the delay parameter nd and the output tracking error.
This completes our proposed methodology for inversionbased output tracking. In the next section, we show that
our proposed solution provides a framework for systematic treatment of the systems with transmission zeros on
the unit circle.
5
Systems having transmission zeros on the unit
circle
In this section, we show that our proposed approach can
be extended and applied for handling the output tracking problem in systems that have transmission zeros on
the unit circle in addition to MP and NMP zeros. This
problem has not been addressed and solved in the literature. For simplicity of the discussion, we only consider
a SISO system that is described by,
Y (z) = G(z)U (z) = (z + 1)G′ (z)U (z)
(49)
The function G(z) is given by the z-transform of the
system S that has a transmission zero at z = −1, and
therefore can be written as (z + 1)G′ (z), where all the
zeros and poles of G′ (z) are inside the unit circle. The
extension of our proposed solution to a general MIMO
system with non-minimum phase transmission zeros
and multiple zeros on the unit circle is straightforward
and not included here for simplicity. Our objective is
to determine u(k) such that y(k) follows the desired
and known trajectory yd (k). Note that if such a u(k)
is found, it will also satisfy Yd (z) = G(z)U (z), where
Yd (z) denotes the z-transform of yd (k).
Let us rewrite Yd (z) as Yd (z) = zYd′ (z) + Yd′ (z), where
Yd′ (z) = G′ (z)U (z). Clearly, Yd′ (z) is given by
Yd′ (z) = (z + 1)−1 Yd (z)
(50)
The exact computation of yd′ (k) = Z −1 {Yd′ (z)} requires
the initial condition yd′ (0), which is not known. If we
rewrite equation (50) in the time-domain, we have,
yd′ (k + 1) = −yd′ (k) + yd (k)
(51)
Therefore, we have yd′ (1) = −yd′ (0) + yd (0) at k = 0.
Although yd (0) is known, but yd′ (0) is unknown. On the
other hand, the pole of the system (51) is at -1, thus
the effect of unknown initial condition will not die out
overtime. If yd′ (0) was known, since U (z) satisfies both
Yd (z) = G(z)U (z) and Yd′ (z) = G′ (z)U (z), then one
would simply use the algorithm of Table 1 to compute
u(k) by using G′ (z) and the exact value of yd′ (k) instead
of G(z) and yd (k). Consequently, the problem of handling transmission zeros on the unit circle would be easily resolved. In fact, the Algorithm provided in Table 1,
similar to the other work, e.g. in [8] can generate an almost exact duplicate of the unknown states and inputs
that are biased by the value of the unknown initial condition. However, our objective here is to diminish the effects of unknown initial conditions ! so that the system
outputs converge to the desired trajectory starting from
an arbitrary initial condition.
Let us assume that the system S′ is specified by the
quadruple Σ′ := (A′ , B ′ , C ′ , D′ ) and goverend by
Z −1 (G′ (z)). Note that the system S′ is minimum phase.
Therefore, according to the equation (18), the estimate
of all the system states is given by,
X̂(z) = (zI − A)−1 F ZM Yd′ (z)
(52)
−1
−1
(1)
(1)
(1)
where A = Mq
ÂMq , F = Mq
F and
h
iT
ZM = z 0 . . . z n−1 , where z −1 denotes the delay
(1)
Mq ,
operator. The matrices
 and F are computed
from Theorem 2 by using the quadruple Σ′ . The transfer
function in equation (52) is stable since the poles of A
are the MP zeros of the system.
Let us introduce the controller H(z) having the order nc
that can be written as (z + 1)H ′ (z). It is necessary for
H(z) to have (z + 1) as a factor in order to cancel out
the pole of the inverse system at z = −1. Moreover, let
us assume that an approximation to the system S′ states
is given by X̃(z) = H(z)X̂(z). Therefore, according to
equation (38), an approximation Ũ (z) to the unknown
input is given by,
Ũ (z) = (B ′(1) )† (zI − A′(1) )H(z)X̂(z)
(53)
where B ′(1) = QB ′ and A′(1) = QA′ Q−1 . Combining
equations (50), (52) and (53) yield,
Ũ (z) = (B ′(1) )† (zI−A′(1) )H(z)(zI−A)−1 F ZM (z+1)−1 Yd (z)
(54)
By canceling out the pole at z = −1 by the (z + 1) factor
of H(z), one obtains,
Ũ (z) = (B ′(1) )† (zI − A′(1) )H ′ (z)(zI − A)−1 F ZM Yd (z)
= G̃i (z)Yd (z)
(55)
Equation (55) provides a direct and biased estimate of
the system input by using yd (k) with an additional delay
of nc . Theoretically, the unbiased estimate of the system
input is given by (H(z) = I),
Û (z) = (B ′(1) )† (zI − A′(1) )(zI − A)−1 F ZM (z + 1)−1 Yd (z)
= Ĝi (z)Yd (z)
(56)
Therefore, the output tracking error is approximated by,
Ec (z) = Ŷ (z) − Ỹ (z)
= G(z)(Ĝi (z) − G̃i (z))Yd (z)
(57)
The above result provides an important criterion for
selecting the controller H(z). Note that the governing
error dynamics (57) does not have a pole at z = −1 since
the pole of Ĝi (z) has been canceled out by the (z + 1)
factor of G(z). Therefore, the system (57) is stable.
Let us assume that for the particular case we considered
here the function H(z) is expressed as H(z)I, where H(z)
represents a SISO transfer function of a controller that
contains a factor (z + 1). Therefore, the error dynamics
(57) is given by,
Ec (z) = G(z)Gi (z)(1 − H(z))Yd (z)
(58)
where Gi (z) = (B ′(1) )† (zI − A′(1) )(zI − A)−1 F ZM .
Therefore, the design problem is reduced to that of
solving the following optimization problem,
min k1 − H(z)k
H(z)
(59)
which is also known as the H∞ (or H2 ) norm minimization problem.
Clearly, the above problem is associated with numerous trade-off considerations. For instance, one may not
be able to achieve a minimum value over all frequencies. One may directly introduce the controller H(z) by
defining an approximation of the system input given by
Ũ (z) = H(z)Û (z). The choice depends on the design
preference requirements. Using H(z) provides more degrees of freedom, however it also complicates the mini-
mization problem and the trade-off studies.
0.9
0.8
0.7
0.6
|e x |2
A significant advantage of our solution is derived from
the fact that both minimum and non-minimum phase
zeros will not be involved in the design process of H(z) to
handle the transmission zeros on the unit circle. In other
words, the controller H(z) should only cancel out the
transmission zeros on the unit circle, and therefore one
can always find such a controller. However, the actual
challenge is due to the fact that the controller should
also minimize the k1 − H(z)k over a desired range of
frequencies.
1
0.5
0.4
0.3
0.2
Let us further assume that the particular case we consider here has also a non-minimum phase transmission
zero. In this case, we exactly follow the algorithm that
is provided in Table 1 subject to two modifications,
namely: i) we use the quadruple Σ′ to determine the
algorithm parameters, and ii) we use the following relationship for estimation of the MP states instead of
equation (18), namely
X̃(z) = H ′ (z)(zI − A)−1 F ZM Yd (z)
Therefore, the inverse dynamics poles are the same as
the poles of H ′ (z) and A which are inside the unit circle.
Note that the order of A is less than n, therefore H ′ (z)
should be selected such that the above transfer function
becomes proper.
Finally, following along a similar approach allows one to
relax the condition of simplicity of the MP transmission
zeros in Theorem 2. As an illustration, let us assume that
the transfer function of a SISO system can be expressed
as G(z) = (z − p)2 G′′ (z), where kpk < 1 and G′′ (z) does
not have (z − p) as a factor in its denominator. Therefore, the algebraic multiplicity of z = p is two. In this
case, we exactly follow the algorithm that is provided in
Table 1 subject to two modifications, namely: i) we use
the quadruple Σ′ as described by Z −1 {(z − p)G′′ (z)} to
determine the algorithm parameters, and ii) we use the
signal yd′ (k) instead of yd (k) where yd′ (k) is governed
d (z)
by Yd′ (z) = Yz−p
. The above transfer function has an
arbitrary initial condition. Since kpk < 1, the effects
of the unknown initial condition will die out quickly. !
The solution also does not require the introduction of a
controller.
This now completes our proposed methodology for
handling transmission zeros on the unit circle. In the
next section, we will provide illustrative simulations to
demonstrate the merits of our proposed methodologies.
6
Numerical Case Studies Simulations
Case I: Consider the following discrete-time linear system,
(z − 1.5)(z − 0.5)
(60)
G(z) =
z2
or in its equivalent state space representation given by,
"
#
" #
0
0
0
x(k + 1) =
x(k) +
u(k)
1 0
1
(61)
h
i
y(k) = −2 0.75 x(k) + u(k)
0.1
0
0
2
4
6
8
10
12
14
16
18
20
nd
Fig. 1. Upper bound of the NMP state estimation error versus
nd .
Using Lemma 1, the solution hto the conditions
(i)-(iii)
i
is given by  = 0.5, F = −0.5547 0 , and M =
h
i
−0.5574 0.8321 . Therefore, the unknown input observer is now given by equation (18),
(
η(k + 1) = 0.5η(k) − 0.5574y(k)
(62)
(1)
x̂1 (k) = η(k)
Moreover, we have from the LQ decomposition of M and
equations (28) and (29), the following
"
#
h
i
−0.5547 0.8321
(1)
T =
;L = 1 0
0.8321 0.5547
h
i
Azd = 1.5; Bzd = −1.5 −1 .
The upper bound for the state estimation error versus nd is shown in Figure 1. We have applied a
non-smooth random input to the system in order to
illustrate and demonstrate the effects of nd on the estimation error. A smooth input, as stated in Remark
4, will be estimated in an almost unbiased manner for
(1)
any nd ≥ 2. Figure 2 depicts that x1 (1) is perfectly
estimated by using the unknown input observer (UIO)
as expected. Figure 3 shows that a perfect reconstruc(1)
tion can be achieved for x2 (1) by selecting nd = 15, as
expected from Figure 1. According to the Proposition
1, the unknown input should also be almost perfectly
reconstructed with nd = 15, which is also verified in
Figure 4.
Case II: In another simulation case study, consider that
a non-smooth yd (k) is required to be followed. The unknown input is reconstructed by using the Algorithm
that is detailed in Table 1, and the results are depicted in
Figure 5. This figure demonstrates that an almost perfect output tracking is achieved by selecting nd = 15.
Finally, consider a smooth yd (k) as given by yd (k) =
k 2 sin(5πk). The output tracking result for this smooth
desired trajectory (not shown due to space limitations)
confirms and validates the statements made in Remark
4.
Case III: To provide a comparative study, consider a
MIMO system that is taken from the reference [14] with
0.6
1
true state
estimated state
0.8
0.4
0.6
0.3
0.4
0.2
0.2
0.1
0
y
x (1)
1
0.5
0
-0.2
-0.1
-0.4
-0.2
-0.6
-0.3
-0.8
-0.4
-1
0
5
10
15
20
25
30
35
40
45
50
yd
tracking performance with n d =5
tracking performance with n d =15
0
5
time step
true state
estimated state with n d =5
1.2
estimated state with n d =15
1
x (1)
2
0.8
0.6
n d =15
n d =5
0.4
0.2
0
-20
-10
0
10
20
30
40
50
time step
Fig. 3. The estimation of the NMP state (the graphs are
shifted by nd − n time steps to the left for the purpose of
comparison) by utilizing the filter (36).
1.2
true input
estimated input with n d =5
1
estimated input with n d =15
20
25
30
35
40
45
50
Fig. 5. The output tracking performance corresponding to
different values of nd by utilizing equations (42), (43) and
(45).
The system (63) has two zeros at z1 = 0.6072 and
z1 = 1.9928. Therefore, it has three MP states and one
NMP state. Authors of [14] proposed a geometric approach and applied it to the system (63) to achieve an
almost perfect estimation of the states and unknown
inputs with a delay of 20 time steps (nd = 20). For comparison, our simulation results for the same example is
shown in Figure 6 (the numerical values of the estimation filter parameters are given in Appendix H), which
demonstrate that by using our proposed methodology
the unknown states and inputs are almost perfectly reconstructed with only a delay of nd = 10, which is half
of the delay that was used in [14]. Moreover, as shown
in Figure 6b, by using our approach the three MP states
of the system are estimated without any delay when the
transient response due to the unknown initial condition
dies out quick! ly. This is in contrast to the delayed results that are shown in the work [14].
However, the most important contribution of our work
over that in [14] is derived from the fact that our
methodology unlike the one in [14] can handle transmission zeros on the unit circle as illustrated in the next
case study.
Case IV: Consider the following system,
0.8
u
15
time step
Fig. 2. The estimation of the MP state by utilizing the filter
(18).
1.4
10
0.6
G(z) =
(z + 1)(z + 3)(z + 0.5)(z − 0.5)
z 2 (z 2 − z + 0.5)
(64)
0.4
The above system has both MP and NMP transmission
zeros as well as a zero on the unit circle. We follow the
procedure that was introduced in Section 5 for designing an inversion-based output tracking controller. We
selected the controller H(z) having the structure,
0.2
0
-20
-10
0
10
20
30
40
50
time step
Fig. 4. The estimation of the unknown input (the graphs are
shifted by nd − n time steps to the left for the purpose of
comparison) by utilizing equation (39).
A ∈ R4×4 , B ∈ R4×2 and C ∈ R2×4 as follows,
0.6 −0.3
0.1
1
x(k + 1) = −0.4 −1.5
0.3 1.1
"
#
1
2
3
4
y(k) =
x(k)
2 1 5 6
0
0
0
0.4
0
u(k)
x(k) + 0
0.4 −0.3
0 −0.1
0.1 0.1
0.2 0.9
0
0
(63)
H(z) =
z+1
I
2z
(65)
The numerical values for the other parameters are given
in Appendix I. Figure 7 shows the output tracking performance of our proposed solution with nc = 1 and
nd = 10. The result demonstrates the significant advantage of our proposed solution for handling all types of
transmission zeros within a single framework. Specifically, Figure 7 shows that the desired trajectory is approximately followed by an error that is governed by
equation (57). On the other hand, the proposed methodology in [14] essentially fails under this case. Note that
a
20
yd
x (1)
(1)
1
15
estimated state
true state
0.5
tracking performance
10
0
5
-0.5
0
10
20
30
40
50
60
70
0
y
x (1)
(2)
1
1
-5
0
-10
-1
0
10
20
30
40
50
60
70
-15
x (1)
(3)
1
5
-20
0
-25
-5
0
10
20
30
40
50
60
70
time step
-30
0
20
40
60
80
100
120
time step
b
2
Fig. 7. The output tracking performance for the system (64).
estimated state
true state
1.5
1
x (1)
2
0.5
0
-0.5
-1
-1.5
-2
-10
0
10
20
30
40
50
60
40
50
60
time step
c
4
estimated input
true input
u(1)
2
0
values of Γ = (A − BIn D+
n Cn ) may or may not coincide
with the transmission zeros of the system. Also, it may or
may not have the same characteristics, namely the MP
transmission zero of S remains the stable eigenvalue of Γ.
However, if one determines the matrices Â, F and M by
using a different method for these systems, then the remainder of our procedur! e for unknown state and input
reconstruction, as described in this paper, will remain
applicable and unchanged. We have also demonstrated
that our proposed methods can provide an almost perfect tracking of any desired output trajectory by using
data and information that correspond to a small preview
time. An important contribution of our methodology is
the fact that we have provided a single framework that
can handle the problem of output tracking for systems
that have transmission zeros on the unit circle in addition to MP and NMP zeros. However, further research
is required to address issues of robustness and tracking
error performance in presence of disturbances and modeling uncertainties. These issues are left as topics of future research.
-2
-4
-10
0
10
20
30
References
4
u(2)
2
0
-2
-4
-10
0
10
20
30
40
50
60
time step
Fig. 6. Simulation results for the MIMO system (63) taken
from [14], (a) The MP state estimates, (b) The NMP state
estimates, and (c) The unknown input estimates.
the H(z) that is selected in equation (65) can be used
for all SISO systems that have a transmission zero at
z = −1, in addition to MP and NMP zeros.
7 Conclusion
In this paper, we have shown that one can almost perfectly estimate and reconstruct the unknown state and
inputs of a system if i) the system S is square, and ii)
(1)
B1 or D is full column rank. Non-square systems rarely
have transmission zeros [23], and therefore it is straightforward to design an unknown input observer (UIO) to
estimate all the system states. We excluded non-square
systems from our analysis since Theorem 1 is not guaranteed for this class of systems. In other words, the eigen-
[1] R. Brockett and M. Mesarovic, “The reproducibility of
multivariable systems,” Journal of Mathematical Analysis
and Applications, vol. 11, pp. 548–563, 1965.
[2] L. Silverman, “Inversion of multivariable linear systems,”
IEEE Transactions on Automatic Control, vol. 14, pp. 270 –
276, jun 1969.
[3] J. Massey and M. Sain, “Inverses of linear sequential
circuits,” IEEE Transactions on Computers, vol. C-17,
pp. 330 – 337, april 1968.
[4] P. Moylan, “Stable inversion of linear systems,” IEEE
Transactions on Automatic Control, vol. 22, pp. 74 – 78, feb
1977.
[5] S. Gilijns, Kalman filtering techniques for system inversion
and data assimilation. PhD thesis, K.U.Leuven, Leuven,
Belgium, 2007.
[6] R. A. Chavan and H. J. Palanthandalam-Madapusi, “Delayed
recursive state and input reconstruction,” arXiv preprint
arXiv:1509.06226, 2015.
[7] H.
J.
Palanthandalam-Madapusi
and D. S. Bernstein, “Unbiased minimum-variance filtering
for input reconstruction,” in American Control Conference,
2007. ACC’07, pp. 5712–5717, 2007.
[8] S. Kirtikar, H. Palanthandalam-Madapusi, E. Zattoni,
and D. S. Bernstein, “l-delay input and initial-state
reconstruction for discrete-time linear systems,” Circuits,
Systems, and Signal Processing, vol. 30, no. 1, pp. 233–262,
2011.
[9] Y. Xiong and M. Saif, “Unknown disturbance inputs
estimation based on a state functional observer design,”
Automatica, vol. 39, no. 8, pp. 1389–1398, 2003.
[10] S. Wahls and H. Boche, “Novel system inversion algorithm
with application to oversampled perfect reconstruction filter
banks,” IEEE Transactions on Signal Processing, vol. 58,
pp. 3008–3016, June 2010.
[11] T. Floquet and J.-P. Barbot, “A sliding mode approach of
unknown input observers for linear systems,” in Decision
and Control, 2004. CDC. 43rd IEEE Conference on, vol. 2,
pp. 1724–1729, 2004.
[12] Q. Zou and S. Devasia, “Preview-based stable-inversion for
output tracking of linear systems,” Journal of dynamic
systems, measurement, and control, vol. 121, no. 4, pp. 625–
630, 1999.
[13] K. George, M. Verhaegen, and J. M. Scherpen, “Stable
inversion of mimo linear discrete time nonminimum phase
systems,” in Proc. 7th Mediterranean Conference on Control
and Automation, pp. 267–281, 1999.
[14] G. Marro and E. Zattoni, “Unknown-state, unknown-input
reconstruction in discrete-time nonminimum-phase systems:
Geometric methods,” Automatica, vol. 46, no. 5, pp. 815 –
822, 2010.
[15] G. Marro, E. Zattoni, and D. S. Bernstein, “Geometric insight
and structure algorithms for unknown-state, unknown-input
reconstruction in linear multivariable systems,” {IFAC}
Proceedings Volumes, vol. 44, no. 1, pp. 11320 – 11325, 2011.
18th {IFAC} World Congress.
[16] Q. Zou and S. Devasia, “Preview-based optimal inversion
for output tracking: application to scanning tunneling
microscopy,” IEEE Transactions on Control Systems
Technology, vol. 12, no. 3, pp. 375–386, 2004.
[17] Q. Zou and S. Devasia, “Precision preview-based stableinversion for nonlinear nonminimum-phase systems: The vtol
example,” Automatica, vol. 43, no. 1, pp. 117–127, 2007.
[18] Q. Zou, “Optimal preview-based stable-inversion for output
tracking of nonminimum-phase linear systems,” Automatica,
vol. 45, no. 1, pp. 230–237, 2009.
[19] H. Wang, K. Kim, and Q. Zou, “B-spline-decompositionbased output tracking with preview for nonminimum-phase
linear systems,” Automatica, vol. 49, no. 5, pp. 1295–1303,
2013.
[20] B. Kiumarsi, F. L. Lewis, H. Modares, A. Karimpour, and
M.-B. Naghibi-Sistani, “Reinforcement q-learning for optimal
tracking control of linear discrete-time systems with unknown
dynamics,” Automatica, vol. 50, no. 4, pp. 1167–1175, 2014.
[21] M. Duan, K. S. Ramani, and C. E. Okwudire, “Tracking
control of non-minimum phase systems using filtered basis
functions: A nurbs-based approach,” in ASME 2015 Dynamic
Systems and Control Conference, pp. V001T03A006–
V001T03A006, American Society of Mechanical Engineers,
2015.
[22] S. Sundaram and C. N. Hadjicostis, “Delayed observers for
linear systems with unknown inputs,” IEEE Transactions on
Automatic Control, vol. 52, pp. 334–339, Feb 2007.
[23] E. Davison and S. Wang, “Properties and calculation
of transmission zeros of linear multivariable systems,”
Automatica, vol. 10, no. 6, pp. 643 – 658, 1974.
[24] D. Carlson, E. Haynsworth, and T. Markham, “A
generalization of the schur complement by means of
the moorepenrose inverse,” SIAM Journal on Applied
Mathematics, vol. 26, no. 1, pp. 169–175, 1974.
A
Proof of Theorem 1
The eigenvalues of A − BIn D+
n Cn are obtained by solving |zI−A+BIn D+
n Cn | = 0. If the system is square, then
D+
n is a nonzero square matrix. Therefore, one can equiv+
alently solve the equation |D+
n | |zI−A+BIn Dn Cn | = 0.
On the other hand from the Schur identity [24], we have,
#
"
zI − A −BIn
+
+
Dn |zI−A+BInDn Cn | =
(A.1)
Cn
Dn
Let us partition Cn and Dn as follows,
Cn =
C
!
CA
C
=
..
C−
.
n−1
CA
(A.2)
0
0
D 0
Dn =
..
−
−
D21
D22
.
D
(A.3)
Then, the right hand side of equation (A.1) can be partitioned as,
#
"
zI − A −B 0
zI − A −BIp
(A.4)
= C
D 0
Cn
Dn
−
−
C − D21
D22
!
D
0
D
CB
=
..
..
.
.
n−1
n−2
CA
B CA
B
...
...
..
.
...
−
Thus, if D22
is full row rank, then according to the Schur
identity, equation |zI− A+ BIn D"+
n Cn | = 0 has only
# one
zI − A −BIp
set of solution that is given by
= 0,
Cn
Dn
which is exactly the transmission zeros of the system
−
S. hHowever, if Di22
is rank deficient, then certain rows
−
of C − D21
D− are linearly dependent on the rows of
h
i 22
−A −B 0 . Hence, z = 0 is also a solution. On the
other hand, since |zI − A + BIn D+
n Cn | = 0 must have
n eigenvalues, if the system S has p transmission zeros,
then z = 0 is a solution of multiplicity n − p.
B
Proof of Theorem 2
Since the system S has α1 transmission zeros having an
algebraic multiplicity of 1, therefore Γ has α1 linearly
independent eigenvectors. Therefore, M0 has at least α1
linearly independent rows. On the other hand, the set Z
(as defined in Theorem 1) has αz zeros, where αz = n −
β − α1 . Therefore, I − Dn D+
n has αz independent rows.
This implies that M# has αz linearly independent rows.
Therefore, M has αz + α1 = n − β linearly independent
rows.
C
Proof of Lemma 2
Since the system S has at least one NMP zero (q < n),
then by the definition of transmission zeros, there exists
a nonzero u(k) that yields a zero output (y(k) = 0 for all
(1)
k). On the other hand, according to Theorem 3, x1 (k)
approaches to zero when y(k) = 0 for k = k0 , k0 + 1, . . ..
Therefore, from the first and third equations of (19), we
have for k → ∞,
#"
#
" # "
(1)
(1)
C2
D
0
x2 (k − n)
.
(C.1)
=
(1)
(1)
0
u(k − n)
A12 B1
"
#
(1)
x2 (k − n)
Since
is nonzero, it implies that the
u(k − n)
"
#
(1)
C2
D
columns of
are linearly dependent.
(1)
(1)
A12 B1
D
Proof of Lemma 3
(1)
First note that A11 in equation (19) is a Hurwitz matrix,
(1)
otherwise x1 (k) → ∞ as k → ∞. Next, consider,
(
(1)
(1) (1)
(1)
x2 (k − n + 1) = A22 x2 (k − n) + B2 u(k − n)
(z1)
S
:
(1) (1)
(1)
ξ(k − n) = A12 x2 (k − n) + B1 u(k − n)
(D.1)
If there exists a nonzero u(k) that yields ξ(k) = 0, then
this implies that from the first equation of (19) we have,
(1)
x1 (k) → 0 as k → ∞. Therefore, y(k) → 0 as k →
∞ according to Theorem 3. Therefore, the transmission
zeros of S (z1) are also the transmission zeros of S (1) .
E
Proof of Lemma 4
Consider the following system,
(1)
(1) (1)
(1)
x2 (k − n + 1) = A22 "x2 (k − n) +#B2 u(k − n)
h
i x(1) (k − n)
S (z2) :
(1)
2
ξ(k − n) = C2 D
u(k − n)
(E.1)
If there
exists
a
nonzero
u(k)
that
yields
ξ(k)
=
0,
then
h
i
since I C1 is full row rank, the third equation of (19)
(1)
yields x1 (k) = 0, and y(k) = 0. Therefore, the transmission zeros of S (z2) are also the transmission zeros of
S (1) .
F
Proof of Theorem 4
Note that Cz2 = 0 is an immediate result of the Schur
identity [24] and Lemma 2. The eigenvalues
of#Az are a
"
(1)
(1)
A22 B2
, which
subset of the transmission zeros of
(1)
(1)
A12 B1
are a subset of the system S zeros according to Lemma
3. According to Theorem 3 and Theorem 4 (Cz2 = 0),
the output of the system (22) goes to zero as k → ∞
(1)
(1)
if and only if x1 (k), and consequently, X1 (k) goes to
zero as k → ∞. The first equation of (22) implies that if
(1)
Az is a Hurwitz matrix, then x2 (k) must approach to
(1)
zero when X1 (k) is zero. However, we know that there
(1)
exists nonzero x2 (k) and u(k) that yield a zero y(k)
for all k. Therefore, since the response of an unforced
line! ar system can approach to zero or infinity (recall
we excluded systems with transmission zeros on the unit
(1)
circle in Assumption 1), therefore x2 (k) must approach
to infinity. This implies that the eigenvalues of Az are
the NMP zeros of S.
G
Proof of Theorem 5
e NMP zeros of S. Proof follows along the same lines
as those for Theorem 4 using Lemma 4 and is therefore
omitted for brevity.
H
Numerical values for estimation filter parameters for the system (63)
0.6072 0 0
 = 0
0 0 , Az = 1.9928
0
0 0
0 0 −0.662 0.0184
F = 0 0 −0.0206 0.1365
0 0 0.1337 0.0338
0.0488 0.9650 0.2063 −0.1547
M = 0.2523 0.0953 0.6205 0.7364
0.2013 0.3012 0.5700 0.7375
−0.0488 −0.9650 −0.2063 0.1547
0.2483 −0.0190 0.6003 0.7600
T(1) =
−0.4645 0.2474 −0.5833 0.6187
−0.8487 −0.0855 0.5067 −0.1252
−1.0000
0
0
0
L = −0.1183 0.9930
0
0
−0.3039 0.9469 0.1048 0
h
i
Bz = −0.2463 −2.0822 2.4171 5.3035 −0.0678 −2.3507
I Numerical values for estimation filter parameters for the system (64)
0 1.042 0 0
0 0
0
 = 0 0.5 0 , F = 0 −0.4119 0 0
0 0.4823 0 0
0 0 −0.5
0.368 −0.368 0.792 −0.317
M= 0
0.395 −0.852 0.341
0 −0.395 −0.852 0.341
−0.3680 0.3680 −0.7928 0.3171
−0.9298 −0.1456 0.3138 −0.1255
T(1) =
0
−0.9184 −0.3674 0.1470
0
0
0.3714
0.9285
−1.0000
0
0
0
L = 0.9298 −0.3680
0
0 , Az = 3
0.6386 −0.2527 0.7269 0
h
i
Bz = 2.1348 −0.8448 0.9894 −0.2135 0.0845 −0.0989
| 3cs.SY
|
Contextuality in multipartite pseudo-telepathy
graph games
Anurag Anshu1 , Peter Høyer2 , Mehdi Mhalla3 , and Simon Perdrix4
arXiv:1609.09689v2 [quant-ph] 24 Jun 2017
1
3
Centre for Quantum Technologies, National University of Singapore
2
University of Calgary, Canada
Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG, F-38000 Grenoble France
4
CNRS, LORIA, Université de Lorraine, Nancy, France
Abstract. Analyzing pseudo-telepathy graph games, we propose a way
to build contextuality scenarios exhibiting the quantum supremacy using
graph states. We consider the combinatorial structures that generate
equivalent scenarios. We introduce a new tool called multipartiteness
width to investigate which scenarios are harder to decompose and show
that there exist graphs generating scenarios with a linear multipartiteness
width.
1
Introduction
Contextuality is an active area of research that describes models of correlations
and interpretations, and links to some fundamental questions about the natural
world. It also provides a framework where one can utilize the understanding
of quantum mechanics (and quantum information) in order to better analyze,
understand, and interpret macroscopic phenomena [5, 11, 12, 22, 30].
The theoretical and experimental study of quantum world has proven that a
scenario involving many parties (each having access to a local information) can
contain correlations that do not possess any classical interpretation that relies
on decomposition of these correlations using local functions. Contextuality can
be viewed as a tool to describe the combinatorial structures present in these
correlations.
A model of contextuality scenario is due to Abramsky and Brandenburger [1]
and uses sheaf theory to naturally translate the consistency of interpretation
by the pre-sheaf structure obtained by a distribution functor on the sheaf of
events. The authors introduce three levels of contextuality: (i) Probabilistic
contextuality, which corresponds to the possibility of simulating locally and
classically a probability distribution. It extends the celebrated Bell’s theorem [7]
which shows that quantum probabilities are inconsistent with the predictions of
any local realistic theory; (ii) Logical contextuality or possibilistic contextuality,
which extends Hardy’s construction [20] and considers only the support of a
probability distribution; (iii) Strong contextuality, which extends the properties
of the GHZ state [16] and relies on the existence of a global assignment consistent
with the support.
More recently Acı́n, Fritz, Leverrier, and Belén Sainz [3] have presented
contextuality scenarios defined as hypergraphs, in which vertices are called
outcomes and hyperedges are called measurements. A general interpretation
model is an assignment of non negative reals to the vertices that can be interpreted
as a probability distribution for any hyperedge (weights of the vertices of each
hyperedge sum to 1). Each hypergraph H admits a set C(H) (resp. Q(H), G(H))
of classical (resp. quantum, general probabilistic) models with C(H) ⊆ Q(H) ⊆
G(H).
They have shown that the Foulis Randall product of hypergraphs [13] allows
one to describe the set of no-signaling models in product scenarios G(H1 ⊗ H2 ).
They have also investigated the multipartite case, showing that the different
products for composition produce models that are observationally equivalent.
A particular case of contextuality scenarios is the pseudo-telepathy games [9],
which are games that can be won by non-communicating players that share
quantum resources, but cannot be won classically without communication. A
family of pseudo-telepathy games based on graph states have been introduced
in [4]. The pseudo-telepathy game associated with a graph G of order n (on n
vertices), is a collaborative n-player game where each player receives a binary
input (question) and is asked to provide, without communication, a binary output
(answer). Some global pairs of (answers—questions) are forbidden and correspond
to losing positions. Given such a scenario, to quantify its multipartiteness, we
define the multipartiteness width: a model on n parties has a multipartiteness
width less than k if it has an interpretation (assignment of real positive numbers to
the vertices) that can be obtained using as ressources interpretations of contextual
scenarios on less than k parties.
It has been shown in [10] that even though GHZ type scenarios are maximally
non local (strongly contextual), they can be won with 2 partite nonlocal boxes. So
the multipartiteness width is different from the usual measures of contextuality
[2, 15]. However, it has potential application for producing device independent
witnesses for entanglement depth [25].
In section 2, we define the graph pseudo-telepathy games, investigate in detail
the quantum strategy and link them to contextuality scenarios. We show in section
3 that provided that the players share multipartite randomness, it is enough
to surely win the associated pseudo-telepathy game, in order to simulate the
associated quantum probability distribution. In section 4, we prove that graphs
obtained by a combinatorial graph transformation called pivoting correspond
to equivalent games. Finally, we prove that there exist graphs for which the
multipartiteness width is linear in the number of players, improving upon the
previous logarithmic bound given in [4].
Note that even though the rules of these graph games appear non-trivial, they
naturally correspond to the correlations present in outcomes of a quantum process
that performs X and Z measurements on a graph state. Thus, they might be
easy to produce empirically. Furthermore even if the space of events is quite large,
the scenarios have the advantage of possessing concise descriptions, quite similar
to the separating scenarios using Johnson graphs in [14]. Requiring such large
structures to achieve possibilistic contextuality for quantum scenarios seems to
be unavoidable. Indeed, it has been shown that multiparty XOR type inequalities
involving two-body correlation functions cannot achieve pseudo-telepathy [17].
2
Pseudo-telepathy graph games, multipartiteness and
contextuality scenarios
Graph notations. We consider finite simple undirected graphs. Let G = (V, E)
be a graph. For any vertex u ∈ V , NG (u) = {v ∈ V | (u, v) ∈ E} is the
neighborhood of u. For any D ⊆ V , the odd neighborhood of D is the set
of all vertices which are oddly connected to D in G: Odd(D) = {v ∈ V :
|D ∩ N (v)| = 1 mod 2}. Even(D) = V \ Odd(D) is the even neighborhood of D,
and loc(D) = D ∪ Odd(D) is the local set of D which consists of the vertices
in D and those oddly connected to D. For any D ⊆ V , G[D] = (D, E ∩ D×D)
is the subgraph induced by D, and |G[D]| its size, i.e. the number of edges
of G[D]. Note that Odd can be realized as linear map (where we consider
subsets as binary vectors), which implies that for any two subset of vertices A, B,
Odd(A ⊕ B) = Odd(A) ⊕ Odd(B) where ⊕ denotes the symmetric difference.
We introduce the notion of involvement:
Definition 1 (Involvement). Given a graph G = (V, E), a set D ⊆ V of
vertices is involved in a binary labelling x ∈ {0, 1}V of the vertices if D ⊆
supp(x) ⊆ Even(D), where supp(x) = {u ∈ V, xu = 1}.
In other words, D is involved in the binary labelling x, if all the vertices in D
are labelled with 1 and all the vertices in Odd(D) are labelled with 0. Notice
that when G[D] is not a union of Eulerian graphs1 , there is no binary labelling
in which D is involved. On the other hand, if G[D] is a union of Eulerian graphs,
there are 2|Even(D)|−|D| binary labellings in which D is involved.
Collaborative games. A multipartite collaborative game G for a set V of
players is a scenario characterised by a set L ⊆ {0, 1}V × {0, 1}V of losing pairs:
each player u is asked a binary question xu and has to produce a binary answer au .
The collaborative game is won by the players if for a given question x ∈ {0, 1}V
they produce an answer a ∈ {0, 1}V such that the pair formed by a and x,
denoted (a|x), is not a losing pair, i.e. (a|x) ∈
/ L.
A game is pseudo-telepathic if classical players using classical resources cannot
perfectly win the game (unless they cheat by exchanging messages after receiving
the questions) whereas using entangled states as quantum resources the players
can perfectly win the game, giving the impression to a quantum non believer
that they are telepathic (as the only classical explanation to a perfect winning
strategy is that they are communicating).
Example 1: ThePlosing set associatedP
with thePMermin parity game [26] is
LMermin = {(a|x) :
xi = 0 mod 2 and
ai + ( xi )/2 = 1 mod 2}. Notice
1
The following three properties are equivalent: (i) D ⊆ Even(D); (ii) every vertex
of G[D] has an even degree; (iii) G[D] is a union of Eulerian graphs. Notice that
D ⊆ Even(D) does not imply that G[D] is Eulerian as it may not be connected.
that the losing set admits the following simpler description: LMermin = {(a|x) :
2|a| = |x| + 2 mod 4}, where |x| = |supp(x)| is the Hamming weight of x.
Collaborative graph games MCG(G): A multipartite collaborative game
MCG(G) associated with a graph G = (V, E), where V is a set of players, is the
collaborative
P game where the set of losing pairs is LG := {(a|x) : ∃D involved
in x s.t. u∈loc(D) au = |G[D]| + 1 mod 2}. In other words, the collaborative
game is won by the players if for a given question x ∈ {0, 1}V they
P produce an
answer a ∈ {0, 1}V such that for any non-empty D involved in x, u∈loc(D) au =
|G[D]| mod 2.
Example 2: Consider MCG(Kn ) the collaborative game associated with
the complete graph Kn of order n. When a question x contains an even number
of 1s the players trivially win since there is no non-empty subset of vertices
involved in such a question. When x has an odd number of 1s, the set of players
(vertices) involved in this question is D = supp(x). In this case, all the players
are either in D or Odd(D) thus the sum of all the answers has to be equal to
|G[D]| = |D|(|D|−1)
= |D|−1
mod 2. Thus for the complete graph Kn , LKn =
2
2
{(a|x) : |x| = 1 mod 2 and |a| = |x|−1
2 +1 mod 2} = {(a|x) : 2|a| = |x|+1 mod 4}.
Note that for this particular graph, the constraints are global in the sense that
the sum of the answers of all the players is used for all the questions. Notice also
that the set of losing pairs LKn = {(a|x) : 2|a| = |x| + 1 mod 4} is similar to the
one of the Mermin parity game, LMermin = {(a|x) : 2|a| = |x| + 2 mod 4}. In
section 4, we actually show the two games simulate each other.
Quantum strategy (Qstrat): In the following we show that for any graph G,
the corresponding multipartite collaborative game can be won by the players if
they share a particular quantum state.
P More precisely the state they share is the
so-called graph state |Gi = √ 1|V | y∈{0,1}V (−1)|G[supp(y)]| |yi, and they apply
2
the following strategy: every player u measures his qubit according to X if xu = 1
or according to Z if xu = 0. Every player answers the outcome au ∈ {0, 1} of
this measurement.
This quantum strategy QStrat, not only produces correct answers, but
provides all the good answers uniformly:
Lemma 1. Given a graph G = (V, E) and question x ∈ {0, 1}V , the probability
p(a|x) to observe the outcome a ∈ {0, 1}V when each qubit u of a graph state |Gi
is measured according to Z if xu = 0 or according to X if xu = 1 satisfies:
(
p(a|x) =
0
|{D involved in x}|
2|V |
if (a|x) ∈ L
otherwise.
Proof. According to the Born rule, the probability to get the answer a ∈ {0, 1}V
to a given question x ∈ {0, 1}V is:
av
O
O I + (−1)au Xu
I
+
(−1)
Z
v
⊗
|Gi
p(a|x) = hG|
2
2
v∈V \supp(x)
u∈supp(x)
P
1 X
(−1) u∈D au hG| ZD\supp(x) XD∩supp(x) |Gi
= n
2
D⊆V
The basic property which makes this strategy work is that for any u ∈ V ,
Xu |Gi = ZN (u) |Gi. As a consequence, since X and Z anti-commute and X 2 =
Z 2 = I, for any D ⊆ V , XD |Gi = (−1)|G[D]| ZOdd(D) |Gi. Thus,
P
1 X
(−1)|G[D∩supp(x)]|+ u∈D au hG| Z(Odd(D∩supp(x)))⊕(D∩\supp(x)) |Gi
n
2
p(a|x) =
D⊆V
(
1 if C = ∅
,
Where ⊕ denotes the symmetric difference. Since hG| ZC |Gi =
0 otherwise
p(a|x) =
1
2n
1
= n
2
1
= n
2
=
(x,a)
1
2n
X
P
(−1)|G[D∩supp(x)]|+
u∈D
au
D⊆V,D\supp(x)=Odd(D∩supp(x))
X
P
(−1)|G[D1 ]|+
X
u∈D0 ∪D1
au
D1 ⊆supp(x) D0 ⊆V \supp(x),D0 =Odd(D1 )
P
(−1)|G[D1 ]|+
X
u∈loc(D1 )
au
D1 ⊆supp(x),Odd(D1 )∩supp(x)=∅
(x,a)
X
P
(−1)|G[D1 ]|+
u∈loc(D1 )
D1 involved in x
=
|R0
(x,a)
| − |R1
2n
|
P
/
u∈loc(D) au = d mod 2}. If (a|x) ∈
|{D involved in x}|
L, then
= ∅, so p(a|x) =
> 0 since ∅ is involved in
2n
(x,a)
(x,a)
x. Otherwise, there exists D0 ∈ R1 . Notice that R0
is a vector space
(x,a)
(x,a)
(x,a)
(x,a)
(∀D1 , D2 ∈ R0 , D1 ⊕ D2 ∈ R0 ) and R1
an affine space R1
= {D0 ⊕
(x,a)
(x,a)
(x,a)
D | D ∈ R0 }. Thus |R0 | = |R1 | which implies p(a|x) = 0.
t
u
where Rd
= {D involved in x : |G[D]| +
au
(x,a)
R1
The probability distribution produced by QStrat depends on the number
of sets D involved in a given question x. Notice that a set D ⊆ supp(x) is
involved in x if and only if D ∈ Ker(Lx ), where Lx linearly2 maps A ⊆ supp(x)
to Odd(A) ∩ supp(x). Thus |{D involved in x}| = 2|x|−rkG (x) , where rkG (x) =
log2 (|{Lx (A) : A ⊆ supp(x))}|) is the rank of Lx = A 7→ Odd(A) ∩ supp(x).
2
Lx is linear for the symmetric difference: Lx (D1 ⊕ D2 ) = Lx (D1 ) ⊕ Lx (D2 )
Contextuality scenario. Following the hypergraph model of [3], we associate
with every graph G a contextuality scenario, where each vertex is a pair (a|x)
and each hyperedge corresponds, roughly speaking, to a constraint. There are
two kinds of hyperedges, those (HNsigV ) which guarantee no-signaling and those
(HG ), depending on the graph G, which avoid the losing pairs.
– HNsigV is the hypergraph representing the no-signaling polytope. It corresponds [3] to the Bell scenario BV,2,2 where |V | parties have access to 2 local
measurements each, each of which has 2 possible outcomes (see Figure 1),
which is obtained as a product3 of the elementary scenario B1,2,2 .
Fig. 1. HNsig2 : hyperedges of the Bell
scenario B2,2,2 from [14]
Fig. 2. Paley Graph of order 13
– The hypergraph HG defined on the same vertex set, corresponds to the game
constraints: for each question4 x ∈ {0, 1}V we associate an hyperedge ex
containing all the answers which make the players win on x i.e., ex = {(a|x) ∈
{0, 1}V × {0, 1}V , (a|x) ∈
/ L}.
Given a graph G = (V, E), MCG(G) is a pseudo-telepathy game if it admits a
quantum model (Q(HG ∪ HNsigV ) 6= ∅) but no classical model (C(HG ∪ HNsigV ) =
∅). It has been proven in [4] that MCG(G) is pseudo-telepathic if and only if G
is not bipartite.
3
4
The Foulis Randall product of scenarios [3] is the scenario HA ⊗ HB with vertices
V (HA ⊗ HB ) = V (HA ) × V (HB ) and edges E(HA ⊗ HB ) = EA→B ∪ EA←B where
EA→B := {∪a∈eA {a} × f (a) : ea ∈ EA , f : eA → EB } and EA←B := {∪b∈eA f (b) ×
{b} : eb ∈ Eb , f : EB → EA }. In the multipartite case there are several ways to
define products, however they all correspond to the same non-locality constraints [3].
Therefore one can just consider the minimal product min ⊗n
i=1 Hi which has vertices in
the cartesian product V = ΠVi and edges ∪k∈[1,n] Ek where Ek = {(v1 . . . , vn ), vi ∈
→
ei ∀i =
6 k, vk ∈ f (−
v )} for some edge ei ∈ E(Hi ) for every party i 6= k and a function
−
→
−
→
→
v 7→ f ( v ) which assigns to every joint outcome −
v = (v1 . . . vk−1 , vk+1 , . . . vn ) an
→
edge f (−
v ) ∈ E(Hk ) (the kth vertex is replaced by a function of the others).
Note that for the questions x for which there exists no D involved in x, all the answers
are allowed thus the constraints represented by the associated edge is a hyperedge of
no-signaling scenario HN sig .
Example 3: In a complete graph Kn of order n, there exists a non-empty set
D involved in a question x ∈ {0, 1}V if and only if |x| = 1 mod 2. With each such
question x, the associated hyperedge is ex = {(a|x) ∈ {0, 1}V × {0, 1}V s.t. 2|a| =
6
|x| + 1 mod 4}.
Example 4: In the graph Paley 13 (see Figure 2), Odd({0, 1, 4}) =
{2, 7, 8, 9, 11, 12} thus if {0, 1, 4} is involved in x i.e. xi = 1 for i ∈ {0, 1, 4}
and xi = 0 for i ∈ {2, 7, 8, 9, 11, 12} then the associatedP
pseudo-telepathy game
requires that the sum of the outputs of these nine players i6∈{3,5,6,10} ai has to be
odd. This corresponds to 8 hyperedges
ejkl for j, k, l ∈ {0, 1} in the contextuality
P
scenario where ejkl = {(a|x), i6∈{5,6,10} ai = 1 mod 2, xi = 1 for i ∈ {0, 1, 4},
xi = 0 for i ∈ {2, 7, 8, 9, 11, 12}, x5 = j, x6 = k, x10 = l}.
The probabilistic contextuality is what was considered in [4] as it corresponds
to investigating the possibility of simulating a probability distribution of a
quantum strategy playing with graph states. The two other levels of contextuality
gain some new perspectives when viewed as games: indeed the possibilistic
contextuality coincides with the fact that the players cannot give all the good
answers with non zero probability using classical local strategies, and strong
contextuality just means that classical players cannot win the game (even by
giving a strict subset of the good answers).
Definition 2. An interpretation p : {0, 1}V × {0, 1}V → [0, 1] is k-multipartite
if it can be obtained by a strategy without communication using nonlocal boxes
that are at most k-partite: for any set I ⊂ V with |I| ≤ k, each player has access
to one bit of a variable λI (aI |xI ) that has a no-signaling probability distribution.
In other words, a k-multipartite interpretation can be obtained with nosignaling correlations involving at most k players. For example the strategy to
win the Mermin game proposed in [10] where each pair among n players share a
(2-partite) non localbox and each player outputs the sum of his boxes’ ouputs
is a 2-multipartite interpretation. Similarly, the result in [8] where they prove
that a probability distribution that can be obtained by 5 players measuring a
quantum state cannot be simulated without communication using any number of
bi-partite non local boxes shows that it is not a 2-multipartite interpretation.5
Definition 3 (multipartiteness width). A scenario has a multipartiteness
width k if it admits a k-multipartite interpretation but no (k − 1)-multipartite
interpretation.
In a contextual scenario, the more hyperedges one adds the less possible
interpretations exist. A scenario has a multipartiteness width k if its hyperedges
already forbids all the interpretations of a product of Bell scenarios on less
than k parties. For a scenario, having a classical interpretation means being
decomposable : one can think of the probability distribution as local actors acting
each on his bit and that’s a classical interpretation. The multipartiteness width
5
The probability distribution described in [8] corresponds to the quantum winning
strategy on the graph state obtained from a cycle with 5 vertices.
measure how non-decomposable a scenario is : it can not be decomposed with
interpretations where each subspace has a small width.
It implies that the players cannot perfectly win the game if they have only
quantum systems on less than k qubits, this corresponds to using k seprable
states as ressources as defined in [19].
Note that from the observations in [4] the multipartiteness width of the
scenario generated by the Paley graph on 13 (see figure 2) is strictly larger than
4.
In the next section, we will show how for the scenarios we describe, being able
to give only good answers allows for simulation of the quantum distribution with
random variables. Thus, the contextuality lies in the combinatorial structure of
the graph and the three levels collapse for these games.
3
Simulating a probability distribution is the same as
winning the pseudo-telepathy graph game
In [4] it was proven that for some graphs, the probability distributions of the
quantum strategy using the graph states cannot be simulated using non local
boxes on less than k parties, we show here that any strategy that allows to win
the game can be extended using random variables shared between neighbors (in
the graph) to simulate this uniform probability distribution coming from the
quantum strategy.
We start by describing a classical strategy CStrat based on shared random
variables rather than quantum states. We show that CStrat is a winning strategy
if and only if the graph is bi-partite. We also show that CStrat can be used to
make any winning strategy a uniform winning strategy, i.e. each valid answer to
a given question are equiprobable. We show that CStrat can be locally adapted
to collaborative games on graphs that can be obtained by a sequence of local
complementations.
Classical strategy (Cstrat): Given a graph G = (V, E), pick uniformly at
V
random
bits (λu , µu ), where
Pλ ∈ {0, 1} . Each player u ∈ V receives a pair of
µu = v∈NG (u) λu mod 2. Given a question x ∈ {0, 1}V , each player u ∈ V
locally computes and answers au = (1 − xu ).λu + xu .µu mod 2.
Lemma 2. Given a graph G = (V, E) and a question x ∈ {0, 1}V , CStrat
produces an answer uniformly at random in {a ∈ {0, 1}V | ∃D ⊆ S, (A ⊕
Odd(A ⊕ D))∩S = ∅ where A = supp(a) and S = supp(x)}.
Proof. Given a graph G = (V, E), a question x ∈ {0, 1}V and a ∈ {0, 1}V , the
probability that CStrat outputs a is
p(a|x) = p (∀u ∈ V \S, au = λu ) p(∀u ∈ S, au =
X
λv mod 2 | ∀u ∈ V \S, au = λu )
v∈N (u)
= p (A \ S = Λ \ S) p(A ∩ S = Odd(Λ) ∩ S|A \ S = Λ \ S)
where S = supp(x), A = supp(a) and Λ = supp(λ). Since p (A \ S = Λ \ S) =
1
,
2n−|x|
p(a|x) =
=
1
2n−|x|
p (A ∩ S = Odd(Λ ∩ S ⊕ Λ \ S) ∩ S|A \ S = Λ \ S)
1
p (A ∩ S = Odd(D ⊕ (A \ S)) ∩ S|A \ S = Λ \ S)
2n−|x|
where D = Λ ∩ S. If A ∩ S 6= Odd(D ⊕ (A \ S)) ∩ S for all D ⊆ S, then p(a|x) = 0.
Otherwise, the set of subsets D of S which satisfy the condition is the affine
space {D0 ⊕ D| D ⊆ S ∧ Odd(D) ∩ S = ∅}, where D0 is a fixed set which satisfies
1
A ∩ S = Odd(D0 ⊕ A \ S) ∩ S. Thus the p(a|x) = 2n−|x|
. |{D⊆S|Odd(D)∩S=∅}|
=
2|x|
|x|−rkG (x)−n
2
, which is independent of a, proving the uniformity of the answer.
Finally notice ∃D0 ⊆ S, A ∩ S = Odd(D0 ⊕ (A \ S)) ∩ S if and only if
∃D1 ⊆ S, (A ⊕ Odd(A ⊕ D1 ))∩S = ∅, by taking D1 = D0 ⊕ (A ∩ S).
t
u
We consider some standard graph transformations : Given a graph G = (V, E)
the local complementation on a vertex u ∈ V produces the graph G ∗ u =
(V, E ⊕ KN (v) ) where the sum is taken modulo 2 (it is the symmetric difference)
and KU is the complete graph on U ⊂ V . G ∗ u is obtained from G by exchanging
the edges by non edges and vice versa in the neighborhood of the vertex u.
Pivoting using an edge (u, v), is a sequence of three local complementations
G ∧ uv = G ∗ u ∗ v ∗ u. We denote by δloc (G) (δpiv (G)) the minimum degree
taken over all graphs that can be obtained from G through some sequence of
local complementations (edge pivots).
Given the shared randomness (λv , µv )v∈V associated with G, if player u
replaces its first bit by the XOR of its two bits, and each of his neighbors replaces
his second bit by the XOR of his two bits, one gets the shared randomness
associated with G ∗ u.
Lemma 3. Given the probability distribution (λv , µv )v∈V associated with G, if
player u replaces its first bit by the XOR of its two bits, and each of its neighbors
replaces their second bit by the XOR of their two bits, one gets the probability
distribution associated with G ∗ u.
(
λu + µu mod 2 if u = v
0
Proof. For any v ∈ V , let λv =
and
λv
otherwise
(
µv + λv mod 2 if v ∈ NG (u)
µ0v =
.
µv
otherwise
(
NG (v) ⊕ NG (u) ⊕ {v} if v ∈ NG (u)
Since NG∗u (v) =
, one gets that for
NG (v)
otherwise
P
P
0
any v ∈
/ NG (u), µ0v = µv =
w∈NG (v) λw mod 2 =P w∈NG∗u (v) λw mod 2
0
and for any v ∈ NG (u), µv = µv + λv = λv + w∈NG (v) λw mod 2 =
P
P
P
λv + µu + w∈NG (v) λ0w mod 2 = λv + w∈NG (u) λ0w + w∈NG (v) λ0w mod 2 =
P
0
t
u
w∈NG∗u (v) λw mod 2.
Thus the probability distribution corresponding to the classical strategy
for G can be locally transformed into the probability distribution associated
with the G ∗ u, thus one can use local complementation to optimise the cost of
preparing the shared randomness. For instance the classical strategy CStrat
for a graph G requires shared random bits on at most ∆loc (G) + 1 players,
where ∆loc (G) = min(∆(G0 ), s.t. ∃u1 , . . . , uk , G0 = G ∗ u1 ∗ . . . ∗ uk ) and ∆(G)
is its maximum degree. If there is no pre-shared random bits, the probability
distribution can be prepared using at most 2|G|loc communications in-between
the players, where |(G)|loc = min(|G0 |, s.t. ∃u1 , . . . , uk , G0 = G ∗ u1 ∗ . . . ∗ uk ) is
the minimum number of edges by local complementation.
Now we show how, using the classical strategy CStrat , one can simulate the
quantum strategy QStrat given an oracle that provides only good answers.
Lemma 4. For any collaborative game on a graph G, for any strategy Q that
never loses, there exists a strategy Q0 using the outputs of Q and shared random
variables that simulate QStrat .
Proof. Given a collaborative graph game on a graph G, let Q be a strategy that
always outputs permissible outputs for any set of inputs x, so we have pairs
(a|x) 6∈ L. We consider the strategy which combines Q and CStrat for this graph:
For a given question x, Q0 outputs the XOR of the Q answer and CStrat answer
for x. First we prove that such an answer is a valid answer and then the uniform
probability among the possible answer to a given question. Given a question
x ∈ {0, 1}V , suppose Q0 outputs a0 ∈ {0, 1}V : ∀u ∈ V , a0u = au +(1−xu )λu +xu µu
where au is the answer produced by Q and λ and µ are as defined in the classical
strategy.
assume (a0 |x) ∈ L, so there exists D involved in x such that
P By contradiction,
0
u∈loc(D) au = |G[D]| + 1 mod 2
X
a0u =
u∈loc(D)
X
(au + (1 − xu )λu + xu µu ) mod 2
u∈loc(D)
=
X
u∈loc(D)
=
X
au +
X
u∈loc(D)
=
X
X
λu +
X
u∈Odd(D)
µu mod 2
u∈loc(D)∩supp(x)
X X
λv mod 2
u∈D v∈N (u)
u∈Odd(D)
au +
X
λu +
u∈loc(D)\supp(x)
u∈loc(D)
=
X
au +
λu +
X
λv mod 2
v∈Odd(D)
au mod 2
u∈loc(D)
Thus (a|x) ∈ L which is a contradiction thus p(a0 |x) = 0 if (a0 |x) ∈ L
Now we prove that p(a0 |x) = 2|x|−n−rkG (x) . First assume Q is
determinist, thus p(a0 |x) is the probability that the classical strategy outputs a + a0 := (au + a0u mod 2)u∈V . Since this probability is
|x|−n−rkG (x)
non
. If Q is probabilistic, p(a0 |x) =
P zero it must be 2
0
≤
a∈{0,1}V p(Q outputs a on x)p(classical strategy outputs a + a on x)
P
|x|−n−rkG (x)
|x|−n−rkG (x)
2
≤ 2
. Thus each
a∈{0,1}V p(Q outputs a on x)
answer a produced by the strategy on a given question x is s.t.
(a|x) ∈
/ L and occurs with probability at most 2|x|−n−rkG (x) . Since
|{a ∈ {0, 1}V | (a|x) ∈
/ L}| = 2|x|−n−rkG (x) , each of the possible answers
is produced by the strategy and occurs with probability 2|x|−n−rkG (x) .
t
u
4
Locally equivalent games
A pseudo telepathy game G locally simulates another pseudo telepathy game G 0
if any winning strategy for G can be locally turned into a winning strategy for G 0 :
Definition 4 (Local Simulation). Given two pseudo telepathy games G and
G 0 on a set V of players which sets of losing pairs are respectively LG and LG 0 ,
G locally simulates G 0 if for all u ∈ V , there exist f1 , . . . , fn : {0, 1} → {0, 1}
and g1 , . . . , gn : {0, 1} × {0, 1} → {0, 1} s.t. ∀x, a ∈ {0, 1}V (g(a, x), x) ∈ LG 0 ⇒
(a|f (x)) ∈ LG where f (x) = (fu (xu ))u∈V and g(a, x) = (gu (au , xu ))u∈V .
Assuming G locally simulates G 0 and that the players have a strategy to win
G, the strategy for G 0 is as follows: given an input x of G 0 , each player u applies
the preprocessing fu turning her input xu into fu (xu ), then they collectively play
the game G with this input f (x) getting an output a s.t. (a|f (x)) ∈
/ LG . Finally
each player u applies a postprocessing gu which depends on her output au and
her initial input xu to produce the output gu (au , xu ) to the game G 0 . This output
is valid since, by contradiction, (g(a, x), x) ∈ LG 0 would imply (a|f (x)) ∈ LG .
Definition 5 (Local Equivalence). G and G 0 are locally equivalent games if
G locally simulates G 0 and G 0 locally simulates G.
In the following we give two examples of locally equivalent games : first we
show that the games associated with the complete graphs are locally equivalent to
Mermin parity games, and then that pivoting, a graph theoretical transformation,
produces a graph game locally equivalent to the original one:
Lemma 5. For any n, the game associated with the complete graph Kn is locally
equivalent to the Mermin parity game on n players.
Proof. The set of losing pairs of the two games are LMermin = {(a|x), 2|a| =
|x| + 2 mod 4} (see example 1 ) and LKn = {(a|x), 2|a| = |x| + 1 mod 4} (see
example 2 ).
[Kn simulates
Mermin] Let u ∈ V be a fixed (
player. We define for all v ∈ V ,
(
1 − xv if v = u
av + xv − 2av xv if v = u
fv (xv ) =
and gv (av , xv ) =
. If
xv
otherwise
av
otherwise
(g(a, x), x) ∈ LMermin then 2|g(a, x)| = |x|+2 mod 4 which implies 2(|a|−au +au +
xu −2au xu ) = 2|a|+2xu = |x|+2 mod 4, so 2|a| = |x|+2−2xu = |f (x)|+1 mod 4,
thus (a|f (x)) ∈ LKn .
[Mermin simulates Kn ] Let u ∈ V be a fixed player. f and g are defined like in
the previous case except gu (au , xu ) = 1 − au − xu + 2au xu . If (g(a, x), x) ∈ LKn
then 2|g(a, x)| = |x| + 1 mod 4 which implies 2(|a| − au + 1 − au − xu + 2au xu ) =
2|a| + 2 − 2xu = |x| + 1 mod 4, so 2|a| = |x| + 1 − 2xu + 2 = |f (x)| + 2 mod 4,
thus (a|f (x)) ∈ LMermin .
t
u
Lemma 6. Given a graph G = (V, E) and (u, v) ∈ E, the games associated with
G and G ∧ uv are locally equivalent.
Proof. Since pivoting is its self inverse ((G ∧ uv) ∧ uv = G) it is enough to prove
that G locally simulates G ∧ uv. The proof is based on the existence of a quantum
strategy for any graph G which consists in sharing the quantum state |Gi, and
for each player w, in measuring her qubit according to X if xw = 1 or according
to Z if xw = 0, and then output the outcome aw ∈ {0, 1} of this measurement.
A quantum strategy for G can be turned into a quantum strategy for G ∧ uv due
to the following property of graph states: |G ∧ uvi = Hu,v ZN (u)∩N (v) |Gi [27].
The unitary map H exchanges X- and Z- measurements – i.e. for any state
|φi, apply H on |φi followed by a Z-measurement (resp. X) produces the same
classical outcome as measuring |φi according to X (resp. Z) – while the unitary
Z leaves invariant the classical outcomes of a Z-measurement and exchanges
the two possible outcomes of a X-measurement. As a consequence, (g(a, x), x) ∈
L
(G∧uv ⇒ (a|f (x)) ∈ LG , where f and
( g are defined as follows fw (xw ) =
1−xw if w∈{u, v}
aw +xw −2aw xw if w∈N (u)∩N (v)
and gw (aw , xw ) =
xw
otherwise
aw
otherwise
t
u
Therefore, the important quantity for the pre-shared randomness for the strategies defined with a graph is ∆piv (G) = min{∆(G0 ), G0 pivot equivalent to G}.
5
Scenarios with linear multipartiteness width
We prove that there exist contextuality scenarios with linear multipartiteness
width. We use a graph property called k-odd domination which is related [4]
to the classical simulation of the quantum probability distribution obtained by
playing the associated graph game. Since bipartite graphs correspond to graph
games that can be won classically [4], we focus on the non-bipartite case by
showing that there exist non-bipartite 0.11n-odd dominated graphs of order n.
Definition 6 (k-odd domination
[4]). A graph G = (V, E) is k-odd dominated
(k-o.d.) iff for any S ∈ Vk , there exists a labelling of the vertices in S =
{v1 , . . . , vk } and C1 , . . . Ck , s.t. ∀i, Ci ⊆ V \ S and Odd(Ci ) ∩ {vi , . . . vk } = {vi }
and Ci ⊆ Even(Ci ).
Lemma 7. For any k ≥ 0, r ≥ 0 and any graph G = (V, E) a graph of order
n having two distinct independent sets V0 and V1 of order |V0 | = |V1 | = b n−r
2 c,
G is k-odd dominated if for any i ∈ {0, 1}, and any non-empty D ⊆ V \ Vi ,
|OddG (D) ∩ Vi | > k − |D|
Proof. Given S0 ⊆ V0 , S1 ⊆ V1 , and S2 ⊆ V2 = V \(V0 ∪V1 ) s.t. |S0 |+|S1 |+|S2 | =
k, we show that for any u ∈ S = S0 ∪ S1 ∪ S2 , there exists Cu ⊆ V \ S s.t.
Odd(Cu ) ∩ S = {u} and Cu ⊆ Even(Cu ). For any u ∈ S, there exists i ∈ {0, 1}
s.t. u ∈ Si ∪ S2 . Let Li : 2Si ∪S2 → 2V1−i \S1−i be the function which maps
D ⊆ Si ∪ S2 to Li (D) = OddG (D) ∩ (V1−i \ S1−i ).
Li is linear according to the symmetric difference. Li is injective: for any
D ⊆ Si ∪ S2 , Odd(D) ∩ (V1−i \ S1−i ) = ∅ implies Odd(D) ∩ V1−i ⊆ S1−i , thus
|Odd(D) ∩ V1−i | ≤ |S1−i |. notice that |D| ≤ |Si | + |S2 |, so |Odd(D) ∩ V1−i | ≤
|S1−i | ≤ |S0 | + |S1 | + |S2 | − |D| = k − |D|, so D = ∅.
The matrix representing Li is nothing but the submatrix Γ[Si ∪S2 ,V1−i \S1−i ]
of the adjacency matrix Γ of G. So its transpose Γ[V1−i \S1−i ,Si ∪S2 ] is surjective
which means that the corresponding linear map LTi : 2V1−i \S1−i → 2Si ∪S2 = C 7→
OddG (C) ∩ (V1−i \ S1−i ) is surjective, so ∃Cu ⊆ V1−i \ S1−i s.t. OddG (Cu ) ∩ (Si ∪
S2 ) = {u}, which implies, since V1−i is an independent set, that OddG (Cu ) ∩ S =
{u} and Cu ⊆ Even(Cu ).
t
u
Theorem 1. For any even n > n0 , there exists a non-bipartite b0.110nc-odd
dominated graph of order n.
Proof. Given n, r ≤ n s.t. r = n mod 2, and k ≥ 0. Let p = (n − r)/2, and
let G = (V0 ∪ V1 ∪ V2 , E) s.t. |V0 | = |V1 | = p, |V2 | = r be a random graph on
n vertices s.t. for any u ∈ Vi , v ∈ Vj there is an edge between u and v with
probability 0 if i = j and with probability 1/2 otherwise.
(i)
For any i ∈ {0, 1}, and any non empty D ⊆ V \ Vi s.t. |D| ≤ k, let AD be
the bad event |OddG (D) ∩ Vi | ≤ k − |D|. Since each vertex of Vi is in OddG (D)
k−|D|
Pk−|D| p −p
(i)
with probability 1/2, P r(AD ) = j=0
≤ 2p[H( p )−1] .
j 2
Another bad event is that G is bipartite which occurs with probability less
than ( 78 )pr . Indeed, the probability that given u ∈ V0 , v ∈ V1 , w ∈ V2 , (u, v, w)
do not form a triangle is 78 , so given a bijection f : V0 → V1 , the probability that
∀u ∈ V0 , ∀w ∈ V2 , (u, f (u), w) do not form a triangle is ( 78 )pr .
Let X be the number of bad events.
k−d
k
X
7
p + r X p −p
2 + ( )pr
E[X] = 2
j
d
8
j=0
d=1
≤2
k
X
d
2(p+r)H( p+r )+pH(
k−d
p )−p
d=1
≤2
k
X
d=1
d
2pH( p+r )+pH(
k−d
p )−p+r
7
+ ( )pr
8
7
+ ( )pr
8
d
The function d 7→ pH( p+r
) + pH( k−d
p ) − p + r is maximal for d =
k
2pH( 2p+r
)−p+r
k(p+r)
2p+r .
Thus,
E[X] ≤ 2k2
+ ( 78 )pr .
By taking r = 1, and k = 0.11n = 0.11(2p + 1), E[X] < 1 when p large
enough, thus G has no bad event with a non zero probability.
t
u
Corollary 1. There exist contextuality scenarios with linear multipartiteness
width: for any even n > n0 , there exist graph games on n players producing
contextuality scenarios of multipartiteness width at least b0.11nc.
Proof. Using the result from [4], for any non bipartite graph of order n being 0.11no.d ensures that the probability distribution obtained by using the quantum
strategy cannot be simulated using non local boxes involving at most 0.11n
parties. Thus lemma 4 allows to conclude that the associated pseudo-telepathy
game cannot be won classically. Therefore there is no interpretation that is
k-multipartite with k < 0.11n which means that the contextuality scenario has
linear width.
t
u
6
Conclusion
We have shown that there exist graphs with linear multipartiteness width, however
the proof is non constructive and the best known bound for explicit families
is logarithmic.A natural future direction of research would be to find explicit
families with linear multipartiteness width or to improve the bounds proven
for the Paley graph states. An other important question is to consider lower
bounds for the scenarios associated with the graph games.A promising area of
investigation for multipartite scenarios is: what happens if we limit the width
of shared randomness? Indeed, for the proof of how winning the game allows to
simulate the quantum probability distributions, one needs only shared random
variables that are correlated in local neighborhoods in the graph. One can
also consider the link with building entanglement witnesses for graph states,
generalizing the construction of [21]. It would be also very interesting to link the
multipartiteness width with the structures of the groups of the associated binary
linear system defining the two-player bipartite non-local games [28]. Finally, one
can expect that the multipartiteness width of the Paley graph states might have
cryptographic applications to ensure security against cheating for some protocols
for example.
References
1. S. Abramsky and A. Brandenburger The sheaf theoretic structure of non locality
and contextuality. New Journal of Physics 13 113036 (2011)
2. S. Abramsky, R.S. Barbosa and S. Mansfield Quantifying contextuality via linear
programming, Informal Proceedings of Quantum Physics & Logic, (2016)
3. A. Acı́n, T. Fritz, A. Leverrier, and A. Belén Sainz A combinatorial approach to
nonlocality and contextuality. Comm. Math. Phys. 334(2), 533-628 (2015)
4. A. Anshu and M. Mhalla Pseudo-telepathy games and genuine NS k-way nonlocality
using graph states Quantum Information and Computation, Vol. 13, No. 9, 10
0833-0845 Rinton Press (2013)
5. A. Badanidiyuru, J. Langford, A. Slivkins Resourceful Contextual Bandits.
http://arxiv.org/abs/1402.6779, COLT (2014)
6. J. Barrett, S. Pironio, J. Bancal and N. Gisin, The definition of multipartite
nonlocality. Phys. Rev. A 88, 014102 (2013)
7. J.S. Bell On the Einstein-Podolsky-Rosen paradox. Physics 1 195-200 (1964)
8. J. Barrett and S. Pironio, Popescu-Rohrlich correlations as a unit of nonlocality.
Phys. Rev. Lett. 95, 140401, (2005)
9. G. Brassard, A. Broadbent and A. Tapp Multi-Party Pseudo-Telepathy 8th International Workshop, WADS 2003, Ottawa, Ontario, Canada, July 30 - August 1,
Proceedings (2003).
10. A. Broadbent and A. A. Methot, On the power of non-local boxes Theoretical
Computer Science C 358: 3-14, (2006)
11. B. Coecke From quantum foundations via natural language meaning to a theory of
everything. arXiv:1602.07618v1 (2016)
12. E. Dzhafarov , S. Jordan , R. Zhang , V. Cervante Contextuality from Quantum
Physics to Psychology. Advanced Series on Mathematical Psychology: Volume 6
(2016)
13. D. J. Foulis and C. H. Randall, Empirical logic and tensor products. J. Mathematical
Phys. 5 , 9:20. MR683888. (1981)
14. T. Fritz, A. B. Sainz, R. Augusiak, J. Bohr Brask, R. Chaves, A. Leverrier, and A.
Acı́n Local orthogonality as a multipartite principle for quantum correlations Nature
Communications 4, 2263 (2013)
15. A. Grudka, K. Horodecki, M. Horodecki, P. Horodecki, R. Horodecki, P. Joshi, W.
Klobus, A. Wójcik Quantifying contextuality Phys. Rev. Lett. 112, 120401 (2014)
16. D. M. Greenberger, M. A. Horne, A. Shimony, and A. Zeilinger, Bell’s theorem
without inequalities. American Journal of Physics 58 , 1131.(1990)
17. P. Gnaciński, M. Rosicka, R. Ramanathan, K. Horodecki, M. Horodecki, P.
Horodecki, S. Severini Linear game non-contextuality and Bell inequalities - a
graph-theoretic approach. e-print arXiv:1511.05415 Nov (2015)
18. C. Godsil, G. Royle. Algebraic graph theory, volume 207 of Graduate Texts in
Mathematics. volume 207 of Graduate Texts in Mathematics - Springer-Verlag,
(2001)
19. O. Gühne, G.Tóth, H.J. Briegel Multipartite entanglement in spin chains. New
Journal of Physics, 7(1), 229. (2005)
20. L. Hardy Nonlocality for two particles without inequalities for almost all entangled
states. Physical Review Letters 71 , 1665:1668.(1993)
21. M. Hein,W. Dür, J. Eisert, R. Raussendorf, M. Nest and H. J. Briegel Entanglement
in graph states and its applications. arXiv preprint quant-ph/0602096. (2006)
22. M. Howard, J.J. Wallman, V. Veitch and J. Emerson Contextuality supplies the
”magic” for quantum computation. Nature 510, 351-355 (2014)
23. V. Kumar. Algorithms for constraint-satisfaction problems: A survey. AI magazine,
13(1):32, (1992).
24. R. Luce and H. Rai. Games and Decisions: Introduction and Critical Survey.
Doverbooks on advanced mathematics, Dover Publications (1957)
25. Y.-C.Liang, D. Rosset, J-D. Bancal, G.Pütz, T.J. Barnea and N. Gisin Family of
Bell-like inequalities as device-independent witnesses for entanglement depth. Physical
review letters, 114(19), 190401. (2015)
26. N. D. Mermin, Extreme quantum entanglement in a superposition of macroscopically
distinct states. Physical Review Letters, 65(15):1838-1849, (1990).
27. M. Mhalla and S. Perdrix Graph States, Pivot Minor, and universality of (X-Z)
Measurements. IJUC 9 (1-2) : 153-171, (2013).
28. W. Slofstra Tsirelson’s problem and an embedding theorem for groups arising from
non-local games e-print arXiv:1606.03140 (2106)
29. N. Pflueger. Graph reductions, binary rank, and pivots in gene assembly. Discrete
Applied Mathematics, 2011.
30. W. Zeng and P. Zahg, Contextuality and the weak axiom of the theory of choice
Quantum Interactions proc. Volume 9535 of the series Lecture Notes in Computer
Science pp 24-35, (2016)
Acknowledgements.
We would like to thanks an anonymous reviewer for noticing a mistake in an
earlier version and helpful comments.
| 7cs.IT
|
On Global Warming
(Softening Global Constraints)
Willem Jan van Hoeve1 , Gilles Pesant2,3 and Louis-Martin Rousseau2,3
arXiv:cs/0408023v1 [cs.AI] 9 Aug 2004
1
CWI, P.O. Box 94079, 1090 GB Amsterdam, The Netherlands
[email protected]
2
École Polytechnique de Montréal, Montreal, Canada
3
Centre for Research on Transportation (CRT),
Université de Montréal, C.P. 6128, succ. Centre-ville, Montreal, H3C 3J7, Canada
{pesant,louism}@crt.umontreal.ca
Abstract. We describe soft versions of the global cardinality constraint
and the regular constraint, with efficient filtering algorithms maintaining
domain consistency. For both constraints, the softening is achieved by
augmenting the underlying graph. The softened constraints can be used
to extend the meta-constraint framework for over-constrained problems
proposed by Petit, Régin and Bessière.
1
Introduction
Constraint Programming (CP) is a widely used and efficient technique to solve
combinatorial optimization problems. However in practice many problems are
over-constrained (intrinsically or from being badly stated). Several frameworks
have been proposed to handle over-constrained problems, mostly by introducing
soft constraints that are allowed to be (partially) violated. The most well-known
framework is the Partial Constraint Satisfaction Problem framework (PCSP [8]),
which includes the Max-CSP framework that tries to maximize the number of
satisfied constraints. Since in this framework all constraints are either violated
or satisfied, this objective is equivalent to minimizing the number of violations.
It has been extended to the Weighted-CSP [10,11], associating a degree of violation (not just a boolean value) to each constraint and minimizing the sum
of all weighted violations. The Possibilistic-CSP [18] associates a preference to
each constraint (a real value between 0 and 1) representing its importance. The
objective of the framework is the hierarchical satisfaction of the most important
constraints, that is, the minimization of the highest preference level for a violated
constraint. The Fuzzy-CSP [6,7] is somewhat similar to the Possibilistic-CSP
but here a preference is associated to each tuple of each constraint. A preference
value of 0 means the constraint is highly violated and 1 stands for satisfaction.
The objective is the maximization of the smallest preference value induced by
a variable assignment. The last two frameworks are different from the previous
ones since the aggregation operator is a min/max function instead of addition.
Max-CSPs are typically encoded and solved with one of two generic paradigms:
valued-CSPs [19] and semi-rings [5].
Another approach to model and solve over-constrained problems involves
Meta-Constraints [13]. The idea behind this technique is to introduce a set
of domain variables Z that capture the violation cost of each soft constraint.
By correctly constraining these variables it is possible to replicate the previous
frameworks and even to extend the modeling capability to capture other types
of violation measures. Namely the authors argue that although the Max-CSP
family of frameworks is quite efficient to capture local violation measures it is
not as adequate to model violation costs involving several soft constraints simultaneously. By defining (possibly global) constraints on Z such a behaviour
can be easily achieved. The authors propose to replace each soft constraint Si
present in a model by a disjunctive constraint specifying that either zi = 0 and
the constraint Si is hard or zi > 0 and Si is violated. This technique allows the
resolution of over-constrained problem within traditional CP solvers.
Comparatively few efforts have been invested in developing soft versions of
common global constraints [14,4,9]. Global constraints are often key elements in
successfully modeling real applications and being able to easily and effectively
soften such constraints would yield a significant improvement in flexibility. In
this paper we study two global constraints: the widely known global cardinality
constraint (gcc) [15] and the new regular [12] constraint. For each of these we
propose new violation measures and provide the corresponding filtering algorithms to achieve domain consistency. All the constraint softening is achieved by
enriching the underlying graph representation with additional arcs that represent
possible relaxations of the constraint. Violation costs are then associated to these
new arcs and known graph algorithms are used to achieve domain consistency.
The two constraints studied in this paper are useful to model and solve
personnel rostering problems (PRP). The PRP objective is typically to distribute
a set of working shifts (or days off) to a set of employees every day over a
planning horizon (a set of days). The gcc is a perfect tool to restrict the number
of work shifts of each type (Day, Evening, and Night for instance) performed by
each employee. Other types of constraints involve sequences of shifts over time,
typically forbidding non ergonomic schedules. The regular constraint has the
expressive power necessary to cope with the complex regulations found in many
organizations. Since most real rostering applications are over-constrained (due
to lack of personnel or over-optimistic scheduling objectives), soft versions of
the gcc and regular constraints promise to significantly improve our modelling
flexibility.
This paper is organized as follows. Section 2 presents background information on Constraint Programming and the softening of (global) constraints. In
Section 3 and 4 we describe the softening of the gcc and the regular constraint
respectively. Both constraints are softened with respect to two violation measures. We also provide corresponding filtering algorithms achieving domain consistency. Section 5 discusses the aggregation of several soft (global) constraints
by meta-constraints. Finally, a conclusion is given in Section 6.
2
Background
We assume familiarity with the basic concepts of constraint programming. For
a thorough explanation of constraint programming, see [2].
A constraint satisfaction problem (CSP) consists of a finite set of variables
X = {x1 , . . . , xn } with finite domains D = {D1 , . . . , Dn } such that xi ∈ Di
for all i, together with a finite set of constraints C, each on a subset of X. A
constraint C ∈ C is defined as a subset of the Cartesian product of the domains
of the variables that are in C. A tuple (d1 , . . . , dn ) ∈ D1 × · · · × Dn is a solution
to a CSP if for every constraint C ∈ C on the variables xi1 , . . . , xik we have
(di1 , . . . , dik ) ∈ C. A constraint optimization problem (COP) is a CSP together
with an objective function to be optimized. A solution to a COP is a solution
to the corresponding CSP that has an optimal objective function value.
Definition 1 (Domain consistency). A constraint C on the variables x1 , . . . ,
xk is called domain consistent if for each variable xi and value di ∈ Di , there
exist values d1 , . . . , di−1 , di+1 , . . . , dk in D1 , . . . , Di−1 , Di+1 , . . . , Dk , such that
(d1 , . . . , dk ) ∈ C.
Our definition of domain consistency corresponds to hyper-arc consistency or
generalized arc consistency, which are also often used in the literature.
Definition 2 (Consistent CSP). A CSP is domain consistent if all its constraints are domain consistent. A CSP is inconsistent if it has no solution. Similarly for a COP.
When a CSP is inconsistent it is also said to be over-constrained. It is then
natural to identify soft constraints, that are allowed to be violated, and minimize
the total violation according to some criteria. For each soft constraint C, we
introduce a function that measures the violation, and has the following form:
violationC : D1 × · · · × Dn → N.
This approach has been introduced in [14] and was developed further in [4].
There may be several natural ways to evaluate the degree to which a global
constraint is violated and these are not equivalent usually. A standard measure
is the variable-based cost:
Definition 3 (Variable-based cost). Given a constraint C on the variables
x1 , . . . , xk and an instantiation d1 , . . . , dk with di ∈ Di , the variable-based cost
of violation of C is the minimum number of variables that need to change their
value in order to satisfy the constraint.
Alternative measures exist for specific constraints. For example, if a constraint
is expressible as a conjunction of binary constraints, the cost may be defined as
the number of these binary constraints that are violated. For the soft gcc and
the soft regular constraint, we will introduce new violation measures, that are
likely to be more effective in practical applications.
3
Soft Global Cardinality Constraint
A global cardinality constraint (gcc) on a set of variables specifies the minimum
and maximum number of times each value in the union of their domains should
be assigned to these variables. Régin developed a domain consistency algorithm
for the gcc, making use of network flows [15]. A variant of the gcc is the costgcc, which can be seen as a weighted version of the gcc [16,17]. For the cost-gcc
a weight is assigned to each variable-value assignment and the goal is to satisfy
the gcc with minimum total cost.
Throughout this section, we will use the following notation (unless specified
otherwise). Let X denote a set of variables {x1 , . . . , xn } with respective finite
domains D1 , . . . , Dn . We define DX = ∪i∈{1,...,n} Di and we assume a fixed but
arbitrary ordering on DX . For d ∈ DX , let ld , ud ∈ N, with ld ≤ ud . Finally, let
z be a variable with finite domain Dz , representing the cost of violation of the
gcc.
Definition 4 (Global cardinality constraint).
gcc(X, l, u) = {(d1 , . . . , dn ) | di ∈ Di , ld ≤ |{di | di = d}| ≤ ud ∀ d ∈ DX }.
We first give a generic definition for a soft version of the gcc.
Definition 5 (Soft global cardinality constraint).
˜ | di ∈ Di , d˜ ∈ Dz ,
soft gcc[⋆](X, l, u, z) = {(d1 , . . . , dn , d)
˜
violationsoft gcc[⋆] (d1 , . . . , dn ) ≤ d},
where ⋆ defines a violation measure for the gcc.
In order to define measures of violation for the gcc, it is convenient to introduce
the following functions.
Definition 6 (Overflow, underflow). Given gcc(X, l, u), define for all d ∈
DX
|{xi | xi = d}| − ud
if |{xi | xi = d}| ≥ ud ,
overflow(X, d) =
0
otherwise,
l − |{xi | xi = d}|
if |{xi | xi = d}| ≤ ld ,
underflow(X, d) = d
0
otherwise.
Let violationsoft gcc[var] denote the variable-based cost of violation (see Definition 3) of the gcc. The next lemma expresses violationsoft gcc[var] in terms of the
above functions.
Lemma 1. Given gcc(X, l, u),
violationsoft gcc[var] (X) = max
X
overflow(X, d),
X
d∈DX
d∈DX
!
underflow(X, d)
provided that
X
d∈DX
ld ≤ |X| ≤
X
d∈DX
ud .
(1)
Proof. The variable-based cost of violation corresponds
to the minimal numP
overflow(X,
d) = 0 and
ber
of
re-assignments
of
variables
until
both
d∈D
X
P
d)
=
0.
d∈DX underflow(X,
P
P
Assume d∈DX overflow(X, d) ≥ d∈DX underflow(X, d). Variables assigned
′
to values d′ ∈ DX with overflow(X, dP
) > 0 can be assigned to values d′′ ∈
′′
DX withPunderflow(X, d ) > 0, until d∈DX underflow(X, d) = 0. In order to
achieve d∈DX overflow(X, d) = 0, we still need to re-assign the other variables
assigned to valuesPd′ ∈ DX with overflow(X, d′ ) > 0. Hence, in total we need to
re-assign exactly d∈DX overflow(X,
d) variables.
P
P
Similarly when we assume d∈DX overflow(X, d) ≤ d∈DX underflow(X, d).
If (1) does not hold, there is no variable assignment that satisfies the gcc.
Without assumption (1), the variable-based violation measure for the gcc cannot
be applied. Therefore, we introduce the following value-based violation measure,
which can also be applied when assumption (1) does not hold.
Definition 7 (Value-based cost). For gcc(X, l, u) the value-based cost of violation is
X
overflow(X, d) + underflow(X, d).
d∈DX
We denote the value-based violation measure for the gcc by violationsoft gcc[val] .
3.1
Graph Representation
First, we introduce the concept of a flow in a directed graph, following Schrijver [20, pp. 148–150].
A directed graph is a pair G = (V, A) where V is a finite set of vertices and
A is a family1 of ordered pairs from V , called arcs. For v ∈ V , let δ in (v) and
δ out (v) denote the family of arcs entering and leaving v respectively.
A (directed) walk in G is a sequence P = v0 , a1 , v1 , . . . , ak , vk where k ≥ 0,
v0 , v1 , . . . , vk ∈ V , a1 , . . . , ak ∈ A and ai = (vi−1 , vi ) for i = 1, . . . , k. If there
is no confusion, P may be denoted as P = v0 , v1 , . . . , vk . A (directed) walk is
called a (directed) path if v0 , . . . , vk are distinct. A closed (directed) walk, i.e.
v0 = vk , is called a (directed) circuit if v1 , . . . , vk are distinct.
Let s, t ∈ V . We apply a capacity function c : A → R+ , a demand function
d : A → R+ and a cost function w : A → R+ on the arcs. A function f : A → R
is called a feasible flow from s to t, or an s − t flow, if
d(a) ≤ f (a) ≤ c(a)
for each a ∈ A,
f (δ out (v)) = f (δ in (v)) for each v ∈ V \ {s, t},
(2)
(3)
P
where f (S) = a∈S f (a) for all S ⊆ A. Property (3) ensures flow conservation,
i.e. for a vertex v 6= s, t, the amount of flow entering v is equal to the amount of
flow leaving v. The value of an s − t flow f is defined as
value(f ) = f (δ out (s)) − f (δ in (s)).
1
A family is a set in which elements may occur more than once.
In other words, the value of a flow is the net amount of flow leaving s, which can
be shown to be equal to the net amount of flow entering t. The cost of a flow f
is defined as
X
cost(f ) =
w(a)f (a).
a∈A
A minimum-cost flow is a feasible s−t flow of minimum cost. The minimum-cost
flow problem is the problem of finding such a minimum-cost flow.
Theorem 1 ([15]). A solution to gcc(X, l, u) corresponds to a feasible s − t
flow of value n in the graph G = (V, A) with vertex set
V = X ∪ DX ∪ {s, t}
and edge set
A = As→X ∪ AX→DX ∪ ADX →t ,
where
As→X = {(s, xi ) | i ∈ {1, . . . , n}},
AX→DX = {(xi , d) | d ∈ Di , i ∈ {1, . . . , n}},
ADX →t = {(d, t) | d ∈ DX },
with demand function
and capacity function
1 if a ∈ As→X ,
d(a) = 0 if a ∈ AX→DX ,
ld if a = (d, t) ∈ ADX →t ,
1 if a ∈ As→X ,
c(a) = 1 if a ∈ AX→DX ,
ud if a = (d, t) ∈ ADX →t .
Example 1. Consider the CSP
x1 ∈ {1, 2}, x2 ∈ {1}, x3 ∈ {1, 2}, x4 ∈ {1},
gcc(X, l, u)
where X = {x1 , . . . , x4 }, l1 = 1, l2 = 3, u1 = 2 and u2 = 5. In Figure 1.a the
corresponding graph G for the gcc by applying the above procedure is presented.
3.2
Variable-Based Violation
For the variable-based violation measure, we adapt the graph G in the following
/ Di , i ∈ {1, . . . , n}}, with
way. We add the arc set ÃX→DX = {(xi , d) | d ∈
demand d(a) = 0, capacity c(a) = 1 for all arcs a ∈ ÃX→DX . Further, we apply
a cost function w : A → R, where
1 if a ∈ ÃX→DX ,
w(a) =
0 otherwise.
Let the resulting graph be denoted by Gvar .
(0,1) w =1
x1
(1,1)
,1)
x3
t
s
)
(3,5
(0,1)
(0
,1
)
,1)
)
(0
(1,1)
(1
(1,2
(0,1)
(1,2
(0,1)
x2 (
0,1
)w
=
(1
,1)
x3
2
x4
x4
a. original gcc
(1,1)
)
s
1
(1,1)
t
,1)
(0
(0,1)
)
,1
(1
1
x2
,1)
(0,1)
w =1
2
b. soft gcc[var]
1
(0,i
nf)
(1,2
w=
1
)
,1)
t
(0
(1,1)
(1
)
(3,5
(0,1)
(0,1)
x3
)
(3,5
(0,1)
(0
,1)
s
x2
1)
,
(1
(0,1)
(0
,1)
(1,1)
1
,1)
(0
,1)
(0
1)
,
(1
x1
(0,1)
,1)
(0
x1
2
nf)
w
=1
(0,i
x4
=1
(0,3) w
c. soft gcc[val]
Fig. 1. Graph representation for the gcc, the variable-based soft gcc and the
value-based soft gcc. Demand and capacity are indicated between parentheses
for each arc. Dashed arcs indicate the inserted weighted arcs.
Example 2. Consider the CSP
x1 ∈ {1, 2}, x2 ∈ {1}, x3 ∈ {1, 2}, x4 ∈ {1}, z ∈ {0, 1, . . . , 4}
soft gcc[var](X, l, u, z)
minimize z
where X = {x1 , . . . , x4 }, l1 = 1, l2 = 3, u1 = 2 and u2 = 5. In Figure 1.b the
graph Gvar for the soft gcc[var] is presented.
Theorem 2. A minimum-cost flow in the graph Gvar corresponds to a solution
to the soft gcc[var], minimizing the variable-based violation.
Proof. An assignment xi = d corresponds to the arc a = (xi , d) with f (a) = 1.
By construction, all variables need to be assigned to a value and the cost function exactly measures the variable-based cost of violation.
The graph Gvar corresponds to a particular instance of the cost-gcc [16,17].
Hence, we can apply the filtering procedures developed for that constraint directly to the soft gcc[var]. The soft gcc[var] also inherits from the cost-gcc
the time complexity
of achieving domain consistency, being O(n(m + n log n))
Pn
where m = i=1 |Di | and n = |X|.
Note that [4] also consider the variable-based cost measure for a different
version of the soft gcc. Their version considers the parameters l and u to be
variables too. Hence, the variable-based cost evaluation becomes a rather poor
measure, as we trivially can change l and u to satisfy the gcc. They fix this by
restricting the set of variables to consider to be the set X, which corresponds to
our situation. However, they do not provide a filtering algorithm for that case.
3.3
Value-Based Violation
For the value-based violation measure, we adapt the graph G in the following way.
We add arc sets Aunderflow = {(s, d) | d ∈ DX } and Aoverflow = {(d, t) | d ∈ DX },
with demand d(a) = 0 for all a ∈ Aunderflow ∪ Aoverflow and capacity
l if a = (s, d) ∈ Aunderflow ,
c(a) = d
∞ if a ∈ Aoverflow .
Further, we again apply a cost function w : A → R, where
1 if a ∈ Aunderflow ∪ Aoverflow ,
w(a) =
0 otherwise.
Let the resulting graph be denoted by Gval .
Example 3. Consider the CSP
x1 ∈ {1, 2}, x2 ∈ {1}, x3 ∈ {1, 2}, x4 ∈ {1}, z ∈ {0, 1, . . . , 5}
soft gcc[val](X, l, u, z)
minimize z
where X = {x1 , . . . , x4 }, l1 = 1, l2 = 2, u1 = 3 and u2 = 2. In Figure 1.c the
graph Gval for the soft gcc with respect to value-based cost is presented.
Theorem 3. A minimum-cost flow in the graph Gval corresponds to a solution
to the soft gcc[val], minimizing the value-based violation.
Proof. An assignment xi = d corresponds to the arc a = (xi , d) with f (a) = 1.
By construction, all variables need to be assigned to a value and the cost function exactly measures the value-based cost of violation.
Unfortunately, the graph Gval does not preserve the structure of the costgcc because of the arcs Aunderflow . Therefore we cannot blindly apply the same
filtering algorithms. However, it is still possible to design an efficient filtering
algorithm for the value-based soft gcc (in the same spirit of the filtering algorithm for the cost-gcc), based again on flow theory. For this, we need to introduce
the residual graph G f = (V, Af ) of a flow f on G = (V, A) (with respect to c and
d), where
Af = {a | a ∈ A, f (a) < c(a)} ∪ {a−1 | a ∈ A, f (a) > d(a)}.
Here a−1 = (v, u) if a = (u, v). We extend w to A−1 = {a−1 | a ∈ A} by defining
w(a−1 ) = −w(a) for each a ∈ A.
Theorem 4. Let f be a minimum-cost flow in Gval . Then soft gcc[val](X, l, u, z)
is domain consistent if and only if
min Dz ≥ cost(f )
and
cost(f ) + cost(SP(d, xi )) ≤ max Dz ∀(xi , d) ∈ AX→DX ,
where cost(SP(d, xi )) denotes the cost of a shortest path from d to xi in the
f
residual graph Gval
.
Proof. From flow theory [1] we know that, given a minimum-cost flow f in
Gval , if we enforce arc (xi , d) to be in a minimum-cost flow f˜ in Gval , cost(f˜) =
f
cost(f ) + cost(SP(d, xi )) where SP(d, xi ) is the shortest d − xi path in Gval
.
In order for a value d ∈ Di to be consistent, the cost of a minimum-cost flow
that uses (xi , d) should be less than or equal to max Dz . By the above fact, we
only need to compute a shortest path from d to xi instead of a new minimumcost flow.
A minimum-cost flow f inPGval can be computed in O(m(m + n log n)) time
(see [1]), where again m = ni=1 |Di | and n = |X|. Compared to the complexity
of the soft gcc[var], we have a factor m instead of n. This is because computing the flow for soft gcc[val] is dependent on the number of arcs m rather than
on the number variables n. A shortest d − xi path in Gval can be computed in
O(m + n log n) time. Hence the soft gcc with respect to the value-based violation measure can be made domain consistent in O((m − n)(m + n log n)) time
as we need to check m − n arcs for consistency.
When l = 0 in soft gcc[val](X, l, u, z), the arc set Aunderflow is empty. In
that case, Gval has a particular structure, i.e. the only costs appear on arcs from
DX to t. As pointed out in [9] for the soft alldifferent constraint, constraints
with this structure can be checked for consistency in O(nm) time, and domain
consistency can be achieved in O(m) time. The result is obtained by exploiting
the strongly connected components2 in Gval restricted to vertex sets X and DX .
4
Soft Regular Constraint
A regular constraint [12] on a fixed-length sequence of finite-domain variables
requires that the corresponding sequence of values taken by these variables belong to a given regular language. A deterministic finite automaton (DFA) may
be described by a 5-tuple M = (Q, Σ, δ, q0 , F ) where Q is a finite set of states,
Σ is an alphabet, δ : Q × Σ → Q is a partial transition function, q0 ∈ Q is
the initial state, and F ⊆ Q is the set of final (or accepting) states. A finite
sequence of symbols from an alphabet is called a string. Strings processed by
M and ending in an accepting state from F are said to belong to the language
defined by M , denoted L(M ). The languages recognized by DFAs are precisely
regular languages.
Given a sequence x = hx1 , x2 , . . . , xn i of finite-domain variables with respective domains D1 , D2 , . . . , Dn ⊆ Σ, there is a natural interpretation of the set
of possible instantiations of x, D1 × D2 × · · · × Dn , as a subset of all strings of
length n over Σ, Σ n . We are now ready to state the constraint.
Definition 8 (Regular language membership constraint). Let M =
(Q, Σ, δ, q0 , F ) denote a deterministic finite automaton and x a sequence of
finite-domain variables hx1 , x2 , . . . , xn i with respective domains D1 , D2 , . . . ,
2
A strongly connected component in a directed graph G = (V, A) is a subset of vertices
S ⊆ V such that there exists a directed u − v path in G for all u, v ∈ S.
Dn
⊆
Σ. Under a regular language membership constraint
regular(x, M ), any sequence of values taken by the variables of x corresponds
to a string in L(M ).
In [12], a domain consistency algorithm for the regular constraint processed
the sequence x with the automaton M , building a layered directed multi-graph
i
G = (N 1 , N 2 , . . . , N n+1 , A) where each layer N i = {q0i , q1i , . . . , q|Q|−1
} contains
a different node for each state of M and arcs only appear between consecutive
layers. Each arc corresponds to a consistent variable-value pair: there is an arc
from qki to qℓi+1 if and only if there exists some vj ∈ Di such that δ(qk , vj ) = qℓ
and the arc belongs to a path from q0 in the first layer to a member of F in
the last layer. The existence of such an arc, labeled vj , constitutes a support for
variable xi taking value vj .
For example, consider a sequence x of five variables with D1 = {a, b, c, o},
D2 = {b, o}, D3 = {a, c, o}, D4 = {a, b, o}, and D5 = {a}. Figure 2 gives
an automaton M (with its initial state labeled 1) and the resulting graph for
constraint regular(x, M ). As a result, value b is removed from D2 and D4 .
x1
1
a
o
2
5
b
a
b
1
b
3
6
c
4
7
c
x4
1
o
a
2
a
2
2
o
3
4
4
x5
1
a
1
2
a
2
o
3
3
5
o
a
4
3
4
3
a
6
6
6 a
7
7
7
N1
N2
N3
a
4
o
o
5
a
4
5 o
5
5
c
c
o
o
x3
o
a
1
c
3
5
b
c
o
o
o
o
b
c
a
o
x2
1
a
o
2
o
a
c
6
6
6
a
o
7
N4
o
7
7
N5
N6
Fig. 2. A DFA (left) and its layered directed graph G (right).
4.1
Cost Definition
We first give a generic definition for a soft version of the regular constraint.
Definition 9 (Soft regular language membership constraint). Let M =
(Q, Σ, δ, q0 , F ) denote a deterministic finite automaton and x a sequence of
finite-domain variables hx1 , x2 , . . . , xn i with respective domains D1 , D2 , . . . ,
Dn ⊆ Σ. Let z be a finite-domain variable of domain Dz ⊂ N representing the
cost of a violation and let d : Σ ⋆ ×Σ ⋆ → N be some distance function over strings.
Under a soft regular language membership constraint soft regular[d](x, M, z),
for any sequence of values σ taken by the variables of x we have
minσ′ ∈L(M) {d(σ, σ ′ )} = z.
Our first instantiation of the distance function yields the variable-based cost:
Definition 10 (Hamming distance). The number of positions in which two
strings of same length differ is called their Hamming distance.
Intuitively, such a distance represents the number of symbols we need to change
to go from one string to the other, or equivalently the number of variables whose
value must change. Using the Hamming distance for d in the previous definition,
z becomes the variable-based cost.
Another distance function that is often used with strings is the following:
Definition 11 (Edit distance). The smallest number of insertions, deletions,
and substitutions required to change one string into another is called the edit
distance.
It captures the fact that two strings that are identical except for one extra or
missing symbol should be considered close to one another. For example, the edit
distance between strings “bcdea” and “abcde” is two: insert an ’a’ at the front of
the first string and delete the ’a’ from its end. The Hamming distance between
the same strings is five: every symbol must be changed. Edit distance is probably
a better way to measure violations of a regular constraint. We provide a more
natural example in the area of rostering. Given a string, we call stretch a maximal
substring of identical values. We often need to impose restrictions on the length
of stretches of work shifts, and these can be expressed with a regular constraint.
Suppose stretches of a’s and b’s must each be of length 2 and consider the string
“abbaabbaab”: its Hamming distance to a string belonging to the corresponding
regular language is 5 since changing either the first a to a b or b to an a has a
domino effect on the following stretches; its edit distance is just 2 since we can
insert an a at the beginning to make a legal stretch of a’s and remove the b at
the end. In this case, the edit distance reflects the number of illegal stretches
whereas the Hamming distance is proportional to the length of the string.
4.2
Cost Evaluation and Cost-Based Filtering
For both cost measures, we proceed by modifying the layered directed graph
G built for the “hard” version of regular into graph Gvar . Before, we added
an arc from qki to qℓi+1 if δ(qk , vj ) = qℓ for some vj ∈ Di ; now we relax it
slightly to any vj ∈ Σ. This only makes a difference if the domains of the
variables are not initially full. Arcs are never removed in Gvar but their labels
are updated instead. The label of an arc (qki , qℓi+1 ) is generalized to the invariant
Vikℓ = {vj ∈ Di | δ(qk , vj ) = qℓ }; as values are removed from the domain of
variable xi , they are also removed from the corresponding Vikℓ ’s. The cost of
using an arc (qki , qℓi+1 ) for variable-value pair hxi , vj i will be zero if vj belongs to
Vikℓ and some positive integer cost otherwise. This cost represents the penalty for
an individual violation. In the remainder of the section we will consider unit costs
but the framework also makes it possible to use varying costs, e.g. to distinguish
xi
xi
1
1
1
1
2
2
2
2
3
3
3
3
4
4
4
4
5
5
5
5
6
6
6
6
7
7
Ni
Ni+1
i=2,...,5
7
7
Ni
Ni+1
i=2,...,5
Fig. 3. Shorthand versions of Gvar (left) and Gedit (right) for the DFA of Figure
2.
between insertions and substitutions when using the edit distance. The graph on
the left at Figure 3 is a shorthand version of Gvar for the automaton of Figure 2.
Since all values in Σ are considered, the same arcs appear between consecutive
layers. What changes from one layer to the other are the Vikℓ labels.
Taking into account substitutions, common to both Hamming and edit distances, is immediate from the previous modification. It is not difficult to see that
the introduction of costs transforms a supporting path in the domain consistency
algorithm for regular into a zero-cost path in the modified graph. The cost of
a shortest path from q0 in the first layer to a member of F in the last layer
corresponds to the smallest number of variables forced to take a value outside
of their domain.
Theorem 5. A minimum-cost path from u ∈ N 1 to v ∈ N n+1 in Gvar corresponds to a solution to soft regular[var] minimizing the variable-based cost
(Hamming distance).
Just as the existence of a path through a given arc representing a variable-value
pair constituted a support for that pair in the filtering algorithm for regular,
the existence of a path whose cost doesn’t exceed max Dz constitutes a support
for that variable-value pair in a cost-based filtering algorithm for soft regular.
Theorem 6. soft regular[var](x, M, z) is domain consistent on x and bound
consistent on z if and only if
min {cost(SP(q01 , qfn+1 ))} ≤ min Dz
qf ∈F
and
min
qk ∈Q,qf ∈F
{cost(SP(q01 , qki ))+cost(SP(qℓi+1 , qfn+1 ))} ≤ max Dz ,
∀xi ∈ x, vj ∈ Di
where δ(qk , vj ) = qℓ and cost(SP(u, v)) denotes the cost of a shortest path from
u to v in Gvar .
Computing shortest paths from the initial state in the first layer to every
other node and from every node to a final state in the last layer can be done in
O(n |δ|) time3 through topological sorts because of the special structure of the
graph. That computation can also be made incremental in the same way as in
[12]. Recently, that same result was independently obtained in [3]. We however
go further by considering edit distance, for which insertions and deletions are
allowed as well.
For deletions we need to allow “wasting” a value without changing the current
state. To this effect, we add to Gvar an arc (qki , qki+1 ) ∀ 1 ≤ i ≤ n, qk ∈ Q, with
Vikk = ∅, if it isn’t already present in the graph. To allow insertions, inspired by
ǫ-transitions in DFAs, we introduce some special arcs between nodes in the same
layer: if ∃v ∈ Σ such that δ(qk , v) = qℓ then we further add an arc (qki , qℓi ) ∀ 1 ≤
i ≤ n + 1 with fixed positive cost. Figure 3 provides an example of the resulting
graph (on the right). Unfortunately, those special arcs modify the structure of
the graph since cycles (of strictly positive cost) are introduced. Consequently
shortest paths can no longer be computed through topological sorts. An efficient
implementation of Dijkstra’s algorithm increases the time complexity to O(n |δ|+
n |Q| log(n |Q|)). Regardless of this increase in computational cost, Theorems 5
and 6 can be generalized to hold for soft regular[edit] as well.
5
Aggregating Soft Constraints
The preceding sections have introduced filtering algorithms based on different
violation measures for two soft global constraints. If these filtering techniques
are to be effective, especially in the presence of soft constraints of a different
nature, they must be able to cooperate and communicate. Even though there
are many avenues for combining soft constraints, the objective almost always
remains to minimize constraint violations. We propose here a small extension
to the approach of [13], where meta-constraints on the cost variables of soft
constraints are introduced. We illustrate this approach with the newly introduced
soft gcc.
Definition 12 (Soft global cardinality aggregator). Let S be a set of soft
constraints and zi ∈ Dzi the variable indicating the violation cost of Si ∈ S. The
soft global cardinality aggregator (sgca) is defined as soft gcc[⋆](Z, l, u, zagg )
where Z = {z1 , . . . , z|S| }, li , ui is the interval defining the allowed number of
occurrences of each value in the domain of zi and zagg ∈ Dzagg ⊆ N the cost
variable based on the violation measure ⋆.
When all constraints are either satisfied or violated (Z ∈ {0, 1}|S| ) the MaxCSP approach can be easily obtained by setting l1 = 0, u1 = 0, violation(Z) =
3
|δ| refers to the number of transitions in the automaton.
P
d∈DZ overf low(Z, d) and reading the number of violations in zagg . The sgca
could also be used as in [13] to enforce homogeneity (in a soft manner) or to
define other violation measures like restricting the number of highly violated
constraint. For instance, we could wish to impose that no more then a certain
number of constraints are highly violated, but since we cannot guarantee that
this is possible the use of sgca allows to state this wish without risking to
create an inconsistent problem. More generally, by defining the values of l and
u accordingly it is possible to limit (or at least attempt to limit) the number
violated constraints by violation cost. Another approach could be to set all u
to 0 and adjust the violation function so that higher violation costs are more
penalized. The use of soft meta-constraints, when possible, is also an alternative
to the introduction of disjunctive constraints since they need not be satisfied for
the problem to be consistent.
In the original meta-constraint framework, similar behaviour can be established by applying a cost-gcc to Z. For instance, we can define for each pair (zi , d)
(d ∈ Dzi ) a cost d which penalizes higher violations
P more. With the soft gcc,
this cost function can be stated as violation(Z) = d∈DZ d·overflow(Z, d). However, as for this variant of the soft gcc we have l = 0, the soft gcc will be
much more efficient than the cost-gcc, as was discussed at the end of Section 3.
In fact, the sgca can be checked for consistency in O(nm) time and made domain
consistent in O(m)
P time (where n = |S| and m = ∪i |Dzi | whenever l = 0 and
violation(Z) = d∈DZ F (d) · overflow(Z, d) for any cost function F : DZ → R+ .
6
Conclusion
We have presented soft versions of two global constraints: the global cardinality
constraint and the regular constraint. Different violation measures have been
presented and the corresponding filtering algorithms achieving domain consistency have been introduced. These new techniques are based on the addition
of “relaxation arcs” in the underlying graph and the use of known graph algorithms. We also have proposed to extend the Meta-Constraint framework for
combining constraint violations by using the soft version of gcc.
Since these two constraints are very useful to solve Personnel Rostering Problems the next step is thus the implementation of these algorithms in order to
model such problems and benchmark these new constraints.
References
1. R.K. Ahuja, T.L. Magnanti, and J.B. Orlin. Network Flows. Prentice Hall, 1993.
2. K.R. Apt. Principles of Constraint Programming. Cambridge University Press,
2003.
3. N. Beldiceanu, M. Carlsson, and T. Petit. Deriving Filtering Algorithms from
Constraint Checkers. In Proceedings of the Tenth International Conference on
Principles and Practice of Constraint Programming (CP 2004), volume 3258 of
LNCS. Springer, 2004.
4. N. Beldiceanu and T. Petit. Cost Evaluation of Soft Global Constraints. In
CPAIOR 2004: Proceedings of the First International Conference, volume 3011
of LNCS, pages 80–95. Springer, 2004.
5. S. Bistarelli. Semirings for Soft Constraint Solving and Programming, volume 2962
of LNCS. Springer, 2004.
6. D. Dubois, H. Fargier, and H. Prade. The calculus of fuzzy restrictions as a basis
for flexible constraint satisfaction. In Proceedings of the Second IEEE International
Conference on Fuzzy Systems, volume 2, pages 1131–1136, 1993.
7. H. Fargier, J. Lang, and T. Schiex. Selecting preferred solutions in fuzzy constraint
satisfaction problems. In Proceedings of the first European Congress on Fuzzy and
Intelligent Technologies, 1993.
8. R. Freuder and M. Wallace. Partial Constraint Satisfaction. Artificial Intelligence,
58:21–70, 1992.
9. W.J. van Hoeve. A Hyper-Arc Consistency Algorithm for the Soft Alldifferent
Constraint. In Proceedings of the Tenth International Conference on Principles and
Practice of Constraint Programming (CP 2004), volume 3258 of LNCS. Springer,
2004.
10. J. Larrosa. Node and Arc Consistency in Weighted CSP. In Proceedings of
the Eighteenth National Conference on Artificial Intelligence, pages 48–53. AAAI
Press/The MIT Press, 2002.
11. J. Larrosa and T. Schiex. In the quest of the best form of local consistency for
Weighted CSP. In Proceedings of the Eighteenth International Joint Conference
on Artificial Intelligence, pages 239–244. Morgan Kaufmann, 2003.
12. G. Pesant. A Regular Language Membership Constraint for Finite Sequences of
Variables. In Proceedings of the Tenth International Conference on Principles and
Practice of Constraint Programming (CP 2004), volume 3258 of LNCS. Springer,
2004.
13. T. Petit, J.-C. Régin, and C. Bessière. Meta constraints on violations for over
constrained problems. In Proceedings of the 12th IEEE International Conference
on Tools with Artificial Intelligence (ICTAI), pages 358–365, 2000.
14. T. Petit, J.-C. Régin, and C. Bessière. Specific Filtering Algorithms for Over
Constrained Problems. In Proceedings of the Seventh International Conference on
Principles and Practice of Constraint Programming (CP 2001), volume 2239 of
LNCS, pages 451–463. Springer, 2001.
15. J.-C. Régin. Generalized Arc Consistency for Global Cardinality Constraint. In
Proceedings of AAAI/IAAI, volume 1, pages 209–215. AAAI Press/The MIT Press,
1996.
16. J.-C. Régin. Arc Consistency for Global Cardinality Constraints with Costs. In
Proceedings of the Fifth International Conference on Principles and Practice of
Constraint Programming (CP’99), volume 1713 of LNCS, pages 390–404. Springer,
1999.
17. J.-C. Régin. Cost-Based Arc Consistency for Global Cardinality Constraints. Constraints, 7:387–405, 2002.
18. T. Schiex. Possibilistic Constraint Satisfaction Problems or “How to handle soft
constraints ?”. In Proceedings of the 8th Annual Conference on Uncertainty in
Artificial Intelligence, pages 268–275. Morgan Kaufmann, 1992.
19. T. Schiex, H. Fargier, and G. Verfaillie. Valued Constraint Satisfaction Problems:
Hard and Easy Problems. In Proceedings of the Fourteenth International Joint
Conference on Artificial Intelligence, pages 631–639. Morgan Kaufmann, 1995.
20. A. Schrijver. Combinatorial Optimization - Polyhedra and Efficiency. Springer,
2003.
| 6cs.PL
|
S EQ 2SQL: G ENERATING S TRUCTURED Q UERIES
FROM NATURAL L ANGUAGE USING R EINFORCEMENT
L EARNING
arXiv:1709.00103v7 [cs.CL] 9 Nov 2017
Victor Zhong, Caiming Xiong, Richard Socher
Salesforce Research
Palo Alto, CA
{vzhong,cxiong,rsocher}@salesforce.com
A BSTRACT
Relational databases store a significant amount of the worlds data. However, accessing this data currently requires users to understand a query language such as
SQL. We propose Seq2SQL, a deep neural network for translating natural language questions to corresponding SQL queries. Our model uses rewards from inthe-loop query execution over the database to learn a policy to generate the query,
which contains unordered parts that are less suitable for optimization via cross entropy loss. Moreover, Seq2SQL leverages the structure of SQL to prune the space
of generated queries and significantly simplify the generation problem. In addition
to the model, we release WikiSQL, a dataset of 80654 hand-annotated examples
of questions and SQL queries distributed across 24241 tables from Wikipedia that
is an order of magnitude larger than comparable datasets. By applying policybased reinforcement learning with a query execution environment to WikiSQL,
Seq2SQL outperforms a state-of-the-art semantic parser, improving execution accuracy from 35.9% to 59.4% and logical form accuracy from 23.4% to 48.3%.
1
I NTRODUCTION
Relational databases store a vast amount of today’s information and provide the foundation of applications such as medical records (Hillestad et al., 2005), financial markets (Beck et al., 2000),
and customer relations management (Ngai et al., 2009). However, accessing relational databases
requires an understanding of query languages such as SQL, which, while powerful, is difficult to
master. Natural language interfaces (NLI), a research area at the intersection of natural language
processing and human-computer interactions, seeks to provide means for humans to interact with
computers through the use of natural language (Androutsopoulos et al., 1995). We investigate one
particular aspect of NLI applied to relational databases: translating natural language questions to
SQL queries.
Our main contributions in this work are two-fold. First, we introduce Seq2SQL, a deep neural
network for translating natural language questions to corresponding SQL queries. Seq2SQL, shown
in Figure 1, consists of three components that leverage the structure of SQL to prune the output
space of generated queries. Moreover, it uses policy-based reinforcement learning (RL) to generate
the conditions of the query, which are unsuitable for optimization using cross entropy loss due
to their unordered nature. We train Seq2SQL using a mixed objective, combining cross entropy
losses and RL rewards from in-the-loop query execution on a database. These characteristics allow
Seq2SQL to achieve state-of-the-art results on query generation.
Next, we release WikiSQL, a corpus of 80654 hand-annotated instances of natural language questions, SQL queries, and SQL tables extracted from 24241 HTML tables from Wikipedia. WikiSQL is an order of magnitude larger than previous semantic parsing datasets that provide logical forms along with natural language utterances. We release the tables used in WikiSQL both in
raw JSON format as well as in the form of a SQL database. Along with WikiSQL, we release a
query execution engine for the database used for in-the-loop query execution to learn the policy.
On WikiSQL, Seq2SQL outperforms a previously state-of-the-art semantic parsing model by Dong
1
Generated SQL
Seq2SQL
Database
Reward
Predicted results
Question, schema
Ground truth results
Figure 1: Seq2SQL takes as input a question and the columns of a table. It generates the corresponding SQL query, which, during training, is executed against a database. The result of the
execution is utilized as the reward to train the reinforcement learning algorithm.
Table: CFLDraft
Pick # CFL Team
Question:
Player
Position College
27
Hamilton Tiger-Cats Connor Healy
DB
Wilfrid Laurier
28
Calgary Stampeders Anthony Forgone OL
York
29
Ottawa Renegades
L.P. Ladouceur
DT
California
30
Toronto Argonauts
Frank Hoffman
DL
York
…
…
…
…
…
How many CFL teams are from York College?
SQL:
SELECT COUNT CFL Team FROM
CFLDraft WHERE College = “York”
Result:
2
Figure 2: An example in WikiSQL. The inputs consist of a table and a question. The outputs consist
of a ground truth SQL query and the corresponding result from execution.
& Lapata (2016), which obtains 35.9% execution accuracy, as well as an augmented pointer network baseline, which obtains 53.3% execution accuracy. By leveraging the inherent structure of
SQL queries and applying policy gradient methods using reward signals from live query execution,
Seq2SQL achieves state-of-the-art performance on WikiSQL, obtaining 59.4% execution accuracy.
2
M ODEL
The WikiSQL task is to generate a SQL query from a natural language question and table schema.
Our baseline model is the attentional sequence to sequence neural semantic parser proposed by Dong
& Lapata (2016) that achieves state-of-the-art performance on a host of semantic parsing datasets
without using hand-engineered grammar. However, the output space of the softmax in their Seq2Seq
model is unnecessarily large for this task. In particular, we can limit the output space of the generated
sequence to the union of the table schema, question utterance, and SQL key words. The resulting
model is similar to a pointer network (Vinyals et al., 2015) with augmented inputs. We first describe
the augmented pointer network model, then address its limitations in our definition of Seq2SQL,
particularly with respect to generating unordered query conditions.
2.1
AUGMENTED P OINTER N ETWORK
The augmented pointer network generates the SQL query token-by-token by selecting from an input
sequence. In our case, the input sequence is the concatenation of the column names, required for the
selection column and the condition columns of the query, the question, required for the conditions
of the query, and the limited vocabulary of the SQL language such as SELECT, COUNT etc. In
the example shown in Figure 2, the column name tokens consist of “Pick”, “#”, “CFL”, “Team”
etc.; the question tokens consist of “How”, “many”, “CFL”, “teams” etc.; the SQL tokens consist of
SELECT, WHERE, COUNT, MIN, MAX etc. With this augmented input sequence, the pointer network
can produce the SQL query by selecting exclusively from the input.
Suppose we have a list of N table columns and a question such as in Figure 2, and want to produce
the corresponding SQL query. Let xcj = [xcj,1 , xcj,2 , ...xcj,Tj ] denote the sequence of words in the
name of the jth column, where xcj,i represents the ith word in the jth column and Tj represents the
total number of words in the jth column. Similarly, let xq and xs respectively denote the sequence
of words in the question and the set of unique words in the SQL vocabulary.
2
We define the input sequence x as the concatenation of all the column names, the question, and the
SQL vocabulary:
x = [<col>; xc1 ; xc2 ; ...; xcN ; <sql>; xs ; <question>; xq ]
(1)
where [a; b] denotes the concatenation between the sequences a and b and we add sentinel tokens
between neighbouring sequences to demarcate the boundaries.
The network first encodes x using a two-layer, bidirectional Long Short-Term Memory network (Hochreiter & Schmidhuber, 1997). The input to the encoder are the embeddings corresponding to words in the input sequence. We denote the output of the encoder by henc , where henc
is the
t
state of the encoder corresponding to the tth word in the input sequence. For brevity, we do not write
out the LSTM equations, which are described by Hochreiter & Schmidhuber (1997). We then apply
a pointer network similar to that proposed by Vinyals et al. (2015) to the input encodings henc .
The decoder network uses a two layer, unidirectional LSTM. During each decoder step s, the decoder
LSTM takes as input ys−1 , the query token generated during the previous decoding step, and outputs
ptr
the state gs . Next, the decoder produces a scalar attention score αs,t
for each position t of the input
sequence:
ptr
αs,t
= W ptr tanh U ptr gs + V ptr ht
(2)
We choose the input token with the highest score as the next token of the generated SQL query,
ys = argmax(αsptr ).
2.2
S EQ 2SQL
While the augmented pointer network can
How many
solve the SQL generation problem, it does
Seq2SQL
SELECT
engine types did
not leverage the structure inherent in SQL.
Aggregation
Val Musetti use?
COUNT
Typically, a SQL query such as that shown
classifier
Entrant
SELECT column
in Figure 3 consists of three components.
Engine
Constructor
pointer
The first component is the aggregation opChassis
WHERE
clause
WHERE
erator, in this case COUNT, which proEngine
pointer
Driver =
duces a summary of the rows selected by
No
Val Musetti
decoder
Driver
the query. Alternatively the query may request no summary statistics, in which case
an aggregation operator is not provided.
Figure 3: The Seq2SQL model has three components,
The second component is the SELECT
corresponding to the three parts of a SQL query (right).
column(s), in this case Engine, which
The input to the model are the question (top left) and
identifies the column(s) that are to be inthe table column names (bottom left).
cluded in the returned results. The third
component is the WHERE clause of the query, in this case WHERE Driver = Val Musetti,
which contains conditions by which to filter the rows. Here, we keep rows in which the driver is
“Val Musetti”.
Seq2SQL, as shown in Figure 3, has three parts that correspond to the aggregation operator, the
SELECT column, and the WHERE clause. First, the network classifies an aggregation operation
for the query, with the addition of a null operation that corresponds to no aggregation. Next, the
network points to a column in the input table corresponding to the SELECT column. Finally, the
network generates the conditions for the query using a pointer network. The first two components
are supervised using cross entropy loss, whereas the third generation component is trained using
policy gradient to address the unordered nature of query conditions (we explain this in the subsequent WHERE Clause section). Utilizing the structure of SQL allows Seq2SQL to further prune
the output space of queries, which leads to higher performance than Seq2Seq and the augmented
pointer network.
Aggregation Operation. The aggregation operation depends on the question. For the example
shown in Figure 3, the correct operator is COUNT because the question asks for “How many”. To
compute the aggregation operation, we first compute the scalar attention score, αtinp = W inp henc
t ,
for each tth token in the input sequence. We normalize the vector of scores αinp = [α1inp , α2inp , ...] to
3
produce a distribution over the input encodings, β inp = softmax αinp . The input representation
κagg is the sum over the input encodings henc weighted by the normalized scores β inp :
X inp
κagg =
βt henc
(3)
t
t
Let αagg denote the scores over the aggregation operations such as COUNT, MIN, MAX, and the noaggregation operation NULL. We compute αagg by applying a multi-layer perceptron to the input
representation κagg :
αagg = W agg tanh (V agg κagg + bagg ) + cagg
(4)
We apply the softmax function to obtain the distribution over the set of possible aggregation operations β agg = softmax (αagg ). We use cross entropy loss Lagg for the aggregation operation.
SELECT Column. The selection column depends on the table columns as well as the question.
Namely, for the example in Figure 3, “How many engine types” indicates that we need to retrieve
the “Engine” column. SELECT column prediction is then a matching problem, solvable using a
pointer: given the list of column representations and a question representation, we select the column
that best matches the question.
In order to produce the representations for the columns, we first encode each column name with a
LSTM. The representation of a particular column j, ecj , is given by:
hcj,t = LSTM emb xcj,t , hcj,t−1
ecj = hcj,Tj
(5)
hcj,t
Here,
denotes the tth encoder state of the jth column. We take the last encoder state to be ecj ,
column j’s representation.
To construct a representation for the question, we compute another input representation κsel using the same architecture as for κagg (Equation 3) but with untied weights. Finally, we apply a
multi-layer perceptron over the column representations, conditioned on the input representation, to
compute the a score for each column j:
αjsel = W sel tanh V sel κsel + V c ecj
(6)
We normalize the scores with a softmax
function to produce a distribution over the possible SELECT
columns β sel = softmax αsel . For the example shown in Figure 3, the distribution is over the
columns “Entrant”, “Constructor”, “Chassis”, “Engine”, “No”, and the ground truth SELECT column “Driver”. We train the SELECT network using cross entropy loss Lsel .
WHERE Clause. We can train the WHERE clause using a pointer decoder similar to that
described in Section 2.1. However, there is a limitation in using the cross entropy loss
to optimize the network: the WHERE conditions of a query can be swapped and the query
yield the same result. Suppose we have the question “which males are older than 18”
and the queries SELECT name FROM insurance WHERE age > 18 AND gender =
"male" and SELECT name FROM insurance WHERE gender = "male" AND age
> 18. Both queries obtain the correct execution result despite not having exact string match. If
the former is provided as the ground truth, using cross entropy loss to supervise the generation
would then wrongly penalize the latter. To address this problem, we apply reinforcement learning to
learn a policy to directly optimize the expected correctness of the execution result (Equation 7).
Instead of teacher forcing at each step of query generation, we sample from the output distribution to
obtain the next token. At the end of the generation procedure, we execute the generated SQL query
against the database to obtain a reward. Let y = [y 1 , y 2 , ..., y T ] denote the sequence of generated
tokens in the WHERE clause. Let q (y) denote the query generated by the model and qg denote the
ground truth query corresponding to the question. We define the reward R (q (y) , qg ) as
−2, if q (y) is not a valid SQL query
R (q (y) , qg ) = −1, if q (y) is a valid SQL query and executes to an incorrect result
+1, if q (y) is a valid SQL query and executes to the correct result
4
(7)
The loss, Lwhe = −Ey [R (q (y) , qg )], is the negative expected reward over possible WHERE clauses.
We derive the policy gradient for Lwhe as shown by Sutton et al. (2000) and Schulman et al. (2015).
∇Lwhe
= −∇Θ Ey∼py [R (q (y) , qg )]
(8)
Θ
"
#
X
= −Ey∼py R (q (y) , qg ) ∇Θ
(log py (yt ; Θ))
(9)
t
≈
−R (q (y) , qg ) ∇Θ
X
(log py (yt ; Θ))
(10)
t
Here, py (yt ) denotes the probability of choosing token yt during time step t. In equation 10, we
approximate the expected gradient using a single Monte-Carlo sample y.
Mixed Objective Function. We train the model using gradient descent to minimize the objective
function L = Lagg + Lsel + Lwhe . Consequently, the total gradient is the equally weighted sum of
the gradients from the cross entropy loss in predicting the SELECT column, from the cross entropy
loss in predicting the aggregation operation, and from policy learning.
3
W IKI SQL
WikiSQL is a collection of questions, corresponding SQL queries, and SQL tables. A single example in WikiSQL, shown in Figure 2,
contains a table, a SQL query, and the natural language question corresponding to the SQL
query. Table 1 shows how WikiSQL compares
to related datasets. Namely, WikiSQL is the
largest hand-annotated semantic parsing dataset
to date - it is an order of magnitude larger than
other datasets that have logical forms, either in
terms of the number of examples or the number
of tables. The queries in WikiSQL span over Figure 4: Distribution of questions in WikiSQL.
a large number of tables and hence presents an
unique challenge: the model must be able to not only generalize to new queries, but to new table
schema. Finally, WikiSQL contains realistic data extracted from the web. This is evident in the distributions of the number of columns, the lengths of questions, and the length of queries, respectively
shown in Figure 5. Another indicator of the variety of questions in the dataset is the distribution of
question types, shown in Figure 4.
Figure 5: Distribution of table, question, query sizes in WikiSQL.
We collect WikiSQL by crowd-sourcing on Amazon Mechanical Turk in two phases. First, a worker
paraphrases a generated question for a table. We form the generated question using a template,
filled using a randomly generated SQL query. We ensure the validity and complexity of the tables
by keeping only those that are legitimate database tables and sufficiently large in the number of
rows and columns. Next, two other workers verify that the paraphrase has the same meaning as
5
the generated question. We discard paraphrases that do not show enough variation, as measured
by the character edit distance from the generated question, as well as those both workers deemed
incorrect during verification. Section A of the Appendix contains more details on the collection
of WikiSQL. We make available examples of the interface used during the paraphrase phase and
during the verification phase in the supplementary materials. The dataset is available for download
at https://github.com/salesforce/WikiSQL.
The tables, their paraphrases, and SQL queries
are randomly slotted into train, dev, and test
splits, such that each table is present in exactly
one split. In addition to the raw tables, queries,
results, and natural utterances, we also release
a corresponding SQL database and query execution engine.
3.1
Dataset
WikiSQL
Geoquery
ATIS
Freebase917
Overnight
WebQuestions
WikiTableQuestions
E VALUATION
Size
80654
880
5871
917
26098
5810
22033
LF
yes
yes
yes*
yes
yes
no
no
Schema
24241
8
141
81*
8
2420
2108
Let N denote the total number of exComparison between WikiSQL
amples in the dataset, Nex the number Table 1:
The datasets are
of queries that, when executed, result in and existing datasets.
the correct result, and Nlf the number of GeoQuery880 (Tang & Mooney, 2001),
queries has exact string match with the ATIS (Price, 1990), Free917 (Cai & Yates,
ground truth query used to collect the para- 2013), Overnight (Wang et al., 2015), WebQuesphrase.
We evaluate using the execution tions (Berant et al., 2013), and WikiTableQuesaccuracy metric Accex = NNex .
One tions (Pasupat & Liang, 2015). “Size” denotes
downside of Accex is that it is possible the number of examples in the dataset. “LF”
to construct a SQL query that does not indicates whether it has annotated logical forms.
correspond to the question but neverthe- “Schema” denotes the number of tables. ATIS is
less obtains the same result.
For exam- presented as a slot filling task. Each Freebase API
ple, the two queries SELECT COUNT(name) page is counted as a separate domain.
WHERE SSN = 123 and SELECT COUNT(SSN) WHERE SSN = 123 produce the same result if no two people with different names share the SSN 123. Hence, we also use the logical form
accuracy Acclf = NNlf . However, as we showed in Section 2.2, Acclf incorrectly penalizes queries
that achieve the correct result but do not have exact string match with the ground truth query. Due
to these observations, we use both metrics to evaluate the models.
4
E XPERIMENTS
We tokenize the dataset using Stanford CoreNLP (Manning et al., 2014). We use the normalized
tokens for training and revert into original gloss before outputting the query so that generated queries
are executable on the database. We use fixed GloVe word embeddings (Pennington et al., 2014) and
character n-gram embeddings (Hashimoto et al., 2016). Let wxg denote the GloVe embedding and wxc
the character embedding for word x. Here, wxc is the mean of the embeddings of all the character ngrams in x. For words that have neither word nor character embeddings, we assign the zero vector.
All networks are run for a maximum of 300 epochs with early stopping on dev split execution
accuracy. We train using ADAM (Kingma & Ba, 2014) and regularize using dropout (Srivastava
et al., 2014). All recurrent layers have a hidden size of 200 units and are followed by a dropout of
0.3. We implement all models using PyTorch 1 . To train Seq2SQL, we first train a version in which
the WHERE clause is supervised via teacher forcing (i.e. the policy is not learned from scratch) and
then continue training using reinforcement learning. In order to obtain the rewards described in
Section 2.2, we use the query execution engine described in Section 3.
4.1
R ESULT
We compare results against the attentional sequence to sequence neural semantic parser proposed
by Dong & Lapata (2016). This model achieves state of the art results on a variety of semantic parsing datasets, outperforming a host of non-neural semantic parsers despite not using hand-engineered
grammars. To make this baseline even more competitive on our new dataset, we augment their input
1
https://pytorch.org
6
Model
Dev Acclf
Dev Accex
Test Acclf
Test Accex
Baseline (Dong & Lapata, 2016)
Aug Ptr Network
23.3%
44.1%
37.0%
53.8%
23.4%
43.3%
35.9%
53.3%
Seq2SQL (no RL)
Seq2SQL
48.2%
49.5%
58.1%
60.8%
47.4%
48.3%
57.1%
59.4%
Table 2: Performance on WikiSQL. Both metrics are defined in Section 3.1. For Seq2SQL (no RL),
the WHERE clause is supervised via teacher forcing as opposed to reinforcement learning.
with the table schema such that the model can generalize to new tables. We describe this baseline in
detail in Section 2 of the Appendix. Table 2 compares the performance of the three models.
Reducing the output space by utilizing the augmented pointer network improves upon the baseline
by 17.4%. Leveraging the structure of SQL queries leads to another improvement of 3.8%, as is
shown by the performance of Seq2SQL without RL compared to the augmented pointer network.
Finally, training using reinforcement learning based on rewards from in-the-loop query executions
on a database leads to another performance increase of 2.3%, as is shown by the performance of the
full Seq2SQL model.
4.2
A NALYSIS
Limiting the output space via pointer network leads to more accurate conditions. Compared
to the baseline, the augmented pointer network generates higher quality WHERE clause. For example, for “in how many districts was a successor seated on march 4, 1850?”, the baseline generates the condition successor seated = seated march 4 whereas Seq2SQL generates
successor seated = seated march 4 1850. Similarly, for “what’s doug battaglia’s
pick number?”, the baseline generates Player = doug whereas Seq2SQL generates Player
= doug battaglia. The conditions tend to contain rare words (e.g. “1850”), but the baseline
is inclined to produce common words in the training corpus, such as “march” and “4” for date, or
“doug” for name. The pointer is less affected since it selects exclusively from the input.
Incorporating structure reduces invalid
Model
Precision Recall F1
queries. Seq2SQL without RL directly preAug Ptr Network 66.3%
64.4% 65.4%
dicts selection and aggregation and reduces
Seq2SQL
72.6%
66.2% 69.2%
invalid SQL queries generated from 7.9% to
4.8%. A large quantity of invalid queries
Table 3: Performance on the COUNT operator.
result from column names – the generated
query refers to selection columns that are not present in the table. This is particularly helpful when
the column name contain many tokens, such as “Miles (km)”, which has 4 tokens. Introducing a
classifier for the aggregation also reduces the error rate. Table 3 shows that adding the aggregation
classifier improves the precision, recall, and F1 for predicting the COUNT operator. For more queries
produced by the different models, please see Section 3 of the Appendix.
RL generates higher quality WHERE clause that are ordered differently than ground truth.
Training with policy-based RL obtains correct results in which the order of conditions is differs
from the ground truth query. For example, for “in what district was the democratic candidate first
elected in 1992?”, the ground truth conditions are First elected = 1992 AND Party =
Democratic whereas Seq2SQL generates Party = Democratic AND First elected
= 1992. When Seq2SQL is correct and Seq2SQL without RL is not, the latter tends to produce an
incorrect WHERE clause. For example, for the rather complex question “what is the race name of the
12th round trenton, new jersey race where a.j. foyt had the pole position?”, Seq2SQL trained without
RL generates WHERE rnd = 12 and track = a.j. foyt AND pole position =
a.j. foyt whereas Seq2SQL trained with RL correctly generates WHERE rnd = 12 AND
pole position = a.j. foyt.
5
R ELATED W ORK
Semantic Parsing. In semantic parsing for question answering (QA), natural language questions
are parsed into logical forms that are then executed on a knowledge graph (Zelle & Mooney, 1996;
7
Wong & Mooney, 2007; Zettlemoyer & Collins, 2005; 2007). Other works in semantic parsing
focus on learning parsers without relying on annotated logical forms by leveraging conversational
logs (Artzi & Zettlemoyer, 2011), demonstrations (Artzi & Zettlemoyer, 2013), distant supervision (Cai & Yates, 2013; Reddy et al., 2014), and question-answer pairs (Liang et al., 2011). Semantic parsing systems are typically constrained to a single schema and require hand-curated grammars
to perform well2 . Pasupat & Liang (2015) addresses the single-schema limitation by proposing
the floating parser, which generalizes to unseen web tables on the WikiTableQuestions task. Our
approach is similar in that it generalizes to new table schema. However, we do not require access to
table content, conversion of table to an additional graph, nor hand-engineered features/grammar.
Semantic parsing datasets. Previous semantic parsing systems were designed to answer complex
and compositional questions over closed-domain, fixed-schema datasets such as GeoQuery (Tang
& Mooney, 2001) and ATIS (Price, 1990). Researchers also investigated QA over subsets of largescale knowledge graphs such as DBPedia (Starc & Mladenic, 2017) and Freebase (Cai & Yates,
2013; Berant et al., 2013). The dataset “Overnight” (Wang et al., 2015) uses a similar crowdsourcing process to build a dataset of natural language question, logical form pairs, but has only 8
domains. WikiTableQuestions (Pasupat & Liang, 2015) is a collection of question and answers, also
over a large quantity of tables extracted from Wikipedia. However, it does not provide logical forms
whereas WikiSQL does. In addition, WikiSQL focuses on generating SQL queries for questions
over relational database tables and only uses table content during evaluation.
Representation learning for sequence generation. Dong & Lapata (2016)’s attentional sequence
to sequence neural semantic parser, which we use as the baseline, achieves state-of-the-art results
on a variety of semantic parsing datasets despite not utilizing hand-engineered grammar. Unlike
their model, Seq2SQL uses pointer based generation akin to Vinyals et al. (2015) to achieve higher
performance, especially in generating queries with rare words and column names. Pointer models have also been successfully applied to tasks such as language modeling (Merity et al., 2017),
summarization (Gu et al., 2016), combinatorial optimization (Bello et al., 2017), and question answering (Seo et al., 2017; Xiong et al., 2017). Other interesting neural semantic parsing models are
the Neural Programmer (Neelakantan et al., 2017) and the Neural Enquirer (Yin et al., 2016). Mou
et al. (2017) proposed a distributed neural executor based on the Neural Enquirer, which efficiently
executes queries and incorporates execution rewards in reinforcement learning. Our approach is
different in that we do not access table content, which may be unavailable due to privacy concerns.
We also do not hand-engineer model architecture for query execution and instead leverage existing
database engines to perform efficient query execution. Furthermore, in contrast to Dong & Lapata
(2016) and Neelakantan et al. (2017), we use policy-based RL in a fashion similar to Liang et al.
(2017), Mou et al. (2017), and Guu et al. (2017), which helps Seq2SQL achieve state-of-the-art
performance. Unlike Mou et al. (2017) and Yin et al. (2016), we generalize across natural language
questions and table schemas instead of across synthetic questions on a single schema.
Natural language interface for databases. One prominent works in natural language interfaces is
PRECISE (Popescu et al., 2003), which translates questions to SQL queries and identifies questions
that it is not confident about. Giordani & Moschitti (2012) translate questions to SQL by first
generating candidate queries from a grammar then ranking them using tree kernels. Both of these
approaches rely on high quality grammar and are not suitable for tasks that require generalization
to new schema. Iyer et al. (2017) also translate to SQL, but with a Seq2Seq model that is further
improved with human feedback. Seq2SQL outperforms Seq2Seq and uses reinforcement learning
instead of human feedback during training.
6
C ONCLUSION
We proposed Seq2SQL, a deep neural network for translating questions to SQL queries. Our model
leverages the structure of SQL queries to reduce the output space of the model. To train Seq2SQL,
we applied in-the-loop query execution to learn a policy for generating the conditions of the SQL
query, which is unordered and unsuitable for optimization via cross entropy loss. We also introduced
WikiSQL, a dataset of questions and SQL queries that is an order of magnitude larger than comparable datasets. Finally, we showed that Seq2SQL outperforms a state-of-the-art semantic parser
on WikiSQL, improving execution accuracy from 35.9% to 59.4% and logical form accuracy from
23.4% to 48.3%.
2
For simplicity, we define table schema as the names of the columns in the table.
8
R EFERENCES
I. Androutsopoulos, G.D. Ritchie, and P. Thanisch. Natural language interfaces to databases - an
introduction. 1995.
Yoav Artzi and Luke S. Zettlemoyer. Bootstrapping semantic parsers from conversations. In
EMNLP, 2011.
Yoav Artzi and Luke S. Zettlemoyer. Weakly supervised learning of semantic parsers for mapping
instructions to actions. TACL, 1:49–62, 2013.
Thorsten Beck, Asli Demirgüç-Kunt, and Ross Levine. A new database on the structure and development of the financial sector. The World Bank Economic Review, 14(3):597–605, 2000.
Irwan Bello, Hieu Pham, Quoc V. Le, Mohammad Norouzi, and Samy Bengio. Neural combinatorial
optimization with reinforcement learning. ICLR, 2017.
Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. Semantic parsing on freebase from
question-answer pairs. In EMNLP, 2013.
Chandra Bhagavatula, Thanapon Noraset, and Doug Downey. Methods for exploring and mining
tables on wikipedia. In IDEA@KDD, 2013.
Qingqing Cai and Alexander Yates. Large-scale semantic parsing via schema matching and lexicon
extension. In ACL, 2013.
Li Dong and Mirella Lapata. Language to logical form with neural attention. ACL, 2016.
Alessandra Giordani and Alessandro Moschitti. Translating questions to SQL queries with generative parsers discriminatively reranked. In COLING, 2012.
Jiatao Gu, Zhengdong Lu, Hang Li, and Victor O. K. Li. Incorporating copying mechanism in
sequence-to-sequence learning. ACL, 2016.
Kelvin Guu, Panupong Pasupat, Evan Zheran Liu, and Percy Liang. From language to programs:
Bridging reinforcement learning and maximum marginal likelihood. In ACL, 2017.
Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsuruoka, and Richard Socher. A Joint Many-Task
Model: Growing a Neural Network for Multiple NLP Tasks. arXiv, cs.CL 1611.01587, 2016.
Richard Hillestad, James Bigelow, Anthony Bower, Federico Girosi, Robin Meili, Richard Scoville,
and Roger Taylor. Can electronic medical record systems transform health care? potential health
benefits, savings, and costs. Health affairs, 24(5):1103–1117, 2005.
Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural computation, 1997.
Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, Jayant Krishnamurthy, and Luke Zettlemoyer.
Learning a neural semantic parser from user feedback. In ACL, 2017.
Diederik P. Kingma and Jimmy Ba.
abs/1412.6980, 2014.
Adam: A method for stochastic optimization.
arXiv,
Chen Liang, Jonathan Berant, Quoc V. Le, Ken Forbus, and Ni Lao. Neural symbolic machines:
Learning semantic parsers on freebase with weak supervision. In ACL, 2017.
Percy Liang, Michael I. Jordan, and Dan Klein. Learning dependency-based compositional semantics. Computational Linguistics, 39:389–446, 2011.
Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David
McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pp. 55–60, 2014.
Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture
models. ICLR, 2017.
9
Lili Mou, Zhengdong Lu, Hang Li, and Zhi Jin. Coupling distributed and symbolic execution for
natural language queries. In ICML, 2017.
Arvind Neelakantan, Quoc V. Le, Martı́n Abadi, Andrew McCallum, and Dario Amodei. Learning
a natural language interface with neural programmer. In ICLR, 2017.
Eric WT Ngai, Li Xiu, and Dorothy CK Chau. Application of data mining techniques in customer
relationship management: A literature review and classification. Expert systems with applications,
36(2):2592–2602, 2009.
Panupong Pasupat and Percy Liang. Compositional semantic parsing on semi-structured tables. In
ACL, 2015.
Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word
representation. In EMNLP, 2014.
Ana-Maria Popescu, Oren Etzioni, and Henry Kautz. Towards a theory of natural language interfaces
to databases. In Proceedings of the 8th International Conference on Intelligent User Interfaces,
pp. 149–157. ACM, 2003.
Patti J. Price. Evaluation of spoken language systems: The ATIS domain. 1990.
Siva Reddy, Mirella Lapata, and Mark Steedman. Large-scale semantic parsing without questionanswer pairs. TACL, 2:377–392, 2014.
John Schulman, Nicolas Heess, Theophane Weber, and Pieter Abbeel. Gradient estimation using
stochastic computation graphs. In NIPS, 2015.
Min Joon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. Bidirectional attention
flow for machine comprehension. ICLR, 2017.
Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning
Research, 15:1929–1958, 2014.
Janez Starc and Dunja Mladenic. Joint learning of ontology and semantic parser from text. Intelligent Data Analysis, 21:19–38, 2017.
Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient
methods for reinforcement learning with function approximation. In Advances in neural information processing systems, pp. 1057–1063, 2000.
Lappoon R. Tang and Raymond J. Mooney. Using multiple clause constructors in inductive logic
programming for semantic parsing. In ECML, 2001.
Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In NIPS, 2015.
Yushi Wang, Jonathan Berant, and Percy Liang. Building a semantic parser overnight. In ACL,
2015.
Yuk Wah Wong and Raymond J. Mooney. Learning synchronous grammars for semantic parsing
with lambda calculus. In ACL, 2007.
Caiming Xiong, Victor Zhong, and Richard Socher. Dynamic coattention networks for question
answering. ICLR, 2017.
Pengcheng Yin, Zhengdong Lu, Hang Li, and Ben Kao. Neural enquirer: Learning to query tables.
In ACL, 2016.
John M. Zelle and Raymond J. Mooney. Learning to parse database queries using inductive logic
programming. In AAAI/IAAI, Vol. 2, 1996.
Luke S. Zettlemoyer and Michael Collins. Learning to map sentences to logical form: Structured
classification with probabilistic categorial grammars. In Uncertainty in Artificial Intelligence,
2005.
Luke S. Zettlemoyer and Michael Collins. Online learning of relaxed ccg grammars for parsing to
logical form. In EMNLP-CoNLL, 2007.
10
A
C OLLECTION OF W IKI SQL
WikiSQL is collected in a paraphrase phases as well as a verification phase. In the paraphrase
phase, we use tables extracted from Wikipedia by Bhagavatula et al. (2013) and remove small tables
according to the following criteria:
• the number of cells in each row is not the same
• the content in a cell exceed 50 characters
• a header cell is empty
• the table has less than 5 rows or 5 columns
• over 40% of the cells of a row contain identical content
We also remove the last row of a table because a large quantity of HTML tables tend to have summary statistics in the last row, and hence the last row does not adhere to the table schema defined by
the header row.
For each of the table that passes the above criteria, we randomly generate 6 SQL queries according
to the following rules:
• the query follows the format SELECT agg op agg col from table where
cond1 col cond1 op cond1 AND cond2 col cond2 op cond2 ...
• the aggregation operator agg op can be empty or COUNT. In the event that the aggregation
column agg col is numeric, agg op can additionally be one of MAX and MIN
• the condition operator cond op is =. In the event that the corresponding condition column
cond col is numeric, cond op can additionally be one of > and <
• the condition cond can be any possible value present in the table under the corresponding
cond col. In the event that cond col is numerical, cond can be any numerical value
sampled from the range from the minimum value in the column to the maximum value in
the column.
We only generate queries that produce a non-empty result set. To enforce succinct queries, we
remove conditions from the generated queries if doing so does not change the execution result.
For each query, we generate a crude question using a template and obtain a human paraphrase via
crowdsourcing on Amazon Mechanical Turk. In each Amazon Mechanical Turk HIT, a worker is
shown the first 4 rows of the table as well as its generated questions and asked to paraphrase each
question.
After obtaining natural language utterances from the paraphrase phase, we give each questionparaphrase pair to two other workers in the verification phase to verify that the paraphrase and
the original question contain the same meaning.
We then filter the initial collection of paraphrases using the following criteria:
• the paraphrase must be deemed correct by at least one worker during the verification phrase
• the paraphrase must be sufficiently different from the generated question, with a characterlevel edit distance greater than 10
B
ATTENTIONAL S EQ 2S EQ N EURAL S EMANTIC PARSER BASELINE
We employ the attentional sequence to sequence model for the baseline. This model by Dong &
Lapata (2016) achieves state of the art results on a variety of semantic parsing datasets despite not
using hand-engineered grammar. We implement a variant using OpenNMT and a global attention
encoder-decoder architecture (with input feeding) described by Luong et al.
11
We use the same two-layer, bidirectional, stacked LSTM encoder as described previously. The
decoder is almost identical to that described by Equation 2 of the paper, with the sole difference
coming from input feeding.
gs = LSTM
emb (ys−1 ) ; κdec
s−1 , gs−1
(11)
where κdec
s is the attentional context over the input sequence during the sth decoding step, computed
as
dec
αs,t
= hdec
W dec henc
s
t
κs =
|
βsdec = softmax αsdec
X
βs,t henc
t
(12)
(13)
t
To produce the output token during the sth decoder step, the concatenation of the decoder state and
the attention context is given to a final linear layer to produce a distribution αdec over words in the
target vocabulary
dec
αdec = softmax U dec [hdec
s ; κs ]
(14)
During training, teacher forcing is used. During inference, a beam size of 5 is used and generated
unknown words are replaced by the input words with the highest attention weight.
C
P REDICTIONS BY S EQ 2SQL
Q
P
S’
S
G
when connecticut & villanova are the regular season winner how many tournament venues (city) are there?
SELECT COUNT tournament player (city) WHERE regular season winner city ) = connecticut & villanova
SELECT COUNT tournament venue (city) WHERE tournament winner = connecticut & villanova
SELECT COUNT tournament venue (city) WHERE regular season winner = connecticut & villanova
SELECT COUNT tournament venue (city) WHERE regular season winner = connecticut & villanova
Q
P
S’
S
G
what are the aggregate scores of those races where the first leg results are 0-1?
SELECT aggregate WHERE 1st . = 0-1
SELECT COUNT agg. score WHERE 1st leg = 0-1
SELECT agg. score WHERE 1st leg = 0-1
SELECT agg. score WHERE 1st leg = 0-1
Q
P
S’
S
G
what is the race name of the 12th round trenton, new jersey race where a.j. foyt had the pole position?
SELECT race name WHERE location = 12th AND round position = a.j. foyt, new jersey AND
SELECT race name WHERE rnd = 12 AND track = a.j. foyt AND pole position = a.j. foyt
SELECT race name WHERE rnd = 12 AND pole position = a.j. foyt
SELECT race name WHERE rnd = 12 AND pole position = a.j. foyt
Q
P
S’
S
G
what city is on 89.9?
SELECT city WHERE frequency = 89.9
SELECT city of license WHERE frequency = 89.9
SELECT city of license WHERE frequency = 89.9
SELECT city of license WHERE frequency = 89.9
Q
P
S’
S
G
how many voters from the bronx voted for the socialist party?
SELECT MIN % party = socialist
SELECT COUNT the bronx where the bronx = socialist
SELECT COUNT the bronx WHERE the bronx = socialist
SELECT the bronx WHERE party = socialist
Q
P
S’
S
G
in what year did a plymouth vehicle win on february 9 ?
SELECT MIN year (km) WHERE date = february 9 AND race time = plymouth 9
SELECT year (km) WHERE date = plymouth 9 AND race time = february 9
SELECT year (km) WHERE date = plymouth 9 AND race time= february 9
SELECT year (km) WHERE manufacturer = plymouth AND date = february 9
Table 4: Examples predictions by the models on the dev split. Q denotes the natural language
question and G denotes the corresponding ground truth query. P, S’, and S denote, respectively, the
queries produced by the Augmented Pointer Network, Seq2SQL without reinforcement learning,
Seq2SQL. We omit the FROM table part of the query for succinctness.
12
| 2cs.AI
|
arXiv:1509.01568v2 [math.GR] 8 Apr 2018
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI
NUMBER OF A GROUP
AKRAM YOUSOFZADEH
Abstract. The Tarski number of a group G is the minimal number of the
pieces of paradoxical decompositions of that group. Using configurations along
with a matrix combinatorial property we construct paradoxical decompositions. We also compute an upper bound for the Tarski number of a given
non-amenable group by counting the number of paths in a diagram associated
to the group.
1. Introduction
Rosenblatt and Willis introduced a concept for groups to show that for an infinite
discrete amenable group or a non-discrete amenable group G a net of positive,
normalized functions in L1 (G) can be constructed such that this net converges
weak* to invariance but does not converge strongly to invariance [8]. This concept
which is called configuration is also used to classify some group theoretical properties
(see for example [1, 2]).
Configurations are strongly linked to the amenability of groups and by Tarski’s
alternative, a discrete group is non-amenable if and only if it admits a paradoxical
decomposition. Therefore it is valuable to construct the paradoxical decomposition
for such a group, using configurations. This problem which was originally asked
by Willis is answered partially in [7]. In that paper the paradoxical decomposition
was constructed under a paradoxical condition.
In the present paper we define a general matrix combinatorial property under
which the paradoxical decomposition is completely constructed. We also find a new
upper bound for the Tarski number of a given non-amenable group.
Notations. The following notations are used throughout this paper
•
•
•
•
•
•
N,
F Z and R are the sets of natural, integer and real numbers, respectively,
is the disjoint union of sets,
gA = {ga; a ∈ A}, for a group G, A ⊆ G and g ∈ G,
P(X) is the power set of the set X,
|X| is the the cardinal number of the set X,
A (0, 1)-matrix is a matrix with entries in {0, 1}.
If v is a vector of real numbers and n ∈ Z, by v ≥ n we mean that v has entries
equal to or greater than n.
1
2
AKRAM YOUSOFZADEH
2. Preliminaries
2.1. Matrix theory. We use Gordan’s theorem to find nontrivial nonnegative
solutions to a homogenous system of linear equations with real coefficients. This
theorem has also applications in linear programming [4].
Theorem 2.1. (Gordan 1873) Either a linear homogenous system of equations
AX = 0 possesses a nontrivial solution in nonnegative variables or there exists an
equation, formed by taking some linear combination of equations, that all positive
coefficients. That is, either there exists an x such that
Ax = 0,
0 6= x ≥ 0
t
or there exists a vector m such that m A > 0 (has positive entries).
Remark 2.2. In Theorem 2.1, if A is a matrix with rational entries, then the
entries of m can be chosen in Z.
The main theorem of this paper is proved under a certain condition. To clarify
this condition, we need some definitions.
Let π : {1, . . . , n} → {1, . . . , n} be a permutation for the set {1, . . . , n}. Then
eπ(1)
eπ(2)
Pπ = .
..
eπ(n)
is called the permutation matrix associated to π, where ei denotes the row vector
of length n with 1 in the i-th position and 0 otherwise. When the permutation
matrix Pπ is multiplied with a matrix M from left, Pπ M will permute the rows of
M by π.
P1
P2
If P = . is a permutation matrix, by P + we mean the matrix with
..
Pn
shifted rows, i.e.
P2
P3
P + = Pρ P = ... ,
Pn
P1
in which ρ is the cyclic permutation (1 2 . . . n). Throughout we use the notation
1 0 0 ... 0
1 1 0 ... 0
X
T = .
=
Eij ,
..
1≤j≤i≤n
1 1 1 ... 1
where Eij is the matrix with 1 in ij position and 0 otherwise. When the matrix T
is multiplied with a matrix M from left, j-th row of T M will be the sum of j first
rows of M .
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
Definition 2.3. Let ℓ ∈ N and
A1
A2
..
.
and
B1
B2
..
.
3
be two (0, 1)-matrices with
A
Bn
Pn n
rows Ai , Bi . Let also the vector i=1 (Bi − Ai ) has strictly positive entries. If there
exists a permutation matrix Pπ such that the matrix
(1)
T Pπ (B − A) − Pπ+ A
has integer entries equal or greater than −1, we say that the homogenous system
of equations (B − A)X = 0 is normal.
Pn
It is apparent that if i=1 (Bi − Ai ) has strictly positive entries, then the system
has no non-zero non-negative solution. Conversely, if (B − A)X = 0 is a system
of equations with no non-zero non-negative solution, then by Gordan’s theorem
there exists a vector m = (m1 , . . . , mn ) such that mt (B − A) has strictly positive
entries. If we permit B − A to have repeated rows and we insert the opposite of a
row (exchanging the corresponding rows of A and B) if necessary, then mi can be
chosen in {0, 1}. Omit Bi − Ai , whereP
mi = 0 and denote the modified matrix by
n
B − A again. So, we can assume that i=1 (Bi − Ai ) has strictly positive entries.
It is to be noted that there exist examples of both normal and non-normal
systems of equations with no nonzero non-negative solution. We will explain the
relation between normality and paradoxical decompositions in section 3.
2.2. Non-amenable discrete groups. Let G be a discrete group. Then G is
called amenable if it admits a finitely additive probability measure µ on the σ−algebra
P(G) such that
µ(gA) = µ(A),
(A ⊆ G, g ∈ G).
Definition 2.4. [10] Let G be a group acting on a set X and suppose E ⊆ X.
E is G-paradoxical (or, paradoxical with respect to G) if for some positive integers m, n there are pairwise disjoint subsets A1 , . . . , An , B1 , . . . , Bm of E and
g1 . . . , gn , h1 . . . , hm ∈ G such that
(2)
E=
n
G
i=1
gi Ai =
m
G
h j Bi .
j=1
A group G is called paradoxical if it is G-paradoxical, where G acts on itself
by left multiplication. Clearly if G is a paradoxical group satisfying the above
definition, then it cannot be amenable. Indeed if µ is a G−invariant probability
measure, then
X
X
1 = µ(G) =
µ(Ai ) +
µ(Bj )
X
X
=
µ(gi Ai ) +
µ(hj Bj ) = µ(G) + µ(G) = 2.
In fact there is the following remarkable alternative due to Alfred Tarski.
Theorem 2.5. Let G be a discrete group. Exactly one of the following happens
1) G is paradoxical,
2) G is amenable.
AKRAM YOUSOFZADEH
4
The number τ = n + m for n and m in (2) is called the Tarski number of that
paradoxical decomposition; the minimum of all such numbers over all the possible
paradoxical decompositions of G, is called the Tarski number of G and denoted by
τ (G). In the case that there is no paradoxical decomposition, we set τ (G) = ∞.
The Tarski number of a group is of real interest and has been estimated for some
classes of groups. But it is not so easy to compute in many cases. For more details
see [9] and [3].
We draw the reader’s attention to the next proposition for different types of
paradoxical decompositions.
Proposition 2.6. [7, Proposition 1.2] Let G be a group. Then the following statements are equivalent
1) There exist a partition {A1 , . . . , An , B1 , . . . , Bm } of G and g1 , . . . , gn and
h1 , . . . , hm in G such that {gi Ai }ni=1 and {hj Bj }m
j=1 form partitions of G.
2) There exist pairwise disjoint subsets A1 , . . . , An , B1 , . . . , Bm of G and elements g1 , . . . , gn and h1 , . . . , hm in G such that {gi Ai }ni=1 and {hj Bj }m
j=1 form
partitions of G.
3) There exist pairwise disjoint subsets A1 , . . . , AnS, B1 , . . . , BmSof G and elen
m
ments g1 , . . . , gn and h1 , . . . , hm in G such that G = i=1 gi Ai = j=1 hj Bj (not
necessarily pairwise disjoint).
Because of the above equivalence, G admits a paradoxical decomposition, if any
condition of Proposition 2.6 holds. The decomposition in condition 1 is called a
complete paradoxical decomposition. By [5, Theorem A.1] if the conditions 2 or 3
are satisfied for a group G, then τ (G) ≤ m + n.
2.3. Configuration of groups. Let G be a discrete group. The configurations
of G are defined in terms of finite generating sets and finite partitions of G. If
g = (g1 , . . . , gn ) is a string of elements of G and E = {E1 , . . . , Em } is a partition
of G, a configuration corresponding to (g, E) is an (n + 1)−tuple C = (c0 , . . . , cn ),
where 1 ≤ ci ≤ m for each i, such that there is x in G with x ∈ Ec0 and gi x ∈ Eci for
each 1 ≤ i ≤ n. The set of all configurations corresponding to the pair (g, E) will be
denoted by Con(g, E). It is shown that groups with the same set of configurations
have some common properties. For example they obey the same semigroup laws
and have the same Tarski numbers (see [1] and [11]).
In the case that g = {g1 , . . . , gn } is a generating set for G, the configuration
C = (c0 , . . . , cn ) may be described as a labelled tree which is a subgraph of the
Cayley graph of the finitely generated group G and configuration set Con(g, E) is
a set of rooted trees having height 1. In last section of the paper we assign a new
graph to G that depends on the pair (g, E).
If (g, E) is as above and for each C ∈ Con(g, E)
x0 (C) = Ec0 ∩ (∩nj=1 gj−1 Ecj )
and
xj (C) = gj x0 (C),
then it is seen that for any 0 ≤ j ≤ n, {xj (C); C ∈PCon(g, E)} is a partition for
G. Let C ∈ Con(g, E) and f ∈ ℓ1 (G). Define fC = x∈x0 (C) f (x). Then we have
(see [8])
(1 ≤ j ≤ n, 1 ≤ i ≤ m)
hf −gj f, χEi i = 0,
if and only if
X
X
fC
(1 ≤ j ≤ n, 1 ≤ i ≤ m).
fC =
x0 (C)⊆Ei
xj (C)⊆Ei
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
5
For each pair (g, E) for G, the system of equations
X
X
fC ,
(1 ≤ i ≤ m, 0 ≤ j, k ≤ n)
fC =
xj (C)⊆Ei
xk (C)⊆Ei
with variables fC , C ∈ Con(g, E) is called the system of configuration equations
corresponding to (g, E) and is denoted by Eq(g, E). By a normalized solution
P to this
system, we mean a solution (fC )C such that for each C, fC ≥ 0 and C fC = 1.
A group G isR amenable if and only if there is a net (fα )α of positive functions in
L1 (G) with fα = 1 which is weak* convergence to invariance, that is for any
F ∈ L∞ (G) and any g ∈ G
limhfα −g fα , F i = 0
α
(see [6]). Using this fact, Rosenblatt and Willis proved the following theorem
Theorem 2.7. [8, Proposition 2.4] There is a normalized solution of every possible
instance of the configuration equations if and only if G is amenable.
We will apply this theorem to find paradoxical decompositions.
3. Main Theorem
Throughout this section by G, g = (g1 , . . . , gn ) and E = {E1 , . . . , Em } we mean
a group, a finite string of elements of G and a finite partition for G, respectively.
If D is a subset of Con(g, E), we use the following notation
G
D̃ :=
x0 (C).
C∈D
˜ E) = G. Clearly D1 , D2 ⊆ Con(g, E) are disjoint if and only
In particular Con(g,
P
P
if D̃1 and D̃2 are. The configuration equation xj (C)⊆Ei fC = xk (C)⊆Ei fC is
written in the form aX = bX, where Con(g, E) = {C1 , . . . , Cℓ },
fC1
fC2
X = . ,
..
fCℓ
a is the coefficient vector of the left hand side and b is the coefficient vector of the
right hand side of the equation.
Theorem 3.1. If a subsystem of Eq(g, E) is normal, then G is non-amenable and
a paradoxical decomposition of G can be written in terms of g and E.
V1
W1
V2
W2
Proof. Let V = . and W = . be (0, 1)-matrices in Mp×ℓ (R) such
..
..
Vp
Wp
that (W − V )X = 0 is the normal subsystem of Eq(g, E) satisfying (1)for a permutation matrix Pπ . Write
Vt = (Vt (C))C∈Con(g,E)
and
Wt = (Wt (C))C∈Con(g,E) ,
(1 ≤ t ≤ p).
AKRAM YOUSOFZADEH
6
Then by the normality of (W − V )X = 0, for each C ∈ Con(g, E)
(3)
αC :=
p
X
Wt (C) −
t=1
p
X
Vt (C) > 0.
t=1
We consider only the case that Pπ = I, the identity matrix. Otherwise defining
Vt∗ = Vπ(t) and Wt∗ = Wπ(t) we see that
T Pπ (W − V ) − Pπ+ V = T (W ∗ − V ∗ ) − V ∗ +
and the proof is carried out in a similar way. Therefore throughout the proof we
assume
W1 (C) − (V1 (C) + V2 (C)) ≥ − 1,
(W1 (C) + W2 (C)) − (V1 (C) + V2 (C) + V3 (C)) ≥ − 1,
..
.
p
p−1
X
X
Vt (C)) ≥ − 1.
Wt (C)) −
(
t=1
t=1
Suppose that (i1 , i2 , . . . , ip ), (ji1 , ji2 , . . . , jip ) and (ki1 , ki2 , . . . , kip ) are strings such
that is ∈ {1, . . . , m} and jis , kis ∈ {0, 1, . . . , n} and the modified system is
X
X
fC ,
(1 ≤ t ≤ p).
fC =
xki (C)⊆Eit
xji (C)⊆Eit
t
t
Note that the strings are used instead of subsets since the repetition is not excluded
for the equations. For convenience we use the following notations for 1 ≤ t ≤ p
At = {C; xkit (C) ⊆ Eit } and Bt = {C; xjit (C) ⊆ Eit },
In other words, the system can be written as
X
X
fC ,
fC =
(4)
(1 ≤ t ≤ p).
(1 ≤ t ≤ p).
C∈Bt
C∈At
It is not difficult to see that for 1 ≤ t ≤ p
G
G
(5)
ġt
x0 (C ∗ ) =
x0 (C),
C ∗ ∈At
C∈Bt
gjit . For C ∈ Bt and C ∗ ∈ At put
where ġt = gk−1
i
t
Et (C ∗ , C) = x0 (C ∗ ) ∩ ġt−1 x0 (C).
Then we have by (5)
(6)
x0 (C) =
G
ġt Et (C ∗ , C),
(C ∈ Bt ).
C ∗ ∈At
Case 1. The Fsets At are pairwise disjoint. According to (3) and (4), for
p
C
each C ∈ A := t=1 At there exist at least two numbers 1 ≤ tC
1 , t2 ≤ p such that
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
7
C ∈ BtC1 and C ∈ BtC2 . Also for each C ∈ B := Con(g, E) \ A there exists at least
one 1 ≤ tC ≤ p such that C ∈ BtC . This fact together with (6) implies that
G
G
ġtC1 EtC1 (C ∗ , C) =
ġtC2 EtC2 (C ∗ , C),
(C ∈ A)
x0 (C) =
C ∗ ∈AtC
C ∗ ∈AtC
1
2
and
x0 (C) =
G
C ∗ ∈A
Consequently the fact that
G=
"
ġtC EtC (C ∗ , C),
(C ∈ B).
tC
G
#
x0 (C)
C∈A
G
"
G
#
x0 (C)
C∈B
necessitates the following two equations
"
#
G
G
G G
∗
G=
ġtC1 EtC1 (C , C)
(7)
x0 (C)
C∈A C ∗ ∈AtC
C∈B
1
and
(8)
G
G=
G
C∈A C ∗ ∈AtC
2
G
G
ġtC2 EtC2 (C ∗ , C)
G
C∈B C ∗ ∈AtC
ġtC EtC (C ∗ , C) .
We emphasize that all the sets in the following families are pairwise disjoint
(1) x0 (C), with C ∈ B,
(2) EtC (C ∗ , C), with C ∈ B, C ∗ ∈ AtC ,
(3) EtC1 (C ∗ , C), with C ∈ A, C ∗ ∈ AtC1 ,
(4) EtC2 (C ∗ , C), with C ∈ A, C ∗ ∈ AtC2 .
We know that A ∩ B = ∅ and Ai ∩ Aj = ∅ for i 6= j. On the other hand for
C ∗ ∈ At , Et (C ∗ , C) ⊆ x0 (C ∗ ) ⊆ Ãt ⊆ Ã. Therefore the sets of types (2), (3) and
C
(4) are all disjoint from the sets of type (1). Since tC
(C ∗ , C) ⊆ AtC1
1 6= t2 , EtC
1
∗
∗
and EtC2 (C , C) ⊆ AtC2 and AtC1 ∩ AtC2 = ∅, we have EtC1 (C , C) ∩ EtC2 (C ∗ , C) = ∅.
′
Now let i ∈ {1, 2} and tC
= tC , for C ∈ B and C ′ ∈ A. Let C ∗ ∈ AtC and
i
−1
′∗
′
C ∈ AtC ′ . Applying A ∩ B = ∅ once again, we see gt−1
C x0 (C) ∩ g C ′ x0 (C ) = ∅.
t
i
i
Thus EtC ′ (C ′∗ , C ′ ) ∩ EtC (C ∗ , C) = ∅. Therefore all the sets of types (3) and (4) are
i
disjoint from the sets of type (2). It remains to show that for distinct configurations
C, D ∈ A, i, j ∈ {1, 2}, C ∗ ∈ AtCi and D∗ ∈ AtD
one has EtCi (C ∗ , C)∩EtD
(D∗ , D) =
j
j
D
∅. But it is clear since if tC
i 6= tj , then
EtCi (C ∗ , C) ∩ EtD
(D∗ , D) ⊆ x0 (C ∗ ) ∩ x0 (D∗ ) ⊆ ÃtCi ∩ ÃtD
=∅
j
j
D
and if tC
i = tj , then
−1
EtCi (C ∗ , C) ∩ EtD
(D∗ , D) ⊆ ġt−1
C (x0 (C) ∩ x0 (D)) = ġtC (∅) = ∅.
j
i
i
Consequently the equations (7) and (8) form a paradoxical decomposition of G.
Case 2. The sets At are not pairwise disjoint. The procedure in case 1 does
not work here because the sets in a paradoxical decomposition must be disjoint.
AKRAM YOUSOFZADEH
8
In the sequel we shall replace the sets A1 , . . . , Ap with new sets P1 , . . . , Pp , which
are
Fp disjoint and then a paradoxical decomposition with respect to a partition of
i=1 P̃i will be given.
For C ∈ Con(g, E) if t0 is the least number in {1, . . . , p} such that Vt0 (C) 6= 0,
define Vt′0 (C) = 1 and for t 6= t0 put Vt′ (C) = 0. Then
(
Pp
p
X
1, if
′
t=1 Vt (C) > 0
Vt (C) =
(9)
0, otherwise.
t=1
We need a counter to compute the number of appearance of a configuration in
V1 , . . . , Vp and W1 , . . . , Wp . To achieve this purpose define
(P
Pp
p
( t=1 Vt (C)) − 1, if
0
t=1 Vt (C) > 0
(10)
mC =
0,
otherwise
and for 1 ≤ t ≤ p
mtC
(11)
=
(
t−1
mt−1
>0
C − 1, if Wt (C) = 1, mC
t−1
mC ,
otherwise.
Consequently by (10) we have for C ∈ Con(g, E)
p
X
(12)
Vt (C) −
p
X
Vt′ (C) = m0C .
t=1
t=1
Consider the vectors (Wt′ (C))Con(g,E) , 1 ≤ t ≤ p with the following definition
(
Wt (C) − 1, if Wt (C) = 1, mt−1
>0
′
C
Wt (C) =
Wt (C),
otherwise.
It is seen by (11) that for every C ∈ Con(g, E)
p
X
Wt (C) −
t=1
p
X
Wt′ (C) =
p
X
[Wt (C) − Wt′ (C)]
t=1
t=1
=
=
=
p
X
t
[mt−1
C − mC ]
t=1
m0C −
m0C .
mpC
Thus by (12) and (3)
(13)
p
X
t=1
Put
Wt′ (C) −
p
X
Vt′ (C) =
p
X
t=1
t=1
Wt (C) −
p
X
Vt (C) = αC .
t=1
Pt := {C : Vt′ (C) 6= 0} and Qt := {C : Wt′ (C) 6= 0}.
Then by (9) it is evident that P1 , . . . , Pp are pairwise disjoint. It is also clear for
S
1 ≤ t ≤ p that Pt ⊆ At , Qt ⊆ Bt and At \ Pt = t−1
j=1 (At ∩ Aj ). Let 1 ≤ t < j ≤ p
and define
# t−1
"
j−1
G
G j
j
Dts ) ∩ (Aj \ Pj ) \
Dk ,
(14)
Dt = (Bt \
s=t+1
k=1
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
using the conventions
′
Sl2
l=l1
9
El = ∅, when 0 ≤ l2 < l1 . Obviously
Dtj ∩ Dtj = ∅
Dtj ∩ Dtj′ = ∅,
and
(t 6= t′ , j 6= j ′ ).
Next for 1 ≤ t ≤ p − 1 we prove that
Q t = Bt \ (
(15)
p
G
Dtj ).
j=t+1
To see this, let C ∈ Qt . Then C ∈ Bt and Wt′ (C) = 1, which means that mt−1
= 0.
C
If m0C = 0, then for each 2 ≤ j ≤ p, C ∈
/ Aj \ Pj . In particular C ∈
/ Dtj , for every
t + 1 ≤ j ≤ p. Assume that mt−1
= k > 0. Then by (10) there are exactly k + 1
C
numbers j0 < j1 < · · · < jk such that C ∈ ∩ki=0 Aji or equivalently
(1 ≤ i ≤ k).
C ∈ Aji \ Pji ,
Pt−1
Pt
This along with the normality condition ( j=1 Wj − j=1 Vj ≥ −1) implies that
there exist natural numbers t1 < t2 < · · · < tk < t such that C ∈ Bti , (1 ≤ i ≤ k),
where t1 is the smallest number with this property. Therefore by (14) and the
F
T
/ pj=t+1 Dtj . The converse is
sentence after that C ∈ ki=1 Dtjii . So, we have C ∈
obtained by a similar argument. We continue the proof of the theorem in the following four steps.
P
Pt
Step 1. By the normality condition if 1 ≤ t ≤ p, then t−1
j=1 Wj −
j=1 Vj ≥ −1.
This means (At \ Pt ) ⊆ ∪t−1
B
.
We
shall
prove
that
for
1
≤
t
≤
p
j
j=1
At \ Pt =
(16)
t−1
G
Dst .
s=1
Firstly we show that
At \ Pt ⊆
t−1
[
s=1
"
(Bs \
j
∪t−1
j=s+1 Ds )
\
s−1
G
Dkt
k=1
#
.
Let C ∈ At \ Pt and pick the smallest natural number t0 < t such that C ∈ Bt0
(it is possible by the normality condition). If for each t0 < j < t and 1 ≤ k < t0 ,
Ft0 −1 t
Ft−1
/ Dkj , we are done. If C ∈ k=1
Dk , then C ∈ k=1 Dkt because
C∈
/ Dtj0 and C ∈
t0 < t. Assume that for some t0 < j < t, C ∈ Dtj0 . Then by (14) C ∈ Aj \ Pj . By the
definition of Pj , this means that there exists at least one natural number s < j such
that C ∈ Aj ∩ As but on the other side C ∈ At . Consequently C ∈ Aj ∩ As ∩ At .
Pt−1
Pt
Since by the normality condition j=1 Wj − j=1 Vj ≥ −1, there are at least two
sets of the form Bi containing C. Accordingly there exists the smallest positive
integer t′0 < t such that C ∈ Bt′0 . This process is obviously finite and therefore we
achieve the purpose. Hence
"
# t−1
s−1
t−1
G
G
[
t−1
j
t
(Bs \ ∪j=s+1 Ds ) \
Dk =
Dst .
At \ Pt = (At \ Pt ) ∩
s=1
k=1
s=1
Step 2. We show by induction that for all j ≥ 1 there exists a family {Aσj , σj ∈
Fj−1
Λj } with cardinality 2j−1 −1 of disjoint subsets (possibly the empty sets) of t=1 P̃t
AKRAM YOUSOFZADEH
10
σj ∈ Λj } of G such that
G
Ãj \ P̃j =
gσj Aσj .
and the members {gσj ,
σj ∈Λj
Since Ã1 \ P̃1 = ∅, it is natural to set Aσ1 = ∅ and gσ1 = e, the identity element of
G. For j = 2, we have Ã2 \ P̃2 = Ã1 ∩ Ã2 ⊆ B̃1 = ġ1 Ã1 . Hence Ã2 \ P̃2 = ġ1 (ġ1−1 D̃12 ).
Set σ2 = 1, Λ2 = {σ2 } and gσ2 = ġ1 . Note that |Λ2 | = 1 = 22−1 − 1. Now let for
1≤t≤j
G
and |Λt | = 2t−1 − 1.
Ãt \ P̃t =
gσt Aσt
σt ∈Λt
By (16)
Ãj+1 \ P̃j+1 =
j
G
D̃tj+1
t=1
=
j
G
ġt (ġt−1 D̃tj+1 )
t=1
=
j
G
t=1
=
j
G
G
ġt [ġt−1 D̃tj+1 ∩ P̃t ] [ġt−1 D̃tj+1 ∩ (Ãt \ P̃t )]
ġt
t=1
=
j
G
[ġt−1 D̃tj+1
G
G
∩ P̃t ] [ġt−1 D̃tj+1 ∩ (
gσt Aσt )]
σt ∈Λt
!
ġt (ġt−1 D̃tj+1 ∩ P̃t )
t=1
Dtj+1
G
j
G
G
!
!
ġt−1 D̃tj+1 ∩ Aσt ] .
ġt gσt [gσ−1
t
t=2 σt ∈Λt
We have used the fact that
⊆ Pt . Recall for the last equation that Aσ1 = ∅.
Pj
Note that the number of pieces in the decomposition above is j + t=2 |Λt |, so by
the induction hypothesis
|Λj+1 | = j +
j
X
(2t−1 − 1) = 2j − 1.
t=2
Hence it is done for j + 1. This completes the Step 2. Before proceeding to the rest
of the proof, we draw the reader’s attention to the next two remarks.
Remark 3.2. In step 2 we explained how the sets Aσj+1 are obtained from the sets
Pt and Aσt , where 1 ≤ t < j + 1. In fact each Aσj+1 is in one of the following two
types:
ġt−1 D̃tj+1 ∩ P̃t
(I)
(II)
ġt−1 D̃tj+1 ∩ Aσt ,
gσ−1
t
for suitable t < j + 1 and σt ∈ Λt . So, one can easily see that Aσj+1 ∩ Aσt 6= ∅ if
and only if Aσj+1 ⊆ Aσt . In this case for every σt′ 6= σt , Aσj+1 ∩ Aσt′ = ∅.
Remark 3.3. For 1 ≤ r < t ≤ p and σt ∈ Λt , either Aσt and P̃r are disjoint or
Aσt ⊆ P̃r . To see this we apply Remark 3.2. Let Aσt ∩ P̃r 6= ∅. If Aσt is of type
(I), then Aσt ⊆ P̃k , for some k < t. In this case P̃r ∩ P̃k 6= ∅, which is impossible
unless k = r. Therefore Aσt ⊆ P̃r . Now suppose that Aσt is of type (II) and l be
the smallest natural number such that for some σl ∈ Λl , Aσt ⊆ Aσl and Aσl is of
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
11
l
type (II). Then Aσl = P̃m ∩ ġ −1 D̃m
⊆ P̃m , for some m < l, which is not possible
unless m = r. This implies that Aσt ⊆ P̃r .
Step 3. We show that for 1 ≤ t ≤ p and C ∈ Qt there exists a family {MC
δt ∈
δt ,
F
Γt } with cardinality 2t−1 of disjoint subsets of tj=1 P̃j and the members {hC
,
δt ∈
δt
Γt } of G such that
G
C
hC
(17)
x0 (C) =
δt Mδt .
δt ∈Γt
Since Qt ⊆ Bt , by Step 1 we have
Q̃t ⊆ B̃t = ġt Ãt
G
= ġt P̃t ġt (Ãt \ P̃t )
G
G
= ġt P̃t ġt (
gσt Aσt )
σt ∈Λt
G G
ġt gσt Aσt ).
= ġt P̃t (
σt ∈Λt
Clearly by Step 1 the number of pieces in this decomposition is 1 + |Λt | = 2t−1 .
Therefore for C ∈ Qt , one has
G G
x0 (C) = ġt (P̃t ∩ ġt−1 x0 (C)) (
ġt−1 x0 (C))).
ġt gσt (Aσt ∩ gσ−1
t
σt ∈Λt
Now set MδC0 = P̃t ∩ ġt−1 x0 (C) and MδCt = Aσt ∩ gσ−1
ġt−1 x0 (C)), where σt ∈ Λt .
t
C
Note that these sets are pairwise disjoint; Mδ0 is a subset of P̃t whereas the sets
Ft−1
Ft−1
MδCt , δt ∈ Λt are disjoint subsets of j=1 P̃j and besides ( j=1 P̃j ) ∩ P̃t = ∅. Define
C
C
Γt = {δ0 } ∪ Λt , hC
δt = ġt gσt , where δt = σt ∈ Λt and hδt0 = ġt . Thus {Mδt , δt ∈ Γt }
is a family of pairwise disjoint subsets of G and
G
C
x0 (C) =
hC
δt Mδt .
δt ∈Γt
′
It is noticeable that for 1 ≤ t 6= s ≤ p, MδCt ∩ MδCs = ∅, where C ∈ Qt and C ′ ∈ Qs .
We show this below. Without loss of generality assume that t < s. Letting C ∈ Qt
and C ′ ∈ Qs , we need to prove that the following four equalities are satisfied
′
MδC0 ∩ MδC0 = ∅,
′
MδCt ∩ MδC0 = ∅,
′
MδC0 ∩ MδCs = ∅,
′
MδCt ∩ MδCs = ∅.
But it is equivalent to show that
(18)
[P̃t ∩ ġt−1 Q̃t ] ∩ [P̃s ∩ ġs−1 Q̃s ] = ∅,
(19)
ġt−1 Q̃t )] = ∅,
[P̃s ∩ ġs−1 Q̃s ] ∩ [Aσt ∩ gσ−1
t
(20)
ġs−1 Q̃s )] = ∅,
[P̃t ∩ ġt−1 Q̃t ] ∩ [Aσs ∩ gσ−1
s
(21)
ġs−1 Q̃s )] = ∅.
ġt−1 Q̃t )] ∩ [Aσs ∩ gσ−1
[Aσt ∩ gσ−1
s
t
AKRAM YOUSOFZADEH
12
The first
two equalities are trivial, since for each r 6= s, Pr ∩ Ps = ∅ and besides,
Ft−1
Aσt ⊆ r=1 P̃r and t < s. We now prove (20). If Aσs is of type (I), then there
exists l < s such that As = ġl−1 D̃ls ∩ P̃l . So
ġs−1 Q̃s )] ⊆ P̃l ∩ P̃t .
[P̃t ∩ ġt−1 Q̃t ] ∩ [Aσs ∩ gσ−1
s
The set P̃l ∩ P̃t is clearly empty if l 6= t. If l = t,
ġs−1 Q̃s ] ⊆ P̃l ∩ ġl−1 Q̃l ∩ ġl−1 D̃ls
[P̃t ∩ ġt−1 Q̃t ] ∩ [Aσs ∩ gσ−1
s
and P̃l ∩ ġl−1 Q̃l ∩ ġl−1 D̃ls is also empty since Q̃l ∩ D̃ls , by (15). Now let Aσs be of
type (II) and k be the smallest natural number such that Aσs ⊆ Aσk and Aσk is
of type (II). This implies that Aσk ⊆ P̃l ∩ ġl−1 D̃lk , for some l < k. If (20) does
not satisfy, then by Remark 3.3, Aσs ⊆ P̃t and this is impossible unless t = l. So
Aσs ⊆ Aσk ⊆ ġt−1 D̃tk ∩ P̃t . Taking into account that Dtk ∩ Q̃t = ∅, we have
[P̃t ∩ ġt−1 Q̃t ] ∩ Aσs ⊆ ġt−1 Q̃t ∩ ġt−1 D̃tk = ∅.
This contradicts our assumption.
To prove (21) let Aσs ∩ Aσt 6= ∅. By Remark 3.2 Aσs ⊆ Aσt . In fact Aσs =
ġt−1 D̃ts . Thus we have
Aσt ∩ gσ−1
t
ġt−1 D̃ts = ∅,
ġt−1 Q̃t ∩ gσ−1
ġs−1 Q̃s ] ⊆ Aσt ∩ gσ−1
ġt−1 Q̃t ] ∩ [Aσs ∩ gσ−1
[Aσt ∩ gσ−1
t
t
s
t
because Q̃t ∩ D̃ts = ∅.
Step 4. P1 , P2 , . . . , Pp are pairwise disjoint and for each C ∈ Con(g, E)
p
X
Vt′ (C) ∈ {0, 1}.
t=1
P
P
For C ∈ Con(g, E) we put zC = αC + pt=1 Vt′ (C). This way, if pt=1 Vt′ (C) = 1,
C
then zc ≥ 2. So, by (13) there are at least two numbers 1 ≤ tC
1 , t2 ≤ p such that
C ∈ QtC1 and C ∈ QtC2 . Therefore by Step 2,
G
G
C
C
hC
hC
x0 (C) =
δ C Mδ C .
δ C Mδ C =
1
t
1
t
1
δtC ∈ΓtC
δtC ∈ΓtC
2
1
t
2
t
2
2
Pp
On the other hand, if t=1 Vt′ (C) = 0, then zc ≥ 1, so by (13) there is at least one
number 1 ≤ tC ≤ p such that C ∈ QtC . Therefore by Step 2,
G
C
hC
x0 (C) =
δtC MδtC .
δtC ∈ΓtC
Thus
G⊇
P
G
(
G
MC
δ C)
t
1
Vi′ (C)=1 δtC ∈ΓtC
1
1
G
G
P
(
G
δtC ∈ΓtC
G
2
2
MC
δ C )
G
t
2
Vi′ (C)=0 δtC ∈ΓtC
MC
δtC
G
P
G
Vi′ (C)=0
x0 (C) .
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
13
Observe that the sets in the right hand side of above inclusion are pairwise disjoint
because
p
G
X
Pi ,
{C ∈ Con(g, E),
Vi′ (C) = 0} ⊆ Con(g, E) \
i=1
i.e.,
P
G
Vi′ (C)=0
x0 (C)
p
G
\
P̃i
i=1
!
=∅
F
and the sets MδCt , δt ∈ Γt are pairwise disjoint subsets of pi=1 P̃i (take into account
C
that tC
1 , and t2 are distinct numbers). Based on our choice of sets
G
G
G
G
G
C
C
G=
(
hC
(
hC
δtC MδtC )
δtC MδtC )
P
G=
P
Vi′ (C)=1 δtC ∈ΓtC
P
1
1
1
G
(
G
Vi′ (C)=1 δtC ∈ΓtC
2
2
G
C
hC
δtC MδtC )
2
2
P
Vi′ (C)=0 δtC ∈ΓtC
G
Vi′ (C)=0
x0 (C) ,
which is a paradoxical decomposition of G after omitting the empty sets from this
decomposition. This decomposition is complete if for each C ∈ Con(g, E),
n
X
Vi′ (C) = aC = 1.
i=1
Otherwise the decomposition is not complete but by a process described in the
proof of [7, Proposition 1.2] it can be changed into a complete one.
Corollary 3.4. Using the notations of the proof of Theorem 3.1, if for every C ∈
Con(g, E), aC = 1, then τ (G) ≤ (ℓ − 1)(2p − 1).
Proof. By the definition of configuration the cardinality of each Bt cannot be more
than ℓ − 1, where |Con(g, E)| = ℓ. The Step 3 in the proof of Theorem 3.1 and the
explanation following Proposition 2.6 yields
τ (G) ≤
≤
p
X
t=1
p
X
|Λt ||Qt |
2t−1 |Bt |
t=1
≤
p
X
2t−1 (ℓ − 1)
t=1
≤ (ℓ − 1)(2p − 1).
Remark 3.5. In [7] the authors constructed the paradoxical decomposition under
the paradoxical condition. That is, a subsystem of Eq(g, E) is equivalent to BX = 0,
AKRAM YOUSOFZADEH
14
Pm
where each row of B has nonnegative entries and is of the form i=1 (Lji i − Lki i ),
for some ji , ki ∈ {0, . . . , n}, where Lji is the coefficient vector of the equation
X
X
fC = 0
fC −
x0 (C)⊆Ei
xj (C)⊆Ei
Pm
ks
js
and in addition B has no zero column. Let Rs = i=1 (Li i − Li i ) be s-th row of
B. Then
m
X
X
X
X
fC ) =
Rs (C)fC ,
fC −
(
i=1 xjs (C)⊆Ei
so
m
X
(
i
X
fC ) =
X
fC ) =
m
X
(
X
fC ) +
i=1 xks (C)⊆Ei
i=1 xjs (C)⊆Ei
i
C
xks (C)⊆Ei
i
i
X
Rs (C)fC
C
therefore we have
m
X
(
But Con(g, E) =
where
i=1
W = (
x0 (C)⊆Ei
1, 1, . . . , 1
1, 1, . . . , 1
..
.
1, 1, . . . , 1
X
(
i=1 x0 (C)⊆Ei
i=1 x0 (C)⊆Ei
Fm F
m
X
fC ) +
X
Rs (C)fC .
C
x0 (C). Consequently we can replace B by W − V,
+
R1
R2
..
.
Rp
) and V =
1, 1, . . . , 1
1, 1, . . . , 1
..
.
1, 1, . . . , 1
.
Since by assumption each Rs has nonnegative
Pentries, Ws − Vs − Vs+1 has entries
≥ −1. Finally B has no zero column, so if s (Ws − Vs ) = (αC )C , then αC > 0.
This shows that the new system satisfies the normality condition in the statement
of Theorem 3.1.
4. Graph interpretation
In the current section we assign a graph to a group which helps us construct the
paradoxical decompositions and compute the Tarski numbers.
Definition 4.1. Let G be a group and g = (g1 , . . . , gn ) and E = {E1 , . . . , Em } be a
finite string of elements of G and a finite partition for G, respectively. Γ = Γ(G, g, E)
is a graph constructed as follows
The vertex set of Γ is identified with ℓ-tuples of nonnegative integers, where
|Con(g, E)| = ℓ:
V (Γ) := (aC )C∈Con(g,E) ; aC ∈ N ∪ {0} .
There exists a directed edge from the vertex A = (aC ) to B = (bC ) only if
aC ∈ {0, 1} and there are disjoint subfamilies
C
AC
1 , . . . , AbC
(C ∈ Con(g, E))
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
of a partition of
F
aC 6=0
15
x0 (C) such that for all C ∈ Con(g, E)
G
x0 (C) =
gA A, 1 ≤ j ≤ bC ,
A∈AC
j
for suitable subsets {gA , A ∈ AC
j }, 1 ≤ j ≤ bC , of G.
Proposition 4.2. If Γ contains adjacent vertices A = (aC )C∈Con(g,E) and B =
(bC )C∈Con(g,E) with aC ∈ {0, 1} and αC := bC −aC > 0, then G admits a paradoxical
decomposition in terms of g and E.
Proof. By the assumption for each C, bC ≥ aC + 1. In other words
(
1, if aC = 0
bC ≥
2, if aC =
6 0.
F
Hence by Definition 4.1 there exists a partition P of aC 6=0 x0 (C) with the following
properties:
C
C
C
• If aC 6= 0, there are AC
1 , A2 ⊆ P and {gA , A ∈ A1 }, {gB , B ∈ A2 } ⊆ G
such that
G
G
gB .B.
gA .A =
x0 (C) =
B∈AC
2
A∈AC
1
• If aC = 0, there are AC ⊆ P and {gD , D ∈ AC } ⊆ G such that
G
gD .D.
x0 (C) =
D∈AC
• All the above subfamilies of P are pairwise disjoint.
F
F
For convenience denote the set aC =0 x0 (C) by E. We know that G = C∈Con(g,E) x0 (C).
Then we have
G=E
G
and
G=
G
G
aC =0 D∈AC
G
G
aC 6=0
A∈AC
1
!
gD .D
G
gA .A
G
G
aC 6=0 B∈AC
2
gB .B
which give a (not necessarily complete) paradoxical decomposition for G. To make
it complete, one can apply the proof of [7, Proposition 1.2] as usual.
Theorem 4.3. Let G be a group and g = (g1 , . . . , gn ) and E = {E1 , . . . , Em } be a
finite string of elements of G and a finite partition for G, respectively. If Eq(g, E)
has no nonnegative nonzero solution with a normal subsystem, then Γ = Γ(G, g, E)
includes vertices A = (aC )C∈Con(g,E) and B = (bC )C∈Con(g,E) with aC ∈ {0, 1} and
αC := bC − aC > 0.
P
P
Proof. Using the notations of the proof of Theorem
A = Vi′ and B = Wi′ ,
P 3.1,
are desired vertices because first, B − A > 0 and
V ′ (C) ∈ {0, 1}, and second, by
(17) they satisfy the second condition of the definition of Γ(G, g, E).
AKRAM YOUSOFZADEH
16
4.1. Diagrams associated with configuration equations. Here we present a
kind of diagram associated with configuration equations. It could be of great importance in order to reduce the complexity of the proofs in the previous section.
• Let A, B1 , . . . , Bt be subsets of Con(g, E). If there is g ∈ G and there is a
F
partition {A1 , . . . , At } for à = C∈A x0 (C) such that
B̃j = gAj ,
1 ≤ j ≤ t,
we use the following tree diagram
A
B1
B2
...
Bt
Clearly if D ⊆ A, and Ej = B̃j ∩ g D̃, we also have a similar diagram with
directed paths from D to E1 , . . . , Et−1 , Et .
• By the diagram
B1
B2
...
Bt
A
F
we mean A = ti=1 Bi .
• Let A, B ⊆ Con(g, E). In the sequel we denote a directed path of any kind
of above two types from A to B by [A, B]. Inductively [A, B1 , B2 , . . . , Bn ]
is a path which connects the sets A, B1 , B2 , . . . , Bn through the sequence
of n single paths [A, B1 ], [B1 , B2 ], . . . , [Bn−1 , Bn ].
In the proof of Theorem 3.1 according to the normality condition we have the
possibility to borrow the configurations in At \ Pt from B1 , . . . , Bt−1 so that the
(not necessarily disjoint) subsets A1 , . . . , Ap turn into disjoint subsets P1 , . . . , Pp
and it is enough for us to construct a paradoxical decomposition of G and to get
an upper bound for τ (G). We Apply the above diagrams to make the proof of
that theorem more clear to the reader. Using the same notations, by (15) we
F
F
F
have P̃1 = Ã1 = ġ1−1 B̃1 = ġ1−1 Q̃1 ġ1−1 D̃12 · · · ġ1−1 D̃1p . So, there are p directed
F
paths from P1 = A1 to D12 , . . . , D1p and Q1 . Also D̃12 P̃2 = Ã2 = ġ2−1 B̃2 =
F
F
F
ġ2−1 Q̃2 ġ2−1 D̃23 · · · ġ2−1 D̃2p . So, there are p − 1 directed paths from D12 to
D23 , . . . , D2p and Q2 and the same number of paths from P2 to D23 , . . . , D2p and
Q2 .
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
P1
Q1
p
D1
...
2
D1
A2
2
D1
Q2
p
D2
···
3
D2
17
P2
A2
Gluing these diagrams we obtain two paths from D12 and P2 to Q2 and so on.
P1
Q1
p
D1 ...
3
D1
P2
2
D1
A2
Q2
D2
P3
p
···
3
D2
A3
..
.
Ap
Qp
Pp
AKRAM YOUSOFZADEH
18
Our diagram contains the following 2p − 1 paths
[P1 , Q1 ],
[P1 , D12 , A2 , Q2 ], [P2 , A2 , Q2 ],
[P1 , D13 , A3 , Q3 ], [P1 , D12 , A2 , D23 , A3 , Q3 ], [P2 , A2 , D23 , A3 , Q3 ], [P3 , A3 , Q3 ],
..
.
[P1 , D1p , Ap , Qp ], . . . , [P1 , D12 , A2 , D2p , Ap , Qp ], . . . , [Pp , Ap , Qp ].
Depending onFthe members of Q1 , . . . , Qp , these paths can be continued one more
p
step. In fact t=1 Pt admits a partitionFwith at most 2p − 1 pieces such that with
suitable coefficients construct the sets C∈Qt x0 (C), 1 ≤ t ≤ p, separately (see
proof of Theorem 3.1). 2t−1 pieces in the partition are used to make Q̃t . These
pieces are divided into 2t−1 |Qt | new pieces to construct all the sets x0 (C), C ∈ Qt .
To have a paradoxical P
decomposition each x0 (C),
Pp C ∈ Con(g, E) must be obtained
p
at least two times. If t=1 Vt′ (C) = 1, then t=1 Wt′ (C) ≥ 2. In other words, C
belongs to at least two Qt s and it is done. Otherwise, C may appear in only one
Pp
′
Qt and we have to add the single set {C,
= 0˜} to the pieces of our
t=1
PpVt (C)
t−1
paradoxical
decomposition. Therefore τ (G) ≤ t=1 2 |Qt |+1. It is to be noted
Pp
that t=1 2t−1 |Qt | is the maximum possible number of mentioned paths.
Naturally to make this observation more accurate, we have to choose a procedure
in which the number of pieces in the paradoxical decomposition is as small as
possible. This process depends directly on the set of configurations given in the
statement of problem. We explain it in the next subsection.
4.2. Examples. We end the paper with some examples. Our examples will be
based on the sets of configurations instead of the elements of the group. As the
paradoxical decomposition of a group is not unique, different diagrams also exist.
One can find a more precise Tarski number’s upper bound by counting the paths
from the top to the bottom in a minimal diagram (a diagram with minimum number
of paths). It is noticeable that a minimal diagram is not necessarily a connected
one.
Example 4.4. Let C1 = (1, 2, 3, 2), C2 = (1, 3, 1, 3), C3 = (2, 1, 2, 2), C4 =
(3, 3, 1, 2) and C5 = (3, 3, 2, 1) be configurations corresponding to the configuration
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
19
pair (g, E) of a group G. Consider the following equations
X
X
fC =
fC
x0 (C)⊆E1
X
fC =
fC ,
x2 (C)⊆E2
X
fC =
fC
x2 (C)⊆E2
x1 (C)⊆E2
X
fC ,
x0 (C)⊆E1
x0 (C)⊆E2
X
X
fC =
x3 (C)⊆E1
X
fC
x0 (C)⊆E1
x3 (C)⊆E1
X
X
fC =
x1 (C)⊆E1
X
fC ,
x0 (C)⊆E1
x1 (C)⊆E1
X
X
fC =
fC =
x3 (C)⊆E3
X
fC
x0 (C)⊆E3
which are the next equations, respectively
fC3 =fC1 + fC2 ,
fC3 =fC1 + fC2 ,
fC5 =fC3 ,
fC5 =fC3 ,
fC3 =fC3 + fC5 ,
fC1 =fC3 + fC5 ,
fC2 =fC4 + fC5 .
0 0
0 0
0 0
Then A =
0 0
0 0
1 0
0 1
sponding coefficient
1 0 0
1 0 0
0 0 1
0 0 1
1 0 0
0 0 0
0 0 0
matrices.
1 1 0 0 0
1 1 0 0 0
0 0 1 0 0
and B = 0 0 1 0 0
0 0 1 0 1
0 0 1 0 1
0 0 0 1 1
Setting π = (2 7)(4 6), one has
1 0 −1 0 0
1 0 −1 1 0
0 0 0 1 0
T Pπ (B − A) − Pπ+ A =
0 0 0 1 1 .
0 0 1 1 1
0 0 1 1 1
1 1 0 1 1
are the corre
AKRAM YOUSOFZADEH
20
Therefore this system is normal. The equations corresponding to T Pπ (B − A)
are
fC3 =fC1 + fC2 ,
fC2 =fC4 + fC5 ,
fC5 =fC3 ,
fC1 =fC3 + fC5 ,
fC3 =fC3 + fC5 ,
fC5 =fC3 ,
fC3 =fC1 + fC2 .
We use the notations of the proof of Theorem 3.1. In this example p = 7, A1 =
{C3 }, A2 = {C2 }, A3 = {C5 }, A4 = {C1 }, A5 = {C3 }, A6 = {C5 }, A7 = {C3 },
B1 = {C1 , C2 }, B2 = {C4 , C5 }, B3 = {C3 }, B4 = {C3 , C5 }, B5 = {C3 , C5 }, B6 =
{C3 }, B7 = {C1 , C2 }. The initial diagram is
C3
C2
C1
C5
C2
C1
C3
C4
C5
C3
C3
C5
C5
C3
C5
C3
C3
C1
C2
Applying the process in the proof of Theorem, we have P1 = {C3 }, P2 = {C2 }, P3 =
{C5 }, P4 = {C1 }, P5 = ∅, P6 = ∅, P7 = ∅, Q1 = {C1 , C2 }, Q2 = {C4 }, Q3 = ∅, Q4 =
{C3 , C5 }, Q5 = {C5 }, Q6 = {C3 }, Q7 = {C1 , C2 }. The associated diagram is
C3
C1
C2
C2
C4
C5
C1
C5
C3
C3
C3
C3
C5
C1
C2
C5
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
21
The condition of normality enables us to have disjoint subsets P1 , . . . , Pp which
produce Con(g, E) two times (see proof of Theorem 3.1). By Corollary 3.4, τ (G) ≤
(ℓ − 1)(2p − 1) = 508 (ℓ = 5 and p = 7). This is not a very accurate upper bound.
Changing the diagram to a minimal one, helps us compute the most precise bound
τ (G). Look at the following diagram
C3
C2
C5
C1
C4
C3
C3
C5
C3 , C5
C3
C1 , C2
F
F
According to (5) we haveFg1 x0 (C3 ) = x0 (C1 ) x0 (C2 ), g3 F
x0 (C2 ) = x0 (C4 ) x0 (C5 ),
g2−1 g1 x0 (C1 ) = x0 (C3 ) x0 (C5 ), g2−1 x0 (C3 ) = x0 (C3 ) x0 (C5 ) and g3 x0 (C5 ) =
x0 (C3 ).
F
F
F
Then x0 (C3 ) = ∆1 ∆2 ∆3 ∆4 , where
∆1 = g1−1 g3−2 x0 (C3 )
∆2 = g1−1 g3−1 x0 (C4 )
[
∆3 = g1−2 g22 x0 (C3 ) x0 (C5 )
[
∆4 = g1−2 g2 g3−1 g1−1 x0 (C1 ) x0 (C2 ) .
Therefore it is obvious that
and besides
G
(G \ x0 (C3 ))
G
= g3 g1 ∆2 (G \ x0 (C3 ))
G = x0 (C3 )
G = g32 g1 ∆1
G
g2−2 g12 ∆3
G
g1 g3 g2−1 g12 ∆4 .
In other words, we have a paradoxical decomposition consisting of 5 pieces. Thus
τ (G) ≤ 5. This number can be obtained from the above diagram by counting the
paths from the beginning points to the end points and adding the result by 1. The
AKRAM YOUSOFZADEH
22
suitable part of Γ(G, g, E) corresponding to this example is
(0, 0, 1, 0, 0)
(1, 1, 2, 1, 1)
We have a very special case of Theorem 3.1 which eventuates a result for computing the Tarski number. We use the notations of section 3.
Pp
Theorem 4.5. Suppose that i=1 (Wi − Vi ) = (1, 1, . . . , 1) and there exists a permutation matrix Pπ such that the first p − 1 rows of Pπ B − Pπ+ A have nonnegative
entries. Then τ (G) ≤ 1 + |Aπ(1) | + ℓ.
Proof. The assumption says that for 1 ≤ i ≤ n, Aπ(i+1) ⊆ Bπ(i) . Set Aπ(n+1) = ∅,
−1
−1
−1
hi = ġπ(1)
.ġπ(2)
. . . . .ġπ(i)
and Ri = hi (B̃π(i) \ Ãπ(i+1) ), for 1 ≤ i ≤ p. By induction
on p we see
Ãπ(1) =
p
G
Ri .
i=1
Since by assumption
F
Pp
i=1 (Wi
− Vi ) = (1, 1, . . . , 1), the equations
C∈(Bπ(i−1) \Aπ(i) )
x0 (C) = h−1
i Ri ,
(1 ≤ i ≤ p)
define a paradoxical decomposition. On the other hand we have
n
X
|Bπ(i−1) \ Aπ(i) |) + 1
τ (G) ≤|Bπ(p) | + (
i=2
n
X
|Bπ(i) − Aπ(i) |) + |Aπ(1) | + 1
=(
i=1
=ℓ + |Aπ(1) | + 1.
In the special case where G does not contain the free group on two generators
and ℓ = 3 and |Aπ (1)| = 1 we have τ (G) = 5.
Example 4.6. Let Con(g, E) = {(1, 2, 2, 2), (2, 1, 2, 1), (2, 2, 1, 1)}. Then τ (G) ≤ 5.
A minimal diagram associated with Eq(g, E) is
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
23
C1
C2
C3
C1
C2
C3
C1
C1
C2
C3
C1
The number of paths from the top to the bottom is 4. These paths are [C1 , C3 ],[C1 , C2 , C1 , C2 , C1 ],
[C1 , C2 , C1 , C3 , C1 , C2 ] and [C1 , C2 , C1 , C3 , C1 , C3 , C1 ]. Therefore τ (G) ≤ 4 + 1 = 5.
Indeed according to (5) we have
x0 (C1 ) =g3 x0 (C2 ) ∪ g3 x0 (C3 )
x0 (C2 ) =g1−1 x0 (C1 )
x0 (C3 ) =g2 x0 (C1 ).
Define
∆1 =g3 g1−1 g3 g1−1 g3 x0 (C2 ),
∆2 =g3 g1−1 g3 g1−1 g3 g2 x0 (C1 , )
∆3 =g3 g1−1 g3 g2 x0 (C1 ),
∆4 =g3 x0 (C3 ),
∆5 =x0 (C2 ) ∪ x0 (C3 ).
AKRAM YOUSOFZADEH
24
Then x0 (C1 ) =
F4
∆i . So, G =
F5
∆i and finally we have
G
G =g2−1 g3−1 g1 g3−1 ∆3 g3−1 g1 g3−1 g1 g3−1 ∆1 g3−1 ∆4
G
G =g2−1 g3−1 g1 g3−1 g1 g3−1 ∆2
e∆5 ,
i=1
i=1
G
where e is the identity element of G. We have a complete paradoxical decomposition
with five pieces. So, τ (G) ≤ 5.
Example 4.7. [8] Let F2 = ha, bi be the free group with generators a and b and
identity element e. Suppose that g = (a, b) and E = {E1 , E2 , E3 }, where
E1 = {x, x is a reduced word starting with a},
E2 = {x, x is a reduced word starting with b} and
F
E3 = {x, x is a reduced word starting with a−1 or b−1 } {e}.
Then Con(g, E) = {C1 , . . . , C7 }, where C1 = (1, 1, 2), C2 = (2, 1, 2), C3 = (3, 1, 2), C4 =
(3, 2, 2), C5 = (3, 3, 2), C6 = (3, 1, 3) and C7 = (3, 1, 1). One can see that this example satisfies that condition of the pervious theorem. Consider the following
configuration equations
X
X
X
X
fC
fC =
fC and
fC =
x0 (C)⊆E2
x1 (C)⊆E1
x0 (C)⊆E1
x2 (C)⊆E2
i.e.
fC1 = fC1 + fC2 + fC3 + fC6 + fC7 and fC2 = fC1 + fC2 + fC3 + fC4 + fC5 ,
which imply that each fC is zero. The corresponding diagram is
C1
C7
C6
C3
C2
C1
C1
C2
C3
C4
C5
Put A = a−1 x0 (C1 ), B = b−1 a−1 [∪5i=1 x0 (Ci )], C = a−1 (x0 (C6 ) ∪ x0 (C7 )) and
M = F2 \ x0 (C1 ). Then we have the non-complete paradoxical decomposition
F2 = A ⊔ B ⊔ C ⊔ M ⊔ a−1 x0 (C3 ) = eM ⊔ aA = abB ⊔ aC.
Hence τ (F2 ) = 4.
We now summarize what we have done for estimating the upper bound of the
Tarski number of a group.
(1) Change the order of the configuration equations based on the permutation
matrix in the definition of the normality condition.
A CONSTRUCTIVE WAY TO COMPUTE THE TARSKI NUMBER OF A GROUP
25
(2) Define the initial diagrams consisting of p connected components.
(3) Make each set in the top of the above diagrams disjoint from the previous
ones by borrowing their intersection from the sets in the bottom of the
former diagrams.
(4) Glue these diagrams as it is explained in subsection 4.1.
(5) Find the configurations of the top sets of the initial diagrams in the bottom
of the last gluing diagram.
(6) Count all the directed pats from top to the bottom. This number plus one
is an upper bound for τ (G).
To get the best bound for τ (G) using this procedure, we borrow the maximum
number of configurations that we can get from the former bottom sets (not only
the intersections). It is applied in the examples.
Definition 4.8. Let m, n ∈ N. A set of (n+1)-tuples C = {(ci0 , . . . , cin ), 1 ≤ i ≤ ℓ}
with 1 ≤ cij ≤ m and
∪nj=1 {cij } = {1, . . . , m},
(i = 1, . . . , ℓ)
is called a pre-configuration set if there exist a group G, a string g of elements of
G and a partition E of G such that Con(g, E) = C.
In [5] the authors give examples of groups with Tasrki numbers 5 and 6. Now
the question is whether we can construct such groups using configurations. Responding to this question depends on knowing that given well-behaved sets are
pre-configuration ones. In particular
Question 4.9. Is C = {(1, 2, 3, 2), (1, 3, 1, 3), (2, 1, 2, 2), (3, 3, 1, 2), (3, 3, 2, 1)} a preconfiguration set (see Example 4.4)?
References
[1] A. Abdollahi, A. Rejali and G. A. Willis, Group properties characterized by configuration,
Illinois J. Mathematics, 48 (2004) No. 3, 861–873.
[2] A. Abdollahi, A. Rejali and A. Yousofzadeh, Configuration of nilpotent groups and isomorphism, J. Algebra Appl, 8, No. 3 (2009) 339–350.
[3] T. G. Ceccherini-Silberstein, R. I. Grigorchuk and P. de la Harpe, Amenability and paradoxical decompositions for pseudogroups and for discrete metric spaces, Proc. Steklov. Inst.
Math. 224 (1999), 57–97.
[4] G. B. Dantzig and M. N. Thapa, Linear Programming 2: Theory and Extensions. SpringerVerlag, New York, 2003.
[5] M. Ershov, G. Golan and M. Sapir, The Tarski numbers of groups, Adv. Math. 284 (2015)
21–53.
[6] A. l. T. Paterson, Amenability Mathematical Surveyes and Monographs., 29, American Mathematical Society, Providence, RI, 1988.
[7] A. Rejali and A. Yousofzadeh, Configuration of groups and paradoxical decompositions, Bull.
Belg. Math. Soc. Simon Stevin 18 (2011) 157–172.
[8] J. M. Rosenblatt and G. A. Willis, Weak convergence is not strong for amenable groups,
Canad. math. Bull, 44 (2001) No 2, 231–241.
[9] M. Sapir, Combinatorial algebra: Syntax and Semantics. Springer International Publishing,
Suitzerland, 2014.
[10] S. Wagon, The Banach-Tarski Paradox., vol. 24 of Encyclopedia of Mathematics and its
Applications, Cambridge University Press, Cambridge-New York, 1985.
[11] A. Yousofzadeh, A. Tavakoli and A. Rejali, On configuration graph and paradoxical decomposition,J. Algebra. Appl, 13, No. 2 (2014), 1350086 (11 pages).
E-mail address: [email protected]
| 4math.GR
|
1
Random Coding Error Exponents for the Two-User
Interference Channel
arXiv:1503.02389v2 [cs.IT] 20 Jun 2017
Wasim Huleihel and Neri Merhav
The Andrew & Erna Viterbi Faculty of Electrical Engineering
Technion - Israel Institute of Technology
Haifa 3200003, ISRAEL
E-mail: {wh@campus, merhav@ee}.technion.ac.il
Abstract—This paper is about deriving lower bounds on the
error exponents for the two-user interference channel under
the random coding regime for several ensembles. Specifically,
we first analyze the standard random coding ensemble, where
the codebooks are comprised of independently and identically
distributed (i.i.d.) codewords. For this ensemble, we focus on
optimum decoding, which is in contrast to other, suboptimal
decoding rules that have been used in the literature (e.g., joint
typicality decoding, treating interference as noise, etc.). The
fact that the interfering signal is a codeword, rather than an
i.i.d. noise process, complicates the application of conventional
techniques of performance analysis of the optimum decoder.
Also, unfortunately, these conventional techniques result in loose
bounds. Using analytical tools rooted in statistical physics, as well
as advanced union bounds, we derive single-letter formulas for
the random coding error exponents. We compare our results with
the best known lower bound on the error exponent, and show
that our exponents can be strictly better. Then, in the second part
of this paper, we consider more complicated coding ensembles,
and find a lower bound on the error exponent associated with
the celebrated Han-Kobayashi (HK) random coding ensemble,
which is based on superposition coding.
Index Terms—Random coding, error exponent, interference
channels, superposition coding, Han-Kobayashi scheme, statistical physics, optimal decoding, multiuser communication.
I. I NTRODUCTION
T
HE two-user interference channel (IFC) models a general
scenario of communication between two transmitters and
two receivers (with no cooperation at either side), where each
receiver decodes its intended message from an observed signal,
which is interfered by the other user, and corrupted by channel
noise. The information-theoretic analysis of this model has
begun over more than four decades ago and has recently
witnessed a resurgence of interest. Most of the previous
work on multiuser communication, and specifically on the
IFC, has focused on obtaining inner and outer bounds to
the capacity region (see, for example, [1, Ch. II.7]). In a
nutshell, the study of this kind of channel started in [2] and
continued in [3], where simple inner and outer bounds to
the capacity region were given. Then, in [4], by using the
well-known superposition coding technique, the inner bound
of [3] was strictly improved. In [5], various inner and outer
bounds were obtained by transforming the IFC model into
∗ This research was partially supported by The Israeli Science Foundation
(ISF), grant no. 412/12. This paper was presented in part at the 2016
International Zurich Seminar on Communications.
some multiple-access or broadcast channel. Unfortunately, the
capacity region for the general interference channel is still
unknown, although it has been solved for some special cases
[6, 7]. The best known inner bound is the Han-Kobayashi (HK)
region, established in [8], and which will also be considered
in this paper.
To our knowledge, [9, 10] are the only previous works
which treat the error exponents for the IFC under optimal
decoding. Specifically, [9] derives lower bounds on error
exponents of random codebooks comprised of i.i.d. codewords
uniformly distributed over a given type class, under maximum
likelihood (ML) decoding at each user, that is, optimal decoding. Contrary to the error exponent analysis of other multiuser
communication systems, such as the multiple access channel
[11], the difficulty in analyzing the error probability of the
optimal decoder for the IFC is due to statistical dependencies
induced by the interfering signal. Indeed, for the IFC, the
marginal channel determining each receiver’s ML decoding
rule is induced also by the codebook of the interfering user.
This extremely complicates the analysis, mostly because the
interfering signal is a codeword and not an i.i.d. process.
Another important observation, which was noticed in [9], is
that the usual bounding techniques (e.g., Gallager’s bounding
technique) on the error probability fail to give tight results.
To alleviate this problem, the authors of [9] combined some
of the ideas from Gallager’s bounding technique [12] to get
an upper bound on the average probability of decoding error
under ML decoding, the method of types [13], and used the
method of type class enumerators, in the spirit of [14], which
allows to avoid the use of Jensen’s inequality in some steps.
The main purpose of this paper is to extend the study of
achievability schemes to the more refined analysis of error
exponents achieved by the two users, similarly as in [9].
Specifically, we derive single-letter expressions for the error
exponents associated with the average error probability, for
the finite-alphabet two-user IFC, under several random coding
ensembles. The main contributions of this paper are as follows:
• Similarly as in recent works (see, e.g., [11, 15-18] and
references therein) on the analysis of error exponents, we
derive single-letter lower bounds for the random coding error
exponents. For the standard random coding ensemble, considered in Subsection III-B, we analyze the optimal decoder
for each receiver, which is interested solely in its intended
message. This is in contrast to usual decoding techniques
2
analyzed for the IFC, in which each receiver decodes, in
addition to its intended message, also part of (or all) the
interfering codeword (that is, the other user’s message), or
other conventional achievability arguments [1, Ch. II.7], which
are based on joint-typicality decoding, with restrictions on the
decoder (such as, “treat interference as noise” or to “decode
the interference”). This enables us to understand whether there
is any significant degradation in performance due to the suboptimality of the decoder. Also, since [9] analyzed the optimal
decoder as well, we compare our formulas with those of [9],
and show that our error exponent can be strictly better, which
implies that the bounding technique in [9] is not tight. It is
worthwhile to mention that the analytical formulas of our error
exponents are simpler than the lower bound of [9].
• As was mentioned earlier, in [9] only random codebooks
comprised of i.i.d. codewords (uniformly distributed over a
type class) were considered. These ensembles are much simpler than the superposition codebooks of [8]. Unfortunately,
it is very tedious to analyze superposition codebooks using
the methods of [9]. In this paper, however, the new tools
that we have derived enable us to analyze more involved
random coding ensembles. Indeed, we can consider the coding
ensemble used in the HK achievability scheme [8] and derive
the respective error exponents. We also discuss an ensemble
of hierarchical/tree codes [19].
• The analysis of the error exponents, carried out in this paper,
turns out to be much more difficult than in previous works
on point-to-point and multiuser communication problems, see,
e.g., [11, 15-18]. Specifically, we encounter two main difficulties in our analysis: First, typically, when analyzing the
probability of error, the first step is to apply the union bound.
Usually, for point-to-point systems, under the random coding
regime, the average error probability can be written as a
union of pairwise independent error events. Accordingly, in
this case, it is well known that the truncated union bound is
exponentially tight [20, Lemma A.2]. This is no longer the
case, however, when considering multiuser systems, and in
particular, the IFC. For the IFC, the events comprising the
union are strongly dependent, especially due to the fact that we
are considering the optimal decoder. To alleviate this difficulty,
following the ideas of [11], we derived new upper bounds on
the probability of a union of events, which take into account
the dependencies among the events. The second difficulty that
we have encountered in our analysis is that in contrast to
previous works, applying the type class enumerator method
[14] is not simple, due to the reason mentioned above. Using
some methods from large deviations theory, we were able to
tackle this difficulty.
• Recently, in [21, 22], the authors independently suggested
lower bounds on the error exponents of both standard and cognitive multiple-access channels (MACs), assuming suboptimal
successive decoding scheme, and using the standard random
coding ensemble (considered in Subsection III-B). Although
the motivation in [21] is different, the codebook construction
and the decoding rule are the same as in the first part of
this paper, and thus, essentially, their results apply also for
the IFC. It is important to emphasize that while we believe
that our error exponent analysis is somewhat simpler, at least
conceptually, there is strong resemblance between our analysis
and [21], as they both based on type enumeration techniques.
Note, however, that while in [21] the standard union bound
was used, here, the new upper bounds mentioned above,
provide some potential gain over [21], even for the ordinary
ensemble. Also, as was mentioned above, we consider also
the more complicated ensemble pertaining to the HK scheme.
The derivation of the lower bound on the error exponent of this
ensemble is built upon the derivation of the lower bound on the
error exponent of the standard random coding ensemble, and
thus it makes useful and convenient to start with the analysis
of the latter ensemble. We emphasize that the extension of [21]
to the HK ensemble is non-trivial. Finally, we mention that the
focus in [21] was on achievable rate region, rather than error
exponents, and thus no comparison to [9] was provided.
• We believe that by using the techniques and tools derived
in this paper, other multiuser systems, such as the IFC with
mismatched decoding, the MAC [11], the broadcast channel,
the relay channel, etc., and accordingly, other coding schemes,
such as binning [15], and hierarchical codes [19], can be
analyzed.
The paper is organized as follows. In Section II, we establish
notation conventions. In Section III, we formalize the problem
and assert the main theorems. Specifically, in Subsections
III-B and III-C, we give the resulting error exponents under
the standard random coding ensemble and the HK coding
ensemble, respectively. Finally, Section IV is devoted to the
proofs of our main results.
II. N OTATION C ONVENTIONS
Throughout this paper, scalar random variables (RVs) will
be denoted by capital letters, their sample values will be
denoted by the respective lower case letters, and their alphabets
will be denoted by the respective calligraphic letters, e.g. X, x,
and X , respectively. A similar convention will apply to random
vectors of dimension n and their sample values, which will be
denoted with the same symbols in the boldface font. We also
use the notation Xij (j > i) to designate the sequence of RVs
(Xi , Xi+1 , . . . , Xj ). The set of all n-vectors with components
taking values in a certain finite alphabet, will be denoted by the
same alphabet superscripted by n, e.g., X n . Generic channels
will be usually denoted by the letters P , Q, or W . We shall
mainly consider joint distributions of two RVs (X, Y ) over
the Cartesian product of two finite alphabets X and Y. For
brevity, we will denote any joint distribution, e.g. QXY , simply
by Q, the marginals will be denoted by QX and QY , and
the conditional distributions will be denoted by QX|Y and
QY |X . The joint distribution induced by QX and QY |X will
be denoted by QX × QY |X , and a similar notation will be
used when the roles of X and Y are switched.
The expectation operator will be denoted by E {·}, and when
we wish to make the dependence on the underlying distribution
Q clear, we denote it by EQ {·}. Information measures induced
by the generic joint distribution QXY , will be subscripted
by Q, for example, IQ (X; Y ) will denote the corresponding
mutual information, etc. The divergence (or, Kullback-Liebler
distance) between two probability measures Q and P will be
3
denoted by D(Q||P ). The weighted divergence between two
channels, QY |X and PY |X , with weight PX , is defined as
X
D(QY |X ||PY |X |PX ) ,
PX (x)
We assume that the message pair (m1 , m2 ) is uniformly
distributed over M1 ×M2 . It is clear that the optimal decoder
of the first user, for this problem, is given by
m̂1 = arg max P (y 1 |x1,i )
x∈X
X
y∈Y
i∈M1
QY |X (y|x)
QY |X (y|x) log
. (1)
PY |X (y|x)
For a given vector x, let Q̂x denote the empirical distribution,
that is, the vector {Q̂x (x), x ∈ X }, where Q̂x (x) is the
relative frequency of the letter x in the vector x. Let T (PX )
denote the type class associated with PX , that is, the set of
all sequences x for which Q̂x = PX . Similarly, for a pair of
vectors (x, y), the empirical joint distribution will be denoted
by Q̂xy , or simply by Q̂, for short. All previously defined
notation rules for regular distributions will also be used for
empirical distributions.
The cardinality of a finite set A will be denoted by |A|,
its complement will be denoted by Ac . The probability of an
event E will be denoted by Pr {E}. The indicator function of
an event E will be denoted by I {E}. For two sequences of
·
positive numbers, {an } and {bn }, the notation an = bn means
that {an } and {bn } are of the same exponential order, i.e.,
n−1 log an /bn → 0 as n → ∞, where logarithms are defined
with respect to (w.r.t.) the natural basis, that is, log (·) = ln (·).
Finally, for a real number x, we denote [x]+ , max {0, x}.
III. P ROBLEM F ORMULATION
AND
M AIN R ESULTS
We divide this section into three subsections. In the first, we
present the model and formulate the problem. In the second,
we present a lower bound on the IFC error exponent, assuming
a simple random coding ensemble where random codebooks
comprised of i.i.d. codewords are uniformly distributed over a
type class. It is well-known [10] that this coding scheme can
be improved by using superposition coding and introducing the
notion of “private” and “common” messages (to be defined in
the sequel). Accordingly, in the third subsection, we consider
the HK coding scheme [8], and derive lower bounds on the
error exponents. Finally, we discuss other ensembles that can
be analyzed using the same methods.
A. The IFC Model
Consider a two-user interference channel of two senders,
two receivers, and a discrete memoryless channel (DMC),
defined by a set of single-letter transition probabilities, WY1 Y2 |X1 X2 (y1 , y2 |x1 , x2 ), with finite input alphabets,
X1 , X2 , and finite output alphabets, Y1 , Y2 . Here, each sender,
k ∈ {1, 2}, communicates an independent message mk ∈
{1, 2, . . . , Mk , 2nRk } at rate Rk , and each receiver,
l ∈ {1, 2}, decodes its respective message. Specifically, a
(2nR1 , 2nR2 , n) code Cn consists of:
sets M1 , 0, . . . , 2nR1 − 1 and M2 ,
• Two message
0, . . . , 2nR2 − 1 for the first and second users, respectively.
• Two encoders, where for each k ∈ {1, 2}, the k-th encoder
assigns a codeword xk,i to each message i ∈ Mk .
• Two decoders, where each decoder l ∈ {1, 2} assigns an
estimate m̂l to ml .
= arg max
i∈M1
M2 −1
1 X
P (y 1 |x1,i , x2,j )
M2 j=1
(2)
(3)
where P (y 1 |x1,i , x2,j ) is the marginal channel defined as
P (y 1 |x1,i , x2,j ) ,
n
Y
WY1 |X1 X2 (y1k |x1,i,k , x2,j,k ),
(4)
k=1
and
WY1 |X1 X2 (y1,k |x1,i,k , x2,j,k )
X
,
WY1 Y2 |X1 X2 (y1,k , y2,k |x1,i,k , x2,j,k ).
(5)
y2,k ∈Y2
The optimal decoder of the second user is defined similarly.
Since there is no cooperation between the two receivers, the
error probabilities for the code Cn , are defined as
Pe,i (Cn ) , 2−n(R1 +R2 )
X
·
Pr {m̂i (Yin ) 6= m̃i |m1 = m̃1 , m2 = m̃2 } ,
m̃1 ,m̃2
(6)
for i = 1, 2.
B. The Ordinary Random Coding Ensemble
In this subsection, we consider the ordinary random coding
ensemble: For each k ∈ {1, 2}, we select independently Mk
codewords {xk,i }, for i ∈ Mk , under the uniform distribution
across the type class T (PXk ), for a given distribution PXk
(n)
on Xk . Our goal is to assess the exponential rate of P̄e,1 ,
E {Pe,1 (Cn )}, where the average is over the code ensemble,
that is,
1
(n)
(7)
E1∗ (R1 , R2 ) , lim inf − log P̄e,1 ,
n→∞
n
and similarly for the second user. Before stating the main
result, we define some quantities. Given a joint distribution
QX1 X2 Y1 over X1 × X2 × Y1 , consider the definitions in (8),
shown at the top of the next page. We devote Appendix C for
a discussion on aspects of the computation of (8j). We have
the following result.
Theorem 1 Let R1 and R2 be given, and let E ∗ (R1 , R2 ) be
defined as in (7). Consider the ensemble of fixed composition
codes of types PX1 and PX2 , for the first and second users,
respectively. For a discrete memoryless two-user IFC, we have
E1∗ (R1 , R2 ) ≥ Ẽ1 (R1 , R2 ),
(9)
for any R1 , R2 ≥ 0.
Several remarks on Theorem 1 are in order.
Due to symmetry, the error exponent for the second user,
that is, Ẽ2 (R1 , R2 ) is simply obtained from Theorem 1 by
•
4
f (QX1 X2 Y1 ) , EQ log WY1 |X1 X2 (Y1 |X1 , X2 ) ,
h
i
f (Q̂) − IQ̂ (X2 ; X1 , Y1 ) ,
max
t0 (QX1 Y1 ) , R2 +
(8a)
(8b)
Q̂: Q̂X2 =PX2 , Q̂X1 Y1 =QX1 Y1
IQ̂ (X2 ;X1 ,Y1 )≤R2
E1 (Q̃X1 X2 Y1 , QX1 X2 Y1 ) ,
E2 (Q̃X1 X2 Y1 , QX1 X2 Y1 ) ,
min
Q̂: Q̂X2 =PX2 , Q̂X1 Y1 =Q̃X1 Y1
Q̂∈L(Q̃X1 X2 Y1 ,QX1 X2 Y1 )
min
Q̂: Q̂X2 =PX2 , Q̂X1 Y1 =Q̃X1 Y1
Q̂∈L̂(Q̃X1 X2 Y1 ,QX1 X2 Y1 )
i
h
IQ̂ (X2 ; X1 , Y1 ) − R2 ,
(8c)
i
h
IQ̂ (X2 ; Y1 ) − R2 ,
(8d)
+
+
L(Q̃X1 X2 Y1 , QX1 X2 Y1 ) , Q̂ : max [t0 (QX1 X2 Y1 ), f (QX1 X2 Y1 )]
h
i
,
≤ max f (Q̃X1 X2 Y1 ), f (Q̂) + R2 − IQ̂ (X2 ; X1 , Y1 )
+
L̂(Q̃X1 X2 Y1 , QX1 X2 Y1 ) , Q̂ : max [t0 (QX1 X2 Y1 ), f (QX1 X2 Y1 )]
h
i
,
≤ max f (Q̃X1 X2 Y1 ), f (Q̂) + R2 − IQ̂ (X2 ; Y1 )
+
i
h
IQ̃ (X1 ; X2 , Y1 ) + E1 (Q̃X1 X2 Y1 , QX1 X2 Y1 ) ,
min
Ê1 (QX1 X2 Y1 , R2 ) ,
(8e)
(8f)
(8g)
Q̃: Q̃X1 =PX1 , Q̃X2 Y1 =QX2 Y1
Ê2 (QX1 X2 Y1 , R2 ) ,
min
Q̃: Q̃X1 =PX1 , Q̃X2 Y1 =QX2 Y1
E(QX1 X2 Y1 , R1 , R2 ) , max
Ẽ1 (R1 , R2 ) ,
min
E2 (Q̃X1 X2 Y1 , QX1 X2 Y1 ),
h
i
Ê1 (QX1 X2 Y1 , R2 ) − R1 , Ê2 (QX1 X2 Y1 , R2 ) ,
+
D(QY1 |X1 X2 ||WY1 |X1 X2 |PX1 × PX2 ) + E(QX1 X2 Y1 , R1 , R2 ) .
QY1 |X1 X2 :
QX1 =PX1 ,QX2 =PX2
swapping the roles of X1 , Y1 , and R1 , with those of X2 , Y2 ,
and R2 , respectively.
• An immediate byproduct of Theorem 1 is finding the set of
rates (R1 , R2 ) for which Ẽ1 (R1 , R2 ) > 0, namely, the rates
for which the probability of error vanishes exponentially as
n → ∞. We show in Appendix D, that this set is given by:
(8i)
(8j)
R2
I(X2 ; Y1 |X1 )
Rordinary,1
Rordinary,1 = {R1 < I (X1 ; Y1 )} ∪
{{R1 + R2 < I (X1 , X2 ; Y1 )} ∩ {R1 < I (X1 ; Y1 |X2 )}}
(10)
evaluated with PX1 X2 Y1 = PX1 × PX2 × WY1 |X1 X2 . Note
that this region can be obtained also by using standard
typicality-based achievability arguments (see, e.g., [23]).
Fig. 1 demonstrates a qualitative description of this region. The interpretation is as follows: The corner point
(I (X1 ; Y1 |X2 ) , I (X2 ; Y1 )) is achieved by first decoding the
interference (the second user), canceling it, and then decoding
the first user. The sum-rate constraint can be achieved by
joint decoding the two users (similarly to MAC), and thus,
obviously, also by our optimal decoder. Finally, the region
R1 < I (X1 ; Y1 ) and R2 ≥ I (X2 ; Y1 |X1 ) means that we
decode the first user while treating the interference as noise.
Evidently, from the perspective of the first decoder, which is
interested only in the message transmitted by the first sender,
the second sender can use any rate, and thus there is no bound
on R2 whenever R1 < I (X1 ; Y1 ). Now, it was shown in [10]
(8h)
I(X2 ; Y1 )
I(X1 ; Y1 )
Fig. 1.
I(X1 ; Y1 |X2 )
R1
Rate region Rordinary,1 for which Ẽ1 (R1 , R2 ) > 0.
that the error exponent achievable for the first user under the
ordinary random coding regime is zero outside the closure
of Rordinary,1 . Whence, this fact and the above conclusion,
characterize the rate region where the attainable exponent with
ordinary random coding is positive. Notice that Rordinary,1 is
well-known to be contained in the HK region [10, 23].
• Existence of a single code: our result holds true on the
average, where the averaging is done over the random choice
of codebooks. It can be shown (see, for example, [24, p. 2924])
that there exists deterministic sequence of fixed composition
codebooks of increasing block length n for which the same
5
asymptotic error performance can be achieved for both users
simultaneously.
• About the proof: it is instructive to discuss (in some more
detail than earlier) one of the main difficulties in proving
Theorem 1, which is customary to multiuser systems, such
as the IFC. Without loss of generality, we assume throughout,
that the transmitted codewords are x1,0 and x2,0 . Accordingly,
the average probability of error associated with the decoder
(3) is given by (11), shown at the top of the next page, where
F0 , (X 1,0 , X 2,0 , Y 1 ). In contrast to previous works, applying the type class enumerator1 method [14], is not a simple
task. Since we are interested in the optimal decoder, each
event of the union in (11), depends on the whole codebook
of the second user. One may speculate that this problem can
be tackled by conditioning on the codebook of the second
user, and then (12). However, the cost of this conditioning is a
very complicated (if not intractable) large deviations analysis
of some quantities. The consequence of this situation is that
in order to analyze the probability of error, it is required to
analyze the joint distribution of type class enumerators, and
not just rely on their marginal distributions, as is usually done,
e.g., [15-18].
Another difficulty is handling the union in (11). By the
union bound and Shulman’s inequality [20, Lemma A.2], we
know that for a sequence of pairwise independent events,
N
{Ai }i=1 , the following holds
)
)
(N
( N
[
X
1
Ai
Pr {Ai } ≤ Pr
min 1,
2
i=1
i=1
)
( N
X
Pr {Ai } , (12)
≤ min 1,
i=1
which is a useful result when assessing the exponential behavior of such probabilities. Equation (12) is one of the
building blocks of tight exponential analysis of previously
considered point-to-point systems (see, e.g., [15-18], and many
references therein). However, in our case the various events are
not pairwise independent, and therefore this result cannot be
applied. To alleviate this problem, following the techniques of
[11], we derive new upper bounds on the probability of a union
of events, which takes into account such dependencies among
the events.
• As was mentioned in the Introduction, in [21], lower
bounds on the error exponents of both standard and cognitive
multiple-access channels (MACs) were suggested. Although
the motivation in [21] is different, their results apply also
for the IFC. Now, while in [21] the standard truncated union
bound was used, here our new upper bound on the probability
of a union of events, provides some potential gain over
[21]. Specifically, the lower bound in [21] is the same as
1 For a given y n ∈ Y n , and a given joint probability distribution Q
XY
on
X
n × Y, the type class enumerator, N (QXY ), is the number ofncodewords
xi in Cn whose conditional empirical joint distribution with y is QXY ,
namely, N (QXY ) = xn ∈ Cn : Q̂xn y n = QXY , where Q̂xn y n is the
empirical joint distribution of xn and y n , and |A| designates the cardinality
of a finite set A. Type class enumeration method refers to the process of
converting a sum of exponentially many terms (usually likelihood functions)
into polynomial number of type class enumerators, which are easier to analyze.
(8j) but
i.e., it is given by
without the
Ê2 (Q, R2 ) term,
minQ D(Q||W )+ Ê1 (Q, R2 )−R1 + , and thus, in general,
our result may be tighter. It should be stressed, however, that
we have not identified specific examples where the new term,
namely, Ê2 (Q, R2 ), dominated the maximum in (8i).
• The lower bound in [9] is extremely complicated, and it is
very difficult to compare it analytically to the lower bound in
Theorem 1. Nonetheless, we can still claim (in general) that
our lower bound is at least as good as the lower bound in [9].
Indeed, the first step in the analysis of the error exponent in
both our paper and in [9] is applying the union bound (actually,
here, we employ a tighter union bound). However, it will be
seen that every other passage in our analysis is exponentially
exact, while in [9], some steps are associated with inequalities
that may cause gaps in the exponential scale, and thus in
general, Ẽ(R1 , R2 ) ≥ E[9] (R1 , R2 ), for any (R1 , R2 ) ∈ R2+ ,
where E[9] (R1 , R2 ) is the lower bound in [9].
• Comparison with [9]: Similarly to [9], we present results for
the following channel: Y1 = X1 · X2 ⊕ Z and Y2 = X2 , where
X1 , X2 , Y1 , Y2 ∈ {0, 1}, Z ∼ Bern(p), “·” is multiplication,
and “⊕” is modulo-2 addition. In the numerical calculations,
we fix p = 0.01. Fig. 2 presents the lower bound on the
error exponent under optimal decoding, derived in this paper,
compared to the lower bound E[9] (R1 , R2 ) of [9], as a function
of R1 , for different values of PX1 , PX2 , and R2 . It can be seen
that our exponents are strictly better than those of [9].
C. The Han-Kobayashi Coding Scheme
Consider the channel model of Subsection III-B. The best
known inner bound on the capacity region is achieved by the
HK coding scheme [8]. The idea of this scheme is to split
the message m1 into “private” and “common” messages, m11
and m12 at rates R11 and R12 , respectively, such that R1 =
R11 + R12 . Similarly, m2 is split into m21 and m22 at rates
R21 and R22 , with R2 = R21 + R22 . The intuition behind
this splitting is based on the receiver behavior at low and high
signal-to-noise-ratio (SNR). Specifically, it is well-known [1]
that: (1) when the SNR is low, treating the interference as
noise is an optimal strategy, and (2) when the SNR is high,
decoding and then canceling the interference is the optimal
strategy. Accordingly, the above splitting captures the general
intermediate situation where the first decoder, for example, is
interested only in partial information from the second user, in
addition to its own intended message.
Next, we describe explicitly the coding strategy of [8]. Fix a
distribution PZ11 PZ12 PZ21 PZ22 PX1 |Z11 Z12 PX2 |Z21 Z22 , where
the latter two conditional distributions represent deterministic
mappings. For each k, k ′ ∈ {1, 2}, randomly and conditionally
independently generate a sequence z k,k′ (mk,k′ ) under the
uniform distribution across the type class T (PZkk′ ) for a given
PZk,k′ . To communicate a message pair (m11 , m12 ), sender 1
transmits x1 (z 11 , z 12 ), and analogously for sender 2. All our
results can be extended to the setting in which the codewords
are generated conditionally on a time-sharing sequence q.
However, this leads to more complicated notation. Thus, we
focus primarily on the case without time-sharing.
Let us now describe the operation of each receiver. Receiver
k = 1, 2, recovers its intended message mk and the common
6
(n)
P̄e,1 = Pr
=E
M[
2 −1
1 −1 M
X
i=1
P (Y 1 |X 1,i , X 2,j ) ≥
j=0
j=0
M[
2 −1
1 −1 M
X
Pr
i=1
P (Y 1 |X 1,i , X 2,j ) ≥
P (Y 1 |X 1,0 , X 2,j )
M
2 −1
X
j=0
j=0
P (Y 1 |X 1,0 , X 2,j ) F0
(11)
M11 7→ Z11
0.45
X1
M12 7→ Z12
0.4
Ẽ1 for R2 = 0.139, PX1 (1) = 0.6, PX2 (1) = 0.9
Ẽ1 for R2 = 0.277, PX1 (1) = 0.6, PX2 (1) = 0.7
Y2 → (M̂12 , M̂21 , M̂22 )
M21 7→ Z21
E[9] for R2 = 0.277, PX1 (1) = 0.6, PX2 (1) = 0.7
0.3
Y1 → (M̂11 , M̂12 , M̂21 )
P (Y12 |X12 )
E[9] for R2 = 0.139, PX1 (1) = 0.6, PX2 (1) = 0.9
0.35
Error Exponents
M
2 −1
X
X2
M22 7→ Z22
0.25
0.2
Fig. 3.
Han-Kobayashi coding scheme.
0.15
arg
0.1
0.05
= arg
0
0
max
(i,j,k)∈M11 ×M12 ×M21
0.1
0.2
0.3
0.4
0.5
0.6
R1 [nats/channel use]
P (y 1 |z 11,i , z 12,j , z 21,k )
(13)
max
(i,j,k)∈M11 ×M12 ×M21
M22 −1
1 X
P (y 1 |z 11,i , z 12,j , z 21,k , z 22,l ) . (14)
M22
l=0
Fig. 2. Comparison between Ẽ1 (R1 , R2 ) and E[9] (R1 , R2 ) of [9], as a
function of R1 for two different values of R2 and fixed choices of PX1 and
P X2 .
Accordingly, the probability of error for the code Cn and for
the first user, is defined as
Pe,1 (Cn ) , Pr {(m̂11 , m̂12 ) 6= (m11 , m12 )} ,
message from the other sender (although it is not required to).
This scheme is illustrated in Fig. 3. Note that this decoding
operation is the one that was used in [8], but there, the suboptimal non-unique simultaneous joint typical decoder [1, Ch.
II.7] was used. Here, by contrast, we use sub-optimal ML
decoding (the sub-optimality is due to the fact that our decoder
recovers also the common message from the other sender).
As will be explained in the sequel, analyzing the optimal ML
decoder is a challenging task, and therefore we will focus on
sub-optimal ML decoding.
We wish to find a lower bound on the error exponent,
achieved by the HK encoding functions, in conjunction with
the above described decoding functions. To this end, note
that by combining the channel and the deterministic mappings as indicated by the dashed box in Fig. 3, the channel (Z11 , Z12 , Z21 , Z22 ) 7→ (Y1 , Y2 ) is just a four-sender,
two-receiver, DMC interference channel, with virtual inputs.
Note that this formulation induces the Markovian structure
(Z11 , Z12 , Z21 , Z22 )−
◦ (X1 , X2 )−
◦ (Y1 , Y2 ), where the (virtual) input distributions, i.e., PZk,k′ for k, k ′ ∈ {1, 2},
can be optimized. We assume that the message quadruple
(m11 , m12 , m21 , m22 ) is uniformly distributed over M11 ×
M12 × M21 × M22 . Following the above descriptions, our
decoder for this problem is given by
(m̂11 , m̂12 , m̂21 ) =
(15)
and similarly for the second user. Our goal is to assess the
(n)
exponential rate of P̄e,1 , E {Pe,1 (Cn )}, where the average
is over the code ensemble, namely,
∗
EHK
(R1 , R2 ) , lim inf −
n→∞
1
(n)
log P̄e,1 ,
n
(16)
and similarly for the second user. In order to facilitate the
presentation of the following result, we move the technical
definitions to Appendix A. Our second main result is the
following.
∗
Theorem 2 Let EHK
(R1 , R2 ) be defined as in (16). Consider
the HK encoding scheme described above. For a discrete
memoryless two-user IFC, we have:
∗
EHK
(R1 , R2 ) ≥
max
(R11 ,R12 ,R21 ,R22 ):
R11 +R12 =R1
R21 +R22 =R2
ẼHK (R11 , R12 , R21 , R22 ),
(17)
for any R1 , R2 ≥ 0, where ẼHK (R11 , R12 , R21 , R22 ) is given
in (A.30).
Several remarks on Theorem 2 are in order.
• As before, an immediate byproduct of Theorem 2 is
finding the set of rates (R11 , R12 , R21 , R22 ) for which
7
ẼHK (R11 , R12 , R21 , R22 ) > 0, namely, for which the probability of error vanishes exponentially as n → ∞. It can be
shown that this set is given by the HK region, that is,
R11 ≤ I(Z1 ; Y1 |Z2 , Z3 ),
R12 ≤ I(Z2 ; Y1 |Z1 , Z3 ),
(18a)
(18b)
R21 ≤ I(Z3 ; Y1 |Z1 , Z2 ),
(18c)
R11 + R12 ≤ I(Z1 , Z2 ; Y1 |Z3 ),
R11 + R21 ≤ I(Z1 , Z3 ; Y1 |Z2 ),
(18d)
(18e)
R12 + R21 ≤ I(Z2 , Z3 ; Y1 |Z1 ),
R11 + R12 + R21 ≤ I(Z1 , Z2 , Z3 ; Y1 ),
(18f)
(18g)
evaluated with
PZ14 Y1 = PZ1 PZ2 PZ3 PZ4 PX1 |Z1 Z2 PX2 |Z3 Z4 WY1 |X1 X2
and similarly for the second user, where PX1 |Z1 Z2 and
PX2 |Z3 Z4 represent deterministic mappings. As was mentioned
earlier, it is possible to introduce a time-sharing sequence
q, and accordingly, (18) remains almost the same, but with
a time-sharing RV Q (with alphabet size bounded by eight
[8]), appearing at the conditioning of each the above mutual
information terms. Finally, we mention that in [23] it was
shown that by using the optimal ML decoder (given in (20))
instead of the non-unique simultaneous joint typical decoder
[8], we cannot improve the achievable region. This observation
do not for the error exponent.
2
• It can be shown that the error exponent in Theorem 2
is no worse than the error exponent in Theorem 1, namely,
ẼHK (R11 , R12 , R21 , R22 ) ≥ Ẽ1 (R1 , R2 ) for any (R1 , R2 )
such that R1 = R11 +R12 and R2 = R21 +R22 . Moreover, it is
well-known that upon optimizing the auxiliary RVs, {Zij }, the
HK region in (18) is strictly better than Rordinary,1 . Therefore,
this necessarily implies that for a certain region of high rates,
the HK error exponent in Theorem 2 will be positive while
the standard random coding error exponent in Theorem 1 will
be zero. On the other extreme, it is easy to show that for
(R1 , R2 ) = (0, 0) the error exponent in Theorem 2 equals to
the error exponent in Theorem 1, so for small rates there is
no improvement in the error exponents.
• Contrary to the ordinary ensemble, described in Subsection
III-B, the HK ensemble depends on some auxiliary RVs which
should be optimized. For a give pair of rates (R1 , R2 ), our
error exponent formula provides a criterion for the choice of
the optimal auxiliary RVs: maximize the lower bound on the
error exponent in Theorem 2, w.r.t. the auxiliaries, subject to
some relevant constraints. As a matter of fact, for a given pair
of rates (R1 , R2 ), it is very likely that the optimal choice of
these auxiliaries will be different from the optimal choice for
the same pair in the achievable region. Indeed, even in the
single-user case, the capacity achieving distribution is usually
different from the optimal distribution in the error exponent
sense.
2 By definition, the ordinary ensemble is a simple instance of the HK
ensemble, and thus the latter is indeed better upon optimization of the auxiliary
RVs {Zij }. To see that the ordinary ensemble is a special case of the HK
ensemble, we take Z11 = X1 , Z12 = Z21 = ∅, and Z22 = X2 .
• Using the same techniques and tools derived in this paper, we can consider other random coding ensembles. For
example, we can analyze the error exponents resulting from
the hierarchical code ensemble. Specifically, in this ensemble,
the message m1 is split into common and private messages
m11 , m12 at rates R11 and R12 , respectively, such that
R1 = R11 + R12 . Similarly m2 is split into m21 , m22 at
rates R21 and R22 , respectively, such that R2 = R21 + R22 .
Then, we first randomly draw a rate R11 codebook of block
length n according to a given distribution. Then, for each such
codeword, we randomly and conditionally independently generate a rate R12 codebook of block length n. In other words,
the code has a tree structure with two levels, where the first
serves for “cloud centers”, and the second for the “satellites”.
We do the same for the second user. Under this ensemble,
we can analyze the optimal decoder. Note, however, that this
ensemble is different from the product ensemble considered
in Theorem 2. Indeed, while for the former for each first
stage codeword (cloud center) we independently draw a new
codebook (satellites), for the latter, for each cloud center we
have the same satellite codebook. Loosely speaking, this means
that the product ensemble is “less random”. From the point of
view of achievable region, however, the hierarchical ensemble
is equivalent to the product ensemble used in HK scheme
[1, Ch. II.7]. Nonetheless, the error exponents associated with
these ensembles could be different.
• In Theorem 2 we assumed the sub-optimal decoder given
in (14). Indeed, the optimal decoder for our problem is given
by:
(m̂11 , m̂12 ) = arg max P (y 1 |z 11,i , z 12,j )
i,j
(19)
= arg max
i,j
1
M21 M22
MX
21 −1 MX
22 −1
k=0
P (y 1 |z 11,i , z 12,j , z 21,k , z 22,l ) . (20)
l=0
Unfortunately, it turns out that analyzing the HK scheme (in
conjunction with (20)) is much more difficult, and requires
some more delicate tools from large deviations theory. Specifically, the main difficulty in the derivations, is to analyze the
large deviations behavior of a two-dimensional sum (due to the
double summation in (20)) involving binomial RVs which are
strongly dependent (contrary to the standard one-dimensional
version, see, e.g., [15, p. 6027-6028]). Nonetheless, we note
that for the hierarchical code ensemble described above, the
optimal decoder can be analyzed. Indeed, for this ensemble, it
is clear that the optimal decoder is given by
(m̂11 , m̂12 ) = arg max P (y 1 |x1 (i, j))
i,j
= arg max
i,j
(21)
MX
21 −1 MX
22 −1
1
P (y 1 |x1 (i, j), x2 (k, l))
M21 M22
k=0
l=0
(22)
where x1 (i, j) , f1 (x′1 (i), x′′1 (i, j)) and x2 (i, j) ,
f2 (x′2 (i), x′′2 (i, j)) due to the hierarchical structure. Now,
while here too, we will deal with two-dimensional summation,
the summands will be independent, given the cloud centers
codebook, and the proof can be carried out smoothly.
8
IV. P ROOFS OF M AIN R ESULTS
A. Proof of Theorem 1
Without loss of generality, we assume throughout, that the
transmitted codewords are x1,0 and x2,0 , and due to the fact
that we analyze the first decoder, for convenience, we use y
instead of y 1 . Accordingly, the average probability of error associated with the optimal decoder (3), is given by (24), shown
at the top of the next page, where F0 , (X 1,0 , X 2,0 , Y ). In
the following, we propose new upper bound on the probability
of a union of events, which are suitable for some structured
dependency between the events, as above.
In order to give some motivation for this new bound, we
first rewrite (23) in another (equivalent) form. Specifically, we
express (24) in terms of the joint types of (X 1,0 , X 2,0 , Y )
and {(Y , X 1,i , X 2,j )}i,j . First, for a given joint distribution
QX1 X2 Y of (x1 , x2 , y), let
1
f (QX1 X2 Y ) , log P (y|x1 , x2 )
(25)
n
(26)
= EQ log WY |X1 X2 (Y |X1 X2 ) .
For a given joint type QX1,0 X2,0 Y of the random
vectors (X 1,0 , X 2,0 , Y ), define the set TI QX1,0 X2,0 Y , given
in (27), shown at the top of the next page. The set
TI (QX1,0 X2,0 Y ) is the set of all possible types of (X 1,i , C2 ),
where C2 denotes the codebook of the second user, which lead
to a decoding error when (X 1,0 , X 2,0 , Y ) ∈ T (QX1,0 X2,0 Y )
is transmitted. The various marginal constraints in (28) and
(29) arise from the fact that we are assuming constantcomposition random coding and, of course, fixed marginals
due to the given fixed joint distribution QX1,0 X2,0 Y . Finally,
the constraint
M
2 −1 h
i
X
k
k
nf (Q̃0X1 X2,0 Y )
e
+
enf (Q̃X1 X2 Y ) − enf (Q̂X1 X2 Y )
k=1
≥ enf (QX1,0 X2,0 Y )
(30)
in (27), represents a decoding error event, that is, it holds if
and only if
M
2 −1
X
P (y|x1,i , x2,j ) ≥
M
2 −1
X
P (y|x1,0 , x2,j ) ,
(31)
j=0
j=0
or, equivalently,
P (y|x1,i , x2,0 ) +
M
2 −1
X
[P (y|x1,i , x2,j ) − P (y|x1,0 , x2,j )]
j=1
≥ P (y|x1,0 , x2,0 ) ,
(32)
for (x1,0 , x2,0 , y) ∈ T (QX1,0 X2,0 Y ), (x1,i , x2,0 , y) ∈
T (Q̃0X1 X2,0 Y ), (x1,i , x2,j , y)
∈
T (Q̃jX1 X2 Y ), and
(x1,0 , x2,j , y) ∈ T (Q̂jX1,0 X2 Y ), for j = 1, 2, . . . , M2 − 1.
Now, with these definitions, fixing QX1,0 X2,0 Y , and letting
(x1,0 , x2,0 , y) be an arbitrary triplet of sequences such that
(x1,0 , x2,0 , y) ∈ T (QX1,0 X2,0 Y ), it follows, by definition,
that the error event
M
M[
2 −1
2 −1
1 −1 M
X
X
P (Y |X 1,0 , X 2,j )
P (Y |X 1,i , X 2,j ) ≥
i=1
j=0
j=0
(33)
can be rewritten, in terms of types, as follows
M[
1 −1
i=1
n
Q̃jX
1 X2 Y
,Q̂jX
[
1 X2 Y
o
j
∈TI (QX1,0 X2,0 Y )
(X 1,i , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
oM2 −1
n
(X 1,i , X 2,j , y) ∈ T (Q̃jX1 X2 Y )
,
j=1
n
o
M2 −1
(x1,0 , X 2,j , y) ∈ T (Q̂jX1,0 X2 Y )
j=1
.
(34)
We wish to analyze the probability of the event in (34),
conditioned on F0 . Note that the inner union in (34) is over
vectors of types (an exponential number of them). Finally, for
the sake of convenience, we simplify the notations of (34),
and write it equivalently as
X 1,i ∈ Al,0 ,
M[
1 −1 [
(X 1,i , X 2,j ) ∈ Al,j , for j = 1, . . . , M2 − 1,
i=1 l
X 2,j ∈ Ãl,j , for j = 1, . . . , M2 − 1
(35)
where, again, the index “l” in the inner union runs over the
combinations of types (namely, l = {Q̃jX1 X2 Y , Q̂jX1 X2 Y }j )
that belong to TI (QX1,0 X2,0 Y ), and the various sets
{Al,j , Ãl,j }l,j correspond to the typical sets in (34) (recall
that (x1,0 , x2,0 , y) are given at this stage). Next, following
the ideas of [11], we provide a new upper bound on a generic
probability which has the form of (35). The proof of this
lemma is relegated to Appendix B.
L
1
Lemma 1 Let {V1 (i)}i=1
, V2 , V3 , . . . , VK be independent sequences of independently and identically distributed (i.i.d.)
RVs on the alphabets V1 × V2 × . . . × VK , respectively, with
V1 (i) ∼ PV1 , V2 ∼ PV2 , . . . , VK ∼ PVK . Fix a sequence of
N
N
N
sets {Ai,1 }i=1 , {Ai,2 }i=1 , . . . , {Ai,K−1 }i=1 , where Ai,j ⊆
V1 ×Vj+1 , for 1 ≤ j ≤ K−1 and for all 1 ≤ i ≤ N . Also, fix a
N
set {Ai,0 }i=1 where Ai,0 ⊆ V1 for all 1 ≤ i ≤ N , and another
N
N
N
sequence of sets {Gi,2 }i=1 , {Gi,3 }i=1 , . . . , {Gi,K }i=1 , where
Gi,j ⊆ Vj , for 2 ≤ j ≤ K and for all 1 ≤ i ≤ N . Define
K−1
\
(v1 , vj+1 ) ∈ Am,j ,
Bm,1 , v1 : v1 ∈ Am,0 ,
j=1
K
\
(36)
vj ∈ Gm,j for some {vj }K
j=2 ,
j=2
and
Bm,2 ,
K
{vj }j=2 : v1 ∈ Am,0 ,
K
\
j=2
K−1
\
(v1 , vj+1 ) ∈ Am,j ,
j=1
vj ∈ Gm,j for some v1
,
(37)
for m = 1, 2, . . . , N . Then, a general upper bound is given in
(38), shown at the top of the next page, with (V1 , . . . , VK ) ∼
PV1 · · · × PVK .
9
(n)
P̄e,1 = Pr
=E
M[
2 −1
1 −1 M
X
i=1
P (Y |X 1,i , X 2,j ) ≥
j=0
j=0
M[
2 −1
1 −1 M
X
Pr
i=1
M
2 −1
X
P (Y |X 1,i , X 2,j ) ≥
P (Y |X 1,0 , X 2,j )
M
2 −1
X
j=0
j=0
(23)
P (Y |X 1,0 , X 2,j ) F0
n
oM2 −1 n
oM2 −1
TI QX1,0 X2,0 Y , Q̃0X1 X2,0 Y ∈ S0 ,
Q̃kX1 X2 Y
, Q̂kX1,0 X2 Y
∈ S1 :
k=1
e
nf (Q̃0X1 X2,0 Y )
+
M
2 −1 h
X
e
k=1
nf (Q̃k
X1 X2 Y )
−e
nf (Q̂k
X1 X2 Y )
k=1
i
≥e
nf (QX1,0 X2,0 Y )
o
n
S0 (QX1,0 X2,0 Y ) , Q̃0X1 X2,0 Y : Q̃0X1 = PX1 , Q̃0X2 = PX2 , Q̃0X2,0 Y = QX2,0 Y ,
n
oM2 −1
oM2 −1 n
: Q̃kX1 = PX1 , Q̃kX2 = PX2 , Q̃kY = QY ,
, Q̂kX1,0 X2 Y
S1 (QX1,0 X2,0 Y ) ,
Q̃kX1 X2 Y
=
Pr
[
i
(
N
[
m=1
(
V1 (i) ∈ Am,0 ,
(
K−1
\
≤ min 1, L1 Pr
L1 Pr
(
,
(27)
(28)
= PX2 , Q̂kX1,0 Y = QX1,0 Y , ∀1 ≤ k ≤ M2 − 1
PX1 , Q̂kX2
Q̃kX2 Y
(
)
k=1
k=1
Q̂kX1,0
(24)
m=1
(
, Q̃kX1 Y
(V1 (i), Vk+1 ) ∈ Am,k ,
k=1
( N
[
)
{V1 ∈ Bm,1 } , Pr
m=1
N
[
=
Q̂kX2 Y
V1 ∈ Am,0 ,
K−1
\
k=1
K
\
k=2
( N
[
=
Q̃m
X1 Y
Vk ∈ Gm,k
m=1
, ∀k, m .
(29)
)))
)
o
n
K
{Vj }k=2 ∈ Bm,2
,
(V1 , Vk+1 ) ∈ Am,k ,
K
\
k=2
Vk ∈ Gm,k
)))
(38)
Next, we apply Lemma 1 to the problem at hand. To this types. Accordingly, we have:
end, we choose the following parameters in accordance to the 1) Am,i = Al,i , for 0 ≤ i ≤ M2 − 1,
notations used in Lemma 1. Recall that we deal with
2) Gm,i = Ãl,i−1 , for 2 ≤ i ≤ M2 .
0
X
∈
A
,
1,i
l,0
M[
1 −1 [
These sets correspond to each of the typical sets T (Q̃X1 X2,0 Y ),
M
−1
M
−1
2
2
(X 1,i , X 2,j ) ∈ Al,j , for j = 1, . . . , M2 − 1,
.
, and {T (Q̂kX1,0 X2 Y )}k=1
{T (Q̃kX1 X2 Y )}k=1
• According
i=1 l
to
(36)
and
(37)
we
need
to define
X 2,j ∈ Ãl,j , for j = 1, . . . , M2 − 1
j
j
0
,
{
Q̃
,
Q̂
and
B
=
B
(
Q̃
m,1
1
(39)
X1 X2,0 Y
X1 X2 Y
X1 X2 Y }j )
j
j
0
Bm,2
=
B2 (Q̃X1 X2,0 Y , {Q̃X1 X2 Y , Q̂X1 X2 Y }j ). Using
and in Lemma 1 we have considered:
(36) and (37), we get (41) and (42), given at the top of the
V1 (i) ∈ Am,0 ,
L1 [
N
next page.
[
(V1 (i), Vj+1 ) ∈ Am,j , for j = 1, . . . , K − 1
.
Thus, invoking Lemma 1, we have (44), where each of
i=1 m=1
V2 ∈ Gm,2 , . . . , VK ∈ Gm,K
the probabilities at the r.h.s. of (44) are conditioned on F0 .
(40)
Therefore, we were able to simplify the problematic union
Thus, comparing (39) and (40), we readily notice the following over the codebook of the first user. Note, however, that we
parallels:
cannot (directly) apply here the method of types due to the
fact that the union is over an exponential number of types, and
• The numbers of events in the unions over i is L1 = M1 − 1.
Also, we have K = M2 independent random vectors V1 (i) = thus a more refined analysis is needed. We start by analyzing
X 1,i and Vl+1 = X 2,l , for 1 ≤ i ≤ M1 − 1 and 1 ≤ l ≤ the last term at the r.h.s. of (44). To this end, we will invoke
the type enumeration method, but first, the main observation
M2 − 1.
here is that similarly to the passage from (33) to (34), the last
• The union over m corresponds to a union over l, which as
was mentioned before, is actually a union over a vector of term at the r.h.s. of (44) can be rewritten as (46), shown at the
10
(x1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
n
oM2 −1
j
x
:
(x
,
x
,
y)
∈
T
(
Q̃
)
,
1
1
2,j
X1 X2 Y
Bm,1 =
j=1
n
o
M
−1
2
(x1,0 , x2,j , y) ∈ T (Q̂jX1,0 X2 Y )
for some {x2,j }j
j=1
(x1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
n
oM2 −1
{x2,j }j≥1 :
(x1 , x2,j , y) ∈ T (Q̃jX1 X2 Y )
,
Bm,2 =
n
oM2 −1j=1
(x1,0 , x2,j , y) ∈ T (Q̂jX1,0 X2 Y )
for some x1
j=1
(n)
P̃e,1
M[
2 −1
1 −1 M
X
, Pr
≤ min
i=1
Pr
P (Y |X 1,i , X 2,j ) ≥
j=0
j=0
1, M1 · Pr
n
n
Q̃jX
1 X2 Y
,Q̂jX
Q̃jX
1 X2 Y
[
1 X2 Y
o
j
Q̃jX X Y
1 2
[
,Q̂jX
1 X2 Y
o
j
∈TI (QX1,0 X2,0 Y )
∈TI (QX1,0 X2,0 Y )
n
[
,Q̂jX X Y
1 2
o
∈TI (QX1,0 X2,0 Y )
j
(47)
1
log P (y|x1,1 , x2,0 ) .
n
(48)
For a given (y, x1,0 , x1,1 , x2,0 ), and a given joint probability
distribution QX1 X2 Y on X1 × X2 × Y, let N1 (QX1 X2 Y )
designate the number of codewords {X 2,j }j (excluding x2,0 )
whose conditional empirical distribution with y and x1,1 is
QX1 X2 Y , that is,
N1 (QX1 X2 Y ) ,
{X 2,j }j≥1 ∈ B2 Q̃0X1 X2,0 Y , (Q̃jX1 X2 Y , Q̂jX1 X2 Y )j )
,
1
log P (y|x1,0 , x2,0 ) ,
n
M
2 −1
X
(43)
X 1,1 ∈ B1 Q̃0X1 X2,0 Y , (Q̃jX1 X2 Y , Q̂jX1 X2 Y )j )
,
(X 1,1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
oM2 −1
(X 1,1 , X 2,j , y) ∈ T (Q̃jX1 X2 Y )
,
n
oj=1
M2 −1
(x1,0 , X 2,j , y) ∈ T (Q̂jX1,0 X2 Y )
j=1
I {(x1,1 , X 2,j , y) ∈ T (QX1 X2 Y )} ,
j=1
(49)
(44)
and let N2 (QX1 X2 Y ) designate the number of codewords
{X 2,j }j (excluding x2,0 ) whose conditional empirical distribution with y and x1,0 is QX1 X2 Y , that is
N2 (QX1 X2 Y ) ,
M
2 −1
X
I {(x1,0 , X 2,j , y) ∈ T (QX1 X2 Y )} .
j=1
(50)
Also, recall that
1
log P (y|x1 , x2 )
n
= EQ log WY |X1 X2 (Y |X1 , X2 )
f (QX1 X2 Y ) =
and
r,
(42)
P (Y |X 1,0 , X 2,j ) F0
next page. That is, we returned back to the structure of the
original probability in (24), but now, without the union over
the codebook of the first user. Note that the conditioning on
the random vector X 1,1 in (46), is due to the fact that X 1,1 is
common to all the summands in the inner summation over the
codebook of the second user. We next evaluate the exponential
behavior of the probability in (46). For a given realization of
Y = y, X 1,0 = x1,0 , X 1,1 = x1,1 , and X 2,0 = x2,0 , let us
define
s,
(41)
M1 · Pr
n
M
2 −1
X
(51)
(52)
where QX1 X2 Y is understood to be the joint empirical distribution of (x1 , x2 , y) ∈ X1n × X2n × Y n . Thus, in terms of the
above notations, we may write:
M
2 −1
X
P (y|x1,1 , X 2,j ) = enr
j=0
+
X
N1 (QX1 X2 Y ) enf (QX1 X2 Y )
(53)
QX2 |X1 Y ∈S(QX1 Y )
, enr + N1 (QX1 Y ).
(54)
where for a given QX1 Y , S(Q
X1 Y ) is defined as
the set of all distributions
QX2 |X1 Y , such that
11
(X 1,1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
oM2 −1
[
j
(X
,
X
,
y)
∈
T
(
Q̃
)
,
1,1
2,j
X1 X2 Y
Pr
j=1
o
n
o
n j
M
−1
2
j
∈TI (QX1,0 X2,0 Y )
Q̃X X Y ,Q̂jX X Y
(x1,0 , X 2,j , y) ∈ T (Q̂X1,0 X2 Y )
1 2
1 2
j
j=1
M
−1
2
2 −1
MX
X
P (Y |X 1,0 , X 2,j ) F0
P (Y |X 1,1 , X 2,j ) ≥
= Pr
j=0
j=0
M
2 −1
2 −1
MX
X
P (Y |X 1,0 , X 2,j ) F0 , X 1,1 F0 .
P (Y |X 1,1 , X 2,j ) ≥
= E Pr
n
(x1 ,y)∈X1 ×Y QX1 Y (x1 , y) QX2 |X1 Y
for all x2 ∈ X2 , namely,
Similarly,
P (y|x1,0 , X 2,j ) = ens
j=0
X
+
QX2 |X1,0 Y ∈S(QX1,0 Y )
N2 QX1,0 X2 Y enf (QX1,0 X2 Y ) (56)
, ens + N2 (QX1,0 Y ).
(57)
where for a given QX1,0 Y , S(Q
X1,0 Y ) is defined as
the
set
of
all
distributions
QX2 |X1,0 Y , such that
P
=
(x1 ,y)∈X1 ×Y QX1,0 Y (x1 , y) QX2 |X1,0 Y (x2 |x1 , y)
PX2 (x2 ) for all x2 ∈ X2 (similarly as in (55)). For simplicity
of notation, in the following, we use Q and Q̃ to denote
QX1 X2 Y and QX1,0 X2 Y , respectively. Therefore, with these
definitions in mind, we wish to calculate (given (F0 , X 1,1 ))
M
M
2 −1
2 −1
X
X
P (Y |x1,0 , X 2,j )
Pr
P (Y |x1,1 , X 2,j ) ≥
j=0
j=0
= Pr N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr
(58)
where s, r, N1 (Q) and N2 (Q) are given in (47), (48), (54),
and (57), respectively.
Let ε > 0 be arbitrarily small, and
1
define i1 , nǫ
log P (y|x1,0 , x2,0 ) . Then,
Pr N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr
⌈R2 /ε⌉
=
X
i=i1
⌈R2 /ε⌉
≤
X
i=i1
n
Pr eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε ,
N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr
n
Pr eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε ,
N1 (QX1 Y ) ≥ e
⌈R2 /ε⌉
=
X
i=i1
n
niε
× Pr N1 (QX1 Y ) ≥ eniε + ens − enr
(x2 |x1 , y) = PX2 (x2 )
S(QX1 Y ) = Q′X1 X2 Y : Q′X1 Y = QX1 Y , Q′X2 = PX2 .
(55)
M
2 −1
X
+e
ns
−e
(46)
j=0
j=0
P
(45)
nr
Pr eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε
o
o
o
o
| eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε .
(60)
It is not difficult to show that (see, e.g., [15, p. 6028])
o
n
Pr ent ≤ N2 (QX1,0 Y ) ≤ en(t+ε)
t < t0 (QX1,0 Y ) − ε
0
·
= 1
t0 (QX1,0 Y ) − ε ≤ t ≤ t0 (QX1,0 Y )
exp −nE(t, QX1,0 Y )
t > t0 (QX1,0 Y )
(61)
where
t0 (QX1,0 Y ) , R2 +
max
Q̃∈S(QX1,0 Y ): IQ̃ (X2 ;X1,0 ,Y )≤R2
i
h
f (Q̃) − IQ̃ (X2 ; X1,0 , Y ) ,
(62)
in which S(Q) is defined in (55), f (Q) is given in (52), and
h
i
E(t, QX1,0 Y ) , min IQ̃ (X2 ; X1,0 , Y ) − R2 :
+
h
i
f (Q̃) + R2 − IQ̃ (X2 ; X1,0 , Y ) ≥ t . (63)
+
Now, in the exponential scale, the term at the r.h.s. of
(60) is dominated by one of the summands, and we
claim that the dominant contribution to the sum over i
is due to the
first term3 , i = t0 (QX1,0 Y )/ε. Indeed,
nkε
let Ak , e
≤ N2 (QX1,0 Y ) ≤ en(k+1)ε and Bk ,
nkε
N1 (QX1 Y ) ≥ e
+ ens − enr , and notice that the summands in (60) correspond to Pr {Ak ∩ Bk }. According to
(61), Pr {At0 } → 1 (the exponent E(kε, QX1,0 Y ) vanishes),
and note that Pr {Bk } is monotonically decreasing with k.
Therefore,
Pr {At0 ∩ Bt0 } ≤ max Pr {Ak ∩ Bk }
k≥t0
≤ max Pr {Bk } = Pr {Bt0 } .
k≥t0
(64)
(59)
3 Note that according to (61), Pr{eniε ≤ N (Q
n(i+1)ε }
2
X1,0 Y ) ≤ e
vanishes (in the exponential scale) for i < t0 (QX1,0 Y )/ε. Thus, to asses
the exponential scale of (60) we consider only the indices correspond to i ≥
t0 (QX1,0 Y )/ε.
12
We claim that (79) can be rewritten as4
On the other hand,
·
Pr {At0 ∩ Bt0 } = Pr {Bt0 } − Pr Atc0 ∩ Bt0
≥ Pr {Bt0 } − Pr Atc0 .
Pe,3 =
max
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y ) Q∈S(QX1 Y )
(80)
which
follows
from
the
fact
that
the
set
{x2 : (x1,1 , x2 , y) ∈ T (QX1 X2 Y ), for some x1,1 }
equals
T (QX2 Y ) (see, e.g., [11, eqs. (24)-(25)]). Thus, by using
standard large deviations techniques (see, e.g., [15, p. 6027])
·
Pr {At0 ∩ Bt0 } = Pr {Bt0 } .
(66)
Combining (60), (64) and (66), and the fact that ε is arbitrarily
small, we get (68) (shown at the top of the next page) by
using standard large deviations techniques (see, e.g., [15, p.
6027]), where N1 (Q) and S(Q) are defined in (49) and (55),
respectively, and
o
n
L̃ , Q : max [t0 , s] − f (Q) ≤ [R2 − IQ (X2 ; X1 , Y )]+ .
(69)
Thus,
Pr N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr
·
= exp −nE1 (QX1 X2,0 Y , QX1,0 X2,0 Y )
o
n
Pr Ñ1 (Q) ≥ enγ ,
(65)
Thus, due to the fact that Pr {At0 } → 1 super-exponentially
fast [15, p. 6028], we may conclude that
·
max
(70)
where E1 (·, ·) is defined in (8c). Note that when r >
max [t0 , s], the r.h.s. term of the inequality in the probability
in (67) is negative, and due to the fact that the enumerator
is nonnegative, the overall probability is unity. Finally, we
average over X 1,1 given F0 . Using the method of types, we
readily obtain (73), given at the top of the next page, where
Ŝ(QX2 Y ) , Q′X1 X2 Y : Q′X2 Y = QX2 Y , Q′X1 = PX1 ,
(74)
and Ê1 (·, ·) is defined in (8g). This completes the analysis of
the last term at the r.h.s. of (44).
Next, we analyze the second and third terms at the r.h.s.
of (44). Recall that the latter is given by (75). Accordingly,
in the spirit of (46), we note that Pe,3 can be equivalently
rewritten as (78), shown at the top of the next page, where
s, r, N1 (Q), N2 (Q), and Ŝ(Q), are given in (47), (48),
(54), (57), and (74), respectively, and the second passage
follows by using the method of types. Now, due to the fact
that only N1 (and not N2 ) in (78) depends on x1,1 , and
since the analysis in (58)-(67) is independent of x1,1 , it can
be repeated here, and we obtain (79), shown at the top of
page 14, where N1 (Q) is given in (49), and we have defined
] + en[s−f (Q)] − en[r−f (Q)] . Recall
enγ , en[t0 (QX1,0 Y )−f (Q)P
M2 −1
(49), and let Ñ1 (Q) ,
j=1 I {(X 2,j , y) ∈ T (QX2 Y )}.
Pe,3 =
where
max
max
(81)
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y ) Q∈S(QX1 Y )
1
e−n[IQ (X2 ;Y )−R2 ]+
0
r > max [t0 , s]
r ≤ max [t0 , s] , Q ∈ L̂
(82)
c
r ≤ max [t0 , s] , Q ∈ L̂
o
n
L̂ , Q : max [t0 , s] − f (Q) ≤ [R2 − IQ (X2 ; Y )]+ .
(83)
Therefore,
n
o
·
Pe,3 = exp −nÊ2 (QX1,0 X2,0 Y , R2 )
(84)
where Ê2 (QX1,0 X2,0 Y , R2 ) is defined in (8h). This completes
the analysis of the third term at the r.h.s. of (44). Finally, recall
that the second term at the r.h.s. of (44) is given by
[
A , M1 · Pr
TI (QX1,0 X2,0 Y )
X 1,1
∈ B1 Q̃0X1 X2,0 Y , (Q̃jX1 X2 Y , Q̂jX1 X2 Y )j ) (85)
and is equivalent to (86), given at the top of page 14. This
term can be analyzed as before, but, we claim that it is
actually larger than the fourth term at the r.h.s. of (44), and
thus, essentially, does not affect the minimum in (44). Indeed,
recall that the fourth term is given by (87), shown at the top
of page 14, and since the factor M1 is common to both A
and B, we just need to compare the probabilities in these
terms. However, it is obvious that the probability term in
B is smaller than the probability in A, due to the fact that
events in the former are contained in the events in the latter.
Indeed, this is equivalent to comparing Pr {(Z1 , Z2 ) ∈ Z}
and Pr {(Z1 , z2 ) ∈ Z, for some z2 ∈ Z2 }, where Z1 and Z2
4 It is easy to see that (80) is an upper bound on (79). The other direction
follows from:
o
n
o
n
∗
max Pr Ñ1 (Q̂) ≥ enγ(Q) = max Pr Ñ1 (Q̂) ≥ enγ(Q )
Q̂∈S(Q∗ )
Q,Q̂∈S(Q)
=
max
Q̂∈S(Q∗ )
n
∗
o
Pr N1 (Q̂) ≥ enγ(Q ) , for some x1,1 ∈ T (Q∗ )
n
o
≤ max max Pr N1 (Q̂) ≥ enγ(Q) , for some x1,1 ∈ T (Q) ,
Q
Q̂∈S(Q)
where in the first equality we designate Q∗ as the maximizer,
and
the second equality follows from the fact that T (QX2 Y ) =
x2 : (x1,1 , x2 , y) ∈ T (QX1 X2 Y ), for some x1,1 .
13
·
Pr N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr = Pr {Bt0 }
o
n
= Pr N1 (QX1 Y ) ≥ ent0 (QX1,0 Y ) + ens − enr
o
n
·
=
max Pr N1 (Q) ≥ en[t0 (QX1,0 Y )−f (Q)] + en[s−f (Q)] − en[r−f (Q)]
Q∈S(QX1 Y )
(67)
1
·
=
max
e−n[IQ (X2 ;X1 ,Y )−R2 ]+
Q∈S(QX1 Y )
0
r > max [t0 , s]
r ≤ max [t0 , s] , Q ∈ L̃
(68)
r ≤ max [t0 , s] , Q ∈ L̃c
M
2 −1
2 −1
MX
X
P (Y |X 1,0 , X 2,j ) F0 , X 1,1 F0
P (Y |X 1,1 , X 2,j ) ≥
E Pr
j=0
j=0
)
(
·
= exp −n
min
IQ (X1 ; X2,0 , Y ) + E1 (QX1 X2,0 Y , QX1,0 X2,0 Y )
(71)
(72)
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y )
n
o
, exp −nÊ1 (QX1,0 X2,0 Y , R2 )
Pe,3 , Pr
n
Pe,3 = Pr
Q̃jX X Y
1 2
[
[
,Q̂jX X Y
1 2
o
{X 2,j }j≥1 ∈ B2
j
∈TI (QX1,0 X2,0 Y )
P (y|x1,1 , x2,0 ) +
M
2 −1
X
j=1
·
=
max
Pr P (y|x1,1 , x2,0 ) +
+
M
2 −1
X
j=1
max
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y )
M
2 −1
X
(88)
=
X
P (z1 ) = Pr {(Z1 , z2 ) ∈ Z, for some z2 } .
(77)
(78)
(91)
z1 ∈Ṽ
Wrapping up, using (24), (44), and the last results, after
averaging w.r.t. F0 , we get (96), shown at the top of the next
page, as required.
(89)
B. Proof of Theorem 2
(z1 ,z2 )∈Z
P (z1 , z2 )
Pr N1 (QX1 Y ) − N2 (QX1,0 Y ) ≥ ens − enr , for some x1,1 ∈ T (QX1 X2,0 Y )|F0
Then, it is obvious that Z ⊆ V, and thus
X
P (z1 , z2 )
Pr {(Z1 , Z2 ) ∈ Z} =
(z1 ,z2 )∈V
P (y|x1,1 , X 2,j ) ≥ P (y|x1,0 , x2,0 )
P (y|x1,0 , X 2,j ) , for some x1,1 ∈ T (QX1 X2,0 Y ) F0
Ṽ , {z1 ∈ Z1 : (z1 , z2 ) ∈ Z, for some z2 ∈ Z2 } .
≤
(76)
j=1
are RVs that are defined over the alphabets Z1 and Z2 ,
respectively, and Z ⊆ Z1 × Z2 . Let V , Ṽ × Z2 , in which
X
(75)
P (y|x1,1 , X 2,j ) ≥ P (y|x1,0 , x2,0 )
P (y|x1,0 , X 2,j ) , for some x1,1 ∈ T (QX1 X2,0 Y ) F0 ,
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y )
=
Q̃0X1 X2,0 Y , (Q̃jX1 X2 Y , Q̂jX1 X2 Y )j )
.
j=1
QX1 |X2,0 Y
+
M
2 −1
X
(73)
(90)
Without loss of generality, we assume throughout, that the
transmitted codewords are x1,0 and x2,0 which correspond
14
⌈R2 /ε⌉
·
Pe,3 =
X
max
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y ) i=i
1
× Pr
·
=
max
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y )
·
=
max
o
n
Pr eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε
[
x1,1
o
n
Pr N1 (QX1 Y ) ≥ ent0 (QX1,0 Y ) + ens − enr , for some x1,1
max
QX1 |X2,0 Y ∈Ŝ(QX2,0 Y ) Q∈S(QX1 Y )
n
N1 (QX1 Y ) ≥ eniε + ens − enr eniε ≤ N2 (QX1,0 Y ) ≤ en(i+1)ε
[
A = M1 · Pr
TI (QX1,0 X2,0 Y )
Pr {N1 (Q) ≥ enγ , for some x1,1 } ,
(79)
(X 1,1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
oM2 −1
(X 1,1 , x2,j , y) ∈ T (Q̃jX1 X2 Y )
, for some {x2,j }
j=1
n
oM2 −1
(x1,0 , x2,j , y) ∈ T (Q̂jX1,0 X2 Y )
j=1
(X 1,1 , x2,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
n
oM2 −1
[
j
(X
,
X
,
y)
∈
T
(
Q̃
)
,
1,1
2,j
X1 X2 Y
B , M1 · Pr
j=1
n
o
TI (QX1,0 X2,0 Y )
M
−1
2
(x1,0 , X 2,j , y) ∈ T (Q̂jX1,0 X2 Y )
j=1
.
,
n
n
oo
(n) ·
P̄e,1 ≤ E min 1, e−n(Ê1 (QX1,0 X2,0 Y ,R2 )−R1 ) , e−nÊ2 (QX1,0 X2,0 Y ,R2 )
−n[Ê1 (QX1,0 X2,0 Y ,R2 )−R1 ]
−nÊ2 (QX1,0 X2,0 Y ,R2 )
+, e
= E min e
h
i
= E exp −n max Ê1 (QX1,0 X2,0 Y , R2 ) − R1 , Ê2 (QX1,0 X2,0 Y , R2 )
+
#)
(
"
·
D(QY |X1,0 X2,0 ||WY |X1,0 X2,0 |PX1,0 × PX2,0 ) + E(Q, R1 , R2 )
= exp −n
min
QY |X1,0 X2,0
h
i
E(Q, R1 , R2 ) , max Ê1 (QX1,0 X2,0 Y , R2 ) − R1 , Ê2 (QX1,0 X2,0 Y , R2 ) .
+
to z 11,0 , z 12,0 , z 21,0 and z 22,0 . Here, we distinguish between
several types of errors. Recall that the overall error probability
is given by
(n)
P̄e,1 = Pr {(m̂11 , m̂12 ) 6= (0, 0)} ,
(98)
which can be divided into six possible types of errors: (m̂11 6=
0, m̂12 = 0, m̂21 = 0), (m̂11 = 0, m̂12 6= 0, m̂21 = 0),
(m̂11 6= 0, m̂12 6= 0, m̂21 = 0), (m̂11 6= 0, m̂12 = 0, m̂21 6=
0), (m̂11 = 0, m̂12 6= 0, m̂21 6= 0), and (m̂11 6= 0, m̂12 6=
0, m̂21 6= 0). Note that the event (m̂11 = 0, m̂12 = 0, m̂21 6=
0) will not result in an error, and thus ignored. Obviously,
the exponent of the overall error probability in (98) is given
by the minimum between the error exponents corresponding to
each type of error individually. We start with analyzing the last
error event, which is also the most involved one. For this event,
(86)
(87)
(93)
(94)
(95)
(96)
(97)
the average probability of error, associated with the decoder in
(14), is given by (100), given at the top of the next page, where
Z̃ ijk , (Z 11,i , Z 12,j , Z 21,k ), Z̃ 0 , (Z 11,0 , Z 12,0 , Z 21,0 ),
and F0 , (Z̃ 0 , Z 22,0 , Y ). We will assess the exponential
behavior of (100) in the same manner as we did for (24).
Specifically, we start with expressing (100) in terms of types.
First, for a given joint distribution QZ14 Y , we let
f (QZ14 Y ) ,
1
log P (y|x1 (z 1 , z 2 ), x2 (z 3 , z 4 )) .
n
(101)
4 Y of the random vectors
Now, for a given joint type QZ1,0
4 Y ) in
(Z 1,0 , Z 2,0 , Z 3,0 , Z 4,0 , Y ), we define the set TI (QZ1,0
(102) given at the top of the next page. Now, with these
4 Y , it follows, by definition, that the
definitions, fixing QZ1,0
15
M[
21 −1
12 −1 M[
11 −1 M[
Pe(7) , Pr
=E
j=1
i=1
k=1
(M −1
22
X
l=0
M[
21 −1
12 −1 M[
11 −1 M[
Pr
j=1
i=1
P (Y |Z̃ ijk , Z 22,l ) ≥
k=1
l=0
(M −1
22
X
P (Y |Z̃ ijk , Z 22,l ) ≥
l=0
1
e
1,0 Z4 Y
l=1
nf (Q̃0Z 3 Z
4,0 Y
1
)
)
P (Y |Z̃ 0 , Z 22,l )
MX
22 −1
+
MX
22 −1
e
oM22 −1
l=1
nf (Q̃lZ 4 Y )
1
−e
(99)
)
P (Y |Z̃ 0 , Z 22,l )
l=0
n
oM22 −1 n
4 Y ),
TI (QZ1,0
Q̃0Z 3 Z4,0 Y ∈ S0 ,
Q̃lZ 4 Y
, Q̂lZ 3
1
MX
22 −1
F0
∈ S1 :
nf (Q̂lZ 3
1,0
Z4 Y
)
l=1
≥e
nf (QZ 4
1,0
Y
)
)
n
o
4 Y ) ,
S0 (QZ1,0
Q̃0Z 3 Z4,0 Y : Q̃0Zi = PZi , Q̃0Z4,0 Y = QZ4,0 Y , ∀1 ≤ i ≤ 4 ,
n 1
oM22 −1
oM22 −1 n
4 Y ) ,
: Q̃lZi = PZi , Q̃lY = QY ,
, Q̂lZ 3 Z4 Y
S1 (QZ1,0
Q̃lZ 4 Y
1
1,0
l=1
Q̂lZi
=
PZi , Q̂lZ 3 Y
1,0
l=1
3 Y , ∀1 ≤ i ≤ 4, ∀1 ≤ l ≤ M22 − 1
= QZ1,0
1
error event
j=1
i=1
(M −1
22
X
k=1
P (Y |Z̃ ijk , Z 4,l ) ≥
l=0
MX
22 −1
)
P (Y |Z̃ 0 , Z 4,l )
l=0
(105)
can be rewritten, in terms of types, as follows
M[
21 −1
12 −1 M[
11 −1 M[
i=1
j=1
k=1
[
(106)
TI (QZ 4
1,0 Y
)
(Z̃ ijk , z 4,0 , y) ∈ T (Q̃0Z 3 Z4,0 Y ),
1
oM22 −1
n
l
(Z̃ ijk , Z 4,l , y) ∈ T (Q̃Z 4 Y )
,
1
ol=1
n
M22 −1
(z̃ 0 , Z 4,l , y) ∈ T (Q̂l 3
)
Z Z4 Y
1,0
l=1
.
(107)
We next analyze the probability of (107), conditioned on F0 .
Note that the inner union in (107) is over vectors of types (an
exponential number of them). Finally, as before, we simplify
the notations of (107), and write it equivalently as
M[
11 −1 M[
12 −1 M[
21 −1 [
i=1
j=1
k=1
(108)
l
Z̃ ijk ∈ Al,0 ,
(Z̃ ijk , Z 4,m ) ∈ Al,m ,
Z 4,m ∈ Ãl,m ,
for m = 1, . . . , M22 − 1
for m = 1, . . . , M22 − 1
(109)
where, again, the index “l” in the inner union runs over the
combinations of types (namely, l = {Q̃lZ 4 Y , Q̂lZ 3 Z4 Y }l ) that
1
(102)
(103)
o
Q̃lZ4 Y = Q̂lZ4 Y , Q̃lZ 3 Y = Q̃m
Z 3 Y , ∀l, m .
M[
21 −1
12 −1 M[
11 −1 M[
(100)
1,0
4 Y ), and the various sets {Al,j , Ãl,j }l,j
belong to TI (QZ1,0
(104)
1
correspond to the typical sets in (107) (recall that
(z 1,0 , z 2,0 , z 3,0 , z 4,0 , y) are given in this stage). Similarly as
in the proof of Theorem 1, we derive upper bound on a generic
probability which have the form of (109). In the following,
we give a generalization of Lemma 1 to the probability of a
union indexed by K values. The proof is very similar to the
proof of Lemma
1, and thus omitted for brevity. For a given
subset J = j1 , . . . , k|J | of {1, . . . , J} we write Z J as a
shorthand for (Zj1 , . . . , Zj|J | ).
N
N
1
J
Lemma 2 Let
{Z1 (i)}i=1
, . . . , {ZJ (i)}i=1
and
NJ+1
NJ+1
NJ+1
be
{V1 (i)}i=1 , {V2 (i)}i=1 , . . . , {VK (i)}i=1
independent
sequences
of
independently
and
identically distributed (i.i.d.) RVs on the alphabets
Z1 × . . . × ZJ × V1 × . . . × VK , respectively, with Z1 (i) ∼
PZ1 , . . . , ZJ (i) ∼ PZJ , V1 (i) ∼ PV1 , . . . , VK (i) ∼ PVK . Fix
N
N
N
a sequence of sets {Ai,1 }i=1 , {Ai,2 }i=1 , . . . , {Ai,K }i=1 ,
where Ai,j ⊆ Z1 × . . . × ZJ × Vj , for 1 ≤ j ≤ K
N
and for all 1 ≤ i ≤ N . Also, fix a set {Ai,0 }i=1 where
Ai,0 ⊆ Z1 × . . . × ZJ for all 1 ≤ i ≤ N , and another
N
N
N
sequence of sets {Gi,1 }i=1 , {Gi,2 }i=1 , . . . , {Gi,K }i=1 , where
Gi,j ⊆ Vj , for 1 ≤ j ≤ K and for all 1 ≤ i ≤ N . Let
U = (Z1 , Z2 , . . . , ZJ , UJ+1 ) with UJ+1 , (V1 , . . . , VK ).
Finally, define Bl,J given in (110), for l = 1, 2, . . . , N , and
Z(iJ1 ) = (Z1 (i1 ), . . . , ZJ (iJ )). Then, a general upper bound
is given by (111), shown at the top of the next page.
Applying Lemma 2 on (107) (or, (109)) we obtain (112),
shown at the top of the next page, where N1 = M11 , N2 =
M12 , N3 = M21 , N4 = 1, and
U = (Z 11 , Z 12 , Z 21 , U 4 )
(113)
in which U 4 = (Z 4,1 , . . . , Z 4,M22 −1 ), and Bl,J is given in
(114), also shown at the top of the next page. The various
16
Bl,J ,
uJ : z1J ∈ Al,0 ,
Pr
i1 ,j
Z(iJ1 ) ∈ Al,0 ,
l=1
K
\
k=1
≤ min
1,
Pr
z1J , vj ∈ Al,j ,
j=1
(N (
[ [
J
K
\
i=1
j=1
·
≤ min
k=1
1,
Bl,J
vj ∈ Gl,j for some uJ c
k=1
min
Y
j∈J
Nj Pr
min
(N
[
P (Y |Z̃ ijk , Z 4,l ) ≥
U J ∈ Bl,J
J ⊆{1,...,4}J 6=∅
Y
j∈J
Nj Pr
(
[
U J ∈ Bl,J
l
)
l=1
,
(115)
namely, we have 15 possibilities. We claim that possibilities {1, 2, 3, 12, 13, 23, 123} do not affect the outer
minimum in (112), and so we left with possibilities
{4, 14, 24, 34, 124, 134, 234, 1234}. This observation follows
from the same arguments used in (86)-(91) for the second term
at the r.h.s. of (44). For example, possibilities {1, 2, 3} do not
affect the outer minimum due to the fact that the probabilities
that correspond to possibilities {14, 24, 34}, respectively, are
smaller. Indeed, the multiplicative factors in (112) for each of
the pairs (1, 14), (2, 24), and (3, 34), are the same, but the
respective probabilities in (112) are smaller for {14, 24, 34}
(due to the same reason used in (91)). Similarly, possibilities
{12, 13, 23, 123} do not affect the outer minimum due to
possibilities {124, 134, 234, 1234}, respectively.
In the following, we analyze the remaining terms. For
example, the term that corresponds to possibility “1234”, is
given by
(
)
[
Pe,1234 , M11 M12 M21 Pr
(116)
U ∈ Bl,1234 ,
l
.
(111)
F0
(112)
(z̃ 111 , z 4,0 , y) ∈ T (Q̃0X1 X2,0 Y ),
n
oM22 −1
uJ : (z̃ 111 , z 4,l , y) ∈ T (Q̃lZ 4 Y )
,
=
1
l=1
n
oM22 −1
(z̃ , z , y) ∈ T (Q̂j
, for some uJ c
0
4,l
X1,0 X2 Y )
)
P (Y |Z̃ 0 , Z 4,l )
possibilities for the set J are,
1; 2; 3; 4;
12; 13; 14; 23; 24; 34;
123; 124; 134; 234;
1234
(110)
))
)
l=0
,
l=1
MX
22 −1
l=0
j=1
K
\
Z(iJ1 ), Vk (j) ∈ Al,k ,
Vk (j) ∈ Gl,k
J ⊆{1,...,J+1}J 6=∅
(M −1
21 −1
22
12 −1 M[
11 −1 M[
M[
X
K
\
(114)
which similarly to the passage from (105) to (107), can be
rewritten as in (117), shown at the top of the next page.
Equation (117) has the same form of the probability in (71),
which we already analyzed. Accordingly, we similarly obtain
(120), where E7 (·, ·) is defined in (A.4), and S{4} (Q) is given
in (A.1).
The other terms are handled in a similar fashion. Specifically, let Ẑ , {Z1 , Z2 , Z3 }, and define
the sets U
= {1, 2, 3, 12, 13, 23, 123}, and Ũ
=
{14, 24, 34, 124, 134, 234, 1234}. Then, define for any5 u ∈
{1, 2, . . . , 7}:
(
)
[
(6)
Pe,u , MU (u) · Pr
U Ũ(u) ∈ Bl,Ũ (u) ,
(121)
l
where
MU (1) , M11 ; MU (2) = M12 ; MU (3) = M21 ;
MU (4) = M11 M12 ; MU (5) , M11 M21 ;
MU (6) = M12 M21 MU (7) = M11 M12 M21 .
(122)
Accordingly, following (49)-(84), we get (123), shown at the
top of the next page, where Eu (·, ·) is defined in (A.4). Note
that the mutual information term in the above exponent is due
to the averaging over Ẑ U (u) , and it is resulted by using the
method of types as in (72). This concludes the analysis for
5 Note
(6)
that Pe,7 correspond to Pe,1234 in (117).
17
Pe,1234 = M11 M12 M21 Pr
= M11 M12 M21 E
E
(
Pr
(M −1
22
X
P (Y |Z̃ 111 , Z 4,l ) ≥
l=0
(M −1
22
X
(
Pr
= exp −n
min
∈S{4} (QZ 4
1,0
n
o
(6)
4 Y , R22 )
, exp −nÊ7 (QZ1,0
(6) ·
=
Pe,u
(
exp −n
QZ 3 |Z
min
4,0 Y
1
∈S{4} (QZ 4
1,0 Y
o
4 Y , R22 )
, exp −nÊu(6) (QZ1,0
n
MX
22 −1
l=0
4,0 Y
)
Y
)
= exp −n
QZ 3 |Z
1
4,0 Y
min
∈S{4} (QZ 4
1,0
n
o
(6)
4 Y , R22 )
, exp −nÊ8 (QZ1,0
Y
)
P (Y |Z̃ 0 , Z 4,l ) F0 , Z̃ 111
)
F0
)
F0
)
)
(117)
(118)
)
(119)
(120)
)
h
i
4 Y )
IQ (Ẑ U (u) ; Z4,0 , Y |Ẑ 123\U (u) ) + Eu (QZ13 Z4,0 Y , QZ1,0
(123)
l
·
P (Y |Z̃ 0 , Z 4,l ) F0 , Z̃ 111
h
i
4 Y )
IQ (Z13 ; Z4,0 , Y ) + E7 (QZ13 Z4,0 Y , QZ1,0
possibilities {14, 24, 34, 124, 134, 234, 1234}, and we left with
possibility {4}, which is very similar to (75). Accordingly,
using the same arguments in (78)-(84), we obtain
(
)
[
(6)
Pe,8 , Pr
U 4 ∈ Bl,4
(
)
l=0
P (Y |Z̃ 111 , Z 4,l ) ≥
1
P (Y |Z̃ 0 , Z 4,l ) F0
P (Y |Z̃ 111 , Z 4,l ) ≥
l=0
QZ 3 |Z
MX
22 −1
l=0
MX
22 −1
l=0
(
·
( M −1
22
X
)
4 Y )
E0 (QZ13 Z4,0 Y , QZ1,0
(124)
where E0 (·, ·) is, again, defined in (A.4). Wrapping up, using
(100), (112), and the last results, after averaging w.r.t. F0 , we
get (129), shown at the top of the next page, where Ru for
u = 1, 2, . . . , 7 is defined in (A.6).
This concludes the analysis of the error event (m̂11 6=
0, m̂12 6= 0, m̂21 6= 0) in (98). The other types of errors
are analyzed in a similar manner. For (m̂11 6= 0, m̂12 =
0, m̂21 = 0), the average probability of error, associated
with the decoder in (14) is given in (131), given at the
top of the next page, where F0 , (Z̃ 0 , Z 22,0 , Y ). Thus,
due to the fact that (Z 12,0 , Z 21,0 ) are now fixed, they play
a same role as Y and Z 22,0 . Accordingly, following the
same steps as in (78)-(84), we get (132), presented at the
top of the next page, where E6 (·, ·) and E7 (·, ·) are defined
in (A.4), and S{2,3,4} (Q) is given in (A.1). Again, since
(Z 12,0 , Z 21,0 ), which correspond to (Z2,0 , Z3,0 )), are fixed,
(2)
they are conjugated to (Z4,0 , Y ). The error exponent of Pe
which corresponds to (m̂11 = 0, m̂12 6= 0, m̂21 = 0) can
be derived in the same way. We get that the exponent of
(2)
Pe is obtained by replacing the role of Z1 with Z2 and
(3)
R1 with R2 , in (132)-(134). Similarly, Pe , corresponding
to (m̂11 6= 0, m̂12 6= 0, m̂21 = 0), is upper bounded by
(136), presented at the top of the next page, where S{3,4} (Q)
is defined in (A.1).
(4)
(5)
Finally, the error exponents of Pe and Pe , corresponding
to (m̂11 6= 0, m̂12 = 0, m̂21 6= 0) and (m̂11 = 0, m̂12 6=
0, m̂21 6= 0), respectively, are obtained in the same way. The
(4)
exponent of Pe is obtained by replacing the role of Z2 with
Z3 , and changing the minimization in (137) to over the indexes
(5)
{1, 3, 5}, and the exponent of Pe is obtained by replacing the
role of Z1 with Z3 , and changing the minimization in (137)
to over the indexes {2, 3, 6}.
A PPENDIX A
D EFINITIONS FOR T HEOREM 2
In this appendix, we give the definitions of the various
parameters appearing in Theorem 2. For simplicity of notation,
in the following, we use the indexes {1, 2, 3, 4} instead of
{11, 12, 21, 22}, respectively. Let Z , (Z1 , Z2 , Z3 ), and
U , {1, 2, 3, 12, 13, 23, 123}. For u ∈ {0, 1, 2, . . . , 7}, Z U (u)
is a random vector consisting of the RVs corresponding to the
indexes in U(u), for example, Z 1 , Z U (1) = Z1 , Z 12 ,
Z U (4) = (Z1 , Z2 ), Z 123 , Z U (7) = (Z1 , Z2 , Z3 ), and so on,
where we define Z U (0) = ∅. Let also Z̃ , {Z1 , Z2 , Z3 , Z4 },
I ⊆ {1, 2, 3, 4}, and Z̃ I be the restriction of the entries of Z̃
on the set I. Then, let
o
n
SI (Q) , Q̃ : Q̃z̃I Y1 = Qz̃ I Y1 , Q̃Zi = PZi , for i ∈ I c .
(A.1)
18
Pe(6)
·
(
(
≤ E min 1, min e
=E
min
(6)
−n Êu
(QZ 4
(6)
−n Êu
(QZ 4
(
= exp −n
(6)
4 Y
EHK (QZ1,0
) , max
Pe(1) = Pr
=E
,R22 )−n−1 log MŨ (u)
"
,R22 )−n−1 log MŨ (u)
1,0 Y
= E exp −n max max
·
Y
u∈{1:7}
min e
u∈{1:7}
"
(
(
1,0
u
4 Y
Êu(6) (QZ1,0
(6)
,e
−nÊ8 (QZ 4
1,0 Y
(6)
+
,e
−nÊ8 (QZ 4
1
, R22 ) − log MŨ(u)
n
1,0
+
))
(125)
(126)
,R22 )
,R22 )
Y
(6)
4 Y
, Ê8 (QZ1,0
)#)
, R22 )
#)
h
i
(6)
4 ||WY |Z 4 |PZ 4 ) + E
4 Y )
min D(QY |Z1,0
HK (QZ1,0
1,0
1,0
(128)
QY |Z 4
1,0
i
h
(6)
(6)
4 Y , R22 )
4 Y , R22 ) − Ru
, Ê8 (QZ1,0
max Êu (QZ1,0
u∈{1:7}
P (Y |Z̃ i00 , Z 22,l ) ≥
i=1
l=0
"M −1 (M −1
22
11
X
[
Pr
i=1
(
≤ exp −n
"
MX
22 −1
)#
P (Y |Z̃ 0 , Z 22,l )
l=0
MX
22 −1
P (Y |Z̃ i00 , Z 22,l ) ≥
l=0
·
Pe(1)
(129)
+
"M −1 (M −1
22
11
X
[
(
min
QY |Z 4
1,0
)
P (Y |Z̃ 0 , Z 22,l )
l=0
h
(130)
4 ||WY |Z 4 |PZ 4 ) +
D(QY |Z1,0
1,0
1,0
F0
(1)
4 Y
EHK (QZ1,0
#)
i
)
#)
h
i
(1)
(1)
4 Y , R22 )
4 Y , R22 ) − R1
) , max Ê (QZ1,0
, Ê8 (QZ1,0
+
h
i
4
4 Y , R22 ) =
4 Y , QZ 4 Y )
Ê (1) (QZ1,0
min
IQ (Z1 ; Z2,0
, Y ) + E7 (QZ1 Z2,0
1,0
(1)
4 Y
EHK (QZ1,0
QZ
∈S{2,3,4} (QZ 4
QZ
∈S{2,3,4} (QZ 4
4
1 |Z2,0 Y
1,0 Y
(1)
4 Y , R22 ) =
Ê8 (QZ1,0
Pe(3)
min
4
1 |Z2,0 Y
(
·
≤ exp −n
1,0 Y
"
)
)
4 Y , QZ 4 Y )
E6 (QZ1 Z2,0
1,0
#)
h
i
(3)
4 ||WY |Z 4 |PZ 4 ) + E
4 Y )
min D(QY |Z1,0
HK (QZ1,0
1,0
1,0
QY |Z 4
QZ 2 |Z 4
1
3,0
Y
(3)
4 Y , R22 ) =
Ê2 (QZ1,0
QZ 2 |Z 4
1
3,0 Y
(3)
4 Y , R22 ) =
Ê4 (QZ1,0
∈S{3,4} (QZ 4
1,0 Y
min
∈S{3,4} (QZ 4
3,0 Y
Y
min
QZ 2 |Z 4
1
f (QZ14 Y1 ) , EQ
1,0
∈S{3,4} (QZ 4
3,0 Y
(3)
4 Y , R22 ) =
Ê8 (QZ1,0
∈S{3,4} (QZ 4
QZ 2 |Z 4
1
(131)
(132)
(133)
(134)
(135)
(136)
1,0
i
h
(3)
(3)
4 Y , R22 )
4 Y , R22 ) − Ru
, Ê8 (QZ1,0
) , max
max Êu (QZ1,0
+
u∈{1,2,4}
h
i
(3)
4
4 Y , QZ 4 Y )
4 Y , R22 ) =
IQ (Z1 ; Z3,0
, Y |Z2 ) + E5 (QZ12 Z3,0
Ê1 (QZ1,0
min
1,0
(3)
4 Y
EHK (QZ1,0
Define
(127)
1,0 Y
min
1,0 Y
)
)
)
)
h
i
4
4 Y , QZ 4 Y )
IQ (Z2 ; Z3,0
, Y |Z1 ) + E6 (QZ12 Z3,0
1,0
h
i
4
4 Y , QZ 4 Y )
IQ (Z1 , Z2 ; Z3,0
, Y ) + E7 (QZ12 Z3,0
1,0
4 Y , QZ 4 Y )
E3 (QZ12 Z3,0
1,0
(137)
(138)
(139)
(140)
(141)
and let
f (Q̂) − IQ̂ (Z4 ; Z13 , Y1 ).
max
r0 (QZ13 Y1 ) , R22 +
Q̂:
Q̂∈S
(Q)
{1,2,3}
log WY1 |X1 X2 (Y1 |X1 (Z1 , Z2 ), X2 (Z3 , Z4 ))
IQ̂ (Z4 ;Z13 ,Y1 )≤R22
(A.2)
(A.3)
19
inequality follows from the fact that min {1, x} ≤ 1. The third
term in (38) follows in a similar fashion, and the forth term
follows from the fact that min {1, x} ≤ x, and thus we get
(B.5), which concludes the proof.
For u ∈ {0, 1, 2, . . . , 7}, define
Eu (Q̃Z14 Y1 , QZ14 Y1 ) ,
min
Q̂: Q̂∈S{1,2,3} (Q̃)
Q̂∈Du (Q̃Z 4 Y ,QZ 4 Y )
1
1
i
h
IQ̂ (Z4 ; Z U (u) , Y1 ) − R22 ,
+
1 1
(A.4)
where
A PPENDIX C
C OMPUTATIONAL A SPECTS OF
THE
E XPONENTS
In this appendix, we discuss the computation of (8j), similarly as in [21]. We start with an alternative formulation of
(8j). Recall that
h
i
Du (Q̃Z14 Y1 , QZ14 Y1 ) , Q̂ : max r0 (QZ14 Y1 ), f (QZ14 Y1 )
Ẽ1 (R1 , R2 ) = min {D(Q||W ) + E(Q, R1 , R2 )} , (C.1)
Q
h
i
where
≤ max f (Q̃Z14 Y1 ), f (Q̂) + R22 − IQ̂ (Z4 ; Z U (u) , Y1 )
h
+
i
(A.5)
E(Q, R1 , R2 ) = max Ê1 (Q, R2 ) − R1 , Ê2 (Q, R2 ) .
+
Finally, we let
(C.2)
R1 , R11 ; R2 , R12 ; R3 , R21 ; R4 , R11 + R12 ;
R5 , R11 + R21 ; R6 , R12 + R21 ;
R7 , R11 + R12 + R21 .
(A.6)
Using all the above definition, we define (A.7)-(A.30), shown
at the top of page 20.
A PPENDIX B
P ROOF OF L EMMA 1
We prove a generalized version of Lemma 1, where we conL2
L2
sider random sequences, {V2 (i)}i=1
, . . . , {VK (i)}i=1
, rather
than single RVs V2 , . . . , VK . Lemma 1 is then obtained on
substituting L2 = 1. We start with the following result which
can be thought of as an extension of [11, Lemma 2].
L2
L2
1
Lemma 3 Let
{V1 (i)}L
i=1 , {V2 (i)}i=1 , . . . , {VK (i)}i=1
be independent sequences of independently and
identically distributed (i.i.d.) RVs on the alphabets
V1 × V2 × . . . × VK , respectively, with V1 (i) ∼
PV1 , V2 (i) ∼ PV2 , . . . , VK (i) ∼ PVK . Fix a sequence
N
N
N
of sets {Ai,1 }i=1 , {Ai,2 }i=1 , . . . , {Ai,K−1 }i=1 , where
Ai,j ⊆ V1 × Vj+1 , for 1 ≤ j ≤ K − 1 and for all
N
1 ≤ i ≤ N . Also, fix a set {Ai,0 }i=1 where Ai,0 ⊆ V1
for all 1 ≤ i ≤ N , and another sequence of sets
N
N
N
{Gi,2 }i=1 , {Gi,3 }i=1 , . . . , {Gi,K }i=1 , where Gi,j ⊆ Vj ,
for 2 ≤ j ≤ K and for all 1 ≤ i ≤ N . We have (B.1), shown
at the top of page 21, with (V1 , . . . , VK ) ∼ PV1 · · · × PVK .
Proof of Lemma 3: The second term in (B.1) follows by
first applying the union bound over i as in (B.2), shown at the
top of page 21, and then we apply the truncated union bound to
the union over j, and obtain (B.3). The third term is obtained
similarly by applying the union bounds in the opposite order,
and the upper bound of 1 is trivial.
We are now in a position to prove Lemma 1.
Proof of Lemma
S 1: To obtain (38) we weaken (B.1) as
follows. Let F , N
l=1 {V1 ∈ Bl,1 }. The second term in (38)
follows from (B.4), shown at the top of page 21, where the
second equality follows from T
the fact that the inner term in
N
the expectation vanishes over l=1 {V1 ∈
/ Bl,1 }, and the third
In the following, for a given QY1 |X1 X2 , we show that
Ê1 (Q, R2 ) and Ê2 (Q, R2 ) can be calculated efficiently. For
˜ Q̃) ≡ I (X1 ; X2 , Y1 ) and I(
ˆ Q̂) ≡
brevity, we let I(
Q̃
IQ̂ (X2 ; X1 , Y1 ). Recall that
i
h
˜ Q̃) + I(
ˆ Q̂) − R2
,
I(
Ê1 (Q, R2 ) =
min
+
Q̃∈S(Q),Q̂∈Ŝ(Q̃),
Q̂∈L(Q̃,Q)
Ê2 (Q, R2 ) =
min
Q̃∈S(Q),Q̂∈Ŝ(Q̃),
Q̂∈L̂(Q̃,Q)
(C.3)
i
h
IˆQ̂ (X2 ; Y1 ) − R2 , (C.4)
+
where L and L̂ are defined in (8e) and (8f), respectively,
S(Q) = {Q̃ : Q̃X1 = PX1 , Q̃X2 Y1 = QX2 Y1 }, and
Ŝ(Q̃) = {Q̂ : Q̂X2 = PX2 , Q̂X1 Y1 = Q̃X1 Y1 }. In [21],
it was shown that Ê1 (Q, R2 ) can be equivalently expressed in
terms of the minimum between the following terms:
Ê1′ (Q, R2 ) ,
Ê1′′ (Q, R2 ) ,
Ê1′′′ (Q, R2 )
,
min
˜ Q̃),
I(
(C.5)
Q̃∈S(Q),max[t0 (Q),f (Q)]≤f (Q̃)
min
ˆ Q̂)≤R2
Q̃∈S(Q),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
max[t0 (Q),f (Q)]≤f (Q̂)+R2 −I(
min
Q̃∈S(Q),Q̂∈Ŝ(Q̃)
t0 (Q),f (Q)≤f (Q̂)
˜ Q̃),
I(
(C.6)
i
h
˜
ˆ
.
I(Q̃) + I(Q̂) − R2
+
(C.7)
Using the same arguments as in [21], it can be shown that
Ê2 (Q, R2 ) can be equivalently be expressed as
i
h
IQ̂ (X2 ; Y1 ) − R2 .
Ê2′′ (Q, R2 ) ,
min
Q̃∈S(Q),Q̂∈Ŝ(Q̃)
max[t0 (Q),f (Q)]≤f (Q̂)
+
(C.8)
Accordingly, from (C.5)-(C.7) and (C.8), we see that (C.3)
and (C.4) can be expressed in terms of convex optimization
problems, namely, for a given QY1 |X1 X2 , the terms Ê1 (Q, R2 )
and Ê2 (Q, R2 ) (i.e., the inner terms of the minimization
problem in (C.1)) can be calculated efficiently, as desired.
Finally, we discuss the computation of (C.1). Generally
speaking, the minimization over QY1 |X1 X2 might not be a
20
Ê (1) (QZ14 Y1 , R22 ) ,
Ê (2) (QZ14 Y1 , R22 ) ,
(1)
Ê8 (QZ14 Y1 , R22 ) ,
(2)
Ê8 (QZ14 Y1 , R22 ) ,
(3)
Ê1 (QZ14 Y1 , R22 ) ,
(3)
Ê2 (QZ14 Y1 , R22 ) ,
(3)
Ê4 (QZ14 Y1 , R22 ) ,
(3)
Ê8 (QZ14 Y1 , R22 ) ,
(4)
Ê1 (QZ14 Y1 , R22 ) ,
(4)
Ê3 (QZ14 Y1 , R22 ) ,
(4)
Ê5 (QZ14 Y1 , R22 ) ,
(4)
Ê8 (QZ14 Y1 , R22 ) ,
(5)
Ê2 (QZ14 Y1 , R22 ) ,
(5)
Ê3 (QZ14 Y1 , R22 ) ,
(5)
Ê6 (QZ14 Y1 , R22 ) ,
(5)
Ê8 (QZ14 Y1 , R22 ) ,
Êu(6) (QZ14 Y1 , R22 ) ,
(6)
Ê8 (QZ14 Y1 , R22 ) ,
min
Q̃: Q̃∈S{2,3,4} (Q)
min
Q̃: Q̃∈S{1,3,4} (Q)
min
Q̃: Q̃∈S{2,3,4} (Q)
min
Q̃: Q̃∈S{1,3,4} (Q)
min
Q̃: Q̃∈S{3,4} (Q)
min
Q̃: Q̃∈S{3,4} (Q)
min
Q̃: Q̃∈S{3,4} (Q)
min
Q̃: Q̃∈S{3,4} (Q)
min
Q̃: Q̃∈S{2,4} (Q)
min
Q̃: Q̃∈S{2,4} (Q)
min
Q̃: Q̃∈S{2,4} (Q)
min
Q̃: Q̃∈S{2,4} (Q)
min
Q̃: Q̃∈S{1,4} (Q)
min
Q̃: Q̃∈S{1,4} (Q)
min
Q̃: Q̃∈S{1,4} (Q)
min
Q̃: Q̃∈S{1,4} (Q)
min
Q̃: Q̃∈S{4} (Q)
min
Q̃: Q̃∈S{4} (Q)
h
i
IQ̃ (Z1 ; Z24 , Y1 ) + E7 (Q̃Z14 Y1 , QZ14 Y1 ) ,
h
i
IQ̃ (Z2 ; Z1 , Z34 , Y1 ) + E7 (Q̃Z14 Y1 , QZ14 Y1 ) ,
(A.7)
(A.8)
E6 (Q̃Z14 Y , QZ14 Y1 ),
(A.9)
E5 (Q̃Z14 Y , QZ14 Y1 ),
(A.10)
h
i
IQ̃ (Z U (1) ; Z34 , Y1 |Z 12\U (1) ) + E5 (Q̃Z14 Y1 , QZ14 Y ) ,
h
i
IQ̃ (Z U (2) ; Z34 , Y1 |Z 12\U (2) ) + E6 (Q̃Z14 Y1 , QZ14 Y ) ,
h
i
IQ̃ (Z U (4) ; Z34 , Y1 |Z 12\U (4) ) + E7 (Q̃Z14 Y1 , QZ14 Y ) ,
(A.11)
(A.12)
(A.13)
E3 (Q̃Z14 Y , QZ14 Y1 ),
(A.14)
h
i
IQ̃ (Z U (1) ; Z2 , Z4 , Y1 |Z 13\U (1) ) + E4 (Q̃Z14 Y1 , QZ14 Y1 ) ,
(A.15)
h
i
IQ̃ (Z U (3) ; Z2 , Z4 , Y1 |Z 13\U (3) ) + E6 (Q̃Z14 Y1 , QZ14 Y1 ) ,
h
i
IQ̃ (Z U (5) ; Z2 , Z4 , Y1 |Z 13\U (5) ) + E7 (Q̃Z14 Y1 , QZ14 Y1 ) ,
(A.16)
(A.17)
E2 (Q̃Z14 Y1 , QZ14 Y1 ),
(A.18)
h
i
IQ̃ (Z U (2) ; Z1 , Z4 , Y1 |Z 23\U (2) ) + E4 (Q̃Z14 Y1 , QZ14 Y1 ) ,
(A.19)
h
i
IQ̃ (Z U (3) ; Z1 , Z4 , Y1 |Z 23\U (3) ) + E5 (Q̃Z14 Y1 , QZ14 Y1 ) ,
h
i
IQ̃ (Z U (6) ; Z1 , Z4 , Y1 |Z 23\U (6) ) + E7 (Q̃Z14 Y1 , QZ14 Y1 ) ,
E1 (Q̃Z14 Y1 , QZ14 Y1 ),
h
i
IQ̃ (Z U (u) ; Z4 , Y1 |Z 123\U (u) ) + Eu (Q̃Z14 Y1 , QZ14 Y1 ) , u ∈ {1, . . . , 7} ,
E0 (Q̃Z14 Y1 , QZ14 Y1 ),
h
i
(u)
Ê (u) (QZ14 Y1 , R22 ) − Ru , Ê8 (QZ14 Y1 , R22 ) , u ∈ {1, 2} ,
+
i
h
(3)
(3)
(3)
max Êu (QZ14 Y1 , R22 ) − Ru , Ê8 (QZ14 Y1 , R22 ) ,
EHK (QZ14 Y1 ) , max
+
u∈{1,2,4}
i
h
(4)
(4)
(4)
max Êu (QZ14 Y1 , R22 ) − Ru Ê8 (QZ14 Y1 , R22 ) ,
EHK (QZ14 Y1 ) , max
u∈{1,3,5}
,
i
h
(5)
(5)
max Êu(5) (QZ14 Y1 , R22 ) − Ru , Ê8 (QZ14 Y1 , R22 ) ,
EHK (QZ14 Y1 ) , max
+
u∈{2,3,6}
i
h
(6)
(6)
EHK (QZ14 Y1 ) , max max Êu(6) (QZ14 Y1 , R22 ) − Ru , Ê8 (QZ14 Y1 , R22 ) ,
u∈{1:7}
+
(u)
D(QY1 |Z14 ||WY1 |Z14 |PZ14 ) + min EHK (QZ14 Y1 ) .
ẼHK (R11 , R12 , R21 , R22 ) ,
min
(u)
EHK (QZ14 Y1 ) , max
QY
4:
1 |Z1
u∈{1:6}
(A.20)
(A.21)
(A.22)
(A.23)
(A.24)
(A.25)
(A.26)
(A.27)
(A.28)
(A.29)
(A.30)
QZi =PZi , 1≤i≤4
convex problem, and thus one should resort to global opti-
mization methods (e.g., a simple algorithm is an exhaustive
21
Pr
( (
N
[ [
i,j
K−1
\
V1 (i) ∈ Al,0 ,
l=1
(
(V1 (i), Vk+1 (j)) ∈ Al,k ,
k=1
"
(
≤ min 1, L1 E min 1, L2 Pr
N
[
K
\
Vk (j) ∈ Gl,k
k=2
N
[
(
l=1
(
V1 ∈ Al,0 ,
K−1
\
K
\
(V1 , Vk+1 ) ∈ Al,k ,
k=1
K−1
\
))
Vk ∈ Gl,k
k=2
K
\
)
)
V1
))#
,
))#)
(B.1)
Vk (j) ∈ Gl,k
j
l=1
k=1
k=2
( (
))
N
K−1
K
[ [
\
\
≤ L1 E Pr
V1
V1 ∈ Al,0 ,
(V1 , Vk+1 (j)) ∈ Al,k ,
Vk (j) ∈ Gl,k
(B.2)
"
(
L2 E min 1, L1 Pr
Pr
( (
N
[ [
i,j
≤ L1 Pr
(
l=1
(
V1 ∈ Al,0 ,
(V1 , Vk+1 ) ∈ Al,k ,
V1 (i) ∈ Al,0 ,
l=1
K−1
\
K
\
(V1 (i), Vk+1 (j)) ∈ Al,k ,
k=1
( (
N
[ [
j
Vk (j) ∈ Gl,k
l=1
(V1 , Vk+1 (j)) ∈ Al,k ,
K
\
k=1
))
))
k=2
K−1
\
V1 ∈ Al,0 ,
Vk ∈ Gl,k
k=2
k=1
{Vk }K
k=2
k=2
( (
))
N
K−1
K
[ [
\
\
Vk (j) ∈ Gl,k
V1 (i) ∈ Al,0 ,
(V1 (i), Vk+1 (j)) ∈ Al,k ,
Pr
i,j
l=1
k=1
k=2
)#
"
(
(N (
))
K−1
K
[
\
\
≤ L1 E min 1, L2 Pr
V1
V1 ∈ Al,0 ,
(V1 , Vk+1 ) ∈ Al,k ,
Vk ∈ Gl,k
l=1
(
min 1, L2 Pr
(
(
N
[
l=1
(
(
+ I {F } min 1, L2 Pr
= I {F } min 1, L2 Pr
(
L1 E min 1, L2 Pr
≤ L1 L2 Pr
(N (
[
l=1
(V1 , Vk+1 ) ∈ Al,k ,
(
(
(
l=1
(V1 , Vk+1 ) ∈ Al,k ,
V1 ∈ Al,0 ,
(V1 , Vk+1 ) ∈ Al,k ,
search over a fine grid of probability simplex). Nonetheless,
in the following we somewhat simplify these optimizations.
Vk ∈ Gl,k
k=2
(V1 , Vk+1 ) ∈ Al,k ,
k=1
Vk ∈ Gl,k
K
\
K
\
k=2
Vk ∈ Gl,k
)
Vk ∈ Gl,k
k=2
K
\
k=2
k=1
K−1
\
V1
))
Vk ∈ Gl,k
(V1 , Vk+1 ) ∈ Al,k ,
k=1
K−1
\
K−1
\
K
\
)
k=2
K
\
k=1
(
V1 ∈ Al,0 ,
Vk ∈ Gl,k
(V1 , Vk+1 ) ∈ Al,k ,
V1 ∈ Al,0 ,
V1 ∈ Al,0 ,
N
[
K−1
\
k=1
K−1
\
l=1
( N (
[
K
\
k=2
V1 ∈ Al,0 ,
l=1
(N
[
l=1
≤ I {F }
"
N
[
(B.3)
k=2
k=1
(
(
K−1
\
V1 ∈ Al,0 ,
= I {F } min 1, L2 Pr
c
k=1
V1
)
)
))
V1
V1
))
))
(B.4)
)
V1
))#
))
We first see that (C.1) can be rewritten as
n
o
Ẽ1 (R1 , R2 ) , min Ẽ1′ (R1 , R2 ), Ẽ1′′ (R1 , R2 )
(B.5)
(C.9)
22
where
Ẽ1′ (R1 , R2 )
and
=
inf
Q: R1 <Ê1 (Q,R2 )−Ê2 (Q,R2 )
o
n
D(Q||W ) + Ê1 (Q, R2 ) − R1 ,
Ẽ1′′ (R1 , R2 ) =
inf
n
o
D(Q||W ) + Ê2 (Q, R2 ) .
(C.10)
Obviously, we should get that R < IW (X; Y ), namely,
the minimum in (D.4) should be equal to IW (X; Y ). To
see that this is indeed the case, note that since the above
optimization problem is convex, the linear constraint is met
with equality, and we note that f (W ) = −HW (Y |X),
f (Q) = −D(Q||W |PX ) − HQ (Y |X), and IQ (X; Y ) =
HW (Y ) − HQ (Y |X). Using the last facts, we get
R<
Q: R1 ≥Ê1 (Q,R2 )−Ê2 (Q,R2 )
(C.11)
=
Let us analyze Ẽ1′ (R1 , R2 ) as a function of R1 . For R1 = 0,
we have
n
o
(C.12)
Ẽ1′ (0, R2 ) = inf D(Q||W ) + Ê1 (Q, R2 ) .
min
Q:QY =WY ,QX =PX ,f (W )=f (Q)
min
Q:QY =WY ,QX =PX ,
HW (Y |X)=D(Q||W |PX )+HQ (Y |X)
= IW (X; Y )
+
R1 =0
ing the critical rate R1,crit = Ê1 ( Q∗Y1 |X1 X2
R1 =0
, R2 ), it is easily noticed that for R1 ≤
Ê2 ( Q∗Y1 |X1 X2
R1 =0
R1,crit , the exponent is an affine function
Ẽ1′ (R1 , R2 ) = Ẽ1′ (0, R2 ) − R1 .
D(Q||W |PX )
as required.
In our case, using the equivalent representation of our error
exponent in (C.5)-(C.7), we readily get that the error exponent
in Theorem 1 is positive if6 :
R1 <
(C.13)
R1 <
Furthermore, for R1 ≤ R1,crit , it is readily seen that
Ẽ1 (R1 , R2 ) = Ẽ1′ (R1 , R2 ).
min
= IW (X; Y ),
, and defin, R2 ) −
HW (Y ) − HQ (Y |X)
Q:QY =WY ,QX =PX ,
HW (Y |X)=D(Q||W |PX )+HQ (Y |X)
Q
Now, letting the minimizer be Q∗Y1 |X1 X2
IQ (X; Y )
(C.14)
R1 <
For R1 > R1,crit , however, since the optimization of Q is not
convex, the term cannot be simplified anymore.
˜ Q̃),
I(
min
(D.5)
Q̃∈S(W ),max[t0 (W ),f (W )]≤f (Q̃)
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
max[t0 (W ),f (W )]≤f (Q̂)+R2 −I(
min
Q̃∈S(W ),Q̂∈Ŝ(Q̃)
max[t0 (W ),f (W )]≤f (Q̂)
˜ Q̃),
I(
(D.6)
i
h
˜
ˆ
I(Q̃) + I(Q̂) − R2
+
(D.7)
A PPENDIX D
P ROOF OF (10)
First, note that the following region:
R1 < IW (X1 ; Y1 |X2 ),
(D.1)
R1 < IW (X1 ; Y1 ) + [IW (X2 ; Y1 |X1 ) − R2 ]+ ,
(D.2)
evaluated with PX1 X2 Y1 = PX1 × PX2 × WY1 |X1 X2 , is
equivalent to (10). Thus, to show that (10) is achievable, it
suffices to show that the above region is achievable. Now,
recall that the ordinary random coding exponent, in the singleuser setting, is given by
o
n
Er (R) = min D(QY |X ||W |PX ) + [IQ (X; Y ) − R]+ .
Q
From [25, Lemma 9] it can be shown that Er (R) can be
rewritten as
n
D(Q̃Y |X ||W |PX )
Er (R) =
min
(Q,Q̃)∈D:f (Q̃)≤f (Q)
o
+ [IQ (X; Y ) − R]+ , (D.3)
where f (Q) = EQ {log W (Y |X)}, and D is the set of (Q, Q̃)
distributions such that QY = Q̃Y , and QX = Q̃X = PX . The
last representation is very similar, in some sense, to the error
exponent formula in Theorem 1. It can be seen that Er (R) is
positive as long as
R<
min
Q:QY =WY ,QX =PX ,f (W )≤f (Q)
IQ (X; Y ).
(D.4)
˜ Q̃) ≡ I (X1 ; X2 , Y1 ) and I(
ˆ Q̂) ≡
where we recall that I(
Q̃
IQ̂ (X2 ; X1 , Y1 ). In the following, we show that (D.5) and
(D.7) correspond to (D.1) and (D.2), respectively. Finally, we
show that (D.5) is dominated by (D.5) and (D.7), and thus
superfluous. Indeed, for (D.5), we have
˜ Q̃)
I(
min
Q̃∈S(W ),max[t0 (W ),f (W )]≤f (Q̃)
=
min
Q̃Y |X2 =WY |X2 ,max[t0 (W ),f (W )]≤f (Q̃)
≥
min
Q̃Y |X2 =WY |X2 ,f (W )≤f (Q̃)
IQ̃ (X1 ; X2 , Y )
IQ̃ (X1 ; X2 , Y ).
(D.8)
Now, as before, we note that (using the fact that the minimization over Q̃ in (D.8) is such that Q̃Y |X2 = WY |X2 )
IQ̃ (X1 ; X2 , Y ) = IQ̃ (X1 ; X2 ) + IQ̃ (X1 ; Y |X2 )
(D.9)
= IQ̃ (X1 ; X2 ) + HW (Y |X2 )
− HQ̃ (Y |X1 , X2 ),
(D.10)
f (W ) = −HW (Y |X1 , X2 ),
(D.11)
f (Q̃) = −D(Q̃||W |PX ) − HQ̃ (Y |X1 , X2 ).
(D.12)
and
Thus, we have (D.13), shown at the top of the next page, where
6 To show that (10) is achievable, we consider a lower bound
on Ẽ(R
the contribution of Ê2 (Q, R2 ), namely,
1 , R2 ), which ignores
h
i
minQ D(Q||W |PX ) + Ê1 (Q, R2 ) − R1
.
+
23
min
Q̃Y |X2 =WY |X2 ,f (W )≤f (Q̃)
=
IQ̃ (X1 ; X2 , Y )
min
Q̃Y |X2 =WY |X2 ,
D(Q̃||W |PX )+HQ̃ (Y |X1 ,X2 )≤HW (Y |X1 ,X2 )
≥ IW (X1 ; Y |X2 ) +
IQ̃ (X1 ; X2 ) + HW (Y |X2 ) − HQ̃ (Y |X1 , X2 )
min
Q̃Y |X2 =WY |X2 ,
D(Q̃||W |PX )+HQ̃ (Y |X1 ,X2 )≤HW (Y |X1 ,X2 )
IQ̃ (X1 ; X2 ) + D(Q̃||W |PX )
= IW (X1 ; Y |X2 )
the inequality follows from the fact that −HQ̃ (Y |X1 , X2 ) ≥
D(Q̃||W |PX ) − HW (Y |X1 , X2 ) induced by the optimization
constraint, and the last equality is achieved by taking Q̃ = W
and Q̃X1 ,X2 = PX1 PX2 . The constraint in (D.7) is handled
in a similar manner. Indeed, using the same manipulations,
we get (D.14), shown at the top of the next page, where
the inequality is due to the fact that −HQ̃ (Y |X1 , X2 ) ≥
D(Q̃||W |PX ) − HW (Y |X1 , X2 ), the second equality follows
by taking Q̂ = W , and the last equality follows by taking Q̃
such that X2 − Y − X1 is a Markov chain.
Finally, we show that the constraint in (D.6) is superfluous.
To this end, we will show that for R2 < IW (X2 ; Y |X1 ),
the r.h.s. of (D.6) reduces to R1 + R2 < I(X1 , X2 ; Y ),
which is dominated by (D.5) and (D.7) (or, equivalently, by
(D.13) and (D.14)), and for R2 ≥ IW (X2 ; Y |X1 ), (D.6)
reduces to R1 < IW (X1 ; Y ), already supported by (D.7)
(see (D.14)). Whence, (D.6) is redundant. Indeed, for R2 ≥
IW (X2 ; Y |X1 ), the r.h.s. of (D.6) can be lower bounded
as in (D.15), presented at the top of the next page, where
in (a) we use the definition of t0 (W ) in (8b), (b) follows
from the assumption that R2 ≥ IW (X2 ; Y |X1 ), (c) is due to
(D.11)-(D.12), (d) follows from the fact that −HQ̃ (Y |X1 ) ≥
D(Q̂||W |PX ) + IQ̂ (X1 ; X2 ) − HW (Y |X1 ) induced by the
optimization constraint, and (e) is achieved by taking Q̂ = W ,
Q̂X1 X2 = Q̂X1 Q̂X2 , and Q̃ such that X2 − Y − X1 is a
Markov chain. Thus, for R2 ≥ IW (X2 ; Y |X1 ), we obtained
that R1 < IW (X1 ; Y ), as required. On the other hand,
for R2 < IW (X2 ; Y |X1 ), the r.h.s. of (D.6) can be lower
bounded as shown in (D.16), given in the next page, where
(a) is due to (D.11)-(D.12), (b) is because Q̂ ∈ S(Q̃) and
thus HQ̃ (Y |X1 ) = HQ̂ (Y |X1 ), and (c) follows from the
fact that −HQ̂ (Y |X1 ) ≥ D(Q̂||W |PX ) + IQ̂ (X1 ; X2 ) −
HW (Y |X1 , X2 ) − R2 induced by the optimization constraint.
Whence, we obtained that R1 + R2 ≤ IW (X1 , X2 ; Y ), as
required.
ACKNOWLEDGMENT
The authors would like to thank the associate editor, Jun
Chen, and the anonymous referees for their suggestions and
comments which helped improving the content of this paper.
The authors are also grateful to Jonathan Scarlett for very
helpful comments on a previous version of this paper, and for
drawing our attention to [21] (and its extended version which
(D.13)
is not published yet), which was unknown to the authors at
the time of writing the original version of this paper.
R EFERENCES
[1] A. El Gamal and Y. H. Kim, Network Information Theory. Cambridge
University Press, 2011.
[2] E. C. Shannon, “Two-way communication channels,” in Proc. 4th
Berkeley Symp. on Mathematical Statistics and Probability, vol. 1.
Berkeley, CA: Univ. California Press, 1961, pp. 611–64.
[3] R. Ahlswede, “The capacity region of a channel with two senders and
two receivers,” Annals Probabil., vol. 2, no. 5, pp. 805–814, 1974.
[4] A. B. Carleial, “Interference channels,” IEEE Trans. on Inf. Theory, vol.
IT-24, pp. 60–70, Jan. 1978.
[5] H. Sato, “Two-user communication channels,” IEEE Trans. on Inf.
Theory, vol. IT-23, pp. 295–304, May. 1977.
[6] A. B. Carleial, “A case where intereference does not reduce capacity,”
IEEE Trans. on Inf. Theory, vol. IT-21, pp. 569–570, Sep. 1975.
[7] R. Benzel, “The capacity region of a class of discrete additive degraded
interference channels,” IEEE Trans. on Inf. Theory, vol. IT-25, pp. 228–
231, Mar. 1979.
[8] T. S. Han and K. Kobayashi, “A new achievable rate region for the
interference channel,” IEEE Trans. on Inf. Theory, vol. IT, no. 27, pp.
49–60, Jan. 1981.
[9] R. Etkin, N. Merhav, and E. Ordentlich, “Error exponents of optimum
decoding for the interference channel,” IEEE Trans. on Inf. Theory,
vol. 56, no. 1, pp. 40–56, Jan. 2010.
[10] C. Chang, R. Etkin, and E. Ordentlich, “Interference
channel capacity region for randomized fixed-composition
codes,” HP Labs Technical Report, 2009. [Online]. Available:
http://www.hpl.hp.com/techreports/2008/HPL-2008-194R1.html
[11] J. Scarlett, A. Martinez, and A. G. i Fábregas, “Multiuser coding
techniques for mismatched decoding,” submitted to IEEE Trans. on Inf.
Theory, Nov. 2013. [Online]. Available: arxiv.org/pdf/1311.6635
[12] R. G. Gallager, Information Theory and Reliable Communication. New
York: Wiley, 1968.
[13] I. Csiszár and J. Korner, Information Theory: Coding Theorems for
Discrete Memoryless Systems. Academic Press, 2011.
[14] N. Merhav, “Relations between random coding exponents and the
statistical physics of random codes,” IEEE Trans. on Inf. Theory, vol. 55,
no. 1, pp. 83–92, Jan. 2009.
[15] ——, “Exact random coding exponents of optimal bin index decoding,”
IEEE Trans. on Inf. Theory, vol. 60, no. 10, pp. 6024–6031, Oct. 2014.
[16] ——, “List decoding-random coding exponents and expurgated exponents,” IEEE Trans. on Inf. Theory, vol. 60, no. 11, pp. 6749–6759,
Nov. 2014.
[17] ——, “Exact correct-decoding exponent for the wiretap channel decoder,” IEEE Trans. on Inf. Theory, vol. 60, no. 12, pp. 7606–7615,
Dec. 2014.
[18] W. Huleihel, N. Weinberger, and N. Merhav, “Erasure/list random
coding error exponents are not universally achievable,” submitted
to IEEE Trans. on Inf. Theory, Dec. 2015. [Online]. Available:
http://arxiv.org/abs/1410.7005
[19] N. Merhav, “The generalized random energy model and its application to
the statistical physics of ensembles of hierarchical codes,” IEEE Trans.
on Inf. Theory, vol. 55, no. 3, pp. 1250–1268, May 2009.
[20] N. Shulman, “Communication over an unknown channel via common broadcasting,” Ph.D. dissertation, Tel-Aviv University, 2003,
http://citeseerx.ist.psu.edu/viewdoc.
24
min
Q̃Y |X2 =WY |X2 ,Q̂Y |X1 =Q̃Y |X1
D(Q̂||W |PX )+HQ̂ (Y |X1 ,X2 )≤HW (Y |X1 ,X2 )
≥
i
h
˜ Q̃) + I (X1 ; X2 ) + H (Y |X1 ) − H (Y |X1 , X2 ) − R2
I(
Q̃
Q̂
Q̂
min
Q̃Y |X2 =WY |X2 ,Q̂Y |X1 =Q̃Y |X1
D(Q̂||W |PX )+HQ̂ (Y |X1 ,X2 )≤HW (Y |X1 ,X2 )
+
n
h
˜ Q̃) + I (X1 ; X2 ) + D(Q̂||W |PX ) + H (Y |X1 )
I(
Q̃
Q̂
− HW (Y |X1 , X2 ) − R2
=
min
Q̃Y |X2 =WY |X2 ,Q̃Y |X1 =WY |X1
=
n
min
Q̃Y |X2 =WY |X2 ,Q̃Y |X1 =WY |X1
i
+
i
h
IQ̃ (X1 ; Y ) + IQ̃ (X1 ; X2 |Y ) + HQ̃ (Y |X1 ) − HW (Y |X1 , X2 ) − R2
+
IW (X1 ; Y ) + IQ̃ (X1 ; X2 |Y ) + [HW (Y |X1 ) − HW (Y |X1 , X2 ) − R2 ]+
= IW (X1 ; Y ) + [IW (X2 ; Y |X2 ) − R2 ]+ ,
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
max[t0 (W ),f (W )]≤f (Q̂)+R2 −I(
(a)
=
˜ Q̃) ≥
I(
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
t0 (W )≤f (Q̂)+R2 −I(
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
maxQ̂: Q̂∈S(W ),Î(Q̂)≤R [f (Q̂)−IQ̂ (X2 ;X1 ,Y )]≤f (Q̂)−IQ̂ (X2 ;X1 ,Y )
o
(D.14)
˜ Q̃)
I(
IQ̃ (X1 ; X2 , Y )
2
(b)
≥
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
f (W )−IW (X2 ;X1 ,Y )≤f (Q̂)−IQ̂ (X2 ;X1 ,Y )
(c)
=
IQ̃ (X1 ; X2 , Y )
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
−HW (Y |X1 )≤−D(Q̂||W |PX )−IQ̂ (X1 ;X2 )−HQ̃ (Y |X1 )
(d)
≥ IW (X1 ; Y ) +
HW (Y ) − HQ̃ (Y |X1 ) + IQ̃ (X1 ; X2 |Y )
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
−HW (Y |X1 )≤−D(Q̂||W |PX )−IQ̂ (X1 ;X2 )−HQ̃ (Y |X1 )
D(Q̂||W |PX ) + IQ̂ (X1 ; X2 ) + IQ̃ (X1 ; X2 |Y )
(e)
= IW (X1 ; Y )
(D.15)
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
max[t0 (W ),f (W )]≤f (Q̂)+R2 −I(
(a)
=
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
ˆ Q̂)
f (W )≤f (Q̂)+R2 −I(
˜ Q̃)
I(
min
HW (Y ) − HQ̃ (Y |X1 ) + IQ̃ (X1 ; X2 |Y )
min
HW (Y ) − HQ̂ (Y |X1 ) + IQ̃ (X1 ; X2 |Y )
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
−HW (Y |X1 ,X2 )≤−D(Q̂||W |PX )−HQ̂ (Y |X1 )−IQ̂ (X1 ;X2 )+R2
(b)
=
˜ Q̃) ≥
I(
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
−HW (Y |X1 ,X2 )≤−D(Q̂||W |PX )−HQ̂ (Y |X1 )−IQ̂ (X1 ;X2 )+R2
(c)
≥ IW (X1 , X2 ; Y ) − R2
+
min
ˆ Q̂)≤R2
Q̃∈S(W ),Q̂∈Ŝ(Q̃),I(
−HW (Y |X1 ,X2 )≤−D(Q̂||W |PX )−HQ̂ (Y |X1 )−IQ̂ (X1 ;X2 )+R2
≥ IW (X1 , X2 ; Y ) − R2
D(Q̂||W |PX ) + IQ̂ (X1 ; X2 ) + IQ̃ (X1 ; X2 |Y )
(D.16)
25
[21] J. Scarlett, A. Martinez, and A. G. i Fábregas, “Mismatched multi-letter
successive decoding for the multiple-access channel,” in Proc. ISIT 2014,
Jul. 2014, pp. 2539–2543.
[22] J.
Scarlett,
“Reliable
Communication
Under
Mismatched
Decoding,” Ph.D. dissertation, University of Cambridge, 2014,
http://itc.upf.edu/biblio/1061.
[23] B. Bandemer, A. El Gamal, and Y. K. Kim, “Optimal achievable rates for
interference networks with random codes,” IEEE Trans. on Inf. Theory,
vol. 61, no. 12, pp. 6536–6549, Dec. 2015.
[24] L. Weng, S. S. Pradhan, and A. Anastasopoulos, “Error exponent regions
for Gaussian broadcast and multiple-access channels,” IEEE Trans. on
Inf. Theory, vol. 54, no. 7, pp. 2919–2942, July 2008.
[25] R.
Gallager,
“Fixed
composition
arguments
and
lower
bounds to error probability,” July 1994. [Online]. Available:
www.rle.mit.edu/rgallager/documents/notes5.pdf
| 7cs.IT
|
Compact Representations of Event Sequences
Nieves R. Brisaboa∗ , Guillermo de Bernardo∗ , Gonzalo Navarro† , Tirso V. Rodeiro∗
and Diego Seco‡
arXiv:1803.02576v1 [cs.DS] 7 Mar 2018
∗ University
of A Coruña
A Coruña, Spain
[email protected]
[email protected]
[email protected]
† CeBiB
and DCC
University of Chile
Santiago, Chile
[email protected]
‡ University
of Concepción
Concepción, Chile
[email protected]
Abstract
We introduce a new technique for the efficient management of large sequences of multidimensional data, which takes advantage of regularities that arise in real-world datasets
and supports different types of aggregation queries. More importantly, our representation
is flexible in the sense that the relevant dimensions and queries may be used to guide the construction process, easily providing a space-time tradeoff depending on the relevant queries in
the domain. We provide two alternative representations for sequences of multidimensional
data and describe the techniques to efficiently store the datasets and to perform aggregation queries over the compressed representation. We perform experimental evaluation on
realistic datasets, showing the space efficiency and query capabilities of our proposal.
Introduction
In recent years, the amount of information produced in different domains has increased exponentially. In addition to transactional data in operational databases,
other areas such as Geographic Information Systems (GIS) or bioinformatics produce
huge amounts of information obtained from sensor networks or other tools. In many
cases, the information produced is multi-dimensional, in the sense that each entry
contains a set of features that can be regarded as dimensions. For instance, in a sales
database, each transaction involves a salesman, a client, a product, etc.
In many domains, this data collection is then used to answer different kinds of
analytic queries. Frequently queries involve grouping and filtering operations on the
data, by specific dimensions, to extract relevant information from cumulative values.
This is performed for instance in OLAP systems [1], designed to provide information
about aggregated values according to multiple characteristics or dimensions of the
data. For instance, a system storing the activity log of employees will contain entries,
including time, employee and activity. From this dataset, we can obtain the time
devoted to activity X for one/all the employees in a period. While this information
Funded in part by European Union’s Horizon 2020 research and innovation programme under
the Marie Sklodowska-Curie grant agreement No 690941. NB, GdB and TR supported by: CDTI,
MINECO grant ITC-20151247; MINECO grants TIN2016-78011-C4-1-R, TIN2016-77158-C4-3-R
and TIN2013-46238-C4-3-R; and Xunta de Galicia grant ED431C 2017/58 GRC. DS supported
by: MINECO (PGE and FEDER) grants TIN2016-77158-C4-3-R and TIN2013-46801-C4-3-R, and
Fondecyt-Conicyt grant number 1170497. GN supported by Basal Funds FB0001, Conicyt, Chile.
can be computed from the original data, in most cases the information is processed
and stored in a separate data warehouse to answer those aggregation queries.
However, many times data has some kind of ordering that can be exploited. For
example, a traditional data warehouse cannot answer how many times activity X
(meeting with a customer) has been followed by the activity Y (sell a product). In a
GIS dataset, we may have the elevations of roads. A traditional data warehouse could
answer queries about how many Km. a road X traverses at elevation 1000m, 1050m
or 1010m, but we could also be interested in knowing if those Kms. at elevation
1000m are contiguous (the road has no slopes) or if they are short sectors because
the road has many changes of elevation. In these cases where the order of the values
of some variable is relevant, typical data warehouses are not capable of answering all
the relevant queries because they cannot exploit the sequence of events (or values of
the variable of interest: activity or elevation in our examples).
In this paper we introduce a new representation for sequences of multi-dimensional
data that is designed to efficiently solve aggregation queries, while keeping all the information of the original sequence. Our proposal is specifically designed to take
advantage of repetitiveness in the original sequence. This repetitiveness occurs in
many real-world sequences due to the locality of the data. For instance, in the example of activities performed by employees, an employee is likely to perform the same
activity in consecutive time instants. Our proposal provides a space/time tradeoff,
allowing us to use only the space necessary to provide indexing capabilities in the relevant dimensions. Additionally, in many cases our representation can also be easily
“extended” to add query capabilities in new dimensions without altering the original
structure and features. Our proposal is based on the wavelet tree [2]. We combine
wavelet trees with other compact data structures in order to exploit regularities in
the datasets and to perform different types of counting/aggregation queries. We experimentally evaluate our representations, comparing them with a more traditional
alternative based on storing separately the original data sequence and building an
indexed representation to answer aggregation queries.
Background
Due to space constraints, we introduce here the compact data structures we use
as building blocks of our solutions, but we refer to [3] for details about them. An
ubiquitous building block of most compact data structures is a bitmap with support
for rank/select operations. Given a binary sequence of length n, a bitmap B[1, n]
provides access to any position i of the sequence, denoted as access(B, i), counts the
number of occurrences of bit v up to position i, rankv (B, i), and retrieves the position
of the j-th occurrence of bit v, selectv (B, j). All these operations can be supported in
O(1) time with n + o(n) space [4]. When the binary sequence is very sparse (i.e. the
number of 1s, m, is much smaller than n), space can be improved to nH0 (B) + O(m)
n
bits, supporting select in O(1) time and rank in O(min(log m, log m
)) time [5].
The wavelet tree [2], hereinafter wt, is a data structure for the representation
of sequences. Given a sequence S[1, n] over an alphabet Σ[1, σ], a wt is built as a
perfectly balanced binary tree that subdivides the elements in the sequence according
to the position of their symbol in Σ. At the root node of the tree, the original alphabet
is divided into two halves ΣL and ΣR . Then, a bitmap B[1, n] is associated with the
node, so that B[i] = 0 ⇐⇒ S[i] ∈ ΣL , and B[i] = 1 otherwise. The process is
repeated recursively, considering in the left sub-tree the subsequence of symbols in
ΣL and in the right sub-tree the subsequence of symbols in ΣR . The tree has ⌈log σ⌉
levels, and a total of n bits per level, for a total size of n⌈log σ⌉ bits.
The wt supports several operations, being access and rank the ones we use in this
work. To access S[i] in the wt, we start at the root node of the tree and check its
bitmap B. If B[i] = 0, the symbol is in the left subtree. The offset of the symbol
in the left subtree can be easily computed as the number of 0s in the current node
up to position i (i.e. we compute the new i as rank0 (B, i). Conversely, if B[i] = 1
we move to the right subtree and set i = i + rank1 (B, i). After ⌈log σ⌉ steps the
path traversed corresponds to the position of [i] in the alphabet. On the other hand,
rankc (S, i) counts the occurrences of symbol c in S up to position i and it can be also
implemented in O(log σ) time with a similar traversal.
In addition to plain representations of wts, that essentially require the same space
as the original uncompressed sequence, several other variants allow the compressed
representation, achieving space proportional to the zero-order entropy of the input.
In particular, we use a run-length compressed representation of the bitmaps of the
wt [6] to exploit the regularities of our sequences.
Our proposal
Formulation of the problem
Consider a sequence S of data where each datum is a multidimensional entry, from
a given set of dimensions, of the form (d1 , d2 , . . . , dk ) ∈ D1 × D2 × · · · × Dk . This
sequence may be a set of measurements, entries in a database, or any other source of
multi-dimensional data. We also may assume that the order of the elements in the
sequence could be random or according to some ordering, either time or any other
value related to one or more dimensions Di .
The queries of interest in this domain involve aggregations of values for specific
values, or ranges of values, in one or more of the dimensions of the dataset. That is,
we are interested in counting the number of entries in the sequence for a given value
of Di , or for a given set of constraints across multiple dimensions, such as d1 = x,
d2 ∈ [d2ℓ , d2r ], etc.
A plain representation of the original sequence is quite inefficient to retrieve any
kind of aggregate information. In order to answer these queries, additional data structures are used to keep track of the specific values, thus storing only accumulated data.
Hence, both the representation of the sequence and the additional data structures are
required to keep the ability to access and decode the sequence of operations.
Wavelet tree composition for aggregate queries
Our proposal is based on using wts to represent the sequence and provide efficient
aggregation on a subset of the dimensions. The wt representation of a sequence provides a reordering of the elements in the original sequence according to the “alphabet
order”. In other words, when a wt of the sequence is built on an alphabet Σ, the
elements of the sequence are stably-sorted by their position in this alphabet, so that
the sequence can be accessed in the original order, using the root of the wt, or in
alphabet order, on the leaves.
Consider a sequence of tuples (d1 , . . . , dk ) sorted by D1 , then D2 , and so on. If we
build a wt, SDi , on this sequence according to Di , we are in practice stably-sorting all
the elements of the sequence by Di . Hence, this wt provides the support to efficiently
restrict our queries to specific values of Di , thus effectively providing aggregation
capabilities in dimension Di . For example, in order to count the number of entries
for a given c ∈ di , we just need to compute rankc (SDi , n).
In order to provide aggregation capabilities in another dimension, we can easily repeat the process, creating a wt over the re-ordered sequence using a different
dimension. If we build a new wt, SDj , using the values of Dj , we are in practice
grouping elements according to that dimension, while keeping the previous ordering
by Di . Using rank operations, we can now easily compute rankc′ (SDj , n) to count the
occurrences of any c′ ∈ Dj . In addition, we can easily restrict the previous query to
the range covered by any c in Di , thus we can also answer queries involving any pair
of values in both dimensions.
Notice that, since after reordering by a second dimension the data are still sorted
by the previous one in each leaf (that is, they are sorted by Dj , and by Di inside each
group of Dj ), we can also efficiently look for the elements in the sequence that have
a specific value for Dj and a range of values of Di : we just need to restrict the search
in the second wt to the range of values for the other dimension, which can be easily
computed. Table 1 shows an example in which the data are originally sorted by day,
then reordered by activity and finally by employee.
Original order
Day
Emp
Act
1
1
C
1
1
E
1
1
A
1
2
E
1
2
B
2
1
A
2
1
C
2
2
E
3
2
E
Reordering by Act
Day
Emp
Act
1
1
A
2
1
A
1
2
B
1
1
C
2
1
C
1
1
E
1
2
E
2
2
E
3
2
E
Reordering by Emp
Day
Emp
Act
1
1
A
2
1
A
1
1
C
2
1
C
1
1
E
1
2
B
1
2
E
2
2
E
3
2
E
Table 1: Example of reorderings with dimensions Day, Employee and Activity
Our representation can efficiently compute queries involving consecutive elements
in any of the reorderings. For example, count the number of entries involving activity
E, which are 4, can be answered in the first reordering. In the same reordering, the
query can be also restricted to a range of days (e.g. entries involving activity E during
days [1..2], which is 2). As for the second reordering, we may count the number of
entries in which employee 2 was performing activity E during days [1..2], which is 2.
Interesting enough, the different wts need not be just stacked, but can be combined
in different ways according to the relevant queries in the domain. For instance, if we
have a 4-dimensional dataset, over A × B × C × D, we may first process the sequence
of values of D, sorting it by that dimension. The result can then be used in two
different wts, one that uses C as the vocabulary and another that uses B. In this
way, we shall be able to easily count occurrences for a given d, a given pair (c, d⋆)
or a given pair (b, d⋆), where x⋆ denotes the ability to search for a specific value or
a range of values. An alternative representation that builds the 3 wts in sequence,
would be able to answer aggregation queries for given d, (c, d⋆) or (b, c, d⋆). In this
way, the final representation can include as many combinations as needed to provide
the necessary query capabilities, at a cost of extra space. Given a set of queries,
the problem of computing the minimum number of necessary wts is an extension of
the shortest common superstring problem [7], i.e. obtaining the shortest string that
contains a set of given strings.
Compressible sequences
In many real-world scenarios, sequences of events are highly repetitive in terms of
number of repeated k-tuples. This effect, called locality, is increased with the granularity of the sequence. In the example described in the Introduction, if activities are
recorded every minute, the sequence will have more locality than if they are recorded
every hour. We propose now two implementations that take advantage of this property, providing solutions that are insensitive to the granularity of the sequences. This
is interesting because it makes them suitable to represent sequences with a high level
of detail, i.e. resolution, in little space and supporting aggregation queries.
We assume that the original sequence is sorted in such a way that aforementioned
locality exists (e.g. time or space depending on the domain). Then, locality produces
runs of symbols in the sequence. An out-of-the-box solution to exploit the existence
of these runs is to use run-length compressed wts [6]. We refer to this solution as
wtrle. Runs on the sequence produce runs on the bitmaps of the wt, which can be
represented in few space and still support rank/select operations [8]. Overall, these
wts require space proportional to the number of runs in the sequence, instead of to
the length of the sequence itself. Counting and access queries can be implemented
as rank and access operations on the wt, respectively. When queries involve more
dimensions, an additional component is necessary. In the wt composition explained
above, a query on a wt is refined on a second wt, and so on. To do that, we need
to store for each leaf of the wt the number of elements that are lower, in the order
defined by such wt, than its corresponding symbol. This component is used to restrict
the query in the subsequent wt and it can be implemented as a plain array or as a
sparse bitmap, depending on the size of the dimension.
This simple approach represents the runs at each level of the wt, which is somehow
a waste of space. Thus, our second approach, wtmap, uses a technique to remove the
runs from the original sequence. We store a bitmap BM , of length n, that marks
the start of each run. Then, a wt is built over a sequence S ′ of length n′ < n,
in which each run of value v is represented as a single value v. As this sequence
does not contain runs, the wt is built using plain bitmaps, which have the additional
benefit of being faster for querying. A query on the original sequence can be easily
mapped to a query on this new wt using rank operations on BM . However, this
wt does not contain information about repetitions of symbols, which is necessary to
support counting operations. Hence, a second bitmap BC is built, storing in unary
Figure 1: Schema of the wtmap solution for a sequence of 2 days, 2 employees, 4 activities
and 10 time instants per day. Dimension Di represents activities. Highlighted elements are
those visited to count the time-instants devoted by e2 to activity A during d2 , which are 4.
the length of each run. This bitmap considers the runs in the reordering performed
by the wt, i.e. it is aligned with the leaves of the wt. Rank/select support is also
necessary on this bitmap to solve counting queries. The two bitmaps, BM and BC , are
sparse, which is directly related with the existence of runs on the original sequence,
and are implemented using Elias-Fano representation [5]. The same final component
described for the wtrle is necessary here to do the composition between wts. This
component is shown as bitmap BL in Figure 1, which illustrates this approach.
Experimental evaluation
We design a set of experiments focused on a realistic use case, involving the storage
and access to employee tracking information. We assume that we have a set of entries
storing the following information: day, employee, time instant and activity. We want
to keep track of all the information of our employees, hence being able to recover
the specific activity that was being performed at a given time instant by a specific
employee. In addition, typical queries in this domain involve checking the amount of
time devoted to a specific activity by an employee, or by all employees. Hence we
need to store the original sequence but also to efficiently answer aggregate queries.
We shall consider for instance aggregations of time per activity/employee/day, and
per activity/day.
Problem setup
In this kind of data, time is a natural ordering of entries, and also provides compressibility due to locality: consecutive entries for the same employee and day should
usually return the same activity, especially if the measurement frequency (resolution)
is high. In this sense, storing the sequence of activities sorted by employee and time
should lead to long runs of similar activities. We will consider the sequence of activities as sorted first by day, and then by employee and time, since days are a natural
way of grouping data, the locality of activities is not affected and it becomes easier to
answer queries for a specific day. Even though not all employees work every day, we
can ignore this problem by storing every possible combination of day, employee and
time instant, and using a special activity 0 to denote the fact that the employee was
not working at that point. This increases the length of the sequence, but it makes
computations simpler, and our representations will compress these long runs of values
in little space.
The queries that we will consider in our experiments involve aggregation operations and retrieval of specific entries in the sequence. We name our counting queries
C-x D-yE-z A, where D, E and A are days, employees and activities, and x,y,z can
denote a single value (1), a range (“r”) or all possible values (“a”). Hence, C-1D1E-1A counts the time devoted on 1 day, by 1 employee, to a specific activity. A
more complex aggregation is C-1D-aE-1A (1 day, all employees, 1 activity), that can
be trivially extended to a range of days. We also test ranges of days with query
C-rD-1E-1A. In addition to these aggregation queries, we also test the ability of our
proposal to access specific positions of the original sequence: query Acc retrieves the
activity being performed, given an employee, day and time instant.
Taking into account the characteristics of the dataset, we build our representation
sorting entries by activity first, in a first wt decomposition; then we group again by
employee. The first decomposition allows efficient counting operations by activity and
day (or range of days); the second one works for queries on activity-employee-day.
Notice also that since we actually store every possible combination of values, an Acc
query can be easily translated into a position to extract in the sequence.
Baseline representation
As a baseline for comparison we build a simpler representation that uses the same
ordering of the information and takes advantage of it to answer the same relevant
queries with additional data structures. The baseline consists of two separate components: a representation of the original sequence, which can be used to access the
original data, and an additional data structure for aggregation queries. This baseline
thus follows the usual approach of separating the original data and precomputing
aggregated values in a separate structure.
The sequence of tuples is stored using two arrays and two additional bitmaps.
Initially, the sequence is “compressed” removing repeated entries: each group of
consecutive entries with the same values is replaced with a single entry storing the
activity and the amount of time devoted to it. The compact representation works
with this shorter sequence. A bitmap BD marks the positions in the sequence where
a change of day occurs; a second bitmap BE is used to locate positions where each
employee sequence begins. Using these two bitmaps, we only need to store activities
and times for each position: two arrays store the activities and time values for each
entry. With this representation the original contents can be recovered using the
bitmaps to locate the appropriate run and processing the sequence of times to locate
the desired position. Operations that require counting (time spent in an activity by
an employee, or by all employees in a day) can also be computed using this sequence,
but may require the traversal of several different regions of the array to compute the
overall result.
To provide efficient counting operations, the baseline also stores a data structure
specifically designed to compute aggregated values. The data structure we use is
called CMHD [9], and is a compact representation designed for aggregation queries
in hierarchical domains such as OLAP databases. This representation essentially
decomposes the n-dimensional datasets, according to predefined hierarchies, and is
able to store accumulated values at each level of decomposition (for instance, given
products and places, it can store in the same representation accumulated values by
category and country, and also by individual products and places). In our domain
no natural hierarchies appear apart from days and time. Nevertheless, we can build
fictitious hierarchy levels to efficiently answer queries: decomposing first some dimensions we can obtain a first level in the CMHD that stores cumulative values for all
employees, 1 activity and 1 day; then, decomposition can continue in a second level
to store cumulative values for 1 employee, 1 activity, 1 day. In this way, queries that
involve a single value or all values can be answered efficiently, as long as a hierarchy
is appropriately built in the CMHD.
Experiments and results
We built several synthetic datasets following the expected distribution for the domain.
All our datasets have similar realistic characteristics but different size in different
dimensions, to measure the evolution of the proposals. The changing parameters
in the datasets are the number of activities A, the number of employees E and the
number of time instants per day, or resolution, R. The number of days is set to
a fixed value (500), since it has little effect on query times and only changes the
length of the sequence. Datasets T-rrr (where rrr is the resolution) are built with
E = 50, A = 16 and R ∈ {720, 1120, 5760, 11520}. Datasets Dat-ee-aa are built
with R = 5760, E ∈ {20, 50, 100} and A ∈ {16, 32, 64, 256}. In all the datasets, we
consider that employees work on shifts (50% of the time each working day), and have
free days (only 80% of the employees work each day). For each dataset and query
type we generated sets of 1 million queries, selecting random values and intervals.
In this section we compare our two proposals, wtrle and wtmap, with the alternative baseline representation, for all the test datasets. All the representations are
implemented in C++ and compiled with g++ with full optimization enabled. All the
experiments in this section were performed in an Intel Core [email protected] and
16GB of RAM, running Ubuntu 16.04.4.
Due to space constraints, we only show space required by some of the datasets,
for selected numbers of employees, activities and resolutions, since behavior in all
the datasets is similar. Table 2 shows this summary. For each dataset, the input
size is computed considering a completely plain representation of all the tuples in
the sequence, with each value stored using 32 bits, so this size increases linearly with
the number of entries. All the compressed representations are much smaller than the
original sequence, thanks to the efficient compression of the runs of repeated values.
Both of our representations are in all cases considerably smaller (5 to 10 times smaller)
than the baseline. The most efficient representation is wtmap, since it removes runs
of repeated values in a single step, while wtrle has a small overhead to compress the
runs in every level of the wt.
Dataset
T-2880
T-11520
Dat-20-16
Dat-20-64
Dat-20-256
Dat-50-16
Dat-50-64
Dat-50-256
Dat-100-16
Dat-100-64
Dat-100-256
Input size
1,098.63
4,394.53
878.91
878.91
878.91
2,197.27
2,197.27
2,197.27
4,394.53
4,394.53
4,394.53
baseline
30.73
30.85
12.41
13.20
13.74
30.65
33.14
34.43
61.76
65.78
67.96
wtrle
3.58
4.09
1.50
2.04
2.35
3.82
5.35
6.20
7.72
10.78
12.45
wtmap
3.33
3.70
1.35
1.51
1.65
3.51
3.99
4.35
7.27
8.15
8.84
Table 2: Space required by all the datasets (sizes in MB)
Next we compare our proposals with the baseline for the different queries described
for this domain. Again, due to space constraints, we only show the results for some
of the studied datasets. Comparison results for the other datasets used are similar to
those introduced here, with query times varying slightly depending on the size of the
sequence and the different dimensions.
Table 3 contains the query times for all approaches for all queries. The first query
we study is the access query Acc, that recovers random positions in the sequence. In
this query both of our proposals are faster than the baseline in all cases, and again
wtmap is the fastest of our alternatives.
Acc
C-1D-1E-1A
C-1D-aE-1A
C-rD-1E-1A
base wtrle wtmap
base
wtrle wtmap base wtrle wtmap base wtrle wtmap
Dat-20-16
0.51 0.36
0.23 1.46 (0.45∗ ) 1.12
0.74 1.35 2.21
1.67
— 7.25
3.49
0.56 0.53
0.32 1.57 (0.58∗ ) 1.58
1.15 1.38 3.30
2.28
— 8.60
4.65
Dat-50-64
0.46 1.50 (0.66∗ ) 1.97
1.57 1.31 4.50
2.97
— 11.78 7.19
Dat-100-256 0.61 0.56
∗ Values in parentheses are performed counting sequentially in the baseline sequence representation
Dataset
Table 3: Query times for access (Acc) and counting queries. Times in µs/query
Next, also in Table 3, we study aggregation queries. For C-1D-1E-1A we show
two different times for the baseline: the first time is obtained querying the CMHD
representation, and the time in parentheses is obtained traversing the sequence representation. In this query, the naive approach that traverses the sequence obtains
in practice the best results. This is due to the relatively small number of activities
per day in this domain. Note however that the cost of the naive sequence traversal
is linear on the number of entries to search, so it may be slower than any of the indexed proposals even for simple queries if the number of entries to search is relatively
large. In any case, our representations are still competitive in query times and use up
to 10 times less space. Moreover, our proposals are faster than the indexed CMHD
representation in many cases.
In the last two queries a naive traversal of the sequence in the baseline becomes
too slow (5 to 20 times slower than our proposals) so we only compare with the
CMHD. In query C-1D-aE-1A the CMHD can be faster by storing cumulative values
for the corresponding query, but our proposals are still competitive. In addition, our
proposals can answer a generalization of the query to a range of days (C-rD-aE-1A)
in roughly the same time, whereas the CMHD would have to resort to summing up
multiple entries, becoming much slower. This is also shown in query C-rD-1E-1A: our
proposals are a bit slower in this query, but query times are not too high in comparison
with previous queries; however, the CMHD would have to obtain and sum cumulative
values for each day, becoming too slow in practice, with times again comparable
to a naive traversal of the sequence. A partial improvement on the CMHD could
be obtained by generating fixed divisions as time hierarchies (by week, by month),
reducing the number of sums to perform, but this requires previous knowledge of the
desired ranges, and unless queries involve ranges exactly matched with hierarchical
values, the CMHD is expected to be much slower in these queries. While the CMHD is
much more dependent on a previous definition of hierarchies of interest, our proposals
are more flexible and can efficiently answer range queries on the different dimensions.
Conclusions
We have presented two different compact representations for multidimensional sequences with support for aggregation queries. We show that our representations can
take advantage of locality to store very large datasets in reduced space. Our experiments show that our proposals are smaller, and faster to access, than simpler
representations of multidimensional sequences. Also, and even if storing much more
information, our proposals are competitive on aggregation queries with state-of-theart data structures designed specifically for those queries.
An interesting line of future work is to explore the effect of the encoding of the
symbols in real domains. In our running example, some activities may be more likely
to be performed after a specific one than others. For example, take-a-nap activity
after lunch. By encoding similar activities with closer symbols, these regularities are
automatically exploded by the wtrle, but not for any other of the solutions.
References
[1] S. Chaudhuri and U. Dayal, “An overview of data warehousing and olap technology,”
SIGMOD Rec., vol. 26, no. 1, pp. 65–74, Mar. 1997.
[2] R. Grossi, A. Gupta, and J. S. Vitter, “High-order entropy-compressed text indexes,”
in Proc. 14th SODA, 2003, pp. 841–850.
[3] G. Navarro, Compact Data Structures: A Practical Approach. Cambridge University
Press., 2016.
[4] G. Jacobson, “Space-efficient static trees and graphs,” in Proc. SFCS, 1989, pp. 549–554.
[5] D. Okanohara and K. Sadakane, “Practical entropy-compressed rank/select dictionary,”
in Proc. 9th ALENEX, 2007.
[6] V. Mäkinen and G. Navarro, “Succinct suffix arrays based on run-length encoding,” in
Proc. 16th CPM, 2005, pp. 45–56.
[7] K. Räihä and E. Ukkonen, “The shortest common supersequence problem over binary
alphabet is np-complete,” Theor. Comput. Sci., vol. 16, pp. 187–198, 1981.
[8] O. Delpratt, N. Rahman, and R. Raman, “Engineering the LOUDS succinct tree representation,” in Proc. 5th WEA, 2006, pp. 134–145.
[9] N. R. Brisaboa, A. Cerdeira-Pena, N. Lopez Lopez, G. Navarro, M. R. Penabad, and
F. Silva Coira, “Efficient representation of multidimensional data over hierarchical domains,” in Proc. 23rd SPIRE, 2016, pp. 191–203.
| 8cs.DS
|
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
arXiv:1612.06260v2 [math.NT] 28 Jun 2017
ZACHARY CHARLES
Abstract. We present a randomized polynomial-time algorithm to generate an ideal
and its factorization uniformly at random in a given number field. We do this by generating a random integer and its factorization according to the distribution of norms of
ideals at most N in the given number field. Using this randomly generated norm, we can
produce a random factored ideal in the ring of algebraic integers uniformly at random
among ideals with norm up to N , in randomized polynomial time. We also present a
variant of this algorithm for generating random factored ideals in function fields.
1. Introduction
We consider a generalization of the following problem: Given an integer N > 0, generate
an integer in [1, N] uniformly at random, along with its prime factorization, in polynomial
time. Since there are currently no known polyonmial time factorization algorithms, we
cannot simply generate an integer and factor it. Instead, we can generate the prime
factorization uniformly at random.
In his thesis, Bach gave a randomized polynomial time method to uniformly produce
a factored integer in [ N2 , N] [2]. Bach’s method uses only an expected number of log N
primality tests in [1, N] [3]. Since we can test for primality in polynomial time by the
work of Agrawal, Kayal, and Saxena [1], Bach’s algorithm runs in randomized polynomial
time. In 2003, Kalai gave another method for doing this using a conceptually simpler
but slower method. Kalai’s algorithm uses an expected number of O(log(N)2 ) primality
tests [7]. In 2016, Lebowitz-Lockard and Pomerance gave a variant of Kalai’s algorithm
to produce random factored elements on Z[i] with norm at most N [9]. This paper gives
a generalization of these algorithms that will produce random factored ideals in OK for
any number field K. The method given will be polynomial in N and the degree d, where
we fix d and assume N tends to infinity.
We first discuss Kalai’s algorithm and give a brief analysis. We then use these ideas to
generate, in polynomial time, an integer in [1, N] according to the distribution of norms of
ideals in OK . We then discuss how to use this algorithm to produce random factored ideals
2
of OK with norm up to N uniformly at random using O((log N)d +d+1 ) primality tests.
This algorithm is then modified to generate random factored ideals in function fields using
2
O((log N)d +d+1 ) primality tests. While the algorithm described by Agrawal, Kayal, and
Saxena runs in O(log 15/2 N), Lenstra and Pomerance later developed a primality testing
algorithm with run time O(log6 N) [10]. Using the latter algorithm, our algorithms for
generating random factored ideals in number fields and function fields have expected run
2
times that are O((log N)d +d+7 ).
2. Kalai’s Algorithm
In his 2003 paper [7], Kalai presents the following algorithm:
The author would like to thank Eric Bach for his invaluable advice, suggestions, and encouragement.
The author was supported by the National Science Foundation grant DMS-1502553.
1
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
2
Algorithm 1. Input: A positive integer N.
Output: A positive integer r ∈ [1, N] and its prime factorization, produced according to
a uniform distribution.
1. Generate a sequence N ≥ s1 ≥ s2 ≥ . . . ≥ sl = 1 where s1 ∈ {1, . . . , N}, si+1 ∈
{1, . . . , si } are chosen uniformly at random. Stop when sj = 1.
2. Let r be the product of the prime si ’s.
3. If r ≤ N output r with probability Nr .
4. If r > N or we do not accept r ≤ N, return to step 1.
Let p be a prime number. Then there is a first number si produced in the range
{1, 2, . . . , p}. It is chosen uniformly at random, so the probability that si = p is p1 . Generating e factors p then occurs with probability p1e (1 − p1 ). The probability of outputting
a given r is then:
v (r)
Y
1
r Y 1 p
r
vp (r)
1−
=
Pr r =
p
N
N
p
p
p≤n
p≤N
r 1Y
1
=
1−
N r p≤n
p
=
MN
N
Q
Here MN = p≤N (1−1/p). Therefore, if the algorithm terminates, it produces r and its
prime factorization with probability N1 . Moreover, the probability that Kalai’s algorithm
terminates after a single round is MN . We therefore expect MN−1 trials before we output
a number.
We now show that the expected number of primality tests is O(log(N)2 ). Note that
given some si > 1, the expected value of si+1 is s2i . Therefore, a given list s1 , . . . , sl has
expected length O(log N). We need an expected number of O(log N) primality tests for
every round of the algorithm. Hence, we do an expected number of O(MN−1 log N) primality tests overall. By Mertens’ theorem, MN−1 = O(log N). Therefore Kalai’s algorithm
uses O(log(N)2 ) primality tests in expectation before the algorithm terminates.
3. Generating Random Factored Norms
We now give a generalization of Kalai’s algorithm that produces a random norm of an
ideal in OK , along with its prime factorization. Let K be a number field of degree d,
and N an integer satisfying d << log(N). The algorithm will use a polynomial number
of primality tests. In order to do so, we need knowledge of how the rational primes
split. In general, this is computationally efficient for primes not dividing the discriminant
of K [11]. If f (x) is the monic irreducible polynomial over Q determining K, then
we can determine how p splits by factoring f (x) modulo Z/pZ. Factoring polynomials
over finite fields can be accomplished in randomized polynomial time [8]. The remaining
prime numbers can be factored using a method of Chistov to factor polynomials in Qp in
polynomial time [4]. We therefore assume we can factor rational primes in K.
For any integer r > 0, let D(r) denote the number
of ideals in K of norm r. Let p ∈ Z
Q
ej
fj
e
be a rational prime with factorization pOK = m
p
j ) = p . Finding D(p )
j=1 j . Let N(p
P
m
can now be reduced to determining the number of solutions to i=1 ci fi = e. This is an
instance of the subset sum problem. We know that e is O(logP
N). Therefore, we can solve
the subset sum problem in O(m log N) operations [5]. Since m
i=1 ei fi = d, we know that
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
3
m ≤ d. The runtime for this becomes O(log N) as a result. Since D(r) is multiplicative,
we can calculate D(r) for any r relatively efficiently (certainly in polynomial time).
We will generate r and keep r with probability proportional to rD(r)
. In order to
N
guarantee a well-defined probability, we need to bound D(r) by a factor that can be
incorporated in to the selection of the factors. We use the following result. Let Ω(r)
denote the number of prime factors of r with multiplicity, and let Ωd (r) be the number of
prime factors (counting multiplicity) of r that are greater than d.
Proposition 1. D(r) ≤ dΩ(r) .
Proof. Note that D(r) is multiplicative. This is because prime ideals in O
norms
QK have
ep
that are powers of prime numbers [11]. Constructing an ideal of norm r = p p amounts
ep
to constructing an
Qmidealei of norm p for each prime p|r.
Say pOK =
⊂ OK prime. The norm is multiplicative, and since
i=1 pi for p P
m
ei
d
= d. Fix e. Note that D(pe ) will be at
N(pOK ) = p , we must have
i=1 N(pi )
its largest when the p are all distinct (ie. ei = 1) and all the pi have norm p. Then D(pe )
will be the number of unordered sets of e elements taken from the d prime ideals. This is
bounded above by the number ordered sets, which is given by de = dΩ(r) .
We now present the main algorithm. For simplicity of analysis, we demonstrate the
case where d is even and N ≡ d − 1 mod d. Let k ∈ Z be such that N = kd + (d − 1).
These assumptions are not necessary, but help simplify minor details.
Algorithm 2. Input: A positive integer N.
Output: A random integer in [1, N]. The integer is generated according to the distribution
of norms of ideals in OK with norm up to N.
1. Generate ⌊ d2 ⌋ lists of integers as follows. For b ∈ {1, 3, 5, . . . , d − 1}, generate a list
N ≥ s1,b ≥ s2,b ≥ . . . ≥ slb ,b = 1. We take s1,b ∈ {1, d + b, 2d + b, . . . , kd + b}, where
d
1
and is any other element with probability
. Take
s1,b = 1 with probability
kd + 1
kd + 1
1
si+1,b in {1, d + b, 2d + b, . . . , si,b} as 1 with probability
and any other element
si,b − b + 1
d
with probability
.
si,b − b + 1
2. Let r be the product of the prime si,b .
3. For all primes p between 1 and d, do the following: Multiply r by p with probability
p−1
, and continue to multiply by p with this probability until your first failure.
p
ψ(r)D(r)
4. If r ≤ N, keep r with probability Md (r) Ω (r) .
d d N
5. If you did not keep r, go to step 1.
If d is odd, then we take b ∈ {1, 3, 5, . . . , d − 2}. This way we ensure that we are only
picking odd numbers. If N is some other value mod d, say N = kd + j for j < d − 1
then we can instead do the following. Let N ′ be the smallest number above N such that
N ′ ≡ d − 1 mod d. Run steps 1-3 of the algorithm above with N ′ substituted for N. Then
run step 4 by rejecting all r > N instead. For simplicity, we will analyze the algorithm
in the case that N ≡ d − 1 mod d.
Let n denote the residue of n mod d. We define ψ(r) as follows. Let ψ(r) have the same
factors as r, except that for any prime p|r, p > d, replace the factor of p with a factor of
p − p + 1. Note that ψ(r) ≤ r and they share the same prime factors, with multiplicity,
for 2 ≤ p < d.
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
4
We define Md (r) by:
vp (r)
1
α
Md (r) =
p−1
2≤p<d
−1
d + ⌊log N⌋ − 1
α=
⌊log N⌋
Y
4. Analysis of the algorithm
We wish to show that the probability of accepting r is a well-defined probability. Let
g(r) be the product of all prime factors p of r, with multiplicity satisfying p > d. Note
that since D(r) is multiplicative, we have
D(r) = D(g(r))
Y
D(pvp (r) )
2≤p≤d
(r)−1
By the discussion in section 3, D(pvp (r) ) ≤ d+vvpp (r)
. This follows from the fact that
vp (r)
D(p
) is maximized when all the primes ideals lying above p are all distinct and have
norm p. Then D(pe ) equals the number of unordered sets of size e taken from the d prime
ideals lying above p.
Using the above we find
ψ(r)D(r)
dΩd (r) N
Y 1 vp (r)
ψ(r) D(g(r))
vp (r)
α
D(p
)
=
p−1
N dΩd (r)
2≤p≤d
Y
ψ(r) D(g(r))
vp (r)
αD(p
)
≤
N dΩd (r)
2≤p≤d
Y d + ⌊log N⌋ − 1−1 d + vp (r) − 1 ψ(r) D(g(r))
≤
vp (r)
⌊log N⌋
N dΩ(g(r))
2≤p≤d
Md (r)
≤1
Let p > d be an odd prime with p ≡ p mod d. Then we will produce exactly e factors
of p with probability given by:
e
d
d
1−
p−p+1
p−p+1
Let r be some integer at most N. Recall that ψ(r) is formed from r by replacing all
r and ψ(r) have the same prime
prime factors p|r, p > d by p − p + 1. In particular,
Q
divisors for p ≤ d. Let P r s = d<p≤N pvp (r) denote the probability that after step 2 we
have generated an integer s that is the product of the prime factors of r that are larger
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
5
than d. The probability can be worked out as follows:
vp (r)
Y
Y
d
d
vp (r)
1−
=
Pr s =
p
p
−
p
+
1
p−p+1
d<p≤N
d<p≤N
vp (r) Y
Y
d
d
=
1−
p−p+1
p−p+1
d<p≤N
d<p≤N
vp (r) Y
Y
1
d
Ωd (r)
=d
1−
p
−
p
+
1
p−p+1
d<p≤N
d<p≤N
dΩd (r) Y vp (r) Y
d
=
1−
p
ψ(r) 2≤p≤d
p−p+1
d<p≤N
Note that this last step used the fact that by definition of ψ(r), we have
vp (r) Q
vp (r)
Y
1
2≤p≤d p
=
p−p+1
ψ(r)
d<p≤N
Given any r ≤ N, the probability that we generate r after step 3 is then given by:
v (r)
Y
Y
d
p−1 p
p−1
dΩd (r) Y
vp (r)
1−
1−
p
ψ(r) 2≤p≤d
p
p
p−p+1
2≤p≤d
d<p≤N
dΩd (r) Y
d
1 Y
1−
=
(p − 1)vp (r)
ψ(r) 2≤p≤d
p d<p≤N
p−p+1
Finally, the probability that we accept this r is then given by:
Y
ψ(r)D(r) dΩd (r) Y
d
vp (r) 1
Md (r) Ω (r)
1−
(p − 1)
d d N ψ(r) 2≤p≤d
p d<p≤N
p−p+1
vp (r)
Y
d
1
D(r) Y
vp (r) 1
1−
α
(p − 1)
=
N 2≤p≤d
p−1
p d<p≤N
p−p+1
Y
D(r) Y α
d
=
1−
N 2≤p≤d p d<p≤N
p−p+1
Note that other than D(r), all terms depend only on d and N. Therefore, this generates
a number with probability proportional to D(r).
We now show that the algorithm above runs in polynomial time, with polynomial many
primality tests and factorizations of rational primes. Summing over all r at most N, the
probability that we generate an ideal is
P
Y α Y
d
r≤N D(r)
1−
N
p
p−p+1
2≤p<d
d<p≤N
Y
d
1−
Let Zn =
. By the Wiener-Ikehara Tauberian theorem (see [6]
p−p+1
d<p≤N
P
for reference), r≤N D(r) asymptotically approaches CK N, where CK is the residue of
the DedekindQ
zeta function of K at 1. Asymptotically then, the expected number of trials
−d −1
is O(α ZN
2≤p<d p).
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
6
By direct computation, we have:
d + ⌊log N⌋ − 1
−1
α =
⌊log N⌋
= (d − 1 + ⌊log N⌋)(d − 2 + ⌊log N⌋) . . . (1 + ⌊log N⌋)
≤ (2 log N)d−1
2
We have d factors of α−1 , Q
so this contributes O(log(N)d −d ) to the expected number
of trials. Note that the term 2≤p<d p contributes a term that is bounded by a constant
dd , and is therefore O(1) in terms of N. We now wish to find the contribution of the
remaining term in the probability calculation above.
Simple estimates show:
Y
ZN ≥
1−
Y
d
d
1−
d+1
p−d
d<p≤2d
2d<p≤N
Y
2d
1−
≥c
p
2d<p≤N
Q
d
Here c = d<p≤2d 1 − d+1 = (d + 1)−d . By standard estimates, such as in [12], we
have:
−1
Y
2d
1−
= O(log(N)2d )
p
2d<p≤N
2
Hence we need O(log(N)d +d ) trials before success. Since any given list has expected
2
length O(log(N)), this leads to O(log(N)d +d+1 ) primality tests.
We now use the above algorithm to generate random ideals of OK , uniformly at random
among all ideals with norm up to N. As previously stated,
we assume that we for any
Q
ei
fi
p
rational prime p, we can find the factorization pOK = m
i=1 i with N(pi ) = p . To find
an ideal of norm pe , it suffices to solve, as previously discussed, the subset sum problem:
m
X
ci fi = e
i=1
Q
ci
Any solution to this corresponds to the ideal m
i=1 p . Since e ≤ log N and m ≤ d, we
d
can clearly find all solutions in O(log N operations [5]. While this is not optimal, even
the naive approach is dwarfed by the effort needed to generate the norm. After we find all
solutions, we can then choose uniformly at random one of the solutions, which will give
us one of the ideals of the desired norm. Therefore, the algorithm runs in a number of
2
operations that is O(logd +d+1 N) primality tests and factorizations of pOK . The primality
testing is the dominant part of this in terms of run-time. Since we can perform primality
2
tests in O(log6 N), this gives us a run time that is O(logd +d+7 N) overall.
5. Function Fields
The analogy between number fields and function fields suggests that there should be
an analogous algorithm for function fields. In particular, there are well-known randomized polynomial-time algorithms for factoring over Fq [t]. To generate a factored random
polynomial in Fq [t], we could simply generate one at random and then factor it in randomized polynomial time. Much more elegant ways exist that generate the factorization
at random, the same idea used by the methods above.
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
7
Therefore, we would expect the ability to translate the algorithm above to arbitrary
function fields. Fix a function field K of degree d and N > 0. We want to generate
a random ideal I ⊂ OK with norm r(t) ∈ Fq [t] of degree at most N, along with the
factorization of I. We will use the fact that we can perform primality testing over Fq [t]
in polynomial time. We will also assume that for any irreducible polynomial f (t) ∈ Fq [t],
we can factor f (t)OK in polynomial time. Since OK is a Dedekind domain, this holds for
all f (t) not dividing the discriminant, so there are only finitely many f (t) that need to
be factored as a one-time operation.
As in number fields, the main obstacle is generating the norm g(t) of I with probability
proportional to the number of ideals with this norm. Once we can do this, then we can
use our ability to factor g(t) over OK and solve the corresponding subset sum problem to
generate an ideal of OK uniformly at random with its factorization.
Let g ∈ Fq [t]. Then we can consider g to be a number written base q by looking at
its coefficients. Let n(g) denote this number. Let D(g) denote the number of ideals in
OK with norm g. For any number n ∈ Z, using its q-ary expansion, we can form a
corresponding element β(n) ∈ Fq [t].
We present the following algorithm for generating a norm g ∈ Fq [t] with probabiity
proportional to D(g). It is virtually identical to the algorithm above, except that our
concept of primality of a number p is replaced by primality of the corresponding element
β(p) ∈ Fq [t].
Algorithm 3. Input: A positive integer N.
Output: A random element in Fq [t] with degree at most N. The polynomial is generated
according to the distribution of norms of ideals in OK with degree of their norm at most
N.
1. Generate ⌊ d2 ⌋ lists of integers as follows. For b ∈ {1, 3, 5, . . . , d − 1}, generate a list
q N +1 > s1,b ≥ s2,b ≥ . . . ≥ slb ,b = 1. We take s1,b ∈ {1, d + b, 2d + b, . . . , kd + b}, where
1
d
s1,b = 1 with probability
and is any other element with probability
. Take
kd + 1
kd + 1
1
si+1,b in {1, d + b, 2d + b, . . . , si,b} as 1 with probability
and any other element
si,b − b + 1
d
.
with probability
si,b − b + 1
2. Let r be the product of the si,b such that β(si,b ) is prime.
3. For all integers p between 1 and d such that β(p) is prime, do the following: Multiply
p−1
r by p with probability
, and continue to multiply by p with this probability until
p
your first failure.
ψ(r)D(β(r))
4. If r < q N +1 , return β(r) with probability Md (r) Ω (α(r)) N .
d d
q
5. If you did not return β(r), go to step 1.
Let n denote the residue of n mod d. We define ψ(r) as follows. Factor β(r) in to primes
gi . For each gi such that n(gi ) > d, replace the factor of n(gi ) with n(gi ) − n(gi ) + 1. Note
that ψ(r) ≤ r and β(r), β(ψ(r)) share the same factors gi such that 2 ≤ n(gi ) ≤ d.
We define Md (r) by:
vn(f ) (r)
Y
1
α
Md (r) =
n(f ) − 1
f ∈Fq [t], f is prime
2≤n(f )<d
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
8
−1
d + N⌊log q⌋ − 1
α=
N⌊log q⌋
An almost identical argument to the one above shows that this produces r with probability proportional to D(β(r)). Moreover, the algorithm produces a norm with probability
P
g∈Fq [t],deg(g)≤N
qN
D(g)
Y
g∈Fq [t]
2≤n(g)≤d
α
n(g)
Y
g∈Fq [t]
d<n(g)<q N+1
1−
d
n(g) − n(g) + 1
P
+
We first want to analyze g∈Fq ,deg(g)≤N D(g). Let DK
denote the set of effective divisors
of V , where K is the function field of the projective variety V . Note that the zeta function
for K can be written as:
X
1
ζK (s) =
deg(D)s
q
+
D∈DK
If we let an denote the number of g ∈ Fq [t] of degree n, then this becomes:
X an
ζK (s) =
q ns
n≥0
Since ζK (s) converges absolutely for Re(s) > 1 and has a simple pole at s = 1, the
analogue of the Wiener-Ikehara Tauberian theorem for function fields implies that:
X
an = Θ(q N )
n≤N
Let ZN be given by:
ZN =
Y
d<n(g)<q N+1
1−
d
n(g) − n(g) + 1
g prime
Note that the expected number of trials of the algorithm is:
O(α−d ZN−1
Y
n(g))
2≤n(g)≤d
g prime
By an analogous argument to the above, we have at most d factors of α which contribute
at most (2N log(q))d−1 to the expecte number of trials, while the last product in the
estimate contributes an amount bounded by dd . Therefore, it suffices to bound ZN−1 .
Using a version of the prime number theorem for Fq [x], one can prove in an analogous
way to the proof above that ZN−1 contributes O((N log(q))2d ) iterations in expected value.
2
Therefore, we require O((N log q)d +d ) trials before success in expectation. Each of the
2
d
lists has expected length O(N log q), so we require O((N log q)d +d+1 ) primality tests.
2
Note that this gives us a randomized algorithm that is polynomial in the logarithm of the
size of the input (as there are q N possible norms of the ideal to generate). Since primality
testing can be performed in time that is O((N log q)6 ), this gives us an overall runtime
2
that is O((N log q)d +d+7 ).
GENERATING RANDOM FACTORED IDEALS IN NUMBER FIELDS
9
6. Further Work
The algorithm above uses similar ideas to Kalai’s algorithm. Bach’s algorithm for
generating factored integers runs in fewer primality tests, and so one could ask whether the
ideas of Bach could be adapted to this setting in order to reduce the number of primality
tests required. In general, one could determine ways to make the above algorithm run
faster, in particular by reducing the number of primality tests required.
The algorithms above work for a fixed number or function field. In particular, their
runtime is polynomial treating the degree of the field extension as constant. It remains
an open question whether there is an algorithm for generating random factored ideals
that runs in polynomial time, irrespective of the degree of the field. The methods above
would have to be altered significantly to do so, due to their exponential dependence on
the degree d.
There are other generalizations of this problem that could be considered. Due to the
natural way in which principal ideals arise in number fields, one could ask for a variant
of this algorithm that generates principal ideals uniformly at random. Clearly we could
use the above algorithm to generate an ideal uniformly at random, and then only accept
if it is principal. Since this occurs with probability 1/h, where h is the class number, this
would result in a polynomial time algorithm provided we had a polynomial time way to
recognize principal ideals. Unfortunately, this is a difficult question in arbitrary number
fields, as there are no known polynomial time algorithms to detect whether an ideal in an
arbitrary number field is principal.
References
1. Manindra Agrawal, Neeraj Kayal, and Nitin Saxena, Primes is in p, Annals of mathematics (2004),
781–793.
2. Eric Bach, Analytic methods in the analysis and design of number-theoretic algorithms, MIT press
Cambridge, 1985.
3. Eric Bach, How to generate factored random numbers, SIAM Journal on Computing 17 (1988), no. 2,
179–193.
4. Alexandre Chistov, Efficient factoring polynomials over local fields and its applicatons, Proceedings
of the International Congress of Mathematicians, Kyoto, Japan, vol. 2, Springer Verlag, August 1990,
pp. 1509–1519.
5. Thomas H. Cormen, Clifford Stein, Ronald L. Rivest, and Charles E. Leiserson, Introduction to
algorithms, 2nd ed., McGraw-Hill Higher Education, 2001.
6. Shikao Ikehara, An extension of landau’s theorem in the analytical theory of numbers, Journal of
Mathematics and Physics 10 (1931), no. 1-4, 1–12.
7. Adam Kalai, Generating random factored numbers, easily, J. Cryptology 16 (2003), 287–289.
8. Donald Knuth, Seminumerical algorithms, vol. 2, Addison-Wesley, Reading, MA, 1997.
9. Noah Lebowitz-Lockard and Carl Pomerance, Generating random factored Gaussian integers, easily,
Math. Comp. 85 (2016), 503–516.
10. Hendrik W. Lenstra, Jr., Primality testing with gaussian periods, Proceedings of the 22Nd Conference
Kanpur on Foundations of Software Technology and Theoretical Computer Science (London, UK,
UK), FST TCS ’02, Springer-Verlag, 2002, pp. 1–.
11. James S. Milne, Algebraic number theory, 2013, Available at www.jmilne.org/math/, p. 161.
12. J Barkley Rosser, Lowell Schoenfeld, et al., Approximate formulas for some functions of prime numbers, Illinois Journal of Mathematics 6 (1962), no. 1, 64–94.
University of Wisconsin-Madison, Dept. of Mathematics
E-mail address: [email protected]
| 8cs.DS
|
arXiv:1403.7793v1 [math.OC] 30 Mar 2014
True Global Optimality of the Pressure Vessel Design
Problem: A Benchmark for Bio-Inspired Optimisation
Algorithms
Xin-She Yang, Christian Huyck, Mehmet Karamanoglu, Nawaz Khan
School of Science and Technology, Middlesex University,
The Burroughs, London NW4 4BT, UK.
Abstract
The pressure vessel design problem is a well-known design benchmark for validating bioinspired optimization algorithms. However, its global optimality is not clear and there has been
no mathematical proof put forward. In this paper, a detailed mathematical analysis of this
problem is provided that proves that 6059.714335048436 is the global minimum. The Lagrange
multiplier method is also used as an alternative proof and this method is extended to find the
global optimum of a cantilever beam design problem.
Citation detail: Xin-She Yang, Christian Huyck, Mehmet Karamanoglu, Nawaz Khan, rue
Global Optimality of the Pressure Vessel Design Problem: A Benchmark for Bio-Inspired Optimisation Algorithms, Int. J. Bio-Inspired Computation, vol. 5, no. 6, pp. 329-335 (2013).
1
Introduction
Engineering optimization is often non-linear with complex constraints, which can be very challenging
to solve. Sometimes, seemingly simple design problems may in fact be very difficult indeed. Even
in very simple cases, analytical solutions are usually not available, and researchers have struggled
to find the best possible solutions. For example, the well-known design benchmark of a pressure
vessel has only four design variables [[Cagnina et al. (2008), Gandomi et al. (2013), Yang (2010)]];
however, the global optimum solution for pressure vessel design benchmark is still unknown to
the research community, despite a large number of attempts and studies, i.e. [Annaratone (2007),
Deb and Gene (1997)]. Thus, a mathematical analysis will help to gain some insight into the problem
and thus guide researchers to validate if their solutions are globally optimal. This paper attempts to
provide a detailed mathematical analysis of the pressure vessel problem and find its global optimum.
To the best of the author’s knowledge, this is a novel result in the literature.
Therefore, the rest of the paper is organised as follows: Section 2, introduces the basic formulation
of the pressure vessel design benchmark and then highlights the relevant numerical results from the
literature. Section 3 provides an analysis of the global optimum for the problem, whereas Section
4 uses Lagrange multipliers as an alternative method to prove that the analysis in Section 3 indeed
gives the global optimum. Section 5 extends the same methodology to analyse the optimal solution
of another design benchmark: cantilever beam. Finally, we draw our conclusions in Section 6.
2
Pressure Vessel Design Benchmark
Bio-inspired optimization algorithms have become popular, and many new algorithms have emerged
in recent years [[Yang and Gandomi (2012), Che and Cui (2011), Cui et al. (2013), Gandomi et al. (2012),
Yang and Deb (2013), Yang (2013)]]. In order to validate new algorithms, a diverse set of test functions and benchmarks are often used [Gandomi and Yang (2011), Jamil and Yang (2013)]. Among
the structural design benchmarks, the pressure vessel design problem is one of the most widely used.
1
In fact, the pressure vessel design problem is a well-known benchmark for validating optimization algorithms [[Cagnina et al. (2008), Yang (2010)]]. It has four design variables: thickness (d1 ),
thickness of the heads (d2 ), the inner radius (r) and the length (L) of the cylindrical section. The
main objective is to minimize the overall cost, under the nonlinear constraints of stresses and yield
criteria. The thickness can only take integer multiples of 0.0625 inches.
This optimization problem can be written as
minimize f (x) = 0.6224d1rL + 1.7781d2r2
+ 3.1661d21 L + 19.84d21 r,
(1)
subject to
g1 (x) = −d1 + 0.0193r ≤ 0
g2 (x) = −d2 + 0.00954r ≤ 0
The simple bounds are
(2)
3
g3 (x) = −πr2 L − 4π
3 r + 1296000 ≤ 0
g4 (x) = L − 240 ≤ 0.
0.0625 ≤ d1 , d2 ≤ 99 × 0.0625, 10.0 ≤ r, L ≤ 200.
2
(3)
3
Table 1: Summary of main results. [Results marked with ∗ are not valid, see text].
Authors
Results
Authors
Results
[Coello (2000a)]
6288.745
[Lee and Geem (2005)]
7198.433
[Li and Chou (1994)]
7127.3
[Cai and Thierauf (1997)]
7006.931
[Cagnina et al. (2008)]
6059.714 [Li and Chang (1998)]
7127.3
[Cao and Wu (1999)]
7108.616
[Hu et al. (2003)]
6059.131∗
[He et al. (2004)]
6059.714 [Coello and Mezura Montes (2001)] 6059.946
[Huang et al. (2007)]
6059.734
[He and Wang (2006)]
6061.078
[Litinetskiand Abramzon (1998)] 7197.7
[Coello (2000b)]
6263.793
[Sandgren (1998)]
7980.894
[Kannan and Kramer (1994)]
7198.042
[Akhtar et al. (2002)]
6171
Yun [Yun (2005)]
7198.424
[Tsai et al. (2002)]
7079.037
[Cao and Wu (1997)]
7108.616
[Deb and Gene (1997)]
6410.381
[Coello (1999)]
6228.744
[Montes et al. (2007)]
6059.702∗
[Parsopoulos and Vrahatis (2005)]
6544.27
[Shih and Lai (1995)]
7462.1
[Kaveh and Talatahari (2010)]
6059.73
[Sandgren (1990)]
8129.104
[Santos Ceolho (2010)]
6059.714
[Wu and Chow (1995)]
7207.494
[Rat and Liew (2003)]
6171
[Zhang and Wang (1993)]
7197.7
[Coello and Cortés (2004)]
6061.123
[Joines and Houck (1994)]
6273.28
[Michalewicz and Attia (1994)]
6572.62
[Hadj-Alouane and Bean (1997)] 6303.5
[Fu et al. (1991)]
8048.6
[Yang and Gandomi (2012)]
6059.714 [Gandomi et al. (2013)]
6059.714
This is a mixed-integer problem, which is usually challenging to solve. However, there are extensive studies in the literature, and details can be found in several good survey papers [[Thanedar and Vanderplaats (1995),
Gandomi and Yang (2011)]]. The main results are summarized in Table 1. It is worth pointing out
that some results are not valid and marked with * in the footnote. These seemingly lower results
actually violated some constraints and/or used different limits.
As it can be seen from this table, the results vary significantly from the highest value of 8129.104
by Sandgren [Sandgren (1998)] to the lowest value of 6059.714 by a few researchers [[Cagnina et al. (2008),
Santos Ceolho (2010), He et al. (2004), Gandomi et al. (2013), Gandomi et al. (2011), Yang and Gandomi (2012)]].
However, nobody is sure that 6059.714 is the globally optimal solution for this problem.
The best solution by [Gandomi et al. (2013)] and [Yang and Gandomi (2012)] is
f∗ = 6059.714,
(4)
x∗ = (0.8125, 0.4375, 42.0984, 176.6366).
(5)
with
The rest of the paper analyses this problem mathematically and proves that this solution is
indeed near the global optimum and concludes that the true globally minimal solution is fmin =
6059.714335048436 at
x∗ = (0.8125, 0.4375,
42.0984455958549, 176.6365958424394).
3
(6)
Analysis of Global Optimality
As all the design variables must have positive values and f is monotonic in all variables, the minimization of f requires the minimization of all the variables if there is no constraint. As there
are 4 constraints, some of the constraints may become tight or equalities. As the range of L is
10 ≤ L ≤ 200, the constraint g4 automatically satisfies L ≤ 240 and thus becomes redundant, which
means that the upper bound for L is
L ≤ 200.
(7)
This is a mixed integer programming problem, which often requires special techniques to deal
with the integer constraints. However, as the number of combinations of d1 and d2 is not huge (just
1002 = 10, 000), it is possible to go through all the cases for d1 and d2 , and then focus on solving
the optimization problems in terms of r and L.
The first two constraints are about stresses. In order to satisfy these conditions, the hoop stresses
d1 /r and d2 /r should be as small as possible. This means that r should be reasonably large. For
any given d1 and d2 , the first two constraints become
r≤
d2
.
0.00954
(8)
d1
d2
,
}.
0.0193 0.00954
(9)
d1
,
0.0193
r≤
So the upper bound or limit for r becomes
Ur = min{
The above argument that r should be moderately high, may imply that one of the first two
constraints can become tight, or an equality.
The third constraint g3 can be rewritten
πr2 L +
4π 3
r ≥ K,
3
K = 1296000.
(10)
In fact, this is essentially the requirement that the volume of the pressure vessel must be greater
than a fixed volume. This provides the lower boundary in the search domain of (r, L).
4
Since f (r, L) is monotonic in r and L, the global solution must be on the lower boundary for any
given d1 and d2 . In other words, the inequality g3 becomes an equality
πr2 L +
4π 3
r = K.
3
(11)
Using equation (10), with L ≤ 200, r can be derived using Newton’s method (or an online
polynomial root calculator). The only positive root implies that
r ≥ 40.31961872409872 = r1 .
(12)
r ≤ 65.22523261350128 = r2 .
(13)
Similarly, L ≥ 10 means that
So the true value of r must lie in the interval of [r1 , r2 ].
From the first inequality with r = r1 , we have
d1 ≥ 0.7782.
(14)
d2 ≥ 0.3846.
(15)
The second inequality gives
As both d1 and d2 must be integer multiples I and J, respectively, of d = 0.0625, the above two
inequalities mean
0.7782
0.3846
I=⌈
⌉ = 13, J = ⌈
⌉ = 7.
(16)
0.0625
0.0625
In other words, we have
d1 ≥ 13d = 0.8125, d2 ≥ 7d = 0.4375.
(17)
From the objective function (Eq. 1), both d1 and d2 should be as small as possible, so as to
get the minimum possible f . This means that the global minimum will occur at d1 = 0.8125 and
d2 = 0.4375.
Now the objective function with these d1 and d2 values can be written as
f (r, L) = 0.5057rL + 0.77791875r2
+ 2.090120703125L + 13.0975r.
(18)
As d1 = 0.8125 and d2 = 0.4375, the first inequalities (g1 and g2 ) will give an upper bound of r
R∗ = min{
d1
d2
,
}
0.0193 0.00954
= min{42.0984455958549, 45.859538784067}
= 42.0984455958549.
(19)
Again from the objective function, which is monotonic in terms of r and L, the optimal solution
should occur at the two extreme ends of the boundary governed by Eq. (11).
The one end at r = R∗ gives
L∗ =
4R∗
K
−
= 176.6365958424394.
πR∗2
3
(20)
This is the point for the global optimum with
fmin = 6059.714335048436.
(21)
The other extreme point is at r′ = 40.31961872409872 and L = 200, which leads to an objective
value of
f ′ = 6288.67704565344,
(22)
and clearly is not the global optimum.
5
4
Method of Lagrange Multipliers
The optimal solution (21) can alternatively be proved by solving the following constrained problem
with one equality because all of the upper bounds or inequalities are automatical satisfied. By
minimizing d1 , and d2 , the objective function becomes:
minimize f (r, L) = 0.5057rL + 0.77791875r2
+ 2.090120703125L + 13.0975r.
subject to
ge = πr2 L +
4π 3
r − K = 0,
3
(23)
(24)
with the simple bounds
40.31961872409872 ≤ r ≤ 42.098445595854919,
10 ≤ L ≤ 176.6365958424394.
(25)
To avoid writing long numbers, let us define
a = 0.5057,
b = 0.77791875,
c = 2.090120703125,
d = 13.0975.
(26)
We have
minimize f (r, L) = arL + br2 + cL + dr.
(27)
This problem can be solved by the Lagrange multiplier method, and we have
minimize φ = f + λge ,
(28)
where λ is the Lagrange multiplier.
The optimum should occur when
∂f
∂ge
∂φ
=
+λ
∂r
∂r
∂r
= aL + 2br + d + λ(2πrL + 4πr2 ) = 0,
∂f
∂ge
∂φ
=
+λ
= ar + c + λ(πr2 ) = 0,
∂L
∂L
∂L
∂φ
4π 3
= πr2 L +
r − K = 0.
∂λ
3
Now we have three equations for three unknowns
aL + 2br + d + λ(2πrL + 4πr2 ) = 0,
ar + c + λ(πr2 ) = 0,
3
πr2 L + 4πr
3 − K = 0.
(29)
(30)
(31)
(32)
The equation in the middle gives
ar + c
,
πr2
Substituting this, together with L = K/(2r2 ) − 4r/3, into the first equation, we have
λ=−
(33)
4πr3
) + 2bπr4 + πdr3
3
4πr3
) = 0,
(34)
− (ar + c)(2K +
3
which is a quartic equation with four roots in general. The only feasible solution within [r1 , r2 ]
is 42.098445595854919, which corresponds to L = 176.6365958424394. This solution is indeed the
global best solution as given in (21).
a(Kr −
6
5
Cantilever Beam Design Benchmark
Another widely used benchmark for validating bio-inspired algorithms is the design optimization
of a cantilever beam, which is to minimize the overall weight of a cantilever beam with square
cross sections [[Fleury and Braibant (1986), Gandomi et al. (2013), Gandomi et al. (2011)]]. It can
be formulated as
minimize f (x) = 0.0624(x1 + x2 + x3 + x4 + x5 ),
(35)
subject to the inequality
g(x) =
7
1
61 37 19
+ 3 + 3 + 3 + 3 − 1 ≤ 0.
3
x1
x2
x3
x4
x5
(36)
The simple bounds/limits for the five design variables are
0.01 ≤ xi ≤ 100,
i = 1, 2, ..., 5.
(37)
Since the objective f (x) is linear in terms of all design variables, and g(x) encloses a hypervolume,
it can be thus expected that the global optimum occurs when the inequality becomes tight. That
is, the inequality becomes an equality
g(x) =
61 37 19
7
1
+ 3 + 3 + 3 + 3 − 1 = 0.
x31
x2
x3
x4
x5
(38)
For ease of analysis, we rewrite the above equation as
g(x) =
5
X
ai
3 − 1 = 0,
x
i=1 i
(39)
where
a = (a1 , a2 , a3 , a4 , a5 ) = (61, 37, 19, 7, 1).
(40)
Hence, the cantilever beam problem becomes
minimize f (x) = k
5
X
xi ,
k = 0.0624,
(41)
i=1
subject to
g(x) =
5
X
ai
− 1 = 0.
x3
i=1 i
(42)
Using the method of Lagrange multipliers, we have
Minimize φ = f + λg
=k
n
X
ai
−
1
.
xi + λ
x5
i=1 i
i=1
5
X
(43)
Then, the optimality conditions give
ai
∂φ
= k + λ(−3) 4 = 0,
∂xi
xi
i = 1, 2, ..., 5,
(44)
5
∂φ X ai
− 1 = 0.
=
∂λ
x3
i=1 i
(45)
From Eq. (44), we have
x4i =
3λai
,
k
or
7
kxi
ai
=
.
x3i
3λ
(46)
Substituting it into Eq. (45), we have
5
5
X
k X
kxi
xi ) − 1 = 0.
)−1 =
(
(
3λ
3λ i=1
i=1
(47)
After rearranging and using results from (46), now we have
5
3λ X 3λai 1/4
,
=
k
k
i=1
(48)
which is a nonlinear equation for λ. However, it is straightforward to find that
λ ≈ 0.4466521202,
(49)
which leads to the optimal solution
x∗ = (6.0160159, 5.3091739,
4.4943296, 3.5014750, 2.15266533),
(50)
with the minimum
fmin (x∗ ) = 1.339956367.
(51)
This is the global optimum. However, the authors have not seen any studies that have found
this solution in the literature. Slightly higher values have been found by cuckoo search and other
methods [[Chickermane and Gea (1996), Gandomi et al. (2013)]]. The best solution found so far by
[Gandomi et al. (2013)] is
xbest = (6.0089, 5.3049, 4.5023, 3.5077, 2.1504),
(52)
fbest = 1.33999,
(53)
and
which is near this global optimum.
The above mathematical analysis can be very useful to guide future validation of new optimization
methods when the above design benchmarks are used.
6
Conclusions
Pressure vessel design problem is a well-tested benchmark that has been used for validating optimization algorithms and their performance. We have provided a detailed mathematical analysis and
obtained its global optimality. We have also used the method of Lagrange multipliers to double-check
that the obtained optimum is indeed the global optimum for the pressure vessel design problem. By
using the same methodology, we also analysed the design optimization of a cantilever beam.
However, it is worth pointing out that the method of Lagrange multipliers is only valid for optimization problems with equalities or when an inequality becomes tight. For general nonlinear
optimization problems, we have to use the full Karush-Kuhn-Tucker (KKT) conditions to analyze
their optimality [[Yang (2010)]], though such KKT can be extremely challenging to analyse in practice.
Even for design problems with only a few design variables, an analytical solution will provide
greater insight into the problem and thus can act as better benchmarks for validating new optimization algorithms. Further work can focus on the analysis of other nonlinear design benchmarks.
8
References
[Akhtar et al. (2002)] Akhtar, S., Tai, K., Ray, T. (2002). A socio-behavioural simulation model for
engineering design optimization, it Engineering Optmization, 34(4), pp. 341–354.
[Annaratone (2007)] Annaratone, D. (2007) Pressure Vessel Design, Springler-Verlag Berlin Heidelberg.
[Cagnina et al. (2008)] Cagnina L. C., Esquivel S. C., and Coello C. A., (2008). Solving engineering
optimization problems with the simple constrained particle swarm optimizer, Informatica, 32,
319-326.
[Cai and Thierauf (1997)] Cai, J., Thierauf, G., (1997). Evolution strategies in engineering optimization, Eng Optimization, 29(1), pp. 177–199.
[Cao and Wu (1997)] Cao, Y. J., Wu, Q. H., (1997). Mechanical design optimization by mixed
variable evolutionary programming. In: Proceedings of the 1997 International Conference on
Evolutionary Computation, Indianapolis, pp. 443–446.
[Cao and Wu (1999)] Cao, Y. J., Wu, Q. H., (1999). A mixed variable evolutionary programming
for optimization of mechanical design, Int J Eng Intel Syst Elect Eng Commun, 7(2), pp. 77–82.
[Che and Cui (2011)] Che, Z. H., Cui, Z. H., (2011). Unbalanced supply chain design using the analytic network process and a hybrid heuristic-based algorithm with balance modulating mechanism, Int. J. Bio-inspired Computation, 3(1), 56–66.
[Chickermane and Gea (1996)] Chickermane, H., Gea, H. C., (1996). Structural optimization using
a new local approximation method, Int J Numer Method Eng, 39, pp.829–846.
[Santos Ceolho (2010)] dos Santos Coelho, L, (2010). Gaussian quantum-behaved particle swarm
optimization approaches for constrained engineering design problems, Expert. Syst. Appl., 37(2),
pp. 16761683.
[Coello (2000a)] Coello C. A. C., (2000a). Use of a self-adaptive penalty approach for engineering
optimization problems. Comput. Ind., 41(2), pp. 113-127.
[Coello (2000b)] Coello, C. A. C., (2000b). Constraint-handling using an evolutionary multiobjective
optimization technique, Civil Engrg Environ Syst, 17, pp. 319–346.
[Coello (1999)] Coello C. A. C., (1999). Self-adaptive penalties for GA based optimization, Proc.
Congr. Evol. Comput., 1, pp. 573–580.
[Coello and Cortés (2004)] Coello, C. A. C., Cortés, N. C., (2004). Hybridizing a genetic algorithm
with an artificial immune system for global optimization, Engineering Optmization, 36(5), pp.
607634.
[Coello and Mezura Montes (2001)] Coello, C. A. C., Mezura Montes, E., (2001). Use of dominancebased tournament selection to handle constraints in genetic algorithms. In: Intelligent Engineering Systems through Artificial Neural Networks (ANNIE2001), 11, ASME Press, St. Louis,
pp. 177–182.
[Cui et al. (2013)] Cui, Z. H., Fan S. J., Zeng J. C., Shi, Z. Z., (2013). Artificial plant optimisation
algorithm with three-period photosynthesis, Int. J. Bio-Inspired Computation, 5(2), 133–139.
[Deb and Gene (1997)] Deb, K., Gene, A. S., (1997). A robust optimal design technique for mechanical component design. in: Evolutionary algorithms in engineering applications. Springer-Verlag,
Berlin, pp. 497–514.
[Fleury and Braibant (1986)] Fleury, C., Braibant, V., (1986). Structural optimization: a new dual
method using mixed variables, Int J Numer Meth Eng, 23, pp. 409–428.
9
[Fu et al. (1991)] Fu, J., Fenton, R. G., Cleghorn, W. L., (1991). A mixed integer-discrete continuous programming method and its application to engineering design optimization, Engeering
Optmization, 17, pp. 263–280.
[Gandomi et al. (2013)] Gandomi, A. H., Yang, X.-S., and Alavi, A. H., (2013). Cuckoo search
algorithm: a metaheuristic approach to solve structural optimization problems, Engineering
with Computers, 29(1), pp. 17-35.
[Gandomi et al. (2011)] Gandomi, A. H., Yang, X. S., Alavi, A. H., (2011). Mixed variable structural
optimization using Firefly Algorithm, Computers & Structures, 89 (23-24), 2325–2336.
[Gandomi and Yang (2011)] Gandomi, A. H. and Yang, X. S., (2011). Benchmark problems in structural optimization, in: Computational Optimization, Methods and Algorithms, Studies in Computational Intelligence (Eds. Koziel S. and Yang X. S.), Springer, Heidelberg, vol. 356, pp.
259–281.
[Gandomi et al. (2012)] Gandomi, A. H., Yang, X. S., Talatahari, S., Deb, S., (2012). Coupled
eagle strategy and differential evolution for unconstrained and constrained global optimization,
Computers & Mathematics with Applications, 63(1), 191–200.
[Hadj-Alouane and Bean (1997)] Hadj-Alouane, A. B., Bean, J. C., (1997). A genetic algorithm for
the multiple-choice integer program. Oper Res, 45, pp. 92–101.
[He et al. (2004)] He, S., Prempain, E., Wu, Q. H., (2004). An improved particle swarm optimizer
for mechanical design optimization problems, Engineering Optimization, 36(5), pp. 585–605.
[He and Wang (2006)] He, Q., Wang, L., (2006). An effective co-evolutionary particle swarm optimization for engineering optimization problems, Eng Appl Artif Intel, 20, pp.89–99.
[Hu et al. (2003)] Hu, X., Eberhart, R. C., Shi, Y. (2003). Engineering optimization with particle
swarm. In: Proc. 2003 IEEE Swarm Intelligence Symposium, pp. 53–57.
[Huang et al. (2007)] Huang, F. Z., Wang, L., He, Q. (2007). An effective co-evolutionary differential
evolution for constrained optimization, Appl. Math. Comput., 186, pp. 340–356.
[Jamil and Yang (2013)] Jamil, M., and Yang, X. S., (2013). A litrature survey of benchmar functions for global optimisation problems, Int. J. of Mathematical Modelling and Numerical Optimisation, 4(2), 150–194.
[Joines and Houck (1994)] Joines, J., Houck, C. (1994). On the use of non-stationary penalty functions to solve nonlinear constrained optimization problems with GAs. In: Proceedings of the
first IEEE Conference on Evolutionary Computation, Orlando, Florida. D. Fogel (ed.). IEEE
Press, pp 579–584
[Kannan and Kramer (1994)] Kannan, B. K., Kramer, S. N., (1994). An augmented Lagrange multiplier based method for mixed integer discrete continuous optimization and its applications to
mechanical design. J Mech Des. Trans., 116, pp. 318–320
[Kaveh and Talatahari (2010)] Kaveh, A., Talatahari, S., (2010). An improved ant colony optimization for constrained engineering design problems, Engineering Computations, 27(1), pp.155–
182.
[Lee and Geem (2005)] Lee K. S., Geem Z. W. (2005), A new meta-heuristic algorithm for continuous engineering optimization: harmony search theory and practice, Comput. Methods Appl.
Mech. Eng., 194, pp. 3902–3933.
[Li and Chou (1994)] Li, H.-L., Chou C.-T., (1994). A global approach for nonlinear mixed discrete
programming in design optimization. Engineering Optimization, 22, pp. 109–122.
10
[Li and Chang (1998)] Li, H. L., Chang, C. T., (1998). An approximate approach of global optimization for polynomial programming problems, Eur J. Oper Res, 107(3), pp.625–632.
[Litinetskiand Abramzon (1998)] Litinetski, V. V., Abramzon, B. M., (1998). Multistart adaptive
random search method for global constrained optimization in engineering applications, Engineering Optimization, 30(2), pp. 125–154.
[Michalewicz and Attia (1994)] Michalewicz, Z., Attia, N. (1994). Evolutionary optimization of constrained problems. Proceedings of the 3rd Annual Conference on Evolutionary Programming,
World Scientific, pp. 98–108.
[Montes et al. (2007)] Montes, E. M., Coello, C. A. C., Velázquez-Reyes, J, Muñoz-Dávila, L.,
(2007). Multiple trial vectors in differential evolution for engineering design, Engineering Optimization, 39(5), pp. 567–589.
[Parsopoulos and Vrahatis (2005)] Parsopoulos, K. E., Vrahatis, M. N.,. (2005). Unified particle
swarm optimization for solving constrained engineering optimization problems. In: Lecture
Notes in Computer Science (LNCS), 3612, pp. 582–591.
[Rat and Liew (2003)] Ray, T., Liew, K., (2003). Society and civilization: An optimization algorithm based on the simulation of social behavior, IEEE Trans Evol Comput, 7(4), pp.386–396.
[Sandgren (1998)] Sandgren, E., (1988). Nonlinear integer and discrete programming in mechanical
design, Proceedings of the ASME Design Technology Conference, Kissimine, FL, pp. 95–105.
[Sandgren (1990)] Sandgren, E. (1990). Nonlinear integer and discrete programming in mechanical
design optimization, J Mech Design, 112(2), pp. 223–229.
[Shih and Lai (1995)] Shih, C. J., Lai, T. K., (1995). Mixed-discrete fuzzy programming for nonlinear engineering optimization, Engineering Optimization, 23(3), pp. 187–199.
[Thanedar and Vanderplaats (1995)] Thanedar, P. B., Vanderplaats, G. N., (1995). Survey of discrete variable optimization for structural design, Journal of Structural Engineering ASCE, 121
(2), 301–306.
[Tsai et al. (2002)] Tsai, J.-F., Li, H.-L., Hu, N.-Z., (2002). Global optimization for signomial discrete programming problems in engineering design, Engineering Optmization, 34(6), pp. 613–
622.
[Wu and Chow (1995)] Wu, S. J., Chow, P. T., (1995). Genetic algorithms for nonlinear mixed
discrete-integer optimization problems via meta-genetic parameter optimization, Engineering
Optimization, 24, pp. 137–159.
[Yang (2010)] Yang X. S., (2010). Engineering Optimisation: An Introduction with Metaheuristic
Applications, John Wiley and Sons.
[Yang and Gandomi (2012)] Yang, X. S. and Gandomi, A. H., (2012). Bat algorithm: a novel apporach for global engineering optimization, Engineering Computations, 29(5), pp. 464–483.
[Yang (2013)] Yang, X. S., (2013). Multiobjective firefly algorithm for continuous optimization, Engineering with Computers, 29(2), 175–184.
[Yang and Deb (2013)] Yang, X. S. and Deb, S., (2013). Multiobjective cuckoo search for design
optimization, Computers & Operations Research, 40(6), 1616–1624.
[Yun (2005)] Yun, Y. S., (2005). Study on Adaptive Hybrid Genetic Algorithm and Its Applications
to Engineering Design Problems, Waseda University, MSc Thesis.
[Zhang and Wang (1993)] Zhang, C., Wang, H. P., (1993). Mixed-discrete nonlinear optimization
with simulated annealing. Engineering Optmization, 17(3), pp. 263280.
11
| 9cs.NE
|
Multi-category Angle-based Classifier Refit
Guo Xian Yau1 and Chong Zhang
arXiv:1607.05709v1 [math.ST] 19 Jul 2016
1
2
Department of Mathematics
University of Waterloo
2
Amazon.com, Inc.
Abstract
Classification is an important statistical learning tool. In real application, besides high
prediction accuracy, it is often desirable to estimate class conditional probabilities for new
observations. For traditional problems where the number of observations is large, there exist
many well developed approaches. Recently, high dimensional low sample size problems are
becoming increasingly popular. Margin-based classifiers, such as logistic regression, are well
established methods in the literature. On the other hand, in terms of probability estimation,
it is known that for binary classifiers, the commonly used methods tend to under-estimate the
norm of the classification function. This can lead to biased probability estimation. Remedy
approaches have been proposed in the literature. However, for the simultaneous multicategory
classification framework, much less work has been done. We fill the gap in this paper. In
particular, we give theoretical insights on why heavy regularization terms are often needed
in high dimensional applications, and how this can lead to bias in probability estimation. To
overcome this difficulty, we propose a new refit strategy for multicategory angle-based classifiers.
Our new method only adds a small computation cost to the problem, and is able to attain
prediction accuracy that is as good as the regular margin-based classifiers. On the other hand,
the improvement of probability estimation can be very significant. Numerical results suggest
that the new refit approach is highly competitive.
1
1
Introduction
Classification is one of the founding pillars in the statistical learning literature. It has achieved many
successes on real applications in various scientific disciplines, such as artificial intelligence, cancer
research, and econometrics. Given a training data set with both predictors and labels observed,
one can obtain a classifier for future prediction. Typical examples of classification methods include
logistic regression, Fisher’s linear discriminant analysis, nearest neighbours, and many more. See
Hastie et al. (2009) for a comprehensive review of existing methods.
For classification problems, the primary goal is to train the classifier with a high prediction
accuracy. A classifier is optimal if its corresponding prediction error rate achieves the Bayes error
rate, i.e., the smallest error rate among all possible classifiers. Besides prediction accuracy, class
probability estimation is also very important in practice, especially for non-separable classification
problems. In particular, when the density of two or more classes overlaps, there are random errors
involved in class label prediction. In this case, class probability estimation can provide practitioners
with valuable information on the confidence of classifying one subject into a certain category. For
example, given the prediction that the stock market index would rise (as opposed to fall), the
future investment strategies can be quite different if the estimated probabilities are 0.51 and 0.99
respectively.
In practice, if the dimension of the classification problem is low, traditional classifiers such as
nearest neighbours are known to work well. Furthermore, class probability estimation can also be
very accurate. However, recent technology development has enabled scientists to collect big data
sets with very high dimensions. For example, microarray data can measure the expression levels of
millions of genes simultaneously, whereas the number of observations is often limited. In this case,
traditional learning methods are known to suffer from the potential danger of overfitting. That
is, even if the classifier has a high prediction accuracy on the training data set, its generalization
error on future testing data sets can be very large. To overcome this difficulty, many new classifiers
have been proposed in the literature. For example, Cai and Liu (2011) proposed a generalization
of Fisher’s LDA method to handle high dimensional problems, and Lin et al. (2000) suggested the
penalized logistic regression.
In this paper, we focus on margin-based classification methods. This family was originally introduced in the machine learning literature, and has become increasingly popular partly because of
their ability to handle high dimensional classification problems. A typical margin-based classifier
2
solves an optimization problem in the loss + regularization form. Here the loss term is used to measure the goodness of fit of the classifier on the training data set. The regularization term prevents
overfitting by imposing penalty on complexity of the functional space, aiming to obtain a classifier
with accurate prediction on future testing data sets. Many existing margin-based classifiers have
been proposed in this framework, including penalized logistic regression (Lin et al., 2000), Support Vector Machines (SVM, Boser et al., 1992; Cortes and Vapnik, 1995), import vector machines
(Zhu and Hastie, 2005), Adaboost in Boosting (Freund and Schapire, 1997; Friedman et al., 2000),
Ψ-learning (Shen et al., 2003), and large margin unified machines (Liu et al., 2011).
Binary margin-based classifiers have been well studied. To address a multicategory classification
problem, one can first consider a sequence of binary margin-based classifiers, such as the oneversus-one and one-versus-rest approaches. Although these methods are simple in concept and
implementation, they have certain drawbacks. For example, the one-versus-one method has a
potential tie-in-vote problem, and the one-versus-rest method can be inconsistent (Liu, 2007).
Therefore, in this paper, we focus on simultaneous margin-based classifiers. In the literature, most
simultaneous margin-based classifiers use k classification functions for a k-class problem. These
methods often employ a sum-to-zero constraint in the corresponding optimization to reduce the
parameter space and to ensure some theoretical properties such as Fisher consistency (also known as
classification calibration, see Bartlett et al., 2006). Many proposed classification methods follow this
procedure. See, for example, Lee et al. (2004), Zhu and Hastie (2005), Wang and Shen (2007), Park
et al. (2010), Liu and Yuan (2011), Zhang and Liu (2013), among others. However, this approach
of using k functions and reducing to k − 1 by the sum-to-zero constraint, despite its simplicity, can
be inefficient. To circumvent this difficulty, Zhang and Liu (2014) proposed a new framework for
simultaneous margin-based classification, namely, the angle-based classification, which only uses
k − 1 functions for classification. More details of angle-based classifiers are given in Section 2.
Zhang and Liu (2014) showed that their method is free of the sum-to-zero constraint, hence for
high dimensional problems, the angle-based method can enjoy faster computational speed and
higher prediction accuracy. In this paper, we employ the angle-based framework for multicategory
classification.
For margin-based classifiers, to estimate class probabilities, one commonly used approach is to
explore their relationship with respect to the theoretical minimizer, and use the obtained classification functions for probability estimation (more details are given in Section 2). Many existing
classifiers in the literature use this method to estimate class probabilities. See, for example, Lin
3
et al. (2000), Park et al. (2010), Liu et al. (2011), and Zhang and Liu (2013), among others. Despite the simplicity of this approach, it does not take into account the effect of the regularization
term on classification function estimation. For binary problems, Zhang et al. (2013) showed that
the regularization terms can lead to severe bias in class probability estimation. In particular, for
high dimensional problems, we often need a heavy penalty in the corresponding optimization to
ensure a good generalization ability of the resulting classifier. However, heavy penalties can introduce significant shrinkage effect on the fitted classification functions. Consequently, the estimated
probabilities would be biased. To alleviate this difficulty, for binary margin-based classifiers, Zhang
et al. (2013) proposed a two-stage refit method. The idea of Zhang et al. (2013) is that despite the
bias in probability estimation, margin-based classifiers can often find near optimal classification
boundaries. Consequently, one can project all observations in the training data set onto the norm
vector of the separating hyperplane, and perform a second step classifier training using the projected coordinates as the only predictor. Because there is only one predictor in the refit step, the
regularization term is no longer needed, thus it does not suffer from the shrinkage effect. Through
numerical examples, Zhang et al. (2013) showed that the refit method works very well for binary
problems, in the sense that the second step classifier has similar prediction accuracy compared with
the original one, and the corresponding probability estimation can be significantly improved.
Despite the progress of obtaining accurate probability estimation for binary margin-based classifiers, the effect of regularization terms on probability estimation for multicategory problems remains
unclear. In this paper, we fill the gap by showing the shrinkage effect of the penalty on probability
estimation for multicategory classification, and propose a new multicategory refit method to alleviate this difficulty. Our contribution is two fold. First, we provide some theoretical insights on why
heavy penalty is often needed for high dimensional problems, and show that the fitted classification
functions would have small norms. This leads to that the estimated probabilities would shrink towards 1/k for a k-class problem. Our second contribution is that we propose a multicategory refit
method for simultaneous margin-based classifiers. Note that for multicategory problems, because
there are more than one separating boundaries, how to perform appropriate projection becomes
more involved. We propose to use the k − 1 classification functions introduced in Zhang and Liu
(2014) as the coordinates for the second step classifier training, and show that our new method can
significantly improve class probability estimation in empirical studies.
The rest of this article is organized as follows. In Section 2, we give a brief review of both
binary and multicategory margin-based classifiers, and introduce how to estimate class conditional
4
probability. In Section 3, we first give some theoretical insights on why heavy penalty is often
needed for high dimensional low sample size problems, then discuss the corresponding consequence
on class probability estimation. Then, we propose our multicategory refit method to alleviate
such difficulties. Numerical examples including simulation and real data analysis in Section 4
demonstrates the usefulness of our new method. Some concluding remarks are given in Section 5.
All technical proofs are collected in the Supplementary Materials.
2
Review of Margin-based Classifiers
For a classification problem, suppose the training data set {(xi , yi ); i = 1, . . . , n} is obtained from
an unknown underlying distribution P (X, Y ), where X ∈ Rp is a vector of predictors, and Y ∈
{+1, −1} is the label for a binary problem. The typical learning goal is to obtain a classifier from the
training data set, such that the future prediction error rate is minimized. In other words, we aim to
find a prediction rule ŷ(·) that minimizes the expected 0-1 loss E[I{ŷ(X) 6= Y }], where I(·) is the
indicator function. Binary margin-based classifiers use a single function f (x) for classification, and
the corresponding prediction rule is ŷ(x) = sign{f (x)}. Consequently, minimizing the expected
0-1 loss is equivalent to minimizing E[I{Y f (X) < 0}]. However, because the indicator function
is non-convex and discontinuous, direct minimization of the empirical 0-1 loss can be difficult. To
overcome this challenge, a common approach is to use a surrogate convex loss function ` in place of
the indicator function. In particular, a typical binary margin-based classifier solves the optimization
n
min
f ∈F
1X
`{yi f (xi )} + λJ(f ),
n
i=1
where F is the functional space of f , J(f ) is the regularization term on f to prevent overfitting, and
λ is a tuning parameter to balance the loss and regularization terms. In practice, an appropriate
choice of λ can be critical for the performance of the resulting classifier. Note that for an observation
(x, y), yf (x), known as the functional margin, is positive if and only if the prediction is correct.
Therefore, many binary margin-based classifiers use a non-increasing loss function ` to encourage
large functional margins yf .
When the number of classes k > 2, it is common to have Y ∈ {1, . . . , k}. In this case, one
cannot use a single function f for classification. In the literature, many simultaneous margin-based
classifiers use k functions f = (f1 , . . . , fk ) for learning, and the corresponding prediction rule is
5
ŷ = argmaxj=1,...,k fj . A typical simultaneous classifier solves the following optimization
n
min
f ∈F
1X
φ{yi , f (xi )} + λJ(f ),
n
i=1
subject to
k
X
fj = 0,
(1)
j=1
where φ is a multicategory margin-based loss function. Here, to reduce the parameter space and to
ensure theoretical properties such as Fisher consistency, the sum-to-zero constraint on f , namely
Pk
j=1 fj (x) = 0 for all x, is often imposed. Different loss functions correspond to different simultaneous margin-based classifiers. For example, the Multicategory SVM (MSVM) proposed by
P
Lee et al. (2004) uses φ(y, f ) = j6=y [1 + fj ]+ , where [u]+ = max(0, u), the MSVM by Liu and
P
Shen (2006) uses φ(y, f ) = j6=y [1 − minj (fy − fj )]+ , and the multicategory large margin unified
P
machine by Zhang and Liu (2013) uses φ(y, f ) = γV (fy ) + (1 − γ) j6=y V (−fj ), where γ ∈ [0, 1]
is a convex combination parameter, and V is the binary large margin unified machine loss function
(see the appendix for more details). Note that these methods all encourage a large value of fyi (xi ),
which can be regarded as the functional margin in multicategory problems.
Recently, Zhang and Liu (2014) pointed out that for a k-class problem, using k classification
functions and the sum-to-zero constraint can be inefficient, and proposed an angle-based classification framework to overcome this challenge. The details of angle-based classifiers are as follows.
Consider a centered simplex with k vertices W = {W 1 , . . . , W k } in Rk−1
(k − 1)−1/2 1
Wj =
−(1 + k 1/2 )/{(k − 1)3/2 }1 + {k/(k − 1)}1/2 e
j = 1,
j−1
2 ≤ j ≤ k,
where 1 is a vector of 1, and ej is a vector with jth element 1 and 0 elsewhere. One can verify
that W defines a symmetric simplex in Rk−1 . Without loss of generality, assume class j is assigned
to W j . Zhang and Liu (2014) proposed to use f (x) = {f1 (x), . . . , fk−1 (x)} : Rp → Rk−1 as
the classification function vector. Each f ∈ Rk−1 defines k angles with respect to W 1 , . . . , W k ,
namely, ∠(W j , f ); j = 1, . . . , k. The prediction rule is ŷ(x) = argminj∈{1,...,k} ∠(W j , f ). Note
that the smaller ∠(W j , f ), the larger the corresponding inner product hW j , f i. Therefore, the
goal of the angle-based classifier is equivalent to maximizing hW y , f (x)i, which can be regarded as
functional margins in the angle-based classification. In particular, an angle-based classifier solves
6
the optimization problem
n
min
f ∈F
1X
`{hW yi , f (xi )i} + λJ(f ),
n
(2)
i=1
where ` is a binary margin-based loss function. The angle-based classifiers are free of the sum-to-zero
constraint, therefore can enjoy a more efficient computational speed. Zhang and Liu (2014) showed
that angle-based classifiers are highly competitive for high dimensional classification problems.
Besides class prediction accuracy, estimation of class probabilities can be very important. For
example, in clinical trials, physicians may have different treatments for patients based on their
estimated probabilities of a certain disease. In the margin-based classification literature, to estimate
the probabilities, a common approach is to use the fitted classification functions. In particular, we
denote by Pj (x) = pr(Y = j | X = x) the class conditional probability for class j. Define the
conditional expected loss
S(x) = E[φ{Y, f (X)} | X = x],
(3)
where φ is a margin-based loss function. Note that if f is one dimensional and one chooses
φ appropriately, (3) can include binary classifiers as special cases. Next, define the theoretical
minimizer of (3) to be
f ∗ (x) = arginf S(x) = arginf E[φ{Y, f (X)} | X = x].
f
(4)
f
For a classification problem with Y ∈ {1, . . . , k}, under mild conditions (e.g., the loss function φ is
differentiable), one can find the relationship between Pj (x) and f ∗ (x), namely, Pj (x) = gj {f ∗ (x)}
for all j. When the fitted classification function f̂ is obtained for a new observation x, one can use
P̂j (x) = gj {f̂ (x)} to estimate the probability of class j for x. For example, in binary margin-based
classification, under mild conditions, one can estimate P+1 (x) by
`0 (−fˆ)/{`0 (−fˆ) + `0 (fˆ)},
(5)
where ` is the corresponding binary margin-based loss function (Zou et al., 2008; Zhang et al., 2013).
For the angle-based classification, Zhang and Liu (2014) showed that the relationship between Pj
7
and f ∗ is such that
`0 (hW j , f ∗ i)−1
,
Pj = Pk
0 (hW , f ∗ i)−1
`
i
i=1
(6)
for all j ∈ {1, . . . , k} and differentiable loss function `. Consequently, one can estimate Pj by
P
P̂j = {`0 (hW j , f̂ i)−1 }/{ ki=1 `0 (hW i , f̂ i)−1 }. Note that Wang et al. (2008) and Wu et al. (2010)
proposed methods for class conditional probability estimation using SVM classifiers.
3
Methodology
In this section, we first explore the effect of the regularization term on class probability estimation
in Section 3.1, then propose a new refit method for multicategory classification in Section 3.2.
3.1
Shrinkage Effect of Probability Estimation in Multicategory Classification
For binary margin-based classifiers, Zhang et al. (2013) observed that the L2 regularization term
can have a shrinkage effect on the estimated classification function, which can further lead to biased
probability estimation. Despite the progress, many challenges remain. For example, whether L1
type and other regularization terms such as the smoothly clipped absolute deviation (SCAD, Fan
and Li, 2001) penalty have a similar shrinkage effect was not explored, and the consequence of
regularization terms for multicategory classification is unclear. In this section, we extend the
discussion to angle-based classifiers for various regularization methods, and demonstrate that the
shrinkage effect is common for high dimensional problems.
To begin with, we first provide some theoretical insights on why heavy penalty is often needed
for high dimensional low sample size classification problems. Recall that the primary goal of
classification is to obtain a classifier with high prediction accuracy for future testing data sets.
Traditional methods such as ordinary logistic regression minimize the empirical loss function on
the training data set, and it is known that this approach can result in a classifier that fits the training
data too well with suboptimal future prediction. For high dimensional problems, the overfitting
issue is more severe. To overcome this difficulty, it is common to use regularization terms in the
optimization to guarantee the generalization ability of the resulting classifier. In the next theorem,
we focus on the angle-based classifier (2) with linear learning and L1 and L2 penalties, and show
that in order to have a good classifier for problems with large p and small n, we often need λ to
8
be large. Note that fj (x) = β Tj x + β0,j for all j ∈ {1, . . . , k − 1} in linear learning. In this paper,
we assume the domain of X is bounded, and the loss function ` is continuously differentiable. One
can verify that this assumption is mild and can be satisfied by many practical problems.
Theorem 1. Define I{y, f (x)} to be the indicator function of misclassification for the observation
(x, y), using f (·) and the angle-based prediction rule. We have that, with probability at least 1 − δ,
n
E[I{Y, f̂ (X)}] ≤
1X
I{yi , f̂ (xi )} + 3
n
i=1
log(2/δ)
n
1/2
+ Z,
where
Z=
C1
λ
C2
λ
q
q
log(p)
n
p
n1/2
for linear learning with L1 penalty,
for linear learning with L2 penalty,
and C1 , C2 are constants independent of λ, n and p.
Theorem 1 shows that for a trained classification function vector f̂ from (2), the corresponding
prediction error rate on a large testing data set can be bounded by the error rate on the training
set, plus a penalty term that depends on λ, n and p. For a problem with large p and small n, one
can verify that we need a large λ in order to guarantee that the future error rate is small. This
helps to shed some light on why heavy penalty is often needed for high dimensional problems.
Remark 1. For brevity, in Theorem 1 we focus on linear learning in angle-based classifiers using
L1 and L2 regularization. One can verify that similar results hold for many other settings, for
example learning in a Reproducing Kernel Hilbert Space (RKHS, Schölkopf and Smola, 2002; ShaweTaylor and Cristianini, 2004), using other penalties such as the SCAD (Fan and Li, 2001) and the
elastic-net (Zou and Hastie, 2005) methods, and classification with k functions and the sum-to-zero
constraint (1). See the proof of Theorem 1 in the appendix for more details.
Next, we show that the regularization term with large λ can significantly shrink the norm of f̂ ,
which further leads to the bias in probability estimation using (6) for angle-based classifiers. As in
Theorem 1, we focus on L1 and L2 penalties in linear learning in the following proposition.
Proposition 1. The solution f̂ to (2) satisfies that kf̂ k1 ≤ C/λ for the L1 and L2 regularized
learning, where C is a universal constant independent of λ, n and p.
Because every element of W j is bounded by 1, we have that hW j , f̂ (x)i is bounded by C/λ
for all j and x. For high dimensional problems, Theorem 1 shows that the classifier tends to
9
use a large λ for the regularization term. As a result, the functional margins in the angle-based
classification framework, hW j , f̂ (x)i, tend to shrink towards zero. This, however, does not affect
the label prediction performance of the classifier, as a correct prediction only requires that the
corresponding angle-based functional margin hW y , f̂ (x)i is the largest. Therefore, with a large
λ, the future prediction error rate EI{Y, f̂ (X)} can be close to the optimal (assuming that the
empirical error rate on the training set is small), although the estimated f̂ has a small norm. On
the other hand, the estimated probabilities tend to shrink towards 1/k. In particular, recall from
(6) that we use
`0 {hW j , f̂ (x)i}−1
P̂j (x) = Pk
0
−1
i=1 ` {hW i , f̂ (x)i}
(7)
to estimate Pj (x). By Proposition 1, we have that hW j , f̂ (x)i → 0 for large λ. Consequently,
`0 {hW j , f̂ (x)i} → `0 (0). Therefore, for a high dimensional problem where the selected λ is large,
P̂j (x) → 1/k for all j and k.
Remark 2. Using similar proofs as that of Proposition 1, one can verify that the shrinkage effect
also exists for many other regularization methods, such as the SCAD and elastic-net penalties,
and for learning in nonlinear functional spaces such as RKHS. See the proof and discussion of
Proposition 1 in the appendix for more details.
To alleviate the bias in probability estimation, in the next section, we propose a new refit
method for angle-based classifiers. We show that our method can improve probability estimation
accuracy, at the cost of an additional optimization problem that can be solved very efficiently.
3.2
A New Refit Method for Angle-based Classifiers
With appropriate regularization terms and tuning parameters in high dimensional problems, Theorem 1 and Proposition 1 show that margin-based classifiers can deliver accurate class prediction,
however, the estimated functional margins could shrink towards 0. This can result in suboptimal
probability estimation. To overcome this difficulty, in this section, we propose a new multicategory
refit method that can significantly alleviate the bias in probability estimation.
To begin with, recall that it is common to perform probability estimation using (5) in binary
problems. The following proposition shows that all the information about class conditional probability for binary margin-based classification is contained in the estimated fˆ(x).
10
Proposition 2. Suppose for two observations x and z, we have that P̂1 (x) = P̂1 (z) using (5).
Then, it must be that fˆ(x) = fˆ(z).
Notice that the mapping x → fˆ(x) is equivalent to finding the separating hyperplane in the
feature space, and projecting all training observations onto its corresponding norm vector (see
Figure 1 for an illustration). From Theorem 1 and Proposition 1, one can conclude that the sign
of fˆ(x) can be accurate, yet the scale would shrink towards 0 for high dimensional problems.
To circumvent this challenge, Zhang et al. (2013) proposed to learn a second step classifier using
{(fˆ(xi ), yi ); i = 1, . . . , n} as the training data set, and employ probability estimation based on the
result from the refit step. In other words, we employ the projected coordinates on the norm vector
as the new variable. Zhang et al. (2013) proved that for binary classifiers where there is shrinkage
effect, the refit step can help to “restore” the norm of estimated classification functions, which can
lead to better probability estimation.
Figure 1: For binary margin-based classifiers, the classification function f (x) maps x onto the norm vector
of the separating hyperplane.
Despite the progress, how to perform the projection step in multicategory classification problems
is challenging. For example, in the simultaneous margin-based classification with k functions and
the sum-to-zero constraint, one can verify from similar arguments as in Theorem 1 and Proposition 1
that the label prediction based on (fˆ1 , . . . , fˆk ) can be accurate, yet the probability estimation is
biased because the norm of (fˆ1 , . . . , fˆk ) shrinks. To alleviate this difficulty, a possible approach is
11
to use (fˆ1 , . . . , fˆk ) as the second step input for classification without penalty terms. Similarly, in
the angle-based classification framework, because class conditional probability estimation depends
on hf̂ , W j i for all j, one can use hf̂ , W j i; j = 1, . . . , k as the predictors for a naive refit approach.
However, because of the commonly used sum-to-zero constraint for regular classifiers and the sumto-zero property for angle-based methods, one of the k predictors in the second step refit approach
mentioned above is unnecessary. It is unclear how to choose predictors that are not redundant,
and simultaneously contain all information about class conditional probability. To overcome this
challenge, for angle-based classifiers, we propose to use the k−1 dimensional {f̂ (xi ); i = 1, . . . , n} =
{(fˆ1 (xi ), . . . , fˆk−1 (xi )); i = 1, . . . , n} as predictors for our second step refit method. The next
theorem shows that the k − 1 functions contain the same information about class conditional
probabilities as hf̂ , W j i; j = 1, . . . , k.
Theorem 2. There is a one-to-one correspondence between f ∈ Rk−1 and hf , W j i; j = 1, . . . , k,
such that f = g if and only if hf , W j i = hg, W j i; j = 1, . . . , k. Moreover, if for two observations
x and z, we have that (P̂1 (x), . . . , P̂k (x)) = (P̂1 (z), . . . , P̂k (z)) using (7), then f̂ (x) = f̂ (z).
Because the norm of hf̂ , W j i is an increasing function with respect to the norm of f̂ , Theorem 2 suggests that if we can restore the norm of f̂ , we can alleviate the shrinkage effect on class
conditional probability estimation. We summarize our refit step for probability estimation in the
angle-based classification framework as follows.
Algorithm for Multicategory Refit Method
1. Fit the classifier on the training data set to obtain f̂ with proper tuning on λ.
2. Create the training data set for the second step refit method. In particular, for the ith
observation, we use f̂ (xi ) as the corresponding predictor vector with k − 1 elements, and yi
as the label.
3. Fit the classifier again on the training data set from Step 2, without any penalty terms.
Denote by f̃ the obtained classifier.
4. Estimate the class conditional probability Pj (x) by P̃j (x) =
`0 [hW j ,f̃ {f̂ (x)}i]−1
Pk
.
0
−1
q=1 ` [hW q ,f̃ {f̂ (x)}i]
Note that in Step 3 in the above algorithm, we train the refit classifier without regularization
terms. This is because in most applications, the number of observations n is much larger than
the number of classes k. Hence, there is little danger for overfitting, and no penalty is needed.
12
Consequently, there would be no shrinkage effect on the estimated f̃ , and we can obtain better
probability estimation accuracy.
In the next section, we perform numerical analysis on simulated and real data sets to demonstrate the effectiveness of our refit approach for multicategory classification problems.
4
Numerical Examples
In this section, we conduct numerical experiments to demonstrate the effectiveness of our new refit
method for multicategory classification. In particular, we perform simulation analysis in Section 4.1,
and test on some real data sets in Section 4.2.
4.1
Simulation
We investigate the performance of our refit method using two simulated examples with high dimensional predictors in this section. In particular, we compare the class prediction accuracy, as well as
probability estimation accuracy, for the traditional method and our new refit method. To measure
the probability estimation accuracy, we report the mean absolute difference (MAD) between the
true class conditional probability and the estimated value. The MAD is defined to be
n
MAD =
k
1 XX
|Pj (xi ) − P̂j (xi )|.
n
i=1 j=1
For all examples and classifiers, we train the model on a training data set, and select the best tuning
parameters from a candidate set {2−10 , 2−9 , . . . , 29 , 210 } that correspond to the smallest prediction
error on a separate tuning data set. Then, we measure the class prediction and probability estimation accuracy on a large testing data set. For the choice of loss functions, we use the logistic
regression deviance loss (logi) and the soft large-margin unified machine loss (soft, Liu et al., 2011)
for demonstration. We emphasize that our method can be generalized in a straightforward manner
to other differentiable convex loss functions. We also train support vector machines using kernlab
(KSVM) and report the corresponding classification accuracy as a comparison. Note that traditional SVM does not provide class conditional probability estimation. For each example, the prior
probabilities for each class are the same.
Example 1. In this example, we let k = 3. The classification signal depends on 10 independent
predictors X1 , . . . , X10 . Given Y , the marginal distribution of each predictor is normal with variance
13
Bayesian
Soft
Refit Soft
Logi
Refit Logi
KSVM
Example 1
Error MAD
5.51
5.58
15.4
7.23
4.22
5.53
23.6
7.13
3.94
28.87
-
Example 2
Error MAD
35.0
36.9
13.9
38.0
5.80
38.5
12.0
43.0
8.81
87.4
-
Table 1: The classification error rates (Error) and mean absolute differences (MAD) probability
estimation for simulated examples 1 and 2.
σ = 2. For class 1, the mean vector of the normal distribution is (3, 3, 3, 3, 0, 0, 0, 0, 0, 0)T . For class
2, the mean vector of the normal distribution is (0, 0, 0, 3, 3, 3, 3, 0, 0, 0)T . For class 3, the mean
vector of the normal distribution is (0, 0, 0, 0, 0, 0, 3, 3, 3, 3)T . Next, we generalize noise predictors
for high dimensional learning. In particular, we use 1490 variables, each following i.i.d. N (0, 0.02).
The training and tuning sets have size 300 each, and the testing set has 29400 observations.
Example 2. We generalize a 10-class problem for this example, where the classification signal
depends on two predictors. The marginal distribution of X1 and X2 contains classification signal,
and we generate another 498 variables that follow i.i.d. N (0, 0.01). For each class, the marginal
distribution of (X1 , X2 ) is joint normal with constant variance matrix I2 , and the corresponding
mean vectors are uniformly distributed on the circle X12 + X22 = 9.
The results for simulated Examples 1 and 2 are reported in Table 1. From the outcome, one
can conclude that our refit method can significantly improve the accuracy of class conditional
probabilities, compared to the original method proposed by Zhang and Liu (2014). Moreover, the
corresponding classification accuracy of the original angle-based method is close to that of our refit
ones. Hence, the refit approach is highly competitive. Note that the SVM performs poorly.
4.2
Real Data Analysis
In this section, we investigate the performance of our proposed refit method by applications on
some real data sets. In particular, we apply the original angle-based classifier with probability
estimation, as well as our refit method, on two benchmark data sets that can be found on the
UCI machine learning repository. The first data set that we consider is Wine, which depicts
the quality of 3 categories of wine with 13 variables. The second data set is Phishing websites,
which contains 30 attributes and 2456 observations. More details for these data can be found at
https://archive.ics.uci.edu/ml/datasets/.
14
Soft
Refit Soft
Logi
Refit Logi
Wine
Error CRE
52.9
5.85
28.1
3.76
58.4
6.40
31.6
5.79
Phishing
Error CRE
47.0
2.03
47.8
0.77
36.6
1.92
42.3
1.29
Table 2: The classification error rates (Error) and cross entropy error (CRE) probability estimation
for real data sets.
For both examples, we split the entire data into 6 parts. We use one of the parts as the testing
set, and perform a four-fold cross validation on the remaining training data to select the best
classifiers. The selection of tuning parameters is similar to that of the simulation examples. Since
we do not know the underlying class conditional probability, we use the Cross Entropy Error (CRE,
Zhang and Liu, 2014) to measure the goodness of fit of the estimated probabilities. In particular,
for a testing set with n observations, the CRE is given as
n
CRE = −
1X
p̂yi (xi ) log{p̂yi (xi )}.
n
i=1
We report the average results for 100 replicates in Table 2. Notice that one can draw similar
conclusions as in the simulation section, which demonstrates the effectiveness of our refit approach.
5
Discussion
In this paper, we investigate the effect of penalty terms for multicategory margin-based classifiers.
In particular, we provide some theoretical insights on why heavy penalty is often needed for high
dimensional low sample size classification problems. Then we show that such penalty can lead
to severely biased class conditional probability. To overcome the difficulty, we propose a new
multicategory refit approach for the angle-based classifiers. Numerical experiments show that the
new method can improve class probability estimation with very low computation burden, hence
can be very useful.
15
References
Bartlett, P. L., Jordan, M. I., and McAuliffe, J. D. (2006). Convexity, Classification, and Risk
Bounds. Journal of the American Statistical Association, 101, 138–156.
Boser, B. E., Guyon, I. M., and Vapnik, V. N. (1992). A Training Algorithm for Optimal Margin
Classifiers. In D. Haussler, editor, Proceedings of the Fifth Annual Workshop on Computational
Learning Theory, COLT ’92, pages 144–152. Association for Computing Machinery, New York,
NY, U.S.A.
Cai, T. and Liu, W. (2011). A Direct Estimation Approach to Sparse Linear Discriminant Analysis.
Journal of the American Statistical Association, 106(496), 1566–1577.
Cortes, C. and Vapnik, V. N. (1995). Support Vector Networks. Machine Learning, 20, 273–297.
Fan, J. and Li, R. (2001). Variable Selection via Nonconcave Penalized Likelihood and its Oracle
Properties. Journal of the American statistical Association, 96(456), 1348–1360.
Freund, Y. and Schapire, R. E. (1997). A Desicion-theoretic Generalization of On-line Learning
and an Application to Boosting. Journal of Computer and System Sciences, 55(1), 119–139.
Friedman, J. H., Hastie, T. J., and Tibshirani, R. J. (2000). Additive Logistic Regression: a
Statistical View of Boosting. Annals of Statistics, 28(2), 337–407.
Hastie, T. J., Tibshirani, R. J., and Friedman, J. H. (2009). The Elements of Statistical Learning.
New York: Springer, 2nd edition.
Lee, Y., Lin, Y., and Wahba, G. (2004). Multicategory Support Vector Machines, Theory, and
Application to the Classification of Microarray Data and Satellite Radiance Data. Journal of the
American Statistical Association, 99, 67–81.
Lin, X., Wahba, G., Xiang, D., Gao, F., Klein, R., and Klein, B. (2000). Smoothing Spline Anova
Models for Large Data Sets with Bernoulli Observations and the Randomized GACV. Annals of
Statistics, 28(6), 1570–1600.
Liu, Y. (2007). Fisher Consistency of Multicategory Support Vector Machines. In Eleventh International Conference on Artificial Intelligence and Statistics, pages 289–296.
Liu, Y. and Shen, X. (2006). Multicategory ψ-learning. Journal of the American Statistical Association, 101, 500–509.
Liu, Y. and Yuan, M. (2011). Reinforced Multicategory Support Vector Machines. Journal of
Computational and Graphical Statistics, 20(4), 901–919.
Liu, Y., Zhang, H. H., and Wu, Y. (2011). Soft or Hard Classification? Large Margin Unified
Machines. Journal of the American Statistical Association, 106, 166–177.
Park, S. Y., Liu, Y., Liu, D., and Scholl, P. (2010). Multicategory composite least squares classifiers.
Statistical Analysis and Data Mining, 3(4), 272–286.
Schölkopf, B. and Smola, A. J. (2002). Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond (Adaptive Computation and Machine Learning). The MIT
Press.
16
Shawe-Taylor, J. S. and Cristianini, N. (2004). Kernel Methods for Pattern Analysis. Cambridge
University Press, 1st edition.
Shen, X., Tseng, G. C., Zhang, X., and Wong, W. H. (2003). On ψ-learning. Journal of the
American Statistical Association, 98, 724–734.
Wang, J., Shen, X., and Liu, Y. (2008). Probability Estimation for Large Margin Classifiers.
Biometrika, 95(1), 149–167.
Wang, L. and Shen, X. (2007). On L1 -norm Multi-class Support Vector Machines: Methodology
and Theory. Journal of the American Statistical Association, 102, 595–602.
Wu, Y., Zhang, H. H., and Liu, Y. (2010). Robust Model-free Multiclass Probability Estimation.
Journal of the American Statistical Association, 105(489), 424–436.
Zhang, C. and Liu, Y. (2013). Multicategory Large-margin Unified Machines. Journal of Machine
Learning Research, 14, 1349–1386.
Zhang, C. and Liu, Y. (2014). Multicategory Angle-based Large-margin Classification. Biometrika,
101(3), 625–640.
Zhang, C., Liu, Y., and Wu, Z. (2013). On the Effect and Remedies of Shrinkage on Classification
Probability Estimation. The American Statistician, 67(3), 134–142.
Zhu, J. and Hastie, T. J. (2005). Kernel Logistic Regression and the Import Vector Machine.
Journal of Computational and Graphical Statistics, 14(1), 185–205.
Zou, H. and Hastie, T. J. (2005). Regularization and Variable Selection via the Elastic Net. Journal
of the Royal Statistical Society, Series B,, 67(2), 301–320.
Zou, H., Zhu, J., and Hastie, T. J. (2008). New Multicategory Boosting Algorithms Based on
Multicategory Fisher-Consistent Losses. Annals of Applied Statistics, 2(4), 1290–1306.
17
| 10math.ST
|
Robust Facial Landmark Detection under Significant Head Poses and Occlusion
Yue Wu
Qiang Ji
ECSE Department, Rensselaer Polytechnic Institute
arXiv:1709.08127v1 [cs.CV] 23 Sep 2017
110 8th street, Troy, NY, USA
{wuy9,jiq}@rpi.edu
Abstract
There have been tremendous improvements for facial
landmark detection on general “in-the-wild” images. However, it is still challenging to detect the facial landmarks on
images with severe occlusion and images with large head
poses (e.g. profile face). In fact, the existing algorithms
usually can only handle one of them. In this work, we propose a unified robust cascade regression framework that can
handle both images with severe occlusion and images with
large head poses. Specifically, the method iteratively predicts the landmark occlusions and the landmark locations.
For occlusion estimation, instead of directly predicting the
binary occlusion vectors, we introduce a supervised regression method that gradually updates the landmark visibility
probabilities in each iteration to achieve robustness. In addition, we explicitly add occlusion pattern as a constraint to
improve the performance of occlusion prediction. For landmark detection, we combine the landmark visibility probabilities, the local appearances, and the local shapes to iteratively update their positions. The experimental results
show that the proposed method is significantly better than
state-of-the-art works on images with severe occlusion and
images with large head poses. It is also comparable to other
methods on general “in-the-wild” images.
(a) General “in-the-wild”images (b) occlusion (c) Profile face
Figure 1. Predicted landmark locations and occlusions (red dots in
(c) indicate occluded points) with the proposed method. Images
are from Helen [16], LFPW [2], COFW [4], and FERET [19] sets.
The existing algorithms usually can only handle either
images with occlusion [4][11][28][10][14] or images with
large head poses [27][30]. In addition, they treat them differently. For example, pose dependent [30] or occlusion dependent [4][28] models are trained to handle different cases.
However, if we regard the landmarks on the self-occluded
facial parts as occluded points, where the face itself is the
occluder, we can consider images with large head poses as
special cases of images with occlusion and treat them similarly. Based on this intuition, we propose a novel method to
handle both images with severe occlusion and images with
large head poses.
The general framework of the proposed robust cascade
regression method is shown in Figure 2 and 3. First, we
initialize the landmark locations using the mean face shape
and assume all the points are visible in the first iteration.
Then, to achieve robustness, instead of directly predicting
the binary landmark occlusion vectors and landmark locations, we gradually update the landmark visibility probabilities and locations iteratively in a coarse to fine manner. When updating the visibility probability, we utilize
the appearance and shape information that depend on the
currently estimated landmark locations. In addition, we explicitly add occlusion pattern as a constraint. When updating the landmark locations, we consider the appearance and
1. Introduction
Facial landmark detection refers to the localization of the
fiducial points on facial images. With the detected points,
human facial shape and appearance information can be utilized for facial analysis. Recently, there are tremendous improvements of the facial landmark detection algorithms on
general “in-the-wild” images (Figure 1(a)). However, it is
still challenging to detect the facial landmarks on images
with severe occlusion and large head poses (e.g. pose > 60
degree, self-occlusion)(Figure 1(b)(c)).
1
(a) Initialization
(b) Iteration 1
(c) Iteration 2
(d) Last iteration
(e) final output
Figure 2. Facial landmark detection and occlusion prediction in different iterations. First row: image with severe occlusion. Second row:
image with large head pose. In (a)-(d), the radius of point is inversely proportional to the landmark visibility probabilities and the point
is marked with red color if its visibility probability is smaller than a threshold. In (e), for images with large head pose, we only show the
visible points as the final output. (Better see in color)
shape information with weights that depend on the landmark visibility probabilities. The landmark locations and
visibility probabilities interact to reach convergence. We
highlight the major contributions of the proposed work:
• General framework: The proposed method is the first
algorithm that can handle both images with severe occlusion and images with large head poses. It treats selfocclusion in images with large head poses as a special
case of image occlusion.
• Occlusion prediction: Our occlusion prediction
method is different from the previous works [4][28].
While they train several occlusion dependent models
(e.g. mouth is occluded), we handle them in one unified framework. In addition, we explicitly add occlusion pattern as a constraint.
• Landmark localization: For facial landmark detection with occlusion, we treat points differently based
on their visibility probabilities. We explicitly add the
shape features for prediction. For images with large
head poses, since the landmark annotations are missing
on the self-occluded facial parts, we propose a learning
method to handle this issue. The facial shape pattern is
implicitly embedded in the model.
• Experimental results: The proposed method performs well on general “in-the-wild” images, and it
is significantly better than the other state-of-the-art
works on images with severe occlusion and large head
poses.
• Database: We annotated some images with large head
poses from FERET database [19]. 1
The remaining part of the paper is organized as follows.
In section 2, we review the related work. In section 3, we
introduce the proposed method. In section 4, we discuss the
experiments and we conclude the paper in section 5.
1 Landmark annotations can be downloaded: http://www.ecse.
rpi.edu/˜cvrl/wuy/FERET_annotation.rar
2. Related Work
Facial landmark detection algorithms can be classified
into three major categories, including the holistic methods, the constrained local methods and the regression based
methods. The proposed method follows the regression
framework, but it is specifically designed to handle occlusion and large head poses.
The holistic methods build global appearance and shape
models during learning and fit testing image by estimating
the model parameters. The differences among the holistic
methods lie in the fitting procedure and they usually follow
either the least squares formulation [17][1] or the linear regression formulation [6].
The constrained local methods [8] combine global face
shape model and local appearance model for facial landmark detection. This approach is usually superior to the
holistic methods, since it relies on more flexible local appearance model that is easier to learn. Typical constrained
local methods usually focus on the representations of the
face shapes [2][24] or the local appearances [22][7].
Recently, the regression based methods show more
promising performance than the holistic methods and the
constrained local methods. Unlike the methods in the other
two categories, the regression based methods do not explicitly build the global appearance or shape models. Instead,
they directly map the local facial appearance to the landmark locations. For example, the absolute coordinates of
the facial landmarks can be estimated directly from the facial appearance with the conditional regression forests [9]
or deep convolutional neural networks [23]. Different
from [9][23], most of the other regression based methods [26][20][5][15] start from an initial face shape, and they
gradually update the landmark locations based on the local
appearances. For those regression based methods, cascade
techniques are usually embedded in the framework to improve both the robustness and accuracy. One limitation of
Appearance features
Shape features
Occlusion patterns: L
,
Appearance features
Shape features
.
,
Landmark visibility probabilities
Update the visibility
probabilities
Initial landmark locations
and visibility probabilities
Update the landmark
locations
Figure 3. Facial landmark detection in one iteration. (Better see in color)
the regression based methods is that they usually can only
be trained with fully supervised data. Therefore, they cannot learn one unified model to handle both frontal face and
profile face with missing landmark location annotations. In
this work, we improve the method to solve this issue.
Despite the fact that facial landmark detection accuracy
has been dramatically improved on general “in-the-wild”
images, facial landmark detection remains challenging for
facial images with severe occlusion or large head poses.
There are only a few algorithms focusing on those challenging cases. For example, for images with severe occlusion,
occlusion dependent models are built based on the partial
appearances from pre-defined non-occluded facial parts for
landmark detection and occlusion estimation in [4] and [28].
During testing, the detection results from all occlusion dependent models are merged together with weights depending on the landmark occlusion prediction results. In [11], a
probabilistic graphical model is built to infer the landmark
occlusions and locations jointly. For images with large head
poses, there are even fewer works [30][27]. In [30], several
models are built to handle facial landmark detection with
different pre-defined discrete head poses. The detection result with the highest score is outputted as the final result.
In [27], 3D model is used to handle images with large head
poses. Different from the previous works, we propose a
general framework to handle both cases and we do not build
pose dependent or occlusion dependent models.
3. Approach
3.1. The general framework
The goal of the facial landmark detection algorithm is
to find out the mapping from image I to landmark locations x ∈ <2Dl , where Dl is the number of facial landmarks. To handle images with occlusion and large head
poses, we introduce the landmark visibility probability variable p ∈ [0, 1]Dl . For one specific image with binary landmark occlusion vector c ∈ {0, 1}Dl , pd measures the probability that a landmark is visible (cd = 1).
The general framework of the proposed robust cascade
regression method is shown in Figure 2, 3 and Algorithm 1.
With the initial landmark locations and the assumption that
all the landmarks are visible at the beginning, the algorithm
Algorithm 1: The general framework
Initialize the landmark locations x0 using the mean face;
Assume all the landmarks are visible p0 = 1
for t=1, 2, ..., T or convergence do
Update the landmark visibility probabilities given the
images, the current landmark locations, and the
occlusion pattern Loss(.).
ft : I, xt−1 , Loss(.) → ∆pt
pt = pt−1 + ∆pt
Update the landmark locations given the images, the
current landmark locations, and the landmark visibility
probabilities.
gt : I, xt−1 , pt → ∆xt
xt = xt−1 + ∆xt
end
Output the estimated landmark locations xT and the binary
occlusion vector based on the predicted visibility
probabilities pT .
updates the visibility probabilities and the landmark locations across iterations to achieve convergence. When updating the visibility probabilities, we introduce a constrained
supervised regression model, denoted as ft , to predict the
landmark visibility probability update ∆pt based on the image, the current landmark locations xt−1 and the occlusion
pattern embedded in a loss function Loss(.). When updating the landmark locations, we use a regression function gt
that predicts the landmark location update ∆xt based on the
image, the current landmark locations xt−1 , and the visibility probabilities pt . In the following, we discuss each part
in details.
3.2. Update the landmark visibility probability
The landmark visibility probability and landmark occlusion are difficult to predict. First, there are large variations
of the facial appearance on the occluded facial parts, since
the occlusion could be caused by arbitrary objects. Second,
due to the poor landmark detection results in the first few
iterations, direct occlusion prediction is infeasible. Thus, a
better choice is to update the landmark visibility probabilities by accumulating information through iterations. Third,
due to the occlusion consistency among nearby points, there
exists certain occlusion pattern, which should be embedded
in the model as a constraint. However, since any part of
the face could be occluded by arbitrary objects, there are a
large number of possible patterns and the occlusion patterns
could be complex. Thus, it’s not appropriate to pre-define
the possible occlusion pattern (e.g. mouth is occluded) as
the existing works [4][28], and more effective model should
be used. Fourth, it’s not enough to learn the occlusion pattern from limited training images with exhaustive human
annotations. In fact, the occlusion pattern can be learned
from synthetic data. Last but not the least, the regression
function should depend on both the local appearances and
the current shapes for joint prediction. Based on those intuitions, we propose to update the landmark visibility probabilities based on the appearance and shape information from
all points and use the learned explicit occlusion pattern as a
constraint.
3.2.1
Landmark visibility prediction model
Landmark visibility prediction depends on the local appearance, the current shape, and the occlusion pattern. To encode the appearance information, we use SIFT features of
the local patches centered at the current landmark locations,
denoted as φ(I, xt−1 ) ∈ <Dl Df (Df =128 is the dimension
of features). To encode the shape information, we calculate
the differences of x, y coordinates for pairwise landmarks
to get the shape features denoted as ϕ(xt−1 ) ∈ <Dl (Dl −1) ,
which provide the scale, pose, and non-ridge information of
the current face. By combining the appearance and shape
information, we can generate a concatenated feature vector
denoted as Ψ(I, xt−1 ) = [φ(I, xt−1 ); ϕ(xt−1 )]. To encode
the occlusion pattern, we learn a loss function Loss(c) for
occlusion vector c (there are 2Dl ). The loss function penalizes the infrequent and infeasible occlusion label configurations (e.g. every other point is occluded). Then, we can
update the landmark visibility probabilities pt for the next
iteration:
minimize
k∆pt − T t Ψ(I, xt−1 )k22 + λEpt [Loss(c)]
subject to
pt = pt−1 + ∆pt
∆pt
(1)
t
0≤p ≤1
Ept [Loss(c)] =
D
2 l
X
Loss(ck )P (ck ; pt )
(2)
D k=1
Y
l
P (c; p) =
d=1
pd cd (1 − pd )1−cd
(3)
In the first term of the objective function, we use linear
regression function with parameter T t to predict the landmark visibility probability update ∆pt from the appearance
and shape features Ψ(I, xt−1 ), and we want to minimize the
standard least squares error. In the second term, we want to
minimize the expected loss of the occlusion vector, where
the expectation is taken over the visibility probabilities pt
we want to infer for the next iteration. The basic idea is
to minimize the discrepancy between the predicted binary
occlusion vector and the prior occlusion pattern encoded in
Loss(c). The expectation is denoted as Ept [Loss(c)] and it
is detailed in Equation 2 and 3. λ is the hyper-parameters.
In the following, we first explain model learning, and then
discuss model inference.
3.2.2
Learning the landmark visibility prediction
model
Model learning refers to the estimation of the linear regression parameter T t for each iteration and the loss function
Loss(.), which should be learned before the cascade training.
We use the Autoencoder model [3] (Figure 4 (a)) to learn
the loss function Loss(.) that captures the prior occlusion
pattern based on the landmark occlusion labels of the training data and the synthetic data. We generate the synthetic
landmark occlusion labels by sampling different numbers
of occluders (up to 4 rectangles with random sizes and locations) in the face region and superimposing them onto the
mean face (see Figure 4 (b) for one example). Then, based
on all the feasible landmark occlusion label ci from the real
training data and the synthetic data, we learn the Autoencoder model with parameters θ = {W1 , b1 , W2 , b2 } that
can minimize the reconstruction errors:
θ∗ = arg min
X
i
kci − σ(W2 σ(W1 ci + b1 ) + b2 )k22 ,
(4)
where σ(.) is the sigmoid function. The model is pretrained with Restricted Boltzmann Machine model and finetuned jointly [13]. After model learning, the loss function is defined as the reconstruction errors Loss(c; θ) =
kc−σ(W2 σ(W1 c+b1 )+b2 )k22 . Figure 4 (c) shows the distributions of reconstruction errors of the feasible occlusion
labels for Auto-encoder training, and random binary data,
consisting of both feasible and infeasible occlusion vectors.
The reconstruction error apparently can penalize the random infeasible occlusion vectors.
For the estimation of linear regression function with parameter T t in each iteration, we use standard least squares
formulation. Specifically, given the training image, the currently estimated landmark locations xt−1
, we can calculate
i
the appearance and shape features Ψ(Ii , xt−1 ). By subtracting the currently estimated landmark visibility probabilities
pt−1
from the ground truth probabilities p∗i , we can get the
i
landmark visibility probability update ∆pti . Then, T t could
be estimated by the standard least-squares formulation with
closed form solution.
∗
T t = arg min
Tt
X
i
k∆pti − T t Ψ(Ii , xt−1
)k22
i
(5)
ࢉො
ࢃ ǡ ࢈
ࢎ
ࢃ ǡ ࢈
WUDLQLQJGDWD
UDQGRPGDWD
ࢉ
SHUFHQWDJHRIGDWD
(a)
UHFRQVWUXFWLRQHUURU
(b)
(c)
Figure 4. (a) Learning the occlusion patterns with Autoencoder.
(b) Generate the synthetic occlusion labels. Green points: mean
face shape. Gray areas: randomly generated occluders. (c) The
distributions of reconstruction errors. Training data: feasible occlusion labels for Auto-encoder training. Random data: randomly
generated binary data consisting of both feasible and infeasible
occlusion vectors (e.g. every other point is occluded).
3.2.3
Inference with the landmark visibility prediction
model
In inference, with Equation 1, we need to estimate ∆pt
given the appearance and shape features Ψ(I, xt−1 ), the currently estimated visibility probabilities pt−1 , model parameter T t , and the loss function Loss(.). The inference is
non-trivial, since the calculation of the expectation (Equation 2) would sum over 2Dl possible occlusion vectors
which would be intractable if the number of landmarks Dl
is large. To tackle this problem, we use Monte Carlo approximation and calculate the second term over the samples
(K=5000 in our experiments).
Ept [Loss(c)] ≈
K
const X
Loss(cek )P t (cek ),
K
(6)
k=1
where cek is sampled by assuming all vectors are equally
likely, and const = 2Dl .
To solve the optimization problem in Equation 1 with the
approximation in Equation 6, we use gradient descent algorithm. If we denote the gradient of the objective function
w.r.t. ∆pt as δ, then:
K
∂P t (cek )
const X
δ = 2(∆pt − T t Ψ(I, xt−1 )) + λ
Loss(cek )
,
K
∂∆pt
k=1
(7)
issues and difficulties for landmark localization on images
with occlusion and large head poses. First, the points should
not be treated equally. For points with low visibility probabilities, the local appearances would be less useful and reliable, since the appearance of the occluder has limited information about the landmark locations. Second, for the detection of the occluded landmarks, the shape features and the
shape constraint from other landmarks are more important
than its local appearance. Third, due to the self-occlusion
issue on images with large head poses, the location annotations for landmarks on the occluded facial parts are not
available.
Based on the intuitions illustrated above, we modify the
regression based method for landmark detection. Specifically, we predict the location update vector ∆xt with linear
regression function as bellow:
p
∆xt = Rt [ pt ◦ Ψ(I, xt−1 )]
(8)
Here, “◦” denotes the block-wise product between the
square of the landmark visibility probabilities and the appearance features from corresponding point (we keep the
shape features unchanged). In this case, the prediction will
rely more on appearance features from points with high visibility probabilities, while treat the shape features equally.
Since all points are estimated together starting from the
mean face, the shape constraint is automatically embedded
in learning. Thus, the algorithm may detect the occluded
points based on the shape constraint. During detection, we
could only output the landmark locations with high visibility probabilities (> 0.6) (second row of Figure 2 (e)).
In model training, we need to estimate the parameter
Rt in each iteration with missing landmark location annotations (e.g. no landmark annotations on self-occluded
facial parts). To tackle this incomplete annotation issue,
for each training image, we introduce the binary variable
w ∈ {0, 1}Dl to indicate whether the location annotation of
a specific landmark is available (wk = 1) or not (wk = 0).
Then, combining wi , the location update ∆xti (subtracting
the current landmark locations xt−1
from the ground truth
i
x∗i ), the currently estimated visibility probability pti , and the
concatenated appearance and shape features Ψ(Ii , xt−1
),
i
parameter learning can be formulated as a weighted least
squares problem with closed form
q solution.
Rt = arg min
Rt
X
i
k∆xti −Rt [
pti ◦Ψ(Ii , xt−1
)]k2diag(wi ) , (9)
i
Given the gradient δ, ∆pt can be updated by moving along
the descent direction, with the constraint that pt = pt−1 +
∆pt is in the range [0, 1]Dl .
where diag(wi ) ∈ <2Dl ×2Dl is a diagonal matrix and the
corresponding elements (for x,y coordinates) are 0 if the
landmark location annotation is missing. Therefore, parameter learning for the corresponding rows of Rt will not depend on the specific data with missing landmark annotation.
3.3. Update the landmark locations
3.4. Discussion
Given the predicted landmark visibility probabilities, we
need to update the landmark locations. There are a few
Differences with Supervised Descent method
(SDM) [26]: 1) SDM learns the descent direction for facial
landmark detection. It is not designed to handle occlusions.
2) SDM cannot handle images with large head pose with
severe self-occlusion.2 3) The derivation in [26] shows
that the regression function should change according to the
current shape, while SDM fix it as constant. In the proposed
method, our shape features compensate this limitation.
Differences with Robust Cascade Pose Regression
method (RCPR) [4]: 1) RCPR builds several occlusion
dependent models which only draw features from 1/9 of the
facial region, assuming the region is not occluded. Based
on the limited information from 1/9 of the facial region,
it’s difficult for RCPR to predict the landmark locations and
their occlusions on the other 8/9 facial region. In addition,
the pre-defined 9 models can not effectively cover all possible occlusion cases. On the contrary, the proposed method
trains one unified model, which is more flexible and robust.
It draws features from the whole regions and considers them
with different weights, which is more effective. 2) RCPR
cannot handle images with large head poses. 3) There is no
explicit occlusion pattern in RCPR.
Differences with Face detection, Pose estimation,
Landmark Localization algorithm (FPLL) [30]: 1)
FPLL follows the Constrained Local Method. It builds several pose dependent models and chooses the detection result
with the highest score from all models. It would lead to poor
result if the model is chose incorrectly. On the contrary, we
propose a unified model that automatically solve the problem. 2) FPLL can not handle images with severe occlusion.
4. Experimental results
In this section, we evaluate the proposed method on images with severe occlusion, images with large head poses,
and general “in-the-wild” images.
4.1. Implementation details
Databases: We use three kinds of databases. The first
kind of databases contain general “in-the-wild” images collected from the internet with near-frontal head poses (less
than 60 degree) and limited occlusion (about 2%). They
are the Labeled Face Parts in the Wild (LFPW) database [2]
with 29 points and the Helen database [16] with 194 points.
For LFPW database, due to the invalid URLs, we only collected 608 images for training and 152 images for testing
from the internet. For Helen database, following the previous works [16][20], we use 2000 images for training and
use the remaining 330 images for testing. The second kind
of database contains “in-the-wild” images with severe occlusion (about 25%). Here, we use the Caltech Occluded
Faces in the Wild (COFW) database [4]. There are 1345
images for training and 507 images for testing. The annotations include the landmark locations and the binary oc2 Global
SDM [25] solves this issue to some extent.
clusion labels for 29 points. The third kind of database
contains images with large head poses (e.g. profile face,
head pose larger than 60 degrees). Most of the facial images comes from the MultiPIE database [12] and the Facial
Recognition Technology (FERET) database [19]. The annotations of 39 visible points on MultiPIE are provided by
the database and [30]. We annotated 11 points on profile images from FERET. Sample images from different databases
can be found in Figure 8.
Evaluation criteria: Following the previous works,
we calculate the error as the distance between detected
landmarks and the ground truth landmarks normalized by
the inter-ocular distance. For the third kind of database
with profile images, we normalized the error by half of
the distance between the outer eye corner and mouth corner. Throughout the paper, we calculate the average error
across all available annotated landmarks from the testing
databases.
Model parameters: When calculating the SIFT features, the radius of the local image patch is about 0.14 of the
face size. There are 4 cascade iterations for the model. To
augment the training images, following the previous works,
we perturb the scale, rotation angle, and position of the initial face shape for parameter learning. The hyper-parameter
λ in Equation 1 is 0.001. We use Autoencoder with one
hidden layer and we set 20 and 25 hidden nodes for experiments in section 4.2 and 4.3, respectively.
4.2. Images with severe occlusion
In this section, we show the performance of the algorithms on the challenging COFW database with severe occlusion (about 25%). For fair comparison, following the
previous work [4], the algorithm is trained with training set
from COFW and tested on the testing set. For the proposed
method, we implemented three versions and they are denoted as ours baseline (no shape features as discussed in
section 3.2.1, no occlusion pattern constraint as defined in
the second term of Equation 1), ours ShapeFea (with shape
features, no occlusion pattern), and ours Full (with shape
features and occlusion pattern).
Facial landmark detection: The experimental results
are shown in Table 1 and Figure 8 (a). The proposed method
is close to human performance and it is significantly better than all the other state-of-the-art works, including the
Cascaded Regression Copse (CRC) [10], the Occlusion Coherence (OC) [11], SDM [26], RCPR [4], Explicit Shape
Regression (ESR) [5], and FPLL algorithm [30]. Note that
for the algorithms that can perform well on general “in-thewild” databases, such as SDM [26] and ESR [5], there are
significant performance drops on the COFW database with
severe occlusion. In addition, the proposed method is significantly better than SDM that ignores the occlusion pattern and shape features. Comparing three versions of the
Table 1. Comparison of facial landmark detection errors and occlusion prediction results on COFW database (29 points) [4]. The
reported results from the original papers are marked with “*”.
Landmark detection error
visible points all points
algorithm
Human
CRC [10]
OC [11]
SDM [26]
RCPR [4]
ESR [5]
FPLL [30]
ours baseline
ours ShapeFea
ours Full
6.69
5.68
5.22
5.18
5.6 [4]
7.30*
7.46*
7.70
8.50*
11.20
14.40
6.54
5.96
5.93
Occlusion prediction
precision/recall%
80.8/37.0%*
80/40%*
80/43.78%
80/46.07%
80/49.11%
Occlusion prediction: The occlusion prediction results
are shown in Table 1 and Figure 8 (a). Following the previous work [4], we fix the precision to be about 80%, and
compare the recall values. As can be seen, the proposed
method is much better than OC [11] and RCPR [4], which
are the state-of-the-art works.
Performance across iterations: In Figure 5, we show
the landmark detection errors and the recall values (fixing
precision as 0.8, following [4]) based on the estimated occlusion probabilities across all four iterations using the proposed method (ours Full). As can be seen, both the landmark detection and the occlusion prediction results improve
over iterations and they converge quickly.
methods that can perform facial landmark detection on images with large head poses, due to the self-occlusion issue.
However, exact fair comparison to them is infeasible, since
they are trained on the combinations of different subsets of
MultiPie databases [12] and other databases. To ensure relatively fair comparison, we use the software provided by
the authors for FPLL and pose-free algorithm, and we train
the proposed method with similar procedure. For training,
we use the MultiPie database with 14 poses, the training set
from Helen and LFPW databases with 51 landmarks. We
test all the algorithms on profile faces from FERET database
(244 right profile face, 221 left profile face).
The experimental results are shown in Figure 6, 7 and
8 (b). In Figure 6, we plot the cumulative error distribution curves. In Figure 7, for different algorithms, we show
the images with the largest fitting errors. As can be seen,
the proposed method is significantly better than the FPLL
algorithm [30] and Pose-free algorithm [27] on the profile
faces. The Pose-free algorithm could result in large facial
landmark detection error if the estimated pose is wrong (e.g.
Figure 7(a)). Among all the 465 testing images, our algorithm only fail to predict the correct landmark occlusions on
one image (last image in Figure 8(b)).
1
Proposed, left faces
Proposed, right faces
FPLL, left faces
FPLL, right faces
Pose−free, left faces
Pose−free, right faces
0.9
0.8
precentage of testing faces
proposed method, we see that the shape features, and occlusion pattern are important for good performance.
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20
25
errors
30
35
40
45
50
Figure 6. Cumulative error distribution curves on profile faces.
(a)
(b)
Figure 5. Performance of the proposed method across iterations
on COFW database. (a) Landmark detection errors. (b) Occlusion
prediction accuracy (recall values at precision=0.8).
(a)
(b)
(c)
Figure 7. For different methods, we show the images with the
worst fitting results. (a) Pose-free [27], (b) FPLL [30], (c) proposed method.
4.3. Images with large head poses
4.4. General “in-the-wild" images
In this section, we evaluate the full model of the proposed method and compare it to other algorithms on challenging images with large head poses (larger than 60 degrees). To the best of our knowledge, the FPLL algorithm [30] and Pose-free algorithm [27] are the only two
Finally, we evaluate the proposed method on general
and less challenging “in-the-wild” images and compare it
to more state-of-the-art works. Note that, most of the algorithms that perform well on general “in-the-wild” images do not work well on challenging images with severe
(a) COFW database
(b) Profile faces from FERET database
(c) Helen database
(d) LFPW database
Figure 8. Facial landmark detection results with the proposed method on four databases. (a) COFW [4] database with severe occlusion.
Red points: the proposed method predicts them as occluded landmarks. (b) Profile faces from FERET database [19]. We show all points
that the algorithm predicted as visible points. (c)(d) General “in-the-wild” Helen [16] and LFPW [2] database. (Better see in color)
occlusion or large head poses. Following the previous
works [16][20][26], for each database, we use the training
set to learn the model and test it on the testing set.
The experimental results on Helen and LFPW databases
are shown in Table 2 and Figure 8 (c)(d). The results on Helen database show that the performance of the
proposed method is better than the other state-of-the-art
works, including the fast version of Local Binary Feature
(LBF) method [20], SDM [26], RCPR [4], ESR [5], the
CompASM [16], and the Extended Active Shape Model
(STASM) [18]. For the experiments on LFPW database,
the training and testing data varies from method to method.
We only can get access to half of the training data comparing to the data used in the original Consensus of exemplars (COE) [2]. However, the performance of the proposed method is still comparable to the other methods. We
also tested the proposed method on the most challenging
ibug set [21] with 135 images. Our method achieves the
detection error of 11.52, which is better than the state-ofthe-art works [20][26][5], among which the best algorithm
achieves error of 11.98.
The speed of the proposed method is comparable to other
state-of-the-art works. For the model without the explicit
occlusion pattern, the proposed algorithm can achieves 15
frames per second running on a single core machine with
matlab implementation. With the occlusion pattern, the full
model of the proposed algorithm achieves 2 frames per second. This is comparable to some state-of-the-art methods
(e.g. [26][4]), but slower than the others (e.g. [20]). But,
again, those fast algorithms may not work well on images
with severe occlusion or large head poses.
5. Conclusion
In this work, we propose a general facial landmark detection algorithm to handle images with severe occlusion and
Table 2. Comparison of facial landmark detection errors on Helen
database [16] (194 points) and LFPW database (29 points) [2].
The reported results from the original papers are marked with “*”.
algorithm
LBF [20]
SDM [26]
RCPR [4]
ESR [5]
EGM [29]
COE [2]
OC [11]
CompASM [16]
STASM [18]
FPLL [30]
ours
Helen
5.41* (fast 5.80*)
5.82
6.50*
5.70*
9.10*
11.1
5.49
LFPW
3.35*(fast 3.35*)
3.47*
3.50*
3.5*
3.98*
3.99*
5.07*
10.91
3.93
(less data)
images with large head poses. We iteratively update the
landmark visibility probabilities and landmark locations.
For occlusion prediction, we train one unified model to handel different kinds of occlusion and we explicitly add the
prior occlusion pattern as the constraint. For landmark detection, we treat points differently and rely more on the
information from points with high visibility probabilities.
The experimental results show that the proposed method is
significantly better than the other state-of-the-art works on
images with severe occlusion and images with large head
poses. It is also comparable to other methods on general
and less challenging “in-the-wild” images.
In the future, we would further improve the algorithm in
two directions. First, we would extend the detection algorithm for realtime tracking. Second, we would improve the
algorithm so that it can handle more challenging cases in
real world conditions (e.g. significant illumination change,
low resolution, etc.).
References
[1] S. Baker, R. Gross, and I. Matthews. Lucas-kanade 20 years
on: A unifying framework: Part 3. International Journal of
Computer Vision, 56:221–255, 2002. 2
[2] P. N. Belhumeur, D. W. Jacobs, D. J. Kriegman, and N. Kumar. Localizing parts of faces using a consensus of exemplars. In The 24th IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), June 2011. 1, 2, 6, 8
[3] Y. Bengio. Learning deep architectures for ai. Found. Trends
Mach. Learn., 2(1):1–127, Jan. 2009. 4
[4] X. P. Burgos-Artizzu, P. Perona, and P. Dollar. Robust face
landmark estimation under occlusion. In IEEE International
Conference on Computer Vision, pages 1513–1520, 2013. 1,
2, 3, 4, 6, 7, 8
[5] X. Cao, Y. Wei, F. Wen, and J. Sun. Face alignment by explicit shape regression. International Journal of Computer
Vision, pages 177–190, 2014. 2, 6, 7, 8
[6] T. F. Cootes, G. J. Edwards, and C. J. Taylor. Active appearance models. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 23(6):681–685, June 2001. 2
[7] T. F. Cootes, M. C. Ionita, C. Lindner, and P. Sauer. Robust
and accurate shape model fitting using random forest regression voting. In Proceedings of the 12th European Conference on Computer Vision - Volume Part VII, ECCV’12, pages
278–291, 2012. 2
[8] D. Cristinacce and T. Cootes. Automatic feature localisation with constrained local models. Pattern Recognition,
41(10):3054 – 3067, 2008. 2
[9] M. Dantone, J. Gall, G. Fanelli, and L. V. Gool. Real-time
facial feature detection using conditional regression forests.
In CVPR, 2012. 2
[10] Z.-H. Feng, P. Huber, J. Kittler, W. Christmas, and X.-J. Wu.
Random cascaded-regression copse for robust facial landmark detection. Signal Processing Letters, IEEE, 22(1):76–
80, Jan 2015. 1, 6, 7
[11] G. Ghiasi and C. Fowlkes. Occlusion coherence: Localizing
occluded faces with a hierarchical deformable part model.
In Computer Vision and Pattern Recognition (CVPR), 2014
IEEE Conference on, pages 1899–1906, June 2014. 1, 3, 6,
7, 8
[12] R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker.
Multi-pie. Image Vision Computing, 28(5):807–813, May
2010. 6, 7
[13] G. E. Hinton and R. R. Salakhutdinov. Reducing the
dimensionality of data with neural networks. Science,
313(5786):504–507, July 2006. 4
[14] X. Jia, H. Yang, K.-P. Chan, and i. Patras. Structured semisupervised forest for facial landmarks localization with face
mask reasoning. In Proceedings of the British Machine Vision Conference. BMVA Press, 2014. 1
[15] V. Kazemi and J. Sullivan. One millisecond face alignment
with an ensemble of regression trees. In Computer Vision
and Pattern Recognition (CVPR), 2014 IEEE Conference on,
pages 1867–1874, June 2014. 2
[16] V. Le, J. Brandt, Z. Lin, L. Bourdev, and T. S. Huang. Interactive facial feature localization. In Proceedings of the 12th
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
European Conference on Computer Vision - Volume Part III,
ECCV’12, pages 679–692, 2012. 1, 6, 8
I. Matthews and S. Baker. Active appearance models revisited. International Journal of Computer Vision, 60(2):135–
164, Nov. 2004. 2
S. Milborrow and F. Nicolls. Locating facial features with an
extended active shape model. In Proceedings of the 10th European Conference on Computer Vision: Part IV, ECCV ’08,
pages 504–513, Berlin, Heidelberg, 2008. Springer-Verlag. 8
P. J. Phillips, H. Moon, P. Rauss, and S. A. Rizvi. The feret
evaluation methodology for face-recognition algorithms. In
Proceedings of the 1997 Conference on Computer Vision and
Pattern Recognition (CVPR ’97), CVPR ’97, pages 137–,
Washington, DC, USA, 1997. IEEE Computer Society. 1,
2, 6, 8
S. Ren, X. Cao, Y. Wei, and J. Sun. Face alignment at 3000
fps via regressing local binary features. In Computer Vision
and Pattern Recognition (CVPR), 2014 IEEE Conference on,
pages 1685–1692, June 2014. 2, 6, 8
C. Sagonas, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic.
A semi-automatic methodology for facial landmark annotation. In Computer Vision and Pattern Recognition Workshops
(CVPRW), 2013 IEEE Conference on, pages 896–903, June
2013. 8
J. M. Saragih, S. Lucey, and J. F. Cohn. Deformable model
fitting by regularized landmark mean-shift. International
Journal of Computer Vision, 91(2):200–215, Jan. 2011. 2
Y. Sun, X. Wang, and X. Tang. Deep convolutional network cascade for facial point detection. In IEEE International Conference on Computer Vision and Pattern Recognition, pages 3476–3483, 2013. 2
Y. Wu and Q. Ji. Discriminative deep face shape model for
facial point detection. International Journal of Computer
Vision, pages 1–17, 2014. 2
X. Xiong and F. De la Torre. Global supervised descent
method. June 2015. 6
X. Xiong and F. De la Torre Frade. Supervised descent
method and its applications to face alignment. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), May 2013. 2, 5, 6, 7, 8
X. Yu, J. Huang, S. Zhang, W. Yan, and D. Metaxas. Posefree facial landmark fitting via optimized part mixtures and
cascaded deformable shape model. In 14th IEEE International Conference on Computer Vision, 2013. 1, 3, 7
X. Yu, Z. Lin, J. Brandt, and D. N. Metaxas. Consensus
of regression for occlusion-robust facial feature localization.
In D. Fleet, T. Pajdla, B. Schiele, and T. Tuytelaars, editors, Computer Vision – ECCV 2014, volume 8692 of Lecture Notes in Computer Science, pages 105–118. Springer,
2014. 1, 2, 3, 4
F. Zhou, J. Brandt, and Z. Lin. Exemplar-based graph matching for robust facial landmark localization. In IEEE International Conference on Computer Vision (ICCV), December
2013. 8
X. Zhu and D. Ramanan. Face detection, pose estimation,
and landmark localization in the wild. In IEEE International
Conference on Computer Vision and Pattern Recognition,
pages 2879–2886, 2012. 1, 3, 6, 7, 8
| 1cs.CV
|
Published as a conference paper at ICLR 2017
D EEP L EARNING WITH
DYNAMIC C OMPUTATION G RAPHS
Moshe Looks, Marcello Herreshoff, DeLesley Hutchins & Peter Norvig
Google Inc.
{madscience,marcelloh,delesley,pnorvig}@google.com
arXiv:1702.02181v2 [cs.NE] 22 Feb 2017
A BSTRACT
Neural networks that compute over graph structures are a natural fit for problems
in a variety of domains, including natural language (parse trees) and cheminformatics (molecular graphs). However, since the computation graph has a different
shape and size for every input, such networks do not directly support batched
training or inference. They are also difficult to implement in popular deep learning libraries, which are based on static data-flow graphs. We introduce a technique
called dynamic batching, which not only batches together operations between different input graphs of dissimilar shape, but also between different nodes within a
single input graph. The technique allows us to create static graphs, using popular libraries, that emulate dynamic computation graphs of arbitrary shape and size.
We further present a high-level library1 of compositional blocks that simplifies the
creation of dynamic graph models. Using the library, we demonstrate concise and
batch-wise parallel implementations for a variety of models from the literature.
1
I NTRODUCTION
Training deep neural networks directly on minimally pre-processed corpora has led to many recent
performance breakthroughs, mainly on problems in domains such as vision (Krizhevsky et al., 2012)
and natural language (Bahdanau et al., 2015) where the inputs can be cast as dense n-dimensional
arrays (henceforth tensors), or sequences of tensors. These successes exploit the effectiveness of
training via gradient descent on mini-batches of tens to hundreds of inputs, implemented using the
parallel SIMD capabilities of modern GPUs (Oh & Jung, 2004) and multi-core CPUs (Vanhoucke
et al., 2011). This, in turn has led to a proliferation of libraries making it easier to train and deploy
such models, by expressing them in terms of differentiable data-flow graphs over tensors (Abadi
et al., 2016; Theano Development Team, 2016; Collobert et al., 2011).
However, there is also a long history of neural networks that compute over structures such as parse
trees (Pollack, 1990), logical terms (Goller & Kuchler, 1996), and molecular graphs (Bianucci et al.,
2000). In these models, each distinct input has a different computation graph structure; we say that
they use dynamic computation graphs (DCGs). Such models continue to be developed and have
recently yielded superior results on problems such as sentiment classification and semantic relatedness (Tai et al., 2015; Li et al., 2015), question-answering (Andreas et al., 2016), and screening of
chemical compounds (Kearnes et al., 2016). Despite these successes, most practitioners avoid DCGs
for implementation reasons. For example, Bowman et al. (2016) assert that “because TreeRNNs use
a different model structure for each sentence ... efficient batching is impossible in standard implementations”. Moreover, even if efficient batching were possible in principle, current libraries such
as TensorFlow (Abadi et al., 2016) assume that the data-flow graph is static (i.e. is the same for each
input) and impose a significant cost to graph construction, which makes it infeasible to build a new
graph for each input.
Section 2 introduces dynamic batching, which enables efficient batching for training and inference
with DCGs. Dynamic batching runs DCGs efficiently with existing libraries that only support static
data-flow graphs; e.g. the same static graph can run a TreeRNN over any parse tree. We present
empirical results for our implementation in TensorFlow. Section 3 presents a combinator library for
concisely implementing models with DCGs using dynamic batching. Section 4 concludes.
1
The library is called TensorFlow Fold and lives at http://github.com/tensorflow/fold.
1
Published as a conference paper at ICLR 2017
2
DYNAMIC BATCHING
In deep learning libraries like TensorFlow, computations are manually batched. The computation
is expressed as a static graph of mathematical operations, such as y = σ(x · w + c), which are
polymorphic in batch size; an input x of dimensions (b, n) will yield an output of dimensions (b, m),
where b is the batch size. With DCGs, the graph of operations is not static, but is assumed to be
different for every input, so multiple inputs no longer naturally batch together in the same way. The
dynamic batching algorithm overcomes this difficulty. Given a set of computation graphs as input,
each of which has a different size and topology, it will rewrite the graphs by batching together all
instances of the same operation that occur at the same depth in the graph. The rewriting process
inserts additional concat and gather operations to move data between the batched operations; the
indices to gather encode the topology of the original input graphs.
We distinguish between individual operations appearing as nodes in the underlying data-flow graph,
such as addition or matrix-multiply, and small sub-graphs that conceptually act as functions over
tensors, such as a feed-forward layer or LSTM cell. We refer to the former as “ops”, and to the
latter as “operations.” Operations, (i.e. sub-graphs), form the building-blocks from which neural
networks with DCGs are composed; dynamic batching schedules operations, not ops. Our algorithm
requires that all operations which might be used be specified in advance, and it enumerates them for
scheduling purposes. For example, a binary TreeRNN for NLP parse trees has two operations:
embedding table lookups for words at the leaves of the tree, and RNN cells for the non-terminals.
The inputs and outputs of operations have tensor types. Each input or output may have a different
type, but all types must be fixed and fully specified in advance. A tensor type consists of a shape,
x1 , . . . xn , together with a scalar data type (e.g. float32). The inputs to an operation shall be
tensors of dimension (b, x1 , . . . xn ), where b is the batch size and x1 . . . xn is the shape of corresponding input tensor type. The outputs must all be tensors of dimension (b, y1 , . . . ym ), where
y1 , . . . ym is the shape of the corresponding output tensor type. Operations must be polymorphic
with respect to the batch size, because the batch size will change each time the operation is invoked,
depending on the topologies of the input graphs. However, their tensor types are fixed, so that it is
possible to assign a known tensor type to each edge in the input computation graph.
The dynamic batching algorithm takes a directed acyclic computation graph as input. A batch of
multiple input graphs can be treated as a single disconnected graph. Source nodes are constant
tensors, and non-source nodes are operations. Edges connect one of the outputs of a node to one of
the inputs of another node. Scheduling is performed using a greedy algorithm:
• Assign a depth to each node in the graph. Nodes with no dependencies (constants) are
assigned depth zero. Nodes with only dependencies of depth zero are assigned depth one,
nodes whose dependencies have a maximum depth of one get assigned depth two, etc.
• Insert pass-through (identity) operations so that an operation at depth d + 1 only refers to
results at depth d.
• Batch together all nodes invoking the same operation at the same depth into a single node.
• Concatenate all outputs which have the same depth and tensor type. The order of concatenation corresponds to the order in which the dynamic batching operations were enumerated.
• Assign a label (d, t, i) to each edge in the original graph, where d is the depth, t is the tensor
type, and i is the integer index for that edge into the (concatenated) outputs for d, t. The
schedule for the graph consists of the indices i for all edges, which are grouped together by
depth and operation.
In our TensorFlow implementation, each dynamic operation is instantiated once in the static
data-flow graph. The inputs to each operation are tf.gather ops, and the outputs are fed
into tf.concat ops, as described above. These TensorFlow ops are then placed within a
tf.while_loop. Each iteration of the loop will evaluate all of the operations at a particular depth.
The loop maintains state variables for each tensor type t, and feeds the output of concat for tensor
type t and iteration d into the input of the gathers at tensor type t and iteration d + 1. The indices
for gather at iteration d are drawn from the edge labels i for depth d in the schedule. The initial
values for the state variables at iteration/depth 0 are the constants in the input graph.
2
Published as a conference paper at ICLR 2017
int [] state
gather
%
gather
&
gather
gather
~
embed lookup
RNN Cell
)
gather
embed
3
embed
5
embed
cell
' x
concat
int [] state
1
float32 [128] state
cell
concat
float32 [128] state
Figure 1: The static data-flow graph created by dynamic batching for a binary TreeRNN over parse
trees (left), and input graph corresponding to the parse tree ((word1 , word3 ), word5 ) (right).
Dynamic batching allows us to construct a static TensorFlow graph that contains a single instance
of each operation, yet can emulate input graphs of arbitrary size and topology where operations may
appear an arbitrary number of times. The TensorFlow concat, gather, and while_loop ops are
all differentiable, so gradients calculations and back-propagation do not require any additional code.
For example, a binary TreeRNN as described above yields a TensorFlow data-flow graph with a
tf.while_loop whose body is shown on the left of Figure 1. Here each gather has an additional
input (the indices for the given op at the given depth) which picks out which elements the operations
are to be called with. The long downward arrows are the pass-throughs. The algorithm consumes a
tree such as the one shown on the right of Figure 1 and turns it into inputs for the gather operations
at each depth (here depth is the loop counter for the tf.while_loop.)
2.1
E XPERIMENTAL RESULTS
We have implemented dynamic batching as part of a new library, TensorFlow Fold, and designed a
synthetic speed benchmark to compare it with manual batching in native TensorFlow. The benchmark uses the same underlying kernels and execution engine in both cases. Native TensorFlow
cannot batch together trees of different shapes so, for testing purposes, we use a batch of random
binary trees, all of which have the same shape. These test results thus represent a best-case scenario,
in which all operations can be batched together perfectly. For the manual batching tests, we construct a static data-flow graph of operations corresponding to the shape of the tree. For the dynamic
batching tests, we traverse each tree to construct a schedule, as described above.
The leaves of the tree are lookups into an embedding table, while the non-terminals implement a
variant of the Tree-LSTM (Tai et al., 2015) equations. The tree size is 128, with a state size of
1024 for the LSTM. The CPU tests were run on a Dell z620 workstation with dual 8-core Intel
Xeon processors (32 hardware threads), and the GPU tests were done using a consumer Nvidia
GeForce GTX-1080 card. We compare manual batching, dynamic batching where all trees have the
same shape, and dynamic batching where each tree has a different shape (the column marked “full
dynamic”). There is no measurable penalty for dealing with trees of different shapes.
The test results shown in Table 1 emphasize the importance of batching, especially on GPUs. TensorFlow will launch a GPU kernel for every node in the tree, so there is a fixed overhead, proportional
to the size of the tree, that dominates execution for small batch sizes. TensorFlow does not begin to
saturate the GPU until relatively large batch sizes – 1024 or higher. The difference in speed between
fully-batched and unbatched is over 160x.
Dynamic batching has less kernel invocation overhead because the data-flow graph is smaller. Dynamic batching instantiates each operation only once, and invokes it once for each depth, so the
number of kernel invocations is log(n), rather than n, where n is tree size. Dynamic batching thus
achieves substantial speedups even at batch size 1, because it batches operations at the same depth
within a single tree.
3
Published as a conference paper at ICLR 2017
Table 1: Inference timing benchmark; times are wall-clock averages in seconds
batch-size
(CPU) 1024
512
256
128
64
32
1
(GPU) 1024
512
256
128
64
32
1
manual
batch
tree
14.62
0.014
7.54
0.014
4.14
0.016
2.48
0.019
1.64
0.025
1.27
0.039
0.52
0.517
0.978 0.0009
0.530 0.0010
0.312 0.0012
0.236 0.0018
0.193 0.0030
0.153 0.0047
0.161 0.1608
dynamic
batch
tree
18.68
0.018
9.84
0.019
5.22
0.020
2.95
0.023
1.76
0.027
1.05
0.032
0.26
0.258
1.590 0.0015
0.715 0.0013
0.323 0.0012
0.164 0.0012
0.093 0.0014
0.061 0.0019
0.038 0.0376
full dynamic
batch
tree
18.37
0.017
9.57
0.018
5.25
0.020
3.08
0.024
1.78
0.027
1.10
0.034
0.26
0.262
1.617 0.0015
0.721 0.0014
0.340 0.0013
0.178 0.0013
0.106 0.0016
0.074 0.0023
0.036 0.0359
cost
ratio
1.27
1.30
1.26
1.18
1.06
0.82
0.49
1.62
1.34
1.03
0.69
0.48
0.40
0.23
speedup
ratio
28.86
27.68
25.23
21.47
18.55
14.94
1.97
101.79
114.15
120.86
115.05
96.40
68.79
4.47
However, the extra concat and gather ops that dynamic batching inserts do have a cost. The “cost
ratio” column above shows the ratio between dynamic and manual batching, in the case where all
trees in the batch have the same shape. The cost is only 20% for inference on GPUs with batch-size
1, but rises to 60% for training with backpropagation. The cost is mainly visible at large batch sizes,
because it is balanced by the benefit of within-tree batching at smaller sizes.
Even with the cost, dynamic batching yields a 120x speedup over using a batch size of 1 on GPU,
and 28x on CPU. The “speedup ratio” column above shows the ratio between the per-tree time for
dynamic batching on random shapes (“full dynamic”), versus manual batching with a batch size of
1. Note that using a batch size of 1 is not actually feasible for TensorFlow, because TensorFlow has
a large graph construction overhead, which is not included in these measurements, but it may apply
to other libraries that lack such overhead.
3
A
COMBINATOR LIBRARY FOR NEURAL NETWORKS
In addition to dynamic batching, the TensorFlow Fold library provides a set of combinators that
simplify the task of constructing neural networks for DCGs. Our goal here is to show how dynamic
batching enables implementing deep learning models (which are growing ever more complex) at a
higher level of abstraction than manual batching. This in turn facilitates a more rapid feedback loop
for trying out novel model variants, and thus obtaining superior results.
The design of the library was inspired by functional programming techniques such as parser combinators (Hutton & Meijer, 1996) and arrows (Hughes, 2000). In a combinator library computations
are structured compositionally, by plugging together simpler computations in various ways. The
basic unit of computation in TensorFlow Fold is a block, essentially a function from input to output.
In a typical DCG model, the input is a graph or tree of some kind, and the output is a vector, which
can be attached to a loss for training.
For example, consider a model where the inputs are sequences of words, of varying lengths, and the
output is a sentence vector. Our library provide several different ways of handling sequences. Given
a simpler block f that operates on elements of the sequence, or g on pairs of elements, we define the
following combinators:
• Map(f ): yields [f (x1 ), f (x2 ), . . . f (xn )]. Applies f to each element of the sequence, e.g.
embedding each of the words of a sentence into RN .
• Fold(g, z): yields g(. . . g(g(z, x1 ), x2 ), . . . xn ). Applies g sequentially in a leftward
chain, e.g. running an RNN over a sequence. By default z = 0.
4
Published as a conference paper at ICLR 2017
• Reduce(g): yields g(Reduce([x1 , . . . xbn/2c ]), Reduce([xbn/2c+1 , . . . xn ])). Applies g
in a balanced tree,2 e.g. max or sum-pooling over the elements.
Note that it is not necessary to pad or truncate sequences to the same length; dynamic batching
handles sequences of differing lengths.
3.1
T YPE SYSTEM
Blocks are statically typed; each block has an input type and an output type. Types are inferred
where possible, but must be explicitly specified in some cases. A type is one of the following:
• Input denotes objects in the host language (Python), such as trees and dictionaries.
• Tensor dtype,shape denotes tensors of a particular dtype and shape.
3
• Tuple(t1 , . . . tn ), denotes a tuple of values of types t1 , . . . tn .
• Sequence(t), denotes a sequence of elements of type t, of any length.
• Void is the unit type.
For example Sequence(Sequence(Tuple(Tensor float32,[] , Tensor int8,[3,4] ))) denotes jagged arrays whose elements are pairs (float32, int83×4 ).
3.2
B LOCKS AND COMBINATORS
Blocks are composed hierarchically; a block expression is always a tree. The non-terminals in the
tree are combinators such as Map and Fold, which take simpler blocks as arguments. The leaves of
the tree are atomic blocks, which include the following:
• Scalar: Input → Tensor
Convert a Python scalar to a tensor.
• Tensor: Input → Tensor
Convert a NumPy array to a tensor.
• Function(h): [Tensor or Tuple(Tensor , . . .)] → [Tensor or Tuple(Tensor , . . .)]
Defines an operation h (see Section 2) over tensors. Operations with multiple inputs and
outputs use tuples of tensors.
• InputTransform(h): Input → Input
Applies a user-defined Python function h to pre-process the input.
In addition to the the sequence combinators described above, important combinators in the library
include the following:
• b1 >> b2 : Function composition; the output of b1 is fed to the input of b2 .
• Record({l1 : b1 , . . . ln : bn }: Input → Tuple(t1 , . . . tn )
Takes a Python dictionary or tuple as input, and applies each block bi to the field labeled
li , to yield an object of type ti . Returns a tuple of the results for all fields.
• OneOf(b1 , . . . bn ): Input → t
Conditionally dispatches on its input to one of the blocks b1 , . . . bn .
• Optional(b): Input → t
Applies b if the input is not None, otherwise returns zeros. A special case of OneOf.
• AllOf(b1 , . . . bn ): t0 → Tuple(t1 , . . . tn )
Passes its input of type t0 to each of the blocks b1 , . . . bn , returning a tuple of results.
2
Reduce uses a balanced tree rather than a chain in order to minimize computation depth and provide more
opportunities for batching.
3
Note that the leading batch size for tensors is not part of the shape of the corresponding Tensor type.
5
Published as a conference paper at ICLR 2017
split
/ word2vec
y
rnn
/ logits
expr
O g
Ac^
h
/e
/α
word
#
pair
aOy `
> pO
y
ax
px
Figure 2: Block architectures for a pipeline (Section 3.3), feed-forward attention (Section 3.4),
binary Tree-LSTMs (Section 3.5), and the weave module for molecule graphs (Section 3.6).
3.3
P IPELINES
Assume we have a set of (text, label) pairs as input and wish to predict the label from the
text. The text consists of words, and we want to use an array of pretrained word embeddings
(word_matrix) and corresponding dictionary mapping words to indices (word_idx). We call
word_idx.get(word) to obtain the index of word in word_matrix, or None if word is unknown.
We start by creating a block which embeds each word into a continuous space:
word2vec = (InputTransform(word_idx.get) >>
Optional(Scalar('int32')) >>
Function(Embedding(initializer=word_matrix)))
This block uses an InputTransform to get the index of a word, which is passed to an Optional
block that converts the scalar index to a tensor (or 0 if None). This in turn gets passed to an
Embedding operation, which performs a lookup into an embedding table.
With word2vec in hand, we can define text2vec, which embeds sentences:
split = InputTransform(str.split)
rnn_cell = Concat() >> Function(FC(d, activation=tf.nn.relu))
text2vec = split >> Map(word2vec) >> Fold(rnn_cell, Zeros(d))
We use an InputTransform to split the string into words. Then we map the words to vectors with
word2vec, and combine the word vectors with a simple RNN, which uses a single fully connected
layer FC with d hidden units. The Zeros block defines the initial state for the RNN.
Assume there are n labels; we use a linear layer with n outputs to get unscaled logits:
text2logits = text2vec >> Function(FC(n, activation=None))
For training, we create a Record block to convert the label to a tensor as well, and calculate loss:
record = Record([('text', text2logits),
('label', Scalar('int32'))])
loss = record >> Function(tf.nn.sparse_softmax_cross_entropy)
Finally, we create a Compiler, which validates a block, performs type-checking, and sets up dynamic batching in TensorFlow. Outputs of a compiled block are available as TensorFlow tensors, so
training now proceeds as it would for any other TensorFlow model:
compiler = Compiler.create(loss)
cross_entropy = Compiler.output_tensors[0]
train_op = tf.train.AdamOptimizer().minimize(cross_entropy)
6
Published as a conference paper at ICLR 2017
3.4
C OMPLEX COMPOSITIONS
Recently, Raffel & Ellis (2016) have introduced an attention model for feed-forward neural networks. The model generalizes average-pooling and is defined as:
T
X
exp(et )
et = a(ht ), αt = PT
,c =
αt ht
t=1
k=1 exp(ek )
(1)
where a is a learnable function.
In this model, the block architecture is not a simple pipeline (i.e. a composition using >>) but instead
forms a directed acyclic graph, as illustrated in Figure 2. A Composition block allows blocks to
be composed into DAGs. The model code and details may be found in Appendix A.
3.5
R ECURSIVE DEFINITIONS
N -ary Tree-LSTMs (Tai et al., 2015, sec. 3.2) generalize LSTMs from 1 to N previous states. In Tai
et al. (2015, sec. 5.1) they are applied to classify sentences from the Stanford Sentiment Treebank.
This corpus consists of binarized constituency parse trees of one-sentence movie reviews, where
every node has a sentiment label. At the leaves of the tree, words are mapped to word-embedding
vectors which serve as the input to a binary tree-LSTM with 0 for the previous states. At the internal
nodes, the LSTM takes 0 as input, and previous states from its two children. More formally,
hword = T reeLST M (Embedding(word), 0, 0)
hlef t,right = T reeLST M (0, hlef t , hright )
(2)
(3)
where T reeLST M (x, hlef t , hright ) is a learnable function corresponding to Tai et al. (2015) eqs.
9-14 with N = 2. Since a tree is a recursive data type, a model that processes trees must be
recursively defined, as illustrated by the cycle in Figure 2. A ForwardDeclaration allows the
creation of recursive models:
expr = ForwardDeclaration()
word = AllOf(Record([('word', word2vec)]),
Zeros((state_size, state_size))
pair = AllOf(Zeros(embedding_size),
Record([('left', expr()), ('right', expr())]))
expr_def = (OneOf(key_fn=len, case_blocks=[(1, word), (2, pair)]) >>
TreeLSTM(state_size))
expr.resolve_to(expr_def)
A forward declaration like expr is not itself a block, but may be called (using the expr() syntax)
to create references – i.e. blocks which refer to the declaration. The subsequent call to resolve_to
then updates all the references to refer to expr_def.
The word2vec block is as defined in Section 3.3.
3.5.1
E XPERIMENTAL R ESULTS
Here we briefly report on some experiments with our implementation of N -ary Tree-LSTMs for
sentiment analysis. While we set a new state-of-the-art, that is not really the point here. Our models
are not particularly original, and could certainly be implemented without using TensorFlow Fold.
What Fold does is to enable simpler and more concise definitions (see Table 3), along with faster
execution, thus making it easier to rapidly explore novel model variants.
We used constituency Tree-LSTMs with tuned Glove vectors for word embedding, which achieved
the best results of all sentiment models presented in Tai et al. (2015). In addition to this specific
model, we have explored several novel variants.4 In particular, Tai et al. (2015) employed non4
Unsuccessful variants included standard LSTMs (i.e. having only a single forget gate) accepting pooled
histories from their children, and models based on character rather than word-level embeddings.
7
Published as a conference paper at ICLR 2017
Table 2: Test set accuracies on the Stanford Sentiment Treebank
model
fine-grained binary
Tai et al. (2015)
51.0 (0.5)
88.0 (0.3)
Munkhdalai & Yu (2016a) 52.8
89.7
Munkhdalai & Yu (2016b) 53.1
89.3
Ours (Single Model)
52.3 (0.7)
89.4 (0.4)
Ours (Ensemble)
53.6
90.2
Table 3: Lines of code comparison
model
Feed-Forward Attention
Tree-LSTM
Graph Convolutions
ours
26
119
32
original
71
219
44
ratio
0.37
0.54
0.73
recurrent dropout and L2 weight regularization. We eliminated weight regularization in favor of the
recurrent dropout scheme introduced by Semeniuta et al. (2016) and increased the LSTM state size
from 150 to 300, leaving all other hyperparameters unchanged.
Results are shown in Table 2, including the best previously reported results. Fine-grained accuracy is
measured for all trees and calculated based on the five possible labels. Binary accuracy is measured
only for trees with non-neutral sentiment, and is based on negative vs. positive classification. The
numbers in parentheses are standard deviations. Tai et al. (2015) report five independent runs, our
results are based on thirty independent runs.5 Noting the small size of this dataset (8544/1101/2210
trees for train/dev/test), we further evaluated an ensemble consisting of these thirty independently
trained models; this variant sets a new state-of-the-art on both subtasks.
3.6
G RAPH CONVOLUTIONS
As a final example, we have used the Fold library to implement the graph convolution model introduced by Kearnes et al. (2016) for molecules, which are represented as undirected graphs of atoms.
The code is more complex than our previous examples because it involves nested Composition
blocks, and is given in Appendix B.
4
D ISCUSSION
Neural architectures with dynamic computation graphs suffer from inefficient batching and poor
tooling. Dynamic batching solves the former problem in full generality, we believe for the first time.
The SPINN architecture (Bowman et al., 2016) is an alternative stack-based approach that also enables efficient batching with DCGs, but it is limited to binary trees, and requires padding/truncation
to handle trees of different sizes. The Fold library addresses the tooling problem by providing a
high-level combinator library which is intended to make it easy for practitioners to rapidly develop
and iterate on architectures with DCGs.
The experimental results presented in section 2.1 quantify the impact of dynamic batching. The
impact of the combinator library is harder to demonstrate quantitatively. One way to approach this
(with a large grain of salt) is by comparing lines of code, which we do in Table 3, vs. the original
author’s sources. See Appendix C for details on the comparison protocol. Of course, a very short
implementation is suboptimal if it comes at the cost of flexibility. The results in Section 3.5.1 show
that models from the literature can be reimplemented in Fold, then extended to achieve superior
performance. We suspect that other models with DCGs will have quite a bit of “head room” as well,
due to simply having less work done tuning them compared with more mainstream architectures.
5
Munkhdalai & Yu (2016a;b) do not report standard deviations or number of runs.
8
Published as a conference paper at ICLR 2017
R EFERENCES
Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. TensorFlow: Large-scale machine
learning on heterogeneous systems, 2015. arXiv, 1603.04467, 2016.
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Learning to compose neural
networks for question answering. In NAACL, 2016.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly
learning to align and translate. In ICLR, 2015.
Anna Maria Bianucci, Alessio Micheli, Alessandro Sperduti, and Antonina Starita. Application of
cascade correlation networks for structures to chemistry. Applied Intelligence, 2000.
Samuel R. Bowman, Jon Gauthier, Abhinav Rastogi, Raghav Gupta, Christopher D. Manning, and
Christopher Potts. A fast unified model for parsing and sentence understanding. In NAACL, 2016.
Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A Matlab-like environment
for machine learning. In BigLearn, NIPS Workshop, 2011.
Christoph Goller and Andreas Kuchler. Learning task-dependent distributed representations by
backpropagation through structure. In ICNN, 1996.
John Hughes. Generalising monads to arrows. Science of Computer Programming, 2000.
Graham Hutton and Erik Meijer. Monadic parser combinators. Technical Report NOTTCS-TR-96-4,
1996.
Steven Kearnes, Kevin McCloskey, Marc Berndl, Vijay Pande, and Patrick Riley. Molecular graph
convolutions: moving beyond fingerprints. Journal of Computer-Aided Molecular Design, 2016.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, 2012.
Jiwei Li, Minh-Thang Luong, Dan Jurafsky, and Eudard Hovy. When are tree structures necessary
for deep learning of representations? arXiv, 1503.00185, 2015.
Tsendsuren Munkhdalai and Hong Yu. Neural semantic encoders. arXiv, 1607.04315, 2016a.
Tsendsuren Munkhdalai and Hong Yu.
1607.04492, 2016b.
Neural tree indexers for text understanding.
arXiv,
Kyoung-Su Oh and Keechul Jung. GPU implementation of neural networks. Pattern Recognition,
2004.
Jordan B Pollack. Recursive distributed representations. Artificial Intelligence, 1990.
Colin Raffel and Daniel PW Ellis. Feed-forward networks with attention can solve some long-term
memory problems. In ICLR (Workshop Track), 2016.
Stanislau Semeniuta, Aliaksei Severyn, and Erhardt Barth. Recurrent dropout without memory loss.
arXiv, 1603.05118, 2016.
Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations
from tree-structured long short-term memory networks. In NAACL, 2015.
Theano Development Team. Theano: A Python framework for fast computation of mathematical
expressions. arXiv, 1605.02688, 2016.
Vincent Vanhoucke, Andrew Senior, and Mark Z. Mao. Improving the speed of neural networks on
CPUs. In Deep Learning and Unsupervised Feature Learning, NIPS Workshop, 2011.
9
Published as a conference paper at ICLR 2017
A
F EED -F ORWARD ATTENTION
The feed-forward attention model from Section 3.4 may be implemented in Fold as follows:
attention = Composition()
with attention.scope():
h = attention.input
exp_e = Map(a >> Function(tf.exp)).reads(h)
z = (Sum() >> Broadcast()).reads(exp_e)
alpha = ZipWith(Function(tf.div)).reads(exp_e, z)
c = (ZipWith(Function(tf.mul)) >> Sum()).reads(alpha, h)
attention.output.reads(c)
Within a composition scope, blocks may be wired together with reads, provided no directed cycles
are formed. The input and output properties are used to define the overall inputs and outputs of
the composition block. This example introduces several additional block types:
• Sum is a specialization of Reduce that performs elementwise addition.
• ZipWith is a variant of Map that accepts n sequences as input and applies an n-ary function
f elementwise (stopping when the end of the shortest input sequence is reached).
• Broadcast creates a Sequence(t) from a single t, repeating the same element endlessly.
B
G RAPH CONVOLUTIONS
This section implements the graph convolution model introduced by Kearnes et al. (2016), for
molecules represented as undirected graphs of atoms. There are real-valued feature vectors for
each atom and for each distinct pair of atoms. For a molecule having N atoms, we index its atom
feature vectors as ai ∈ Rn for 1 ≤ i ≤ N . We index its pair feature vectors as pi,j ∈ Rm for
1 ≤ i, j ≤ N , where pi,j = pj,i and pi,i = 0.
The core of the graph convolution model is the weave module, which combines atom-level and
pair-level features using six learnable functions (typically fully connected ReLU layers). The weave
module can be stacked arbitrarily to create deep graph convolution models. Denoting inputs and
outputs by x and y superscripts respectively, the weave module is:
ayi
=
pyi,j =
fA (fA→A (axi ),
N
X
fP →A (pxi,j ))
j=1
fP (fA→P (axi , axj ) +
fA→P (axj , axi ), fP →P (pxi,j ))
(4)
(5)
where fA , fP , fA→A , fA→P , fP →A and fP →P are learnable functions.
It is noteworthy that the ax → py calculation involves a nested scan over the atoms; for each ai we
must calculate fA→P (axi , axj ) + fA→P (axj , axi ) for all 1 ≤ j ≤ N :
a_i_to_p = Composition()
with a_i_to_p.scope():
a_x_i = Broadcast().reads(a_i_to_p.input[0])
a_x = a_i_to_p.input[1]
f_i_j = ZipWith(Concat() >> f_a_p).reads(a_x_i, a_x)
f_j_i = ZipWith(Concat() >> f_a_p).reads(a_x, a_x_i)
p = ZipWith(Sum()).reads(f_i_j, f_j_i)
a_i_to_p.output.reads(p)
The input to the a_i_to_p composition block is (axi , ax ). It has the type
Tuple(Tensor float32,[n] , Sequence(Tensor float32,[n] )).
axi
We broadcast
over ax twice in succession to compute fA→P (axi , axj ) and fA→P (axj , axi ) for all
1 ≤ j ≤ N , yielding f_i_j and f_j_i, which are length-n sequences of vectors. We join and sum
10
Published as a conference paper at ICLR 2017
each of these vectors elementwise to obtain the ultimate output of the block, which is also a length-n
sequence of vectors. The overall weave module may now be implemented as follows:
weave = Composition()
with weave.scope():
a_x = weave.input[0]
p_x = weave.input[1]
a_to_a = Map(f_a_a).reads(a_x)
p_to_a = Map(Map(f_p_a) >> Sum()).reads(p_x)
a_y = ZipWith(Concat() >> f_a).reads(a_to_a, p_to_a)
a_to_p = ZipWith(a_i_to_p).reads(a_x, Broadcast().reads(a_x))
p_to_p = Map(Map(f_p_p)).reads(p_x)
p_y = ZipWith(ZipWith(Concat() >> f_p)).reads(a_to_p, p_to_p)
weave.output.reads(a_y, p_y)
The input to weave is (ax , px ). It has the type
Tuple(Sequence(Tensor float32,[n] ), Sequence(Sequence(Tensor float32,[m] ))).
The calculation may be understood as follows:
• a_to_a maps over ax with fA→A , going from Sequence(Tensor ) to Sequence(Tensor ).
• p_to_a maps over px with fA→P and sums along the inner dimension, reducing from
Sequence(Sequence(Tensor )) to Sequence(Tensor ).
• a_y zips a_to_a and p_to_a with fA , going from
Tuple(Sequence(Tensor ), Sequence(Tensor )) to Sequence(Tensor ).
• a_to_p broadcasts ax over itself with a_i_to_p, expanding from Sequence(Tensor ) to
Sequence(Sequence(Tensor )).
• p_to_p maps over px with fP →P , going from Sequence(Sequence(Tensor )) to
Sequence(Sequence(Tensor )).
• p_y zips a_to_p and p_to_p with fP , going from
Tuple(Sequence(Sequence(Tensor )), Sequence(Sequence(Tensor ))) to
Sequence(Sequence(Tensor )).
C
C ALCULATING LINES OF CODE
Our protocol for calculating lines6 of code is as follows:
• Define the functional unit of comparison as an input-output mapping.
• Prepare a single file that implements this functionality and nothing else.
• Remove import statements, abstract base classes, logging, file i/o, and validation logic.
• Count lines of code, ignoring blank lines and comments.7 .
F EED - FORWARD ATTENTION
The functional unit of comparison is creating the model for the variable-length experiment described
in Raffel & Ellis (2016, sec. 2.3). This includes the loss and accuracy calculations, but does not
include the training loop or the creation of training data. The original implementation8 is in Python
and uses Theano and Lasagne. The TensorFlow Fold implementation is more concise, partly due
to differences between TensorFlow and Lasagne. Fold itself reduces implementation complexity by
eliminating the need for manual batching, e.g. x.sum(axis=1) where batching is explicit over axis
0, vs. x >> Sum(), which is implicitly batched.
6
All of the implementations we examine are formatted with 80-column lines excepting the Tree-LSTM
implementation, which has a few lines that are slightly longer; we still count these as single lines.
7
The calculations were performed with cloc (https://github.com/AlDanial/cloc).
8
Commit e8fce3e from https://github.com/craffel/ff-attention.
11
Published as a conference paper at ICLR 2017
T REE -LSTM
The functional unit of comparison is creating a (binary) constituency Tree-LSTM and running an
epoch of training for the fine-grained sentiment classification task as described in Tai et al. (2015,
sec. 5.1). This does not include loading the word embeddings or dataset, which are provided as
inputs. The original implementation9 is in Lua and uses Torch. Lua terminates blocks with the end
keyword; we do not count these lines. Here, the use of Python and TensorFlow leads to substantially
more concise code than with Lua and Torch. Unlike the previous example manual batching plays
no role here, because the original implementation computes gradients and losses one tree at a time.
Fold reduces complexity here by using a OneOf block to distinguish between leaves and internal
nodes, rather than a recursive function that explicitly traverses the tree.
G RAPH CONVOLUTION
The functional unit of comparison is creating a single weave module as described in Kearnes
et al. (2016, sec. 3.3). The original implementation10 is in Python and uses TensorFlow. Here,
both implementations use the same language and deep learning library. Fold helps by eliminating the need for manual batching, as in the first example. This is particularly apparent in the
atoms-to-pairs calculation, which requires making n “copies” of an n × d matrix x to get an
n × n × d tensor. In native TensorFlow the first dimension is batch, and the copying is explicit, as
reshape(tile(x, [1, n, 1]), [batch_size, n, n, d]). In Fold, x >> Broadcast()
suffices, because the number of copies needed is determined lazily by subsequent computations.
9
10
Commit b02ad49 from https://github.com/stanfordnlp/treelstm.
Provided by Kearnes et al. (2016).
12
| 9cs.NE
|
arXiv:1611.06714v1 [math.ST] 21 Nov 2016
On the Monotonicity of the Copula Entropy
Yaniv Tenzer, and Gal Elidan
Department of Statistics, The Hebrew University of Jerusalem
Abstract
Understanding the way in which random entities interact is of key interest in numerous scientific fields. This can range from a full characterization of the joint distribution to single scalar summary statistics. In this
work we identify a novel relationship between the ubiquitous Shannon’s
mutual information measure and the central tool for capturing real-valued
non-Gaussian distributions, namely the framework of copulas. Specifically, we establish a monotonic relationship between the mutual information and the copula dependence parameter, for a wide range of copula
families. In addition to the theoretical novelty, our result gives rise to
highly efficient proxy to the expected likelihood, which in turn allows
for scalable model selection (e.g. when learning probabilistic graphical
models).
1
Introduction
Understanding the joint behavior of random entities is of great importance in
essentially all scientific fields ranging from computational biology and health
care to economics and astronomy. Accordingly, the study of joint distributions
is fundamental to all the data sciences and goes back at least to the seminal
studies of Sir Francis Galton [Galton, 1888].
In multivariate modeling, our goal may range from the task of characterizing
the full joint distribution which can be difficult, to specific summary statistics
such as correlation measures, e.g. Pearson’s correlation or mutual information.
In this work, we identify a novel and useful relationship between two central
frameworks for these tasks, namely the frameworks of copulas and information
theory.
In real-valued domains, the most prominent general purpose framework for
going beyond the multivariate normal distribution is that of copulas [Joe, 1997,
Nelsen, 2007] pioneered by Sklar [Sklar, 1959]. In a nutshell, copulas allow us to
separate the modeling of the (possibly nonparametric) univariate marginals and
that of the dependence function. Formally, given a set of univariate marginal
cumulative distribution functions {FXi }, a copula function CU (u1 , . . . , un ) is a
joint distribution over variables that are marginally uniform in the [0, 1] range,
so that
FX (x1 , . . . , xn ) = C{FXi } (FX1 (x1 ), . . . , FXn (xn )) ,
1
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
2
is a valid joint distribution.
This separation between the marginal representation and the copula function
that links them allows us, for example, to easily capture multi-modal or heavytailed distributions. Indeed, the popularity of copulas as a flexible tool for
capturing dependence has grown substantially in recent years [Elidan, 2013].
Most if not all popular copula families are governed by a dependence parameter that spans the range (or part of it) between independence and full
dependence. In fact, in the bivariate case, this is captured by a well known and
fundamental relationship between the dependence parameter of the copula and
rank-based correlation measures, such as Spearman’s rho or Kendall’s tau (see,
for example, chapter 3 in [Nelsen, 2007]).
In the field of information theory, Shanon’s mutual information measures
the reduction in entropy that the knowledge of one variable induces on another.
Mutual information is a fundamental tool used to quantify the strength of dependence between random variables [Shannon, 2001], and is used throughout
the exact sciences. Formally, Shannon’s mutual information is defined as:
Z
fX,Y (x, y)
dxdy.
(1.1)
M I(X; Y ) = fX,Y (x, y) log
fX (x)fY (y)
Whether in machine learning or physics, mutual information is often the de
facto tool for measuring correlation and/or identifying independence [Cover and
Thomas, 2012]. A natural question is thus how does this statistic relate to the
framework of copulas.
In this work we establish a relationship between the mutual information
of two variables and their corresponding copula dependence parameter. Concretely, we prove that the mutual information (equivalently the copula entropy)
is monotonic in the bivariate copula dependence parameter for a wide range of
copula families, covering the vast majority of copulas used in practice. We also
extend our results for the popular class of Archimedean copulas to higher dimensions where other measures of dependence such as Spearman’s rho or Kendall
tau are not well defined.
The monotonicity result is a theoretical one and an obvious question is
whether it has practical merit in the statistical sense. In our simulations we
show that it holds in practice using a modest number of samples. An immediate
implication is that the mutual information between different pairs of variables
can be ranked by evaluating simple statistics that are substantially simpler than
the copula entropy, namely Spearman’s rho or Kendall’s tau.
Thus, for example, given gene expression data for a large number of genes, we
can identify the two genes that have the highest mutual information without ever
evaluating the mutual information or the copula density for all pairs of genes,
a computationally formidable task. More broadly, our results facilitate highly
efficient model selection in scenarios involving many interactions, e.g. when
learning probabilistic graphical models. This practical implication is studied
in depth in our earlier paper [Tenzer and Elidan, 2013], where the theoretical
results were substantially more limited.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
3
The rest of the paper is organized as follows. In Section 2 we briefly review the necessary background on copulas, TP2, super-modular functions and
stochastic orders. In Section 3 we present our main theoretical result: broadcoverage sufficient conditions that guarantee the monotonicity of the mutual information (copula entropy) in the copula dependence parameter. We extend the
results to the class of bivariate two-parameter families in Section 3.3. We then
generalize the results for Archimedean copulas of any dimension in Section 4.
In Section 6 we show the applicability of our theoretical results in the empirical
finite-sample scenario. We finish with concluding remarks in Section 7.
2
Preliminaries
In this section we briefly review basic definitions and results related to copulas,
super modular functions and stochastic orders that will be needed in the sequel.
2.1
Copulas
A copula is a multivariate joint distribution whose univariate marginals are
uniformly distributed. Formally:
Definition 2.1. : Let U1 , . . . , Un be random variables marginally uniformly
distributed on [0, 1]. A copula function C : [0, 1]n → [0, 1] is a joint distribution
Cθ (u1 , . . . , un ) = P (U1 ≤ u1 , . . . , Un ≤ un ),
where θ are the parameters of the copula function.
Now consider an arbitrary set X = {X1 , . . . Xn } of real-valued random variables (typically not marginally uniformly distributed). Sklar’s seminal theorem
[Sklar, 1959] states that for any joint distribution FX (x), there exists a copula
function C such that
FX (x) = C(F1 (x1 ), . . . , Fn (xn )).
When the univariate marginals are continuous, C is uniquely defined.
The constructive converse, which is of central interest from a modeling perspective, is also true. Since Ui ≡ Fi is itself a random variable that is always
uniformly distributed in [0, 1], any copula function taking any marginal distributions {Ui } as its arguments, defines a valid joint distribution with marginals
{Ui }. Thus, copulas are “distribution generating” functions that allow us to
separate the choice of the univariate marginals and that of the dependence
structure.
n
F(x1 ,...,xn )
from the copula construction,
Deriving the joint density f (x) = ∂ ∂x
1 ...∂xn
assuming F has n-order partial derivatives (true almost everywhere when F is
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
4
continuous) is straightforward. Using the chain rule we can write
f (x)
=
≡
∂ n C(F1 (x1 ), . . . , Fn (xn )) Y
fi (xi )
∂F1 (x1 ) . . . ∂Fn (xn )
i
Y
c(F1 (x1 ), . . . , Fn (xn ))
fi (xi ),
(2.1)
i
where c(F1 (x1 ), . . . , Fn (xn )), is called the copula density function.
Copulas are intimately related to the fundamental concept of Mutual Information (MI) of random variables [Cover and Thomas, 2012]. In the bivariate
case, the MI of two random variables X and Y is defined as in Equation (1.1).
Denote U ≡ FX and V ≡ FY so that cU,V is the copula of the joint distribution
of X and Y Applying Equation (2.1) we get
Z
M I(X; Y ) = − cU,V (u, v) log cU,V (u, v)dudv
(2.2)
=
E[− log cU,V ] = H(CU,V ).
In words, the MI between two random variables equals to the entropy of the
corresponding copula. It is easy to see that this also applies to higher dimensions.
Copulas are also closely tied to other measures of association. The following
relationship between the copula function and Spearman’s ρX,Y is well known in
the bivariate case:
Z
ρX,Y = 12 C(FX (x), FY (y))dxdy − 3,
(2.3)
and a similar relationship is known for Kendall’s tau [Nelsen, 2007]
2.2
TP2 and Super-Modular Functions
Total positive of order two functions (TP2) [Olkin and Marshall, 2016] play a
central role in statistics and many common copula families have a TP2 density
function. Below we define the TP2 concept and provide a simply connection
to super-modular functions that will be useful for our developments. Formally,
denote u ∨ v = min(u, v), u ∧ v = max(u, v).
Definition 2.2.: A function Ψ : R2 ⇒ R is called TP2 if
∀ u, v ∈ R2
Ψ(u ∨ v) · Ψ(u ∧ v) ≥ Ψ(u) · Ψ(v).
When the inequality is reversed, the function is reverse rule of order 2 (RR2).
Definition 2.3.: A function Ψ : R2 ⇒ R is said to be super-modular if
∀ u, v ∈ R2 , Ψ(u ∨ v) + Ψ(u ∧ v) ≥ Ψ(u) + Ψ(v)
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
5
Note that if the inequality is reversed the function is called sub-modular [Olkin
and Marshall, 2016]. The following two simple results will be useful in the
sequel.
Lemma 2.4. : Let Ψ(u, v) be a positive TP2 (RR2) function. Then Φ(u, v) =
log(Ψ(u, v)) is super-modular (sub-modular).
Proof: From the definition of a TP2 function Ψ(u ∨ v) · Ψ(u ∧ v) ≥ Ψ(u) · Ψ(v),
we have
Φ(u ∨ v) + Φ(u ∧ v) ≡ log(Ψ(u ∨ v)) + log(Ψ(u ∧ v))
=
log(Ψ(u ∨ v)Ψ(u ∧ v))
≥ log(Ψ(u)Ψ(v)))
=
log(Ψ(u)) + log(Ψ(v))
≡ Φ(u) + Φ(v)
Lemma 2.5. : Let f1 (x, y), f2 (x, y) be two real non-negative TP2 (RR2) functions. Then Ψ(x, y) = f1 f2 is TP2 (RR2).
The proof is immediate and we omit the details.
2.3
PQD and Super-Modular Oredrings
Stochastic orderings introduce the notion of partial orders between random variables. Perhaps the most well known is the standard stochastic order, or using
the name more commonly used in the copula community, the positive quadrant
dependent (PQD) order:
Definition 2.6.: Let X, X0 be bivariate random vectors and let FX (u, v),
FX0 (u, v) be the corresponding distribution functions. X0 is said to be more
PQD then X if:
∀(u, v) ∈ R2 , FX (u, v) ≤ FX0 (u, v)
Another stochastic ordering that will be useful in our development is the
Super-modular (SM) ordering:
Definition 2.7.: Let X, X0 be bivariate random vectors, X0 is said to be greater
than X in the super-modular order, denoted by X 5SM X0 , if∀ Ψ such that Ψ
is super modular:
EX (Ψ(x)) 5 EX0 (Ψ(x0 ))
Note that since most bivariate families are PQD ordered by construction [Joe,
1997], an immediate result of Equation (2.3) is that Spearman’s rho is monotonic
in the copula dependence parameter, within a specific copula family.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
3
6
The Monotonicity of Mutual Information in
the Dependence Parameter for Bivariate Copulas
We now present our central result and establish a novel and elegant connection
between the copula function of the distribution of two random variables and the
mutual information between these variables (equivalently the copula entropy).
In Section 4 we generalize some of these results to higher dimensions.
Theorem 3.1.: Let Cθ (u, v) be an absolutely continuous bivariate copula, and
let X, X’ be two bivariate random vectors distributed according to the same
copula family with two different parameterizations so that X ∼ Cθ1 (u, v), X0 ∼
Cθ2 (u, v), where θ2 ≥ θ1 ≥ 0. (θ2 ≤ θ1 ≤ 0). Then −H(X) 5 −H(X0 ) if one of
the following condition holds:
(a) Cθ (u, v) is increasing (decreasing) in <SM and the copula density cθ (u, v)
is TP2 (RR2) (for all θ).
(b) Cθ (u, v) is an Archimedean copula whose generator φθ is completely monotone and satisfies the boundary condition φθ (0) = 1. In addition the copula
is increasing in <SM .
(c) Cθ,δ (u, v) is a two-parameters Archimedean copula whose generator ηθ,δ (s)
is completely monotone and satisfies the boundary condition ηθ,δ (0) = 1. In
addition the copula is increasing in <SM , with respect to θ, for a fixed δ.
(d) Cθ (u, v) is an elliptical copula.
That is, the negative copula entropy defined in Equation (2.3) is monotonic
increasing (decreasing) in θ if any of the above conditions hold for the copula
family.
We note that in the bivariate case, SM ordering is equivalent to PQD ordering [Shaked and Shanthikumar, 2007] so that the above conditions (a)-(c) can
equivalently be stated using a PQD order condition. We also note that condition (d) was proved in [Elidan, 2012] using an explicit formula of the copula
entropy and is stated here for completeness.
An immediate consequence of Theorem 3.1 and the known monotonicity of
ρs in the dependence parameter θ for PQD ordered families [Nelsen, 2007], is
Corollary 3.2.: If any of the above conditions (a)-(d) hold for a copula family,
then the magnitude of Spearman’s ρs is monotonic in the copula entropy.
Importantly, one of the above conditions holds in most if not all commonly
used bivariate copula families so that our result is widely applicable. In the
sections below, we prove the result for each of the sufficient conditions (a)-(c).
In Section 5, we survey some example families.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
3.1
7
Proof for (a): TP2 Density and SM/PQD Order
To prove the result, we are going to show that the following holds:
Z
Z
cθ1 (u, v)log(cθ1 (u, v))∂u..∂un 5
cθ2 (u, v)log(cθ1 (u, v))∂u∂v
Z
5
cθ2 (u, v)log(cθ2 (u, v))∂u∂v.
Let Ψ(u, v) = log(cθ1 (u, v)). Since cθ1 (u, v) is TP2, Ψ(u, v) is super modular.
X 5SM X0 . Thus, E(Ψ(x)) 5 E(Ψ(x0 )) for all super modular function f (u, v)
and in particular for f = Ψ(u, v), according to Lemma 2.4. Thus,
Z
Z
cθ1 (u, v)Ψ(u, v)∂u∂v 5 cθ2 (u, v)Ψ(u, v)∂u∂v.
Now, substituting the explicit form of Ψ we get the first inequality:
Z
Z
cθ1 (u, v)log(cθ1 (u, v))∂u∂v 5 cθ2 (u, v)log(cθ1 (u, v))∂u∂v.
To prove the second inequality we observe that
Z
Z
cθ2 (u, v)log(cθ1 (u, v))∂u∂v 5
cθ2 (u, v)log(cθ2 (u, v))∂u∂v
Z
cθ2 (u, v)
⇔
cθ2 (u, v)log
∂u∂v = 0
cθ1 (u, v)
⇔ KL(cθ2 ; cθ1 ) = 0
where KL(; ) is the Kullback-Leiber divergence [Cover and Thomas, 2012],
which is always non-negative. Putting these inequalities together we can conclude:
Z
−H(X) ≡
cθ1 (u, v)log(cθ1 (u, v))dudv
Z
5
cθ2 (u, v)log(cθ2 (u, v))dudv ≡ −H(X0 )
3.2
Proof for (b): Bivariate Archimedean Copulas
Archimedean copulas are probably the most common non-elliptical copulas that
allow for asymmetric or heavy-tail distributions.
Condition a in Theorem 3.1 is a fairly general one in the sense that it does not
put any restrictions on the copula family type. In particular, it does not require
the family to be elliptical or Archimedean. As we shall see below, Condition b
actually implies condition (a) in the special case of Archimedean copulas with
completely monotone generator. Among many examples of copula families for
which this holds are the Ali-Mikhail (AMH), Clayton, Frank and the Gumbel
families (see Section 5 for details).
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
8
Definition 3.3. : Let ψθ (x) : [0, ∞) → [0, ∞] be a strictly convex univariate
function that is parametrized by θ ∈ R. In addition assume that ψθ (∞) = 0.
The Archimedean copula that is generated by ψθ (x) is defined as:
Cθ (u, v) = ψθ (ψθ−1 (u1 ) + ψθ−1 (u2 )).
(3.1)
We say that ψθ (x) is the copula generator of Cθ (u, v) [Nelsen, 2007]. We consider
the subclass of copula generators:
L∞ = {ψθ : (−1)i ψθ (x)(i) ≥ 0 ∀i = 0, 1, 2.., ψθ (0) = 1}.
This is the class of generators whose derivatives alternate signs (this property
is widely known as completely monotonicity) and in addition these generators
satisfy the boundary condition ψθ (0) = 1.
The following lemma shows that an Archimedean copula whose generator
ψθ (x) is in L∞ , can be written as a mixture of two univariate CDFs. We will
then use this to show that b implies a.
Lemma 3.4.: Let Cθ (u, v) be an Archimedean copula and let ψθ be its generator
such that ψθ ∈ L∞ . Then there exists a CDF M (α), of a positive random
variable α, and unique CDFs G1 (u), G2 (v) such that:
Z ∞
Cθ (u, v) =
G1 (u)α · G2 (v)α dM (α).
0
For the sack of completeness, we give a simple proof here. Note that a
different proof can be found in [Joe, 1997].
Proof: From Bernstein’s theorem we have that each ψθ ∈ L∞ is a Laplace
transform of some distribution function of a positive random variable α. That
is:
Z ∞
ψθ (s) =
e−sα dM (α), s ≥ 0.
0
Thus
Cθ (u, v) =
ψθ (ψθ−1 (u)
+
ψθ−1 (v))
Z
=
∞
−1
e−α(ψθ
(u)+ψθ−1 (v))
dM (α).
(3.2)
0
In addition, for any arbitrary distribution function F , and any positive random
variable M (α), there exists a unique distribution function G such that [Joe,
1997, p.84]:
Z ∞
Z ∞
F (x) =
G(x)α dM (α) =
e−α(− ln G(x)) dM (α) = ψθ (− ln G(x)),
0
0
−ψθ−1 (F (x))
Thus, G(x) = e
. In particular, if U, V are uniform, then F (u) =
−1
u, F (v) = v and therefore there exist G1 (u), G2 (v) such that G1 (u) = e−ψθ (u)
−1
and G2 (v) = e−ψθ (v) . Substituting this into Equation (3.2), we get
Z ∞
Cθ (u, v) =
G1 (u)α · G2 (v)α dM (α).
0
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
9
Finally, a mixture representation implies a TP2 density [Joe, 1997] . Therefore
if in addition PQD/SM ordering holds for these copula families, then condition
a is implied.
3.3
Proof for (d): Bivariate Two-parameters Archimedean
Copulas
Two-parameter families are used to capture more than one type or aspect of
dependence. For example, one parameter may control the strength of uppertail dependence while the other indicates concordance or the strength of lower
tail dependence. Bivariate Archimedean copulas generalize the one-parameter
families. In particular, these families have the following form:
Cθ,δ (u, v) = ψθ (−log Kδ (e−ψθ (u) , e−ψθ (v) )),
(3.3)
where Kδ is an Archimedean copula, parametrised by δ, as defined in Equation (3.1), and ψθ is a Laplace transform. Let φδ be the generator of Kδ . The
resulting copula is then also an Archimedean copula with generator ηθ,δ (s) =
ψθ (−log φδ (s)). For further details see [Joe, 1997].
For certain families, if we fix the δ parameter, the copula is then increasing
in SM order, with respect to θ. Therefore if the copula generator ηθ,δ (s), is
completely monotone and in addition the boundary condition ηθ,δ (0) = 1 holds,
then we are back to the settings of condition c in Theorem 3.1. As a result we
have the following corollary:
Corollary 3.5.: Let Cθ,δ (u, v) be a two parameters bivariate copula of the above
form specified in 3.3. Let ηθ,δ (s) = ψθ (−log φδ (s)) be its associated generator.
Then, if ηθ,δ (s) is completely monotone such that ηθ,δ (0) = 1 and Cθ,δ (u, v) is
increasing in SM order with respect to θ for a fixed δ, then the copula entropy
is increasing in θ.
The proof is immediate given our previous results and we omit the details.
Examples of bivariate two-parameters families for which the conditions given in
Corollary 3.5 hold are the BB1, BB2 and BB6 families [Joe, 1997]. See Section 5
for details.
4
Multivariate Copulas
In this section we generalize some of our results for higher dimensions. Before
presenting the main results of this section, let us introduce the following class
of positive real univariate functions:
L∗∞ = {φ : R+ → R+ |φ(0) = 0, φ(∞) = ∞, (−1)j−1 φ(j) ≥ 0, j ≥ 1}.
As we shell see, this class plays a central role in our result regarding multivariate
Archimedean copulas. We are now ready to state our main result:
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
10
Theorem 4.1. : Let Cθ (u1 , . . . , un ) be an absolutely continuous bivariate
copula, and let X, X’ be two random n-dimensional vectors such that X ∼
Cθ1 (u1 , . . . , un ), X0 ∼ Cθ2 (u1 , . . . , un ), where θ2 ≥ θ1 ≥ 0. (θ2 ≤ θ1 ≤ 0). Then
−H(X) 5 −H(X0 ) if one of the following condition holds:
(a) Cθ (u1 , . . . , un ) is increasing (decreasing) in <SM and the copula density
cθ (u1 , . . . , un ) is TP2 (RR2) (for all θ).
(b) Cθ (u1 , . . . , un ) is an Archimedean copula whose generator, φθ , is completely
monotone and satisfies the boundary condition φθ (0) = 1. In addition for
∗
all θ1 ≤ θ2 , φ−1
θ1 φθ2 ∈ L∞ .
It can be easily shown that condition a in Theorem 4.1 is sufficient in any
dimension. We omit the proof since it is essentially identical to the bivariate
case. Generalizing condition b, however, requires more work which we present
in this section.
4.1
Proof for (b): Multivariate Archimedean Copulas
Recall that our proof for b in the bivariate case relies on the fact that a completely monotone generator that satisfies the boundary condition implies a mixture representation, which in turn implies a TP2 bivariate density. We now
show that both properties also hold in the general case of Acrhimedean copulas
with completely monotone generators for which the boundary condition holds.
To complete the proof, we will then also need to show that the SM ordering
∗
holds under the composition condition φ−1
θ1 φθ2 ∈ L∞ for θ1 ≤ θ2 .
The following lemma, proved in [Joe, 1997, p.85-89] substantiates the first
property:
Lemma 4.2.: Let Cθ (u1 , ·, un ) be Archimedean copula and let ψθ be its generator
such that ψθ is completely monotone and ψθ (0) = 1. Then there exists a CDF
M (α) of a positive
R ∞ r.v. and a unique CDF-s G1 (u1 ), . . . , Gn (un ) such that:
Cθ (u1 , ..., un ) = 0 G1 (u)α . . . Gn (un )α dM (α)
The following lemma shows that the second property also holds:
Lemma 4.3. : Let Cθ (u1 , . . . , un ) be a copula that can be represented as in
Lemma 4.2. Then Cθ (u1 , . . . , un ) has a TP2 density.
R∞
Proof: We assume that Cθ (u1 , . . . , un ) = 0 G1 (u1 )α . . . Gn (un )α dM (α). Taking the derivative with respect to each argument we then have that the copula
density is:
Z ∞
Y
n
cθ (u1 , . . . , un ) = α
gi (ui )
G1 (u1 )α−1 · .. · Gn (un )α−1 dM (α).
0
i
Qn−1
Gi (ui )α−1 , we can write
Z ∞
Y
n
cθ (u1 , . . . , un ) = α
gi (ui )
φ(u1 , . . . , un−1 , α) · Gn (un )α−1 dM (α).
Using φ(u1 , . . . , un−1 , α) =
i
i=1
0
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
11
Now, for each i, Gi (ui )α−1 is TP2 in (ui , α). Using Observation 2.5 we then have
that φ(u1 , . . . R, un−1 , α) is also TP2. From this and [Karlin and Rinott, 1980]
∞
we have that 0 Q
φ(u1 , . . . , un−1 , α) · Gn (un )α−1 dM (α) is TP2 (with respect to
u1 , . . . , un ). As i gi (ui ) is trivially TP2 [Joe, 1997], using Observation 2.5
again we have that the resulting density is TP2.
We have shown that, as in the bivariate case, a completely monotone generator that satisfies the boundary condition implies a TP2 copula density. We
now characterize the conditions that also ensure SM ordering. Concretely L∗∞
provides us with the needed condition via the following theorem [Wei and Hu,
2002]:
Theorem 4.4. : Let C1 , C2 be two n-dimensional Archimedean copulas and
let φ1 , φ2 be their associated generators, respectively. If φ1 , φ2 are two Laplace
transforms (equivalently φ1 , φ2 are completely monotone and satisfy the boundary condition φi (0) = 1, i = 1, 2, [Joe, 1997]), such that φ1 φ−1
∈ L∗∞ , then
2
C1 ≤SM C2 .
Putting this and Lemma 4.3 together we get the desired result.
The conditions of our theorem for multivariate copulas may seem somewhat
obscure and therefore not of practical interest. However, we note that they
actually apply to some of the most popular multivariate Archimedean copulas,
namely the Clayton, Gumbel, Frank and Joe copulas. See Section 5 for details.
5
Examples
Below we demonstrate the wide applicability of our theory. We begin with
condition a of Theorem 3.1 that provides the broadest coverage and then also
provide examples for the others cases.
5.1
Examples Satisfying Condition a of Theorem 3.1
In all the following examples, the copulas under consideration are known to be
increasing in the PQD order and have a TP2/RR2 density function [Joe, 1997].
Therefore the entropy for these families is monotonic in θ by Theorem 3.1:
• Bivariate Normal:
0 ≤ θ ≤ 1,
Cθ (u, v) = Φθ (Φ−1 (u), Φ−1 (v)),
where Φ is the N (0, 1) cdf, and Φθ is the BVSN cdf with correlation θ.
• Bivariate Farlie-Gumbel-Morgenstern (FGM):
−1 ≤ θ ≤ 1,
Cθ (u, v) = uv + θuv(1 − u)(1 − v).
Note that for this family the density is TP2 when θ ≥ 0 and RR2 when
θ ≤ 0 [Joe, 1997]. Thus its negative entropy is monotonic increasing
in θ, when θ ∈ (0, 1] and monotonic decreasing when θ ∈ [−1, 0) [Joe,
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
12
1997]. Therefore overall the FGM copula negative entropy is monotonic
increasing in |θ|.
• Bivariate Frank:
0 ≤ θ ≤ ∞,
Cθ (u, v) = −θ
−1
−θu
−θv
log τ − 1 − e
(1 − e
/τ ,
where τ = 1 − e−θ .
• Bivariate Gumbel:
1 ≤ θ ≤ ∞,
θ
Cθ (u, v) = e−[(û)
+(v̂)θ )]1/θ
,
where û = −log(u), v̂ = −log(v).
• Bivariate Clayton:
0 ≤ θ ≤ ∞,
5.2
Cθ (u, v) = (u−θ + v −θ − 1)(−1/θ) .
Archimedean Copula Examples Statisfying condition
b in Theorem 3.1
In the following we demonstrate condition b for several Archimedean families
that are not covered by the previous section such as the Clayton/Frank/Gumbel
copulas. These examples are slightly lesser known but still useful Archimedean
copula families. In all the following examples, the copulas under consideration
are known to be increasing in the PQD order [Joe, 1997]. In addition it can be
easily verified that the boundary condition ψθ (0) = 1 holds.
In order to establish the completely monotonicity (and hence also the total
positivity of the corresponding density function) of the copula generator, we use
the following sufficient conditions [Nelsen, 2007, Widder, 1942]:
Lemma 5.1. : Let f (x), g(x) be two real univariate functions and let h1 (x) =
f (x) ◦ g(x), h2 (x) = f (x)g(x). Then:
(i) If g is completely monotonic and f is absolutely monotonic, i.e.,
for k = 0,1,2,... then h1 (x) is completely monotone.
∂f (x)
∂xk
≥0
(ii) If f is completely monotonic and g is a positive function with a completely
monotone derivative, then h1 (x) is completely monotone.
(iii) If f and g are completely monotone, then so is h2 (x).
• Bivariate Ali-Mikhail (AMH):
θ ∈ [−1, 1],
Cθ (u, v) =
uv
,
1 − θ(1 − u)(1 − v)
The generator of this copula is given by:
ψθ (t) =
1−θ
.
exp(t) − θ
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
13
This generator is completely monotone for θ ∈ (0, 1] [Jaworski et al.,
2010]. When θ ∈ [−1, 0), it can easily been shown that the corresponding
densities are RR2, using Lemma 2.5. Therefore overall the AMH copula
negative entropy is monotonic increasing in |θ|.
• Bivariate Joe:
1/θ
Cθ (u, v) = 1 − (1 − u)θ + (1 − v)θ − (1 − u)θ (1 − v)θ
,
θ ∈ [1, ∞),
The generator of this copula is given by:
ψθ (t) = 1 − (1 − exp(−t))1/θ .
Taking f (t) = 1 − t1/θ and g(t) = 1 − exp(−t), we see that f is completely
monotone and g is a positive function whose first derivative completely
monotone. Thus their composite is also completely monotone.
• Family 4.14 [Nelsen, 2007]:
θ ∈ [1, ∞),
Cθ (u, v) =
−1/θ
1 + (u
θ
− 1) + (v
−1/θ
− 1)
θ
θ −θ
,
The generator of this copula is given by:
ψθ (t) = (t1/θ + 1)−θ .
By taking f = (1 + t)−θ , g = t1/θ and repeating the same arguments as
in previous examples, we get that this generator is completely monotone.
• Family 4.19 [Nelsen, 2007]:
θ ∈ (0, ∞),
Cθ (u, v) =
ln(eθ/u
θ
,
+ eθ/v − eθ )
The generator of this copula is given by:
ψθ(t) =
θ
.
ln (t + exp(θ))
By taking f = θ/t, g = ln(t+exp(θ)) and repeating the same arguments as
in previous examples, we get that this generator is completely monotone.
5.3
Examples of Bivariate Two-parameter Families
We now provide some examples of two-parameter copula families that satisfy
condition c of Theorem 3.1. That is, they are all positively PQD/SM ordered
with respect to θ and have a completely monotone generator that satisfies the
boundary condition [Joe, 1997]. Therefore by Corollary 3.5 for fixed δ, their
entropy is monotonic in θ.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
14
Recall that bivariate two-paraneter families have the following form:
Cθ,δ (u, v) = ψ(−log Kδ (e−ψθ (u) , e−ψθ (v) )),
(5.1)
where Kδ is an Archimedean copula, parametrised by δ, as defined in Equation (3.1), and ψθ is a Laplace transform. Let φδ be the generator of Kδ . The
resulting copula is then also an Archimedean copula with generator ηθ,δ (s) =
ψθ (−log φδ (s)). For further details see [Joe, 1997].
• Family BB1 [Joe, 1997]: Taking K to be Gumbel copula and ψθ (s) =
(1 + s)−1/θ , θ ≥ 0. The resulting copula is:
1/δ −1/θ
,
Cθ,δ (u, v) = 1 + (u−θ − 1)δ + (v −θ − 1)δ
where ηθ,δ (s) = (1 + s1/δ )−1/θ .
• [Joe, 1997], bivariate BB2: Taking K to be Clayton copula and ψθ (s) =
(1 + s)−1/θ , θ ≥ 0 the resulting copula is:
−1/θ
−θ
−θ
Cθ,δ (u, v) = 1 + δ −1 log eδ(u −1) + eδ(v −1) − 1
,
where ηθ,δ (s) = [1 + δ −1 log(1 + s)]−1/θ .
• [Joe, 1997], bivariate BB6: Taking K to be Gumbel copula and ψθ (s) =
1 − (1 − e−s )1/θ , θ ≥ 1 the resulting copula is:
1 θ1
,
Cθ,δ (u, v) = 1 − 1 − exp − (− log(1 − ûθ ))δ + (− log(1 − v̂ θ ))δ δ
where û = 1 − u, v̂ = 1 − v,
5.4
ηθ,δ (s) = 1 − [1 − exp −s1/δ ]1/θ .
Examples of Multivariate Copulas
We finish with some multivariate examples that satisfy the conditions of Theorem 4.1. That is, in each example the copula generator, φθ , is completely
∗
monotone and in addition φ−1
θ1 φθ2 ∈ L∞ for θ1 ≤ θ2 [Joe, 1997].
• Multivariate Clayton:
θ1 ≤ θ2 ∈ (0, ∞),
φ−1
θ 1 φθ 2 =
1
1
(θ2 + 1)θ1 /θ2 − .
θ1
θ1
• Multivariate Gumbel:
θ1 ≤ θ2 ∈ (1, ∞),
θ1 /θ2
φ−1
.
θ1 φθ2 = t
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
15
• Multivariate Frank:
θ1 ≤ θ2 ∈ (0, ∞),
φ−1
θ1 φθ2
= − ln
(e−t (e−θ2 − 1) + 1)θ1 /θ2 − 1
e−θ1 − 1
.
• Multivariate Joe :
θ1 ≤ θ2 ∈ (0, 1),
6
θ1 /θ2
φ−1
.
φ
=
−
ln
1
−
(1
−
exp(−t))
θ1 θ2
Simulations
The theory presented in this paper suggests that, in the case of an infinite
number of samples generated from a given copula family, the entropy curve is
monotonic in the copula dependence parameter. In practice, however, we almost
always have access only to a finite number of samples. An obvious empirical
question is thus whether monotonicity approximately holds given a reasonable
number of samples, and what is the impact of the sample size on the entropy
vs. dependence parameter curve.
To answer these questions we explore the finite-sample behaviour of the
copula entropy monotonicity via a simulation study. For each of the different
theoretical scenarios discussed in the previous section, we choose representative
popular copula families. We then generate M = 1000 samples from the copula
for different values of the dependence parameter. For each of these samples, we
compute the entropy using the standard empirical estimator
P
log cθ (u)
b
H(cθ ) = − u∈D
,
M
where D denotes the set of M samples generated from a copula Cθ .
For each value of the dependence parameter, we repeat the above 50 times,
and report the mean empirical entropy along with a 95% confidence interval vs.
the value of the dependence parameter. Figure 1, first two rows, show the results
for single-parameter bivariate copula families. It is clear that with as little as
1000 samples, near-monotonicity consistently holds for all families evaluated
including both elliptical ones (Gaussian, student-T) and Archmedian copula
families (Clayton, Frank, Gumbel, Joe, AMH).
Next we turn to bivariate two parameters families. In this case we also need
to test the monotonicity along the δ axis, since as formalized in Corollary 3.5,
monotonicity holds in θ for each fixed value of δ. Results are shown in Figure 1,
third row, for the BB1 and BB6 copula families, for two different values of δ.
Results for other values of δ as well as for the BB2 family were qualitatively
similar. As before, near monotonicity is evident.
To evaluate the finite-sample monotonicity in the multivariate case, we repeat the same evaluation for several copula families of dimension 5. Results
are shown in the last row of Figure 1, and are qualitatively similar for higher
dimensions. As before, the empirical monotonicity is quiet impressive.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
16
Clayton
Frank
Gumbel
Joe
Gaussian
T, ν = 3
T, ν = 7
AMH
BB1, δ = 1.5
BB1, δ = 3
BB6, δ = 1.5
BB6, δ = 3
Clayton
Frank
Gumbel
Joe
d=2
d=2
d=2
d=5
Figure 1: Simulation study of the monotonicity of the empirical entropy in the copula dependence
parameter. Shown is the mean (red line) and 95% range (black lines) of the empirical entropy over
50 random computations using from M = 1000 samples for each value of the dependence parameter
(x-axis). We consider several popular bivariate (d=2) single parameter copula families (first and second
rows), bivariate two parameters families (third row), as well as multivariate (d=5) copula families
(fourth row).
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
17
Figure 2: The impact of sample size
on the monotonicity of the empirical
entropy as a function of the dependence parameter of the generating
copula. Shown is the mean monotonic trend measure (y-axis) as a
function of the sample size (x-axis).
The monotonicity is measured via
the proportion of consecutive values of the dependence parameter, as
defined in Equation (6.1). Result
shown is the average over 50 repetitions for the Clayton family. Results for other copula families are
qualitatively similar.
Finally, we evaluate the impact of the sample size on the extent to which
the empirical entropy is monotonic in the copula dependence parameter. We
repeat the same evaluation procedure for a wide range of sample sizes (from
500 to 10000). To measure empirical monotonicity, for a given sample size, we
measure the fraction of consecutive parameter values for which monotonicity
holds. That is, we measure the empirical rate at which the following inequality
holds:
P
P
log cθ2 (u)
u∈D log cθ1 (u)
θ1 ≤ θ2 ⇒
≤ u∈D
,
(6.1)
M
M
for two consecutive values of θ1 and θ2 . We repeat this process 50 times and
report the average. Results for the Clayton family are shown in Figure 2.
As expected, with a greater sample size, monotonicity holds more frequently
nearing 100% at just 5000 training instances. Appealingly, even at much smaller
sample sizes, monotonicity is quite appealing.
7
Conclusion
In this work we establish a novel theoretical relationship between the main general purpose framework for capturing non-Gaussian real-valued distributions,
namely copulas, and the ubiquitous Shannon’s mutual information measure that
is used throughout the exact science to quantify the dependence between random variables.
Our main result is that the mutual information between two variables (equivalently the copula entropy) is monotonic in the copula dependence parameter
for a (very) wide range of copula families. Concretely, we provide fairly general
sufficient conditions for this monotonicity that cover the vast majority of com-
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
18
monly used bivariate copulas, as well as a wide class of multivariate Archimedean
copulas.
As we demonstrated in our earlier work where the theory was substantially
less developed [Tenzer and Elidan, 2013], the monotonicity result also has practical merits. Specifically, it allows us to rank the entropy (equivalently the
expected log-likelihood) of copula-based probabilistic models by simple computation of association measures such as Spearman’s rho. This gives rise to highly
efficient model selection in scenarios involving many interactions, e.g. when
learning copula graphical models. The practical merit depends, of course, on
the finite sample behavior of the entropy. Fortunately, as our simulations clearly
show, near perfect monotonicity holds even with modest sample sizes.
Essentially all copula families are constructed so as to span some or all of the
range between the independence and full dependence copula via the dependence
parameter. We have not been able to identify a single family where the above
monotonicity does not hold empirically. However, monotonicity does appear
to hold for the Plackett copula familiy [Nelsen, 2007] and identifying further
sufficient conditions remains a future challenge. Another direction if interesting
is developing finite sample theory that will explicitly quantify the amount by
which the empirical entropy can deviate from the expected monotonic behavior.
References
Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley &
Sons, 2012.
Gal Elidan. Lightning-speed structure learning of nonlinear continuous networks. pages
355–363, 2012.
Gal Elidan. Copulas in machine learning. In Copulae in mathematical and quantitative
finance, pages 39–60. Springer, 2013.
Francis Galton. Co-relations and their measurement, chiefly from anthropometric
data. Proceedings of the Royal Society of London, 45(273-279):135–145, 1888.
Piotr Jaworski, Fabrizio Durante, Wolfgang Karl Härdle, and Tomasz Rychlik. Copula
theory and its applications: proceedings of the workshop held in Warsaw, 25-26
September 2009, volume 198. Springer Science & Business Media, 2010.
H. Joe. Multivariate models and dependence concepts. Monographs on Statistics and
Applied Probability, 73, 1997.
Samuel Karlin and Yosef Rinott. Classes of orderings of measures and related correlation inequalities. i. multivariate totally positive distributions. Journal of Multivariate Analysis, 10(4):467–498, 1980.
R. Nelsen. An Introduction to Copulas. Springer, 2007.
Ingram Olkin and Albert W Marshall. Inequalities: theory of majorization and its
applications, volume 143. Academic press, 2016.
Y. Tenzer and G. Elidan, On the monotonicity of the copula entropy
19
M. Shaked and J. Shanthikumar. Stochastic Orders. Springer, 2007.
Claude Elwood Shannon. A mathematical theory of communication. ACM SIGMOBILE Mobile Computing and Communications Review, 5(1):3–55, 2001.
A. Sklar. Fonctions de repartition a n dimensions et leurs marges. Publications de
l’Institut de Statistique de L’Universite de Paris, 8:229–231, 1959.
Y. Tenzer and G. Elidan. Speedy model selection (SMS) for copula models. Conference
on Uncertainty in Artificial Intelligence, 2013.
Gang Wei and Taizhong Hu. Supermodular dependence ordering on a class of multivariate copulas. Statistics & probability letters, 57(4):375–385, 2002.
DV Widder. Completely convex functions and lidstone series. Transactions of the
American Mathematical Society, 51(2):387–398, 1942.
| 10math.ST
|
Multicuts in Planar and Bounded-Genus Graphs
with Bounded Number of Terminals∗
Éric Colin de Verdière†
November 24, 2016
arXiv:1502.00911v3 [cs.DS] 23 Nov 2016
Abstract
Given an undirected, edge-weighted graph G together with pairs of vertices, called pairs of
terminals, the minimum multicut problem asks for a minimum-weight set of edges such that, after
deleting these edges, the two terminals of each pair belong to different connected components of the
graph. Relying on topological techniques, we provide a polynomial-time algorithm for this problem
in the case where G is embedded on a fixed surface of genus g (e.g., when p
G is planar)
and has a
fixed number t of terminals. The running time is a polynomial of degree O g 2 + gt in the input
size.
In the planar case, our result corrects an error in an extended abstract by Bentz [Int. Workshop
on Parameterized and Exact Computation, 109–119, 2012]. The minimum multicut problem is also
a generalization of the multiway cut problem, a.k.a. multiterminal cut problem; even for this special
case, no dedicated algorithm was known for graphs embedded on surfaces.
1
Introduction
The minimum cut problem is one of the most fundamental problems in combinatorial optimization
(see Schrijver [34] for a fascinating historical account), originally introduced in relation to railway
transshipment problems during the cold war [22]. In this context, the railway network is modeled by a
planar graph, each edge having a weight (its capacity), and the goal is to compute the minimum-weight
set of edges that need to be removed to disconnect two given vertices of the network, the source and
destination for a given commodity. While countless generalizations of this problem have been studied,
we are interested here in two natural extensions:
1. What if there are several commodities, corresponding to different source and destination pairs?
In other words, we are studying an instance of the minimum multicut problem: Given several
pairs of source and destination vertices, how to find a minimum-weight set of edges to disconnect
every destination vertex from its corresponding source?
2. What if the network is not planar, but includes a few tunnels and bridges? In other words, what
happens if the graph is embedded, not in the plane, but on some fixed surface?
More formally, let G = (V, E) be an undirected graph. Furthermore, let T be a subset of vertices
of G, called terminals, and let R be a set of unordered pairs of vertices in T , called terminal pairs.
A subset E 0 of E is a multicut (with respect to (T, R)) if for every terminal pair {t1 , t2 } ∈ R, the
vertices t1 and t2 are in different connected components of the graph (V, E \ E 0 ). In the minimum
multicut problem (also known as the minimum multiterminal cut problem), we assume in addition
that G is positively edge-weighted, and the goal is to find a multicut of minimum weight. We prove
that this problem is polynomial-time solvable if G is embedded on a fixed surface and the number t of
terminals is fixed. More precisely:
∗
Work supported by the French ANR Blanc project ANR-12-BS02-005 (RDAM).
CNRS, Laboratoire d’Informatique Gaspard Monge, Marne-la-Vallée, France. [email protected].
Part of this work was done while the author was at CNRS, Département d’informatique, École normale supérieure, Paris,
France.
†
1
Theorem 1.1. Assume that G is cellularly embedded on a surface (orientable
or not) of Euler genus g.
√
2 +gt)
g
O(g+t)
O
(
Then the minimum multicut problem can be solved in (g + t)
n
time, where t = |T | is
the number of terminals and n is the number of edges of G.
This is the first polynomial-time algorithm for this purpose, even when specialized
√ to either the
multiway cut problem (see below for details) or the planar version. Moreover, the nO( t) dependence
in the number of terminals is unavoidable, assuming the Exponential Time Hypothesis [30], even in
these two special cases.
Comparison with Former Work
Many instances of the minimum multicut problem are hard, even in very restricted cases. In
particular, it is NP-hard in unweighted binary trees [7] and unweighted stars [20, Theorem 3.1], and
even APX-hard in the latter case. In the case where the number of pairs of terminals is fixed and at least
three, Dahlhaus et al. [14] have proved that the problem is APX-hard in general graphs; nonetheless, it
becomes polynomial-time solvable for bounded-treewidth graphs, as proved by Bentz [2, Theorem 1],
and fixed parameter tractable in the size of the solution for unweighted graphs [6]. The problem is
even harder for directed graphs [2].
In the case where the graph is planar, the number of terminals is fixed, and they all lie on the
outer face, Bentz [3] has given a polynomial-time algorithm for the minimum multicut problem. More
recently [4], he has announced an algorithm for the same case, but removing the condition that the
terminals lie on the outer face. Unfortunately, his proof has several flaws, leaving little hope for repair
(see Appendix A). We give a faster algorithm that also works for graphs on arbitrary surfaces.
A special case that is somewhat more tractable is the multiway cut problem (a.k.a. the multiterminal cut problem); this is the case where the set of pairs of terminals comprises all possible pairs of
distinct vertices in the set of terminals T ⊂ V . In the planar case, Dahlhaus et al. [14] have proved that
it is still NP-hard, but Bateni et al. [1] have given a polynomial-time approximation scheme. Again
in the planar case, the problem is also polynomial-time solvable if the number of terminals is fixed,
as proved in the early 1990s [14, 23]. In stark contrast, the complexity of the multicut problem has
remained open until now, although it is a very natural generalization of the multiway cut problem (the
multicut problem is “dual” to the multicommodity flow problem, largely studied [33, Chapters 70–76]).
More recently,
Klein and Marx have shown that the planar multiway cut problem can be solved
√
√
in 2O(t) nO( t) time (where n is the complexity of the graph) [26]; Marx has proved that the nO( t) dependence is the best one could hope for, assuming the Exponential Time Hypothesis (ETH) [30]. Our
algorithm is more general since it deals with √
multicut, not multiway cut, and works on arbitrary surfaces;
its running time, for fixed genus, is tO(t) nO( t) ; while the tO(t) factor is slightly worse than the 2O(t) of
Klein and Marx, the second factor is the same, and optimal unless ETH is false. Since approximability
in the planar case is very different for multicut and multiway cut, our result is surprising, since it
shows that, as far as exact computation is concerned, both are (essentially) equally hard.
As a side note, it is easy to see that, to solve the multicut problem, it suffices to guess the partition
of the terminals into connected components, and to solve a multiway cut problem in a higher genus
surface. However, this reduction is not useful here, since we would get a worse dependence in t if we
were to prove the result for multiway cut on surfaces and use this reduction.
On the other hand, graph algorithms dedicated to graphs embedded on a fixed surface have
flourished during the last decade. One reason is that many graphs arising in geometric settings are
naturally embedded on a surface; another one is that the family of graphs embeddable on a fixed
surface is closed under minor, and such studies can be seen as a first step towards efficient algorithms
for minor-closed families. This line of research is also justified by the fact that testing whether a
graph of complexity n embeds on a surface of genus g can be done efficiently in the complexity of the
graph, namely, in 2O(g) n time [31]. However, the history of flow and cut problems for graphs embedded
on surfaces is rather short. Chambers et al. [9] have given an algorithm to compute a minimum
cut in a graph embedded on an orientable surface of genus g that runs in g O(g) n log n time; a very
different algorithm by Erickson and Nayyeri [17] runs in 2O(g) n log n time. Thus, the minimum cut
problem is solvable in near-linear time on graphs embeddable on a fixed orientable surface. Algorithms
2
are also available for computing global minimum cuts [16] and maximum flows [10] for graphs on
surfaces. To our knowledge, we present here the first algorithm for the minimum multicut problem (or
even the multiway cut problem) for surface-embedded graphs. The main tool for flows and cuts on
surfaces [9, 10, 16, 17] is homology, which is the appropriate algebraic formalism for dealing with graphs
separating two given vertices, but it appears to be insufficient in the multicommodity case.
Overview and Discussion of Proof Techniques
The strategy for proving Theorem 1.1 is the following. In Section 3, we first show that a multicut
corresponds, in a dual sense, to a graph drawn on S that separates all pairs of terminals; such a graph
will be called a multicut dual. Moreover, if the multicut is minimum, then this multicut dual is as
short as possible, when distances on S are measured using the cross-metric: namely, the sum of the
weights of the edges of G crossed by the multicut dual is minimum. The topological structure of the
multicut dual can be described suitably after we cut the surface open into a disk with all terminals on
its boundary (Section 4). We then show that this structure is constrained (Section 5), that we can
enumerate its various possibilities (Section 6), and (roughly) that, for each of these topologies, we can
compute a shortest multicut dual with that topology efficiently (Section 7).
At a high level, our approach follows a similar pattern to Klein and Marx [26], since they also rely
on enumerating the various candidate topologies for the dual solution and find the optimum solution
for each topology separately. This strategy is also present in Dahlhaus et al. [14] and, in a different
context, in Chambers et al. [8], which was our initial source of inspiration. The details are, however,
rather different.
Indeed, Klein and Marx [26] need a reduction to the biconnected case [26, Section 3], which is
shortcut in our approach. Also, the structural properties that we develop for the minimum multicut
problem are more involved than the known ones for the multiway cut problem; indeed, the solution,
viewed in the dual graph, has less structure in the multicut problem than in the multiway cut problem:
in the multiway cut case, it has as many faces as terminals, and thus many cycles, whereas for the
minimum multicut problem, the possible topologies are more diverse (e.g., an optimal solution could
be a single dual cycle).
Chambers et al. [8] have developed related techniques for computing a shortest splitting cycle, which
have been subsequently reused for other topological and computational problems in planar or surface
cases [9, 18, 19]. A key difference, however, is that we extend the method to work with graphs instead
of paths or cycles, which makes the arguments more complicated. We need to encode precisely the
locations of the vertices and edges of the multicut dual with respect to the cut graph; the cross-metric
setting is very convenient for this purpose, since it avoids successive transformations of the input dual
graph to mimic cutting along another graph, as done by Klein and Marx [26].
Moreover, our approach also relies on other techniques from computational topology, in particular,
homology techniques developed for the single commodity minimum cut problem [9], homotopy techniques
for shortest homotopic paths [13, 27], and treewidth techniques for the surface case [15].
Finally, we remark that we are not aware of any significantly simpler proof for the planar case: The
construction of Section 4 can be simplified, and Lemma 7.2 is standard in that case, but the overall
strategy would be the same.
2
Preliminaries
We recall here standard definitions on the topology of surfaces. For general background on topology,
see for example Stillwell [35] or Hatcher [24]. For more specific background on the topology of surfaces
in our context, see recent articles and surveys on the same topic [12, 13].
In this article, S is a compact, connected surface without boundary; g denotes its Euler genus.
Thus, if S is orientable, g ≥ 0 is even, and S is (homeomorphic to) a sphere with g/2 handles attached;
if S is non-orientable, S is a sphere with g ≥ 1 disjoint disks replaced by Möbius strips.
We consider paths drawn on S. A path p is a continuous map from [0, 1] to S; a loop is a path p
such that its two endpoints coincide: p(0) = p(1). A path is simple if it is one-to-one (except, of
course, that its endpoints p(0) and p(1) may coincide). We thus emphasize that, contrary to the
3
standard terminology in graph theory, paths may self-intersect. A simple loop is two-sided if it has
a neighborhood homeomorphic to an annulus; otherwise, it has a neighborhood homeomorphic to a
Möbius strip (which implies that the surface is non-orientable), and is one-sided .
All the graphs considered in this article may have loops and multiple edges. A drawing of a
graph G on S maps the vertices of G to points on S and the edges of G to paths on S whose endpoints
are the images of the incident vertices. An embedding of G is a “crossing-free” drawing: The images
of the vertices are pairwise distinct, and the image of each edge is a simple path intersecting the image
of no other vertex or edge, except possibly at its endpoints. We will mostly consider graph embeddings
on S. A face of an embedding of G on S is a connected component of S minus (the image of) G. A
graph is cellularly embedded on S if every face of the graph is an open disk. A cut graph of S is
a graph G embedded on S whose unique face is a disk. (In particular, a cut graph has no isolated
vertex, unless it is reduced to a single vertex and S is a sphere—this case does not occur in this paper.)
Euler’s formula states that, if G is a graph cellularly embedded on S with v vertices, e edges, and f
faces, then v − e + f = 2 − g.
Algorithmically, we can store graphs cellularly embedded on S by their combinatorial map, which
essentially records the cyclic ordering of the edges around each vertex; there are efficient data structures
for this purpose [15, 28].
3
The Cross-Metric Setting
In this section, we prove that a minimum multicut corresponds, in an appropriate sense, to a shortest
graph satisfying certain properties.
We say that a graph H embedded on S is in general position with respect to our input graph G
if there are finitely many intersection points between G and H, and each such point corresponds to
a crossing between an edge of G and an edge of H. The length of H is the sum, over all crossing
points between G and H, of the weight of the corresponding edge of G. Note that an edge of H can
cross an edge of G several times, and in such cases, the length of the edge of H is computed by taking
into account the multiplicity of intersections. In other words, G is now seen as a graph that provides
a discrete (or cross-metric) distance function on S [13]. Algorithmically, we can store a graph H in
general position with respect to G by recording the combinatorial map of the overlay of G and H,
obtained by adding vertices at each intersection point between G and H and subdividing edges of G
and H.
In the following, unless noted otherwise, all graphs drawn on S will be in general position
with respect to G. Moreover, whenever we consider distances between two points in S (not lying
on G) or lengths of paths in S, we implicitly consider them in the above cross-metric sense. In some
clearly mentioned cases below (see Proposition 4.1), we will need to consider paths p that are in general
position with respect to G, except that some of their endpoints may lie on G. In such cases, the
endpoints of p are not taken into account for determining the length of p.
A multicut dual is a graph C embedded on S such that, for every pair {t1 , t2 } ∈ R, the vertices
t1 and t2 are in different faces of C. As the terminology suggests, we have the following proposition,
which will guide our approach.
Proposition 3.1. Let C be a shortest multicut dual. Then the set E 0 of edges of G crossed at least
once by C is a minimum multicut.
Proof. E 0 is a multicut, because any path in G connecting a pair of terminals {t1 , t2 } ∈ R must cross
the multicut dual, and thus use one edge in E 0 . Moreover, the weight of E 0 is at most the length of C.
To prove that E 0 is a multicut of minimum weight, it suffices to prove that, for every multicut E 00
with weight w, there exists a multicut dual of length w. Consider the dual graph G∗ that has one
vertex inside each face of G and one edge e∗ crossing each edge e of G. The subgraph of G∗ with edge
set {e∗ | e ∈ E 00 } forms a multicut dual whose length is w.
As a side remark, it follows that the minimum multicut problem can be seen as a discrete version
of the following topological problem: Given a metric surface S with boundary, and a set R of pairs of
4
boundary components, compute a shortest graph on S that separates every pair of boundaries in R.
We are exactly solving this problem in the realm of cross-metric surfaces.
4
Planarization
Our algorithm starts by computing a cut graph K of S passing through all the terminals. We will
also need some structural properties for K, detailed in the following proposition. If S is the sphere
(equivalently, if G is planar), we could take for K a shortest spanning tree of T (with respect to the
cross-metric setting), which can be obtained using a simple modification of any algorithm for computing
minimum spanning trees in graphs. For the general case, we use a known construction, a so-called
greedy system of arcs [8]. The following proposition summarizes the properties that we will use.
Proposition 4.1. In O(n log n + (g + t)n) time, we can compute a cut graph K on S, whose O(g + t)
vertices contain T , and with O(g + t) edges, each of which is a shortest path on S. Some vertices of K
may lie on G (either on vertices or on edges).
Proof. We temporarily remove a small disk containing each terminal, the boundary of which crosses
each edge of G incident to that terminal exactly once, and crossing no other edge of G. This yields a
surface with boundary S 0 that is naturally a cross-metric surface, because the intersection of the image
of the graph G with S 0 is also a graph G0 embedded on S 0 . On that surface, we compute a system
of arcs [8, Section 5.1], namely, a set of disjoint, simple paths with endpoints on the boundary of S 0
that cut S 0 into a topological disk. Moreover, there are O(g + t) paths (by Euler’s formula), and the
aforementioned construction guarantees that each path is roughly the “concatenation” of (at most) two
shortest paths. More precisely, each path can be split into two shortest paths by inserting a degree-two
vertex on an intersection point between e and some edge of G0 . (It is actually a shortest system of
arcs [11, Conclusion].)
Putting back the disks containing the terminals, and extending the arcs slightly inside these disks
to the terminals, we obtain a cut graph K satisfying the desired properties. The number of edges of K
is still O(g + t), and the running time is O(n log n + (g + t)n) [8].
At a high level, the algorithm consists in (1) enumerating all possible “topologies” of the multicut
dual with respect to K, (2) for each of these possible topologies, computing a shortest multicut dual
with that topology, and (3) returning the overall shortest multicut dual.
5
Structural Properties of a Shortest Multicut Dual
In this section, we prove some structural properties of a shortest multicut dual.
Consider all shortest multicut duals in general position with respect to K ∪ G. Among all these, let
C0 be one that crosses K a minimum number of times. We can, without loss of generality, assume
that C0 is inclusionwise minimal, in the sense that no edge can be removed from C0 without violating
the fact that it is a multicut dual. Of course, we can assume that C0 has no isolated vertex. If C0
has a degree-one vertex, we can “prune” it, by removing it together with its incident edge. If C0
has a degree-two vertex that is not a loop, we can “dissolve” it, by removing it and identifying the
two incident edges. Thus, we can assume that C0 has minimum degree at least two, and that every
degree-two vertex is the vertex of a connected component that is a loop.
5.1
Crossing Bound
We start with an easy consequence of Euler’s formula.
Lemma 5.1. C0 has O(g + t) vertices and edges.
Proof. We first note that each face of C0 contains at least one terminal: Otherwise, let e be an edge
incident to a face not containing a terminal; we could remove e without violating the multicut dual
5
(a)
(b)
(c)
(d)
(e)
p
Figure 1: (a): The part of the multicut dual C0 close to e (depicted as a horizontal line). (b): its
modified version C00 obtained by pushing all crossings with e to a single point p. The black lines are
the loops in L, the grey ones are the other edges of C00 . (c): The configuration corresponding to a
monogon. The disk is shaded. (d): The new configuration to replace (c). (e): This configuration is not
a monogon, because of the presence of an obstacle (shown as a star).
property (because it would not change which pairs of terminals are separated), contradicting the
minimality of C0 . Thus, C0 has at most t faces.
Let us first assume that every face of C0 is a disk. If C0 contains a degree-two vertex, this means
that some connected component of C0 is a loop `. Since each face of C0 is a disk and S is connected,
the graph C0 must be connected, so it equals `, and the statement of the lemma holds. So we can
assume that C0 has minimum degree at least three. Let v, e, and f be the numbers of vertices, edges,
and faces of C0 ; we deduce that 3v ≤ 2e. Combining this with Euler’s formula v − e + f = 2 − g and
the fact that f ≤ t, we obtain that e = O(g + t), and thus also v = O(g + t), as desired.
If some faces of C0 are not disks, we can iteratively extend C0 by adding edges between existing
vertices of C0 so that no new face is created but each face is cut into a disk (see, e.g., Chambers et
al. [8, proof of Lemma 2.1]; the proof in that article extends verbatim to non-orientable surfaces).
Applying the reasoning of the previous paragraph to this new graph yields that it has O(g + t) vertices
and edges; this is also true for C0 .
The main structural property of C0 is isolated in the following lemma:
Lemma 5.2. There are O(g + t) crossings between C0 and each edge of K.
As in algorithms for other problems using the same approach [8, 9, 18, 19, 26], the proof of this
lemma consists of an exchange argument: If C0 crosses an edge of K too many times, we can replace
C0 with a no longer multicut dual that crosses K fewer times, contradicting the choice of C0 . The
proof ultimately boils down to topological considerations. Let us also mention that the only property
that we are using on the edges of K is that they are disjoint shortest paths (except possibly at their
endpoints).
Proof of Lemma 5.2. We focus on a specific edge e of K crossed by C0 , forgetting about the others.
It is convenient to put an obstacle close to each of the two endpoints of e (since e is a shortest path,
its endpoints are distinct). It is also convenient to temporarily look at the situation differently, by
forgetting about G and by modifying C0 in the vicinity of e by pushing all crossings of C0 with e to a
single point p on e (Figure 1(a, b)). This transforms C0 into another graph C00 that has p as a new
vertex. To prove the lemma, it suffices to prove that the degree of p in C00 is O(g + t). Moreover, every
non-loop edge of C00 corresponds to one of the two endpoints of an edge of C0 , and there are O(g + t)
of these by Lemma 5.1. Hence, if we let L be the one-vertex subgraph of C00 made of the loops of C00
based at p, it suffices to prove that the number of loops in L is O(g + t).
A monogon, resp. a bigon, is a face of L that is, topologically, an open disk with one, resp. two,
copies of p on its boundary and containing in its interior no obstacle, no vertex of C0 , and no terminal.
We first claim that no face of L can be a monogon. Otherwise (Figure 1(c)), some edge e0 of C0
crosses e twice consecutively, at points x and y say, such that the pieces of e and e0 between x and y
bound a disk containing in its interior no obstacle, no vertex of C0 , and no terminal. Since the disk
contains no obstacle, the boundary of the disk lies entirely on one side of e, as in Figure 1(c), and other
cases such as the one shown in Figure 1(e) cannot occur. Since the disk contains no vertex of C0 , it
6
(a)
(b)
(c)
(d)
(e)
Figure 2: The exchange argument in the two-sided case. The horizontal segment represents edge e
of K. The strips are shaded; they represent disks with no terminal, no obstacle, and no piece of C0 in
their interior.
If the sides of the strips are all on the same side of e, there is a single case (a). Replacing the top
configuration of C0 with the bottom configuration (creating two new vertices) still yields a multicut
dual (as all pairs of faces that were separated in the top configuration are still separated in the bottom
configuration, except possibly for the strips, but these contain no terminal), which is no longer than
the original (because e is a shortest path) and has less crossings with K. This is a contradiction with
the choice of C0 .
If the sides of the strips are on different sides of e, we need to distinguish according to four cases (b–e),
depending on how the sides of the strips overlap. In all cases, the same argument shows that we could
find a no longer multicut dual with fewer crossings with K, a contradiction. (We could also remark
that case (e) is impossible because it involves closed curves in C0 without vertex.)
contains no piece of C0 in its interior. We can thus replace the piece of e0 between x and y with a path
that runs along e (Figure 1(d)). This operation does not make e0 longer, since e is a shortest path; it
removes the two intersection points with e and does not introduce other crossings with K. Moreover,
since the disk contains no terminal in its interior, the resulting graph is also a multicut dual. This
contradiction with the choice of C0 proves the claim.
We will prove below that no loop in L can be incident to two bigons. Taking this fact for granted
for now, whenever one face of L is a bigon, we remove one of the two incident loops, and iterate until
there is no bigon any more. The previous fact implies that these iterations remove at most half of the
loops: If L0 is the remaining set of loops, we have |L| ≤ 2|L0 |. Furthermore, L0 has no monogon or
bigon. This latter fact, together with arguments based on Euler’s formula, implies that the number of
loops in L0 is O(g + t) [8, Lemma 2.1], because S has Euler genus g, and the total number of obstacles,
vertices of C0 , and terminals (which are the points that prevent a face that is a disk of degree one or
two to be a monogon or bigon) is O(g + t) (Lemma 5.1). (That article considers the orientable case
only, but the lemma [8, Lemma 2.1] and its proof extend directly to the non-orientable case.) This
implies that |L| = O(g + t), which proves the lemma.
So there only remains to prove that no loop in L can be incident to two bigons. Assume that such
a loop exists. On the original surface S, this corresponds to two “strips” glued together, see Figure 2,
top: Each strip is bounded by two pieces of e and two pieces of edges of C0 , and these two strips share
a common piece of edge of C0 . Since a bigon contains no obstacle, the sides of the strips contain none
of the endpoints of e. Since the interiors of these strips contain no vertex of C0 , they contain no piece
of C0 .
If S is assumed to be orientable, there are five possible cases up to symmetry, see Figure 2, top: (a)
is the case where each strip has its two sides on the same side of e, (b–e) are the various cases where
each strip has its two sides on opposite sides of e. In each case, we change C0 by modifying some edges
and possibly by adding vertices (see Figure 2, bottom). Since e is a shortest path and the new pieces
7
(a)
(b)
(c)
(d)
(e)
Figure 3: The exchange argument in the one-sided case. The argument is the same as in Figure 2; the
sole difference is that the strips are drawn with a “twist”.
“run along” e, one can check that the resulting graph is no longer than C0 ; moreover, it crosses K
fewer times. Also, each replacement may split some faces of the original graph and attach each of the
strips to some of the resulting faces, but pairs of terminals that were initially separated by C0 are still
separated by the new graph, which is thus also a multicut dual. This contradicts the choice of C0 .
If S is non-orientable, there are five other cases, because the loops in L may be one-sided. However,
an entirely similar argument as in the previous paragraph (Figure 3) allows to conclude.
5.2
Some Shortest Multicut Dual is Good
We now give a more precise description of the intersection pattern between C0 and K, using the
properties proved in the previous section. Cutting the surface S along K yields a topological disk D.
The boundary ∂D of D corresponds to copies of vertices and edges of K; there are O(g + t) of these.
The copies of the vertices of K on ∂D are called the corners of D, while the copies of the edges of K
are called the sides of D. The sides of D can be reglued pairwise to obtain S.
Let C be a graph on S in general position with respect to K ∪ G. In particular, C0 is such a graph.
Cutting S along K transforms the overlay of K and C into a graph U drawn in D (Figure 4(a)):
Each edge of that graph corresponds to a piece of an edge of K or of C; each vertex of that graph
corresponds to a vertex of K, a vertex of C, or a (four-valent) intersection point between an edge of K
and an edge of C. We denote by C̄ the subgraph of U made of the edges corresponding to pieces of
edges of C (thus, C̄ lies in the interior of D except possibly for some of its leaves), and by K̄ the
subgraph of U made of the edges corresponding to pieces of edges of K (thus, the image of K̄ is the
boundary of D).
Definition 5.3. We say that C is good if C̄ is the disjoint union of (see Figure 4(a)):
• trees with at least one vertex of degree at least two, with all their leaves on ∂D, and
• arcs, namely, edges with both endpoints on ∂D, on different sides of ∂D,
and, moreover:
• there are O(g + t) intersection points between C and each side of ∂D, and
• the total number of edges of the trees is O(g + t).
Finally, Section 5 can be summarized as follows:
Proposition 5.4. Some shortest multicut dual is good.
8
(b)
(a)
a
a
(c)
1
b
e
1
c
3
1
1
3
c
e
1
d
b
1
d
Figure 4: (a): The surface S, and a (good) graph C, viewed after cutting S along the cut graph K,
resulting in a disk D. The trees of C̄ are in thick lines, the arcs being in thin lines. (b) and (c) are
used to illustrate the proof of Proposition 6.2: (b) shows the disk D, with vertices on the boundary
corresponding to the corners and the leaves of trees of C̄; arcs connecting the same pair of edges of
this polygon are represented by a single arc with an integer indicating their number; (c) represents the
dual polygon (compared from (b), every vertex on the boundary is replaced with an edge, and every
edge is replaced with a vertex); the arcs correspond to the internal or boundary edges of this dual
polygon in thick lines; these edges can be augmented arbitrarily to a weighted triangulation of the
dual polygon (the weights of the edges are zero unless noted otherwise).
Proof. We prove that C0 is indeed good.
If some connected component of C̄0 contains a cycle, that cycle bounds a disk on S containing
no terminal (since the terminals are vertices of K), so we can remove any edge of that cycle without
affecting the fact that C0 is a multicut dual; this contradicts the choice of C0 . Therefore, C̄0 is a forest.
Since C0 has no degree-zero or degree-one vertex, C̄0 has no isolated vertex, and any degree-one
vertex of C̄0 lies on ∂D. Any component of C̄0 that is a single path must have its endpoints on different
sides of K̄; otherwise, we could take an innermost such path and “push it across” the side of C̄0
its endpoints lie in, using the same argument as in Figure 1(a, b): This would not make C0 longer,
and would decrease its number of crossings with K, contradicting the choice of C0 . This proves that
C0 satisfies the first two points of Definition 5.3. The third and fourth points trivially follow from
Lemmas 5.2 and 5.1, respectively.
6
Enumerating Topologies
Let C be a good graph on S; recall that the union of K̄ and C̄ forms a connected planar graph U .
The topology of C is the data of the combinatorial map of U , where the outer face is distinguished,
and the sides are paired. Intuitively, it describes combinatorially the relative positions of C and K.
More generally, a topology is the combinatorial map of a connected, planar graph with a distinguished
outer face and a pairing of the sides (these are the subpaths of the outer cycle connecting consecutive
degree-two vertices).
Lemma 6.1. Given a topology, we can determine whether it is the topology of a good graph that is a
multicut dual in O((g + t)2 ) time.
Proof. First, we can assume that the topology has complexity O((g + t)2 ), since otherwise it is not the
topology of a good graph. We first check that the outer boundary is a cycle. Then we glue the sides of
the outer face according to the pairing. This results in the combinatorial map, on S, of the overlay
of C (corresponding to the interior edges of the topology) and K (corresponding to the edges on the
outer face—we can check that the combinatorial map indeed is the same as that of K). All of this
takes O((g + t)2 ) time.
Deciding whether C is good is easy in time linear in the size of the topology. Deciding whether C
is a multicut dual can also be done in that amount of time: We can determine which elements of T fall
9
into which face of C, since each terminal is a vertex of K, and whether each pair of terminals in R lies
in different faces of C.
This also implies that whether C is a multicut dual or not is completely determined by its topology.
Hence the following terminology: A topology is valid if it is the topology of a good graph that is a
multicut dual. The result of this section can now be stated as follows.
Proposition 6.2. The number of valid topologies is (g + t)O(g+t) ; these topologies can be enumerated
within the same time bound.
Proof of Proposition 6.2. It suffices to prove that there are (g + t)O(g+t) possible topologies for a good
graph, and that we can enumerate them within the same time bound, because we can afterwards select
the ones that correspond to multicut duals by Lemma 6.1.
We first focus on enumerating all the possibilities for the trees of C̄, forgetting about the arcs. We
know that the total number of edges of the forest is O(g + t). Any such forest can be obtained by:
• starting with a tree τ with O(g + t) edges where all internal vertices have degree three,
• contracting an arbitrary number of edges of that tree,
• removing an arbitrary number of edges of that tree.
Moreover, to specify exactly the topology for the trees of C̄, we need to specify the cyclic ordering of
the edges around each vertex of τ , and to specify to which side of K̄ each leaf of τ belongs.
There are 2O(g+t) possibilities for the choice of the initial tree τ together with the cyclic ordering of
the edges around each vertex of τ (this is essentially a Catalan number, counting the number of binary
trees). Once this is determined, there are (g + t)O(g+t) possibilities to determine to which side of K̄
each leaf of τ belongs (actually, 2O(g+t) , but this refinement is useless here). There remains to choose
which of the O(g + t) edges to contract or delete, and to specify to which of the O(g + t) sides of K̄
each leaf of τ belongs. To conclude, there are (g + t)O(g+t) possible choices for the combinatorial map
for the union of the trees, which can be enumerated also in (g + t)O(g+t) time.
Given such a possibility for the trees, we bound the number of possibilities for choosing the arcs.
The O(g + t) leaves of the trees and the corners cut ∂D into O(g + t) intervals. The arcs connect
distinct intervals (by definition of a good graph); moreover, the arcs are pairwise disjoint. Consider the
dual polygon having one vertex per interval, such that each edge connects two consecutive intervals
along K̄ (Figure 4(b, c)). Connect two vertices of the dual polygon by an internal edge of the dual
polygon whenever there exists at least one arc connecting the corresponding pair of intervals. Since the
arcs are non-crossing, the internal edges do not cross, and form a subset of a triangulation of the dual
polygon. There are 2O(g+t) possible triangulations, since the dual polygon has O(g + t) vertices (this is
again a Catalan number). To conclude, the number of possibilities for inserting the arcs, assuming the
trees have already been chosen, is bounded from above by the number of weighted triangulations of
the dual polygon, namely, triangulations where each (internal or boundary) edge bears a non-negative
number that encodes the number of arcs of the corresponding type, which is thus O(g + t). This
number is (g + t)O(g+t) . Moreover, enumerating all these possibilities can be done in the same amount
of time.
7
Dealing With Each Valid Topology
The strategy for our algorithm is roughly as follows: For each valid topology, we compute a shortest
graph C with that topology; then we return the set of edges of G crossed by the overall shortest
graph C. Actually, we do not exactly compute an embedding of a shortest graph C; instead, we compute
a shortest drawing (possibly with crossings) of the same graph, with some homotopy constraints;
in particular, that drawing is no longer than the shortest embedding, and we prove that this also
corresponds to a minimum multicut. To get an efficient algorithm, we use dynamic programming on
small treewidth graphs [5]. The key proposition is the following.
10
√
2
Proposition 7.1. Given a valid topology, we can compute, in nO( g +gt) time, some multicut whose
weight is at most the length of each multicut dual with that topology.
The following lemma allows to use treewidth techniques for surface-embedded graphs; it seems to
be folklore, and is standard in the planar case.
Lemma 7.2. Given a graph H with p vertices, edges, and faces embedded on a surface of genus g, one
√
√
can compute in O(p p) time a path decomposition of the graph with width O( gp).
Proof. We start by removing loops and multiple edges, in time O(p). Note that the resulting graph
(still denoted by H) is, as the original one, not necessarily cellularly embedded on the surface S
of genus g. However, the combinatorial map given by H defines a cellular embedding of H on the
surface S 0 obtained by pasting a disk to each facial walk of H (this idea is recurrent, e.g., in Mohar
and Thomassen [32]). The genus g 0 of S 0 is at most that of S. (Indeed, H could be augmented to a
graph H 0 cellularly embedded on S by adding edges, without increasing the number of vertices and
faces; the cellular embeddings H 0 on S and H on S 0 have the same number of vertices and faces, but
the first one has more edges; so S has genus larger than S 0 , by Euler’s formula.)
Let h be the number of vertices of H. Given the embedding√of H on S 0 , one can compute in O(h)
time a small, balanced, planar separator A for H: a set of O( gh) vertices whose removal leaves a
planar graph with no connected component with more than 2h/3 vertices [15, Theorem 5.1]. (That
theorem states only the existence of the separator, but the proof immediately gives a linear-time
algorithm, as mentioned in the remark right after its proof.)
√
0 := H − A of width O( h0 ) in
We
then
compute
a
path
decomposition
of
the
planar
graph
H
√
O(h0 h0 ) time, where h0 is the number of vertices of H 0 , using standard techniques.
√ For example, one
can compute a small, balanced, planar separator B for H 0 made of at most c1 h0 vertices, for some
constant c1 [29]; by induction, we can assume that each component of H 0 − B has a path decomposition
of width at most 2c2 h0 /3, for some constant c2 to be chosen later; concatenating these paths arbitrarily
and adding B
path gives a path decomposition of H 0 of width at √
most
√ to all the nodes of the resulting
0
0
0
2c2 h /3 + c1 h , which is at most 2c2 h if c2 is chosen large enough. By induction, it takes O(h0 h0 )
time to compute this path decomposition explicitly. (Using more advanced techniques [21], and if one
is willing to accept an implicit representation of the path decomposition, one can achieve the same
result in linear time, but we do not care about this, since this will not improve the running time of the
overall algorithm.)
Finally, adding A to each node of the path decomposition of H 0 = H − A gives a path decomposition
√
√
of H of width O( gp) in O(p p) time.
Let C be a good graph. The crossing sequence of an edge e of C (directed arbitrarily) is the
ordered sequence of edges in K crossed by e when walking along e, together with the indication of the
orientation of each crossing (more precisely, on which side of the edge of K lies the part of e before the
crossing, and on which side lies the part of e after the crossing). Given the topology of C, one can
determine the crossing sequence of every edge of C. We say that a drawing C 0 of the (abstract) graph C
has the same topology as C if each edge of C 0 has the same crossing sequence as the corresponding
edge of C.
Lemma 7.3. Let C 0 be a drawing of a multicut dual C with the same topology as C. Then, the set of
edges of G crossed by C 0 is a multicut.
Proof. Let {t1 , t2 } ∈ R be a pair of terminals, and let f be the face of C containing t1 . The set of
edges of C that are incident exactly once to the face of C containing t1 forms an even subgraph C1
of C, in which every vertex has even degree. Moreover, C1 separates t1 from t2 . Let C10 be the drawing
of the same subgraph in C 0 . To prove our result, it suffices to prove that C10 also separates t1 from t2 ,
using the fact that the crossing sequences are the same in C1 and C10 . Although intuitive, this fact is
non-trivial, and its proof relies on two results involving homology. We will use these results as a black
box; in particular, no knowledge of homology is required here.
We can assume that C 0 has a finite number of self-intersection points, each of which is a crossing.
Therefore, C10 can be seen as an even graph embedded on S (by adding a new vertex at each crossing
11
between two edges of C10 ). Our lemma is implied by two results by Chambers et al. [9], reformulated
here in our setting (that paper only considers orientable surfaces, but the two results we use extend
immediately to non-orientable surfaces):
• if, for every edge e of K, the even graphs C1 and C10 cross e with the same parity, then they are
homologous (over Z/2Z) [9, Lemma 3.4]. In our case, since the crossing sequences are equal, C1
and C10 are homologous;
• an even graph separates t1 from t2 if and only if it is homologous, on the surface S \ {t1 , t2 }, to a
small circle around t1 [9, Lemma 3.1]. Thus, since C1 separates t1 from t2 , it is also the case
for C10 .
Proof of Proposition 7.1. Lemma 7.3 implies that it suffices to compute (the set of edges of G crossed
by) a shortest drawing with the given topology.
Let us first explain how to achieve this, assuming that the locations of the vertices are prescribed.
Let C be any multicut dual with that topology and these vertex locations; we do not know C, but know
the crossing sequence of its edges (they are determined by the topology) and the vertex locations (they
are prescribed). To achieve our goal, it suffices, for each edge e of C, to compute a shortest path with
the same crossing sequence as e and the same endpoints. We remark that algorithms for computing a
shortest homotopic path in S minus the vertex set of K precisely achieve this goal [13]: In short, we
glue copies of the disk D according to the specified crossing sequence, in such a way that edge e “lifts”
to that space Ŝe ; then we compute a shortest path in Ŝe connecting the same endpoints as that lift,
and “project” back to S. (We can assign infinitesimal crossing weights to the edges of K to ensure
that the crossing sequences of e and e0 are the same.) The complexity of D (with its internal structure
defined by the edges of G) is O((g + t)n), and the crossing sequences have total length O((g + t)2 ), so
the total complexity of the spaces Ŝe is O((g + t)3 n). Since Ŝe is planar, and since shortest paths can
be computed in linear time in planar graphs [25], this is also the complexity of computing (the set of
edges of G crossed by) a shortest graph drawing with a given topology and specified vertex locations.
To compute a shortest drawing with the given topology, over all choices of vertex locations, we
can naı̈vely enumerate all possible locations of the O(g + t) vertices. Note that it is only relevant
to consider which face of the overlay of G and K each vertex belongs to, and there are O((g + t)n)
such faces. This yields an nO(g+t) -time algorithm. To get a better running time, we use treewidth
techniques, also used by Klein and Marx in the planar multiway cut case [26]. The (abstract) graph C
defined by the specified topology has O(g + t) vertices and is embedded on a surface
with genus g.
p
Lemma 7.2 gives us, in (g + t)O(1) time, a path decomposition of C of width O( g 2 + gt). We use
standard dynamic programming on the path decomposition (rooted, e.g., at one of its endpoints).
More precisely, at each node N of the path, we have a table that indicates, for every choice of the
locations of the vertices in N , the length of the shortest drawing of the subgraph of C induced by the
vertices in N and its descendents, among those that respect the crossing sequence constraints. We
can fill in the tables by a bottom-up traversal of the path decomposition.
Since each node contains
√
p
2 +gt)
O
(
g
2
O( g + gt) vertices, the running time of the algorithm is n
.
We can now conclude the proof of Theorem 1.1:
Proof of Theorem 1.1. We compute the cut graph K in O(n log n+(g +t)n) time (Proposition 4.1), and
enumerate all valid topologies in (g + t)O(g+t)
time (Proposition 6.2). For each valid topology, we apply
√
2 +gt)
O
(
g
the algorithm of Proposition 7.1 in n
time, and return a shortest multicut found. Therefore,
√
O(g+t)
O
(
g 2 +gt) . The correctness is easy: By Proposition 3.1,
the overall running time is O((g + t)
n
it suffices to compute a multicut whose weight is at most the length of any multicut dual. By
Proposition 5.4, some shortest multicut dual has a valid topology; when this topology is chosen in the
course of the algorithm, Proposition 7.1 guarantees that we have computed a shortest multicut.
12
Acknowledgments
Many thanks to Cédric Bentz for several helpful discussions on his manuscript [4]. Thanks also to
Claire Mathieu for an inspiring discussion in the preliminary stage of this paper, to Arnaud de Mesmay
for a useful remark, and to the anonymous referees for their detailed reports, one of them suggesting
that an improvement might be possible with the treewidth technique of Klein and Marx [26].
References
[1] MohammadHossein Bateni, MohammadTaghi Hajiaghayi, Philip N. Klein, and Claire Mathieu.
A polynomial-time approximation scheme for planar multiway cut. In Proceedings of the 23rd
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 639–655, 2012.
[2] Cédric Bentz. On the complexity of the multicut problem in bounded tree-width graphs and
digraphs. Discrete Applied Mathematics, 156(10):1908–1917, 2008.
[3] Cédric Bentz. A simple algorithm for multicuts in planar graphs with outer terminals. Discrete
Applied Mathematics, 157:1959–1964, 2009.
[4] Cédric Bentz. A polynomial-time algorithm for planar multicuts with few source-sink pairs. In
International Workshop on Parameterized and Exact Computation, pages 109–119, 2012. Also in
arXiv:1206.3999.
[5] Hans Bodlaender. Dynamic programming on graphs with bounded treewidth. In Proceedings of
the 15th International Colloquium on Automata, Languages and Programming (ICALP), 1988.
[6] Nicolas Bousquet, Jean Daligault, and Stéphan Thomassé. Multicut is FPT. In Proceedings of the
43rd Annual ACM Symposium on Theory of Computing (STOC), pages 459–468, 2011.
[7] Gruia Călinescu, Cristina G. Fernandes, and Bruce Reed. Multicuts in unweighted graphs and
digraphs with bounded degree and bounded tree-width. Journal of Algorithms, 48:333–359, 2003.
[8] Erin W. Chambers, Éric Colin de Verdière, Jeff Erickson, Francis Lazarus, and Kim Whittlesey.
Splitting (complicated) surfaces is hard. Computational Geometry: Theory and Applications,
41(1–2):94–110, 2008.
[9] Erin W. Chambers, Jeff Erickson, and Amir Nayyeri. Minimum cuts and shortest homologous
cycles. In Proceedings of the 25th Annual Symposium on Computational Geometry (SOCG), pages
377–385. ACM, 2009.
[10] Erin W. Chambers, Jeff Erickson, and Amir Nayyeri. Homology flows, cohomology cuts. SIAM
Journal on Computing, 41(6):1605–1634, 2012.
[11] Éric Colin de Verdière. Shortest cut graph of a surface with prescribed vertex set. In Proceedings
of the 18th European Symposium on Algorithms (ESA), part 2, number 6347 in Lecture Notes in
Computer Science, pages 100–111, 2010.
[12] Éric Colin de Verdière. Topological algorithms for graphs on surfaces. Habilitation thesis, École
normale supérieure, 2012. Available at http://monge.univ-mlv.fr/~colinde/pub/12hdr.pdf.
[13] Éric Colin de Verdière and Jeff Erickson. Tightening nonsimple paths and cycles on surfaces.
SIAM Journal on Computing, 39(8):3784–3813, 2010.
[14] Elias Dahlhaus, David S. Johnson, Christos H. Papadimitriou, Paul D. Seymour, and Mihalis
Yannakakis. The complexity of multiterminal cuts. SIAM Journal on Computing, 23(4):864–894,
1994.
[15] David Eppstein. Dynamic generators of topologically embedded graphs. In Proceedings of the
14th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 599–608, 2003.
13
[16] Jeff Erickson, Kyle Fox, and Amir Nayyeri. Global minimum cuts in surface embedded graphs. In
Proceedings of the 23rd Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages
1309–1318, 2012.
[17] Jeff Erickson and Amir Nayyeri. Minimum cuts and shortest non-separating cycles via homology
covers. In Proceedings of the 22nd Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 1166–1176, 2011.
[18] Jeff Erickson and Amir Nayyeri. Shortest non-crossing walks in the plane. In Proceedings of the
22nd Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 297–308, 2011.
[19] Kyle Fox. Shortest non-trivial cycles in directed and undirected surface graphs. In Proceedings of
the 24th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 352–364, 2013.
[20] Naveen Garg, Vijay V. Vazirani, and Mihalis Yannakakis. Primal-dual approximation algorithms
for integral flow and multicut in trees. Algorithmica, 18:3–20, 1997.
[21] Michael T. Goodrich. Planar separators and parallel polygon triangulation. Journal of Computer
and System Sciences, 51(3):374–389, 1995.
[22] T. E. Harris and F. S. Ross. Fundamentals of a method for evaluating rail net capacities. Technical
Report Research Memorandum RM-1573, The RAND Corporation, Santa Monica, California,
1955.
[23] David Hartvigsen. The planar multiterminal cut problem. Discrete Applied Mathematics, 85:203–
222, 1998.
[24] Allen Hatcher. Algebraic topology. Cambridge University Press, 2002. Available at http:
//www.math.cornell.edu/~hatcher/.
[25] Monika R. Henzinger, Philip Klein, Satish Rao, and Sairam Subramanian. Faster shortest-path
algorithms for planar graphs. Journal of Computer and System Sciences, 55(1, part 1):3–23, 1997.
√
[26] Philip N. Klein and Dániel Marx. Solving planar k-terminal cut in O(nc k ) time. In
Proceedings of the 39th International Colloquium on Automata, Languages and Programming
(ICALP) volume 1, pages 569–580, 2012.
[27] Martin Kutz. Computing shortest non-trivial cycles on orientable surfaces of bounded genus in
almost linear time. In Proceedings of the 22nd Annual Symposium on Computational Geometry
(SOCG), pages 430–438. ACM, 2006.
[28] Sóstenes Lins. Graph-encoded maps. Journal of Combinatorial Theory, Series B, 32:171–181,
1982.
[29] Richard J. Lipton and Robert Endre Tarjan. A separator theorem for planar graphs. SIAM
Journal on Applied Mathematics, 36(2):177–189, 1979.
[30] Dániel Marx. A tight lower bound for planar multiway cut with fixed number of terminals. In
Proceedings of the 39th International Colloquium on Automata, Languages and Programming
(ICALP) volume 1, pages 677–688, 2012.
[31] Bojan Mohar. A linear time algorithm for embedding graphs in an arbitrary surface. SIAM
Journal on Discrete Mathematics, 12(1):6–26, 1999.
[32] Bojan Mohar and Carsten Thomassen. Graphs on surfaces. Johns Hopkins Studies in the
Mathematical Sciences. Johns Hopkins University Press, 2001.
[33] Alexander Schrijver. Combinatorial optimization. Polyhedra and efficiency, volume 24 of Algorithms
and Combinatorics. Springer-Verlag, Berlin, 2003.
14
[34] Alexander Schrijver. On the history of combinatorial optimization (till 1960). In Handbook of
Discrete Optimization, pages 1–68. Elsevier, 2005.
[35] John Stillwell. Classical topology and combinatorial group theory. Springer-Verlag, New York,
second edition, 1993.
A
Problems with Bentz’ Approach in the Planar Case
Here, we show several flaws in Bentz’ approach [4] for the planar version of the problem, which leaves
little hope for repair. Bentz starts by guessing the clusters, namely, the partition of the terminals
induced by the faces of the optimal multicut dual, and enumerates all possibilities for the locations
of the vertices of the multicut dual that have degree at least three. (He does so in the dual graph,
while we do it in the cross-metric setting to keep track of the crossings between the multicut dual and
the cut graph, but this is the same idea.) Then, his strategy is to consider all the possible homotopy
classes of the closed walks Ci that bound a face of the multicut dual containing the terminals of a
given cluster i. Unfortunately, it is dubious that this strategy could work; at least, the enumeration of
the homotopy classes has to be substantially more complicated than described in his article, and one
would need to consider the homotopy classes of the edges of the multicut dual, not only of the closed
walks Ci . In more detail:
• In Lemma 3, it is claimed that two cycles are homotopic in the plane minus a set of terminals if
and only if they enclose the same terminals in their interior. This is not the case (see Figure 5).
In particular, as soon as there is more than two terminals, the number of homotopy classes is
infinite;
• Lemma 5 claims the following: Assume that Ci passes through vertices ω1 , . . . , ωqi of the multicut
dual. Then we still have a multicut dual if we replace Ci with any cycle Ci0 going through
ω1 , . . . , ωqi that encloses the same set of terminals as Ci . This is not the case, as Figure 6
demonstrates. (Actually, Lemma 5 makes a slightly stronger claim, which is also contradicted by
that figure.) For the same reason, Corollary 1 does not hold.
These counterexamples show that it is not sufficient to determine whether cycles are valid (can be part
of a multicut dual) based only on which terminals they enclose (although this would indeed be very
nice, since the number of such possibilities is nO(t) ). We really need to consider the homotopy classes
of the paths, and to bound their number. This is exactly the purpose of Lemma 5.2 in our technique.
Figure 5: Two non-homotopic cycles enclosing the same set of terminals.
15
Figure 6: In a multicut dual, replacing a cycle (in thick lines) with another cycle passing through the
same vertices and enclosing the same terminals does not always give a multicut dual.
16
| 8cs.DS
|
arXiv:1307.0593v1 [math.AC] 2 Jul 2013
Saturations of powers of certain determinantal ideals
Kosuke Fukumuro, Taro Inagawa and Koji Nishida∗
Graduate School of Science, Chiba University,
1-33 Yayoi-Cho, Inage-Ku, Chiba-Shi, 263-8522, JAPAN
Abstract
Let R be a Noetherian local ring and m a positive integer. Let I be the ideal of
R generated by the maximal minors of an m × (m + 1) matrix M with entries in
R . Assuming that the grade of the ideal generated by the k-minors of M is at least
m − k + 2 for 1 ≤ ∀k ≤ m , we will study the associated primes of I n for ∀n > 0 .
Moreover, we compute the saturation of I n for 1 ≤ ∀n ≤ m in the case where R is
a Cohen-Macaulay ring and the entries of M are powers of elements that form an
sop for R .
1
Introduction
Let (R, m) be a Noetherian local ring and I an ideal of R such that dim R/I > 0 . Let n
be a positive integer. We set
[
(I n )sat =
(I n :R mi )
i>0
and call it the saturation of I n . As (I n )sat /I n ∼
= H0m (R/I n ) , where H0m ( · ) denotes the 0-th
local cohomology functor, we have (I n )sat = I n if and only if depth R/I n > 0 . Moreover,
if J is an m-primary ideal such that depth R/(I n :R J) > 0 , we have (I n )sat = I n :R J .
On the other hand, the n-th symbolic power of I is defined by
\
I (n) =
(I n Rp ∩ R) .
p ∈ MinR R/I
In order to compare (I n )sat and I (n) , let us take a minimal primary decomposition of I n ;
\
In =
Q(p) ,
p ∈ AssR R/I n
where Q(p) denotes the p-primary component. It is obvious that
\
\
(I n )sat =
Q(p) and I (n) =
m 6= p ∈ AssR R/I n
∗
p ∈ MinR R/I
The last author is supported by KAKENHI (23540042)
1
Q(p) .
Hence we have (I n )sat ⊆ I (n) and the equality holds if and only if AssR R/I n is a subset
of { m } ∪ MinR R/I . Therefore (I n )sat = I (n) if dim R/I = 1 . However, if dim R/I ≥ 2 ,
usually it is not easy to decide whether (I n )sat = I (n) or not. Furthermore, describing a
system of generators for (I n )sat /I n precisely is often very hard. In this paper, assuming
that R is an (m + 1)-dimensional Cohen-Macaulay local ring and I is an ideal generated
by the maximal minors of the following m × (m + 1) matrix;
x1
x2 x3 · · · xm xm+1
x2
x3 x4 · · · xm+1 x1
,
M =
x
x
x
·
·
·
x
x
3
4
5
1
2
...............................
xm xm+1 x1 · · · xm−2 xm−1
where x1 , x2 , . . . , xm+1 is an sop for R , we aim to prove
• (I n )sat ( I (n) if n > m ≥ 3 ,
• (I n )sat = I n if 1 ≤ n < m ,
• (I m )sat = I m :R (x1 , x2 , . . . , xm+1 )R ,
• (I m )sat /I m ∼
= R/(x1 , x2 , . . . , xm+1 )R if m ≥ 2 .
Moreover, we describe a generator of (I m )sat /I m using the determinant of a certain matrix
induced from M . The proofs of the assertions stated above are given in Section 3 and 4
taking more general matrices as M .
Throughout this paper R is a commutative ring, and we often assume that R is a
Noetherian local ring with the maximal ideal m . For positive integers m, n and an ideal
a of R , we denote by Mat(m, n ; a) the set of m × n matrices with entries in a . For any
M ∈ Mat(m, n ; R) and any k ∈ Z we denote by Ik (M) the ideal generated by the k-minors
of M . In particular, Ik (M) is defined to be R (resp. (0)) for k ≤ 0 (resp. k > min{m, n}).
If M, N ∈ Mat(m, n ; R) and the (i, j) entries of M and N are congruent modulo a fixed
ideal a for ∀(i, j) , we write M ≡ N mod a .
2
Preliminaries
In this section, we assume that R is just a commutative ring. Let m, n be positive integers
with m ≤ n and M = ( xij ) ∈ Mat(m, n ; R) . Let us recall the following rather well-known
fact.
Lemma 2.1 Suppose Im (M) ⊆ p ∈ Spec R and put ℓ = max{ 0 ≤ k ∈ Z | Ik (M) 6⊆ p } .
Then ℓ < m and there exists N ∈ Mat(m − ℓ, n − ℓ ; pRp ) such that Ik (M)p = Ik−ℓ (N) for
any k ∈ Z .
Proof. We prove by induction on ℓ . The assertion is obvious if ℓ = 0 . So, let us consider
the case where ℓ > 0 . Then I1 (M) 6⊆ p , and so some entry of M is a unit in Rp . Hence,
2
applying elementary operations to M in Mat(m, n ; Rp ) , we get a matrix of the form
1 0 ··· 0
0
,
..
′
.
M
0
where M ′ ∈ Mat(m − 1, n − 1 ; Rp ) . It is easy to see that Ik (M)p = Ik−1 (M ′ ) for any
k ∈ Z . Hence ℓ − 1 = max{ 0 ≤ k ≤ m − 1 | Ik (M ′ ) 6⊆ pRp } . By the hypothesis of
induction, there exists
N ∈ Mat( (m − 1) − (ℓ − 1) , (n − 1) − (ℓ − 1) ; pRp ) = Mat(m − ℓ, n − ℓ ; pRp )
such that It (M ′ ) = It−(ℓ−1) (N) for any t ∈ Z . Then we have Ik (M)p = Ik−ℓ (N) for any
k ∈ Z.
In the rest of this section, we assume n = m + 1 . For 1 ≤ ∀j ≤ m + 1 , Mj denotes
the m × m submatrix of M determined by removing the j-th column. We set aj =
(−1)j−1 · det Mj and I = (a1 , a2 , . . . , am+1 )R = Im (M) . Let us take an indeterminate t
over R and consider the Rees algebra of I ;
R(I) := R[a1 t, a2 t, . . . , am+1 t] ⊆ R[ t ] ,
which is a graded ring such that deg aj t = 1 for 1 ≤ ∀j ≤ m + 1 . On the other hand, let
S = R[T1 , T2 , . . . , Tm+1 ] be a polynomial ring over R with m + 1 variables. We regard S
as a graded ring by setting deg Tj = 1 for 1 ≤ ∀j ≤ m + 1 . Let π : S −→ R(I) be the
homomorphism of R-algebras such that π(Tj ) = aj t for 1 ≤ ∀j ≤ m + 1 . Then π is a
surjective graded homomorphism. Now we set
fi =
m+1
X
xij Tj ∈ S1
j=1
for 1 ≤ ∀i ≤ m . It is easy to see (f1 , f2 , . . . , fm )S ⊆ Ker π . For our purpose, the following
result due to Avramov [1] is very important (Another elementary proof is given in [3]).
Theorem 2.2 Suppose that R is a Noetherian ring. If grade Ik (M) ≥ m − k + 2 for
1 ≤ ∀k ≤ m , then Ker π = (f1 , f2 , . . . , fm )S and grade (f1 , f2 , . . . , fm )S = m .
As the last preliminary result, we describe a technique using determinants of matrices.
Suppose that y1 , y2 , . . . , ym+1 are elements of R such that
y1
0
y2 0
M . = . .
.. ..
ym+1
0
We put y = y1 + y2 + · · · + ym+1 .
3
Lemma 2.3 If y, yk form a regular sequence for some k with 1 ≤ k ≤ m + 1 , then there
exists δ ∈ R such that yj · δ = aj for 1 ≤ ∀j ≤ m + 1 .
Proof. We put a = a1 + a2 + · · · + am+1 . Then the following assertion holds:
Claim y · aj = yj · a for 1 ≤ ∀j ≤ m + 1 .
In order to prove the claim above, let us consider the following (m + 1) × (m + 1) matrix:
1 1 ··· 1
.
N =
M
Expanding det N along the first row, we get det N = a . We fix j with 1 ≤ j ≤ m + 1 .
Multiplying the j-th column of N by yj , we get
j
1
x11
N′ =
.
..
xm1
···
···
`
yj
x1j yj
..
.
···
···
1
x1,m+1
..
.
···
xmj yj
···
xm,m+1
.
Then det N ′ = yj · det N = yj · a . Next, for 1 ≤ ∀ℓ ≤ m + 1 with ℓ 6= j , we add the ℓ-th
column of N ′ multiplied by yℓ to the j-th column, and get
j
1
x11
N ′′ =
.
..
xm1
···
···
`
y ···
0 ···
..
.
1
x1,m+1
..
.
···
0
xm,m+1
···
since our assumption means
,
xi1 y1 + · · · + xij yj + · · · + xi,m+1 ym+1 = 0
for 1 ≤ ∀i ≤ m . Then det N ′′ = det N ′ = yj · a . Finally, replacing the first j columns of
N ′′ , we get
y 1 ··· 1
0
N ′′′ = .
.
..
M
j
0
Then y · aj = y · (−1)j−1 · det Mj = (−1)j−1 · det N ′′′ = det N ′′ = yj · a . Thus we get the
equalities of the claim.
Now we take k with 1 ≤ k ≤ m + 1 so that y, yk form a regular sequence. Because
y ·ak = yk ·a , there exists δ ∈ R such that a = yδ . Then y ·aj = yj ·yδ for 1 ≤ ∀j ≤ m+1 .
As y is an R-NZD, we get aj = yj · δ for 1 ≤ ∀j ≤ m + 1 , and the proof is complete.
4
Lemma 2.4 If R is a Cohen-Macaulay local ring and y1 , y2 , . . . , ym+1 is an ssop for R ,
then y, yk form a regular sequence for 1 ≤ ∀k ≤ m + 1 .
Proof. It is enough to show for k = 1 . Because (y1 , y2 , . . . , ym+1 )R = (y, y1, . . . , ym)R , it
follows that y, y1, . . . , ym is an ssop for R , too. Hence y, y1 is R-regular.
Lemma 2.5 Suppose that a is an ideal of R and xij ∈ a for ∀i, ∀j . We put Q =
(y1 , y2 , . . . , ym+1 )R . Then δ of 2.3 is an element of am :R Q .
Proof. We get this assertion since aj ∈ am for 1 ≤ ∀j ≤ m + 1 .
3
Associated primes of R/I n
Let R be a Noetherian ring and M = (xij ) ∈ Mat(m, m + 1 ; R) , where 1 ≤ m ∈
Z . Let I = Im (M) . Throughout this section, we assume that I is a proper ideal and
grade Ik (M) ≥ m − k + 2 for 1 ≤ ∀k ≤ m . Let us keep the notations of Section 2.
Let K• be the Koszul complex of f1 , f2 , . . . , fm , which is a complex of graded free
S-modules. We denote its boundary map by ∂• . Let e1 , e2 , . . . , em be an S-free basis of
K1 consisting of homogeneous elements of degree 1 such that ∂1 (ei ) = fi for 1 ≤ ∀i ≤ m .
Then, for 1 ≤ ∀r ≤ m ,
{ ei1 ∧ ei2 ∧ · · · ∧ eir | 1 ≤ i1 < i2 < · · · < ir ≤ m }
is an S-free basis of Kr consisting of homogeneous elements of degree r, and we have
∂r (ei1 ∧ ei2 ∧ · · · ∧ eir ) =
r
X
p=1
(−1)p−1 · fip · ei1 ∧ · · · ∧ ec
ip ∧ · · · ∧ eir .
Let 1 ≤ n ∈ Z . Taking the homogeneous part of degree n of K• , we get a complex
∂
∂
m
1
[K• ]n : 0 −→ [Km ]n −→
[Km−1 ]n −→ · · · −→ [K1 ]n −→
[K0 ]n −→ 0
of finitely generated free R-modules. It is obvious that [Kr ]n = 0 if n < r . On the other
hand, if n ≥ r , then
0 ≤ α1 , α2 , . . . , αm+1 ∈ Z ,
αm+1
T1α1 T2α2 · · · Tm+1
· ei1 ∧ ei2 ∧ · · · ∧ eir
α1 + α2 + · · · + αm+1 = n − r ,
1 ≤ i1 < i2 < · · · < ir ≤ m
is an R-free basis of [Kr ]n .
Proposition 3.1 If (R, m) is a local ring and M ∈ Mat(m, m + 1 ; m) , we have
(
n + 1 if n < m ,
proj. dimR R/I n =
m + 1 if n ≥ m .
5
Proof. By 2.2 and [2, 1.6.17], we see that
∂
∂
π
m
1
0 −→ Km −→
Km−1 −→ · · · −→ K1 −→
K0 −→ R(I) −→ 0
is a graded S-free resolution of R(I) . Hence, for 0 ≤ ∀n ∈ Z ,
∂
∂
π
m
1
0 −→ [Km ]n −→
[Km−1 ]n −→ · · · −→ [K1 ]n −→
[K0 ]n −→ I n tn −→ 0
is an R-free resolution of the R-module I n tn . Let us notice I n tn ∼
= I n as R-modules.
Suppose 1 ≤ r ≤ m and n ≥ r . Then, for any non-negative integers α1 , α2 , . . . , αm+1
with α1 + α2 + · · · + αm+1 = n − r and positive integers i1 , i2 , . . . , ir with 1 ≤ i1 < i2 <
· · · < ir ≤ m , we have
α
m+1
∂r (T1α1 T2α2 · · · Tm+1
· ei1 ∧ ei2 ∧ · · · ∧ eir )
m+1
r
X
X
αm+1
α1 α2
p−1
xip ,j Tj ) · ei1 ∧ · · · ∧ ec
= T1 T2 · · · Tm+1 ·
(−1)
·(
ip ∧ · · · ∧ eir
p=1
=
r m+1
X
X
j=1
1+αj
(−1)p−1 xip ,j · T1α1 · · · Tj
p=1 j=1
∈ m · [Kr−1 ]n .
α
m+1
· · · Tm+1
· ei1 ∧ · · · ∧ ec
ip ∧ · · · ∧ eir
Hence [K• ]n gives a minimal R-free resolution of I n . If n < m , we have [Kn ]n 6= 0 and
[Kr ]n = 0 for any r > n , and so proj. dimR I n = n . On the other hand, if n ≥ m , we
have [Km ]n 6= 0 and [Kr ]n = 0 for any r > m , and so proj. dimR I n = m . Thus we get
the required equality as proj. domR R/I n = proj. dimR I n + 1 .
By Auslander-Buchsbaum formula (cf. [2, 1.3.3]), we get the following.
Corollary 3.2 If (R, m) is local and M ∈ Mat(m, m + 1 ; m), we have
(
depth R − n − 1 if n < m ,
depth R/I n =
depth R − m − 1 if n ≥ m .
Here we remark that depth R ≥ grade I1 (M) ≥ m + 1 by our assumption of this section.
As a consequence of 3.2, we see that the next assertion holds.
Corollary 3.3 Suppose that (R, m) is a local ring and M ∈ Mat(m, m + 1 ; m) . Then we
have m ∈ AssR R/I n if and only if n ≥ m and depth R = m + 1 .
The next result is a generalization of 3.3 .
Theorem 3.4 Let I ⊆ p ∈ Spec R and 1 ≤ n ∈ Z . We put ℓ = max{ 0 ≤ k < m |
Ik (M) 6⊆ p } . Then the following conditions are equivalent.
(1) p ∈ AssR R/I n .
(2) n ≥ m − ℓ and depth Rp = m − ℓ + 1 .
6
When this is the case, grade Iℓ+1 (M) = m − ℓ + 1 .
Proof. By 2.1, there exists N ∈ Mat(m − ℓ, m − ℓ + 1 ; pRp ) such that Ik (N) = Ik+ℓ (M)p
for any k ∈ Z . Hence, for 1 ≤ ∀k ≤ m − ℓ , we have
grade Ik (N) = grade Ik+ℓ (M)p ≥ m − (k + ℓ) + 2 = (m − ℓ) − k + 2 .
Therefore, by 3.3, we see that pRp ∈ AssRp Rp /Im−ℓ (N)n if and only if n ≥ m − ℓ and
depth Rp = m − ℓ + 1 . Let us notice Im−ℓ (N) = Ip . Because p ∈ AssR R/I n if and
only if pRp ∈ AssRp Rp /Ip n , we see (1) ⇔ (2) . Furthermore, as Iℓ+1 (M) ⊆ p , we have
grade Iℓ+1 (M) ≤ depth Rp , and so we get grade Iℓ+1 (M) = m − ℓ + 1 if the condition (2)
is satisfied.
For 1 ≤ ∀n ∈ Z , let ΛnM be the set of integers i such that max{1, m − n + 1} ≤ i ≤ m
and grade Ii (M) = m − i + 2 . Then the following assertion holds.
Theorem 3.5 Let R be a Cohen-Macaulay ring. Then, for 1 ≤ ∀n ∈ Z , we have
[
AsshR R/Ii (M) .
AssR R/I n =
i∈Λn
M
Proof. Let us take any p ∈ AssR R/I n and put ℓ = max{0 ≤ k < m | Ik (M) 6⊆ p} .
Then Iℓ+1 (M) ⊆ p . Moreover, by 3.4 we have n ≥ m − ℓ , depth Rp = m − ℓ + 1 and
grade Iℓ+1 (M) = m − ℓ + 1 . Hence ℓ + 1 ∈ ΛnM . Let us notice that ht p = depth Rp and
ht Iℓ+1 (M) = grade Iℓ+1 (M) as R is Cohen-Macaulay. Therefore ht p = ht Iℓ+1 (M) , which
means p ∈ AsshR R/Iℓ+1 (M) .
Conversely, let us take any i ∈ ΛnM and q ∈ AsshR R/Ii (M) . Then ht q = ht Ii (M) =
grade Ii (M) = m−i+2 . As our assumption implies ht Ii−1 (M) ≥ m−i+3 , it follows that
i−1 = max{0 ≤ k < m | Ik (M) 6⊆ q} . Let us notice n ≥ m−(i−1) as m−n+1 ≤ i , which
is one of the conditions for i ∈ ΛnM . Moreover, we have depth Rq = ht q = m − (i − 1) + 1 .
Thus we get q ∈ AssR R/I n by 3.4, and the proof is complete.
Example 3.6 Let 1 ≤ m ∈ Z and let R be an (m + 1)-dimensional Cohen-Macaulay
local ring with the maximal ideal m . We take an sop x1 , x2 , . . . , xm+1 for R and a family
{αij }1≤i≤m,1≤j≤m+1 of positive integers. For 1 ≤ ∀i ≤ m and 1 ≤ ∀j ≤ m + 1 , we set
αij
xi+j−1
if i + j ≤ m + 2 ,
xij =
x αij
i+j−m−2 if i + j > m + 2 .
Let us consider the following matrix;
α11
α1,m+1
α1m
xm+1
x1
x2α12 x3α13 · · · xm
α
α2m
x2α21 x3α22 x4α23 · · · xm+1
x1 2,m+1
α
α
31
x4α32 x5α33 · · · x1α3m x2 3,m+1
M = (xij ) =
.
x3
......................................
αm,m+1
αm2
αmm
αm1
xm−1
x1αm3 · · · xm−2
xm
xm+1
If αij = 1 for ∀i and ∀j , then M is the matrix stated in Introduction. We put I = Im (M) .
Then the following assertions hold.
7
(1) ht Ik (M) ≥ m − k + 2 for 1 ≤ ∀k ≤ m .
n + 1 if n < m ,
(2) proj. dimR R/I n =
m + 1 if n ≥ m .
m − n if n < m ,
n
(3) depth R/I =
0
if n ≥ m .
Furthermore, if αij = 1 for ∀i and ∀j , the following assertions hold.
(4) If m ≥ 2 , then ht I2 (M) = m and AsshR R/I2 (M) ⊆ AssR R/I n for any n ≥ m − 1 .
(5) If m is an odd integer with m ≥ 3 , then ht I3 (M) = m − 1 and AsshR R/I3 (M) ⊆
AssR R/I n for any n ≥ m − 2 .
(6) If m ≥ 3 , then (I n )sat ( I (n) for any n ≥ m − 1 .
Proof. (1) We aim to prove the following.
Claim Jk−1 + Ik (M) is m-primary for 1 ≤ ∀k ≤ m + 1 , where Jk−1 = (x1 , x2 , . . . , xk−1 )R .
If this is true, we have dim R/Ik (M) ≤ k−1 , and so ht Ik (M) ≥ dim R−(k−1) = m−k+2 ,
which is the required inequality.
In order to prove Claim, we take any p ∈ Spec R containing Jk−1 + Ik (M) . It is
enough to show Jm+1 = (x1 , x2 , . . . , xm+1 )R ⊆ p . For that purpose, we prove Jℓ ⊆ p for
k − 1 ≤ ∀ℓ ≤ m + 1 by induction on ℓ . As we obviously have Jk−1 ⊆ p , let us assume
k ≤ ℓ ≤ m + 1 and Jℓ−1 ⊆ p . Because the k-minor of M with respect to the first k rows
and the columns ℓ − k + 1, . . . , ℓ − 1, ℓ is congruent with
α
xℓ 1,ℓ
α
xℓ 2,ℓ−1
det
.
.
.
0
∗
α
xℓ k,ℓ−k+1
mod Jℓ−1 , it follows that Jℓ−1 + Ik (M) includes some power of xℓ . Hence xℓ ∈ p , and so
we get Jℓ ⊆ p .
(2) and (3) follow from 3.1 and 3.2, respectively.
In the rest of this proof, we assume αij = 1 for any ∀i and ∀j .
(4) Let q = (x1 −x2 , x2 −x3 , . . . , xm −xm+1 )R . Then x1 ≡ xi mod q for 1 ≤ ∀i ≤ m+1 .
Hence, any 2-minor of M is congruent with
x1 x1
det
=0
x1 x1
mod q . This means I2 (M) ⊆ q , and so ht I2 (M) ≤ µR (q) = m . On the other hand,
ht I2 (M) ≥ m by (1). Thus we get ht I2 (M) = m . Then, for any n ≥ m − 1 , we have
2 ∈ ΛnM , and so AsshR R/I2 (M) ⊆ AssR R/I n by 3.5.
8
(5) Let p be the ideal of R generated by {xi − xi+2 } , where i runs all odd integers
with 2 ≤ i ≤ m − 2 . Similarly, we set q to be the ideal of R generated by {xj − xj+2 } ,
where j runs all even integers with 2 ≤ j ≤ m − 1 . Let M ′ be the submatrix of M with
the rows i1 , i2 , i3 and the columns j1 , j2 , j3 , where 1 ≤ i1 < i2 < i3 ≤ m and 1 ≤ j1 <
j2 < j3 ≤ m + 1 . We can choose p, q with 1 ≤ p < q ≤ 3 so that ip ≡ iq mod 2 . Then, for
1 ≤ ∀r ≤ 3 , we have ip + jr ≡ iq + jr mod 2 , and so, if ip + jr is odd (resp. even), it follows
that xip ,jr ≡ xiq ,jr mod p (resp. q). Hence, we see that the p-th row of M ′ is congruent with
the q-th row of M ′ mod p + q , which means det M ′ ≡ 0 mod p + q . As a consequence, we
get I3 (M) ⊆ p + q . Therefore ht I3 (M) ≤ µR (p) + µR (q) = (m−1)/2 + (m−1)/2 = m−1 .
(6) Let us take any p ∈ AsshR R/I2 (M) and n ≥ m − 1 . Then, by (4) we have
ht p = m ≥ 3 and p ∈ AssR R/I n . Hence AssR R/I n is not a subset of { m } ∪ MinR R/I .
Therefore, by the observation stated in Introduction, we get (I n )sat ( I (n) and the proof
is complete.
4
Computing (I m)sat
In this section, we assume that R is an (m + 1)-dimensional Cohen-Macaulay local ring
with the maximal ideal m and x1 , x2 , . . . , xm+1 is an sop for R , where 2 ≤ m ∈ Z . Let M
be the matrix stated in 3.6. We put I = Im (M) . Then, by (3) of 3.6, we have (I n )sat = I n
for 1 ≤ ∀n < m . The purpose of this section is to study (I m )sat .
For 1 ≤ ∀j ≤ m + 1 , we set aj = (−1)j−1 · det Mj , where Mj is the submatrix of M
determined by removing the j-th column. Then I = (a1 , a2 , . . . , am+1 )R . Furthermore,
for 1 ≤ ∀k ≤ m + 1 , we denote by βk the minimum of the exponents of xk that appear in
the entries of M . Let us notice that M’s entries which are powers of xk appear as follows:
α
xk 1,k
α
xk 2,k−1
.
.
.
α
x k,1
k
αk+1,m+1
xk
αk+2,m
xk
.
.
.
α
xk m,k+2
if 1 ≤ k < m , and
..
α
xmm,1
.
α
α
xm2,m−1
xm1,m
or
α
α
..
.
1,m+1
xm+1
2,m
xm+1
α
m,2
xm+1
if k = m or m + 1 , respectively. So, we set
min {αi,k−i+1}1≤i≤k ∪ {αi,k−i+m+2}k<i≤m if 1 ≤ k < m ,
βk =
min {α
i,k−i+1 }1≤i≤m if k = m or m + 1 .
9
Then, for 1 ≤ ∀k ≤ m + 1 , we can choose ik with 1 ≤ ik ≤ m so that one of the following
conditions is satisfied:
(i) 1 ≤ ik ≤ k and βk = αik ,k−ik +1
or
(ii) k < ik ≤ m and βk = αik ,k−ik +m+2 .
Now, for 1 ≤ ∀i ≤ m and 1 ≤ ∀k ≤ m + 1 , we set
α
−β
xk i,k−i+1 k
if i ≤ k ,
′
xik =
αi,k−i+m+2 −βk
if i > k .
xk
Then x′ik ,k = 1 for 1 ≤ ∀k ≤ m + 1 . The next assertion can be verified easily.
Lemma 4.1 Suppose 1 ≤ i ≤ m and 1 ≤ j ≤ m + 1 .
(1) If i + j ≤ m + 2 , setting k = i + j − 1 , we have 1 ≤ k ≤ m + 1 , i ≤ k and
xij = xβk k · x′ik .
(2) If i + j > m + 2 , setting k = i + j − m − 2 , we have 1 ≤ k < m , i > k and
xij = xβk k · x′ik .
β
m+1
Let Q be the ideal of R generated by xβ1 1 , xβ2 2 , . . . , xm+1
. Then M ∈ Mat(m, m+ 1 ; Q)
by 4.1. The first main result of this section is the following:
Theorem 4.2 (I m )sat = I m :R Q and (I m )sat /I m ∼
= R/Q .
Proof. Let S be the polynomial ring over R with variables T1 , T2 , . . . , Tm+1 . We regard
S as a graded ring by setting deg Tj = 1 for 1 ≤ ∀j ≤ m + 1 . Let
fi =
m+1
X
xij Tj ∈ S1
j=1
for 1 ≤ ∀i ≤ m and let K• be the Koszul complex of f1 , f2 , . . . , fm . Then K• is a graded
complex. Let ∂• be the boundary map of K• and let e1 , e2 , . . . , em be an S-free basis of
K1 consisting of homogeneous elements of degree 1 such that ∂1 (ei ) = fi for 1 ≤ ∀i ≤ m .
As is stated in the proof of 3.1,
∂
∂
ǫ
m
0 → [Km ]m →
[Km−1 ]m → · · · → [K1 ]m →1 [K0 ]m → R → 0
k
Sm
is an acyclic complex, where ǫ is the R-linear map such that
α
α
m+1
m+1
) = a1α1 aα2 2 · · · am+1
ǫ(T1α1 T2α2 · · · Tm+1
for any 0 ≤ α1 , α2 , . . . , αm+1 ∈ Z with α1 + α2 + · · · + αm+1 = m . We obviously have
Im ǫ = I m . We set e = e1 ∧ e2 ∧ · · · ∧ em and ěi = e1 ∧ · · · ∧ ebi ∧ · · · ∧ em for 1 ≤ ∀i ≤ m + 1 .
10
Let us take { e } and {Tj ěi | 1 ≤ i ≤ m , 1 ≤ j ≤ m + 1} as R-free basis of [Km ]m and
[Km−1 ]m , respectively. Because
(♯)
∂m (e) =
m
X
(−1)i−1 fi · ěi =
i=1
m m+1
X
X
(−1)i−1 xij · Tj ěi ,
i=1 j=1
we have ∂m ([Km ]m ) ⊆ Q · [Km−1 ]m . Hence, by [5, 3.1] we get
(I m :R Q)/I m ∼
= [Km ]m /Q[Km ]m ∼
= R/Q .
Here, for 1 ≤ ∀i ≤ m and 1 ≤ ∀k ≤ m + 1 , we set
Tk−i+1
if i ≤ k ,
Tik =
T
k−i+m+2 if i > k .
Then the following assertion holds:
Claim 1 Suppose 1 ≤ k, ℓ ≤ m + 1 and k 6= ℓ , then Tik 6= Tiℓ .
In order to prove the claim above, we may assume k < ℓ . Then the following three cases
can happen: (i) i ≤ k < ℓ , (ii) k < i ≤ ℓ or (iii) k < ℓ < i . Because k−i+1 < ℓ−i+1
and k − i + m + 2 < ℓ − i + m + 2 , we get Tik 6= Tiℓ in the cases of (i) and (ii). Furthermore,
as m + 1 > ℓ − k , we get k − i + m + 2 > ℓ − i + 1 , and so Tik 6= Tiℓ holds also in the case
of (ii). Thus we have seen Claim 1.
Now, for 1 ≤ ∀k ≤ m + 1 , we set
v(k,e) =
m
X
(−1)i−1 x′ik · Tik ěi .
i=1
Then the following equality holds:
Claim 2 ∂m (e) =
m+1
X
xβk k · v(k,e) .
k=1
In fact, by (♯) and 4.1 we have
m m−i+2
X
X
∂m (e) =
(
(−1)i−1 xij · Tj ěi +
i=1
=
=
j=1
m m+1
X
X
(
(−1)i−1 xβk k x′ik
m+1
X
j=m−i+3
· Tk−i+1ěi +
i=1 k=1
m m+1
X
X
(−1)i−1 xij · Tj ěi )
i−1
X
k=1
(−1)i−1 xβk k x′ik · Tik ěi ,
i=1 k=1
and so the equality of Claim 2 follows.
11
(−1)i−1 xβk k x′ik · Tk−i+m+2 ěi )
Finally, we need the following:
Claim 3 {v(k,e) }1≤k≤m is a part of an R-free basis of [Km−1 ]m .
If this is true, by [5, 3.4] (See [4, 3.4] for the case where m = 2) we get depth R/(I m :R
Q) > 0 , which means (I m )sat = I m :R Q . So, let us prove Claim 3. By Claim 1, we see
that Ti1 ,1 ěi1 , Ti2 ,2 ěi2 , . . . , Tim ,m ěim are different to each other. We set
U = { Tj ěi | 1 ≤ i ≤ m , 1 ≤ j ≤ m + 1} \ {Tik ,k ěik | 1 ≤ k ≤ m}
and aim to prove that U ∪ {v(k,e) }1≤k≤m is an R-free basis of [Km−1 ]m . By [5, 3.3], it is
enough to show that the submodule of [Km−1 ]m generated by U ∪ {v(k,e) }1≤k≤m includes
Tik ,k ěik for 1 ≤ ∀k ≤ m . This can be easily seen since
X
(−1)i−1 x′ik · Tik ěi
v(k,e) = (−1)k−1 · Tik ,k ěik +
i6=ik
and Tik ěi ∈ U if i 6= ik , which follows from Claim 1. Thus the assertion of Claim 3 follows,
and the proof of 4.2 is complete.
If we assume a suitable condition on {αij } , we can describe a generator of (I m )sat /I m .
For 1 ≤ ∀i ≤ m and 1 ≤ ∀k ≤ m + 1 , we set
( ′
if i ≤ k ,
xik ak−i+1
aik =
x′ik ak−i+m+2 if i > k ,
and A = (aik ) ∈ Mat(m, m + 1 ; I) . Then the next equality holds:
Lemma 4.3
A
xβ1 1
xβ2 2
..
.
β
m+1
xm+1
=
0
0
..
.
0
.
Proof. For 1 ≤ ∀i ≤ m , we have
m+1
X
xij aj = 0 .
j=1
Let us divide the left side of this equality as follows:
m−i+2
X
j=1
xij aj +
m+1
X
xij aj = 0 .
j=m−i+3
If 1 ≤ j ≤ m − i + 2 , setting k = i + j − 1 , we have i ≤ k ≤ m + 1 and
xij aj = xβk k x′ik · ak−i+1 = xβk k · aik .
12
On the other hand, if m − i + 3 ≤ j ≤ m + 1 , setting k = i + j − m − 2 , we have 1 ≤ k < i
and
xij aj = xβk k x′ik · ak−i+m+2 = xβk k · aik .
Thus we get
m+1
X
aik · xβk k = 0
k=1
for 1 ≤ ∀i ≤ m , which means the required equality.
For 1 ≤ ∀k ≤ m + 1 , we denote by Ak the submatrix of A determined by removing
the k-th column. We set bk = det Ak .
Example 4.4 Suppose βk = αk,1 for 1 ≤ ∀k ≤ m (For example, this holds if αk,1 = 1 for
1 ≤ ∀k ≤ m). Then, there exists δ ∈ R such that xβk k · δ = bk for 1 ≤ ∀k ≤ m + 1 and
(I m )sat = I m + (δ) .
Proof. The existence of δ such that xβk k · δ = bk for 1 ≤ ∀k ≤ m + 1 follows from 2.3 and
4.3. Then δ ∈ I m :R Q ⊆ (I m )sat . We put Q′ = (x1β1 , xβ2 2 , . . . , xβmm )R . Then
α1,m+1
xm+1
α2,m
xm+1
M1 ≡
mod Q′ ,
.
.
.
αm,2
xm+1
0
0
and so a1 ≡ ±xαm+1 mod Q′ , where α := α1,m+1 + α2,m + · · · + αm,2 . Furthermore,
if 2 ≤ k ≤ m + 1 , we have ak ∈ Q′ since the entries of the first column of Mk are
xβ1 1 , x2β2 , . . . , xβmm . Hence Q′ + I = Q′ + (xαm+1 ) . On the other hand, the assumption of 4.4
implies that, for 1 ≤ ∀k ≤ m, we can take k itself as ik , and then x′kk = 1 . Hence
x′13 a3 · · · x′1m am x′1,m+1 am+1
a1
x′12 a2
x′21 am+1
a1
x′23 a2 · · · x′2m am−1 x′2,m+1 am
′
′
′
′
,
x
a
x
a
a
·
·
·
x
a
x
a
A=
m
m+1
1
m−2
m−1
31
32
3m
3,m+1
.....................................................
x′m1 a3
x′m2 a4 x′m3 a5 · · ·
a1
x′m,m+1 a2
and so
Am+1
≡
±xαm+1
±xαm+1
0
..
0
.
±xαm+1
′
which means bm+1 ≡ ±xmα
m+1 mod Q . Thus we get
mod Q′ ,
β
m+1
′
xm+1
· δ ≡ xmα
m+1 mod Q .
β
m+1
Here we notice βm+1 ≤ α1,m+1 < α . Because xβ1 1 , xβ2 2 , . . . , xm+1
is an R-regular sequence,
it follows that
mα−β
δ ≡ ±xm+1 m+1 mod Q′ ,
13
and so
mα−βm+1
Q′ + (δ) = Q′ + (xm+1
′
m
) ⊇ Q′ + (xmα
m+1 ) = Q + I .
Now we consider the R-linear map
mα−β
f : R −→
Q′ + (xm+1 m+1 )
Q′ + (δ)
=
Q′ + (xmα
Q′ + I m
m+1 )
mα−βm+1
such that f (1) is the class of xm+1
. Then we have the following:
Claim Ker f = Q .
If this is true, then R/Q ∼
= (Q′ + (δ))/(Q′ + I m ) , and so
ℓR ( R/Q ) = ℓR (
Q′ + (δ)
).
Q′ + I m
Because (Q′ + (δ))/(Q′ + I m ) is a homomorphic image of (I m + (δ))/I m and I m + (δ) ⊆
(I m )sat , we have
ℓR (
Q′ + (δ)
I m + (δ)
)
≤
ℓ
(
) ≤ ℓR ( (I m )sat /I m ) = ℓR ( R/Q ) ,
R
Q′ + I m
Im
where the last equality follows from 4.2 . Thus we see
ℓR (
I m + (δ)
) = ℓR ( (I m )sat /I m ) ,
Im
and so I m + (δ) = (I m )sat holds.
Proof of Claim. Let us take any r ∈ Ker f . Then, there exists s ∈ R such that
mα−βm+1
r · xm+1
This congruence implies
mα−βm+1
xm+1
′
≡ s · xmα
m+1 mod Q .
β
m+1
(r − s · xm+1
) ∈ Q′ .
mα−β
β
m+1
Because xβ1 1 , . . . , xβmm , xm+1 m+1 is an R-regular sequence, we have r −s·xm+1
∈ Q′ , which
means r ∈ Q . Hence Ker f ⊆ Q . As the converse inclusion is obvious, we get the equality
of the claim, and the proof of 4.4 is complete.
References
[1] L. Avramov, Complete intersections and symmetric algebras, J. Algebra, 73 (1981),
248–263.
[2] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Stud. Adv. Math. 39,
Cambridge University Press, 1997.
14
[3] K. Fukumuro, On the symmetric and Rees algebras of certain determinantal ideals,
Preprint (2013), arXiv:1306.0993.
[4] K. Fukumuro, T. Inagawa and K. Nishida, On a transform of an acyclic complex of
length 3, J. Algebra, 384 (2013), 84–109.
[5] T. Inagawa, ∗-transforms of acyclic complexes, Preprint (2013).
15
| 0math.AC
|
GenomeFingerprinter and universal genome fingerprint
analysis for systematic comparative genomics
Yuncan Ai*, Hannan Ai, Fanmei Meng, Lei Zhao
State Key Laboratory for Biocontrol, School of Life Sciences, Sun Yat-sen University,
Guangzhou 510275, P. R. China
[email protected]
* Corresponding Author:
Yuncan Ai
Ph.D., Professor, Principal Investigator
[email protected]
State Key Laboratory for Biocontrol
School of Life Sciences
Sun Yat-sen University
Guangzhou 510275
P. R. China
1
Abstract
Background:
How to compare whole genome sequences at large scale has not been achieved via
conventional methods based on pair-wisely base-to-base comparison; nevertheless,
no attention was paid to handle in-one-sitting a number of genomes crossing genetic
category (chromosome, plasmid, and phage) with farther divergences (much less or
no homologous) over large size ranges (from Kbp to Mbp). It should be a priority to
persue comparative genomics at large scale based on geometrical analysis of
sequence in the post-genomic era. However, even how to simply visualize a DNA
sequence has been challenging for decades; little progress has been made to date.
Results:
We created a new method, GenomeFingerprinter, to unambiguously produce
three-dimensional coordinates from a sequence, followed by one three-dimensional
plot and six two-dimensional trajectory projections to illustrate whole genome
fingerprints. We further developed a set of concepts and tools and thereby established
a new method, universal genome fingerprint analysis. We demonstrated their
applications through case studies on over a hundred of genome sequences.
Particularly, we defined the total genetic component configuration (TGCC) (i.e.,
chromosome, plasmid, and phage) for describing a strain as a system, and the
universal genome fingerprint map (UGFM) of TGCC for differentiating a strain as a
universal system, as well as the systematic comparative genomics (SCG) for
comparing in-one-sitting a number of genomes crossing genetic category in diverse
strains. By using UGFM (I), UGFM-TGCC (II), and UGFM-TGCC-SCG (III), we
compared a number of genome sequences with farther divergences (chromosome,
plasmid, and phage; bacterium, archaeal bacterium, and virus) over large size ranges
2
(6Kbp~5Mbp), giving new insights into critical problematic issues in microbial
genomics in the post-genomic era.
Conclusion:
This paper provided a new method for rapidly computing, geometrically visualizing,
and intuitively comparing genome sequences at fingerprint level, and hence
established a new method of universal genome fingerprint analysis for systematic
comparative genomics.
Keywords
Genome fingerprint, fingerprint analysis, systematic comparative genomics,
computational visualization, geometrical analysis
3
Introduction
By using conventional methods based on pair-wisely base-to-base comparison, how
to compare whole genome sequences at large scale has not been achieved;
nevertheless, no attention was paid to handle in-one-sitting a number of genomes
crossing different genetic category (chromosome, plasmid, and phage) with farther
divergences (less or no homologous among genetic components) over large size
ranges (from Kbp to Mbp per sequence). It should be a priority to persue comparative
genomics at large scale based on geometrical analysis of sequence in the
post-genomic era. However, little progress has been made to date; even how to
simply visualize a DNA sequence has been challenging for decades [1].
Pioneering works in computer reading and geometrical visualizing of DNA
sequence had been tried in one-dimension [2, 3], two-dimensions (Z-curve) [4], and
three-dimensions (H-curve) [5, 6]. However, those were valid only for ‘static’ modeling
and visualizing. The ‘dynamic’ modeling and visualizing in a virtual reality environment
had been studied [7, 8]. A comprehensive example was AND-viewer, which provided a
three-dimensional way to dynamically sense the big picture of a large DNA sequence
in a virtual reality environment by using sensor, instead of mouse or keyboard. This
pioneering effort had made fantastic progress for human to mimic 3D visions to
intuitively sense genome sequences [7, 8], but still there was no possibility of using
the datasets created for visions to further explore real biological contexts.
The post-genomic era promoted demanding of data mining and robust reasoning
with huge amount of genome sequences [1]. Comparative genomics was essential to
retrieve and mine genome sequences, there were numerous conventional methods,
which were divided into two types: algebraic approach [9, 10, 11, 12] and geometrical
approach [13, 14].
4
Algebraic approach means that calculating dissimilarity, or similarity or identity
are based on pair-wisely base-to-base comparison; the outputs of calculation are only
used for visualization through graphic techniques, rather than for robust data mining
and reasoning analysis. Of course, tools for genomic data visualization are still
essential for end-users to explore, interpret, and manipulate data [1]. The most
common tools were BLAST [9] and CLUSTALW [10], which were only for pair-wisely
comparisons among a certain number of short fragments at gene level. Recently, a
BLAST-based visualization tool, BRIG, was constructed for genome-wide comparison
to create images of multiple circular genomes among a number of closely related
bacteria strains [11]. The output image showed BLAST-similarity between a central
reference sequence and other sequences in question as a set of concentric rings,
where BLAST-matches were colored on a sliding scale indicating a defined
percentage of BLAST-identity. It had great advantage over other common tools, like
ACT [12], in terms of the numbers of genomes being handled simultaneously and the
ways of comparing and presenting of images in-one-sitting. These features made it a
versatile approach for visualizing a range of genomic data, but it still was only for
visions. Mauve [14, 15] was widely used for comparing and visualizing a number of
genomes of close relatives in linear forms, which combined algebraic calculation and
graphic display. However, even with close relatives only, the number of genomes
being calculated and displayed dramatically depended on computation constraints
causing too much CPU time (at least O(n2) in time complexity) or memory overflow,
which limited to a fewer genomes of close relatives to be compared at once time.
Geometrical approach means that both calculation and visualization are dynamic
for geometrical analysis with the input and output re-useable. One promising example
was Zplotter (in Z-curve method) calculating three-dimensional coordinates from a
linear genome sequence. Those coordinates were used to create a three-dimensional
5
geometrical vision in a rough manner (as open Z-curve) for a given DNA sequence
[16]. Hundreds of such visions for microbial genomes were collected as a database
[17]. The Z-curve method (based on Zplotter) was not only used for simple
visualization [16, 17], but also for geometrical analysis to further explore real contexts
of biology [18, 19, 20, 21]. For example, two replication ori points in archaeal genomes
were predicted by Z-curve method [22, 23] and confirmed later by wet experiments in
other labs [24, 25], showing its promising. Z-curve method was widely used by
researchers around the world and had promoted the progress in understanding of
genomics, starting a new frontier in geometrical analysis of genome sequences.
However, Zplotter algorithm had an inevitable limitation in mispresentation of a
genome sequence but with different cutting-points (explanations in the main text),
which would not be suitable for creating unique genome fingerprints.
In this paper, we present a new method, called GenomeFingerprinter, to
unambiguously produce three-dimensional coordinates from sequence, followed by
one three-dimensional plot and six two-dimensional trajectory projections to illustrate
whole genome fingerprints. We further develop a set of concepts and tools and
thereby establish a new method called universal genome fingerprint analysis. We
demonstrate their applications through case studies on over a hundred of genome
sequences, giving new insights into critical problematic issues in microbial
comparative genomics. We anticipate that our methods could be widely applicable to
systematic comparative genomics in the post-genomic era [1].
Results
Mathematical model and three-dimensional coordinates
6
For geometrical visualization of a given genome sequence, the key step is how to get
its three dimensional coordinates (xn, yn, zn). To do this, the Z-curve [16] defined a
coordinate (xn, yn, zn) for each base in a linear genome sequence (n=1, 2, …, N; N is
the sequence length) by the equation (0). It defined a unique Z-curve from a given
linear sequence, and vice versa. Note, it was designed for a linear genome sequence
and An, Tn, Gn, Cn were the sum of numbers for each of four base-type (A, T, G, C),
respectively, counting from the first base to other bases before and including the nth
base in a linear sequence (n=1, 2, …, N). The calculations could be performed by
using Zplotter program [16]. The main problem here was the ambiguousness of the
“first base” due to cutting-point errors in deposited genome sequences (see
explanations later).
x n (A n G n ) (Cn Tn )
y n (A n Cn ) (G n Tn )
z (A T ) (C G )
n
n
n
n
n
(0)
Here, we take the same definition as equation (0), but with different contents of An,
Tn, Gn, Cn. To do that, we thus propose a new mathematic model, called
GenomeFingerprinter, for geometrical visualization of a circular genome sequence. A
circular sequence contains 40-bps (5’-3’):
ACACTGACGCACACTGACGCACACTGACGCACACTGACGC (Figure 1) as an
artificial example will be used to illustrate the conceptual principals of our method.
Firstly, we randomly select a base (nth) as the first target base (TB). For the given
TB (nth), we define its relative distances (RD) (1) to the other moving mth base (as
focusing base, FB) (m=1, 2, …, N).
7
1,
2,
RD mn ...
N 1,
N,
(m n 1)
(m n 2)
...
(m n n-1)
(m n n)
(1)
Here, the RD concept is critical. Once we have selected the given TB (suppose at
position 1, base A) and the other moving FB (suppose at position 20, base C), the RD
value is 19 (Figure 1). Thus, a collection of RD values (m=1, 2, …, N) will be
generated for the given TB. Particularly, the RD formula (1) can virtually treat an
arbitrary linear sequence as a circular one. For example, when the moving mth FB is
located at the position of n+n, the RD is N, which means the RD value now is N, not
zero, when the moving mth FB going over one circle (i.e., starting from the position at
the nth base and finishing at the same position at the nth base).
Secondly, we define the weighted relative distance (WRD) (2) (N is the sequence
length). The example above will have value 19/40. This is simply to reduce memory
burden and thus release computation constraints for larger sequences.
WRDm
n
RDm
n
N
(2)
Thirdly, for the same chosen TB (nth), we define the sum of the weighted relative
distances (SWRD) (3) from the above collection of WRD (m=1, 2, …, N) for each of
the four base-type (A, G, T, C), respectively.
SWRD An
A
n
G
[WRD nm ]
SWRD Gn n [WRD mn ]
SWRD Tn
n [WRD ]
SWRD Cn
T
C
n
(3)
m
n
[WRD nm ]
Fourthly, we define the coordinate (xn, yn, zn) (4) for the chosen TB (nth). Note,
here we count the sum of the weighted relative distances (SWRD) (unlike Z-curve
method counting the sum of numbers) for each of four base-type (A, T, G, C),
8
respectively. So far, that is only one cycle done for only one chosen TB (nth).
x n (SWRDnA SWRDGn ) (SWRDCn SWRDTn )
A
G
G
T
y n (SWRDn SWRDn ) (SWRDn SWRDn )
z (SWRDA SWRDT ) (SWRDG SWRDG )
n
n
n
n
n
(4)
Finally, we will repeat the above cycle, i.e., selecting the next TB (e.g., n=2 here)
and repeating the process. We will have total N cycles (n=1, 2, …, N); each cycle has
only one chosen TB and creates only one coordinate (xn, yn, zn) for that chosen TB. All
N bases will have their coordinates (xn, yn, zn) after having finished all of these N
cycles. We have developed in-house script, GenomeFingerprinter.exe, to do all.
As an example, the artificial 40-bps genome sequence (Figure 1) had its
coordinates (xn, yn, zn) (Table 1), which were calculated by using our program
GenomeFingerprinter.exe, giving each base with a coordinate (xn, yn, zn) as a point in
the three-dimensional space, in total 40 points for the whole sequence.
Three-dimensional plot (3D-P) and primary genome fingerprint map (P-GFM)
The three-dimensional coordinates (xn, yn, zn) can be plotted out as a
three-dimensional plot (3D-P) to give a geometrical vision. The artificial 40-bps
sequence had only 40 points (Table 1) hence giving a naive vision. As real examples,
we showed visions for fragmental sequences ranging from tens to hundreds of
kilobases (Table 2) of Escherichia coli strains (Figure 2). Clearly, each vision had its
unique genome fingerprint (GF) both globally and locally. We defined such a GF vision
as genome fingerprint map (GFM). The GFM was an intuitive identity for an individual
genome sequence, and vice versa. Therefore, from now on, we can directly operate
and compare GFM for studying sequence. That is, we compare genome sequences
through genome fingerprints (via geometrical analysis) instead of sequence
base-pairs (via algebraic analysis). For convenience, we further defined the
9
three-dimensional plot vision as the primary genome fingerprint map (P-GFM).
Two-dimensional trajectory projections (2D-TPs) and secondary genome
fingerprint maps (S-GFMs)
To demonstrate sophisticated genome fingerprints, we created six two-dimensional
trajectory projections (2D-TPs) for a given P-GFM by combining different components
from its coordinates, including xn ~ n, yn ~ n, zn ~n, xn ~ yn, xn ~ zn, and yn ~ zn. For
convenience, we defined these six 2D-TPs as the secondary genome fingerprint
maps (S-GFMs). For example, six S-GFMs of Escherichia coli K-12/W3110 genome
sequence clearly showed subtle variations both globally and locally (Figure 3). Note
that S-GFMs of xn ~ zn, yn ~ zn, xn ~ yn beared much more sensitive information
compared to those of xn ~ n, yn ~ n, and zn ~ n, respectively. Generally, S-GFMs can
amplify subtle variations that are insensitive or invisible in P-GFMs. Particularly,
S-GFMs of xn ~ yn, xn ~ zn and yn ~ zn are much more sensitive in differentiating local
subtle variations, intuitively identifying unique genome features; whereas S-GFMs of
xn ~ n, yn ~ n and zn ~ n are relatively less informative but useful when focusing on
global patterns (Figure 3).
Universal genome fingerprint map (UGFM)
P-GFM and S-GFMs can be either separately or sequentially used. For convenience,
we defined the universal genome fingerprint map (UGFM) to unify both of them. By
UGFM, we could compare in-one-sitting a number of sequences and display their
GFMs at once time, on which each GFM could be classified into different groups
solely based on its location (Figure 4).
For example, six archaeal genomes and twelve fragmental sequences from E.coli
strains (Table 2) had complex P-GFMs (Figure 4). Within the species Sulfolobus
10
islandicus, strains M.14.25 and M.16.4 shared global similarity in P-GFM (Figure 4, A),
indicating subtle variations at strain level. However, with farther divergences, strain S.
islandicus Y.N.15.51 globally differed from Methanococcus voltae A3 but locally
shared similar regions in P-GFM (Figure 4, B); whereas S. islandicus Y.G.57.14
completely differed from Methanosphaera stadtmanae 3091 (Figure 4, C), confirming
their farther lineages.
On the other hand, within the species Sulfolobus islandicus, two strains M.14.25
and M.16.4 had only subtle variations (Figure 4, A), how could they be precisely
differentiated by P-GFM? We defined geometrical center ( x , y , z ) as a distinctive
indicator for a single P-GFM to compare individual P-GFMs. For example, two strains
M.14.25 and M.16.4 had different geometrical center values (644.00, -2081.00,
388729.14) and (476.50, -1916.50, 387938.64), respectively, and hence were clearly
distinguishable.
Furthermore, those twelve fragmental sequences from E.coli strains (Table 2,
Figure 4, D) were further enlarged and displayed as a UGFM besides their own
individual P-GFMs (Figure 5). Clearly, there were six groups on UGFM (Figure 5, A, B,
C, D, E, F) solely based on the locations of different P-GFMs. Particularly, different
fragmental sequences either from the same strain (e.g., 91.1.1, 91.1.61, 91.6.59) or
from different strains (e.g., 913.5.57, 4431.1.70, 7946.4.7, 10473.1.74, 10498.4.86,
12947.1.50, 13941.2.60) (Table 2) could be revealed as complicated P-GFM patterns.
Some were similar (91.1.61, 913.1.77 and 10473.1.74 (Figure 5, A); 91.6.59, 913.5.57
and 13941.2.60 (Figure 5, B) ) but most were different (Figure 5, C, D, E, F) no matter
what the lineage was, strongly demonstrating the facts that there were modular
domains in these genomes and such mosaic structures probably remained their
tracking of evolutionary history. Most interestingly, a given P-GFM had quite different
11
views between its own P-GFM and that on UGFM simply because of the scale-down
and view-angle rotation effect in UGFM (Figure 5). This feature would ensure UGFM
as a powerful tool for large-scale global comparison in-one-sitting among a large
number of whole genome sequences, theoretically, as many sequences as possible
as long as the computer memory could allow.
Universal genome fingerprint analysis (UGFA)
Now that we had such a powerful tool, UGFM, based on unambiguous genome
fingerprints, to compare a large number of whole genome sequences in-one-sitting
(Figure 5), we further established a new method called universal genome fingerprint
analysis (UGFA) (Figure 6). We anticipated that UGFA would be effective for
systematic comparative genomics at large scale by expanding the scope of genetic
category in question. Briefly, the UGFA method consisted of three subcategories
(Figure 6): UGFM (I) (Figure 7, 8, 9), UGFM-TGCC (II) (Figure 10), and
UGFM-TGCC-SCG (III) (Figure 11) corresponding to three objects: a genome, a strain,
and a set of strains, respectively. For each subcategory, demonstration with examples
of case studies was described in details below.
UGFM (I): Universal genome fingerprint map (UGFM)
Firstly, UGFM (I) was the foundation and the first major component of our UGFA
method. It was proved powerful in global comparison at large scale for prokaryote
bacteria genomes (Figure 5). More examples were from a number of genomes of
archaeal bacterium (Table 2, Figure 7), phage (Table 3, Figure 8), and virus (Table 3,
Figure 9). Five archaeal bacteria strains (Halomonas elongate DSM 2581,
Halorhodospira halophilia SL1, Halorhabdus utahensis DSM 12940, Halothermothrix
orenii H 168 and Halothiobacillus neapolitanus c2) representing five genera of
halophilic Archaea were displayed as a UGFM (Figure 7) showing larger scale-down
12
and view-angle rotation effect. Clearly, each strain had only one chromosome with
size ranging of 2.6 ~4.1 Mbp; and these five archaeal chromosomes had no close
relationships at all (Figure 7) confirming their farther diverse lineages at genus level
(Table 2,). However, forty seven phages of family Microviridae (Table 3) that were
grouped into two major clusters (Figure 8) and twenty four coronavirus strains (Table 3)
that were classified into seven clusters (Figure 9) perfectly matched to their biological
identities among close relatives. Put together, these fingdings from total eight three
genomes (i.e., twelve bacteria, five archaeal bacteria, forty seven phages, and twenty
four viruses) as good examples demonstrated that UGFM (I) could apply to any
genetic category (bacterium, archaeal bacterium, phage, and virus) no matter how
farther (Figure 7) or closer (Figure 5, 8, 9) divergences of genetic components in
comparison.
UGFM-TGCC (II): Universal genome fingerprint map (UGFM) of total genetic
component configuration (TGCC)
Secondly, how to compare a number of genome sequences crossing different genetic
category (e.g., chromosome, plasmid, and phage) in a strain? Accordingly, we defined
the total genetic component configuration (TGCC) as a set of genomes crossing all
genetic category (chromosome, plasmid, and phage, if applicable) in a strain for
describing a strain as a system. We further defined the universal genome fingerprint
map (UGFM) of total genetic component configuration (TGCC) (UGFM-TGCC) for
differentiating a strain in view of a universal system. Therefore, we could use
UGFM-TGCC (II) to compare in-one-sitting among all genetic components in a strain,
regardless the format of category (chromosome, plasmid, and phage).
For example, four strains crossing four genera in haophilic Archaea (Table 2)
including Halogeometricum boringquense DSM 11551 (one chromosome and five
13
plasmids), Haloterrigena turkmenica DSM 5511 (one chromosome and six plasmids),
Natrialba magadii ATCC 43099 (one chromosome and three plasmids), Natrinema
pellirubrum DSM 15624 (one chromosome and two plasmids) were demonstrated by
UGFM-TGCC (II) (Figure 10) clearly indicating their farther lineages at genus level.
Note that the scale-down and view-angle rotation effect revealed the farther
divergences between one chromosome and multiple plasmids in a certain strain,
suggesting that it would be challenging for conventional methods to compare them
due to much less or no homologous. Specifically, in the same figure (Figure 10, H), the
tiny green spot (plasmid NC_008213) and the giant red vision (chromosome
NC_008212) with farther divergences would not be easily compared by any other
conventional methods.
UGFM-TGCC-SCG (III): UGFM-TGCC-based systematic comparative genomics
(SCG)
Thirdly, how to compare a number of genome sequences both crossing genetic
category in a strain (chromosome, plasmid, and phage) and crossing a number of
strains (a cluster of strains) as a system (i.e., in-one-sitting)? To compare a number of
such diverse genomes in-one-sitting, we defined a concept of UGFM-TGCC-SCG (III),
UGFM-TGCC-based systematic comparative genomics (SCG). Note, here we called it
as “systematic comparative genomics (SCG)“ simply because all genomes crossing
different genetic category (chromosome, plasmid, and phage) among diverse strains
should be much less or even no homologous at all, which would be incredibly
challenging to any known conventional methods that principally based on similarity
analysis of homologous. In other words, to our knowledge to date, no one
conventional method could handle such farther diverse genetic components
in-one-sitting; to which even no attention was paid before. In fact, all of the
14
documented researches on comparative genomics to date were automatically based
on the assumption that there was so-called a reference genome sequence for very
close relatives in question; otherwise, they would not bother to do comparison. But, in
our case, we exactly focused on the opposites that had much less or even no
homologous and compared those diverse genetic components crossing farther
divergences regardless the format of genetic category and regardless the extent of
lineage divergence. Therefore, we called our objects in comparison as the “systematic
comparative genomics” in order to distinguish from other traditional routes. This was
one of the core concepts and aims in the present study.
Indeed, our UGFM-TGCC-SCG (III) subcategory method was powerful to handle
those extraordinary situations. For example, total nineteen genomes including six
chromosomes and thirteen plasmids with larger size ranges (6Kbp ~ 4Mbp) could be,
separately, mapped and analyzed by using UGFM-TGCC-SCG (III) (Figure 11). These
nineteen genomes were from four strains crossing four genera of halophilic Archaea
(Table 2) and analyzed in-one-sitting as two sets of comparison (Figure 11):
Halorubrum lacusprofundii ATCC 49239 (two chromosomes and one plasmid) vs.
Haloarcula marismortui ATCC 43049 (two chromosomes and seven plasmids);
Haloferax vocanii DS2 (one chromosome and four plasmids) vs. Halomicrobium
mukohataei DSM 12286 (one chromosome and one plasmid). Obviously, they were
certainly demonstrated as diverse lineages solely based on genome fingerprints. Most
importantly, note that tiny spots (e.g., corresponding to 6Kbp) and giant ones (e.g.,
corresponding to 4Mbp) were harmoniously existed in the same figures (Figure 10, H,
Figure 11, C), either closely or farther away. Such amazing landscapes could be only
revealed by our unique methods under the notions of so-called “total genetic
component configuration” and “systematic comparative genomics”, particularly, as
UGFM-TGCC and UGFM-TGCC-SCG in these cases. These should be more than
15
enough as representatives to prove UGFM-TGCC-SCG (III) effective and powerful.
Case studies: Applications of universal genome fingerprint analysis (UGFA)
Objectives
As for more specific examples, we chose two archaeal halobacteria strains,
Halobacterium sp. NRC-1 and Halobacterium salinarum R1 (Table 2) [26, 27], to
persue systematic comparative genomics by using our UGFA method. It was not only
because they had incredible microbiological features such as genome-wide evolution
events [28, 29] and multiple replication ori points (unlike the common prokaryotes with
only one replication ori point) that could be easily tracking [23], but also because two
genomes were independently sequenced by two labs [28, 29] and had led to
interesting arguments about critical problematic issues in microbial genomics and
taxonomy, such as whether they were the same species or strain [29] or their genome
sequences were correctly assembled particularly considering of mega-plasmids or
minichromosomes [30, 31], and what might be the mechanism for evolutions [29, 31,
32, 33], even what should be considered for refining a “species” in taxonomy [34]. We
expected that the universal genome fingerprint analysis (UGFA) could provide new
insights into these critical problematic issues that would be crucial and invaluable for
modern microbiology in the post-genomic era.
Genome-wide evolution events
From the sophisticated genome fingerprints in S-GFMs (Figure 12), two deposited
chromosomes NRC-1 (NC_002607) and R1 (NC_010364) were very similar but not
identical having subtle differences at strain level (Figure 12, A, B, E, F), supporting the
claim that two strains were virtually from the same ancestor but had undergone
evolutions [31, 32]. The subtle variations (Figure 12, C, D, G) also indicated the
16
genome-wide evolution events (shown by arrow-markers), causing longer of NRC-1
chromosome. It was coincided with the documented facts that the IS-element-rich
regions [27] shuttled between chromosomes and mega-plasmids [30, 31], but the core
genes conserved [33].
Two replication ori points
Again, from S-GFMs, two replication ori points, oriC1 and oriC2, (Figure 12, E, F, H)
were identified and the replication domains in two genomes were demonstrated
identical, and those evolution events were not located in such replication regions.
These evidences also supported the claim that two strains virtually came from the
same ancestor strain [31, 32]. Most interestingly, two replication ori points in strain
NRC-1 were reported as the first representatives of archaeal bacteria, changing the
traditional definition of only one ori point in prokaryotes. In fact, one of two ori points
was predicted by theoretical Z-curve analysis [22, 23] and confirmed by biological
experiments later [24, 25]. Thus the reproducibility in identification of such two
replication ori points in two sequences has proved that our method is as effective and
sensitive as Z-curve analysis [22, 23].
UGFM-TGCC-SCG for differentiating strains
Two strains NRC-1 and R1 were completely different in terms of the numbers of
plasmids and total base-pairs (Table 4). How to concisely describe their differences in
visualization remained challenging. For example, they had eight genetic components
including two chromosomes and six plasmids (Table 4), which made it ambiguous to
only compare any part of them as traditional comparative genomics did. We thus
should compare all of eight genomes in order to differentiate two strains
unambiguously. By using UGFM-TGCC-SCG (Figure 13), we compared eight genome
sequences with farther divergences crossing different genetic category (i.e.,
17
chromosome and plasmid) over large size ranges (40Kb ~ 3Mb) (Table 4) that would
be challenging for conventional methods. In short, the UGFM-TGCC-SCG vision
clearly confirmed that two strains were completely different and eight components had
farther divergences (Figure 13, B). Particularly, two chromosomes were almost the
same (Figure 13, B) but six plasmids had larger divergences (Figure 13, A), strongly
indicating that plasmids had no close lineages with chromosomes in two strains. In
addition, two mega-plasmids in strain NRC-1 had no close lineages with four plasmids
in strain R1, suggesting there was no possibility to misassemble them due to less
homologous. Most interestingly, even within the same strain, chromosome and
plasmid showed distinctive lineage divergences. In other words, there was no
correlation between chromosome and plasmid within a certain strain (i.e., no binding
to a certain strain), indicating possible independent evolution among chromosomes
and plasmids.
Double-check between UGFM-TGCC-SCG and Mauve
To double-check the lineages revealed by UGFM-TGCC-SCG (Figure 13), we used
progressiveMauve mode [14] analysis to make pair-wisely multiple genome alignment
among eight components (Figure 14, A) showing overall bare homologous although it
took much longer time; whereas Mauve mode [15] analysis failed in such a
comparison because it stopped alignment due to no essential homologous, as we
predicted beforehand. Mauve mode [15] analysis yet worked well, separately, with
subsets of six plasmids (Figure 14, B) and two chromosomes (Figure 14, C),
respectively, and confirmed those partial relationships among six plasmids and
between two chromosomes. We thus concluded that the lineage relationships
revealed by UGFM-TGCC-SCG could be partially confirmed by Mauve method and
confirmed that two strains were completely different and remained as sister-strains
18
within one species. In other words, in this case, progressiveMauve mode [14] could
barely compatible to UGFM-TGCC-SCG (III) whereas Mauve mode [15] did not, but it
could be used to deal with subsets, separately.
Discussion
As mentioned before, how to compare whole genome sequences at large scale has
not been achieved by using conventional methods [11, 14] that based on pair-wisely
base-to-base sequence similarity analysis; even no attention was paid to handle
in-one-sitting a number of genomes crossing different genetic category with farther
divergences (e.g., less or no homologous among crossing genetic components:
chromosome, plasmid, and phage; bacterium, archaeal bacterium, and virus) over
large size ranges (e.g., from Kbp to Mbp per genome sequence). We believe that how
to persue comparative genomics at large scale based on geometrical analysis of
sequence, rather than pair-wisely base-to-base comparison, will be a priority in the
post-genomic era. However, little progress has been made to date, even how to
visualize a DNA sequence has been challenging for decades [1]. To our knowledge to
date, no method for creating “unambiguous genome fingerprint (GF)” was
documented; no concept of “universal genome fingerprint analysis (UGFA)”, or “total
genetic component configuration (TGCC)”, or “systematic comparative genomics
(SCG)“ was proposed. Particularly, note that all sequences of components both
crossing different genetic category (e.g., chromosome, plasmid, and phage;
bacterium, archaeal bacterium, and virus) and crossing a number of diverse strains
in-one-sitting should be much less or no homologous at all, which would be incredibly
challenging to any known conventional methods that principally based on pair-wisely
base-to-base homologous analysis. No conventional method could handle
19
in-one-sitting such farther diverse genetic components. Therefore, it would be
impossible to compare our methods, GenomeFingerprinter and universal genome
fingerprint analysis (UGFA), as a whole system with other documented methods in
terms of advantages and disadvantages. However, in the present study, we tried best
to compare partial features with two programs partly related to ours.
GenomeFingerprinter vs. Zplotter
Validity
Zplotter (in Z-curve method as a geometrical-type approach) was mainly used to
create coordinates for subsequent use by Z-curve analysis, but not used for what we
proposed as creating the “genome fingerprint (GF)” and the “universal genome
fingerprint analysis (UGFA)” in the present study. Although Zplotter’s coordinates were
used to produce hundreds of graphs (as Z-curves) of microbial genomes documented
as a database [17], there were no stable features in terms of so-called fingerprints. In
fact, for example, we re-plotted visions for Halobacterium sp. NRC-1 genome
sequence [NC_002607] by using Zplotter’s coordinates of either zn’ (Figure 15, B) or
zn (Figure 15, C) to present as an open rough Z-curve. Note that those visions
themselves created by using zn’ and zn, respectively, were quite different from each
other due to wavelet transform in the algorithm of Zplotter [16]. In contrast, our
method presented a unique circular vision with accurate and delicate fingerprint for
the same genome (Figure 15, A). Also note that using zn’ (Figure 15, B) showed a
similar frame of vision to ours except that it was in an open rough Z-curve with lesser
features whereas using zn (Figure 15, C) gave a complete different vision from ours.
We thus recommend that our GenomeFingerprinter method could be an alternate of
Zplotter to provide more accurate and delicate coordinates for Z-curve analysis, but
should be aware of choosing whether zn from our method or zn’ from Zplotter, referring
20
to the specific questions for various researches.
Reliability
Furthermore, we had found a major problem when using Zplotter to handle circular
genome sequences with cutting-point errors. In fact, Zplotter was designed for a linear
sequence [16] because its algorithm depends on counting the absolute numbers of
bases starting from the “first” base in a given linear sequence. In fact, when a
deposited sequence as a linear form (i.e., no matter what the original form should be
as either linear or circular), even the same circular sequence with cutting-point errors
changing its real “first” base could be quite different for the input to Zplotter so that the
output visions were differently presented (Figure 15, B, C). In contrast, our method
was initially created for a circular sequence (Figure 1), but it could apply to a linear
form since linear one would be a specific form of circular one and particularly because
our method measured the relative distance in a circular form (as discussed with the
formula (1) before, Figure 1), rather than the absolute numbers of bases counting
from the “first” base in a linear sequence. For example, the same circular sequence
Halobacterium sp. NRC-1 (NC_002607) with different cutting-point (e.g.,
NC_002607_RC re-cut at 700 kbps) were incorrectly presented as different visions by
using Zplotter’s coordinates (Figure 15, B, C), whereas the exact same vision was
shown by our method (Figure 15, A). Thus our method was valid for both circular and
linear forms and no matter where the cutting-point was.
Adaptability
Finally, we would like to address the fundamental scientific principals for why dealing
with circular genomes should be critical for microbes. That was overlooked in
literatures before.
21
Theoretically, the circular form [32] would be much more stable than its linear
form in living cells. In nature, most microbial genomes are in circular double strands,
which protect them from natural degradation because of relatively simple structure.
Also, the circular genomes and their linear forms are usually changing into each other
only when they are living at certain functioning stages, such as rolling-model
replication and plasmid-mediated conjunction. Most importantly, circular and linear
forms are functioning both genetically and physiologically in a coordinated way for a
given genome in a given microbe. In other words, their forms are changeable into
each other only when responding to real living conditions [32, 33]. Anyway, we could
catch up the circular form status in life cycles.
Technically, the techniques and people in different groups were not yet unified to
guarantee all deposited genome sequences in correct forms. In fact, most sequences
deposited in public databases so far were neither in their natural orders of starting
from the real “first” base, nor in the direction from 5’ to 3’. We thus had to tackle with
such cutting-point errors, as illustrated by examples (Figure 15). Fortunately, as
mentioned before, the RD formula (1) in our method could virtually treat an arbitrary
linear sequence as a circular one (Figure 1), avoiding the impact of any possible
cutting-point errors existed in public deposited sequences.
Informatively, the closed (or in circular form) fingerprint beared much more
information, concerning with genome-wide comparative genomics at fingerprint level
(Figure 12, 13). Most importantly, our method was initially designed for circular forms
(Figure 1), but finally was proved not ambiguous for linear forms when dealing with
cutting-point errors (Figure 15). In other words, our method could precisely calculate
the three-dimensional coordinates for a given circular or linear sequence with or
without correct cutting-point, could accordingly present a unique genome fingerprint
22
giving a certain geometrical center ( x , y , z ) (Figure 15), and could consequently
guarantee the subsequent unambiguous trajectory projections. In short, our method
guaranteed the validity of universal genome fingerprint analysis.
To sum up, we conclude that GenomeFingerprinter has advantages over Zplotter
in creating unambiguous coordinates and therefore can be an alternate component of
Z-curve method, which can be widely applicable to all aspects established by Z-curve
method to date [18, 19, 20, 21, 22, 23] and beyond.
GenomeFingerprinter vs. Mauve
Efficiency
Mauve (as an typical algebraic-type program), combined computing and plotting
in-one-sitting, is commonly used for pair-wisely comparison and vision [14]. However,
it had difficulty with a number of larger genome sequences due to its inner constraints,
either too slow or memory overflow (MO). In contrast, our method could rapidly
calculate and visualize, separately, tens of large genomes. For example, Mauve had
at least O(n2) whereas our method had O(n) in time complexity (Table 5). By using our
method, only if plotting all larger graphics in-one-sitting would cause memory overflow.
Examples were five bacterial chromosome genomes (Figure 7), forty seven phage
genomes (Figure 8), and twenty four virus genomes (Figure 9), respectively, that
could be easily plotted out in-one-sitting. Particularly, our method calculated and
visualized, separately, and thus not only ensured the higher performance efficiency for
large set of genomes (Table 5) but also offered both inputs and outputs re-usable for
the subsequent processes of universal genome fingerprint analysis and beyond (e.g.,
for Z-curve analysis consequently).
On the other hand, Mauve had two modes: progressiveMauve mode [14] and
23
Mauve mode [15]. As discussed before, only progressiveMauve mode [14] could
partially deal with what we so-called systematic comparative genomics (Figure 14, A)
showing overall bare homologous although it took much longer time; whereas Mauve
mode [15] failed in the comparison because it stopped alignment due to no essential
homologous. Mauve mode [15] analysis yet worked well, separately, with subsets of
six plasmids (Figure 14, B) and two chromosomes (Figure 14, C), respectively, and
confirmed those partial relationships among six plasmids and between two
chromosomes.
Prediction
Mauve [14, 15] can visualize what it is, but can not predict what it should be without a
reference sequence or specific pre-knowledge. In contrast, our method provides
geometrical analysis of genome fingerprints with six trajectory projections, which
intuitively predict unique features such as genome-wide evolution events and
replication ori points (Figure 12), either based on a reference sequence or derived
from common knowledge.
Compatibility
From universal genome fingerprint analysis (UGFA), the subtle variations (Figure 12,
C, D, G) could predict genome-wide evolution events at small scale in chromosomes,
but no direct evidence yet could be drawn. Thus we used Mauve to pair-wisely
compare two genomes and confirmed genome-wide evolution events (Figure 12, C),
demonstrating that our method can rapidly predict evolution events and Mauve can
precisely test and confirm such predictions by showing out the predicted specific
regions (Figure 12, C). The same was true for UGFM-TGCC-SCG by our method
(Figure 13) and the pair-wisely comparison by progressiveMauve (Figure 14, A). Thus,
we recommend that our method and Mauve method are compatible and partners,
24
taking both advantages of our method for rapid and intuitive prediction in general and
Mauve for slow and precise confirmation in details, particularly focusing on the
targeted fragments’ gain, lose, and rearrangement, etc..
To conclude, methodologically, the universal genome fingerprint analysis (UGFA)
through UGFM-TGCC-SCG (Figure 13) and the pair-wisely genome comparison
through Mauve (Figure 14) could be compatible in a manner of sequential operations.
In other words, the UGFM-TGCC-SCG method not only could handle exceptional
situations for a large set of genomes, but also could facilitate the efficiency of
integrating Mauve into performing our so-called systematic comparative genomics,
particularly, in terms of in-one-sitting for a set of sequences with farther divergences
(chromosome, plasmid, and phage, if applicable) over large size ranges (e.g., 6Kbp ~
4Mbp). In other words, any component with too farther divergence could be rapidly
pre-screened out by UGFM-TGCC-SCG, which could guide on the selection of
appropriate subsets of components for subsequent comparison by Mauve.
Prospective in future for universal genome fingerprint analysis
Genome fingerprints and the concepts of strain and species
“Strain” should be the most fundamental unit for taxonomy. The concise definition of
type strain should be crucial for assigning type species, type genus, type family and
beyond. Any deep conflicts in arguable strains would eventually shape the assigned
species or beyond. Unfortunately, it was so critical but had been overlooked by
literatures. To our knowledge to date, no efficient method could provide full description
about a type strain, nor was there common agreement upon how to define a species
[26, 34, 35]. We anticipated that genomics would be the solid foundation for these
issues, as it had re-constructed the concepts of numbers - two or more instead of only
one - of chromosomes and replication ori points. For example, in the present study, by
25
using our method GenomeFingerprinter, we created the whole genome fingerprints
(Figure 12, 13) for all eight genetic components in two arguable strains (Table 4) and
fundamentally demonstrated that they were not identical (Figure 12, 13) and should
probably belong to the same species that needed more characterizations yet. These
findings supported the proposal that genome sequence information should be
considered in refining an arguable “strain” or “species” in the taxonomy of halophilic
Archaea [34]. We agreed with the promotion that in the long run, the definition for a
“species” in modern microbiology needed intensive revisions in light of genomics to
unify inevitable conflicts in nomenclature system, particularly, in halophilic Archaea [34,
35]. We would further recommend that all genetic components should be included
when referring to genomic information for discussing unambiguous taxonomy
although to what extent chromosome, plasmid and phage plays roles, respectively,
still remained unclear and to be negotiated at current knowledge level [26, 34, 35]. We
believed that using UGFM-TGCC-SCG method to concisely resolve the arguments
between closely related strains (Figure 13) as well as among farther divergence
species or genera (Figure 10, 11, 12, 13) would be one of the crucial steps forwarding
to modern microbial nomenclature in the post-genomic era.
Type UGFM-TGCC fingerprint for type strain
We would recommend that any arguable strains should not be judged identical or
different only based upon partial information from bulky traditional features such as
phenotype and genotype including 16S rRNA, AFLP, PCR-RFLP, ISs, MITEs, etc. [26,
27, 31, 34]. It should also be true for defining a type strain. Theoretically, we would
define a type strain or name a new isolate or refine an arguable strain or construct a
refined-version for modern microbial taxonomy based upon all unambiguous
information from total genome sequences (i.e., chromosome, plasmid, and phage, if
26
applicable). Practically, at least, a type strain should have a meaningful genomic
signature. For example, the UGFM-TGCC fingerprint (Figure 10, 11, 12, 13) would be
effective to provide a “type strain” with a “type UGFM-TGCC fingerprint” which is
simple, standard, and meaningful.
Interestingly, to date, the list of genomes sequenced does not include that of the
type strain of Halobacterium salinarum (ATCC 33171), the type species of the type
genus of the family and the order [36]. It is regretted that no genomic information is
available for the nomenclatural type [36]. We expect the community should consider
of sequencing more type strains in order to set up a solid foundation for refining
modern Archaea taxonomy, which would be invaluable for the next generation of
community to understand deeply, research systematically and use efficiently of such
amazing bio-resources. Once the “type genome sequence” for the “type strain” is
available, the “type UGFM-TGCC fingerprint” can be made by using our methods, as
what we did for two related strains Halobacterium NRC-1 and Halobacterium
salinarum R1 (Figure 13) and five diverse strains crossing five genera (Figure 10).
Overall, the family Halobacteriaceae consisted of 36 genera with 129 species
standing in nomenclature (as of November 2011) [26], but only sixteen strains
representing sixteen genera had been sequenced and deposited in GenBank (as of
February 2013), including eighteen chromosomes and thirty-six plasmids (Table 2).
By using our method of UGFM (I), UGFM-TGCC (II), and UGFM-TGCC-SCG (III), we
created the whole genome fingerprints (Figure 10, 11, 12, 13) for all fifty-four genome
sequences. Our results provided new insights into critical problematic issues in
halophilic Archaea genomics, comparative genomics, and taxonomy [26, 34]. That
was a great step on initiatives. We expected more pioneering works to be done. In
short, the present paper provided a new method (GenomeFingerprinter, Figure 1) for
27
rapidly computing, geometrically visualizing, and intuitively comparing sequences at
fingerprint level, and hence established a new method (universal genome fingerprint
analysis (UGFA), Figure 6) for systematic comparative genomics, which would be
invaluable for the first strategic step forwarding to microbial genomics, comparative
genomics, phylogenetics, and taxonomy in the light of post-genomics. We anticipated
that our methods could be widely applicable to systematic comparative genomics.
Conclusions
We created a new method, GenomeFingerprinter, to unambiguously produce
three-dimensional coordinates from a sequence, followed by one three-dimensional
plot and six two-dimensional trajectory projections to illustrate whole genome
fingerprints. We further developed a set of concepts and tools (3D-P, 2D-TP, GF, GFM,
P-GFM, S-GFM, UGFM, TGCC, UGFM-TGCC, SCG, and UGFM-TGCC-SCG), and
thereby established a new method, universal genome fingerprint analysis (UGFA). We
demonstrated their applications through case studies on over a hundred of genome
sequences. Particularly, by using UGFM (I), UGFM-TGCC (II), and
UGFM-TGCC-SCG (III), we compared a number of genome sequences crossing
different genetic category (chromosome, plasmid, and phage; bacterium, archaeal
bacterium, and virus) with farther divergences over large size ranges (6Kbp~5Mbp),
which we called as systematic comparative genomics, giving new insights into critical
problematic issues in microbial genomics. We anticipated that our methods could be
widely applicable to systematic comparative genomics in the post-genomic era.
Materials
28
Genome sequences used in this study were downloaded from NCBI or were derived
from this study were list in Table 2, 3.
Methods
We implemented our method into an in-house script, GenomeFingerprinter.exe. It will
be available upon request to the corresponding author. Zplotter (v1.0) and Mauve
(v2.3.1) used in this study can be downloaded from links: Zplotter.exe at
http://tubic.tju.edu.cn/zcurve/ and Mauve at http://gel.ahabs.wisc.edu/mauve/. To plot
graphics from coordinates, any public graphic tool can be used.
Abbreviations
bps: base pairs; Kbp: kilo-bps; Mbp: Mega-bps; TB: target base; FB: focusing base;
RD: relative distance; WRD: weighted relative distance; SWRD: sum of the weighted
relative distances; 3D-P: three-dimensional plot; 2D-TP: two-dimensional trajectory
projections; GF: genome fingerprint; GFM: genome fingerprint map; P-GFM: primary
genome fingerprint map; S-GFM: secondary genome fingerprint map; UGFM:
universal genome fingerprint map; TGCC: total genetic component configuration;
UGFM-TGCC: universal genome fingerprint map of total genetic component
configuration; SCG: systematic comparative genomics; UGFM-TGCC-SCG: universal
genome fingerprint map of total genetic component configuration based systematic
comparative genomics; UGFA: universal genome fingerprint analysis
Acknowledgements
This work was supported by National High Technology Research & Development
Project (2006AA09Z420), National Science and Technology Major Project of China
(2011ZX08011005) grants to YA. HA was a recipient of Guangzhou Municipal Science
Ambassador Scholarship.
29
Authors’ contributions
Conceived and designed the experiments: YA and FM. Performed experiments: LZ
(Perl) and HA (Perl and Java) initiated models and scripts; HA reconstructed
mathematic models and algorithms, designed, implemented, tested scripts, and
constructed system; YA, FM and HA initiated, developed and confined the conceptual
frameworks for biological research contents; HA and YA performed computing and
collected data. Analyzed the data: HA, YA and FM. Wrote the paper: YA, HA and FM.
Competing interests
The author(s) declare that they have no competing interests.
References
1. Nielsen CB, Cantor M, Dubchak I, Gordon D, Wang T. (2010) Visualizing
genomes: techniques and challenges. Nat Methods 7(Suppl 3): S5-S15.
2. Lathe R, Findlay R. (1984) Machine-readable DNA sequences. Nature 311: 610.
3. Lathe R, Findlay R. (1985) Reply. Nature 314: 585-586.
4. Gates MA. (1986) A simple way to look at DNA. J Theor Biol 119: 319-328.
5. Hamori E, Ruskin J. (1983) H-Curves, a novel method of representation of
nucleotide series especially suited for long DNA sequences. J Biol Chem 258:
1318-1327.
6. Hamori E. (1985) Novel DNA sequence representations. Nature 314: 585.
7. Herisson J, Ferey N, Gros P.E, Gherbi R. (2006) ADN-viewer: A 3D approach for
bioinformatic analyses of large DNA sequences. Cellular Mol Biol 52: 24-31.
8. Herisson J, Payen G, Gherbi R. (2007) A 3D pattern matching algorithm for DNA
30
sequences. Bioinformatics 23: 680-686.
9. Altschul S, Madden TL, Schaffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ.
(1997) Gapped BLAST and Psi-BLAST: A new generation of protein database
search programs. Nucl Acids Res 25: 3389-3402.
10. Thompson JD, Higgins DG, Gibson TJ. (1994) CLUSTALW: improving the
sensitivity of progressive multiple sequence alignment through sequence
weighting, positions specific gap penalties and weight matrix choice. Nucleic
Acids Res 22:4673-4680.
11. Alikhan NF, Petty NK, Ben Zakour NL, Beatson SA. (2011) BLAST Ring Image
Generator (BRIG): simple prokaryote genome comparisons. BMC Genomics
12:402.
12. Carver TJ, Rutherford KM, Berriman M, Rajandream MA, Barrell BG, Parkhill J.
(2005) ACT: The artemis comparison tool. Bioinformatics 21: 3422-3423.
13. Lobry JR. (1996) A simple vectorial representation of DNA sequences for the
detection of replication origins in bacteria. Biochimie 78: 323-326.
14. Aaron E, Darling, Bob Mau, Nicole T. Perna. (2010) progressiveMauve: Multiple
genome alignment with gene gain, loss, and rearrangement. PLoS One 5:
e11147.
15. Darling ACE, Mau B, Blatter FR, Perna NT. (2004) Mauve: multiple alignment of
conserved genomic sequence with rearrangements. Genome Research 14:
1394-1403.
16. Zhang R, Zhang CT. (1994) Z Curves, An intuitive tool for visualizing and
analyzing the DNA sequences. J Biomol Struc Dynamics 11: 767-782.
31
17. Zhang CT, Zhang R, Ou HY. (2003) The Z curve database: a graphic
representation of genome sequences. Bioinformatics 19: 593-599.
18. Guo FB, Ou HY, Zhang CT. (2003) ZCURVE: a new system for recognizing
protein-coding genes in bacterial and archaeal genomes. Nucleic Acids Res 31:
1780-1789.
19. Zheng WX, Chen LL, Ou HY, Gao F, Zhang CT. (2005) Coronavirus phylogeny
based on a geometric approach. Mol Phy and Evo 36: 224-232.
20. Zhang CT, Gao F, Zhang R. (2005) Segmentation algorithm for DNA sequences.
Phys Rev E 72: 041917.
21. Gao F, Zhang CT. (2006) GC-Profile: a web-based tool for visualizing and
analyzing the variation of GC content in genomic sequences. Nucleic Acids Res
34(Web Server issue):W686-W691.
22. Zhang R, Zhang CT. (2002) Single replication origin of the archaeon
Methanosarcina mazei revealed by the Z curve method. Biochem Biophys Res
Commun 297: 396-400.
23. Zhang R, Zhang CT. (2003) Multiple replication origins of the archaeon
Halobacterium species NRC-1. Biochem Biophys Res Commun 302: 728-734.
24. Charkowski AO. (2004) Making sense of an alphabet soup: the use of a new
bioinformatics tool for identification of novel gene islands. Physiol Genomics 16:
180-181.
25. Robinson NP, Dionne I, Lundgren M, Marsh VL, Bernander R, Bell SD. (2004)
Identification of two origins of replication in the single chromosome of the
Archaeon Sulfolobus solfataricus. Cell 116: 25-38.
32
26. Oren, A. (2012) Taxonomy of the family Halobacteriaceae: a paradigm for
changing concepts in prokaryote systematics. Int J Syst Evol Microbiol
62:263-271.
27. File´e J, Siguier P, Chandler M. (2007) Insertion sequence diversity in Archaea.
Microbiol Mol Biol Rev 71:121-157.
28. Ng WV, Kennedy SP, Mahairas GG, Berquist B, Pan M, Shukla HD, Lasky SR,
Baliga NS, Thorsson V, Sbrogna J, Swartzell S, Weir D, Hall J, Dahl TA, Welti R,
Goo YA, Cruz R, Danson MJ, Hough DW, Maddocks DG, Jablonski PE, Krebs
MP, Angevine CM, Dale H, Isenbarger TA, Peck RF, Pohlschroder M, Spudich JL,
Jung KW, Alam M, Freitas T, Hou S, Daniels CJ, Dennis PP, Omer AD, Ebhardt
H, Lowe TM, Liang P, Riley M, Hood L, DasSarma S. (2000) Genome sequence
of Halobacterium species NRC-1. Proc Natl Acad Sci USA 97:12176-12181.
29. Pfeiffer F, Schuster SC, Broicher A, Falb M, Palm P, Rodewald K, Ruepp A,
Soppa J, Tittor J, Oesterhelt D. (2008) Evolution in the laboratory: the genome of
Halobacterium salinarum strain R1 compared to that of strain NRC-1. Genomics
91: 335-346.
30. Berquist BR, DasSarma S. (2003) An archaeal chromosomal autonomously
replicating sequence element from an extreme halophile, Halobacterium sp.
strain NRC-1. J Bacteriol 185: 5959-5966.
31. Ng WV, Ciufo SA, Smith TM, Bumgarner RE, Baskin D, Faust J, Hall B, Loretz C,
Seto J, Slagel J, Hood L, DasSarma S. (1998) Snapshot of a large dynamic
replicon in a halophilic archaeon: megaplasmid or minichromosome? Genome
Res 8: 1131-1141.
32. Kennedy SP, Ng WV, Salzberg SL, Hood L, DasSarma S. (2001) Understanding
33
the adaptation of Halobacterium species NRC-1 to its extreme environment
through computational analysis of its genome sequence. Genome Res
11:1641-1650.
33. Capes MD, DasSarma P, DasSarma S. (2012) The core and unique proteins of
haloarchaea. BMC Genomics 13:39.
34. DasSarma P, DasSarma S. (2008) On the origin of prokaryotic "species": the
taxonomy of halophilic Archaea. Saline Systems 4:5.
35. Oren, A. (2008) Nomenclature and taxonomy of halophilic archaea – comments
on the proposal by DasSarma and DasSarma for nomenclatural changes within
the order Halobacteriales. Int J Syst Evol Microbiol 58: 2245–2246.
36. Oren A, Arahal DR, Ventosa A. (2009) Emended descriptions of genera of the
family Halobacteriaceae. Int J Syst Evol Microbiol 59:637-642.
34
Figures and Legends
Figure 1. A mathematic model for getting coordinates (xn, yn, zn) from a circular
genome sequence. It is arbitrarily starting at the nth base as the chosen target base
(TB) and moving to the mth base as a focusing base (FB).
Figure 2. Three-dimensional plot (3D-P) and primary genome fingerprint map
(P-GFM) of fragmental genome sequences of chromosomes in Escherichia coli
strains. (A). K-12/W3110 [AC_000091]F7; (B). BL21(DE3)pLysS AG [NC_012947]F1;
(C). BL21(DE3)pLysS AG [NC_012947]F5; (D). O55:H7/CB 9615 [NC_013941]F1.
Figure 3. Six two-dimensional trajectory projections (2D-TP) and secondary
genome fingerprint maps (S-GFMs) for E. coli K-12/W3110 chromosome
[AC_000091]. (A). Projection with xn ~ n; (B). Projection with yn ~ n; (C). Projection
with zn ~ n; (D). Projection with xn ~ yn; (E). Projection with xn ~ zn; (F). Projection with
yn ~ zn.
Figure 4. Universal genome fingerprint map (UGFM) for overall comparison of
genome fingerprints. (A). Similar: Sulfolobus islandicus M.14.25 [NC_012588] and
M.16.4 [NC_012726]; (B). Partly similar: S. islandicus Y.N.15.51 [NC_012623] and
Methanococcus voltae A3 [NC_014222]; (C). Different: S. islandicus Y.G.57.14
[NC_012622] and Methanosphaera stadtmanae 3091 [NC_007681]; (D). Mixture:
(total twelve fragmental sequences (Table 2): 91.1.1, 91.1.61, 91.6.59, 913.1.77,
913.5.57, 4431.1.70, 7946.4.7, 10473.1.74, 10473.4.57, 10498.4.86, 12947.1.50,
35
13941.2.60.
Figure 5. Application of universal genome fingerprint map (UGFM) for
comparison among a number of genomes in-one-sitting. The twelve fragmental
genome sequences (list in Table 2) were shown in UGFM. The primary genome
fingerprint map (P-GFM) of each sequence was classified into different groups solely
based upon its location in UGFM: Group (A) (91.1.61, 913.1.77 and 10473.1.74),
Group (B) (91.6.59, 913.5.57 and 13941.2.60), Group (C) (7946.4.7 and 12947.1.50),
Group (D) (10498.4.86), Group (E) (91.1.1), and Group (F) (4431.1.70). Note that
each sequence showed quite different views between its own P-GFM and that in
UGFM simply because of the scale-down and view-angle rotation effect in UGFM,
which ensured for larger number of objects to be compared in-one-sitting.
Figure 6. Diagram of conceptual framework for universal genome fingerprint
analysis (UGFA). Our methods consisted of GenomeFingerprinter and universal
genome fingerprint analysis (UGFA); the former was the fundamental for the latter.
The objects could be a single genome sequence, or a number of genome sequences
crossing different genetic category (e.g., chromosome, plasmid, phage) in a strain, or
a number of sequences of genetic components in a cluster of strains crossing different
genetic category (e.g., bacterium, archaeal bacterium, virus). The UGFA method was
composed by three subcategories of UGFM (I), UGFM-TGCC (II), and
UGFM-TGCC-SCG (III), corresponding to the above three objects, respectively. The
core lied in the systematic concepts and tools, which included 3D-P, 2D-TP, P-GFM,
S-GFM, UGFM, TGCC, UGFM-TGCC, SCG, and UGFM-TGCC-SCG. Abbreviations:
3D-P: three-dimensional plot; 2D-TP: two-dimensional trajectory projections; GF:
36
genome fingerprint; GFM: genome fingerprint map; P-GFM: primary genome
fingerprint map; S-GFM: secondary genome fingerprint map; UGFM: universal
genome fingerprint map; TGCC: total genetic component configuration; UGFM-TGCC:
universal genome fingerprint map of total genetic component configuration; SCG:
systematic comparative genomics; UGFM-TGCC-SCG: universal genome fingerprint
map of total genetic component configuration based systematic comparative
genomics; UGFA: universal genome fingerprint analysis
Figure 7. UGFM (I) of five archaeal strains (each having only one chromosome
with size ranging of 2.6 ~4.1 Mbp) crossing five genera of halophilic Archaea.
Halomonas elongate DSM 2581 [NC_014532], Halorhodospira halophilia SL1
[NC_008789], Halorhabdus utahensis DSM 12940 [NC_013158], Halothermothrix
orenii H 168 [NC_011899] and Halothiobacillus neapolitanus c2 [NC_013422] had no
close lineages confirming their divergences, at genus level.
Figure 8. UGFM (I) of forty seven genomes of phages in the family of
Microviridae. These forty seven phages were close relatives, but most of them were
distinguishable at strain level. They were grouped into two major clusters. Cluster (1)
included twenty nine strains (WA5, ID11, WA3, WA2, ID41, NC10, WA6, ID12, NC13,
NC2, NC6, ID52, ID8, G4, ID2, WA14, ID18, WA45, ID21, NC28, ID62, NC35, NC29,
NC3, alpha3, WA13, phiK, ID32, NC19); Cluster (2) included eighteen strains (NC16,
NC5, NC37, ID1, NC7, NC1, NC11, ID22, S13, phiX174, WA11, WA4, ID34, NC41,
NC56, WA10, NC51, ID45). The details of phage names were list in Table 3.
37
Figure 9. UGFM (I) of twenty four genomes of coronavirus strains. They were
classified into seven clusters. Cluster (1) included the most similar twelve strains of
SARS coronavirus ([AY283796], [AY283797], [AY283798], [AY283794], [AY291451],
[AY278741], [AY283795], [AY278488], [AY278491], [AY278554], [NC_004718],
[AY282752]), tracking with the same UGFM; Cluster (2) included similar four strains of
Murine hepatitis virus ([AF201929], [AF208066], [AF208067], [NC_001846]), tracking
with the similar UGFM; Cluster (3) was a distinctive Porcine epidemic diarrhea virus
strain ([NC_003436]); Cluster (4) was a distinctive Avian infectious bronchitis virus
strain ([NC_001451]); Cluster (5) was a distinctive Feline infectious peritonitis virus
strain ([NC_002306]); Cluster (6) was a distinctive Human coronavirus strain
([NC_002645]); Cluster (7) included four similar strains of Bovine coronavirus
([AF220295], [u00735], [AF391542], [NC_003045]), tracking with the similar UGFM.
These seven clusters were perfectly matched to their biological identity groups (list in
Table 3).
Figure 10. UGFM-TGCC (II) of five archaeal strains crossing four genera of
halophilic Archaea. (A) enlarged vision of those five plasmids and (B)
Halogeometricum boringquense DSM 11551 (one chromosome [NC_014729] and five
plasmids pHBOR02 [NC_014731], pHBOR04 [NC_014732], pHBOR01 [NC_014735],
pHBOR03 [NC_014736], pHBOR05 [NC_014737]); (C) enlarged vision of those six
plasmids and (D) Haloterrigena turkmenica DSM 5511 (one chromosome
[NC_013743] and six plasmids pHTUR01 [NC_013744], pHTUR02 [NC_013745],
pHTUR03 [NC_013746], pHTUR04 [NC_013747], pHTUR05 [NC_013748],
pHTUR06 [NC_013749]); (E) enlarged vision of those three plasmids and (F)
Natrialba magadii ATCC 43099 (one chromosome [NC_013922] and three plasmids
38
pNMAG01 [NC_013923], pNMAG02 [NC_013924], pNMAG03 [NC_013925]; (G)
Natrinema pellirubrum DSM 15624 (one chromosome [NC_019962] and two plasmids
pNATPE02 [NC_019963], pNATPE01 [NC_019967]; (H) Haloquadratum walsbyi DSM
16790 (one chromosome [NC_008212] and one plasmid PL47 [NC_008213]. These
four strains crossing four genera (Table 2) had quite different UGFM-TGCC visions
demonstrating their farther lineages at genus level.
Figure 11. UGFM-TGCC-SCG (III) of four archaeal strains (each having multiple
chromosomes and plasmids) crossing four genera of halophilic Archaea. There
were two sets of in-one-sitting comparison. One set (A-B) : Halorubrum lacusprofundii
ATCC49239 (chromosome I [NC_012029], chromosome II [NC_012028], plasmid
pHLAC01 [NC_012030]) vs. Haloarcula marismortui ATCC43049 (chromosome I
[NC_006396], chromosome II [NC_006397], and seven plasmids pNG100
[NC_006389], pNG200 [NC_006390], pNG300 [NC_006391], pNG400 [NC_006392],
pNG500 [NC_006393], pNG600 [NC_006394], pNG700 [NC_006395]) focusing on
plasmids (A) and a universal system (B); Another set (C-D): Haloferax vocanii DS2
[chromosome [NC_013967], plasmid pHV3 [NC_013964], pHV2 [NC_013965], pHV4
[NC_013966], pHV1 [NC_013968] ) vs. Halomicrobium mukohataei DSM 12286
(chromosome [NC_013202], plasmid pHmuk01[NC_013201] ) focusing on plasmids
(C) and a universal system (D).
Figure 12. Application of universal genome fingerprint analysis (UGFA) for
comparative genomics between two chromosomes of Halobacterium sp. NRC-1
[NC_002607] and Halobacterium salinarum R1 [NC_010364]. Two arguable strains
were compared by using three-dimensional plots (xn~yn~zn) (P-GFM) (A) and
39
two-dimensional trajectory projections (S-GFM) with different combinations of
coordinates: (B) xn~yn; (C) xn~zn; (D) yn~zn; (E) xn~n; (F) yn~n; (G) zn~n; (H) xn~n and
yn~n together. Note two arrows showed replication ori points, oriC1 and ori C2; other
arrows indicated genome-wide evolution events.
Figure 13. Application of universal genome fingerprint analysis (UGFA) for
systematic comparative genomics (SCG). The universal genome fingerprint map
(UGFM) of total genetic component configurations (TGCC) (UGFM-TGCC) was
applied to the systematic comparative genomics (SCG) in-one-sitting. Two strains
Halobacterium sp. NRC-1 and Halobacterium salinarum R1 were compared as a
universal system. (A). UGFM-TGCC-SCG for total six plasmids (Halobacterium sp.
NRC-1 pNRC100 [NC_00001869] and pNRC200 [NC_002608]; Halobacterium
salinarum R1 PHS1 [NC_010366], PHS2 [NC_010369], PHS3 [NC_010368], and
PHS4 [NC_010367]); (B). UGFM-TGCC-SCG for those total six plasmids and two
chromosomes (Halobacterium sp. NRC-1 [NC_002607] and Halobacterium salinarum
R1 [NC_010364]). Note, even within the same strain, chromosome and plasmid
showed distinctive lineage divergences. In other words, there was no correlation
between chromosome and plasmid within a certain strain, i.e., without any binding to a
certain strain, indicating possible independent evolution among chromosomes and
plasmids.
Figure 14. Mauve snapshots for pair-wisely genome comparisons between two
strains Halobacterium sp. NRC-1 and Halobacterium salinarum R1 considering
of TGCC as a universal system. progressiveMauve mode analysis could compare
in-one-sitting all eight components of NRC-1 and R1 strains showing bare
40
homologous (A). Mauve mode analysis failed because it stopped alignment due to no
homologous; but it worked well, separately, with (B) six plasmids (the inner window)
(Halobacterium sp. NRC-1 pNRC100 [NC_00001869] and pNRC200 [NC_002608];
Halobacterium salinarum R1 PHS1 [NC_010366], PHS2 [NC_010369], PHS3
[NC_010368], and PHS4 [NC_010367]), and (C) two chromosomes (Halobacterium
sp. NRC-1 [NC_002607] and Halobacterium salinarum R1 [NC_010364]). Mauve
mode analysis could clearly reveal the relationships among six plasmids and between
two chromosomes, separately.
Figure 15. Comparisons between two chromosomes of Halobacterium sp.
NRC-1 [NC_002607] and its derivative form (NC_002607_RC) with different
cutting-point. (A). Comparison via GenomeFingerprinter; (B). Comparison via
Zplotter with zn’; (C). Comparison via Zplotter with zn.
41
Tables and Captions
Table 1. Coordinates of an artificial sample sequence
# bp / point
xn
yn
zn
1
0.7
8.1
-4.1
2
-0.3
8.7
-4.9
3
0.7
9.3
-3.7
4
-0.3
9.9
-4.5
5
-1.3
8.5
-3.3
6
-0.3
7.1
-4.1
7
0.7
7.7
-2.9
8
-0.3
8.3
-3.7
9
0.7
6.9
-4.5
10
-0.3
7.5
-5.3
11
0.7
8.1
-4.1
12
-0.3
8.7
-4.9
13
0.7
9.3
-3.7
14
-0.3
9.9
-4.5
15
-1.3
8.5
-3.3
16
-0.3
7.1
-4.1
17
0.7
7.7
-2.9
18
-0.3
8.3
-3.7
19
0.7
6.9
-4.5
20
-0.3
7.5
-5.3
21
0.7
8.1
-4.1
22
-0.3
8.7
-4.9
23
0.7
9.3
-3.7
24
-0.3
9.9
-4.5
25
-1.3
8.5
-3.3
26
-0.3
7.1
-4.1
27
0.7
7.7
-2.9
28
-0.3
8.3
-3.7
29
0.7
6.9
-4.5
30
-0.3
7.5
-5.3
31
0.7
8.1
-4.1
32
-0.3
8.7
-4.9
33
0.7
9.3
-3.7
34
-0.3
9.9
-4.5
35
-1.3
8.5
-3.3
36
-0.3
7.1
-4.1
37
0.7
7.7
-2.9
38
-0.3
8.3
-3.7
39
0.7
6.9
-4.5
40
-0.3
7.5
-5.3
42
Table 2. Features of genome sequences from bacteria and archaeal bacteria
used in this study
Species and Strain
Sequence ID
Type
Size (bps)
Downloaded from FTP.ncbi.nlm.nih.gov [GenBank]
Escherichia coli K-12/W3110
Escherichia coli K-12/DH10B
Escherichia coli K-12/MG1655
Escherichia coli BL21 (DE3) pLysSAG
Escherichia coli O55:H7/CB9615
Escherichia coli UTI89
Escherichia coli CFT073
Escherichia coli SMS-3-5
Sulfolobus islandicus M.14.25
Sulfolobus islandicus M.16.4
Sulfolobus islandicus Y.N.15.51
Sulfolobus islandicus Y.G.57.14
Methanococcus voltae A3
Methanosphaera stadtmanae DSM 3091
Halomonas elongate DSM 2581
Halorhodospira halophilia SL1
Halorhabdus utahensis DSM 12940
Halothermothrix orenii H 168
Halothiobacillus neapolitanus c2
Halogeometricum boringquense DSM 11551
Halogeometricum boringquense DSM 11551
Halogeometricum boringquense DSM 11551
Halogeometricum boringquense DSM 11551
Halogeometricum boringquense DSM 11551
Halogeometricum boringquense DSM 11551
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Haloterrigena turkmenica DSM 5511
Natrialba magadii ATCC 43099
Natrialba magadii ATCC 43099
Natrialba magadii ATCC 43099
Natrialba magadii ATCC 43099
Natrinema pellirubrum DSM 15624
Natrinema pellirubrum DSM 15624
Natrinema pellirubrum DSM 15624
Haloquadratum walsbyi DSM 16790
Haloquadratum walsbyi DSM 16790
Halorubrum lacusprofundii ATCC49239
Halorubrum lacusprofundii ATCC49239
Halorubrum lacusprofundii ATCC49239
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Haloarcula marismortui ATCC43049
Halomicrobium mukohataei DSM 12286
Halomicrobium mukohataei DSM 12286
Haloferax vocanii DS2
Haloferax vocanii DS2
Haloferax vocanii DS2
AC_000091
NC_007779
NC_010473
NC_000913
NC_012947
NC_013941
NC_007946
NC_004431
NC_010498
NC_012588
NC_012726
NC_012623
NC_012622
NC_014222
NC_007681
NC_014532
NC_008789
NC_013158
NC_011899
NC_013422
NC_014729
NC_014731
NC_014732
NC_014735
NC_014736
NC_014737
NC_013743
NC_013744
NC_013745
NC_013746
NC_013747
NC_013748
NC_013749
NC_013922
NC_013923
NC_013924
NC_013925
NC_019962
NC_019963
NC_019967
NC_008212
NC_008213
NC_012029
NC_012028
NC_012030
NC_006396
NC_006397
NC_006389
NC_006390
NC_006391
NC_006392
NC_006393
NC_006394
NC_006395
NC_013202
NC_013201
NC_013967
NC_013964
NC_013965
Chromosome*
4646332
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
Chromosome*
a
Chromosome*
a
Chromosome*
a
Chromosome*
a
Chromosome*
a
Chromosome*
b
Chromosome*
b
plasmid pHBOR02
b
plasmid pHBOR04
b
plasmid pHBOR01
b
plasmid pHBOR03
b
plasmid pHBOR05
b
Chromosome*
b
plasmid pHTUR01
b
plasmid pHTUR02
b
plasmid pHTUR03
b
plasmid pHTUR04
b
plasmid pHTUR05
b
plasmid pHTUR06
b
Chromosome*
b
plasmid pNMAG01
b
plasmid pNMAG02
b
plasmid pNMAG03
b
Chromosome*
b
plasmid pNATPE02
b
plasmid pNATPE01
b
Chromosome*
b
plasmid PL47
c
Chromosome I*
c
Chromosome II*
c
plasmid pHLAC01
c
Chromosome I*
c
Chromosome II*
c
plasmid pNG100
c
plasmid pNG200
c
plasmid pNG300
c
plasmid pNG400
c
plasmid pNG500
c
plasmid pNG600
c
plasmid pNG700
c
Chromosome*
c
plasmid pHmuk01
c
Chromosome*
c
plasmid pHV3
c
plasmid pHV2
4686137
4639675
4570938
5386352
5065741
5231428
5068389
2608832
2586647
2812165
2702058
1936387
1767403
4119315
2716716
3161321
2614977
2619785
2860838
343853
197618
367369
213355
17786
3944596
708474
419558
183364
174400
72078
16041
3805456
383753
258593
59323
3844629
279762
291912
3177244
47537
2774371
533457
437500
3176463
292165
33779
33930
40086
50776
134574
157519
416420
3154923
225032
2888440
444162
6450
43
Haloferax vocanii DS2
Haloferax vocanii DS2
Halobacterium sp.NRC-1
Halobacterium sp.NRC-1
Halobacterium sp.NRC-1
Halobacterium salinarum R1
Halobacterium salinarum R1
Halobacterium salinarum R1
Halobacterium salinarum R1
Halobacterium salinarum R1
NC_013966
NC_013968
NC_002607
NC_001869
NC_002608
NC_010364
NC_010366
NC_010369
NC_010368
NC_010367
c
plasmid pHV4
c
plasmid pHV1
d
Chromosome*
d
Plasmid pNRC100
d
Plasmid pNRC200
d
Chromosome*
d
Plasmid PHS1
d
Plasmid PHS2
d
Plasmid PHS3
d
Plasmid PHS4
644869
86308
2014239
191346
365425
2000962
147625
194963
284332
40894
Derivatives created in this study [based on those sequences from GenBank]
Escherichia coli K-12/W3110-91.1.1
Escherichia coli K-12/W3110-91.1.61
Escherichia coli K-12/W3110-91.6.59
Escherichia coli K-12/W3110-91.F7
Escherichia coli K-12/MG1655-913.1.77
Escherichia coli K-12/MG1655-913.5.57
Escherichia coli CFT073-4431.1.70
Escherichia coli UTI89-7946.4.7
Escherichia coli K-12/DH10B -10473.1.74
Escherichia coli K-12/DH10B -10473.4.57
Escherichia coli SMS-3-5-10498.4.86
Escherichia coli BL21 (DE3) pLysSAG-12947.F1
Escherichia coli BL21 (DE3) pLysSAG-12947.1.50
Escherichia coli BL21 (DE3) pLysSAG-12947.F5
Escherichia coli O55:H7/CB9615-13941.F1
Escherichia coli O55:H7/CB9615-13941.2.60
91.1.1
91.1.61
91.6.59
91.7
913.1.77
913.5.57
4431.1.70
7946.4.7
10473.1.74
10473.4.57
10498.4.86
12947.1
12947.1.50
12947.5
13941.1
13941.2.60
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
Chromosome fragment
227694
324260
410186
953958
331163
408963
401260
518065
325622
412818
331536
1759795
470050
43254
1915479
267039
* 32 chromosomes used for calculations as list in Table 5.
a
UGFM (I): five strains and five genomes (Figure 7)
b
UGFM-TGCC (II): five strains and twenty two genomes (Figure 10)
c
UGFM-TGCC-SCG (III): four strains and nineteen genomes (Figure 11)
d
Case studies: two strains and eight genomes (Figure 12, 13, 14)
44
Table 3. Features of genome sequences from viruses and phages used in this
study
Species and Strain
Sequence ID
Type
Size (bps)
Downloaded from FTP.ncbi.nlm.nih.gov [GenBank]
WA5: Coliphage WA5
ID11: Coliphage ID11
WA3: Coliphage WA3
WA2: Coliphage WA2
ID41: Coliphage ID41
NC10: Coliphage NC10
WA6: Coliphage WA6
ID12: Coliphage ID12
NC13: Coliphage NC13
NC2: Coliphage NC2
NC6: Coliphage NC6
ID52: Coliphage ID52
ID8: Coliphage ID8
G4: Enterobacteria phage G4
ID2: Coliphage ID2
WA14: Coliphage WA14
ID18: Coliphage ID18
WA45: Coliphage WA45
ID21: Coliphage ID21
NC28: Coliphage NC28
ID62: Coliphage ID62
NC35: Coliphage NC35
NC29: Coliphage NC29
NC3: Coliphage NC3
alpha3: Enterobacteria phage alpha3
WA13: Coliphage WA13
phiK: Coliphage phiK
ID32: Coliphage ID32
NC19: Coliphage NC19
NC16: Coliphage NC16
NC5: Coliphage NC5
NC37: Coliphage NC37
ID1: Coliphage ID1
NC7: Coliphage NC7
NC1: Coliphage NC1
NC11: Coliphage NC11
ID22: Coliphage ID22
S13: Enterobacteria phage S13
phiX174: Coliphage phiX174
WA11: Coliphage WA11
WA4: Coliphage WA4
ID34: Coliphage ID34
NC41: Coliphage NC41
NC56: Coliphage NC56
WA10: Coliphage WA10
NC51: Coliphage NC51
ID45: Coliphage ID45
SARS coronavirus TW1
SARS coronavirus Sin2679
SARS coronavirus Sin2748
SARS coronavirus Sin2774
NC_007847
NC_006954
NC_007845
NC_007844
NC_007851
NC_007854
NC_007852
NC_007853
NC_007849
NC_007848
NC_007855
NC_007825
NC_007846
NC_001420
NC_007817
NC_007857
NC_007856
NC_007822
NC_007818
NC_007823
NC_007824
NC_007820
NC_007827
NC_007826
DQ085810
NC_007821
NC_001730
NC_007819
NC_007850
NC_007836
NC_007833
NC_007837
NC_007828
NC_007834
NC_007832
NC_007835
NC_007829
NC_001424
NC_001422
NC_007843
NC_007841
NC_007830
NC_007838
NC_007840
NC_007842
NC_007839
NC_007831
AY283796
AY283797
AY283798
AY283794
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
a
Phage chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
5737
5737
5700
5700
5737
5687
5687
5687
5737
5737
5687
5698
5700
5737
5644
5644
5644
6242
6242
6239
6225
6213
6439
6273
6177
6242
6263
6245
5737
5540
5540
5540
5540
5540
5540
5540
5540
5540
5540
5541
5540
5540
5540
5540
5540
5540
5540
30137
30132
30137
30137
45
SARS coronavirus Sin2500
SARS coronavirus Urbani
SARS coronavirus Sin2677
SARS coronavirus BJ01
SARS coronavirus HKU-39849
SARS coronavirus CUHK-W1
SARS coronavirus
SARS coronavirus CUHK-Su10
Murine hepatitis virus strain 2
Murine hepatitis virus strain Penn 97-1
Murine hepatitis virus strain ML-10
Murine hepatitis virus strain A59
Porcine epidemic diarrhea virus
Avian infectious bronchitis virus
Feline infectious peritonitis virus
Human coronavirus 229E
Bovine coronavirus strain Quebec
Bovine coronavirus strain Mebus
Bovine coronavirus isolate BCoV-LUN
Bovine coronavirus
AY291451
AY278741
AY283795
AY278488
AY278491
AY278554
NC_004718
AY282752
AF201929
AF208066
AF208067
NC_001846
NC_003436
NC_001451
NC_002306
NC_002645
AF220295
u00735
AF391542
NC_003045
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
b
Virus chromosome
30155
30153
30131
30151
30168
30162
30178
30162
31724
31558
31681
31806
28435
28004
29776
27709
31546
31477
31473
31473
a
UGFM (I): forty seven strains and forty seven genomes in phage (Figure 8)
b
UGFM (I): twenty four strains and twenty four genomes in virus(Figure 9)
46
Table 4. Features of total genetic component configurations of strains NRC-1
and R1
Components
Chromosome
(bp)
Plasmid 1
(bp)
Plasmid 2
(bp)
NRC-1
NC_002607
2014239
pNRC100
NC_001869
191346
pNRC200
NC_002608
365425
Plasmid 3
(bp)
-
Plasmid 4
(bp)
-
Total (bp)
2571010
R1
NC_010364
2000962
PHS1
NC_010366
147625
PHS2
NC_010369
194963
PHS3
NC_010368
284332
PHS4
NC_0103667
40,894
2668776
47
Table 5. GenomeFingerprinter vs. Mauve
Number
1 chr.
2 chr.
4 chr.
8 chr.
16 chr.
32 chr.
GenomeFingerprinter
1 min cal.
1 min plot
2 min cal.
2 min plot
4 min cal.
4 min plot
8 min cal.
8 min plot
16 min cal. 16 min plot
32 min cal. 32 min plot
Mauve
no valid
2 min
8 min
44 min
332 min
MO
* Notes:
1) MO: memory overflow;
2) Samples: 32 chromosomes (chr.) as list (*) in Table 2;
3) Conditions: HP Proliant server DL580-G5 with 16 CPU/8Gb memory.
48
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
Figure
Click here to download high resolution image
| 5cs.CE
|
A Practical Quantum Instruction Set Architecture
Robert S. Smith, Michael J. Curtis, William J. Zeng
Rigetti Computing
775 Heinz Ave.
Berkeley, California 94710
Email: {robert, spike, will}@rigetti.com
arXiv:1608.03355v2 [quant-ph] 17 Feb 2017
IV
Abstract—We introduce an abstract machine architecture for classical/quantum computations—including
compilation—along with a quantum instruction language called Quil for explicitly writing these computations. With this formalism, we discuss concrete implementations of the machine and non-trivial algorithms
targeting them. The introduction of this machine dovetails with ongoing development of quantum computing
technology, and makes possible portable descriptions
of recent classical/quantum algorithms. Keywords—
quantum computing, software architecture
V
Contents
I
Introduction
1
II
The Quantum Abstract Machine
2
II-A
Qubit Semantics . . . . . . . . . . . .
2
II-B
Quantum Gate Semantics . . . . . . .
3
II-C
Measurement Semantics . . . . . . . .
4
III
Quil: a Quantum Instruction Language
5
III-A
Classical Addresses and Qubits . . . .
5
III-B
Numerical Interpretation of Classical
Memory Segments . . . . . . . . . . .
5
III-C
Static and Parametric Gates . . . . .
5
III-D
Gate Definitions . . . . . . . . . . . .
6
III-E
Circuits . . . . . . . . . . . . . . . . .
6
III-F
Measurement . . . . . . . . . . . . . .
6
III-G
Program Control . . . . . . . . . . . .
7
III-H
Zeroing the Quantum State . . . . . .
7
III-I
Classical/Quantum Synchronization .
7
III-J
Classical Instructions . . . . . . . . .
7
III-J1
Classical Unary Instructions
8
III-J2
Classical Binary Instructions
8
III-K
The No-Operation Instruction . . . .
8
III-L
File Inclusion Semantics . . . . . . .
8
III-M
Pragma Support . . . . . . . . . . . .
8
III-N
The Standard Gates . . . . . . . . . .
9
Copyright c 2016 Rigetti & Co., Inc.—v2.0.20170217
VI
Quil Examples
9
IV-A
Quantum Fourier Transform . . . . .
9
IV-B
Variational Quantum Eigensolver . .
9
IV-B1
Static Implementation . . .
9
IV-B2
Dynamic Implementation .
10
Forest: A Quantum Programming Toolkit 10
V-A
Overview . . . . . . . . . . . . . . . .
10
V-B
Applications and Tools . . . . . . . .
11
V-C
Quil Manipulation . . . . . . . . . . .
11
V-D
Compilation . . . . . . . . . . . . . .
11
V-E
Instruction Parallelism . . . . . . . .
12
V-F
Rigetti Quantum Virtual Machine . .
12
13
Conclusion
VII Acknowledgements
13
Appendix
13
A
The Standard Gate Set . . . . . . . .
13
B
Prior Work . . . . . . . . . . . . . . .
13
B1
Embedded
DomainSpecific Languages . . . . .
13
B2
High-Level QPLs . . . . . .
13
B3
Low-Level Quantum Intermediate Representations . .
14
I.
Introduction
The underlying hardware for quantum computing has
advanced rapidly in recent years. Superconducting chips
with 4–9 qubits have been demonstrated with the performance required to run quantum simulation algorithms [1,
2, 3], quantum machine learning [4], and quantum error
correction benchmarks [5, 6, 7].
Hybrid classical/quantum algorithms—including variational quantum eigensolvers [8, 9, 10], correlated material simulations [11], and approximate optimization [12]—
have much promise in reducing the overhead required
Fig. 1.
Classical Data
Classical Computation
The state of the QAM is specified by the following
elements:
Quantum Computation
A classical/quantum feedback loop.
for valuable applications. In machine learning and quantum simulation, particularly for catalysts [13] and hightemperature superconductivity [9], scalable quantum computers promise performance unrivaled by classical supercomputers.
The promise of hardware and applications must be
matched with advances in programming architectures. The
demands of practical algorithm design, as well as the
shift to hybrid classical/quantum algorithms, necessitate
an update to the quantum Turing machine model [14].
We need new frameworks for quantum program compilation [15, 16, 17, 18, 19] and emulation [20, 21]. For more
details on prior work and its relationship to the topics
introduced here, we refer the reader to Appendix B.
These classical/quantum algorithms require a classical
computer and quantum computer to communicate and
work cooperatively, as in Figure 1. Within the presented
framework, classical information is fed back via a defined
memory model, which can be implemented efficiently in
both hardware and software.
In this paper, we describe an abstract machine which
serves as a model for hybrid classical/quantum computation. Alongside this, we describe an instruction language
for this machine called Quil and its suitability for program
analysis and compilation. Together these form a quantum
instruction set architecture (ISA). Lastly, we give
various examples of algorithms in Quil, and discuss an
executable implementation.
II.
The Quantum Abstract Machine
Turing machines provide a vehicle for studying important concepts in computer science such as complexity and
computational equivalence. While theoretically important,
they do not provide a foundation for the construction
of practical computing machines. Instead, specialized abstract machines are designed to accomplish real-world
tasks, like arithmetic, efficiently while maintaining Turing
completeness. These machines are often specified in the
form of an instruction set architecture. Quantum Turing
machines lie in the same vein as its classical counterpart,
and we follow a similar approach in the creation of a
practical quantum analog.
The Quantum Abstract Machine (QAM) is an abstract representation of a general-purpose quantum computing device. It includes support for manipulating both
classical and quantum state. The QAM executes programs
represented in a quantum instruction language called Quil,
which has well-defined semantics in the context of the
QAM.
• A fixed but arbitrary number of qubits Nq indexed
from 0 to Nq − 1. The k th qubit is written Qk . The
state of these qubits is written |Ψi and is initialized
to |00 . . . 0i. The semantics of the qubits are described
in Section II-A.
• A classical memory C of Nc bits, initialized to zero
and indexed from 0 to Nc − 1. The k th bit is written
C[k].
• A fixed but arbitrary list of static gates G, and a fixed
but arbitrary list of parametric gates G0 . These terms
are defined in Section III-C.
• A fixed but arbitrary sequence of Quil instructions P .
These instructions will be described in Section III.
• An integer program counter 0 ≤ κ ≤ |P | indicating
position of the next instruction to execute when κ 6=
|P | or a halted program when κ = |P |.
The 6-tuple (|Ψi , C, G, G0 , P, κ) summarizes the state of
the QAM.
The QAM may be implemented either classically or on
quantum hardware. A classical implementation is called a
Quantum Virtual Machine (QVM). We describe one
such implementation in Section V-F. An implementation
on quantum hardware is called a Quantum Processing
Unit (QPU).
The semantics of the quantum state and operations
on it are described in the language of tensor products of
Hilbert spaces and linear maps between them. The following subsections give these semantics in meticulous detail.
Readers with intuition about these topics are encouraged
to skip to Section III for a description of Quil.
A. Qubit Semantics
A finite-dimensional Hilbert space over the complex
numbers C is denoted by H . The state space of a qubit is
a two-dimensional Hilbert space over C and is denoted by
B. Each of these Hilbert spaces is equipped with a chosen
orthonormal basis and indexing map on that basis. An
indexing map is a bijective function that maps elements
of a finite set Σ to the set of non-negative integers below
|Σ|, denoted [|Σ|]. For a Hilbert space spanned by {|ui , |vi}
with an indexing map defined by |ui 7→ 0 and |vi 7→ 1, we
write |0i := |ui and |1i := |vi.
In the context of the QAM, each qubit Qk in isolation
has a state space Bk spanned by an orthonormal basis
{|0ik , |1ik } called the computational basis. Since qubits
can entangle, the state space of the system of all qubits is
not a direct product of each constituent space, but rather
a rightward tensor product
Nq −1
|Ψi ∈ H :=
O
BNq −k−1 .
(1)
k=0
The meaning of the tensor product is as follows. Let |pii be
the pth basis vector of Hi according to its indexing map.
The tensor product of two Hilbert spaces is then
o
n X
Hi ⊗Hj :=
Cp,q |pii ⊗ |qij : C ∈ Cdim Hi ×dim Hj .
| {z }
p∈[dim Hi ]
q∈[dim Hj ]
basis element
(2)
The resulting basis elements are ordered by way of the
lexicographic indexing map
|pii ⊗ |qij 7→ q + p dim Hj .
(3)
Having the basis elements of the Hilbert space Hi “dominate” the indexing map is a convention due to the standard
definition of the Kronecker product, in which for matrices
A and B, A ⊗ B is a block matrix (A ⊗ B)i,j = Ai,j B.
This convention, while standard, somewhat muddles the
semantics below with busy-looking variable indexes which
count down, not up.
A basis element of H
|bNq −1 iNq −1 ⊗ · · · ⊗ |b1 i1 ⊗ |b0 i0
can be written shorthand in bit string notation
|bNq −1 . . . b1 b0 i .
The controlled-X or controlled-not gate with control
qubit Qj and target qubit Qk is defined as
1 0 0 0
0 1 0 0
: Bj ⊗ Bk → Bj ⊗ Bk .
CNOT :=
0 0 0 1
0 0 1 0
This gate is common for constructing entanglement in a
quantum system.
It turns out that many different sets of these oneand two-qubit gates are sufficient for universal quantum
computation, i.e., a discrete set of gates can be used
to approximate any unitary matrix to arbitrary accuracy [22, 23]. In fact almost any two-qubit quantum gate
can be shown to be universal [24]. Delving into different
universal gate sets is beyond the scope of this work and
we refer the reader to [23] as a general reference.
We now wish to provide a constructive method for
interpreting operators on a portion of a Hilbert space as
operators on the Hilbert space itself. In the simplest case,
we have a one-qubit gate U acting on qubit Qk , which
induces an operator Ũ on H by tensor-multiplying with
the identity map a total of Nq − 1 times:
This has the particularly useful property that the bit string
corresponds to the binary representation of the index of
that basis element. For clarity, however, we will not use
this notation elsewhere in this paper.
Ũ = INq −1 ⊗ · · · ⊗ |{z}
U ⊗ · · · ⊗ I1 ⊗ I0 .
Example 1. A two-qubit system in the Hilbert space B2 ⊗
B1 has the lexicographic indexing map defined by
We refer to this process as lifting and reserve the tilde
over the operator name to indicate such.
|0i2 ⊗ |0i1 7→ 0,
|1i2 ⊗ |0i1 7→ 2,
|0i2 ⊗ |1i1 7→ 1,
|1i2 ⊗ |1i1 7→ 3.
The standard Bell state in this system is represented by the
element in the tensor space with the matrix
1 1 0
C= √
.
2 0 1
Eqs. (1) and (2) imply that dim H = 2Nq , and as such,
|Ψi can be represented as a complex vector of that length.
(Nq − k − 1)
th
(4)
position (zero-based)
Example 3. Consider a system of four qubits and consider a Hadamard gate acting on Q2 . Lifting this operator
according to (4) gives
H̃ = I3 ⊗ H ⊗ I1 ⊗ I0 .
A two-qubit gate acting on adjacent Hilbert spaces is
just as simple. An operator U : Bk ⊗ Bk−1 → Bk ⊗ Bk−1
is lifted as
Ũ = INq −1 ⊗ · · · ⊗ U ⊗ · · · ⊗ I1 ⊗ I0 .
|
{z
}
(5)
Nq − 1 factors, U at position Nq − k − 1
B. Quantum Gate Semantics
Example 2. The Hadamard gate on Qk is defined as
1 1 1
: Bk → Bk .
H := √
2 1 −1
However, when the Hilbert spaces are not adjacent, lifting involves a bit more bookkeeping because there is no
obvious place to tensor-multiply the identity maps of the
Hilbert spaces indexed between j and k. We can resolve
this by suitably rearranging the space. We need two tools:
a method to “reorganize” an operator’s action on a tensor
product of Hilbert spaces and isomorphisms between these
Hilbert spaces. The general principle to be employed is
to find some operator π : H → H which acts as a
permutation operator on H , and to compute π −1 U 0 π,
where U 0 is isomorphic to U . Simply speaking, π is a
temporary re-indexing of basis vectors and U 0 is just a
trivial reinterpretation of U .
This is unitary because HH† = Ik is the identity map on
Bk .
To reorganize, we use the fact that any permutation can
be decomposed into adjacent transpositions. For swapping
The quantum state of the system evolves by applying a
sequence of operators called quantum gates. Most generally, these are complex unitary matrices of size 2Nq × 2Nq ,
written succinctly as U(2Nq ). However, quantum gates
are typically abbreviated as one- or two-qubit unitary
matrices, and go through a process of “tensoring up” before
application to the quantum state.
two qubits, consider the
1 0 0
0 0 1
SWAPj,k :=
0 1 0
0 0 0
gate
0
0
: Bj ⊗ Bk → Bj ⊗ Bk . (6)
0
1
This is a permutation matrix which maps the basis elements according to
We end this section with an example of a universal
QAM.
Example 4. Define all possible liftings of U within H as
L(U ) := {U lifted for all qubit permutations}.
0
Define the gates S := ( 10 0i ) and T := 10 eiπ/4
. A QAM
with2
|0ij ⊗ |1ik →
7 |1ij ⊗ |0ik
|1ij ⊗ |0ik →
7 |0ij ⊗ |1ik ,
G = L(H) ∪ L(S) ∪ L(T) ∪ L(CNOT) and G0 = {}
and mapping the others identically. For adjacent transpositions in the full Hilbert space, this can be trivially lifted:
can compute to arbitrary accuracy the action of any Nq qubit gate, possibly with P exponential in length. See [23,
§4.5] for details.
τi := SWAPi,i+1 lifted by way of (5).
(7)
We will use a sequence of these operators to arrange for
Bj and Bk to be adjacent by moving one of them next
to the other. There are two cases we need to be concerned
about: j > k and j < k.
For the j > k case, we want to map the state of Bk to
Bj−1 . This is accomplished with the operator1
πj,k :=
j−2
Y
Measurement is a surjective-only operation and is nondeterministic. In the space of a single qubit Qk , there
are two outcomes to a measurement. The outcomes are
determined by lifting and applying—up to a scalar factor—
either of the measurement operators3
M0k := |0ik h0|k
τj+k−i−2 ,
i=k
where the product right-multiplies and is empty when k ≥
j − 1.
For the j < k case, we want to map the state of Bj to
Bk−1 , and then swap Bk−1 with Bk . We can do this with
the π operator succinctly:
0
:= τk−1 πk,j .
πj,k
and
M1k := |1ik h1|k
to the quantum state. These can be interpreted as projections onto either of the basis elements of the Hilbert space.
More generally, in any finite Hilbert space H , we have the
set of measurement operators
M (H ) := |vi hv| : |vi ∈ basis H .
In the QAM, when a qubit is measured, a particular
measurement operator µ is selected and applied according
to the probability
Note the order of j and k in the π factor.
Lastly, for the purpose of correctness, we need to
construct a trivial isomorphism
f : Bj ⊗ Bk → Bj ⊗ Bj−1 ,
C. Measurement Semantics
(8)
which is defined as the bijection between basis vectors with
the same index.
Now we may consider two-qubit gates in their full
generality. Let U : Bj ⊗ Bk → Bj ⊗ Bk be the gate under
consideration. Perform a “change of coordinates” on U to
define
V := f U f −1 : Bj ⊗ Bj−1 → Bj ⊗ Bj−1 ,
and lift V to Ṽ : H → H by way of (5). Then the lifted
U can be constructed as follows:
(
−1
πj,k
Ṽ πj,k
if j > k, and
Ũ =
(9)
0
−1
0
(πj,k ) Ṽ πj,k if j < k.
Since the π operators are essentially compositions of involutive SWAP gates, their inverses are just reversals.
With care, the essence of this method generalizes
accordingly for gates acting on an arbitrary number of
qubits.
1 This is an example of the effects of following the Kronecker
convention. This product is really just “τi in reverse.”
P (µ) := P (µ̃ is applied during meas.) = hΨ| µ̃† µ̃ |Ψi .
(10)
Upon selection of an operator, the quantum state transforms according to
|Ψi ←
1
µ̃ |Ψi .
P (µ)
(11)
This irreversible operation is called collapse of the wavefunction.
In quantum mechanics, measurement is much more
general than the description given above. In fact, any
Hermitian operator can correspond to measurement. Such
operators are called observables, and the quantum state
of a system can be seen as a superposition of vectors in the
observable’s eigenbasis. The eigenvalues of the observable
are the outcomes of the measurement, and the corresponding eigenvectors are what the quantum state collapses to.
For additional details on measurement and its quantum
mechanical interpretation, we refer the interested reader
to [23] and [25].
2 When the context is clear, G is sometimes abbreviated to just be
the set of unlifted gates, but it should be understood that it’s actually
every lifted combination. See Section III-C.
3 In Dirac notation, hv| lies in the dual space of |vi.
III.
Quil: a Quantum Instruction Language
Quil is an instruction-based language for representing quantum computations with classical feedback and
control. In its textual format—as presented below—it is
line-based and assembly-like. It can be written directly
for the purpose of quantum programming, used as an
intermediate format within classical programs, or used as
a compilation target for quantum programming languages.
Most importantly, however, Quil defines what can be done
with the QAM. Quil has support for:
• Applying arbitrary quantum gates,
• Defining quantum gates as optionally parameterized
complex matrices,
• Defining quantum circuits as sequences of other
gates and circuits, which can be bit- and qubitparameterized,
• Expanding quantum circuits,
• Measuring qubits and recording the measurements
into classical memory,
• Synchronizing execution of classical and quantum algorithms,
• Branching unconditionally or conditionally on the
value of bits in classical memory, and
• Including files as modular Quil libraries such as the
standard gate library (see Appendix A).
By virtue of being instruction-based, Quil code effects
transitions of the QAM as a state machine. In the next
subsections, we will describe the various elements of Quil,
using the syntax and conventions outlined in Section II.
refers to a double-precision complex number a + ib where
a is the 64-bit interpretation of [x-(x + 63)], the first
half of the segment, and b is the 64-bit interpretation of
[(x + 64)-(x + 127)], the second half of the segment. The
use of these numbers can be found in Section III-C and
some practical consequences of their use can be found in
Section IV-B2.
C. Static and Parametric Gates
There are two gate-related concepts in the QAM: static
and parametric gates. A static gate is an operator in
U(2Nq ), and a parametric gate is a function4 Cn →
U(2Nq ), where the n complex numbers are called parameters. The implication is that operators in G and G0 are
always lifted to the Hilbert space of the QAM. This is a
formalism, however, and Quil abstracts away the necessity
to be mindful of lifting gates.
In Quil, every gate is defined separately from its invocation. Each unlifted gate is identified by a symbolic name5 ,
and is invoked with a fixed number of qubit arguments.
The invocation of a static (resp. parametric) gate whose
lifting is not a part of the QAM’s G (resp. G0 ) is undefined.
A static two-qubit gate named NAME acting on Q2 and
Q5 , which is an operator lifted from B2 ⊗ B5 , is written in
Quil as the name of the gate followed by the qubit indexes
it acts on, as in
NAME 2 5
Example 5. The Bell state on qubits Q0 and Q1 can be
constructed with the following Quil code:
A. Classical Addresses and Qubits
The central atomic objects on which various operations
act are qubits, classical bits designated by an address, and
classical memory segments.
Qubit A qubit is referred to by its integer index. For
example, Q5 is referred to by 5.
Classical memory address A classical memory address is referred to by an integer index in square
brackets. For example, the address 7 pointing to the
bit C[7] is referred to as [7].
Classical memory segment A classical memory segment is a contiguous range of addresses from a to
b inclusive with a ≤ b. These are written in square
brackets as well, with a hyphen separating the range’s
endpoints. For example, the bits between 0 and 63 are
addressed by [0-63] and represent the concatenation
of bits
C[63]C[62] . . . C[1]C[0],
written in the usual MSB-to-LSB order.
B. Numerical Interpretation of Classical Memory Segments
Classical memory segments can be interpreted as a
numerical value for the purpose of controlling parametric
gates. In particular, a 64-bit classical memory segment
refers to an IEEE-754 double-precision floating point number [26]. A 128-bit classical memory segment [x-(x+127)]
H 0
CNOT 0 1
A parametric three-qubit gate named PNAME with a
single parameter e−iπ/7 acting on Q1 , Q0 , and Q4 , which
is an operator lifted from B1 ⊗ B0 ⊗ B4 , is written in Quil
as
PNAME(0.9009688679-0.4338837391i) 1 0 4
When a parametric gate is provided with a constant
parameter, one could either consider the resulting gate on
the specified qubits to be a part of G, or the parametric
gate itself on said qubits to be a part of G0 .
Parametric gates can take a “dynamic parameter”,
as specified by a classical memory segment. Suppose a
parameter is stored in [8-71]. Then we can invoke the
aforementioned gate with that parameter via
PNAME([8-71]) 1 0 4
In some cases, using dynamic parameters can be expensive
or infeasible, as discussed in Section IV-B2. Gates which
use dynamic parameters are elements of G0 .
4 Calling a parametric gate a “gate” is somewhat of a misnomer.
The quantum gate is actually the image of a point in Cn .
5 To be precise, the symbolic name actually represents the equivalence class of operators under all trivial isomorphisms, as in (8).
D. Gate Definitions
Static gates are defined by their real or complex matrix
entries in the basis described in Section II-A. Matrix entries can be written literally with scientific E-notation (e.g.,
real -1.2e2 or complex 0.3-4.1e-4i = 0.3 − 0.00041i), or
built up from constant arithmetic expressions. These are:
• Simple arithmetic: addition +, subtraction/negation -,
multiplication *, division /, exponentiation ˆ,
• Constants: pi (= pi), i (= 1.0i), and
• Functions: sin, cos, sqrt, exp, cis6 .
The gate is declared using the DEFGATE directive followed
by comma-separated lists of matrix entries indented by
exactly four spaces. Matrices that are not unitary (up to
noise or precision) have undefined7 execution semantics.
With this, Example 5 is replicated by just a single line:
BELL 0 1
Similar to parametric gates, DEFCIRCUIT can optionally
specify a list of parameters, specified as a comma-separated
list in parentheses following the circuit name, as the following example shows.
Example 9. Using the x-y-z convention, an extrinsic
Euler rotation by (α, β, γ) of the state of qubit q on the
Bloch sphere is codified by the following circuit:
DEFCIRCUIT EULER(%alpha, %beta, %gamma) q:
RX(%alpha) q
RY(%beta) q
RZ(%gamma) q
Example 6. The Hadamard gate can be defined by
DEFGATE HADAMARD:
1/sqrt(2), 1/sqrt(2)
1/sqrt(2), -1/sqrt(2)
This gate is included in the collection of standard gates, but
under the name H.
Parametric gates are the same, except for the allowance
of formal parameters, which are names prepended with a
‘%’ symbol. Comma-separated formal parameters are listed
in parentheses following the gate name, as is usual.
Example 7. The rotation gate Rx can be defined by
DEFGATE RX(%theta):
cos(%theta/2), -i*sin(%theta/2)
-i*sin(%theta/2), cos(%theta/2)
This gate is also included in the collection of standard gates.
Defining several gates or circuits with the same name
is undefined.
E. Circuits
Sometimes it is convenient to name and parameterize
a particular sequence of Quil instructions for use as a
subroutine to other quantum programs. This can be done
with the DEFCIRCUIT directive. Similar to the DEFGATE
directive, the body of a circuit definition must be indented
exactly four spaces. Critically, it specifies a list of formal
arguments which can be substituted with either classical
addresses or qubits.
Example 8. In example 5, we constructed a Bell state on
Q0 and Q1 . We can generalize this for arbitrary qubits Qm
and Qn by defining a circuit.
DEFCIRCUIT BELL Qm Qn:
H Qm
CNOT Qm Qn
Within circuits, labels are renamed uniquely per expansion. As a consequence, it is possible to expand the same
circuit multiple times, but it is not possible to jump into
a circuit.
Circuits are intended to be used more as macros than
as specifications for general quantum circuits. Indeed,
DEFCIRCUIT is very limited in its expressiveness, only
performing argument and parameter substitution. It is
included mainly to help with the debugging and human
readability of Quil code. More advanced constructs are
intended to be written on top of Quil, as in Section V-B.
F. Measurement
Measurement provides the “side effects” of quantum
programming, and is an essential part of most practical quantum algorithms (e.g., phase estimation and
teleportation). Quil provides two forms of measurement:
measurement-for-effect, and measurement-for-record.
Measurement-for-effect is a measurement performed on
a single qubit used solely for changing the state of the
quantum system. This is done with a MEASURE instruction
of a single argument. Performing a measurement on Q5 is
written as
MEASURE 5
More useful, however, is measurement-for-record.
Measurement-for-record is a measurement performed
and recorded in classical memory. This form of the
MEASURE instruction takes two arguments, the qubit and
the classical memory address. To measure Q7 and deposit
the result at address 8 is written
MEASURE 7 [8]
The semantics of the measurement operation are described
in Section II-C.
Example 10. Producing a random number between 0 and
3 inclusive can be accomplished with the following program:
6 cis θ
:= cos θ + i sin θ = exp iθ
processing Quil is encouraged to warn or error on such
matrices.
7 Software
H 0
H 1
MEASURE 0 [0]
MEASURE 1 [1]
The memory segment [0-1] is now representative of the
number in binary.
G. Program Control
Program control is determined by the state of the
program counter. The program counter κ determines if the
program has halted, and if not, it determines the location
of the next instruction to be executed. Every instruction,
except for the following, has the effect of incrementing κ.
The exceptions are:
• Conditional and unconditional jumps.
• The halt instruction HALT which terminates execution
and assigns κ ← |P |.
• The last instruction in the program, which—after its
execution—implicitly terminates execution as if by
HALT.
Locations within the instruction sequence are denoted by
labels, which are names that are prepended with an ‘@’
symbol, like @start. The declaration of a new label within
the instruction sequence is called a jump target, and is
written with the LABEL directive.
Unconditional jumps are executed by the JUMP instruction which sets κ to the index of a given jump target.
Example 12. Consider the following two DEFCIRCUIT
declarations and their instantiations. Note the comments
on correct and incorrect usages of JUMP.
DEFCIRCUIT FOO:
LABEL @FOO_A
JUMP @GLOBAL
JUMP @FOO_A
JUMP @BAR_A
# valid, global label
# valid, local to FOO
# invalid
DEFCIRCUIT BAR:
LABEL @BAR_A
JUMP @FOO_A
# invalid
LABEL @GLOBAL
FOO
BAR
JUMP @FOO_A
JUMP @BAR_A
# invalid
# invalid
H. Zeroing the Quantum State
The quantum state of the QAM can be reset to the
zero state with the RESET instruction. This has the effect
of setting
Nq −1
O
|0iNq −k−1 .
|Ψi ←
k=0
Conditional jumps are executed by the JUMP-WHEN
(resp. JUMP-UNLESS) instruction, which set κ to the index
of a given jump target if the bit at a classical memory
address is 1 (resp. 0), and to κ + 1 otherwise. This is a
critical and differentiating element of Quil; it allows fast
classical feedback.
There are no provisions to clear the state of a single qubit,
but we can do this by taking advantage of projective
measurement.
Example 11. Consider the following C-like pseudocode of
an if-statement branching on the bit contained at address
x:
DEFCIRCUIT CLEAR q scratch_bit:
MEASURE q scratch_bit
JUMP-UNLESS @end scratch_bit
X q
LABEL @end
if (*x) {
// instrA...
} else {
// instrB...
}
This can be translated into Quil in the following way:
JUMP-WHEN @THEN [x]
# instrB...
JUMP @END
LABEL @THEN
# instrA...
LABEL @END
Lines starting with the # character are comments and are
ignored.
Labels that are declared within the body of a
DEFCIRCUIT are unique to that circuit. While it is possible
to jump out of a DEFCIRCUIT to a globally declared label,
it is not possible to jump inside of one.
Example 13. We can clear a qubit using a single bit of
classical scratch space.
I. Classical/Quantum Synchronization
Some classical/quantum programs can be constructed
in a way such that at a certain point of a quantum
program, execution must be suspended until some classical
computations are performed and corresponding classical
state is modified. This is accomplished using the WAIT
instruction, a synchronization primitive which signals to
the classical computer that computation will not continue
until some condition is satisfied. WAIT takes no arguments.
The mechanism by which WAIT works is deliberately
unspecified. Some example mechanisms include monitors
and interrupts, depending on the QAM implementation.
An example use of WAIT can be found in Section IV-B2.
J. Classical Instructions
Quil is intended to be a language to manipulate quantum state using quantum operations and classical control.
Classical computation on the classical state should be
done as much as possible with a classical computer, and
using Quil’s classical/quantum synchronization to mediate
the hand-off of classical data between the classical and
quantum processors. However, a few instructions for manipulating the classical state are provided for convenience,
with emphasis on making control flow easier.
1) Classical Unary Instructions: The classical unary
instructions are instructions that take a single classical
address as an argument and modify the bit at that address
accordingly. In each of the following, let a be the address
provided as the sole argument. The three instructions are:
Constant False FALSE, which has the effect C[a] ← 0;
Constant True TRUE, which has the effect C[a] ← 1; and
Negation NOT, which has the effect C[a] ← 1 − C[a].
2) Classical Binary Instructions: The classical binary
instructions are instructions that take two classical addresses as arguments, and modify the bits at those addresses accordingly. In all of the following, let a be the
first address and b be the second address provided as
arguments. The four instructions are:
Conjunction AND, which has the effect C[b] ← C[a]C[b];
Disjunction OR, which has the effect
C[b] ← 1 − (1 − C[a])(1 − C[b]);
Copy MOVE, which has the effect C[b] ← C[a]; and
Exchange EXCHANGE, which has the effect of swapping the
bits at a and b: C[a] ↔ C[b].
Example 14. Exclusive disjunction r ← a + b mod 2 can
be implemented with the following circuit:
DEFCIRCUIT XOR a b r:
# Uses (a | b) & (˜a | ˜b)
MOVE b r
OR a r
# r = a | b
JUMP-UNLESS @end r # short-circuit
MOVE b r
NOT a
NOT r
OR a r
# r = ˜a | ˜b
NOT a
# undo change to a
LABEL @end
Note that r has to be distinct from a and b.
K. The No-Operation Instruction
The no-operation, no-op, or NOP instruction does
not affect the state of the QAM except in the way described in Section III-G, i.e., by incrementing the program
counter. This instruction may appear useless, especially on
a machine with no notion of alignment or dynamic jumps.
However, it has purpose when the QAM is used as the
foundation for hardware emulation. For example, consider
a QAM with some associated gate noise model. If one
were to use an identity gate in place of a no-op, then the
identity gate would be interpreted as noisy while the no-op
would not. Moreover, the no-op has no qubit dependencies,
which would otherwise affect program analysis. Rigetti
Computing has used the no-op instruction as a way to
force a break in instruction parallelization, described in
Section V-E.
L. File Inclusion Semantics
File inclusion is done via the INCLUDE directive. For
example, the library of standard gates—described in Appendix A—can be included with
INCLUDE "stdgates.quil"
File inclusion is not simple token substitution as it is in
languages like the C preprocessor. Instead, the included
file is parsed into a set of circuit definitions, gate definitions, and instruction code. Instruction code is substituted
verbatim, but definitions will be recorded as if they were
originally placed at the top of the file.
Generally, best practice is to include files containing
only contain gate or circuit definitions (in which case the
file is called a library), or only executable code, and not
both. However, this is not enforced.
M. Pragma Support
Programs that process Quil code may want to take
advantage of extra information provided by the programmer. This is especially true when targeting QPUs where
additional information about the machine’s characteristics
affect how the program will be processed. Quil supports a
PRAGMA directive to include extra information in a program
which does not otherwise affect execution semantics. The
syntax of PRAGMA is as follows:
PRAGMA <identifier>+ <string>?
where + indicates one or more instances and ? indicates
zero or one instance.
Example 15. The QAM does not have any notion of
instruction parallelism, but programs are generally parallelized before execution on a QPU. (See Section V-E.)
Programs processing Quil may wish to enforce boundaries
across which parallelization may not occur. An implementation may opt to support a parallelization barrier pragma.
Despite the fact that the X-gates below are commuting, an
implementation may opt to treat the instructions sequentially.
X 0
PRAGMA parallelization_barrier
X 1
Note that this does not change the semantics of Quil vis-àvis the QAM.
Example 16. On modern superconducting qubit architectures, applications of different gates take different times.
This can affect how instructions get scheduled for execution.
Programs processing Quil may wish to allow the physical
time it takes to perform a gate to be defined with a pragma,
like so:
PRAGMA gate_time H "50 ns"
for i in range(1, n):
q = l + n - i
alpha = pi / 2 ** (n - i)
:: CPHASE(alpha) l q
:: H l
PRAGMA gate_time CNOT "150 ns"
H 0
CNOT 0 1
N. The Standard Gates
Quil provides a collection of standard one-, two-,
and three-qubit gates, which are fully defined in Appendix A. The standard gates can be used by including
stdgates.quil.
IV.
Quil Examples
A. Quantum Fourier Transform
In the context of the QAM’s quantum state |Ψi ∈ H ,
the quantum Fourier transform (QFT) [23, §5.1] is a
unitary operator F : H → H defined by the matrix
Fj,k := √
1
2Nq
0 ≤ j, k < dim H = 2Nq
ω jk ,
(12)
with ω := exp(2πi/2Nq ) the complex primitive root of
unity. It can be shown that this operator acts on the basis
vectors (up to permutation) via the map
Nq −1
O
|bk0 ik0 7→ √
k=0
k0 :=Nq −k−1
where φ(k) =
1
2Nq
k
Y
Nq −1
O
[|0ik + φ(k) |1ik ] ,
(13)
k=0
exp 2πibNq −j−1 /2j+1 .
j=0
Here, b is the bit string representation of the basis vectors,
as in Section II-A.
The first thing to notice is that the basis elements get
reversed in this factorization. This is easily fixed via a
product of SWAP gates. In the context of classical fast
Fourier transforms [27], this is called bit reversal.
The second and more
N important thing to notice is that
each factor of the
can be seen as an operation on
the qubit Qk . The factors of φ(k) indicate the operations
are two-qubit controlled-phase or CPHASE gates with Qk
as the target, and each previous qubit as the control. In
the degenerate one-qubit case, this is a Hadamard gate.
Further details on this algorithm can be found in [23].
The core QFT logic can be implemented as a straightforward recursive function QFT’. We write it as one which
transforms qubits Qk for l ≤ k < r. The base case is the
action on a single qubit—a Hadamard. In the general case,
we do a sequence of CPHASE gates targeted on the current
qubit Ql and controlled by all qubits before it, topped off
by a Hadamard. In the Python-like pseudocode below, we
generate Quil code for this algorithm. We prepend lines of
Quil code to be generated with two colons ‘::’.
def QFT’(l, r):
n = r - l
# num qubits
if n == 1:
# base case
:: H l
else:
# general case
QFT’(l + 1, r)
The bit reversal routine can be implemented straightforwardly as exactly bNq /2c SWAP gates.
def revbin(Nq):
for i in range(Nq / 2):
:: SWAP i (Nq - i - 1)
All of this is put together in a final QFT routine.
def QFT(Nq):
QFT’(0, Nq)
revbin(Nq)
B. Variational Quantum Eigensolver
The variational quantum eigensolver (VQE) [8, 9, 10]
is a classical/quantum algorithm for finding eigenvalues of
a Hamiltonian H variationally. The quantum subroutine
has two fundamental steps.
~
1) Prepare the quantum state |Ψ(θ)i,
often called the
ansatz.
~ | H | Ψ(θ)
~ i.
2) Measure the expectation value h Ψ(θ)
The classical portion of the computation is an optimization
loop.
1) Use a classical non-linear optimizer to minimize the
~
expectation value by varying ansatz parameters θ.
2) Iterate until convergence.
We refer to given references for details on the algorithm.
Practically, the quantum subroutine of VQE amounts
to preparing a state based off of a set of parameters θ~ and
performing a series of measurements in the appropriate
basis. Using the QAM, these measurements will end up
in classical memory. Doing this iteratively followed by a
small amount of postprocessing, one may compute a real
expectation value for the classical optimizer to use.
This algorithm can be implemented in at least two
distinct styles which impose different requirements on a
quantum computer. These two styles can serve as prototypical examples for programming a QAM with Quil.
1) Static Implementation: One simple implementation
of VQE is to generate a new Quil listing for every iteration
~ Before calling out to the quantum subroutine, a new
of θ.
Quil program is generated for the next desired θ~ and loaded
onto the quantum computer for immediate execution. For
a parameter θ = 0.00724 . . ., one such program might look
like
# State prep
...
CNOT 5 6
CNOT 6 7
RZ(0.00724195969993) 7
CNOT 6 7
...
# Measurements
MEASURE 0 [0]
MEASURE 1 [1]
...
This technique poses no issue from the perspective of
coherence time, but adds a time penalty to each iteration
of the classical optimizer.
A static implementation of VQE was written using
Rigetti Computing’s pyQuil library and the Rigetti QVM,
both of which are mentioned in Sections V-B and V-F
respectively.
2) Dynamic Implementation: Perhaps the most encapsulated implementation of VQE would be to use dynamic
parameters. Without loss of generality, let’s assume a
single parameter θ. We can define a circuit which takes
our single θ parameter and prepares |Ψi.
Fig. 2. Outline of Forest, Rigetti Computing’s quantum programming toolkit, described in Section V.
DEFCIRCUIT PREP_STATE(%theta):
...
H 3
H 5
...
CNOT 3 5
CNOT 5 6
RZ(%theta) 6
...
optimization and poses particular issues for current quantum computing architectures which have limited natural
gate sets and limited high-speed dynamic tune-up of new
gates or their approximations.
Next, we define a memory layout for classical/quantum
communication:
A. Overview
• [0]: Flag to indicate convergence completion.
• [1-64]: Dynamic parameter θ.
• [100], [101], . . . : Measurements corresponding to
Q0 , Q1 , . . . .
Finally, we can define our VQE circuit.
DEFCIRCUIT VQE:
LABEL @REDO
RESET
PREP_STATE([1-64]) # Dynamic Parameter
MEASURE 0 [100]
MEASURE 1 [101]
...
WAIT
JUMP-UNLESS @REDO [0]
This program has the advantage that the quantum
portion of the algorithm is completely encapsulated. It is
not necessary to dynamically reload Quil code for each
newly varied parameter.
The main disadvantage of this approach is its implementation difficulty in hardware. This is because of
the diminished potential for program analysis to occur
before execution. The actual gates that get applied will
not be known until the runtime of the algorithm (hence
the “dynamic” name). This may limit opportunities for
V.
Forest: A Quantum Programming Toolkit
Quantum computers, and specifically QAM-conforming
QPUs, are not yet widely available. Despite this, software
can make use of the the QAM and Quil to (a) study
practical quantum algorithmic performance in a spirit
similar to MMIX [28], (b) prepare a suite of software
which will be able to be run on QPUs, and (c) provide
a uniform interface to physical and virtualized QAMs.
Rigetti Computing has built a toolkit called Forest for
accomplishing these tasks.
The hierarchy of software roughly falls into four layers,
as in figure 2. Descending in abstraction, we have
Applications & Tools Software written to accomplish
real-world tasks (e.g., study electronic structure problems), and software to assist quantum programming.
Quil The language described in this document and associated software for processing Quil. It acts as an intermediate representation of general quantum programs.
Compiler Software to convert arbitrary Quil to simpler
Quil (or some other representation) for execution on
a QPU or QVM.
Execution Units A QPU, a QVM, or a hardware emulator. Software written at this level will typically incorporate noise models intrinsic to a particular QPU.
We will briefly describe each of these components of the
toolkit in the following sections.
B. Applications and Tools
Quil is an assembly-like language that is intended to
be both human readable and writable. However, more
expressive power comes from being able to manipulate
Quil programs programmatically. Rigetti Computing has
developed a Python library called pyQuil [29] which allows
the construction of Quil programs by treating them as firstclass objects. Using pyQuil along with scientific libraries
such SciPy [30], Rigetti Computing has implemented nontrivial algorithms such as VQE using the abstractions of
the QAM.
C. Quil Manipulation
Quil, as a language in its own right, is amenable
to processing and computation independent of any particular (real or virtual) machine for execution. Rigetti
Computing has written a reusable static analyzer and
parser application for Quil, that allows Quil to easily
be interchanged between programs. For example, Rigetti
Computing’s quil-json program converts Quil into a
structured JSON [31] format:
$ cat bell.quil
H 0
CNOT 0 1
$ quil-json bell.quil
{
"type": "parsed_program",
"executable_program": [
{
"type": "unresolved_application",
"operator": "H",
"arguments": [["qubit", 0]],
"parameters": null
},
{
"type": "unresolved_application",
"operator": "CNOT",
"arguments": [["qubit", 0],
["qubit", 1]],
"parameters": null
}
]
}
Note the two instances of “unresolved application”.
These were generated because of a simple static analysis
determining that these gates were not otherwise defined
in the Quil file. (This could be ameliorated by including
stdgates.quil.)
D. Compilation
In the context of quantum computation, compilation
is the process of converting a sequence of gates into an
approximately equivalent sequence of gates executable on
a quantum computer with a particular qubit topology. This
requires two separate kinds of processing: gate approximation and routing.
Since Quil is specified as a symbolic language, it is
amenable to symbolic analysis. Quil programs can be
decomposed into control flow graphs (CFGs) [32] whose
nodes are basic blocks of straight-line Quil instructions, as
is typical in compiler theory. Arrows between basic blocks
indicate transfers of control. For example, consider the
following Quil program:
LABEL @START
H 0
MEASURE 0 [0]
JUMP-WHEN @END [0]
H 0
H 1
CNOT 1 0
JUMP @START
LABEL @END
Y 0
MEASURE 0 [0]
MEASURE 1 [1]
Roughly speaking, each segment of straight-line control
makes up a basic block. Figure 3 depicts the control flow
graph of this program, with jump instructions elided.
*ENTRY-BLK-793*
*START795*
H 0
MEASURE 0 [0]
*BLK-796*
H 0
H 1
CNOT 1 0
*END798*
Y 0
MEASURE 0 [0]
MEASURE 1 [1]
*EXIT-BLK-794*
Fig. 3.
The control flow graph of a Quil program.
Many of these basic blocks will be composed of gates
and measurements, which themselves can be symbolically
and algebraically manipulated. Gates can go through
approximation to reduce a general set of gates to a
sequence of gates native to the particular architecture, and
then routing so that these simpler gates are arranged
to act on neighboring qubits in a physical architecture.
Another example of a transformation on basic blocks is
parallelization, talked about in the next section.
Both approximation and routing can be formalized
as transformations between QAMs. In this first example,
we show how we can formally describe a transformation
between a QAM to another one with a smaller but computationally equivalent gate set.
Example 17 (Compiling). Let M = (|Ψi , C, GM , G0 , P, κ)
be a one-qubit QAM with
GM = {H, Rx (θ), Rz (θ)}
for some fixed θ ∈ R, and let M0 be a QAM with
GM0 = {H, Rz (θ)}.
Because Rx = HRz H, we can define a compilation function
M 7→ M0 specifically transforming8 ι ∈ P according to
(H, Rz (θ), H) if ι = Rx (θ),
f (ι) =
(ι)
otherwise.
In general, parallelization cannot occur over jumps, resets,
waits, and measurements and dynamic gate applications
with overlapping address ranges. We suggest that NOP is
used as a way to force a parallelization break.
If a control flow graph is constructed as in Section V-D,
then parallelization can be done over each basic block. A
parallelized basic block is called a parallelization schedule. See Figure 4 for an example of quantum parallelization
within a CFG.
In this next example, we show how qubit connectivity
can be encoded in a QAM, and how one can route instructions to give the illusion of a fully connected topology.
Example 18 (Routing). Let M = (|Ψi , C, GM , G0 , P, κ)
be a three-qubit QAM with
GM = L(H) ∪ L(CNOT) ∪ L(SWAP),
where L was defined in Example 4. Consider a three-qubit
QPU with the qubits arranged in a line
Q0 —Q1 —Q2
so that two-qubit gates can only be applied on adjacent
qubits. Then this QPU can be modeled by another threequbit QAM M0 with the lifted gates
(
)
H0 , H1 , H2 ,
CNOT01 , CNOT10 , CNOT12 , CNOT21 , .
GM0 =
SWAP01 , SWAP12
Because of the qubit topology, there are no gates that act
on B2 ⊗ B0 . However, we can reason about transforming
between M and M0 in either direction. We can transform
from M to M0 by way of a transformation f similar to that
in the last example, namely
(SWAP01 , SWAP12 , SWAP01 ) if ι = SWAP02 ,
(SWAP01 , CNOT12 , SWAP01 ) if ι = CNOT02 ,
f (ι) =
(SWAP01 , CNOT21 , SWAP01 ) if ι = CNOT20 ,
(ι)
otherwise.
Similarly, we can transform from M0 to M by adding three
additional gates to GM0 , namely those implied by f .
Many other classes of useful transformations on QAMs
exist, such as G-preserving algebraic simplifications of P ,
an analog of peephole optimization in compiler theory [33].
E. Instruction Parallelism
Instruction parallelism, the ability to apply commuting operations to a quantum state in parallel, is one
of the many benefits of quantum computation. Quil code
as written is linear and serial, but can be interpreted
as an instruction-parallel program. In particular, many
subsequences of Quil instructions may be executed in
parallel. Such sequences include:
• Commuting gate applications and measurements, and
• Measurements with non-overlapping memory addresses.
8 In
the parlance of functional programming, f is applied to P via
a concatmap operation.
*ENTRY-BLK-799*
*START801*
{
H 0
}
{
MEASURE 0 [0]
}
*BLK-802*
{
H 0
H 1
}
{
CNOT 1 0
}
*END804*
{
Y 0
MEASURE 1 [1]
}
{
MEASURE 0 [0]
}
*EXIT-BLK-800*
Fig. 4. The parallelized version of Figure 3. Instruction sequences
which can be executed in parallel are surrounded in curly braces ‘{}’.
F. Rigetti Quantum Virtual Machine
Rigetti Computing has implemented a QVM in ANSI
Common Lisp [34] called the Rigetti QVM. It is a highperformance9 , programmable simulator for the QAM and
emulator for noisy quantum computers. The Rigetti QVM
exposes two interfaces for executing quantum programs:
execution of Quil files directly with POSIX-style shared
memory (“local execution”), and execution of Quil from
HTTP server requests (“remote execution”).
Local execution is useful for high-speed testing of
small-to-medium sized instances of classical/quantum algorithms. It also provides convenient ways of debugging
quantum programs, such as allowing direct inspection of
the quantum state at any point in the execution process, as
well as limited quantum hardware emulation with tunable
noise models.
Remote execution is used for distributed, cloud access
to QVM instances. HTTP libraries exist in nearly all
modern programming languages, and allow programs in
9 The Rigetti QVM has optimized vectorized and parallelized numerics, and has no theoretical limit for the number of qubits it can
simulate. It has been demonstrated to simulate 36 qubits.
these languages to make connections. Rigetti Computing
has built in to pyQuil the ability to send the first-class
Quil program objects to a local or secured remote Rigetti
QVM instance using the Forest API.
VI.
Acknowledgements
The authors would like to thank their colleagues at
Rigetti Computing for their support, especially Nick Rubin. We are also grateful to Erik Davis, Jarrod McClean,
Bill Millar, and Eric Peterson for their helpful discussions
and valuable suggestions provided throughout the development of this work.
Appendix
The following static and parametric gates are defined
in stdgates.quil. Many of these gates are standard gates
used in theoretical quantum computation [23], and some
of them find their origin in the theory of superconducting
qubits [35].
Pauli Gates
X = ( 01 10 )
0 −i
i 0
Y=
Z=
1 0
0 −1
Hadamard Gate
H=
√1
2
1 1
1 −1
Phase Gates
PHASE(θ) =
1 0
0 eiθ
S = PHASE(π/2)
T = PHASE(π/4)
Controlled-Phase Gates
CPHASE00(θ) = diag(eiθ , 1, 1, 1)
iθ
CPHASE01(θ) = diag(1, e , 1, 1)
CPHASE10(θ) = diag(1, 1, eiθ , 1)
CPHASE(θ) = diag(1, 1, 1, eiθ )
Cartesian Rotation Gates
cos θ −i sin θ
RX(θ) = −i sin2 θ cos θ 2
θ 2 θ 2
cos − sin
RY(θ) = sin θ2 cos θ2
2
2
−iθ/2
e
0
RZ(θ) =
iθ/2
0
e
Controlled-X Gates
1 0 0 0 0 0 0 0
1 0 0 0
CNOT =
0100
0001
0010
0
0 0
0 eiθ
0 0
0
eiθ
0
0
0
0
0
1
SWAP = PSWAP(0)
ISWAP = PSWAP(π/2)
1 0 0 0 0 0 0 0
01000000
0 0 1 0 0 0 0 0
CSWAP = 00 00 00 10 01 00 00 00
00000010
00000100
00000001
B. Prior Work
There exists much literature on abstract models, syntaxes, and semantics of quantum computing. Most of
them are in the form of quantum programming languages
(QPLs) and simulators, which achieve various levels of expressiveness. Languages roughly fall into three categories:
• embedded domain-specific languages,
• high-level quantum programming languages, and
• low-level quantum intermediate representations.
In addition, work has been done on designing larger tool
chains for quantum programming [18, 19, 36].
A. The Standard Gate Set
I = ( 10 01 )
1
PSWAP(θ) =
Conclusion
We have introduced a practical abstract machine for
reasoning about and executing quantum programs. Furthermore, we have described a notation for such programs
on this machine, which is amenable to analysis, compilation, and execution. Finally, we have described a pragmatic
toolkit for quantum program construction and execution
built atop these ideas.
VII.
Swap Gates
01000000
0 0 1 0 0 0 0 0
CCNOT = 00 00 00 10 01 00 00 00
00000100
00000001
00000010
In the following subsections, we provide a nonexhaustive account of some previous work within the above
categories, and describe how they relate to the design of
our quantum ISA.
1) Embedded Domain-Specific Languages: An embedded domain-specific language (EDSL) is a language
to express concepts and problems from within another
programming language. Within the context of quantum
programming, two prominent examples are Quipper [16],
which is embedded in Haskell [37], and LIQU i|i [15],
which is embedded in F# [38]. Representation of quantum
programs (and, in particular, the subclass of quantum
programs called quantum circuits) is expressed with data
structures in the host language. Feedback of classical
information is achieved through an interface with the host
language.
Quantum programs written in an EDSL are not directly
executable on a quantum computer, due to the requirement of being present in the host language’s runtime.
However, since quantum programs are represented as firstclass objects, these objects are amenable to processing and
compilation to a quantum computer’s natively executable
instruction set. Quil is one such target for compilation.
2) High-Level QPLs: High-level QPLs are the quantum
analog of languages like C [39] in the imperative case or
ML [40] in the functional case. They provide a plethora of
classical and quantum data types, as well as control flow
constructs. In the case of functional quantum languages,
the lambda calculus for quantum computation with classical control by Selinger and Valiron can act as a theoretical
basis for their semantics of such a language10 [42].
10 This is similar to how the classical untyped lambda calculus
formed the basis for LISP in 1958 [41].
One prominent example of a high-level QPL is Bernhard Ömer’s QCL [43]. It is a C-like language with classical control and quantum data. Among the many that
exist, two important—and indeed somewhat dual—data
types are int and qureg. The following example shows
a Hadamard initialization on eight qubits, and measuring
the first four of those qubits into an integer variable.
// Allocate classical and quantum data
qureg q[8];
int m;
// Hadamard initialize
H(q);
// Measure four qubits into m
measure q[0..3], m;
// Print m, which will be anywhere from 0 to 15.
print m;
Ömer defines the semantics of this language in detail in
his PhD thesis, and presents an implementation of a QCL
interpreter written in C++ [44]. However, a compilation or
execution strategy on quantum hardware is not presented.
Similar to EDSLs, QPLs such as QCL can be compiled
into a lower-level quantum instruction set such as Quil.
3) Low-Level Quantum Intermediate Representations:
In the context of compiler theory, an intermediate representation (IR) is some representation of a language
which is amenable to further processing. IR can be higher
level (as with abstract syntax trees), or lower level (as
with linear bytecodes). Low-level IRs often act as compilation targets for classical programming languages, and are
used for analysis and program optimization. For example,
LLVM IR [45] is used as an intermediate compilation target
for the Clang C compiler [46].
The most well-known example of a low-level quantum
IR is QASM [47]. This was originally a language to describe
the quantum circuits for LATEX output in [23], and hence,
not an IR in that form. However, the syntax was adapted
for executable use in [48]. QASM, however, does not have
any notion of classical control within the language, acting
solely as a quantum circuit description language.
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
Quil is considered a low-level quantum IR with classical
control.
[12]
References
[1] P. J. J. O’Malley, R. Babbush, I. D. Kivlichan,
J. Romero, J. R. McClean, R. Barends, J. Kelly,
P. Roushan, A. Tranter, N. Ding, B. Campbell,
Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth,
A. G. Fowler, E. Jeffrey, E. Lucero, A. Megrant,
J. Y. Mutus, M. Neeley, C. Neill, C. Quintana,
D. Sank, A. Vainsencher, J. Wenner, T. C. White,
P. V. Coveney, P. J. Love, H. Neven, A. AspuruGuzik, and J. M. Martinis, “Scalable quantum
simulation of molecular energies,” Phys. Rev. X,
vol. 6, p. 031007, Jul 2016. [Online]. Available:
http://link.aps.org/doi/10.1103/PhysRevX.6.031007
[2] M. R. Geller, J. M. Martinis, A. T. Sornborger, P. C.
Stancil, E. J. Pritchett, H. You, and A. Galiautdi-
[13]
[14]
[15]
[16]
nov, “Universal quantum simulation with prethreshold superconducting qubits: Single-excitation subspace method,” Physical Review A, vol. 91, no. 6, p.
062309, 2015.
R. Barends, A. Shabani, L. Lamata, J. Kelly, A. Mezzacapo, U. Las Heras, R. Babbush, A. Fowler,
B. Campbell, Y. Chen et al., “Digitized adiabatic
quantum computing with a superconducting circuit,”
Nature, vol. 534, no. 7606, pp. 222–226, 2016.
D. Riste, M. P. da Silva, C. A. Ryan, A. W. Cross,
J. A. Smolin, J. M. Gambetta, J. M. Chow, and B. R.
Johnson, “Demonstration of quantum advantage in
machine learning,” arXiv preprint arXiv:1512.06069,
2015.
J. M. Chow, S. J. Srinivasan, E. Magesan, A. Córcoles,
D. W. Abraham, J. M. Gambetta, and M. Steffen,
“Characterizing a four-qubit planar lattice for arbitrary error detection,” in SPIE Sensing Technology+
Applications. International Society for Optics and
Photonics, 2015, pp. 95 001G–95 001G.
J. Kelly, R. Barends, A. Fowler, A. Megrant, E. Jeffrey, T. White, D. Sank, J. Mutus, B. Campbell,
Y. Chen et al., “State preservation by repetitive error detection in a superconducting quantum circuit,”
Nature, vol. 519, no. 7541, pp. 66–69, 2015.
D. Ristè, S. Poletto, M.-Z. Huang, A. Bruno,
V. Vesterinen, O.-P. Saira, and L. DiCarlo, “Detecting bit-flip errors in a logical qubit using stabilizer
measurements,” Nature communications, vol. 6, 2015.
A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung,
X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L.
O’Brien, “A variational eigenvalue solver on a photonic quantum processor,” Nature communications,
vol. 5, 2014.
D. Wecker, M. B. Hastings, and M. Troyer, “Progress
towards practical quantum variational algorithms,”
Physical Review A, vol. 92, no. 4, p. 042303, 2015.
J. R. McClean, J. Romero, R. Babbush, and
A. Aspuru-Guzik, “The theory of variational hybrid quantum-classical algorithms,” New Journal of
Physics, vol. 18, no. 2, p. 023023, 2016.
B. Bauer, D. Wecker, A. J. Millis, M. B. Hastings, and M. Troyer, “Hybrid quantum-classical
approach to correlated materials,” arXiv preprint
arXiv:1510.03859, 2015.
E. Farhi, J. Goldstone, and S. Gutmann, “A quantum
approximate optimization algorithm,” arXiv preprint
arXiv:1411.4028, 2014.
M. Reiher, N. Wiebe, K. M. Svore, D. Wecker, and
M. Troyer, “Elucidating reaction mechanisms
on
quantum
computers,”
arXiv
preprint
arXiv:1605.03590, 2016.
D. Deutsch, “Quantum Theory, the Church–Turing
Principle and the Universal Quantum Computer,”
Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, vol. 400,
no. 1818, pp. 97–117, 1985.
D. Wecker and K. M. Svore, “LIQUi|i: A Software
Design Architecture and Domain-Specific Language
for Quantum Computing,” 2014. [Online]. Available:
arXiv:1402.4467v1
A. S. Green, P. L. Lumsdaine, N. J. Ross, P. Selinger,
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
and B. Valiron, “Quipper: A scalable quantum
programming language,” SIGPLAN Not., vol. 48,
no. 6, pp. 333–342, Jun. 2013. [Online]. Available:
http://doi.acm.org/10.1145/2499370.2462177
T. Häner, D. S. Steiger, K. Svore, and M. Troyer,
“A software methodology for compiling quantum programs,” arXiv preprint arXiv:1604.01401, 2016.
K. M. Svore, A. V. Aho, A. W. Cross, I. Chuang,
and I. L. Markov, “A layered software architecture for
quantum computing design tools,” IEEE Computer,
vol. 39, no. 1, pp. 74–83, 2006.
A. JavadiAbhari, S. Patil, D. Kudrow, J. Heckey,
A. Lvov, F. T. Chong, and M. Martonosi, “Scaffcc:
Scalable compilation and analysis of quantum programs,” Parallel Computing, vol. 45, pp. 2–17, 2015.
T. Häner, D. S. Steiger, M. Smelyanskiy, and
M. Troyer, “High performance emulation of quantum
circuits,” arXiv preprint arXiv:1604.06460, 2016.
M. Smelyanskiy, N. P. Sawaya, and A. AspuruGuzik, “qhipster: the quantum high performance
software testing environment,” arXiv preprint
arXiv:1601.07195, 2016.
A. Barenco, C. H. Bennett, R. Cleve, D. P. DiVincenzo, N. Margolus, P. Shor, T. Sleator, J. A. Smolin,
and H. Weinfurter, “Elementary gates for quantum
computation,” Physical review A, vol. 52, no. 5, p.
3457, 1995.
M. A. Nielsen and I. L. Chuang, Quantum computation and quantum information. Cambridge University
Press, 2010.
D. Deutsch, A. Barenco, and A. Ekert, “Universality
in quantum computation,” in Proceedings of the Royal
Society of London A: Mathematical, Physical and
Engineering Sciences, vol. 449, no. 1937. The Royal
Society, 1995, pp. 669–677.
S. Aaronson, Quantum computing since Democritus.
Cambridge University Press, 2013.
IEEE standard for binary floating-point arithmetic,
New York, 1985, note: Standard 754–1985.
J. W. Cooley and J. W. Tukey, “An algorithm for
the machine calculation of complex fourier series,”
Mathematics of computation, vol. 19, no. 90, pp. 297–
301, 1965.
D. E. Knuth, The Art of Computer Programming,
Volume 1, Fascicle 1: MMIX – A RISC Computer for
the New Millennium (Art of Computer Programming).
Addison-Wesley Professional, 2005.
Rigetti Computing, “pyQuil,” 2016, accessed:
2016-08-10. [Online]. Available: https://github.com/
rigetticomputing/pyQuil
E. Jones, T. Oliphant, P. Peterson et al., “SciPy:
Open source scientific tools for Python,” 2001–
2016, accessed: 2016-06-21. [Online]. Available: http:
//www.scipy.org/
The
JSON
Data
Interchange
Format,
1st
ed.,
October
2013.
[Online]. Available: http://www.ecma-international.org/
publications/files/ECMA-ST/ECMA-404.pdf
F. E. Allen, “Control flow analysis,” SIGPLAN Not.,
vol. 5, no. 7, pp. 1–19, Jul. 1970. [Online]. Available:
http://doi.acm.org/10.1145/390013.808479
W. M. McKeeman, “Peephole optimization,”
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
Commun. ACM, vol. 8, no. 7, pp. 443–444,
Jul. 1965. [Online]. Available: http://doi.acm.org/
10.1145/364995.365000
American National Standards Institute and Information Technology Industry Council, American National
Standard for Information Technology: programming
language — Common LISP, American National Standards Institute, 1430 Broadway, New York, NY 10018,
USA, 1996, approved December 8, 1994.
J. M. Chow, “Quantum information processing with
superconducting qubits,” Ph.D. dissertation, Yale
University, 2010.
A. J. Abhari, A. Faruque, M. J. Dousti, L. Svec,
O. Catu, A. Chakrabati, C.-F. Chiang, S. Vanderwilt,
J. Black, and F. Chong, “Scaffold: Quantum programming language,” DTIC Document, Tech. Rep., 2012.
S. Marlow, “Haskell 2010 Language Report,” https://
www.haskell.org/definition/haskell2010.pdf, 2010, accessed: 2016-06-21.
D. Syme et al., “The F# 4.0 Language Specification,” http://fsharp.org/specs/language-spec/4.0/
FSharpSpec-4.0-latest.pdf, 2016, accessed: 2016-0621.
B. W. Kernighan, The C Programming Language,
2nd ed., D. M. Ritchie, Ed. Prentice Hall Professional
Technical Reference, 1988.
R. Milner, M. Tofte, and D. Macqueen, The Definition
of Standard ML. Cambridge, MA, USA: MIT Press,
1997.
J. McCarthy, “Recursive functions of symbolic expressions and their computation by machine, part i,”
Communications of the ACM, vol. 3, no. 4, pp. 184–
195, 1960.
P. Selinger and B. Valiron, “A lambda calculus for
quantum computation with classical control,” Mathematical Structures in Computer Science, vol. 16,
no. 03, pp. 527–552, 2006.
B. Ömer, “Structured quantum programming,” Ph.D.
dissertation, Technical University of Vienna, 2003.
[Online]. Available: http://tph.tuwien.ac.at/∼oemer/
doc/structquprog.pdf
——, “QCL - A Programming Language for Quantum Computers,” http://tph.tuwien.ac.at/∼oemer/
qcl.html, 2014, accessed: 2016-06-30.
C. Lattner and V. Adve, “LLVM: A Compilation
Framework for Lifelong Program Analysis & Transformation,” in Proceedings of the 2004 International
Symposium on Code Generation and Optimization
(CGO’04), Palo Alto, California, Mar 2004.
Clang developers, “clang: a C language family frontend for LLVM,” http://clang.llvm.org/, accessed:
2016-06-30.
MIT Quanta Group, “Quantum circuit viewer:
qasm2circ,”
http://www.media.mit.edu/quanta/
qasm2circ/, accessed: 2016-06-30.
——, “qasm-tools,” http://www.media.mit.edu/qua
nta/quanta-web/projects/qasm-tools/,
accessed:
2016-06-30.
| 6cs.PL
|
HEIGHTS OF VARIETIES IN MULTIPROJECTIVE SPACES
AND ARITHMETIC NULLSTELLENSÄTZE
arXiv:1103.4561v2 [math.AG] 20 Oct 2012
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Abstract. We present bounds for the degree and the height of the polynomials
arising in some problems in effective algebraic geometry including the implicitization
of rational maps and the effective Nullstellensatz over a variety. Our treatment is
based on arithmetic intersection theory in products of projective spaces and extends
to the arithmetic setting constructions and results due to Jelonek. A key role is
played by the notion of canonical mixed height of a multiprojective. We study
this notion from the point of view of resultant theory and establish some of its
basic properties, including its behavior with respect to intersections, projections
and products. We obtain analogous results for the function field case, including a
parametric Nullstellensatz.
Contents
Introduction
1. Degrees and resultants of multiprojective cycles
1.1. Preliminaries on multiprojective geometry
1.2. Mixed degrees
1.3. Eliminants and resultants
1.4. Operations on resultants
2. Heights of cycles of multiprojective spaces
2.1. Mixed heights of cycles over function fields
2.2. Measures of complex polynomials
2.3. Canonical mixed heights of cycles over Q
3. The height of the implicit equation
3.1. The function field case
3.2. The rational case
4. Arithmetic Nullstellensätze
4.1. An effective approach to the Nullstellensatz
4.2. Parametric Nullstellensätze
4.3. Nullstellensätze over Z
References
2
7
7
10
16
21
24
24
36
38
50
50
56
61
61
63
68
72
Date: October 23, 2012.
2010 Mathematics Subject Classification. Primary 11G50; Secondary 14Q20,13P15.
Key words and phrases. Multiprojective spaces, mixed heights, resultants, implicitization, arithmetic Nullstellensatz.
D’Andrea was partially supported by the research project MTM2010-20279 (Spain). Krick was
partially supported by the research projects ANPCyT 3671/05, UBACyT X-113 2008-2010 and CONICET PIP 2010-2012 (Argentina). Sombra was partially supported by the research project MTM200914163-C02-01 (Spain) and by a MinCyT Milstein fellowship (Argentina).
1
2
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Introduction
In 1983, Serge Lang wrote in the preface to his book [Lan83]:
It is legitimate, and to many people an interesting point of view, to
ask that the theorems of algebraic geometry from the Hilbert Nullstellensatz to the more advanced results should carry with them estimates
on the coefficients occurring in these theorems. Although some of the
estimates are routine, serious and interesting problems arise in this
context.
Indeed, the main purpose of the present text is to give bounds for the degree and the
size of the coefficients of the polynomials in the Nullstellensatz.
Let f1 , . . . , fs ∈ Z[x1 , . . . , xn ] be polynomials without common zeros in the affine space
An (Q). The Nullstellensatz says then that there exist α ∈ Z \ {0} and g1 , . . . , gs ∈
Z[x1 , . . . , xn ] satisfying a Bézout identity
α = g1 f1 + · · · + gs fs .
As for many central results in commutative algebra and in algebraic geometry, it is a
non-effective statement. By the end of the 1980s, the estimation of the degree and the
height of polynomials satisfying such an identity became a widely considered question
in connection with problems in computer algebra and Diophantine approximation.
The results in this direction are generically known as arithmetic Nullstellensätze and
they play an important role in number theory and in theoretical computer science.
In particular, they apply to problems in complexity and computability [Koi96, Asc04,
DKS10], to counting problems over finite fields or over the rationals [BBK09, Rem10],
and to effectivity in existence results in arithmetic geometry [KT08, BS10].
The first non-trivial result on this problem was obtained by Philippon, who got a
bound on the minimal size of the denominator α in a Bézout identity as above [Phi90].
Berenstein and Yger achieved the next big progress, producing height estimates for the
polynomials gi ’s with techniques from complex analysis (integral formulae for residues
of currents) [BY91]. Later on, Krick, Pardo and Sombra [KPS01] exhibited sharp
bounds by combining arithmetic intersection theory with the algebraic approach in
[KP96] based on duality theory for Gorenstein algebras. Recall that the height of
a polynomial f ∈ Z[x1 , . . . , xn ], denoted by h(f ), is defined as the logarithm of the
maximum of the absolute value of its coefficients. Then, Theorem 1 in [KPS01] reads
as follows: if d = maxj deg(fj ) and h = maxj h(fj ), there is a Bézout identity as above
satisfying
deg(gi ) ≤ 4 n dn , h(α), h(gi ) ≤ 4 n (n + 1) dn h + log s + (n + 7) log(n + 1) d .
We refer the reader to the surveys [Tei90, Bro01] for further information on the history
of the effective Nullstellensatz, main results and open questions.
One of the main results of this text is the arithmetic Nullstellensatz over a variety
below, which is a particular case of Theorem 4.28. For an affine equidimensional
h(V ) the degree and the canonical
variety V ⊂ An (Q), we denote by deg(V ) and by b
height of the closure of V with respect to the standard inclusion An ֒→ Pn . The degree
and the height of a variety are measures of its geometric and arithmetic complexity,
see §2.3 and the references therein for details. We say that a polynomial relation holds
on a variety if it holds for every point in it.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
3
Theorem 0.1. Let V ⊂ An (Q) be a variety defined over Q of pure dimension r and
f1 , . . . , fs ∈ Z[x1 , . . . , xn ] \ Z a family of s ≤ r + 1 polynomials without common zeros
in V . Set dj = deg(fj ) and hj = h(fj ) for 1 ≤ j ≤ s. Then there exist α ∈ Z \ {0}
and g1 , . . . , gs ∈ Z[x1 , . . . , xn ] such that
α = g1 f1 + · · · + gs fs
on V
with
s
Y
dj deg(V ),
• deg gi fi ≤
j=1
• h(α), h(gi ) + h(fi ) ≤
s
Y
j=1
dj
s
X
hℓ
b
h(V ) + deg(V )
+ (4r + 8) log(n + 3) .
dℓ
ℓ=1
For V = An , this result gives the bounds
s
Y
dj
deg gi fi ≤
j=1
,
h(α), h(gi )+ h(fi ) ≤
s Y
X
ℓ=1
j6=ℓ
s
Y
dj .
dj hℓ + (4n + 8) log(n + 3)
j=1
These bounds are substantially sharper than the previously known. Moreover, they
are close to optimal in many situations. For instance, let d1 , . . . , dn+1 , H ≥ 1 and set
n
f1 = x1 − H, f2 = x2 − xd12 , . . . , fn = xn − xdn−1
, fn+1 = xdnn+1 .
This is a system of polynomials without common zeros. Hence, the above result
implies that there is a Bézout identity α = g1 f1 + · · · + gn+1 fn+1 which satisfies
h(α) ≤ d2 · · · dn+1 (log(H) + (4n + 8) log(n + 3)). On the other hand, specializing any
such identity at the point (H, H d2 , . . . , H d2 ···dn ), we get
α = gn+1 (H, H d2 , . . . , H d2 ···dn )H d2 ···dn+1 .
This implies the lower bound h(α) ≥ d2 · · · dn+1 log(H) and shows that the height
bound in Theorem 0.1 is sharp in this case. More examples can be found in §4.3.
It is important to mention that all previous results in the literature are limited to the
case when V is a complete intersection and cannot properly distinguish the influence of
each individual fj , due to the limitations of the methods applied. Hence, Theorem 0.1
is a big progress as it holds for an arbitrary variety and gives bounds depending on
the degree and height of each fj . This last point is more important than it might seem
at first. As it is well-known, by using Rabinowicz’ trick one can show that the weak
Nullstellensatz implies its strong version. However, this reduction yields good bounds
for the strong Nullstellensatz only if the corresponding weak version can correctly
differentiate the influence of each fj , see Remark 4.27. Using this observation, we
obtain in §4.3 the following arithmetic version of the strong Nullstellensatz over a
variety.
Theorem 0.2. Let V ⊂ An (Q) be a variety defined over Q of pure dimension r and
g, f1 , . . . , fs ∈ Z[x1 , . . . , xn ] such that g vanishes on the common zeros of f1 , . . . , fs in
V . Set dj = deg(fj ) and h = maxj h(fj ) for 1 ≤ j ≤ s. Assume that d1 ≥ · · · ≥ ds ≥ 1
Qmin{s,r+1}
and set D = j=1
dj . Set also d0 = max{1, deg(g)} and h0 = h(g). Then there
exist µ ∈ N, α ∈ Z \ {0} and g1 , . . . , gs ∈ Z[x1 , . . . , xn ] such that
α gµ = g1 f1 + · · · + gs fs
on V
4
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
with
• µ ≤ 2D deg(V ),
• deg(gi fi ) ≤ 4d0 D deg(V ),
min{s,r+1}
3h
X
h
0
b
+
+ c(n, r, s) ,
• h(α), h(gi ) + h(fi ) ≤ 2d0 D h(V ) + deg(V )
2d0
dℓ
ℓ=1
where c(n, r, s) ≤ (6r + 17) log(n + 4) + 3(r + 1) log(max{1, s − r}).
Our treatment of this problem is the arithmetic counterpart of Jelonek’s approach to
produce bounds for the degrees in the Nullstellensatz over a variety [Jel05]. To this
end, we develop a number of tools in arithmetic intersection and elimination theory
in products of projective spaces. A key role is played by the notion of canonical
mixed heights of multiprojective varieties, which we study from the point of view of
resultants. Our presentation of mixed resultants of cycles in multiprojective spaces
is mostly a reformulation of the theory developed by Rémond in [Rem01a, Rem01b]
as an extension of Philippon’s theory of eliminants of homogeneous ideals [Phi86].
We also establish new properties of them, including their behavior under projections
(Proposition 1.41) and products (Proposition 1.45).
Let n = (n1 , . . . , nm ) ∈ Nm and set Pn = Pn1 (Q) × . . . × Pnm (Q) for the corresponding
multiprojective space. For a cycle X of Pn of pure dimension r and a multi-index c =
(c1 , . . . , cm ) ∈ Nm of length r + 1, the mixed Fubini-Study height hc (X) is defined as
an alternative Mahler measure of the corresponding mixed resultant (Definition 2.40).
The canonical mixed height is then defined by a limit process as
b
hc (X) := lim ℓ−r−1 hc ([ℓ]∗ X),
ℓ→∞
where [ℓ] denotes the ℓ-power map of Pn (Proposition-Definition 2.45).
To handle mixed degrees and heights, we introduce a notion of extended Chow ring
of Pn (Definition 2.50). It is an arithmetic analogue of the Chow ring of Pn and
nm +1 ). We
can be identified with the quotient ring R[η, θ1 , . . . , θm ]/(η 2 , θ1n1 +1 , . . . , θm
associate to the cycle X an element in this ring, denoted [X]Z , corresponding under
this identification to
X
X
b
hc (X) η θ n1−c1 · · · θ nm −cm +
deg (X) θ n1 −b1 · · · θ nm −bm ,
1
c
m
b
1
m
b
the sums being indexed by all b, c ∈ Nm of respective lengths r and r + 1 such that
b, c ≤ n. Here, degb (X) denotes the mixed degree of X of index b. This element
contains the information of all non-trivial mixed degrees and canonical mixed heights
hc (X) are zero for any other b and c.
of X, since degb (X) and b
The extended Chow ring of Pn turns out to be a quite useful object which allows to
translate geometric operations on multiprojective cycles into algebraic operations on
rings and classes. In particular, we obtain the following multiprojective arithmetic
Bézout’s inequality, see also Theorem 2.58. For a multihomogeneous polynomial f ∈
Z[x1 , . . . , xm ], where xi is a group of ni + 1 variables, we denote by ||f ||sup its supnorm (Definition 2.29) and consider
the element [f ]sup in the extended Chow ring
Pm
corresponding to the element i=1 degxi (f )θi + log ||f ||sup η.
Theorem 0.3. Let X be an effective equidimensional cycle of Pn defined over Q and
f ∈ Z[x1 , . . . , xm ] a multihomogeneous polynomial such that X and div(f ) intersect
properly. Then
[X · div(f )]Z ≤ [X]Z · [f ]sup .
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
5
Statements on classes in the extended Chow ring can easily be translated into statements on mixed degrees and heights. In this direction, the above result implies that,
for any b ∈ Nm of length equal to dim(X),
b
hb (X · div(f )) ≤
m
X
i=1
degxi (f )b
hb+ei (X) + log ||f ||sup degb (X)
where ei denotes the i-th vector of the standard basis of Rm . In a similar way, we also
study the behavior of arithmetic classes (and a fortiori, of canonical mixed heights)
under projections (Proposition 2.64) and products (Proposition 2.66), among other
results.
Jelonek’s approach consists in producing a Bézout identity from an implicit equation of
a specific regular map. In general, the implicitization problem consists in computing
equations for an algebraic variety W from a given rational parameterization of it.
The typical case is when W is a hypersurface: the variety is then defined by a single
equation and the problem consists in computing this “implicit equation”. We consider
here the problem of estimating the height of the implicit equation of a hypersurface
parameterized by a regular map V → W whose domain is an affine variety V , in
terms of the degree and the height of V and of the polynomials defining the map.
To this end, we prove the following arithmetic version of Perron’s theorem over a
variety [Jel05, Thm. 3.3]. It is obtained as a consequence of Theorem 3.15.
Theorem 0.4. Let V ⊂ An (Q) be a variety defined over Q of pure dimension r. Let
q1 , . . . , qr+1 ∈ Z[x1 , . . . , xn ] \ Z such that the closure of the image of the map
V −→ Ar+1 (Q) , x 7−→ (q1 (x), . . . , qr+1 (x))
P
a ∈ Z[y , . . . , y
is a hypersurface. Let E =
1
r+1 ] be a primitive and
a∈Nr+1 αa y
squarefree polynomial defining this hypersurface. Set dj = deg(qj ), hj = h(qj ) for
1 ≤ j ≤ r + 1. Then, for all a = (a1 , . . . , ar+1 ) such that αa 6= 0,
•
r+1
X
i=1
ai di ≤
• h(αa ) +
r+1
Y
r+1
X
i=1
j=1
dj deg(V ),
ai hi ≤
r+1
Y
dj
j=1
r+1
X
hℓ
b
h(V ) + deg(V )
+ (r + 2) log(n + 3) .
dℓ
ℓ=1
For V = An we have r = n, deg(V ) = 1 and b
h(V ) = 0. Hence, the above result extends
the classical Perron’s theorem [Per27,
Satz
57],
P
Q which amounts to the weighted degree
bound for the implicit equation i ai di ≤ j dj , by adding the bound for the height
h(αa ) +
n+1
X
i=1
ai hi ≤
n+1
XY
ℓ=1
j6=ℓ
n+1
Y
dj .
dj hℓ + (n + 2) log(n + 3)
j=1
Our results on the implicitization problem as well as those on mixed resultants and
multiprojective arithmetic intersection theory should be of independent interest, besides of their applications to the arithmetic Nullstellensatz.
The method is not exclusive of Z but can also be carried over to other rings equipped
with a suitable height function. In this direction, we apply it to k[t1 , . . . , tp ], the ring
of polynomials over an arbitrary field k in p variables: if we set t = {t1 , . . . , tp }, the
height of a polynomial with coefficients in k[t] is its degree in the variables t. For
this case, we also develop the corresponding arithmetic intersection theory, including
6
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
the behavior of classes in the extended Chow ring with respect to intersections (Theorem 2.18), projections (Proposition 2.22), products (Proposition 2.23) and ruled joins
(Proposition 2.24). As a consequence, we obtain a parametric analogue of Perron’s
theorem (Theorem 3.1) and then the parametric Nullstellensatz below, which is a particular case of Theorem 4.11. For an affine equidimensional variety V ⊂ An (k(t)), we
denote by h(V ) the t-degree of the Chow form of its closure in Pn (k(t)), see §2.1 for
details.
Theorem 0.5. Let V ⊂ An (k(t)) be a variety defined over k(t) of pure dimension r
and f1 , . . . , fs ∈ k[t][x1 , . . . , xn ]\k[t] a family of s ≤ r+1 polynomials without common
zeros in V . Set dj = degx (fj ) and hj = degt (fj ) for 1 ≤ j ≤ s. Then there exist
α ∈ k[t] \ {0} and g1 , . . . , gs ∈ k[t][x1 , . . . , xn ] such that
α = g1 f1 + · · · + gs fs
with
• degx (gi fi ) ≤
s
Y
j=1
on V
dj deg(V ),
• deg(α), deg t (gi fi ) ≤
s
Y
j=1
dj
s
X
hℓ
.
h(V ) + deg(V )
dℓ
ℓ=1
For V = An (k(t)) we have r = n, deg(V ) = 1 and h(V ) = 0. Hence, this result gives
the following bounds for the partial degrees of the polynomials in a Bézout identity:
s Y
s
X
Y
dj , deg(α), deg t (gi fi ) ≤
degx (gi fi ) ≤
dj hℓ .
j=1
ℓ=1
j6=ℓ
In Theorem 4.22, we give a strong version of the parametric Nullstellensatz over a
variety, which also contains the case of an arbitrary number of input polynomials. Up
to our knowledge, the only previous results on the parametric Nullstellensatz are due
to Smietanski [Smi93], who considers the case when the number of parameters p is at
most two and V = An (k(t)), see Remark 4.21.
To prove both the arithmetic and parametric versions of the effective Nullstellensatz,
we need to consider a more general version of these statements where the input polynomials depend on groups of parameters, see Theorem 4.28. The latter has further
interesting applications. For instance, consider the family F1 , . . . , Fn+1 of general
n-variate polynomials of degree d1 , . . . , dn+1 , respectively. For each j, write
X
Fj =
uj,a xa
a
where each uj,a is a variable. Let uj = {uj,a }a be the group of variables corresponding
to the coefficients of Fj and set u = {u1 , . . . , un+1 }. The corresponding Macaulay
resultant R ∈ Z[u] lies in the ideal (F1 , . . . , Fn+1 ) ⊂ Q[u, x] and Theorem 4.28 gives
bounds for a representation of R in this ideal. Indeed, we obtain that there are
λ ∈ Z \ {0} and gj ∈ Z[u, x] such that λ R = g1 F1 + · · · + gn+1 Fn+1 with
deguj (gi Fi ) ≤
Y
ℓ6=j
dℓ
,
n+1
Y
h(λR), h(gi ) ≤ (6n + 10) log(n + 3)
dℓ ,
ℓ=1
see examples 4.18 and 4.37. The obtained bound for the height of the gi ’s is of the
same order as the sharpest known bounds for the height of R [Som04].
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
7
This text is divided in four sections. In the first one, we recall the basic properties
of mixed resultants and degrees of cycles in multiprojective spaces over an arbitrary
field K. The second section focuses on the mixed heights of cycles for the case when K
is a function field and on the canonical mixed heights of cycles for K = Q. In the third
section, we apply this machinery to the study of the height of the implicit equation,
including generalizations and variants of Theorem 0.4. We conclude in the fourth
section by deriving the different arithmetic Nullstellensätze.
Acknowledgments. We thank José Ignacio Burgos for the many discussions we had
and, in particular, for the statement and the proof of Lemma 1.18. We thank Teresa
Cortadellas, Santiago Laplagne and Juan Carlos Naranjo for helpful discussions and
Matilde Lalı́n for pointing us some references on Mahler measures. We also thank the
referee for his/her comments and, especially, for a simplification of our argument in
the proof of Lemma 1.29. D’Andrea thanks the University of Bordeaux 1 for inviting
him in February 2009, Krick thanks the University of Buenos Aires for her sabbatical
during 2009 and the Universities of Bordeaux 1, Caen, Barcelona and Nice-Sophia
Antipolis for hosting her during that time, and Sombra thanks the University of
Buenos Aires for inviting him during October-December 2007 and in November 2010.
The three authors also thank the Fields Institute, where they met during the Fall 2009
FoCM thematic program.
1. Degrees and resultants of multiprojective cycles
Throughout this text, we denote by N = Z≥0 and by Z>0 the sets of non-negative and
positive integers, respectively. Bold letters denote finite sets or sequences of objects,
where the type and number should be clear from the context: for instance, x might
denote {x1 , . . . , xn } so that if A is a ring, A[x] = A[x1 , . . . , xn ]. For a polynomial
f ∈ A[x] we adopt the usual notation
X
αa xa
f=
a
where, for each index a = (a1 , . . . , an ) ∈ Nn , αa denotes an element of A and xa
the monomial xa11 · · · xann . For a ∈ Nn , we denote by |a| = a1 + · · · + an its length
and by coeff a (f ) = αa the coefficient of xa . We also set a! = a1 ! · · · an !. The
support of f is the set of exponents corresponding to its non-zero terms,
that is,
P
supp(f ) = {a : coeff a (f ) 6= 0} ⊂ Nn . For a, b ∈ Rn , we set ha, bi = ni=1 ai bi . We
say that a ≤ b whenever the inequality holds coefficient wise.
For a factorial ring A, we denote by A× its group of units. A polynomial with coefficients in A is primitive if its coefficients have no common factor in A \ A× .
1.1. Preliminaries on multiprojective geometry. Let A be a factorial ring with
field of fractions K and K the algebraic closure of K. For m ∈ Z>0 and n =
(n1 , . . . , nm ) ∈ Nm we consider the multiprojective space over K
Pn (K) = Pn1 (K) × · · · × Pnm (K).
We also write Pn = Pn (K) for short. For 1 ≤ i ≤ m, let xi = {xi,0 , . . . , xi,ni } be a
group of ni + 1 variables and set
x = {x1 , . . . , xm }.
The multihomogeneous coordinate ring of Pn is K[x] = K[x1 , . . . , xm ]. It is multigraded by declaring deg(xi,j ) = ei ∈ Nm , the i-th vector of the standard basis of Rm .
8
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
For d = (d1 , . . . , dm ) ∈ Nm , we denote by K[x]d its part of multidegree d. Set
Y
Nndii +1 = {ai ∈ Nni +1 : |ai | = di } , Nn+1
=
Nndii +1 .
d
1≤i≤m
A multihomogeneous polynomial f ∈ K[x]d can then be written down as
X
f=
αa xa .
a∈Nn+1
d
Let K ⊂ E be an extension of fields and f ∈ E[x]d . For a point ξ ∈ Pn , the value
×
f (ξ) is only defined up to a non-zero scalar in K which depends on a choice of
multihomogeneous coordinates for ξ.
An ideal I ⊂ K[x] is multihomogeneous if it is generated by a family of multihomogeneous polynomials. For any such ideal, we denote by V (I) the subvariety of Pn
defined as its set of zeros. Along this text, a variety is neither necessarily irreducible
nor equidimensional. Reciprocally, given a variety V ⊂ Pn , we denote by I(V ) the
multihomogeneous ideal of K[x] of polynomials vanishing on V . A variety V is defined
over K if its defining ideal I(V ) is generated by polynomials in K[x].
Let Mn = {x1,j1 · · · xm,jm : 0 ≤ ji ≤ ni } be the set of monomials of multidegree
(1, . . . , 1) ∈ Nm . A multihomogeneous
ideal I ⊂ K[x] defines the empty variety
√
n
of P if and only if Mn ⊂ I, see for instance [Rem01a, Lem. 2.9]. The assignment
V 7→ I(V ) is a one-to-one correspondence between non-empty subvarieties of Pn and
radical multihomogeneous ideals of K[x] not containing Mn .
More generally, we denote by Pn
K the multiprojective space over K corresponding
n
to n. The reduced subschemes of Pn
K will be alternatively called subvarieties of PK
or K-varieties. There is a one-to-one correspondence V 7→ I(V ) between non-empty
subvarieties of Pn
K and radical multihomogeneous ideals of K[x] not containing Mn .
For a multihomogeneous ideal I ⊂ K[x] not containing Mn , we denote by V (I) its
corresponding K-variety. A K-variety V is irreducible if it is an integral subscheme
of Pn
K or, equivalently, if the ideal I(V ) is prime. The dimension of V coincides with
the Krull dimension of the algebra K[x1 , . . . , xm ]/I(V ) minus m.
can be identified with
Remark 1.1. In the algebraically closed case, the scheme Pn
K
n
n
the set of points P (K) and a subvariety V ⊂ PK can be identified with its set of
points V (K) ⊂ Pn (K). Under this identification, a subvariety of Pn (K) defined over
K corresponds to a K-variety. However, a K-variety does not necessarily correspond
to a subvariety of Pn (K) defined over K, as the following example shows. Let t be
a variable and set K = Fp (t), where p is a prime number and Fp is the field with p
elements. The ideal (xp1 − txp0 ) ⊂ K[x0 , x1 ] is prime and hence gives a subvariety
of P1K . Its set of zeros in P1 (K) consists in the point {(1 : t1/p )}, which is not a
variety defined over K. When the field K is perfect (for instance, if char(K) = 0),
the notion of K-variety does coincide, under this identification, with the notion of
subvariety of Pn (K) defined over K.
A K-cycle of Pn
K is a finite Z-linear combination
X
(1.2)
X=
mV V
V
Pn
K.
of irreducible subvarieties of
The subvarieties V such that mV 6= 0 are the
irreducible components of X. A K-cycle is of pure dimension or equidimensional if
its components are all of the same dimension. It is effective (respectively, reduced) if
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
9
it can be written as in (1.2) with mV ≥ 0 (respectively, mV = 1). Given two K-cycles
X1 and X2 , we say that X1 ≥ X2 whenever X1 − X2 is effective. The support of X,
denoted |X|, is the K-variety defined as the union of its components. Reciprocally,
a K-variety is a union of irreducible K-varieties of Pn and we identify it with the
reduced K-cycle given as the sum of these irreducible K-varieties.
n
For 0 ≤ r ≤ |n|, we denote by Zr (Pn
K ) the group of K-cycles of P of pure dimension r
and by Zr+ (Pn
K ) the semigroup of those which are effective. For shorthand, a K-cycle
is called a cycle and we denote the sets of K-cycles and of effective K-cycles of pure
dimension r as Zr (Pn ) and as Zr+ (Pn ), respectively.
Let I ⊂ K[x] be a multihomogeneous ideal. For each minimal prime ideal P of I, we
denote by mP the multiplicity of P in I, defined as the length of the K[x]P -module
(K[x]/I)P . We associate to I the K-cycle
X
X(I) :=
mP V (P ).
P
If V (I) is of pure dimension r, then X(I) ∈ Zr+ (Pn
K ). Let K ⊂ E be an extension of
fields and V an irreducible K-variety. We define the scalar extension of V by E as
the E-cycle VE = X(I(V ) ⊗K E). This notion extends to K-cycles by linearity and
n
induces an inclusion of groups Zr (Pn
K ) ֒→ Zr (PE ).
Each Weil or Cartier divisor of Pn
K is globally defined by a single rational multihomogeneous function in K(x) because the ring K[x] is factorial [Har77, Prop. II.6.2
and II.6.11]. Hence, we will not make distinctions between them. We write Div(Pn
K) =
+
+ n
n
n
n
Z|n|−1 (PK ) for the group of divisors of PK and by Div (PK ) = Z|n|−1 (PK ) for the
semigroup of those which are effective.
Each effective divisor D of Pn
K is defined by a multihomogeneous primitive polynomial
in A[x] \ {0},
unique
up
to
a unit of A. We denote this polynomial by fD . If we
P
write D = H mH H where H is a K-hypersurface of Pn and mH ∈ N, then there
exists λ ∈ A× such that
Y m
fH H .
fD = λ
H
Conversely, given a multihomogeneous polynomial f ∈ A[x] \ {0}, we denote by
div(f ) ∈ Div+ (Pn
K ) the associated divisor.
We introduce some basic operations on cycles and divisors.
Definition 1.3. Let V be an irreducible subvariety of Pn
K and H an irreducible
hypersurface not containing V . Let Y be an irreducible component of V ∩ H. The
intersection multiplicity of V and H along Y , denoted mult(Y |V, H), is the length of
the K[x]I(Y ) -module K[x]/(I(V ) + I(H)) I(Y ) , see [Har77, §I.1.7]. The intersection
product of V and H is defined as
X
mult(Y |V, H) Y,
(1.4)
V ·H =
Y
the sum being over the irreducible components of V ∩ H. It is a cycle of pure dimension dim(V ) − 1.
Let X be an equidimensional cycle and D a divisor. We say that X and D intersect
properly if no irreducible component of X is contained in |D|. By bilinearity, the
intersection product in (1.4) extends to a pairing
n
n
Zr (Pn
K ) × Div(PK ) 99K Zr−1 (PK ) ,
(X, D) 7−→ X · D,
10
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
well-defined whenever X and D intersect properly.
Qℓ
n
Let X ∈ Zr (Pn
K ) and D1 , . . . , Dℓ ∈ Div(PK ). Then X ·
j=1 Dj does not depend
on the order of the divisors, provided that all the intermediate products are welldefined [Ful84, Cor. 2.4.2 and Example 7.1.10(a)].
n2
1
Definition 1.5. Let m1 , m2 ∈ Z>0 and ni ∈ Nmi for i = 1, 2. Let ϕ : Pn
K 99K PK be
1
a rational map and V an irreducible subvariety of Pn
K . The degree of ϕ on V is
(
[K(V ) : K(ϕ(V ))] if dim(ϕ(V )) = dim(V ),
deg(ϕ|V ) =
0
if dim(ϕ(V )) < dim(V ).
The direct image of V under ϕ is defined as the cycle ϕ∗ V = deg(ϕ|V ) ϕ(V ). It is a
cycle of the same dimension as V . This notion extends by linearity to equidimensional
cycles and induces a Z-linear map
n2
1
ϕ∗ : Zr (Pn
K ) −→ Zr (PK ).
n3
2
If ψ : Pn
K 99K PK is a further rational map, then (ψ ◦ ϕ)∗ = ψ∗ ◦ ϕ∗ because of the
multiplicativity of the degree of field extensions.
2
Let H be a hypersuface of Pn
K not containing the image of ϕ. The inverse image
of H under ϕ is defined as the hypersurface ϕ∗ H = ϕ−1 (H). This notion extends to
a Z-linear map
n1
2
ϕ∗ : Div(Pn
K ) 99K Div(PK ),
well-defined for divisors whose support does not contain the image of ϕ.
Direct images of cycles, inverse images of divisors and intersection products are related
n2
1
by the projection formula [Ful84, Prop. 2.3(c)]: let ϕ : Pn
K → PK be a proper map,
n1
n2
X a cycle of PK and D a divisor of PK containing no component of ϕ(|X|). Then
(1.6)
ϕ∗ (X · ϕ∗ D) = ϕ∗ X · D.
1.2. Mixed degrees. We recall the basic properties of mixed degrees of multiprojective cycles. We also study the behavior of this notion under linear projections.
Definition 1.7. Let V ⊂ Pn
K be an irreducible K-variety. The Hilbert-Samuel function of V is the numerical function defined as
HV : Nm −→ N , δ 7−→ dimK (K[x]/I(V ))δ .
Proposition 1.8. Let V ⊂ Pn
K be an irreducible K-variety of dimension r.
(1) There is a unique polynomial PV ∈ Q[z1 , . . . , zm ] such that PV (δ) = HV (δ)
for all δ ≥ δ 0 for some δ 0 ∈ Nm . In addition deg(PV ) = r.
(2) Let b = (b1 , . . . , bm ) ∈ Nm
r . Then b! coeff b (PV ) ∈ N. Moreover, if bi > ni for
some i, then coeff b (PV ) = 0.
Proof. (1) and the second part of (2) follow from [Rem01a, Thm. 2.10(1)]. The first
part of (2) follows from [Rem01a, Thm. 2.10(2)] and its proof.
The polynomial PV in Proposition 1.8 is called the Hilbert-Samuel polynomial of V .
m
Definition 1.9. Let V ⊂ Pn
K be an irreducible K-variety of dimension r and b ∈ Nr .
The (mixed) degree of V of index b is defined as
degb (V ) = b! coeff b (PV ).
It is a non-negative integer, thanks to Proposition 1.8(2). This notion extends by
linearity to equidimensional K-cycles and induces a map degb : Zr (Pn
K ) → Z.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
11
Recall that the Chow ring of Pn
K is the graded ring
n1 +1
nm +1
A∗ (Pn
, . . . , θm
),
K ) = Z[θ1 , . . . , θm ]/(θ1
where each θi denotes the class of the inverse image of a hyperplane of PnKi under the
ni
n
projection Pn
K → PK [Ful84, Example 8.3.7]. Given a cycle X ∈ Zr (PK ), its class in
the Chow ring is
X
nm −bm
∈ A∗ (Pn
[X] =
degb (X) θ1n1 −b1 · · · θm
K ),
b
the sum being over all b ∈ Nm
r such that b ≤ n. It is a homogeneous element of degree
|n| − r. By Proposition 1.8(2), degb (X) = 0 whenever bi > ni for some i. Hence, [X]
contains the information of all the mixed degrees of X, since {θ b }b≤n is a Z-basis of
n
A∗ (Pn
K ). For X1 , X2 ∈ Zr (PK ), we say that [X1 ] ≥ [X2 ] whenever the inequality holds
coefficient wise in terms of this basis.
Given a K-cycle X, its class in the Chow ring is invariant under field extensions. In
particular, [XK ] = [X] and degb (X) = degb (XK ) for all b ∈ Nm
r . If dim(X) = 0, its
degree is defined as the number of points in XK , counted with multiplicity.
Proposition 1.10.
(1) Let X ∈ Zr+ (Pn
K ). Then [X] ≥ 0.
n
n
(2) We have [PK ] = 1. Equivalently, degn (Pn
K ) = 1 and degb (PK ) = 0 for all
m
b ∈ N|n| such that b 6= n.
n
(3) Let X ∈ Z0 (Pn
K ). Then [X] = deg(X)θ . Equivalently, deg0 (X) = deg(X).
(4) Let D ∈ Div+ (Pn
K ) and fD its defining polynomial. Then
[D] =
m
X
degxi (fD )θi .
i=1
Equivalently, degn−ei (D) = degxi (fD ) for 1 ≤ i ≤ m and degb (D) = 0 for all
b ∈ Nm
|n|−1 such that b 6= n − ei for all i.
(5) Let n ∈ N and V ⊂ PnK a K-variety of pure dimension r. Then
[V ] = deg(V )θ n−r ,
where deg(V ) denotes the degree of the projective variety V . Equivalently,
degr (V ) = deg(V ).
Proof.
(1) This follows from the definition of [X] and Proposition 1.8(2).
(2) For δ = (δ1 , . . . , δm ) ∈ Nm ,
m
Y
ni + δi
1
n
=
HPK (δ) = dimK K[x]δ =
δ n + O(||δ|||n|−1 ),
ni
n1 ! · · · nm !
i=1
where || · || denotes any fixed norm on Rm . This implies that degn (Pn
K ) = 1 and thus
n ) = 1, as stated.
[Pn
]
=
deg
(P
n K
K
N
(3) Let ξ = (ξ 1 , . . . , ξ m ) ∈ Pn . We have K[x]/I(ξ) = m
i=1 K[xi ]/I(ξ i ). Hence, for
m
δ∈N ,
Hξ (δ) = dimK
m
Y
dimK K[xi ]/I(ξ i ) δ = 1.
K[x]/I(ξ) δ =
i
i=1
12
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
This implies that deg0 (ξ) P
= 1 and so [ξ] = θ n . For a general zero-dimensional
K-cycle X, write XK =
mξ ξ for some points ξ ∈ Pn and mξ ∈ Z. Hence,
ξP
P
deg0 (X) = ξ mξ deg0 (ξ) = ξ mξ = deg(X) and so [X] = deg(X)θ n .
(4) Write deg(f ) = (degx1 (f ), . . . , degxm (f )). For δ ≥ deg(f ), there is an exact
sequence
×f
0 −→ K[x]δ−deg(f ) −→ K[x]δ −→ (K[x]/(f ))δ −→ 0.
Hence, HD (δ) = HPnK (δ) − HPnK (δ − deg(fD )) and therefore
m
X
degxi (f ) n−ei
δ
+ O(||δ|||n|−2 ).
(n − ei )!
i=1
P
This implies that degn−ei (D) = degxi (f ) and so [D] = m
i=1 degxi (fD )θi , as stated.
(5) This follows readily from the definition of deg(V ) in terms of Hilbert functions.
PD (d) =
The following is the multiprojective version of Bézout’s theorem.
Theorem 1.11. Let X ∈ Zr (Pn
K ) and f ∈ K[x1 , . . . , xm ] be a multihomogeneous
polynomial such that X and div(f ) intersect properly. Then
Equivalently, degb
[X · div(f )] = [X] · [div(f )].
P
m
X · div(f ) = m
i=1 degxi (f ) degb+ei (X) for all b ∈ Nr−1 .
Proof. The equivalence between the two statements follows from Proposition 1.10(4).
The second statement follows for instance from [Rem01b, Thm. 3.4].
Next corollary follows readily from this result together with Proposition 1.10(3).
Corollary 1.12. Let f1 , . . . , f|n| ∈ K[x1 , . . . , xm ] be multihomogeneous polynomials
such that dim V (f1 , . . . , fi ) = |n| − i for all i. Then
|n|
|n|
Y
Y
deg
div(fi ) = coeff θn
degx1 (fi ) θ1 + · · · + degxm (fi ) θm .
i=1
i=1
Example 1.13. How many pairs (eigenvalue, eigenvector) can a generic square matrix
have? Given M = (mi,j )i,j ∈ K n×n , the problem of computing these pairs consists in
n
solving M v = λ v for λ ∈ K and v = (v1 , . . . , vn ) ∈ K \ {0}. Set
f i = s 1 vi − s 0
n
X
j=1
mi,j vj
,
1 ≤ i ≤ n.
The matrix equation M v = λ v translates into the system of n bilinear scalar equations
fi = 0, 1 ≤ i ≤ n, for ((s0 : s1 ), v) ∈ P1 × Pn−1 such that s0 6= 0. If M is generic, the
hypersurfaces V (fi ) intersect properly. By Corollary 1.12, the number of solutions in
P1 × Pn−1 of this system of equations is
n
Y
degs (fi )θ1 + degv (fi )θ2 = coeff θ1 θn−1 (θ1 + θ2 )n = n.
coeff θ1 θn−1
2
i=1
2
We deduce that M admits at most n pairs (eigenvalue, eigenvector) counted with
multiplicities. A straightforward application of the usual Bézout’s theorem would
have given the much larger bound 2n .
The following result shows that mixed degrees can also be defined geometrically.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
13
m
Corollary 1.14. Let X ∈ Zr (Pn
K ) and b ∈ Nr . For 1 ≤ i ≤ m and 0 ≤ j ≤ bi we
ni
n
denote by Hi,j ⊂ PK the inverse image with respect to the projection Pn
K → PK of a
ni
generic hyperplane of PK . Then
bi
m Y
Y
Hi,j .
degb (X) = deg X ·
i=1 j=1
Proof. The variety X and the divisors Hi,j intersect properly and [Hi,j ] = θi . TheoQ Qbi
n
rem 1.11 implies that, for Z = X · m
i=1
j=1 Hi,j ∈ Z0 (PK ),
bi
m Y
X
Y
[Hi,j ] = [X]θ b =
degc (X)θ n−c θ b = degb (X)θ n ,
deg(Z)θ = [Z] = [X]·
n
c∈Nm
r
i=1 j=1
which proves the statement.
Next we show that mixed degrees are monotonic with respect to linear projections.
For 1 ≤ i ≤ m, let 0 ≤ li ≤ ni and set l = (l1 , . . . , lm ) ∈ Nm . Consider the linear
projection which forgets the last ni − li coordinates in each factor of Pn
K:
(1.15)
l
π : Pn
K 99K PK
,
(xi,j ) 1≤i≤m 7−→ (xi,j ) 1≤i≤m
0≤j≤li
0≤j≤ni
This is a rational map, well-defined outside the union of linear subspaces L :=
S
m
n
i=1 V (xi,0 , . . . , xi,li ) ⊂ PK . It induces an injective Z-linear map
: A∗ (PlK ) ֒−→ A∗ (Pn
K) ,
P 7−→ θ n−l P.
l
Proposition 1.16. Let π : Pn
K 99K PK be the linear projection as above and X ∈
+
n
Zr (PK ). Then
[π∗ X] ≤ [X].
Equivalently, degb π∗ X ≤ degb (X) for all b ∈ Nm
r .
The proof of this result relies on the technical Lemma 1.18 below, which was suggested
to us by José Ignacio Burgos. Consider the blow up of Pn along the subvariety L,
denoted BlL (Pn ) and defined as the closure in Pn × Pl of the graph of π. It is an
irreducible variety of dimension |n|. Set x and y for the multihomogeneous coordinates
of Pn and Pl , respectively. The ideal of this variety is
(1.17) I(BlL (Pn )) = {xi,j1 yi,j2 − xi,j2 yi,j1 : 1 ≤ i ≤ m, 0 ≤ j1 < j2 ≤ li } ⊂ K[x, y].
Consider the projections
pr1 : Pn × Pl −→ Pn
,
pr2 : Pn × Pl −→ Pl .
The exceptional divisor of the blow up is supported in the hypersurface E = pr−1
1 (L).
Let V ⊂ Pn be an irreducible variety such that V 6⊂ L and W its strict transform,
n
which is the closure of the set pr−1
1 (V \ L) ∩ BlL (P ). Then
pr1∗ W = V
,
pr2∗ W = π∗ V.
For a multihomogeneous polynomial f ∈ K[y] \ {0}, we write divPl (f ) for the divisor
of Pl defined by f (y) and divPn (f ) for the divisor of Pn defined by f (x).
Lemma 1.18. Let V ⊂ Pn be an irreducible variety of dimension r such that V 6⊂
L and f ∈ K[y] \ {0} a multihomogeneous polynomial. Assume that pr∗2 divPl (f )
intersects W properly and that no component of W · pr∗2 divPl (f ) is contained in E.
Then divPl (f ) (respectively, divPn (f )) intersects π∗ V (respectively, V ) properly and
π∗ (V · divPn (f )) = π∗ V · divPl (f ).
14
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof. Consider the following divisors of Pn × Pl :
D1 = pr∗1 divPn (f ) ,
D2 = pr∗2 divPl (f ).
Write for short B = BlL (Pn ). Since W is irreducible, the hypothesis that D2 intersects
W properly is equivalent to the fact that W is not contained in |D2 |. This implies
that neither V is contained in V (f (x)) nor π(V ) is contained in V (f (y)). Hence all
intersection products are well-defined. We claim that
B · (D2 − D1 )
is a cycle of pure dimension |n| − 1 with support contained in the hypersurface E. To
prove this, write di = degxi (f ) and, for each 1 ≤ i ≤ m, choose an index 0 ≤ ji ≤ li .
Using (1.17) we verify that
m
Y
i=1
di
yi,j
i
m
Y
xdi,ji i f (y) (mod I(B)).
f (x) ≡
i=1
Observe that the ideal of E in K[x, y]/I(B) is generated by the set of monomials
m
Y
i=1
xi,ji
1≤i≤m
0≤ji ≤li
Q
Let (ξ, ξ ′ ) ∈ B \QE. We have that m
i=1 ξi,ji 6= 0 for a choice of ji ’s. From here, we
′
can verify that m
ξ
=
6
0.
This
implies
that f (x) and f (y) generate the same
i=1 i,ji
ideal in the localization (K[x, y]/I(B))(ξ,ξ′ ) for all such (ξ, ξ ′ ) and proves the claim.
Therefore, there exists a cycle Z ∈ Zr−1 (Pn × Pl ) supported on E such that
W · D1 = W · D2 − Z.
Since the map pr1 is proper, the projection formula (1.6) implies that
V · divPn (f ) = pr1∗ (W · D1 ) = pr1∗ (W · D2 ) − pr1∗ Z.
By hypothesis, no component of W · D2 is contained in E. Since pr1 : B \ E → Pn \ L
is an isomorphism, no component of pr1∗ (W · D2 ) is contained in L. Hence,
π∗ (V · divPn (f ))) = π∗ pr1∗ (W · D2 ) − π∗ pr1∗ Z = (π ◦ pr1 )∗ (W · D2 ) = pr2∗ (W · D2 ),
because π∗ pr1∗ Z = 0 as this is a cycle supported on L. Again by the projection
formula,
pr2∗ (W · D2 ) = pr2∗ W · divPl (f ) = π∗ V · divPl (f ),
which proves the statement.
Proof of Proposition 1.16. The equivalence between the two formulations is a direct
consequence of the definitions. We reduce without loss of generality to the case of an
irreducible variety V ⊂ Pn such that dim(π(V )) = r.
We proceed by induction on the dimension. For r = 0, the statement is obvious and
so we assume r ≥ 1. Let 1 ≤ i ≤ m such that bi 6= 0 and ℓ ∈ K[y i ] a linear form. For
each component C of W ∩ E we pick a point ξ C ∈ C and we impose that ℓ(ξ C ) 6= 0,
which holds for a generic choice of ℓ. This implies that pr∗2 (divPl (ℓ)) intersects W
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
15
properly and that their intersection has no component contained in E. Hence
degb (π∗ V ) = degb−ei (π∗ V · divPl (ℓ))
by Bézout’s theorem 1.11,
= degb−ei (π∗ (V · divPn (ℓ)))
≤ degb−ei (V · divPn (ℓ)))
= degb (V )
by Lemma 1.18,
by the inductive hypothesis,
by Bézout’s theorem 1.11,
which completes the proof.
Next result gives the behavior of Chow rings and classes with respect to products.
Proposition 1.19. Let mi ∈ Z>0 and ni ∈ Nmi for i = 1, 2. Then
n2
∗ n1
∗ n2
1
(1) A∗ (Pn
K × PK ) ≃ A (PK ) ⊗Z A (PK ).
i
(2) Let Xi ∈ Zri (Pn
K ) for i = 1, 2. The above isomorphism identifies [X1 × X2 ]
with [X1 ] ⊗Z [X2 ]. Equivalently, for all bi ∈ Nmi such that |b1 | + |b2 | = r1 + r2 ,
(
degb1 (X1 ) degb2 (X2 ) if |b1 | = r1 , |b2 | = r2 ,
degb1 ,b2 (X1 × X2 ) =
0
otherwise.
Proof.
(1) This is immediate from the definition of the Chow ring.
i
(2) We reduce without loss of generality to the case of irreducible K-varieties Vi ⊂ Pn
K,
ni
i = 1, 2. Let xi denote the multihomogeneous coordinates of PK . For δ i ∈ Nmi ,
K[x1 , x2 ]/I(V1 × V2 ) δ ,δ ≃ K[x1 ]/I(V1 ) δ ⊗ K[x2 ]/I(V2 ) δ .
1
2
1
2
Hence HV1 ×V2 (δ 1 , δ 2 ) = HV1 (δ 1 )HV2 (δ 2 ) and therefore PV1 ×V2 = PV1 PV2 . This implies
the equality of mixed degrees, which in turn implies that [V1 × V2 ] = [V1 ] ⊗ [V2 ] under
the identification in (1).
We end this section with the notion of ruled join of projective varieties. Let ni ∈ N
and consider an irreducible K-variety Vi ⊂ PnKi for i = 1, 2. Let K[xi ] denote the
homogeneous coordinate ring of PnKi and I(Vi ) ⊂ K[xi ] the ideal of Vi . The ruled join
of V1 and V2 , denoted V1 #V2 , is the irreducible subvariety of PnK1 +n2 +1 defined by the
homogeneous ideal generated by I(V1 ) ∪ I(V2 ) in K[x1 , x2 ]. In case K is algebraically
closed, identifying Pn1 and Pn2 with the linear subspaces of Pn1 +n2 +1 where the last
n2 + 1 (respectively, the first n1 + 1) coordinates vanish, V1 #V2 coincides with the
union of the lines of Pn1 +n2 +1 joining points of V1 with points of V2 .
The notion of ruled join extends to equidimensional K-cycles by linearity. Given cycles
Xi ∈ Zri (PnKi ), i = 1, 2, the ruled join X1 #X2 is a cycle of PnK1 +n2 +1 of pure dimension
r1 + r2 + 1 and degree
(1.20)
deg(X1 #X2 ) = deg(X1 ) deg(X2 ),
see for instance [Ful84, Example 8.4.5]. For i = 1, 2 consider the injective Z-linear
map i : A∗ (PnKi ) ֒→ A∗ (PnK1 +n2 +1 ) defined by θ l 7→ θ l for 0 ≤ l ≤ ni . Then (1.20) is
equivalent to the equality of classes
[X1 #X2 ] = 1 ([X1 ]) · 2 ([X2 ]).
16
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
1.3. Eliminants and resultants. In this section, we introduce the notions and basic properties of eliminants of varieties and of resultants of cycles in multiprojective
spaces. This is mostly a reformulation of the theory of eliminants and resultants of
multihomogeneous ideals developed by Rémond in [Rem01a, Rem01b] as an extension
of Philippon’s theory of eliminants of homogeneous ideals [Phi86]. We refer the reader
to these articles for a complementary presentation of the subject.
We keep the notation of §1.1. In particular, we denote by A a factorial ring with
field of fractions K. Let V ⊂ Pn
K be an irreducible K-variety of dimension r. Let
d0 , . . . , dr ∈ Nm \ {0} and set d = (d0 , . . . , dr ). For each 0 ≤ i ≤ r, we introduce
a group of variables ui = {ui,a : a ∈ Nn+1
di } and consider the general form Fi of
multidegree di in the variables x:
X
(1.21)
Fi =
ui,a xa ∈ K[ui ][x].
a∈Nn+1
d
i
Set u = {u0 , . . . , ur } and consider the K[u]-module
Md (V ) = K[u][x1 , . . . , xm ]/ I(V ) + (F0 , . . . , Fr ) .
This module inherits a multigraded structure from K[x]. For δ ∈ Nm , we denote by
Md (V )δ its part of multidegree δ in the variables x. It is a K[u]-module multigraded
by setting deg(ui,a ) = ei ∈ Nr+1 , the (i + 1)-th vector of the standard basis of Rr+1 .
For the sequel, we fix a set of representatives of the irreducible elements of K[u] made
out of primitive polynomials in A[u] and we denote it by irr(K[u]). We recall that
the annihilator of a K[u]-module M is the ideal of K[u] defined as
Ann(M ) = AnnK[u] (M ) = {f ∈ K[u] : f M = 0}.
Definition 1.22. Let M be a finitely generated K[u]-module. If Ann(M ) 6= 0, we set
Y
(1.23)
χ(M ) = χK[u] (M ) =
f ℓ(M(f ) ) ,
f ∈ irr(K[u])
where ℓ(M(f ) ) denotes the length of the K[u](f ) -module M(f ) . In case Ann(M ) = 0,
we set χ(M ) = 0.
We have that ℓ(M(f ) ) ≥ 1 if and only if Ann(M ) ⊂ (f ), see [Rem01a, §3.1]. Hence,
the product in (1.23) involves a finite number of factors and χ(M ) is well-defined.
Lemma 1.24. Let V ⊂ Pn
K be an irreducible K-variety of dimension r and d ∈
(Nm \ {0})r+1 . Then there exists δ 0 ∈ Nm such that
Ann(Md (V )δ ) = Ann(Md (V )δ 0 )
,
χ(Md (V )δ ) = χ(Md (V )δ 0 )
for all δ ∈ Nm such that δ ≥ δ 0 .
Proof. Let δ max ∈ Nm be the maximum of the multidegrees of a set of generators of
Md (V ) over K[u]. For δ′ ≥ δ ≥ δ max we have that Md (V )δ ′ = K[u][x]δ′ −δ Md (V )δ
and so Ann(Md (V )δ ′ ) ⊃ Ann(Md (V )δ ). Hence the annihilators of the parts of multidegree ≥ δ max form an ascending chain of ideals with respect to the order ≤ on Nm .
Eventually, this chain stabilizes because K[u] is Noetherian, which proves the first
statement. The second statement is [Rem01a, Lem. 3.2].
We define eliminants and resultants following [Rem01a, Def. 2.14 and §3.2]. The
principal part ppr(I) of an ideal I ⊂ K[u] is defined as any primitive polynomial in
A[u] which is a greatest common divisor of the elements in I ∩ A[u]. If I is principal,
this polynomial can be equivalently defined as any primitive polynomial in A[u] which
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
17
is a generator of I. The principal part of an ideal is unique up to a unit of A and we
fix its choice by supposing that it is a product of elements of irr(K[u]).
Definition 1.25. Let V ⊂ Pn
K be an irreducible K-variety of dimension r ≥ 0 and
m
r+1
d ∈ (N \ {0}) . The eliminant ideal of V of index d is defined as
Ed (V ) = Ann(Md (V )δ )
for any δ ≫ 0. The eliminant of V of index d is defined as
Elimd (V ) = ppr(Ed (V )).
The eliminant ideal is a non-zero multihomogeneous prime ideal in K[u] [Rem01a,
Lem. 2.4(2) and Thm. 2.13(1)], see also Proposition 1.37(1) below. In particular, the
eliminant is a primitive irreducible multihomogeneous polynomial in A[u] \ {0}.
Definition 1.26. Let V ⊂ Pn
K be an irreducible K-variety of dimension r ≥ 0 and
d ∈ (Nm \ {0})r+1 . The resultant of V of index d is defined as
Resd (V ) = χ Md (V )δ
for any δ ≫ 0. It is a non-zero primitive multihomogeneous polynomial in A[u],
because of Definition 1.22 and the
P fact that Ed (V ) is non-zero.
Let X ∈ Zr (Pn
)
and
write
X
=
V mV V . The resultant of X of index d is defined as
K
Y
Resd (X) =
Resd (V )mV ∈ K(u)× .
V
When X is effective, Resd (X) is a primitive multihomogeneous polynomial in A[u].
Eliminants and resultants are invariant under index permutations. Next result follows
easily from the definitions:
n
Proposition 1.27. Let X ∈ Zr (Pn
K ) and V ⊂ PK an irreducible K-variety of dim
r+1
mension r. Let d = (d0 , . . . , dr ) ∈ (N \ {0})
and u = (u0 , . . . , ur ) the group of
variables corresponding to d. Let σ be a permutation of the set {0, . . . , r} and write
σd = (dσ(0) , . . . , dσ(r) ), σu = (uσ(0) , . . . , uσ(r) ). Then Resσd (X)(σu) = Resd (X)(u)
and Elimσd (V )(σu) = Elimd (V )(u).
Eliminants and resultants are also invariant under field extensions.
n
Proposition 1.28. Let X ∈ Zr (Pn
K ), V ⊂ PK an irreducible K-variety of dimension r
m
r+1
and d = (d0 , . . . , dr ) ∈ (N \ {0}) . Let K ⊂ E be a field extension. Then
there exists λ1 ∈ E × such that Resd (XE ) = λ1 Resd (X). Furthermore, if VE is an
irreducible E-variety, then there exists λ2 ∈ E × such that Elimd (VE ) = λ2 Elimd (V ).
To prove this, we need the following lemma.
Lemma 1.29. Let M be a finitely generated K[u]-module and K ⊂ E a field extension. Then AnnE[u] (M ⊗K E) = AnnK[u] (M ) ⊗K E and χE (M ⊗K E) = λ χK (M )
with λ ∈ E × .
Proof. The first statement is a consequence of the fact that E is a flat K-module: for
each m ∈ M , the exact sequence
0 → AnnK[u] (m) → K[u] → K[u]m → 0
yields the tensored exact sequence
0 → AnnK[u] (m) ⊗K E → E[u] → E[u](m ⊗K 1) → 0.
18
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Hence, AnnE[u] (m ⊗K 1) = AnnK[u] (m) ⊗K E. Now, if M = (m1 , . . . , mℓ ), then
\
AnnE[u] (M ⊗K E) =
AnnE[u] (mi ⊗K 1)
i
=
\
i
AnnK[u] (mi ) ⊗K E = AnnK[u] (M ) ⊗K E.
For the second statement, we can reduce to the case when Ann(M ) 6= 0 because
otherwise it is trivial. Let f ∈ irr(K[u]). The localization K[u](f ) is a principal local
LN
νi
domain and so M(fQ
) ≃
i=1 K[u](f ) /(f ) for some νi ≥ 1. In particular, ℓ(M(f ) ) =
P
µ
g
be the factorization of f into elements g ∈ irr(E[u]) and a
i νi . Let f = λf
gg
×
non-zero constant λf ∈ E . On the one hand, for each g in this factorization,
M ⊗K E
(g)
≃ M(f ) ⊗K[u] E[u](g) ≃
N
M
E[u](g) /(gµg νi ).
i=1
P
Hence ℓ((M ⊗K E)(g) ) = ( νi )µg = ℓ(M(f ) )µg . On the other hand, let g ∈ irr(E[u])
be an irreducible polynomial which does not divide any f ∈ irr(K[u]) and suppose
that ℓ((M ⊗K E)(g) ) ≥ 1. If this were the case, we would have AnnE[u] (M ⊗K E) ⊂ (g).
By the (already proved) first part of this proposition,
AnnE[u] (M ⊗K E) = AnnK[u] (M ) ⊗K E.
This implies Ann(M ) ⊂ (g) ∩ K[u] = 0, which contradicts the assumption that
Ann(M ) 6= 0. Therefore, ℓ((M ⊗K E)(g) ) = 0. We deduce
Y
Y
Y
Y
gℓ((M ⊗E)(g) ) =
gℓ(M(f ) )µg = λ
f ℓ(M(f ) )
g∈ irr(E[u])
for λ =
Q
−ℓ(M(f ) )
f
λf
f ∈ irr(K[u])
g|f
f
∈ E × . Hence χE[u] (M ⊗K E) = λ χK[u] (M ), as stated.
Proof of Proposition 1.28. To prove the first part, it is enough to consider the case of
an irreducible K-variety V . By definition, VE is the effective E-cycle defined by the
extended ideal I(V ) ⊗K E. Hence,
Md (VE ) = E[u][x1 , . . . , xm ]/ I(V ) ⊗K E + (F0 , . . . , Fr ) = Md (V ) ⊗K E.
By [Rem01a, Thm. 3.3] and Lemma 1.29,
Resd (VE ) = χE[u] Md (VE )δ ) = λ χK[u] Md (V )δ ) = λ1 Resd (V )
for any δ ≫ 0 and a λ1 ∈ E × , which proves the first part of the statement. The
second part follows similarly from the definition of eliminants and Lemma 1.29.
Proposition 1.30. Let V ⊂ Pn
K be an irreducible K-variety of dimension r and
m
r+1
d ∈ (N \ {0}) . Then there exists ν ≥ 1 such that
Resd (V ) = Elimd (V )ν .
Proof. Let δ ∈ Nm and f ∈ irr(K[u]). We have that (Md (V )δ )(f ) 6= 0 if and only if
Ann(Md (V )δ ) ⊂ (f ). Therefore,
f | Resd (V ) ⇐⇒ f | Elimd (V ).
Thus Elimd (V ) is the only irreducible factor of Resd (V ) and the statement follows.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
19
Example 1.31. The resultant of an irreducible variety is not necessarily an irreducible
polynomial: consider the curve C = V (x21,0 x2,1 − x21,1 x2,0 ) ⊂ P1 × P1 and the indexes
d0 = d1 = (0, 1) with associated linear forms Fi = ui,0 x2,0 + ui,1 x2,1 for i = 0, 1. We
can verify that the corresponding resultant is
Resd0 ,d1 (C) = (u0,0 u1,1 − u0,1 u1,0 )2 .
The partial degrees of resultants can be expressed in terms of mixed degrees.
m
r+1 . Then, for 0 ≤ i ≤ r,
Proposition 1.32. Let X ∈ Zr+ (Pn
K ) and d ∈ (N \ {0})
m
Y X
degui (Resd (X)) = coeff θn [X]
dj,ℓ θℓ .
j6=i
ℓ=1
Proof. This follows from [Rem01a, Prop. 3.4].
For projective varieties, eliminants and resultants coincide:
Corollary 1.33. For n ∈ N let V ⊂ PnK be an irreducible K-variety of dimension r
and d ∈ (Z>0 )r+1 . Then Resd (V ) = Elimd (V ).
ν
Proof. By PropositionQ1.30, Res
d (V ) = Elimd (V ) for some ν ≥ 1. On the one hand,
degui (Elimd (V )) =
j6=i dj deg(V ) for all i [Phi86, Remark to Lem. 1.8] while on
the other hand, Proposition 1.32 implies that
Y
Y
degui (Resd (V )) = coeff θn deg(V )θ n−r
dj θ =
dj deg(V ).
j6=i
j6=i
Thus Resd (V ) and Elimd (V ) have the same total degree. Hence, ν = 1 and the
statement follows.
Q
nj
Given a subset J ⊂ {1, . . . , m} we set πJ : Pn
K →
j∈J PK for the natural projection
and xJ = (xj )j∈J .
Lemma 1.34. Let d ∈ (Nm \ {0})r+1 and Fi the associated general form of multidegree di for 0 ≤ i ≤ r. Let V ⊂ Pn
K be an irreducible K-variety of dimension r ≥ 0.
Then Ed (V ) is a principal ideal if and only if
(1.35) dim(πJ (V )) ≥ #{i : 0 ≤ i ≤ r, Fi ∈ K[ui ][xJ ]} − 1
for all J ⊂ {1, . . . , m}.
If this is the case, Elimd (V ) ∈ K[u] \ K. Otherwise, Ed (V ) is not principal and
Elimd (V ) = 1.
Proof. Assume for the moment that the field K is infinite.
(⇐) If (1.35) holds, [Rem01a, Cor. 2.15(2)] implies that Elimd (V ) generates Ed (V ).
Applying [Rem01a, Cor. 2.15(1)], it follows Elimd (V ) 6= 0 since, for J = {1, . . . , m},
dim(πJ (V )) = dim(V ) = r = #{i ∈ {0, . . . , r} : Fi ∈ K[ui ][x]} − 1.
Now suppose that Elimd (V ) = 1. This is equivalent to the fact that Md (V )δ = 0
for δ ≫ 0, which implies I(V ) ⊃ (Mn )δ and so V = ∅, which is a contradiction.
Therefore, Elimd (V ) ∈ K[u] \ K.
(⇒) Suppose that (1.35) does not hold. By [Rem01a, Cor. 2.15(3)], Elimd (V ) = 1.
Hence Ed (V ) is necessarily not principal, because otherwise we would have that V = ∅.
The case when K is a finite field reduces to the previous case, by considering any
transcendental extension E of K and applying Proposition 1.28.
20
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Given d ∈ (Nm \ {0})r+1 , the space of coefficients of a family of multihomogeneous
polynomials in K[x] \ {0} of multidegrees d0 , . . . , dr can be identified with
P
N
:=
r
Y
P
Qm
j=1
j )−1
(di,jn+n
j
.
i=0
For V ⊂ Pn consider the following subset of PN :
(1.36)
∇d (V ) = {(u0 , . . . , ur ) ∈ PN : V ∩ V (F0 (u0 , x), . . . , Fr (ur , x)) 6= ∅}.
The following results gives a geometric interpretation of eliminant ideals.
Proposition 1.37. Let V ⊂ Pn be an irreducible variety of dimension r ≥ 0 and
d ∈ (Nm \ {0})r+1 . Then
(1) I(∇d (V )) = Ed (V );
(2) the variety ∇d (V ) is a hypersurface if and only if (1.35) holds. If this is the
case, ∇d (V ) = V (Elimd (V )) = V (Resd (V )).
Proof. (1) follows from [Rem01a, Thm. 2.2], while (2) follows from (1) together with
Lemma 1.34 and Proposition 1.30.
The following corollary gives a formula à la Poisson for the resultants of a cycle of
dimension 0. Recall that the evaluation of a multihomogeneous polynomial at a point
×
of Pn is only defined up to a non-zero constant in K which depends of a choice of a
representative of the given point.
P
m
Corollary 1.38. Let X ∈ Z0 (Pn
K ) and d0 ∈ N \ {0}. Write XK =
ξ mξ ξ with
n
ξ ∈ P and mξ ∈ Z and let F0 be the general form of multidegree d0 . Then there
×
exists λ ∈ K such that
Y
Resd0 (X) = λ
F0 (ξ)mξ .
ξ
Proof. Let ξ ∈ Pn . Observe that F0 (ξ) = 0 is the irreducible equation of the hypersur×
face ∇d (ξ). By Proposition 1.37, there exists λ ∈ K such that Elimd0 (ξ) = λ F0 (ξ).
Proposition 1.32 together with Proposition 1.10(3) imply that degu0 (Resd (ξ)) = 1.
Applying Proposition 1.30, we get Resd0 (ξ) = Elimd0 (ξ). The general case follows readily from the definition of the resultant and its invariance under field extensions.
Remark 1.39. The notions of eliminant and resultant of multiprojective cycles include several of the classical notions of resultant.
(1) The Macaulay resultant [Mac1902]. The classical resultant of n + 1 homogeneous polynomials of degrees d0 , . . . , dn coincides both with Elim(d0 ,...,dn ) (Pn )
and with Res(d0 ,...,dn ) (Pn ). This is a consequence of Proposition 1.37(2) and
Corollary 1.33.
(2) Chow forms [CW37]. The Chow form of an irreducible variety V ⊂ Pn of
dimension r coincides both with Elim(1,...,1) (V ) and with Res(1,...,1) (V ). This
follows from Proposition 1.37 and Corollary 1.33.
(3) The GKZ mixed resultant [GKZ94, §3.3]. Let V be a proper irreducible variety over C of dimension r equipped with a family of very ample line bundles L0 , . . . , Lr and RL0 ,...,Lr the (L0 , . . . , Lr )-resultant of V in the sense of
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
21
I. Gelfand, M. Kapranov and A. Zelevinski. Each Li defines an embedding
ψi : V ֒→ Pni . We consider then the map
ψ : V ֒−→ Pn
,
ξ 7−→ (ψ0 (ξ), . . . , ψr (ξ)).
Using Proposition 1.37(2), it can be shown that RL0 ,...,Lr coincides with the
eliminant form Elime0 ,...,er (ψ(V )). Using the formula for the degree of the
GKZ mixed resultant in [GKZ94, Thm. 3.3], we can show that it also coincides
with the resultant Rese0 ,...,er (ψ(V )).
1.4. Operations on resultants. We will now study the behavior of resultants with
respect to basic geometric operations, including intersections, linear projections and
products of cycles.
An important feature of resultants is that they transform the intersection product of
a cycle with a divisor into an evaluation.
m
Proposition 1.40. Let X ∈ Z r (Pn
K ) and d0 , . . . , dr ∈ N \ {0}. Let f ∈ K[x]dr such
that div(f ) intersects X properly. Then there exists λ ∈ K × such that
Resd0 ,d1 ...,dr (X)(u0 , . . . , ur−1 , f ) = λ Resd0 ,...,dr−1 (X · div(f ))(u0 , . . . , ur−1 ),
where the left-hand side denotes the specialization of the last of group of variables of
Resd0 ,d1 ...,dr (X) at the coefficients of f .
Proof. This is [Rem01a, Prop. 3.6].
Next we consider the behavior of resultants with respect to standard projections.
l
Consider the linear projection π : Pn
K 99K PK in (1.15) and let x and y denote the
l
m
r+1 .
multihomogeneous coordinates of Pn
K and of PK , respectively. Let d ∈ (N \ {0})
The general forms of multidegree di in the variables x and y are, respectively,
X
X
Fi =
ui,a y a ∈ K[ui ][y].
ui,a xa ∈ K[ui ][x] , Fi′ =
a∈Nn+1
d
l+1
a∈Nd
i
i
Write
\ Nl+1
u′′i = {ui,a : a ∈ Nn+1
u′i = {ui,a : a ∈ Nl+1
di }.
di
di } ,
Let ≺ be the partial monomial order on K[u0 , . . . , ur ] defined as
{u′i }0≤i≤r ≺ u′′0 ≺ · · · ≺ u′′r .
By this, we mean that the variables in each set have the same weight and that those
in u′′r have the maximal weight, then come those in u′′r−1 , etcetera. Observe that
this order can be alternatively defined as the lexicographic order associated to the
sequence of vectors w0 , . . . , wr defined as
i−1
r−i
}|
{
}|
{
z
z
wi = (0, 0), . . . , (0, 0), (0, 1), (0, 0), . . . , (0, 0) .
Given a polynomial F ∈ K[u] \ {0}, we denote by init≺ (F ) ∈ K[u] \ {0} its initial
part with respect to this order. It consists in the sum of the terms in F whose
monomials are minimal with respect to ≺. This order is multiplicative, in the sense
that init≺ (F G) = init≺ (F )init≺ (G) for all F, G ∈ K[u].
l
Proposition 1.41. Let π : Pn
K 99K PK be the linear projection and ≺ the partial
m
r+1 .
monomial order on K[u] considered above. Let X ∈ Zr+ (Pn
K ) and d ∈ (N \ {0})
Then
Resd (π∗ X) init≺ (Resd (X)) in A[u].
22
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof. It is enough to consider the case π∗ (X) 6= 0, otherwise Resd (π∗ X) = 1. In
addition, we only need to prove that the division holds in K[u], since both polynomials
belong to A[u] and Resd (π∗ X) is primitive.
P
We proceed by induction on the dimension r. For r = 0, write XK = ξ mξ ξ with
ξ ∈ Pn and mξ ∈ Z. Hence,
X
mξ π∗ ξ
π∗ XK =
ξ∈L
/
since π∗ ξ = 0 whenever ξ ∈ L, with L defined in (1.15). Observe that F0′ (π(ξ)) =
init≺ (F0 (ξ)) for each ξ ∈
/ L. Using Corollary 1.38 and the multiplicativity of the
×
order ≺, we deduce that there exist λ, λ′ ∈ K such that
Y
Y
F0 (ξ)mξ = λ init≺ (Resd0 (ξ)).
init≺ (F0 (ξ))mξ init≺
Resd0 (π∗ X) = λ′
ξ
ξ∈L
/
Now let r ≥ 1 and suppose that we have proved the statement for all cycles of pure
dimension up to r − 1 and any base field. Let V be an irreducible K-variety of
dimension r. We suppose that Resd (π∗ V ) 6= 1 because otherwise the statement is
trivial. Thus, the degree of this resultant in some group of variables is ≥ 1 and, up
to a reordering, we can suppose that this holds for the group u′0 . Consider the scalar
extension VK(u′r ) of the K-variety V by the field K(u′r ). By Proposition 1.28, there
exists λ1 ∈ K(u′r )× such that
Resd (π∗ V ) = λ1 Resd (π∗ VK(u′r ) ).
In turn, Proposition 1.40 implies that there exists λ2 ∈ K(u′r )× such that
Resd (π∗ (VK(u′r ) )) = λ2 Resd0 ,...,dr−1 (π∗ (VK(u′r ) ) · divPl (Fr′ ))(u0 , . . . , ur−1 ).
We can verify that the form Fr′ satisfies the hypothesis of Lemma 1.18. Hence, this
lemma implies the equality of cycles π∗ (VK(u′r ) · divPn (Fr′ )) = π∗ (VK(u′r ) ) · divPl (Fr′ ).
In particular,
Resd0 ,...,dr−1 (π∗ (VK(u′r ) ) · divPl (Fr′ )) = Resd0 ,...,dr−1 (π∗ (VK(u′r ) · divPn (Fr′ ))).
Applying the inductive hypothesis to the (r − 1)-dimensional cycle VK(u′r ) · divPn (Fr′ ),
Resd0 ,...,dr−1 (π∗ (VK(u′r ) · divPn (Fr′ ))) init≺ (Resd0 ,...,dr−1 (VK(u′r ) · divPn (Fr′ ))).
The divisor divPn (Fr′ ) intersects VK(u′r ) properly. Proposition 1.40 then implies
Resd0 ,...,dr−1 (VK(u′r ) · divPn (Fr′ ))(u0 , . . . , ur−1 ) = λ3 Resd (V )(u0 , . . . , ur−1 , Fr′ )
for some λ3 ∈ K(u′r )× . This last polynomial is not zero and it satisfies
init≺ (Resd (V )(u0 , . . . , ur−1 , Fr′ )) = init≺ (Resd (V )),
due to the definition of ≺. We conclude
Resd (π∗ V ) init≺ (Resd (V ))
in K(u′r )[u0 , . . . , ur−1 ].
This readily implies that Resd (π∗ V ) | init≺ (Resd (V )) in K[u], because Resd (π∗ V ) is
a power of an irreducible polynomial of positive degree in u0 and r > 0.
For a general K-cycle of pure dimension r, the statement follows by applying this
result to its irreducible components and using the multiplicativity of the order ≺.
This concludes the inductive step.
Remark 1.42. In the projective case (m = 1), this result can be alternatively derived
from [PS93, Prop. 4.1], see also [KPS01, Lem. 2.6].
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
23
Resultants corresponding to general linear forms play an important role in the definition and study of mixed heights of cycles. We introduce a convenient notation for
handling this particular case. Given c = (c1 , . . . , cm ) ∈ Nm we set
c1
(1.43)
cm
z
}|
{
z }| {
e(c) = (e1 , . . . , e1 , . . . , em , . . . , em ) ∈ (Nm \ {0})|c| ,
where ei denotes i-th vector of the standard basis of Rm . For c ∈ Nm
r+1 , both
Elime(c) (V ) and Rese(c) (V ) are polynomials in the coefficients of the r + 1 general
linear forms L0 , . . . , Lr corresponding to the index e(c). In this case, Proposition 1.32
implies that, for 0 ≤ i ≤ r,
(1.44)
degui (Rese(c) (X)) = degc−ej(i) (X),
where j(i) is the index j such that c1 + . . . + cj−1 < i + 1 ≤ c1 + . . . + cj .
i
Proposition 1.45. Let mi ∈ Z>0 , ni ∈ Nmi and Xi ∈ Zri (Pn
K ) for i = 1, 2. Let
m
i
ci ∈ N such that |c1 | + |c2 | = r1 + r2 + 1. Then there exists λ ∈ A× such that
λ Rese(c1 ) (X1 )degc2 (X2 ) if |c1 | = r1 + 1, |c2 | = r2 ,
Rese(c1 ,c2 ) (X1 × X2 ) =
λ Rese(c2 ) (X2 )degc1 (X1 ) if |c1 | = r1 , |c2 | = r2 + 1,
1
otherwise.
Proof. We first prove the statement for the case when K is algebraically closed and
A = K. Let u = {u0 , . . . , ur1 +r2 } be the group of variables associated to the index
e(c1 , c2 ). By (1.44), for each 0 ≤ i ≤ r1 + r2 there is j, 1 ≤ j ≤ m1 + m2 , such that
degui (Rese(c1 ,c2 ) (X1 × X2 )) = deg(c1 ,c2 )−ej (X1 × X2 ).
If either |c1 | ≥ r1 + 2 or |c2 | ≥ r2 + 2, then deg(c1 ,c2 )−ej (X1 × X2 ) = 0 for all j, thanks
to Proposition 1.19. Hence all partial degrees are 0 and Rese(c1 ,c2 ) (X1 × X2 ) = 1.
Consider then the case when |c1 | = r1 + 1, |c2 | = r2 . Again by (1.44), for 1 ≤ i ≤ r2 ,
there exists j > m1 such that
degur
1 +i
(Rese(c1 ,c2 ) (X1 × X2 )) = deg(c1 ,c2 )−ej (X1 × X2 ).
By Proposition 1.19, this mixed degree also vanishes, because |c1 | > r1 . Therefore
Rese(c1 ,c2 ) (X1 × X2 ) ∈ K[u0 , . . . , ur1 ].
(1.46)
Furthermore, suppose that degc2 (X2 ) = 0. In this case, for each 0 ≤ i ≤ r1 there
exists j, 1 ≤ j ≤ m1 , such that
degui (Rese(c1 ,c2 ) (X1 × X2 )) = deg(c1 ,c2 )−ej (X1 × X2 ) = degc1 −ej (X1 ) degc2 (X2 ) = 0.
Hence, Rese(c1 ,c2 ) (X1 × X2 ) = 1 = Rese(c1 ) (X1 )degc2 (X2 ) and the statement holds in
this case. Therefore, we assume that degc2 (X2 ) 6= 0. By linearity, it suffices to prove
the statement for two irreducible varieties Vi ⊂ Pni of dimension ri , i = 1, 2. We
consider first the case when r2 = 0, that is, when V2 = {ξ} is a point. Then
Me(c1 ) (V1 × V2 ) = Me(c1 ) (V1 ) ⊗K K[x2 ]/I(ξ).
Hence, for δ i ∈ Nmi ,
Me(c1 ) (V1 × V2 )δ1 ,δ2 = Me(c1 ) (V1 )δ1 ⊗K (K[x2 ]/I(ξ))δ 2 ≃ Me(c1 ) (V1 )δ 1 .
By the definition of the resultant, there exists λ ∈ K × such that
Rese(c1 ) (V1 × V2 ) = λ Rese(c1 ) (V1 ) = λ Rese(c1 ) (V1 )deg 0 (V2 ) ,
24
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
which proves the statement in this case.
Now let V2 ⊂ Pn2 be an irreducible variety of dimension r2 ≥ 1. Write u′ =
(u0 , . . . , ur1 ) and let ℓi ∈ K[x2 ], 1 ≤ i ≤ r2 , be generic linear forms associated
to e(c2 ) and π2 the projection Pn1 × Pn2 → Pn2 . By (1.46), Rese(c1 ,c2 ) (V1 × V2 ) does
not depend on the groups of variables ur1 +i for 1 ≤ i ≤ r2 . Hence,
Rese(c1 ,c2 ) (V1 × V2 )(u) = Rese(c1 ,c2 ) (V1 × V2 )(u′ , ℓ1 , . . . , ℓr2 )
r2
Y
π2∗ div(ℓi ) (u′ )
= λ Rese(c1 ) (V1 × V2 ) ·
= λ Rese(c1 ) V1 × (V2 ·
i=1
r
2
Y
i=1
Qr2
div(ℓi )) (u′ )
with λ ∈ K × , thanks to Proposition 1.40. The cycle V2 · i=1 div(ℓi ) is of dimension 0
and so we are in the hypothesis of the previous case. By Corollary
1.14, it is a cycle of
Qr2
′
degree degc2 (V2 ). Therefore, Rese(c1 ) V1 ×(V2 · i=1 div(ℓi )) = λ Rese(c1 ) (V1 )degc2 (V2 )
with λ′ ∈ K × , which completes the proof for the case when K is algebraically closed
and A = K.
The case of an arbitrary field K which is the field of fractions of a factorial ring A
follows from Proposition 1.28 and the fact that the resultants of V1 , V2 and V1 × V2
are primitive polynomials in A[u].
2. Heights of cycles of multiprojective spaces
2.1. Mixed heights of cycles over function fields. Throughout this section, we
denote
P by k a field and t = {t1 , . . . , tp } a group of variables. The height of a polynomial
f = a αa xa11 · · · xann ∈ k[t][x1 , . . . , xn ] \ {0} is defined as
h(f ) = degt (f ) = maxa deg(αa ).
For f = 0, we set h(f ) = 0. The following lemma estimates the behavior of the height
of polynomials with respect to addition, multiplication and composition. Its proof
follows directly from the definitions.
Lemma 2.1. Let f1 , . . . , fs ∈ k[t][x1 , . . . , xn ] and g ∈ k[t][y1 , . . . , ys ]. Then
P
(1) h( i fi ) ≤ maxi h(fi );
P
Q
(2) h( i fi ) = i h(fi );
(3) h(g(f1 , . . . , fs )) ≤ h(g) + degy (g) maxi h(fi ).
and study its basic properties.
In the sequel, we extend this notion to cycles of Pn
k(t)
To this end, we specialize the theory in §1 to the case when the factorial ring A is the
polynomial ring k[t] with field of fractions K = k(t). In particular, the resultant of
an effective equidimensional k(t)-cycle is a primitive polynomial in k[t][u].
Definition 2.2. Let V ⊂ Pn
be an irreducible k(t)-variety of dimension r, c ∈ Nm
r+1
k(t)
and e(c) as in (1.43). The (mixed) height of V of index c is defined as
hc (V ) = h(Rese(c) (V )) = degt (Rese(c) (V )).
).
This definition extends by linearity to cycles in Zr (Pn
k(t)
n
For n ∈ N and X ∈ Zr (Pk(t) ), the height of X is defined as h(X) = hr+1 (X).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
25
Definition 2.3. Let η be an indeterminate. The extended Chow ring of Pn
is the
k(t)
graded ring
nm +1
A∗ Pn
; k[t] = A∗ Pn
⊗Z Z[η]/(η 2 ) ≃ Z[η, θ1 , . . . , θm ]/(η 2 , θ1n1 +1 , . . . , θm
),
k(t)
k(t)
i
where θi denotes the class in A∗ Pn
of the inverse image of a hyperplane of Pnk(t)
k(t)
n
n
i
under the projection Pk(t) → Pk(t) . For short, we alternatively denote this ring as
A∗ (Pn ; k[t]). To a cycle X ∈ Zr Pn
we associate an element of this ring, namely
k(t)
X
X
nm −cm
nm −bm
+
[X]k[t] =
hc (X) η θ1n1 −c1 · · · θm
.
degb (X) θ1n1 −b1 · · · θm
c∈Nm
r+1 , c≤n
b∈Nm
r , b≤n
This is a homogeneous element of degree |n| − r.
There is an inclusion of the Chow ring into the extended Chow ring
ı : A∗ Pn
֒→ A∗ Pn
; k[t]
k(t)
k(t)
satisfying [X]k[t] ≡ ı([X]) (mod η). In particular, the class of a cycle in the Chow ring
is determined by its class in the extended Chow ring.
For a cycle X of pure dimension r, we will see in Theorem 2.18(2) that hc (X) = 0 for
every c such that ci > ni for some i. Hence [X]k[t] contains the information about all
mixed degrees and heights, since {θ a , η θ a }a≤n is a Z-basis of A∗ (Pn ; k[t]).
The mixed heights of a k(t)-cycle X can be interpreted as some mixed degrees of
a model of X over Ppk . For simplicity, we will only consider the case of projective
k(t)-cycles, where t is a single variable.
Definition 2.4. Let n ∈ N and V ⊂ Pnk(t) be an irreducible k(t)-variety. Let
s = {s0 , s1 } and x = {x0 , . . . , xn } be groups of variables and I ⊂ k[s, x] the bideg (f )
homogeneous ideal generated by all the polynomials of the form s0 t f (s1 /s0 , x)
for f ∈ I(V ) ∩ k[t, x]. The standard model of V over P1k is defined as the k-variety
V (I) ⊂ P1k ×Pnk . This notion extends by linearity to cycles of Pnk(t) : the standard model
P
P
of a cycle X = V mV V is defined as X = V mV V, where V denotes the standard
model of the irreducible k(t)-variety V .
Remark 2.5. Set V0 and V∞ for the restriction of V to the open subsets (P1k \ {(0 :
1)}) × Pnk and (P1k \ {(1 : 0)}) × Pnk , respectively. These are irreducible k-varieties
which correspond to the prime ideals I0 = I(V ) ∩ k[t][x] and I∞ = I(V ) ∩ k[t−1 ][x],
respectively, and form a covering of V. In particular, V is an irreducible k-variety.
A variety W ⊂ P1k × Pnk is vertical if its projection to P1k consists in a single point. The
following lemma shows that the standard model of a cycle of Pnk(t) of pure dimension
r ≥ 0 is a cycle of P1k × Pnk without vertical components. Moreover, there is a bijection
between Zr (Pnk(t) ) and the set of cycles in Zr+1 (P1k × Pnk) without vertical components.
Lemma 2.6. Let n ∈ N and r ≥ 0.
(1) Let X ∈ Zr (Pnk(t) ). Then its standard model X is a cycle of P1k × Pnk of pure
dimension r + 1, without vertical fibers, and the generic fiber of X → P1k
coincides with X under the natural identification of the generic fiber of P1k ×
Pnk → P1k with Pnk(t) .
(2) Let Y ∈ Zr+1 (P1k × Pnk ) be a cycle without vertical components and Y the
generic fiber of Y → P1k . Then Y ∈ Zr (Pnk(t) ) and Y is its standard model.
26
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof.
(1) It suffices to prove the statement for an irreducible variety V ⊂ Pnk(t) . We keep the
notation in Definition 2.4 and Remark 2.5 and we denote by π the projection V → P1k .
Suppose that V is vertical. This would imply that π(V0 ) is a point. Hence, there exists
f ∈ k[t] \ {0} such that π(V0 ) ⊂ V (f ) or, equivalently, such that f ∈ I0 . But this
would imply that 1 ∈ I(V ) and a fortiori, V = ∅. This is contrary to our assumptions
and so we deduce that V is not vertical.
Choose 0 ≤ i ≤ n such that V is not contained in V (xi ). For simplicity, we suppose
that i = 0. Set Vx0 = V \ V (x0 ) ⊂ Pnk(t) \ V (x0 ) ≃ Ank(t) . Set x′i = xi /x0 , 1 ≤ i ≤ n,
and let J ⊂ k(t)[x′1 , . . . , x′n ] be the ideal of Vx0 . Then J ∩ k[t, x′1 , . . . , x′n ] is the ideal
of the k-variety (V0 )x0 := V0 \ V (x0 ) ⊂ (A1k × Pnk ) \ V (x0 ) ≃ A1k × Ank . Hence,
(V0 )x0 ×A1 Spec(k(t)) = Spec(k[t, x′ ]/I((V0 )x0 ) ⊗k k(t)) = Spec(k(t)[x′ ]/J) = Vx0 .
k
The generic fiber of π coincides with the closure in Pnk(t) of the generic fiber of (V0 )x0 →
A1k , namely, with Vx0 = V . The fact that π is surjective with generic fiber V implies
that dim(V) = dim(V ) + dim(P1k ) = r + 1 by the theorem of dimension of fibers.
(2) It is sufficient to consider the case of an irreducible k-variety W ⊂ P1k × Pnk without
vertical components. Let J0 ⊂ k[t][x] be the ideal of the restriction of W to the open
subset (P1k \ {(0 : 1)}) × Pnk ≃ A1k × Pnk . It is a prime ideal of Krull dimension r + 2
and J0 ∩ k[t] = {0} since π is surjective. Hence, J := k(t) ⊗k[t] J0 is a homogeneous
prime ideal of Krull dimension r + 1 which defines the generic fiber of π. Moreover,
J0 = J ∩ k[t][x] and so W is the standard model of W .
Lemma 2.7. Let I ⊂ k(t)[x] be an equidimensional ideal, X(I) ∈ Z(Pnk(t) ) its associated cycle and X the standard model of X(I). Let I ⊂ k[s, x] be the ideal generated by
deg (f )
all the polynomials of the form s0 t f (s1 /s0 , x) for f ∈ I ∩ k[t, x]. Then X = X(I).
Proof. This can be verified by going through the ideals: the minimal primes of I are in
bijection with the minimal primes of I, and this bijection preserves multiplicities.
There is an isomorphism φ : A∗ (Pnk(t) ; k[t]) → A∗ (P1k × Pnk ) which sends the generators
η, θ1 ∈ A∗ (Pnk ; k[t]) to the generators θ1 , θ2 ∈ A∗ (P1k × Pnk ), respectively. Next result
shows that, via this isomorphism, the class of a projective k(t)-cycle X identifies with
the class of its standard model X . In particular, the height of X coincides with a
mixed degree of X .
Proposition 2.8. Let X ∈ Zr (Pnk(t) ) and X be the standard model of X. Then
φ [X]k(t) = [X ].
Equivalently, deg(X) = deg1,r (X ) and h(X) = deg0,r+1 (X ).
Proof. It is enough to prove the statement for an irreducible k(t)-variety V with
standard model V. Set u = {u0 , . . . , ur }, 1 = (1, . . . , 1) ∈ Nr+1 , e = (1, 0) and
e
e = ((0, 1), . . . , (0, 1)) ∈ (N2 )r+1 . We first claim that there exists λ ∈ k× such that
(2.9)
Res1 (V )(u) = λ Rese,ee (V)((−t, 1), u).
Let I ⊂ k(t)[x] and I ⊂ k[s, x] denote the homogeneous ideal of V and the bihomogeneous ideal of V, respectively. Set
J1 = k(t) ⊗k I + (s1 − ts0 ),
J2 = k(t)[s] ⊗k(t) I + (s1 − ts0 )
⊂ k(t)[s, x].
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
27
These ideals define the subschemes Vk(t) ∩ V (s1 − ts0 ) and {(1 : t)} × V of P1k(t) × Pnk(t) ,
respectively. For f ∈ k[t, x], set
s
deg (f )
1
, x ∈ k[s, x].
f hom = s0 t f
s0
This polynomial is homogeneous of degree degt (f ) with respect to the variables s.
Observe that
(2.10)
degt (f )
f hom ≡ s0
f ≡ (t−1 s1 )degt (f ) f
(mod (s1 − ts0 )).
The ideal I is generated by {f hom : f ∈ I ∩ k[t, x]}. By (2.10), these generators
lie in J2 ; hence I ⊂ J2 and so J1 ⊂ J2 . The equations (2.10) also imply that I is
contained in the localization (J2 )si ⊂ k(t)[s, x]si for i = 0, 1. Hence, (J1 )si = (J2 )si
and so both ideals define the same subschemes of P1k(t) ×Pnk(t) . In particular, div(s1 −ts0 )
intersects Vk(t) properly and we have the equality of cycles
(2.11)
Vk(t) · div(s1 − ts0 ) = {(1 : t)} × V ∈ Zr (P1k(t) × Pnk(t) ).
Applying successively propositions 1.28 and 1.40, (2.11) and Proposition 1.45, we
obtain that there exists µ0 , µ ∈ k[t] \ {0} such that
(2.12) Rese,ee (V)((−t, 1), u) = µ0 Resee (Vk(t) · div(s1 − ts0 ))(u)
= µ0 Resee ({(1 : t)} × V )(u) = µ Res1 (V )(u).
It remains to show that µ ∈ k× . For τ ∈ k, div(s1 − τ s0 ) intersects V properly since
the projection π : V → P1k is surjective (Lemma 2.6). Proposition 1.40 then implies
that Rese,ee (V)((−τ, 1), u) 6= 0. Furthermore, Res1 (V )(u) t=τ 6= 0 as this resultant is a
primitive polynomial in k[t][u]. Specializing (2.12) at t = τ , we deduce that µ(τ ) 6= 0
for all τ ∈ k. Hence, µ ∈ k× and (2.9) follows for λ = µ−1 .
Let v = {v0 , v1 } be a group of variables. By (2.9), for any 0 ≤ i ≤ r,
degui Res1 (V )(u) = degui Rese,ee (V)((−t, 1), u) = degui Rese,ee (V)(v, u) ,
since Rese,ee (V) is homogeneous in the variables v. Proposition 1.32 then implies that
deg(V ) = deg1,r (V).
Applying an argument similar to the one above, we verify that s0 intersects V properly.
Hence, Rese,ee (V)((1, 0), u) = λ Resee (V · div(s0 ))(u) with λ ∈ k× . In particular, this
specialization is not zero, and so the degrees of Rese,ee (V) in v0 and in v coincide.
Therefore,
h(V ) = degt Res1 (V )(u) = degt Rese,ee (V)((−t, 1), u) = degv Rese,ee (V)(v, u) .
By Proposition 1.32, we conclude that h(V ) = deg0,r+1 (V).
Next proposition collects some basic properties of mixed heights and classes in the
extended Chow ring of multiprojective cycles over k(t) with t = {t1 , . . . , tp }.
Proposition 2.13.
(1) Let X ∈ Zr+ Pn
. Then [X]k[t] ≥ 0. In particular, hc (X) ≥ 0 for all
k(t)
c ∈ Nm
.
r+1
(2) Let X ∈ Zr (Pn
k ). Then [Xk(t) ]k[t] = ı([Xk(t) ]) or equivalently, hc (X) = 0 for
m
] = 1.
all c ∈ Nr+1 . In particular, [Pn
k(t) k[t]
28
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
(3) Let ξ = (ξ 1 , . . . , ξ m ) ∈ Pn be a point with coordinates in k(t) and for each
1 ≤ i ≤ m write ξ i = (ξi,j )j for coprime polynomials ξi,j ∈ k[t]. Then
[ξ]k[t] =
m
X
h(ξ i ) η θ n−ei + θ n ,
i=1
with h(ξ i ) := maxj h(ξi,j ) = maxj deg(ξi,j ). In particular, hei (ξ) = h(ξ i ).
(4) Let D ∈ Div+ Pn
and fD ∈ k[t][x] its primitive defining polynomial. Then
k(t)
[D]k[t] = h(fD ) η +
m
X
degxi (fD ) θi .
i=1
In particular, hn (D) = h(fD ).
Proof.
(1) This is immediate from the definition of mixed heights and classes in the extended
Chow ring.
(2) By Proposition 1.28, there exists λ ∈ k(t)× such that Rese(c) (Xk(t) ) = λ Rese(c) (X).
Since the term in the left-hand side is primitive with respect to k[t], we deduce that
λ ∈ k× . Hence, this resultant does not depend on t and hc (X) = 0 for all c. This
readily implies that [Xk(t) ]k[t] = ı([Xk(t) ]). The rest of the statement follows from
Proposition 1.10(2).
(3) By Proposition 1.10(3), it is enough to show that hei (ξ) = h(ξ i ). By Corollary 1.38,
there exists λi ∈ k(t)× such that Resei (ξ) = λi Li (ξ i ), where Li is the general linear
form of multidegree ei . Indeed, λi ∈ k× since Resei (ξ) and Li (ξ i ) are primitive
polynomials in k[t]. Hence, hei (ξ) = degt (Li (ξ i )) = h(ξ i ).
(4) By Proposition 1.10(4), it is enough to show that hn (D) = h(fD ). Set d =
deg(fD ). Consider the general form F of multidegree d and the general linear forms
L = (L0 , . . . , L|n|−1 ) corresponding to e(n). Write for short R = Rese(n),d (Pn
k ). Using
×
propositions 1.40 and 1.28, we deduce that there exists λ ∈ k(t) such that
λ Rese(n) (D)(L) = R(L, fD ).
Observe that λ ∈ k[t] \ {0} because the resultant in the left-hand side is a primitive
polynomial. On the other hand, let H ⊂ Apk be an irreducible k-hypersurface. The
fact that fD is primitive implies that there exists ξ ∈ H such that fD (ξ, x) 6= 0.
Applying Proposition 1.40 to the cycle Pn
k and fD (ξ, x) ∈ k[x] we deduce
R(L, fD (ξ, x)) 6= 0.
Hence, λ(ξ) 6= 0. This implies that V (λ) contains no hypersurface of Apk and so
λ ∈ k× . Hence,
hn (D) = degt Rese(n) (D) = degt R(L, fD ) = degt (fD ) degun (R) = h(fD ),
since degt (fD ) = h(fD ) and, by propositions 1.32 and 1.10(2), degun (R) = 1.
We recall some notions and properties of valuations of fields. Let T0 , . . . , Tp denote
the standard homogeneous coordinates of Ppk . If we identify each variable ti with the
rational function Ti /T0 , we can regard k(t) as the field of rational functions of this
projective space: given α ∈ k(t), then α(T1 /T0 , . . . , Tp /T0 ) is homogeneous of degree
0 and defines a rational function on Ppk .
Given an irreducible k-hypersurface H of Pp and α ∈ k(t), we set ordH (α) for the
order of vanishing of α along H. The map ordH : k(t) → Z is a valuation of k(t).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
29
If H∞ := V (T0 ) is the hyperplane at infinity, then ordH∞ (α) = − deg(α) where
deg(α) = deg(α1 ) − deg(α2 ) for any αi ∈ k[t] such that α = α1 /α2 . If H 6= H∞ , then
ordH (α) coincides with the order of the polynomial fH (1, t1 , . . . , tp ) in the factorization
of α, where fH ∈ k[T0 , . . . , Tp ] is the defining polynomial of H.
Let K be an arbitrary field and v a valuation of K. For a polynomial f over K, we
set v(f ) for the minimum of the valuation of its coefficients. Gauss Lemma says that
for any given polynomials f, g over K,
v(f g) = v(f ) + v(g).
Given a finite extension E of K, there exists a (non-necessarily unique) valuation w
of E extending v.
The height of an arbitrary resultant can be expressed in terms of mixed heights:
Lemma 2.14. Let X ∈ Zr+ Pn
and d ∈ (Nm \ {0})r+1 . Then
k(t)
m
r X
Y
degt (Resd (X)) = coeff ηθ n [X]k[t]
di,j θj .
i=0 j=1
m
In particular, for b ∈ Nm
r and dr ∈ N \ {0},
(2.15)
degt (Rese(b),dr (X)) =
m
X
dr,j hb+ej (X).
j=1
Proof. Write d = (d0 , . . . , dr ) for di ∈ Nm \ {0}. We claim that
(2.16)
degt (Resd0 ,...,dr (X)) =
m
X
d0,j degt (Resej ,d1 ,...,dr (X)).
j=1
We first consider the case r = 0. Let Xk(t) =
P
ξ
mξ ξ with ξ ∈ Pn and mξ ≥ 1. For
each of these points write ξ = (ξi,j )i,j for some ξi,j ∈ k(t). Let F0 and Lj denote the
general forms of degree d0 and ej as in (1.21), respectively. By Proposition 1.28 and
×
Corollary 1.38, there exist µ, λj ∈ k(t) such that
Y
Y
Resd0 (X)(F0 ) = µ
F0 (ξ)mξ , Resej (X)(Lj ) = λj
Lj (ξ j )mξ .
ξ
ξ
Picking a suitable choice of multihomogeneous coordinates for the ξ’s, it is possible
to set λj = 1 for all j. We have
Y
m
m
m
Y
YY
d0,j
Resej (X)(Lj )d0,j .
Lj
Lj (ξ)d0,j mξ = µ
=µ
Resd0 (X)
j=1
j=1
ξ j=1
We deduce that µ ∈ k(t)× since all considered resultants have coefficients in k[t].
Now let E be a sufficiently large finite extension of k(t) containing all the chosen
coordinates ξi,j . For any valuation v of E,
(2.17)
v(F0 (ξ)) = min v(ξ a ) =
|a|=d0
m
X
j=1
d0,j min v(ξj,ℓ ) =
ℓ
m
X
d0,j v(Lj (ξ j )).
j=1
Let H be an irreducible k-hypersurface of Pp different from H∞ , and vH a valuation
of E extending ordH . Since Resd0 (X) and Resej (X) are primitive with respect to k[t],
30
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
we have that vH (Resd0 (X)) = vH (Resej (X)) = 0. By Gauss Lemma,
Y
X
ordH (µ) = −ordH
F0 (ξ)mξ = −
mξ vH (F0 (ξ)),
ξ
0 = ordH
Y
mξ
Lj (ξ j )
ξ
=
X
ξ
mξ vH (Lj (ξ j )).
ξ
Applying (2.17),
ordH (µ) = −
X
ξ
mξ vH (F0 (ξ)) = −
m
X
d0,j
j=1
X
ξ
mξ vH (Lj (ξ j )) = 0.
Since this holds for every H 6= H∞ , we deduce that µ ∈ k× . Now let v∞ be a valuation
of E extending ordH∞ . Applying again Gauss Lemma and (2.17),
X
ordH∞ (Resd0 (X)) =
mξ v∞ (F0 (ξ))
ξ
=
m
X
d0,j
j=1
Hence degt (Resd0 (X)) =
Pm
X
j=1 d0,j
ξ
m
X
d0,j ordH∞ (Resej (X)).
mξ v∞ (Lj (ξ j )) =
j=1
degt (Resej (X)), which concludes the case r = 0.
The case r ≥ 1 follows by reduction to the zero-dimensional case. For 1 ≤ i ≤ r,
let Fi be the generic polynomial of multidegree di and ui the variables corresponding
to its coefficients. Set e
k = k(u1 , . . . , ur ). Observe that div(Fj ) intersects properly
Q
Qj−1
the cycle Xek(t) · i=1 div(Fi ), 1 ≤ j ≤ r. Therefore, Xek(t) · ri=1 div(Fi ) is a cycle of
dimension 0 and, by Proposition 1.40, there exists λ ∈ e
k(t)× such that
r
Y
div(Fi ) ,
Resd (X) = λ Resd0 Xek ·
i=1
Since both resultants are primitive with respect to e
k[t], λ ∈ e
k× and so these resultants
have the same t-degree. Analogous relations hold for Resej ,d1 ,...,dr (X), 1 ≤ j ≤ m.
Hence, (2.16) follows by applying the previously considered zero-dimensional case.
From Proposition 1.27, we deduce that the map
(Nm \ {0})r+1 −→ Z ,
d 7−→ degt (Resd (X))
is multilinear with respect
to the variables
d0 , . . . , dr . The same holds for the map
Q P
d 7→ coeff ηθ n [X]k[t] ri=0 m
.
Both
maps coincide when d = e(c) for c ∈
d
θ
i,j
j
j=1
because
Nm
r+1
degt (Rese(c) (X)) = hc (X) = coeff η θn (θ c [X]k[t] ).
Since the family {e(c)}c is a basis of the semigroup (Nm \{0})r+1 , both maps coincide
for all d. This completes the proof of the statement.
The following is an arithmetic version of Bézout’s theorem for multiprojective cycles
over k(t).
Theorem 2.18. Let X ∈ Zr Pn
and f ∈ k[t][x1 , . . . , xm ] a polynomial, multihok(t)
mogeneous in the variables x, such that X and div(f ) intersect properly.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
(1) If X is effective, then for any b ∈ Nm
r
hb (X · div(f )) ≤ h(f ) degb (X) +
m
X
31
degxi (f )hb+ei (X).
i=1
(2) hc (X) = 0 for any c ∈ Nm
r+1 such that ci > ni for some i.
(3) If X is effective, then [X · div(f )]k[t] ≤ [X]k[t] · [div(f )]k[t] .
Proof.
(1) Set d = deg(f ) and let u = (u0 , . . . , ur−1 , ur ) be the groups of variables corresponding to (e(b), d). By Proposition 1.40, there exists λ ∈ k(t)× such that
Rese(b),d (X)(u0 , . . . , ur−1 , f ) = λ Rese(b) (X · div(f ))(u0 , . . . , ur−1 ).
Indeed, λ ∈ k[t] \ {0} because the resultant in the right-hand side is primitive. Hence,
degt Rese(b) (X · div(f )) ≤ degt Rese(b),d (X)(u0 , . . . , ur−1 , f )
≤ degt (f ) degur (Rese(b),d (X)) + degt (Rese(b),d (X)).
From the definition of the height and Proposition 1.32, we deduce that hb (X ·div(f )) ≤
h(f ) degb (V ) + degt (Rese(b),d (X)). The statement follows then from (2.15).
(2) It is enough to prove the statement for a k(t)-variety V . We first consider the
case when V is determined by a regular sequence. We proceed by induction on its
codimension. Let fj , 1 ≤ j ≤ |n| − r, be a regular sequence of multihomogeneous
polynomials. For r ≤ ℓ ≤ |n|, set
|n|−ℓ
Yℓ =
Y
j=1
.
div(fj ) ∈ Zℓ+ Pn
k(t)
For r = |n|, we have Y|n| = Pn and Proposition 2.13(2) implies that hc (Y|n| ) = 0 for
all c ∈ N|n|+1 . Suppose now that r < |n| and that the statement holds for Yr+1 . Let
c ∈ Nm
r+1 such that ci > ni for some i. By Proposition 2.13(1) and item (1) above,
0 ≤ hc (Yr ) ≤ h(f|n|−r ) degc (Yr+1 ) +
m
X
degxj (f|n|−r )hc+ej (Yr+1 ).
j=1
The inductive hypothesis together with the fact that degc (Yr+1 ) = 0 imply that the
right-hand side of this inequality vanishes, and hence hc (Yr ) = 0.
For the general case, consider |n| − r generic linear combinations of a system of
generators of I(V ). The obtained polynomials form a regular sequence and define
a variety Yr such that Yr − V is effective. By the previous analysis, 0 ≤ hc (V ) ≤
hc (Yr ) = 0, hence hc (V ) = 0 as stated.
(3) This is a direct consequence of (1) and (2) together with Theorem 1.11 and Proposition 2.13(4).
Corollary 2.19. Let V ⊂ Pn
be a k(t)-variety of pure dimension r and f ∈
k(t)
k[t][x1 , . . . , xm ] a multihomogeneous polynomial. Let W denote the union of the components of dimension r − 1 of the intersection V ∩ V (f ). Then
[W ]k[t] ≤ [V ]k[t] · [V (f )]k[t] .
P
m
In particular, hb (W ) ≤ h(f ) degb (V ) + m
i=1 degxi (f )hb+ei (V ) for all b ∈ Nr .
32
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof. Let V ′ ⊂ Pn
be the union of the components of V not contained in |div(f )|.
k(t)
′
Then W ⊂ V ∩ V (f ) and V (f ) intersects V ′ properly. By Theorem 2.18(3),
[W ]k(t) ≤ [V ′ · V (f )]k(t) ≤ [V ′ ]k(t · [V (f )]k(t) ≤ [V ]k(t · [V (f )]k(t) .
The last statement follows from this inequality when b ≤ n and from Theorem 2.18(2)
otherwise.
Corollary 2.20. Let n ∈ Z>0 , X ∈ Zr+ Pnk(t) and fj ∈ k[t][x0 , . . . , xn ] \ k[t] a family
Q
of s ≤ r polynomials homogeneous in the variables x0 , . . . , xn such that X· i−1
j=1 div(fj )
and div(fi ) intersect properly for 1 ≤ i ≤ s. Then
s
s
s
X
Y
Y
degt (fℓ )
.
degx (fj ) h(X) + deg(X)
div(fj ) ≤
h X·
degx (fℓ )
j=1
j=1
Proof. Set Y = X ·
Qs
j=1 div(fj ).
ℓ=1
We have
[Y ]k[t] = h(Y ) η θ n−r+s−1 + deg(Y ) θ n−r+s , [X]k[t] = h(X) η θ n−r−1 + deg(X) θ n−r .
Applying recursively Theorem 2.18(3),
[Y ]k[t] ≤ [X]k[t] ·
s
Y
[div(fj )]k[t] .
j=1
The statement follows by comparing the coefficients corresponding to the monomial
η θ n−r+s−1 in the above inequality.
Next result shows that, for projective k(t)-cycles, the inequality in Theorem 2.18(1)
is an equality in the generic case.
Proposition 2.21. Let X ∈ Zr (Pnk(t) ) with k an algebraically closed field, t a single variable and n, r ≥ 1. Let X ⊂ P1k × Pnk be the standard model of X and
ℓ ∈ k[x0 , . . . , xn ] a generic linear form. Then
(1) X · div(ℓ) is the standard model of X · div(ℓ);
(2) deg(X · div(ℓ)) = deg(X) and h(X · div(ℓ)) = h(X).
Proof.
(1) By Lemma 2.7, it is enough to prove the claim for an irreducible k(t)-variety V of
dimension r ≥ 1. Let V ⊂ P1k × Pnk be the standard model of V . By Lemma 2.6(2), it
suffices to prove that V · div(ℓ) is the generic fiber of V · div(ℓ) and that the support of
V · div(ℓ) has no vertical components. Consider the projection ̟ : V → Pnk . We have
1 ≤ r = dim(V) − dim(P1k ) ≤ dim(̟(V)) ≤ dim(V) = r + 1.
If dim(̟(V)) = 1, then dim(V ) = 1 and ̟ −1 (ξ) = P1k × {ξ} for all ξ ∈ ̟(V) because
of the theorem of dimension of fibers. Hence, in this case,
V = P1k × ̟(V).
Therefore, V · div(ℓ) = P1k × (̟(V) · div(ℓ)) has no vertical components. Moreover,
by considering the generic fiber of π : V → P1k , it follows from Lemma 2.6(1) that
V = ̟(V)k(t) . Hence, the generic fiber of V · div(ℓ) over P1k coincides with V · div(ℓ),
which proves the claim for dim(̟(V)) = 1.
If dim(̟(V)) ≥ 2, V ∩ V (ℓ) is an irreducible k-variety of dimension r by [Jou83,
Thm. 6.3(4)]. Moreover, the induced projection πℓ : V · div(ℓ) → P1k is surjective.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
33
Indeed, for τ ∈ P1 (k), we have that V ∩ π −1 (τ ) is a projective variety of dimension
r ≥ 1 and so
V ∩ V (ℓ) ∩ π −1 (τ ) = V ∩ π −1 (τ ) ∩ V (ℓ) 6= ∅.
Hence, the projection πℓ has no vertical fibers, as it is surjective and the support
of V · div(ℓ) is irreducible. By going through the ideals of definition, we can verify
that its generic fiber coincides with V · div(ℓ), which completes the proof of the claim.
(2) The statement concerning the degree follows from Theorem 1.11. For the height,
h(X) = deg0,r+1 (X ) = deg0,r (X · div(ℓ)) = h(X · div(ℓ)).
The first and third equalities follow from Proposition 2.8 and the second one follows
from Theorem 1.11.
We next show that mixed heights are monotonic with respect to linear projections.
We keep the notation from Proposition 1.16. In particular, we consider the linear
l
∗ l
∗ n
projection π : Pn
K 99K PK in (1.15) and the inclusion : A (P ; k[t]) ֒→ A (P ; k[t])
n−l
defined by (P ) = θ
P.
Proposition 2.22. Let π : Pn
99K Plk(t) be the linear projection defined in (1.15)
k(t)
and X ∈ Zr+ Pn
. Then
k(t)
[π∗ X]k[t] ≤ [X]k[t] .
In particular, hc (π∗ X) ≤ hc (X) for all c ∈ Nm
r+1 .
Proof. The statement is equivalent to the inequalities degb (π∗ X) ≤ degb (X) and
hc (π∗ X) ≤ hc (X) for all b, c. Because of Proposition 1.16, we only need to prove the
latter. Let c ∈ Nm
r+1 . By Proposition 1.41, Rese(c) (π∗ X) divides init≺ (Rese(c) (X)) in
k[t][u]. We deduce that
hc (π∗ X) = degt (Rese(c) (π∗ X)) ≤ degt (Rese(c) (X)) = hc (X),
which proves the statement.
The following result gives the behavior of extended Chow rings and classes with respect
to products.
Proposition 2.23. Let mi ∈ Z>0 and ni ∈ Nmi for i = 1, 2. Then
(1) A∗ (Pn1 × Pn2 ; k[t]) ≃ A∗ (Pn1 ; k[t]) ⊗Z[η] A∗ (Pn2 ; k[t]).
i
(2) Let Xi ∈ Zr+i Pn
for i = 1, 2. The above isomorphism identifies [X1 × X2 ]k[t]
k(t)
with [X1 ]k[t] ⊗ [X2 ]k[t] . In particular, for ci ∈ Nmi such that |c1 | + |c2 | =
r1 + r2 + 1,
degc2 (X2 ) hc1 (X1 ) if |c1 | = r1 + 1, |c2 | = r2 ,
degc1 (X1 ) hc2 (X2 ) if |c1 | = r1 , |c2 | = r2 + 1,
h(c1 ,c2 ) (X1 × X2 ) =
0
otherwise.
Proof. (1) is immediate from the definition of the extended Chow ring while (2) follows
directly from Proposition 1.45.
Finally, we compute the class in the extended Chow ring of the ruled join of two
projective varieties. Let n1 , n2 ∈ N and consider the Z-linear map i : A∗ (Pni ; k[t]) ֒→
A∗ (Pn1 +n2 +1 ; k[t]) defined by i (θ l η b ) = θ l η b for 0 ≤ l ≤ ni and b = 0, 1.
34
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
i
Proposition 2.24. Let Xi ∈ Zri Pnk(t)
for i = 1, 2. Then
[X1 #X2 ]k[t] = 1 ([X1 ]k[t] ) · 2 ([X2 ]k[t] ).
In particular, h(X1 #X2 ) = deg(X1 ) h(X2 ) + deg(X2 ) h(X1 ).
We need some lemmas for the proof of this result. The first of them deals with
0-dimensional cycles. For i = 1, 2 and j = 0, 1, let xi = {xi,0 , . . . , xi,ni } be the
(i)
(i)
i
homogeneous coordinates of Pnk(t)
and uj = {uj,l }0≤l≤ni a group of ni + 1 variables.
(i)
Write Lj for the general linear form in the variables xi .
i ) for i = 1, 2. Then h(X #X ) = deg(X ) h(X ) +
Lemma 2.25. Let Xi ∈ Z0 (Pnk(t)
1
2
1
2
deg(X2 ) h(X1 ).
Q
×
(i)
Proof. For i = 1, 2, write Res1 (Xi ) = λi ξi L0 (ξ i )mξi with λi ∈ k(t) , ξ i ∈ Pni (k(t))
and mξi ∈ Z. We claim that there exists ν ∈ k(t)× such that
Y
m m
deg(X2 ) deg(X1 )
(1)
(2)
(1)
(2)
Res1,1 (X1 #X2 ) = ν λ1
λ2
L0 (ξ 1 )L1 (ξ 2 )− L1 (ξ 1 )L0 (ξ 2 ) ξ1 ξ2 .
ξ1 ,ξ2
Indeed, for each ξ 1 ∈ Pn1 (k(t)) and ξ 2 ∈ Pn2 (k(t)), Proposition 1.37(1) and Corol(1)
(2)
(1)
(2)
lary 1.33 imply that Res1,1 (ξ 1 #ξ 2 ) is equal to L0 (ξ 1 )L1 (ξ 2 ) − L1 (ξ 1 )L0 (ξ 2 ) up
×
×
to a constant in k(t) . We deduce that there exists µ ∈ k(t) such that
Y
m m
(1)
(2)
(1)
(2)
(2.26)
Res1,1 (X1 #X2 ) = µ
L0 (ξ 1 )L1 (ξ 2 ) − L1 (ξ 1 )L0 (ξ 2 ) ξ1 ξ2 .
ξ1 ,ξ2
(1)
By setting u1 = 0 in this equality and comparing its right-hand side with the explicit
expression for Res1 (Xi ) plus the fact that Res1,1 (X1 #X2 ) has coefficients in k(t), we
deg(X2 ) deg(X1 )
get that µ = ν λ1
λ2
with ν ∈ k(t)× .
Let E be a sufficiently large extension of k(t), H ⊂ Ppk a hypersurface and v a valuation
of E extending ordH . For i = 1, 2 and ξ i = (ξi,j )j ∈ E ni +1 set v(ξ i ) = min{v(ξi,j )}.
Observe that
(2)
(1)
(2)
(1)
(2)
(1)
v L0 (ξ 1 )L1 (ξ 2 ) − L1 (ξ 1 )L0 (ξ 2 ) = v(ξ 1 ) + v(ξ 2 ) = v(L0 (ξ 1 )) + v(L0 (ξ 2 )).
Therefore, applying Gauss’ lemma, we obtain that ordH (Res1,1 (X1 #X2 )) is equal to
X
ordH (ν) + deg(X2 )v(λ1 ) + deg(X1 )v(λ2 ) +
mξ1 mξ2 v(ξ 1 ) + v(ξ 2 )
ξ1 ,ξ2
X
X
=ordH (ν) + deg(X2 ) v(λ1 ) +
mξ1 v(ξ 1 ) + deg(X1 ) v(λ2 ) +
mξ2 v(ξ 2 )
ξ1
ξ2
= ordH (ν) + deg(X1 ) ordH (Res1 (X2 )) + deg(X2 ) ordH (Res1 (X1 )).
Let H∞ be the hyperplane at infinity. For H 6= H∞ , ordH (Res1,1 (X1 #X2 )) =
ordH (Res1 (Xi )) = 0 and the identity above implies that ordH (ν) = 0 in this case.
Since this holds for all H 6= H∞ , it follows that ν ∈ k× . Hence, ordH∞ (ν) = 0. For
H = H∞ in the same identity, we get
ordH∞ (Res1,1 (X1 #X2 )) = deg(X1 ) ordH∞ (Res1 (X2 )) + deg(X2 )ordH∞ (Res1 (X1 )).
This implies the statement.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
35
Lemma 2.27. Let k be an algebraically closed field and t a single variable. Let
i
V ⊂ P1k × Pkn1 +n2 +1 be the standard model of V1 #V2 , where each Vi ⊂ Pnk(t)
is an
irreducible k(t)-variety and dim(V2 ) ≥ 1. Let ℓ ∈ k[x2 ] be a generic linear form. Then
V · div(ℓ) is the standard model of V1 #(V2 · div(ℓ)).
Proof. By Lemma 2.6(2), it suffices to prove that V1 #(V2 · div(ℓ)) is the generic fiber
of V · div(ℓ) and that the support of the cycle V ·div(ℓ) has no vertical components.
Let Ii,0 = I(Vi ) ∩ k[t][xi ] and I0 = I(V1 ) + I(V2 ) ∩ k[t][x1 , x2 ] be the prime ideals of
Vi,0 and of V0 respectively, following the notation in Remark 2.5. Then I0 ⊃ I1,0 +I2,0 .
In addition, there is an isomorphism of k[t]-algebras
k[t][x1 , x2 ]/(I1,0 + I2,0 ) ≃ k[t][x1 ]/I1,0 ⊗k[t] k[t][x2 ]/I2,0 .
Hence I1,0 + I2,0 is a prime ideal of Krull dimension r1 + r2 + 2. By Remark 2.5
and Lemma 2.6(1), I0 is also a prime ideal of Krull dimension r1 + r2 + 2. Hence,
I0 = I1,0 + I2,0 . This implies that I0 + (ℓ) = I1,0 + I2,0 + (ℓ) and so
k[t][x1 , x2 ]/(I0 + (ℓ)) ≃ k[t][x1 ]/I1,0 ⊗k[t] k[t][x2 ]/(I2,0 + (ℓ)).
Let τ ∈ P1 (k) \ {(0 : 1)} and write τ = (1 : τ ) with τ ∈ k. The k[t]-algebra
k[t][x1 , x2 ]/(I0 + (ℓ)) ⊗k[t] k(τ ) is isomorphic to
k[t][x1 ]/I1,0 ⊗k[t] k(τ ) ⊗k[t] k[t][x2 ]/(I2,0 + (ℓ)) ⊗k[t] k(τ )
The cycle X(I0 + (ℓ)) coincides with V0 · div(ℓ) because of the definition of the intersection product. Since dim(V2 ) ≥ 1, Proposition 2.21(1) and Lemma 2.7 imply that
the cycle X(I2,0 + (ℓ)) coincides with the standard model of V2 · div(ℓ).
We also deduce that dim(π −1 (τ )) = r1 + r2 = dim(V0 · div(ℓ)) − 1, and so there are
no vertical components. The case when τ = (0 : 1) can be treated in the same way
by considering I∞ and Ii,∞ instead of I0 and Ii,0 , respectively. Finally, applying the
isomorphism above to the generic point of P1k , we deduce that the generic fiber of
π : V · div(ℓ) → P1k coincides with V1 #(V2 · div(ℓ)). This concludes the proof.
Proof of Proposition 2.24. The statement is equivalent to the equalities
deg(X1 #X2 ) = deg(X1 ) deg(X2 ) ,
h(X1 #X2 ) = deg(X1 ) h(X2 ) + deg(X2 ) h(X1 ).
The first one is (1.20) and so we only need to prove the second. It is enough to prove
this equality for irreducible k(t)-varieties V1 , V2 over an algebraically closed field k.
Suppose first that the group of parameters consists in a single variable t. In this case,
the proof will be done by induction on the dimension of the Vi ’s. Write ri = dim(Vi ).
The case r1 = r2 = 0 is covered by Lemma 2.25, so we assume that either r1 or r2
are not 0. By symmetry, we can suppose that r2 ≥ 1. Let Vi and V be the standard
models of Vi and of V1 #V2 , respectively, and ℓ ∈ k[x2 ] a generic linear form. Then,
by Proposition 2.8, Theorem 1.11, Lemma 2.27 and again Proposition 2.8,
h(V1 #V2 ) = deg0,r1 +r2 +2 (V) = deg0,r1 +r2 +1 (V · div(ℓ)) = h(V1 #(V2 · div(ℓ))).
Using the inductive hypothesis and Proposition 2.21(2), we get
h(V1 #(V2 · div(ℓ))) = deg(V1 ) h(V2 · div(ℓ)) + deg(V2 · div(ℓ)) h(V1 )
= deg(V1 ) h(V2 ) + deg(V2 ) h(V1 ),
which proves the statement for this case.
The case of an arbitrary number p of parameters reduces to the previous one as follows.
tp −γp
1
Given an irreducible subvariety W ⊂ Pnk(t) , consider the field e
k = k t1 −γ
w ,..., w
where w is an additional variable and γi ∈ k is a generic element. Observe that e
k(w)
36
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
coincides with k(t)(w) and so it is a transcendental extension of k(t). The scalar
extension Wk(w)
⊂ Pne is an irreducible subvariety of the same dimension and degree
e
k(w)
as W . Let Res1 Wk(w)
be the Chow form of Wk(w)
, primitive with respect to the base
e
e
e
ring k[w].
By Proposition 1.28, there exists λ ∈ e
k(w)× such that
t − γ
t p − γp
1
1
w
+
γ
,
.
.
.
,
w
+
γ
(w)
=
Res
(W
)(t)
=
Res
(W
)
λ Res1 Wk(w)
1
p .
1
1
e
w
w
is primitive with respect
Indeed, Res1 (W ) is a polynomial over e
k[w] and Res1 W
e
k(w)
to this ring. Hence, λ ∈ e
k[w] \ {0}. Furthermore, let ζ be a root of λ in the algebraic
e
closure of k. If ζ 6= 0, setting w = ζ in the equality above we get Res1 (W ) = 0, which
is impossible. If ζ = 0, we get 0 = Res1 (W )(γ), which is also impossible since γ is
generic. We conclude that λ ∈ e
k× . In particular,
(2.28)
h(W ) = degt (Res1 (W )) = degw Res1 (Wk(w)
) = hw Wk(w)
,
e
e
where hw denotes the height with respect to the base ring e
k[w]. Coming back to
our problem, we observe that (V1 #V2 )ek(w) = V1,ek(w) #V2,ek(w) , which follows readily from the definition of the ruled join. Applying (2.28) and the previously considered case when p = 1, we conclude that h(V1 #V2 ) = hw V1,ek(w) #V2,ek(w) =
deg V1,ek(w) hw V2,ek(w) +deg V2,ek(w) hw V1,ek(w) = deg(V1 )h(V2 )+deg(V2 )h(V1 ).
2.2. Measures of complex polynomials. To study cycles defined over Q, we will
use different measures for the size of a complex multivariate polynomial. We introduce
them in this section and collect some of their properties.
P
Definition 2.29. Let f = a αa xa ∈ C[x1 , . . . , xn ]. The ℓ∞ -norm, the ℓ1 -norm and
the sup-norm of f are respectively defined as
X
|αa | , ||f ||sup = supx∈(S 1 )n |f (x)|,
||f ||∞ = maxa |αa | , ||f ||1 =
a
where S 1 = {x ∈ C : |x| = 1} denotes the unit circle of C. The Mahler measure of f
is defined as
Z
Z 1
Z 1
2πiu1
2πiun
log |f |dµn
)| du1 . . . dun =
log |f (e
,...,e
···
m(f ) =
0
0
(S 1 )n
where µ denotes the Haar measure on S 1 of total mass 1.
We list some inequalities comparing these measures.
Lemma 2.30. Let f ∈ C[x1 , . . . , xn ]. Then
(1) log ||f ||∞ ≤ log ||f ||sup ≤ log ||f ||1 ≤ log ||f ||∞ + log(n + 1) deg(f );
(2) m(f ) ≤ log ||f ||sup ;
(3) log ||f ||1 ≤ m(f ) + log(n + 1) deg(f );
(4) | m(f ) − log ||f ||∞ | ≤ log(n + 1) deg(f ).
Proof.
P
(1) Let f = a αa xa . By Cauchy’s formula, for a ∈ supp(f ),
Z
f (x) n
αa =
dµ .
a
(S 1 )n x
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
37
Hence ||f ||∞ = maxa |αa | ≤ supx∈(S 1 )n |f (x)| = ||f ||sup , which gives the first inequality. The second follows easily from the definitions while the third one follows from
X
n + deg(f )
||f ||1 =
|αa | ≤ #supp(f ) ||f ||∞ ≤
||f ||∞ ≤ (n + 1)deg(f ) ||f ||∞ .
n
a
(2) follows easily from the definitions, (3) follows from [Phi86, Lem. 1.13] while (4)
follows from (1), (2) and (3).
We also list some well-known properties of the Mahler measure. Recall that a weight
monomial order on C[x] is a partial order on the monomials of C[x] defined by a
vector v ∈ Rn as xa ≺ xb if and only if hv, ai < hv, bi.
Lemma 2.31. Let f, g ∈ C[x]. Then
(1) m(f g) = m(f ) + m(g).
(2) m(f (ω1 x1 , . . . , ωn xn )) = m(f ) for any (ω1 , . . . , ωn ) ∈ (S 1 )n .
(3) m(f (xℓ1 , . . . , xℓn )) = m(f ) for any ℓ ≥ 1.
(4) Let ≺ be a weight monomial order on C[x]. Then m(init≺ (f )) ≤ m(f ).
Proof.
(1), (2) These follow easily from the definition of the Mahler measure.
(3) For u ∈ Rn set w = (ℓu1 , . . . , ℓun ) and observe that dw1 . . . dwn = ℓn du1 . . . dun .
Hence,
Z
Z
1
log |f (e2πiℓu )| du1 . . . dun = n
log |f (e2πiw )| dw1 . . . dwn
m(f (xℓ )) =
ℓ
n
n
[0,1]
[0,ℓ]
Z
log |f (e2πiw )| dw1 . . . dwn = m(f (x)).
=
[0,1]n
(4) Let v ∈ Rn be a vector defining ≺. The exponents of the monomials of f which
are minimal with respect to ≺ are the vectors in supp(f ) lying in the maximal face F
of the Newton polytope of f which has v as an inner exterior normal. Hence, init≺ (f )
is the face polynomial corresponding to F , that is, the sum of the terms in f whose
exponent lies in that face. The result then follows from the fact that the Mahler
measure of a polynomial is bounded below by the Mahler measure of any of its face
polynomials, see for instance [Smy81].
We collect in the following lemma some further inequalities comparing the measures
of polynomials depending on groups of variables.
Lemma 2.32. Let f ∈ C[y 1 , . . . , y m ] be a polynomial in m groups of ni variables
each. Then
P
(1) log ||f ||sup ≤ log ||f ||1 ≤ log ||f ||∞ + m
i=1 log(ni + 1) deg y i (f );
Pm
(2) | m(f ) − log ||f ||∞ | ≤ i=1 log(ni + 1) degyi (f ).
Proof. (1) can be proved in the same way as the third inequality in Lemma 2.30(1),
while (2) follows from [KPS01, Lem. 1.1].
For multihomogeneous polynomials, we will also need the following variant of the
Mahler measure, introduced and studied by Philippon in [Phi91].
38
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Definition 2.33. Let f ∈ C[x1 , . . . , xm ] be a multihomogeneous polynomial in m
groups of ni + 1 variables each. The Philippon measure of f is defined as
Z
ni
m X
X
1
degxi (f ),
log |f | dµ1 . . . dµm +
ph(f ) =
2j
S 2n1 +1 ×···×S 2nm +1
i=1
j=1
where S 2ni +1 = {x ∈ Cni +1 : |x0 |2 + · · · + |xni |2 = 1} denotes the unit sphere of Cni +1
and µi the Borel measure on S 2ni +1 of total mass 1, invariant under the action of the
unitary group.
The Philippon measure is related to the Mahler measure by the inequalities
(2.34)
0 ≤ ph(f ) − m(f ) ≤
m
X
log(ni + 1) degxi (f ),
i=1
which follow from [Lel94, Thm. 4]. In particular,
(2.35)
ph(f ) ≤ log ||f ||sup +
m
X
log(ni + 1) degxi (f ).
i=1
Definition 2.36. Let f ∈ Z[x1 , . . . , xn ]. The height of f is defined as h(f ) = log kf k∞
if f 6= 0 and as 0 otherwise.
The following lemma estimates the behavior of the height of polynomials with respect
to the arithmetic operations and composition. Its proof follows directly from the
definitions.
Lemma 2.37. Let f1 , . . . , fs ∈ Z[x1 , . . . , xn ]. Then
P
(1) h( i fi ) ≤ maxi h(fi ) + log(s);
P
P
P
Q
(2) h( i fi ) ≤ h(f1 ) + si=2 log ||fi ||1 ≤ i h(fi ) + log(n + 1) si=2 deg(fi ).
(3) Let g ∈ C[y1 , . . . , ys ] and write d = maxi deg(fi ) and h = maxi h(fi ). Then
h(g(f1 , . . . , fs )) ≤ h(g) + deg(g)(h + log(s + 1) + d log(n + 1)).
2.3. Canonical mixed heights of cycles over Q. The projective space Pn = Pn (Q)
has a standard structure of toric variety with torus
×
(Pn )◦ := Pn \ V (x0 · · · xn ) ≃ (Q )n .
The action of this torus on Pn writes down as u · x = (u0 x0 : · · · : un xn ) for u = (u0 :
· · · : un ) ∈ (Pn )◦ and x = (x0 : · · · : xn ) ∈ Pn . This toric structure on Pn allows to
define a notion of canonical height for its subvarieties [BPS11]. Following David and
Philippon, this height can be defined by a limit process à la Tate [DP99]. In precise
terms, for ℓ ≥ 1, consider the ℓ-power map
[ℓ] : Pn −→ Pn
,
(x0 : · · · : xn ) 7−→ (xℓ0 : · · · : xℓn ).
Let V ⊂ Pn be an irreducible subvariety and let h denote the Fubini-Study height of
projective varieties [Phi95, KPS01]. The canonical (or normalized) height of V can
be defined as
h([ℓ] V )
b
.
(2.38)
h(V ) = deg(V ) lim
ℓ→∞ ℓ deg([ℓ] V )
Both h and b
h extend to cycles by linearity. Alternatively, the canonical height can be
defined using Arakelov geometry, as the height of V with respect to canonical metric
on the universal line bundle O(1), see for instance [Mai00, BPS11].
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
39
We collect in the proposition below some of the basic properties of this notion. For a
cycle X of Pn , we denote by [ℓ]∗ X its direct image under the map [ℓ]. We denote by
×
µ∞ the subgroup of Q of roots of unity and by µℓ the subgroup of ℓ-roots. A variety
is called a torsion subvariety of Pn if it is the closure of the orbit of the action of a
subtorus of (Pn )◦ over a point with coordinates in {0} ∪ µ∞ .
Proposition 2.39.
(1) Let X ∈ Zr (Pn ). Then b
h(X) = limℓ→∞ ℓ−r−1 h([ℓ]∗ X).
(2) Let X ∈ Zr (Pn ) and ℓ ∈ Z>0 . Then deg([ℓ]∗ X) = ℓr deg(X) and b
h([ℓ]∗ X) =
ℓr+1 b
h(X).
(3) If X is an effective cycle, then b
h(X) ≥ 0.
(4) If X is a sum of torsion subvarieties, then b
h(X) = 0. In particular, b
h(Pn ) = 0.
7
n
+
h(X) − h(X) ≤ 2 (r + 1) log(n + 1) deg(X).
(5) Let X ∈ Zr (P ). Then b
n
(6) Let ξ ∈ P be a point with rational coordinates and write ξ = (ξ0 : · · · : ξn ) for
some coprime ξi ∈ Z. Then b
h(ξ) = log(maxi |ξi |).
+ n
(7) Let D ∈ Div (PQ ) and fD ∈ Z[x] its primitive defining polynomial. Then
b
h(D) = m(fD ).
Proof.
(1-2) By linearity, we reduce to the case of an irreducible variety V . We can assume
that V ∩ (Pn )◦ 6= ∅, otherwise we restrict to a sufficiently small standard subspace of
Pn , that is, a linear subspace defined by a subset of the coordinates xi .
For shorthand, let [ℓ] denote also the restriction of the ℓ-power map to the torus (Pn )◦ .
This is a group homomorphism with kernel ker[ℓ] ≃ µnℓ . Let stab(V ) = {u ∈ (Pn )◦ :
u · V = V } be the stabilizer of V . On the one hand, by [DP99, Prop. 2.1(i)],
ℓr
deg(V ).
deg([ℓ] V ) =
#(stab(V ) ∩ ker[ℓ])
On the other hand, for a generic point x ∈ V ,
deg([ℓ]
V
) = #{y ∈ V : [ℓ]y = [ℓ]x} = #{ω ∈ ker[ℓ] : ω x ∈ V } = #(stab(V )∩ker[ℓ]).
Therefore, deg([ℓ]∗ V ) = deg([ℓ]
Finally
V
) deg([ℓ]V ) = ℓr deg(V ). Furthermore,
deg([ℓ] V )h([ℓ]V )
deg(V )h([ℓ]V )
h([ℓ]∗ V )
b
h(V ) = lim
= lim
= lim
.
ℓ→∞ ℓ deg([ℓ]V )
ℓ→∞
ℓ→∞
ℓr+1
ℓr+1
h([j]∗ [ℓ]∗ V )
h([j ℓ]∗ V )
b
h([ℓ]∗ V ) = lim
= ℓr+1 lim
= ℓr+1 b
h(V ).
r+1
j→∞
j→∞ (ℓ j)r+1
j
(3) This is a direct consequence of the definition of b
h and the analogous property for
the Fubini-Study height.
(4–7) These follow from [DP99, Prop. 2.1 and display (2)].
In the sequel, we extend the notion of canonical height to the multiprojective setting
and study its behavior under geometric constructions. Our approach relies on the
analogous theory for the Fubini-Study mixed height developed by Rémond in [Rem01a,
Rem01b]. For simplicity, we will restrict to subvarieties of Pn defined over Q or
equivalently, to Q-varieties in Pn
Q , see Remark 1.1. We will apply the resultant theory
in §1 for the case when A = Z. In particular, the resultant of an irreducible Q-variety
is a primitive polynomial in Z[u].
40
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Definition 2.40 ([Rem01b, §2.3]). Let V ⊂ Pn
Q be an irreducible Q-variety of dimenm
sion r, c ∈ Nr+1 and e(c) as defined in (1.43). The Fubini-Study (mixed) height of V
of index c is defined as
hc (V ) = ph(Rese(c) (V )).
This definition extends by linearity to cycles in Zr (Pn
Q ).
Next, we introduce morphisms relating different multiprojective spaces.
i
Definition 2.41. Let δi ∈ Z>0 and set Ni = δi n+n
. The Veronese embedding and
i
the modified Veronese embedding of index δi are the embeddings vδi , veδi : Pni ֒→ PNi −1
respectively defined for xi ∈ Pni as
1/2
δi
ai
ai
xi
vδi (xi ) = (xi )a ∈Nni +1 , veδ i (xi ) =
i
n +1
ai
δi
ai ∈N i
δi
vδm for
For δ = (δ1 , . . . , δm ) ∈ (Z>0 )m , we set vδ = vδ1 × · · · × vδm and veδ = veδ1 × · · · × e
the Veronese embedding and the modified Veronese embedding of index δ, respectively.
These are embeddings of Pn into PN −1 , where N = (N1 , . . . , Nm ) and 1 = (1, . . . , 1).
Consider the Segre
embedding s : PN −1 ֒→ PN1 ···Nm −1 defined as s(y 1 , . . . , y m ) =
y1,j1 · · · ym,jm 1≤j ≤N for y i ∈ PNi −1 . The composed maps s ◦ vδ , s ◦ veδ : Pn ֒−→
i
i
PN1 ···Nm −1 are called the Segre-Veronese embedding and the modified Segre-Veronese
embedding of index δ, respectively.
Let δ = (δ1 , . . . , δm ) ∈ (Z>0 )m . For each i, consider the diagonal endomorphism
∼
∆i : PNi −1 → PNi −1 defined as
1/2
δi
yi,ai
.
∆i ((yi,ai )ai ) =
n +1
ai
ai ∈N i
δi
∼
Consider also the diagonal endomorphism ∆ : PN1 ···Nm −1 → PN1 ···Nm −1 defined as
1/2
1/2
δm
δ1
ya
.
···
(2.42)
∆((ya )a ) =
am
a1
a∈Nn+1
δ
These linear maps allow to write the modified Veronese and the modified SegreVeronese embeddings in terms of the Veronese and Segre-Veronese embeddings as
veδi = ∆i ◦ vδi
(2.43)
,
s ◦ veδ = ∆ ◦ s ◦ vδ .
The degree and the Fubini-Study height of the direct image of a cycle X ∈ Zr (Pn
Q)
under the Segre-Veronese and the modified Segre-Veronese embeddings decompose in
terms of mixed degrees and Fubini-Study mixed heights of X.
m
Proposition 2.44. Let X ∈ Zr (Pn
Q ) and δ ∈ (Z>0 ) . Then
X r
X r + 1
b
deg (s ◦ vδ )∗ X =
degb (X) δ , h (s ◦ e
vδ )∗ X =
hc (X) δ c .
b
c
m
m
c∈Nr+1
b∈Nr
Proof. This follows from [Rem01b, p. 103].
Let ℓ ∈ Z>0 . The ℓ-power map of Pn is defined as
[ℓ] : Pn −→ Pn
,
x = (xi,j )i,j 7−→ xℓ = (xℓi,j )i,j .
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
41
m
Proposition-Definition 2.45. Let X ∈ Zr (Pn
Q ) and c ∈ Nr+1 . Then the sequence
(ℓ−r−1 hc ([ℓ]∗ X))ℓ≥1 converges for ℓ → ∞. The limit
b
hc (X) := lim ℓ−r−1 hc ([ℓ]∗ X)
ℓ→∞
is called the canonical (mixed) height of X of index c. For any δ ∈ (Z>0 )m it holds
X r + 1
b
b
(2.46)
h (s ◦ vδ )∗ X =
hc (X) δ c .
c
m
c∈Nr+1
hr+1 (X) = b
h(X).
In particular, for a projective cycle X ∈ Zr (PnQ ) we have that b
Proof. Proposition 2.44 applied to the cycle [ℓ]∗ X implies that
X r + 1
(2.47)
h (s ◦ veδ )∗ [ℓ]∗ X =
hc ([ℓ]∗ X) δ c .
c
m
c∈Nr+1
Let I ⊂ (Z>0 )m be a subset of cardinality #(Nm
r+1 ) such that the square matrix
r+1 c
is
of
maximal
rank.
Inverting
this matrix, we can write the
δ
c
c∈Nm , δ∈I
r+1
mixed heights in the formula above in terms of heights of projective cycles as
X
(2.48)
hc ([ℓ]∗ X) =
νc,δ h (s ◦ veδ )∗ [ℓ]∗ X .
δ∈I
with νc,δ ∈ Q not depending on ℓ. Observe that
(s ◦ veδ ◦ [ℓ])∗ X = ∆∗ (s ◦ vδ ◦ [ℓ])∗ X,
where ∆ denotes the linear map in (2.42). By [KPS01, Lem. 2.7] applied to the
projective cycle (s ◦ vδ ◦ [ℓ])∗ X, the map ∆ and its inverse, there exists κ(m, r, δ) ≥ 0
such that the quantity |h (s ◦ veδ )∗ [ℓ]∗ X − h (s ◦ vδ )∗ [ℓ]∗ X | is bounded above by
Y
m
δi + ni
deg((s ◦ vδ )∗ [ℓ]∗ X).
κ(m, r, δ) log
ni
i=1
We have that [ℓ] commutes with s ◦ vδ . By Proposition 2.39(2),
deg((s ◦ vδ )∗ [ℓ]∗ X) = deg([ℓ]∗ (s ◦ vδ )∗ X) = ℓr deg((s ◦ vδ )∗ X).
We deduce h (s ◦ veδ )∗ [ℓ]∗ X = h [ℓ]∗ (s ◦ vδ )∗ X + O(ℓr ). Therefore, for each δ ∈ I,
lim ℓ−r−1 h (s ◦ veδ )∗ [ℓ]∗ X = lim ℓ−r−1 h [ℓ]∗ (s ◦ vδ )∗ X = b
h (s ◦ vδ )∗ X .
ℓ→∞
ℓ→∞
This proves that the sequence (ℓ−r−1 hc ([ℓ]∗ X))ℓ≥1 converges for ℓ → ∞, since it is a
linear combination of convergent sequences as shown in (2.48). The formula (2.46)
follows from (2.47) by passing to the limit for ℓ → ∞. The last statement is this
formula applied to X ∈ Zr (PnQ ) and δ = 1.
Remark 2.49. The definition of canonical mixed heights in [PS08, Formula (1.3)]
is different from the one presented here. Nevertheless, both notions coincide as they
both satisfy (2.46). These mixed heights can be alternatively defined using Arakelov
geometry, as explained at the end of [PS08, §I], and they correspond to the canonical
mixed heights induced by the toric structure of Pn
Q , see [BPS11].
42
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Definition 2.50. Let η be an indeterminate. The extended Chow ring of Pn
Q is the
graded ring
∗ n
2
2 n1 +1
nm +1
A∗ (Pn
, . . . , θm
),
Q ; Z) = A (PQ ) ⊗Z R[η]/(η ) ≃ R[η, θ1 , . . . , θm ]/(η , θ1
ni
where θi denotes the class in A∗ (Pn
Q ) of the inverse image of a hyperplane of PQ under
ni
∗ n
the projection Pn
Q → PQ . For short, we alternatively denote this ring as A (P ; Z).
To a cycle X ∈ Zr (Pn
Q ) we associate an element of this ring, defined as
X
X
nm −cm
nm −bm
b
+
hc (X) η θ1n1−c1 · · · θm
[X]Z =
.
degb (X) θ1n1 −b1 · · · θm
c∈Nm
r+1 , c≤n
b∈Nm
r ,b≤n
It is a homogeneous element of degree |n| − r.
∗ n
There is an inclusion ı : A∗ (Pn
Q ) ֒→ A (PQ ; Z) which satisfies [X]Z ≡ ı([X]) (mod η).
In particular, the class of a cycle in the Chow ring is determined by its class in the
extended Chow ring.
b
For X ∈ Zr (Pn
Q ), Theorem 2.58(2) shows that hc (X) = 0 for every c such that ci > ni
for some i. Hence [X]Z contains the information of all mixed degrees and heights,
since {θ a , η θ a }a≤n is a basis of A∗ (Pn ; Z).
Next proposition extends the first properties of the canonical height in Proposition 2.39
to the multiprojective setting. The space Pn is a toric variety with torus
m
Y
×
n ◦
(Pni )◦ ≃ (Q )|n| .
(P ) =
i=1
A variety V is called a torsion subvariety of Pn if it is the closure of the orbit of the
action of a subtorus of (Pn )◦ over a point with coordinates in {0} ∪ µ∞ .
Proposition 2.51.
r
b
(1) Let X ∈ Zr (Pn
Q ) and ℓ ≥ 1. Then degb ([ℓ]∗ X) = ℓ degb (X) and hc ([ℓ]∗ X) =
m
hc (X) for all b ∈ Nm
ℓr+1 b
r and c ∈ Nr+1 .
m
b
(2) Let X ∈ Zr+ (Pn
Q ). Then [X]Z ≥ 0. In particular, hc (X) ≥ 0 for all c ∈ Nr+1 .
(3) If X is a linear combination of torsion subvarieties, then [X]Z = ı([X]) or
n
equivalently, b
hc (X) = 0 for all c ∈ Nm
r+1 . In particular, [PQ ]Z = 1.
m
(4) Let X ∈ Zr+ (Pn
Q ). Then there exists κ(r, m) ≥ 0 such that for all c ∈ Nr+1 ,
X
b
hc (X) − hc (X) ≤ κ(r, m) log(|n| + 1)
degb (X).
b∈Nm
r
Proof.
(1) Recall that [ℓ] commutes with s ◦ vδ . Hence, propositions 2.44 and 2.39(2) imply
that
X r
r
ℓ
degb (X)δ b = ℓr deg (s ◦ vδ )∗ X
b
X r
b∈Nm
r
= deg (s ◦ vδ )∗ [ℓ]∗ X =
degb ([ℓ]∗ X) δ b .
b
m
b∈Nr
Since this holds for all δ ∈ (Z>0 )m , we deduce that degb ([ℓ]∗ X) = ℓr degb (X). The
statement for the height follows analogously by using (2.46) and Proposition 2.39(2).
(2) The non-negativity of the canonical mixed heights is a consequence of the nonnegativity of the Fubini-Study mixed heights. The latter follows from the estimates
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
43
in (2.34) and the non-negativity of the Mahler measure of a polynomial with integer
coefficients. Hence, b
hc (X) ≥ 0 for all c ∈ Nm
r+1 . The rest of the statement follows
from this together with Proposition 1.10(1).
(3) It suffices to prove this statement for a torsion subvariety V of Pn . Given δ ∈
(Z>0 )m , Proposition 2.39(4) implies that b
h((s ◦ vδ )∗ X) = 0 since the image of V under
a Segre-Veronese embedding is a torsion subvariety of a projective space. Therefore,
the right-hand side of (2.46) is equal to 0 for all δ, which implies that all its coefficients
are 0. Hence all canonical mixed heights of V are 0, as stated.
(4) By (2.46) and Proposition 2.44, for each δ ∈ (Z>0 )m ,
X r + 1
b
b
h (s ◦ vδ )∗ X − h (s ◦ veδ )∗ X =
hc (X) − hc (X) δ c .
c
m
c∈Nr+1
m
As in the proof of Proposition-Definition 2.45, we pick asubset
I ⊂ (Z>0 ) of cardir+1
c
m
δ c∈Nm ,δ∈I is of maximal
nality #(Nr+1 ) such that the square matrix MI :=
c
r+1
rank. Inverting this matrix, we obtain for each c
X
b
h (s ◦ vδ )∗ X − h (s ◦ veδ )∗ X
(2.52)
hc (X) − hc (X) =
νc,δ b
δ∈I
with νc,δ ∈ Q. Using (2.43), [KPS01, Lem. 2.7], Proposition 2.44 and the inequality
Qm δi +ni
log
≤ |δ| log(|n| + 1), we get that there exists κ1 (r, m, δ) ≥ 0 such that
i=1
ni
X
h (s ◦ veδ )∗ X − h (s ◦ vδ )∗ X | ≤ κ1 (r, m, δ) log(|n| + 1)
degb (X).
b∈Nm
r
By propositions 2.39(5) and 2.44, there exists κ2 (r, m, δ) ≥ 0 such that
X
b
degb (X).
h (s ◦ vδ )∗ X − h (s ◦ vδ )∗ X ≤ κ2 (r, m, δ) log(|n| + 1)
b∈Nm
r
Therefore, setting κ3 (r, m, δ) = κ1 (r, m, δ) + κ2 (r, m, δ), we get
X
b
h (s ◦ vδ )∗ X − h (s ◦ veδ )∗ X ≤ κ3 (r, m, δ) log(|n| + 1)
degb (X).
b∈Nm
r
Observe that the matrix MI does not depend on n. Using (2.52), we deduce that
there exists κ(r, m) such that
X
degb (X)
|b
hc (X) − hc (X)| ≤ κ(r, m) log(|n| + 1)
b∈Nm
r
for all c ∈ Nm
r+1 , as stated.
The following proposition describes the mixed heights and classes of points and divisors, extending Proposition 2.39(6-7) to the multiprojective setting.
Proposition 2.53.
(1) Let ξ = (ξ 1 , . . . , ξ m ) ∈ Pn be a point with coordinates in Q and for each
1 ≤ i ≤ m write ξ i = (ξi,j )j for coprime ξi,j ∈ Z. Then
[ξ]Z =
m
X
i=1
In particular, b
hei (ξ) = b
h(ξ i ).
b
h(ξ i ) η θ n−ei + θ n .
44
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
(2) Let D ∈ Div+ (Pn
Q ) and fD ∈ Z[x] its primitive defining polynomial. Then
[D]Z = m(fD ) η +
m
X
degxi (fD ) θi .
i=1
In particular, b
hn (D) = m(fD ).
We need the following lemma.
Lemma 2.54. Let H be an irreducible Q-hypersurface of Pn
Q which is not a standard
hyperplane and ℓ ≥ 1. Then
Y
f[ℓ]∗ [ℓ]∗ H =
fH (ω · x).
ω∈ker[ℓ]
Proof. Let stab(H) = {u ∈
(Pn )◦
: u · H = H} be the stabilizer of V . Then
[
[ℓ]−1 [ℓ]H =
ω · H,
ω
the union being over a set of representatives of ker[ℓ]/(stab(H) ∩ ker[ℓ]). As in the
proof of Proposition 2.39(1-2), we can show that deg([ℓ] H ) = #(stab(H) ∩ ker[ℓ]).
Hence,
X
[ℓ]∗ [ℓ]∗ H = deg([ℓ] H ) [ℓ]−1 [ℓ] H =
ω · H.
ω∈ker[ℓ]
This implies that the primitive polynomial defining this divisor satisfies
Y
f[ℓ]∗ [ℓ]∗ H = λ
fH (ω · x)
ω∈ker[ℓ]
with λ ∈ Q× . It only remains to prove that λ = ±1. Let K be the ℓ-th cyclotomic
field, so that fH (ω · x) ∈ K[x] for all ω ∈ ker[ℓ]. Let p ∈ Z be a prime number and v
a valuation of K extending ordp . By Gauss Lemma,
Y
X
v fH (ω · x) = 0
fH (ω · x) =
ordp
ω
ω
since fH is primitive and v(ω) = 0 for all ω ∈ ker[ℓ]. This implies that ordp (λ) = 0
for all p and λ = ±1 as desired.
Proof of Proposition 2.53.
(1) It is enough to prove that b
hei (ξ) = b
h(ξ i ). We have that [ℓ]∗ ξ = ξ ℓ , hence
ℓ
Resei ([ℓ]∗ ξ) = λi Li (ξ i ), where Li is the general linear form of multidegree ei and
λi ∈ Q× , thanks to Corollary 1.38. Since Resei (ξ) and Li (ξ i ) are primitive polynomials, we deduce that λi = ±1. Applying Definition 2.40, the estimates in (2.34) and
Lemma 2.30(4), we obtain
ℓ−1 hei ([ℓ]∗ ξ) = ℓ−1 ph Li (ξ ℓi ) = log(maxj |ξi,j |) + O(ℓ−1 ).
The statement follows by letting ℓ → ∞.
(2) By Proposition 1.10(4),
[D]Z = b
hn (D) η + [D] = b
hn (D) η +
m
X
degxi (fD ) θi .
i=1
Thus, we only have to prove that b
hn (D) = m(fD ). We reduce without loss of generality
to the case of an irreducible hypersuface H. If H is a standard hyperplane of Pn
Q , then
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
45
bn (H) = 0 = m(xi,j ), and the statement is true in this
fH = xi,j for some i, j. Hence h
case.
We can then suppose that H is an irreducible Q-hypersurface which is not a standard
hyperplane. Let ℓ ≥ 1. By definition, f[ℓ]∗[ℓ]∗ H = f[ℓ]∗ H (xℓ ). Applying successively
Lemma 2.31(3), Lemma 2.54 and Lemma 2.31(2), we get
X
m(f[ℓ]∗ H ) = m(f[ℓ]∗H (xℓ )) =
m(fH (ω · x)) = ℓ|n| m(fH ).
ω∈ker[ℓ]
By [Rem01b, Thm. 3.4], the estimates in (2.34) and Proposition 2.51(1),
j
m nX
i −1 X
X
1
degn−ei ([ℓ]∗ H) = m(f[ℓ]∗ H ) + O(ℓ|n|−1 ).
hn ([ℓ]∗ H) = ph(f[ℓ]∗H ) +
2l
i=1
j=1 l=1
Therefore, ℓ−|n| hn ([ℓ]∗ H) = m(fH ) + O(ℓ−1 ). We conclude that b
hn (H) = m(fH ) by
letting ℓ → ∞.
The following result gives the behavior of mixed degrees and heights with respect to
Veronese embeddings.
Proposition 2.55. For δ ∈ (Z>0 )m let vδ be the Veronese embedding of index δ,
m
m
X ∈ Zr (Pn
Q ), b ∈ Nr and c ∈ Nr+1 . Then
degb ((vδ )∗ X) = δ b degb (X)
,
b
hc ((vδ )∗ X) = δ c b
hc (X).
Proof. It is enough to prove the statement for an irreducible Q-variety V . Set Ni =
δi +ni
for 1 ≤ i ≤ m, N = (N1 , . . . , Nm ) and 1 = (1, . . . , 1) ∈ Nm . The embedding
ni
vδ induces an isomorphism of multigraded algebras
M
M
Q[Pn ]/I(V ) (d δ ,...,d δ ) .
Q[PN −1 ]/I(vδ (V )) =
Q[PN −1 ]/I(vδ (V )) d ≃
1 1
d∈Nm
d∈Nm
m m
Hence, the Hilbert-Samuel functions of V and of vδ (V ) satisfy Hvδ (V ) (d1 , . . . , dm ) =
HV (δ1 d1 , . . . , δm dm ) for all d. Comparing the coefficient of the monomial db in the
corresponding Hilbert polynomials and using that vδ is a map of degree 1, it follows
that degb ((vδ )∗ V ) = degb (vδ (V )) = δ b degb (V ).
Concerning the height, consider the embeddings
′
′
s ◦ vdδ : Pn ֒−→ PN1 ···Nm −1 , s ◦ vd ◦ vδ : Pn ֒−→ PM1 ···Mm −1 ,
di +Ni −1
i
. For xi ∈ Pni ,
and
M
=
where Ni′ = di δni +n
i
Ni −1
i
vdi δi (xi ) = (xai i )|ai |=di δi , vdi ◦ vδi (xi ) = ((xbi i )|bi |=δi )ci |c |=d
i
i
Observe that the monomials appearing in the image of both maps are the same. This
′
′
implies that there are linear maps Ai : PMi −1 → PNi −1 and Bi : PNi −1 → PMi −1 such
that Ai ◦ vdi ◦ vδi = vdi δi and vdi ◦ vδi = Bi ◦ vdi δi . In turn, this implies that there exist
′
′
′
′
linear maps A : PM1 ···Mm −1 → PN1 ···Nm −1 and B : PN1 ···Nm −1 → PM1 ···Mm −1 such that
A ◦ s ◦ vd ◦ vδ = s ◦ vdδ
,
s ◦ vd ◦ vδ = B ◦ s ◦ vdδ .
Let ℓ ≥ 1. We apply [KPS01, Lem. 2.7] to compare the Fubini-Study height of the
image of the cycle [ℓ]∗ V under the maps s ◦ vdδ and s ◦ vd ◦ vδ . Using also propositions
2.44 and 2.51(1), we obtain that there exists κ(n, d, δ) ≥ 0 such that
X
(2.56) h(s ◦ vdδ ([ℓ]∗ V )) − h(s ◦ vd ◦ vδ ([ℓ]∗ V )) ≤ κ(n, d, δ)
degb ([ℓ]∗ V ) = O(ℓr ).
b
46
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
The map [ℓ] commutes with vδ , vd , vdδ and s. Hence, s◦vdδ ([ℓ]∗ V ) = [ℓ]∗ s◦vdδ (V ) and
s◦vd ◦vδ ([ℓ]∗ V ) = [ℓ]∗ s◦vd ◦vδ (V ). From (2.56), we deduce that h([ℓ]∗ s◦vd ◦vδ (V )) =
h([ℓ]∗ s ◦ vdδ (V )) + O(ℓr ). Therefore,
b
h(s ◦ vd ◦ vδ (V )) = lim ℓ−r−1 h([ℓ]∗ s ◦ vd ◦ vδ (V ))
ℓ→∞
h(s ◦ vdδ (V )).
= lim ℓ−r−1 h([ℓ]∗ s ◦ vdδ (V )) = b
ℓ→∞
Hence, Proposition 2.44 implies that
X r + 1
b
hc (vδ (V )) dc = b
h(s ◦ vd ◦ vδ (V ))
c
c
h(s ◦ vdδ (V )) =
=b
X r + 1
c
c
b
hc (V ) (d δ)c .
hc (V ), as stated.
hc (vδ (V )) = δ c b
As this hold for all d ∈ Nm , it follows that b
The following result is the arithmetic analogue of Bézout’s theorem for multiprojective
Q-cycles. It contains Theorem 0.3 in the introduction. Given a multihomogeneous
polynomial f ∈ Z[x1 , . . . , xm ], we consider the element in the extended Chow ring
(2.57)
[f ]sup = log ||f ||sup η +
Observe that [div(f )]Z ≤ [f ]sup .
m
X
i=1
degxi (f )θi ∈ A∗ (Pn ; Z).
Theorem 2.58. Let X ∈ Zr (Pn
Q ) and f ∈ Z[x1 , . . . , xm ] a multihomogeneous polynomial such that X and div(f ) intersect properly.
(1) If X is effective, then for any b ∈ Nm
r ,
b
hb (X · div(f )) ≤
m
X
i=1
degxi (f )b
hb+ei (X) + log ||f ||sup degb (X).
bc (X) = 0 for any c ∈ Nm such that ci > ni for some i.
(2) h
r+1
(3) If X is effective, then [X · div(f )]Z ≤ [X]Z · [f ]sup .
Proof.
(1) We reduce without loss of generality to the case of an irreducible Q-variety
m . Consider the Veronese embedV . Let ℓ ≥ 1 and set ℓ = (ℓ, . . . , ℓ) ∈ (Z
>0 ) ℓ+n
ℓ+ni
n
N
−1
i
ding vℓ : P → P
where N =
, 1 = (1, . . . , 1) ∈ Nm . Set
ni , . . . , ni
d = (d1 , . . . , dm ) = deg(f ). We have that f ℓ ∈ Q[Pn ]ℓ d and so there is a unique polynomial Fℓ ∈ Q[PN −1 ]d such that vℓ∗ div(Fℓ ) = div(f ℓ ) = ℓ div(f ). Then, the projection
formula (1.6) implies that
(2.59)
(vℓ )∗ V · div(Fℓ ) = ℓ (vℓ )∗ (V · div(f )).
ℓ (ξ)|
Set ||F ||v (V ) = supξ∈vℓ (V ) ||ξ ||d|F
We
dm , where || · || is the Euclidean norm.
1
···||ξ
ℓ
1
m ||
apply the arithmetic Bézout’s theorem for the Fubini-Study mixed height ([Rem01b,
Thm. 3.4 and Cor. 3.6]) to the variety Zℓ := vℓ (V ) = (vℓ )∗ V and the divisor div(Fℓ )
and we obtain
m
X
di hb+ei (Zℓ ) + degb (Zℓ ) log ||Fℓ ||v (V ) .
hb (Zℓ · div(Fℓ )) ≤
ℓ
i=1
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
47
Applying Proposition 2.51(4) to Zℓ and to Zℓ ·div(Fℓ ) together with the multiprojective
Bézout’s theorem 1.11, it follows that there exist κ(r, m, d) ≥ 0 such that
(2.60) b
hb (Zℓ · div(Fℓ )) ≤
m
X
i=1
di b
hb+ei (Zℓ ) + degb (Zℓ ) log(||Fℓ ||v
ℓ (V )
+ κ(r, m, d) log(|N − 1| + 1)
X
degb (Zℓ ).
b
By Proposition 2.55, degb (Zℓ ) =
degb (V ) and b
hb+ei (Zℓ ) = ℓr+1 b
hb+ei (V ). The
same result together with (2.59) also implies that
b
hb (Zℓ · div(Fℓ )) = b
hb (ℓ vℓ,∗ (V · div(f ))) = ℓr+1 b
hb (V · div(f )).
ℓr
We have that
||Fℓ ||v
ℓ (V )
= supx∈V
|f ℓ (x)|
||vℓ (x1 )||d1 · · · ||vℓ (xm )||dm
|f ℓ (x)|
≤ supx∈Dn+1 |f ℓ (x)|,
ℓdi
max
|x
|
j i,j
i=1
where D = {z ∈ C : |z| ≤ 1} is the unit disk. Hence, log(||Fℓ ||v (V ) ≤ ℓ log ||f ||sup
ℓ
because of the maximum modulus principle applied to f . Besides, log(|N − 1| + 1) =
O(log(ℓ)) and we deduce from (2.60) that
m
log(ℓ)
X
b
di b
hb+ei (V ) + degb (V ) log ||f ||sup + O
hb V · div(f ) ≤
.
ℓ
≤ supx∈V Qm
i=1
The result follows by letting ℓ → ∞.
(2) This follows by adapting the proof of Theorem 2.18(2) to this setting without
major changes.
(3) This is an immediate consequence of (1) and (2), together with Theorem 1.11 and
Proposition 1.10(4).
The following results can be proved by adapting the arguments in the proofs of corollaries 2.19 and 2.20 without major changes.
Corollary 2.61. Let V ⊂ Pn
Q be a Q-variety of pure dimension r and f ∈ Z[x1 , . . . , xm ]
a multihomogeneous polynomial. Let W denote the union of the components of dimension r − 1 of the intersection V ∩ V (f ). Then
In particular, b
hb (W ) ≤
Pm
[W ]Z ≤ [V ]Z · [V (f )]sup .
b
i=1 degxi (f ) hb+ei (V
) + log ||f ||sup degb (V ) for all b ∈ Nm
r .
Corollary 2.62. Let n ∈ Z>0 , X ∈ Zr+ (PnQ ) and fj ∈ Z[x0 , . . . , xn ] be s ≤ r homoQi−1
geneous polynomials such that X · j=1
div(fj ) and div(fi ) intersect properly for all
1 ≤ i ≤ s. Then
s
s
s
Y
X
Y
log ||fℓ ||sup
b
b
.
div(fj ) ≤
deg(fi ) h(X) + deg(X)
h X·
deg(fℓ )
j=1
i=1
ℓ=1
Remark 2.63. Theorem 2.58 is one of the main reasons for considering canonical
mixed heights instead of others. For instance, the fact that the canonical mixed
heights of index c such that ci > ni for some i are zero is quite convenient in the
applications. Observe that the analogue of this statement for the Fubini-Study mixed
48
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
heights does not hold: for instance, hn+1 (Pn ) equals the n-th Stoll number, which
is not zero. This fact implies that the class in A∗ (Pn
Q ; Z) of a cycle X contains the
information about all canonical mixed heights of X. This information is necessary if
one wants to express the arithmetic Bézout’s inequality in Theorem 2.58(1) in terms
of elements in the extended Chow ring.
We next show that canonical mixed heights are monotonic with respect to linear
projections. We keep the notation from Proposition 1.16 and consider also the injective
R-linear map : A∗ (Pl ; Z) ֒→ A∗ (Pn ; Z) defined by (P ) = θ n−l P .
l
Proposition 2.64. Let π : Pn
Q → PQ be the standard projection defined in (1.15) and
X ∈ Zr+ Pn
Q . Then
[π∗ X]Z ≤ [X]Z .
In particular, b
hc (π∗ X) ≤ b
hc (X) for all c ∈ Nm
r+1 .
Proof. The statement is equivalent to degb (π∗ X) ≤ degb (X) and b
hc (π∗ X) ≤ b
hc (X)
for all b, c. Because of Proposition 1.16, we only need to prove the latter inequality.
Let ℓ ≥ 1. By Proposition 1.41, Rese(c) (π∗ [ℓ]∗ X) divides init≺ (Rese(c) ([ℓ]∗ X)) in Z[u].
Using the fact that the Mahler measure of a polynomial with integers coefficients is
≥ 0 and Lemma 2.31(4), we deduce that
m(Rese(c) (π∗ [ℓ]∗ X)) ≤ m init≺ (Rese(c) ([ℓ]∗ X)) ≤ m Rese(c) ([ℓ]∗ X) .
From the estimates in (2.34) and Proposition 2.51(1), it follows that
hc (π∗ [ℓ]∗ X)) = ph(Rese(c) (π∗ [ℓ]∗ X)) = m(Rese(c) (π∗ [ℓ]∗ X)) + O(ℓr ),
hc ([ℓ]∗ X)) = ph(Rese(c) ([ℓ]∗ X)) = m(Rese(c) ([ℓ]∗ X)) + O(ℓr ).
Since π commutes with [ℓ],
hc (π∗ [ℓ]∗ X)
hc ([ℓ]∗ X)
hc ([ℓ]∗ π∗ (X))
b
= lim
≤ lim
=b
hc (X),
hc (π∗ (X)) = lim
r+1
r+1
ℓ→∞
ℓ→∞
ℓ→∞
ℓ
ℓ
ℓr+1
which completes the proof.
Example 2.65. We revisit the problem in Example 1.13 of computing the pairs
(eigenvalue, eigenvector) for the case of a matrix
Q M with entries in Z. We use the
notations therein and furthermore we set Z = ni=1 div(fi ) ∈ Z0+ (P1 × Pn−1 ; Q). By
Corollary 2.62,
n
Y
[div(fi )]sup .
[Z]Z ≤
i=1
We have that [Z]Z = b
h(1,0) (Z) η θ2n−1 + b
h(0,1) (Z) η θ1 θ2n−2 + deg(Z) θ1 θ2n−1 and that
[div(fi )]sup = log ||fi ||sup η + θ1 + θ2 . By comparing coefficients, it follows that
X
X
b
h(1,0) (Z) ≤
log ||fi ||sup , b
h(0,1) (Z) ≤ (n − 1)
log ||fi ||sup , deg(Z) ≤ n.
i
i
Let π1 and π2 denote the projections from
respectively. We have that
Z1 := (π1 )∗ Z ∈ Z0+ (P1Q )
,
P1 × Pn−1
to the first and the second factor,
n−1
)
Z2 := (π2 )∗ Z ∈ Z0+ (PQ
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
49
are the cycles of eigenvalues and of eigenvectors of M , respectively. Applying Proposition 2.64, we deduce that the heights of these cycles satisfy
X
X
b
h(Z1 ) ≤
log ||fi ||sup , b
h(Z2 ) ≤ (n − 1)
log ||fi ||sup .
i
i
A straightforward application of the arithmetic Bézout inequality
Pfor projective cycles
(Corollary 2.62) would have given the much worse bound 2n−1 i log ||fi ||sup .
Next result gives the behavior of extended Chow rings and classes with respect to
products.
Proposition 2.66. Let mi ∈ Z>0 and ni ∈ Nmi for i = 1, 2. Then
(1) A∗ (Pn1 × Pn2 ; Z) ≃ A∗ (Pn1 ; Z) ⊗R[η] A∗ (Pn2 ; Z).
i
(2) Let Xi ∈ Zri (Pn
Q ) for i = 1, 2. The above isomorphism identifies [X1 × X2 ]Z
with [X1 ]Z ⊗[X2 ]Z . In particular, for ci ∈ Nmi such that |c1 |+|c2 | = r1 +r2 +1,
b
degc1 (X1 ) hc2 (X2 ) if |c1 | = r1 , |c2 | = r2 + 1,
b
h(c1 ,c2 ) (X1 × X2 ) =
degc2 (X2 ) b
hc1 (X1 ) if |c1 | = r1 + 1, |c2 | = r2 ,
0
otherwise.
Proof.
(1) This is immediate from the definition of the extended Chow ring.
(2) Consider the case when |c1 | = r1 + 1, |c2 | = r2 . By propositions 1.45 and 2.51(1),
hc ([ℓ]∗ (X1 × X2 )) = ph(Rese(c) ([ℓ]∗ X1 × [ℓ]∗ X2 ))
= ph(Rese(c1 ) ([ℓ]∗ X1 )degc2 ([ℓ]∗ X2 ) ) = ℓr2 degc2 (X2 ) hc1 ([ℓ]∗ X1 ).
Hence,
b
hc (X1 × X2 ) = lim ℓ−r1 −r2 −1 hc ([ℓ]∗ (X1 × X2 ))
ℓ→∞
= degc2 (X2 ) lim ℓ−r1 −1 hc1 ([ℓ]∗ X1 ) = b
hc1 (X1 ) degc2 (X2 ).
ℓ→∞
The other cases can be proved similarly. The equality of classes in the extended Chow
ring follows from this together with Proposition 1.19.
Finally, we compute the class in the extended Chow ring of the ruled join of two
projective varieties. For i = 1, 2 consider the injective R-linear map i : A∗ (Pni ; Z) ֒→
A∗ (Pn1 +n2 +1 ; Z) defined by θ l η b 7→ θ l η b for 0 ≤ l ≤ ni and b = 0, 1.
Proposition 2.67. Let ni ∈ N and Xi ∈ Zri (PnQi ) for i = 1, 2. Then
[X1 #X2 ]Z = 1 ([X1 ]Z ) · 2 ([X2 ]Z ).
In particular, b
h(X1 #X2 ) = deg(X1 ) b
h(X2 ) + deg(X2 ) b
h(X1 ).
Proof. The equality of classes is equivalent to deg(X1 #X2 ) = deg(X1 ) deg(X2 ) and
b
h(X1 #X2 ) = deg(X1 ) b
h(X2 ) + deg(X2 ) b
h(X1 ). The first one is (1.20) while the second
one is [PS08, Prop. 4.9(a)].
50
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
3. The height of the implicit equation
As an application of the results in the previous sections, we obtain sharp bounds for
the exponents and the coefficients of an equation defining the closure of the image of
an affine variety under a rational map. We consider separately the cases K a field of
functions and K = Q. Besides of their intrinsic interest, these results play a central
role in our treatment of the parametric and arithmetic Nullstellensätze.
In the sequel, we will be mostly concerned with affine varieties instead of multiprojective cycles. Given n ∈ Nm , we identify the affine space
n1
nm
An
K := AK × · · · × AK
with the dense open subset Pn
K \ V (x1,0 , . . . , xm,0 ). This identification allows to transn
fer notions and results from Pn
K to AK . For instance, subvarieties, cycles and divisors
n
of AK correspond to the restriction of subvarieties, cycles and divisors of Pn
K to this
open subset. Thus, we can consider their mixed degrees and heights, which will correspond to the analogous notions for their closure in the corresponding multiprojective
space. In particular, the degree and the height of a subvariety of AnK are respectively
defined as the degree and the height of its closure in PnK .
For n ∈ N, the set of K-varieties of AnK is in one-to-one correspondence with the set of
radical ideals of K[x1 , . . . , xn ]. For an affine K-variety V ⊂ AnK , we denote by V (K)
n
the set of closed points of VK . It can be identified with the subset of K defined by
I(V ). A polynomial relation is said to hold on V if it holds for every point of V (K).
Since Q is a perfect field, a Q-variety of AnQ can be identified with a subvariety of
An (Q) defined over Q, see Remark 1.1.
3.1. The function field case. Let k be an arbitrary field and consider the groups of
variables t = {t1 , . . . , tp } and x = {x1 , . . . , xn }. The following result is a parametric
analogue of Perron’s theorem on the size of an equation of algebraic dependence for
r + 1 polynomials over a variety of dimension r.
Theorem 3.1. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and q1 , . . . , qr+1
∈ k[t, x] \ k[t]. Set dj = degx (qj ), hj = degt (qj ) for 1 ≤ j ≤ r + 1 and write
d = (d1 , . . . , dr+1 ), h = (h1 , . . . , hr+1 ). Then there exists
X
E=
αa y a ∈ k[t][y1 , . . . , yr+1 ] \ {0}
a∈Nr+1
satisfying E(q1 , . . . , qr+1 ) = 0 on V and such that, for all a ∈ supp(E),
• hd, ai ≤
r+1
Y
j=1
dj deg(V ),
• deg(αa ) + hh, ai ≤
r+1
Y
j=1
dj
r+1
X
hℓ
.
h(V ) + deg(V )
dℓ
ℓ=1
Our proof below follows Jelonek’s approach in [Jel05, Thm. 3.3] and heavily relies on
the arithmetic intersection theory developed in the previous sections.
Lemma 3.2. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and l ∈ N. Then
deg V × Alk(t) = deg(V ) , h V × Alk(t) = h(V ).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
51
Proof. Consider the standard inclusions ι1 , ι2 and ι3 of An , Al−1 and An+l into Pn ,
Pl−1 and Pn+l , respectively. We can easily verify that ι3 (V × Al ) = ι1 (V )#ι2 (Al−1 ).
Applying propositions 2.24 and 2.13(2),
[V × Al ]k(t) = [V ]k(t) · [Al−1 ]k(t) = [V ]k(t) ,
which implies the statement.
Proof of Theorem 3.1. Consider first the case when the map
ψ : V −→ Ar+1
k(t)
x 7−→ (q1 (x), . . . , qr+1 (x))
,
is generically finite onto its image, that is, when the fiber of a generic point in Im(ψ) is
finite. In our setting, this implies that Im(ψ) is a hypersurface. Let E ∈ k[t][y]\{0} be
a primitive and squarefree polynomial defining this hypersurface. Let v, w1 , . . . , wr+1
be a set of auxiliary variables and write
d
r+1
P = E(w1 + v h1 y1d1 , . . . , wr+1 + v hr+1 yr+1
) ∈ k[t, v][w, y].
Hence,
(3.3)
maxa {hd, ai} = degw,y (P )
,
maxa {deg(αa ) + hh, ai} = degt,v (P ).
We have that E is primitive and squarefree as a polynomial in k[t, v][w, y]. Hence the
d
same holds for P , since the map defined by yj 7→ wj + v hj yj j , wj 7→ yj is the identity
on k[t, v] and an automorphism of k[t, v][w, y] which sends P to E.
Write V (P ) for the k(t, v)-hypersurface of A2r+2
defined by P . By Proposition 2.13(4),
k(t,v)
(3.4)
degw,y (P ) = deg(V (P ))
,
degt,v (P ) = h(V (P )).
We will bound both the degree and the height of this hypersurface. Consider the
following subvarieties of An+2r+2
:
k(t,v)
G = (V × A2r+2 ) ∩
r+1
\
j=1
V (yj − qj ) ,
G(d, h) = (V × A2r+2 ) ∩
r+1
\
j=1
d
V (wj + v hj yj j − qj ).
Both varieties are of pure dimension 2r + 1: indeed, each of them is defined by a set of
r+1 polynomials which form a complete intersection over V ×A2r+2 as they depend on
d
different variables yj . Let ρ : Ar+1 → Ar+1 be the map defined by yj 7→ wj + v hj yj j
and π the projection An+2r+2 ≃ An × A2r+2 → A2r+2 . We have a commutative
diagram
G(d, h)
IdAn ×Ar+1 ×ρ
/G
π
π
V (P )
IdAr+1 ×ρ
/ V (E)
Both G(d, h) → G and G → V (E) are generically finite since ψ is generically finite,
and so this is also the case for π : G(d, h) → V (P ). Proposition 2.22 then implies
that
(3.5)
deg(V (P )) ≤ deg(G(d, h)) ,
h(V (P )) ≤ h(G(d, h)).
52
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
d
d
Observe that degx,w,y (v hj yj j + wj − qj ) = dj and degt,v (v hj yj j + wj − qj ) = hj . Hence,
since G(d, h) is an open set of a component defined by the homogenization of these
equations in Pn+2r+2 ,
r+1
r+1
Y
Y
dj deg(V × A2r+2 ) =
dj deg(V ),
deg(G(d, h)) ≤
j=1
h(G(d, h)) ≤
≤
r+1
Y
j=1
r+1
Y
j=1
j=1
r+1
X
hℓ
2r+2
2r+2
h(V × A
) + deg(V × A
)
dj
dℓ
ℓ=1
r+1
X
hℓ
dj
h(V ) + deg(V )
,
dℓ
ℓ=1
thanks to Corollary 2.19 and Lemma 3.2. The statements follows from these bounds,
together with (3.3), (3.4) and (3.5).
The general case reduces to the generically finite one by a deformation argument.
Choose variables xi1 , . . . , xir+1 among those in the group x in such a way that the
projection V → Ar+1
, x 7→ (xi1 , . . . , xir+1 ) is generically finite onto its image. Adding
k(t)
a further variable z, we consider the map
Vk(z,t) −→ Ar+1
, x 7−→ q1 (x) + z xi1 , . . . , qr+1 (x) + z xir+1 .
k(z,t)
It is also generically finite onto its image. Thus, we are in the hypothesis of the previous
case with respect to the base ring k(z)[t]. We deduce that there is a polynomial
a
e =P α
E
a ea y ∈ k(z)[t][y] \ {0} defining the closure of the image of this map and
e
satisfying, for all a ∈ supp(E),
r+1
r+1
r+1
Y
Y
X
hℓ
.
dj deg(V ) , degt (e
dj
hd, ai ≤
αa )+hh, ai ≤
h(V )+deg(V )
dℓ
j=1
j=1
ℓ=1
After multiplying by a suitable non-zero polynomial in z, we can assume without loss
e lies in k[z][t, y] and that it is primitive as a polynomial in the
of generality that E
e t, y) ∈ k[t, y] \ {0}. We have that E(q
e 1 + z xi , . . . , qr+1 +
variables t, y. Set E = E(0,
1
z xir+1 ) ∈ I(V ) ⊗ k(t)[z] and so E(q) ∈ I(V ) or, equivalently, E(q) = 0 on V . We
deduce that E = 0 is an equation of algebraic dependence for q1 , . . . , qr+1 which
e
satisfies the same bounds as E.
For V = Ank(t) we have r = n, deg(V ) = 1 and h(V ) = 0. The above result gives the
bounds
n+1
n+1
XY
Y
dj hℓ
dj , deg(αa ) + hh, ai ≤
(3.6)
hd, ai ≤
j=1
ℓ=1
j6=ℓ
for the y-degree and the t-degree of an equation of algebraic dependence for n + 1
polynomials in k[t, x] of x-degree dj ≥ 1 and t-degree hj .
Example 3.7. Let d1 , d2 ∈ N be coprime integers and set qj = gj xdj − 1 ∈ k[t, x] for
a generic univariate polynomial gj ∈ k[t] of degree hj , j = 1, 2. The implicit equation
of the closure of the image of
A1k(t) −→ A2k(t)
,
x 7−→ (q1 (x), q2 (x))
is g2d1 (y1 + 1)d2 − g1d2 (y2 + 1)d1 = 0. The bounds (3.6) are optimal in this case.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
53
Theorem 3.1 can be regarded
for the Newton polytope of the equaP as an cestimate
a ∈ k[t, y] with γ
tion E: if we write E =
γ
t
y
a,c ∈ k, the corresponding
a,c a,c
Newton polytope is the convex hull
N(E) = conv (a, c) : γa,c 6= 0 ⊂ Rr+1 × Rp .
Theorem 3.1 is equivalent to the statement that N(E) is contained in the intersection
of the non-negative orthant with the two half-spaces defined by the inequalities
r+1
r+1
r+1
X
Y
Y
hℓ
h(V ) + deg(V )
dj
dj deg(V ) , h1, ci + hh, ai ≤
hd, ai ≤
,
dℓ
j=1
j=1
ℓ=1
Np .
An
where 1 = (1, . . . , 1) ∈
Indeed, it can be shown that, for V = k(t) and generic
polynomials qj of x-degree dj ≥ 1 and t-degree hj , the Newton polytope of E coincides
with the set cut out by these inequalities.
For instance, consider in Example 3.7 the case when d2 h1 ≥ d1 h2 . Then, the Newton polytope of E is the convex hull of the points (0, 0, 0), (d2 , 0, 0), (d2 , 0, d1 h2 ),
(0, 0, d2 h1 ), (0, d1 , d2 h1 ), (0, d1 , 0), as shown in the figure:
t
(0, 0, d2 h1 )
(0, d1 , d2 h1 )
(d2 , 0, d1 h2 )
(0, 0, d1 h2 )
y1
(d2 , 0, 0)
y2
(0, d1 , 0)
It coincides with the subset of R3 cut out by the inequalities (3.6), namely
(a1 , a2 , c) : a1 , a2 , c ≥ 0, d1 a1 + d2 a2 ≤ d1 d2 , c + h1 a1 + h2 a2 ≤ d1 h2 + d2 h1 .
Example 3.8. Consider the elliptic curve C = V (t + 1)x31 + x21 − x22 ⊂ A2k(t) and
the polynomials q1 = x1 + (t + 1)x2 − 1, q2 = x1 x2 + (t − 1)x22 + t ∈ k[t, x1 , x2 ]. The
implicit equation of the closure of the image of
C −→ A2k(t) , (x1 , x2 ) 7−→ q1 (x1 , x2 ), q2 (x1 , x2 )
is defined by a polynomial E ∈ k[t, y1 , y2 ] with 138 terms. Its Newton
polytope is the
polytope conv (0, 0, 0), (6, 0, 0), (6, 0, 5), (0, 0, 11), (0, 3, 8), (0, 3, 0) ⊂ R3 :
t
(0, 0, 11)
(6, 0, 5)
(0, 3, 8)
(6, 0, 0)
y2
y1
(0, 3, 0)
It also coincides with the polytope cut out by the inequalities in the parametric
Perron’s theorem: we have that deg(C) = 3 and h(C) = 1 while degx (q1 ) = 1,
54
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
degx (q2 ) = 2 and degt (qi ) = 1, i = 1, 2. Theorem 3.1 implies the inclusion of polytopes
N(E) ⊂ (a1 , a2 , c) : a1 , a2 , c ≥ 0, a1 + 2a2 ≤ 6, c + a1 + a2 ≤ 11 ,
which turns out to be an equality.
The following result is an upper bound for the degree and the height of the implicit
equation of a hypersurface defined as the closure of the image of a general rational
map. These estimates are not used in the proof of the arithmetic Nullstellensätze.
Nevertheless, we include them because they may have some independent interest.
Recall that a rational map ρ : V 99K Ar+1
is defined by quotients of polynomials
k(t)
whose denominators do not vanish identically on any of the components of V . It is
generically finite onto its image if the fiber of a generic point in Im(ρ) is finite. In our
setting, this implies that the closure of its image is a hypersurface.
Theorem 3.9. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and
q1
qr+1
r+1
(x), . . . ,
(x)
ψ : V 99K Ak(t) , x 7−→
p1
pr+1
a rational map, generically finite onto its image, defined by polynomials qj , pj ∈ k[t, x]
such that qj /pj ∈
/ k(t). Let E ∈ k[t, y] be a primitive and squarefree polynomial
defining Im(ψ). Set dj = max{degx (qj ), degx (pj )} and hj = max{degt (qj ), degt (pj )}
for 1 ≤ j ≤ r + 1. Then
Y
• degyi (E) ≤
dj deg(V ) for 1 ≤ i ≤ r + 1,
j6=i
• degt (E) ≤
Proof. Let U = V \
The graph of ψ is
Sr+1
j=1
r+1
Y
j=1
dj
r+1
X
hℓ
.
h(V ) + deg(V )
dℓ
ℓ=1
V (pj ) be the dense open subset of V where ψ is defined.
{(x, y) : x ∈ U, yj pj (x) = qj (x) for 1 ≤ j ≤ r + 1} ⊂ U × Ar+1 (k(t)).
Let G be the closure of this set in V × Ar+1
. The equations yj pj = qj intersect
k(t)
r+1
properly on U × Ak(t) because they depend on different variables yj . Hence, G is an
equidimensional variety of dimension r, and the projection V × Ar+1 → Ar+1 induces
a generically finite map between G and V (E).
We consider mixed degrees and heights of G and V (E) with respect to the inclusions
V × Ar+1 ֒→ Pn × (P1 )r+1
,
Ar+1 ֒→ (P1 )r+1 .
Set 1 = (1, . . . , 1) ∈ Nr+1 and let ei denote the i-th vector of the standard basis
of Rr+1 . By Proposition 2.13(4),
(3.10)
degyi (E) = deg1−ei (V (E))
,
degt (E) = h1 (V (E)).
Proposition 2.22 applied to the projection Pn × (P1 )r+1 → (P1 )r+1 implies that
(3.11)
deg1−ei (V (E)) ≤ deg0,1−ei (G)
,
h1 (V (E)) ≤ h0,1 (G).
2 ) be the class of G in the extended Chow
Let [G]k[t] ∈ Z[η, θ0 , θ]/(η 2 , θ0n+1 , θ12 , . . . , θr+1
ring of Pn × (P1 )r+1 . Observe that G is contained in the part of dimension r of the
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
intersection
(V × A
r+1
)∩
r+1
\
j=1
55
V (yj pj − qj ).
We will give an upper bound for the class of G by applying Corollary 2.19 recursively
to the variety V × Ar+1 and the polynomials gj = yj pj − qj for j = 1, . . . , r + 1. By
propositions 2.23 and 2.13(2),
[V × Ar+1 ]k[t] = [V ]k[t] ⊗ [A1 ]⊗(r+1)
= [V ]k[t] = h(V )ηθ0n−r−1 + deg(V )θ0n−r .
k[t]
Hence,
[G]k[t] ≤ [V × A
r+1
]k[t]
r+1
Y
degt (gj )η + degx (gj )θ0 +
j=1
r+1
X
degyℓ (gj )θℓ
ℓ=1
= h(V )ηθ0n−r−1 + deg(V )θ0n−r
r+1
Y
(hj η + dj θ0 + θj ).
j=1
Observe that h0,1 (G) and deg0,1−ei (G) are the coefficients in [G]k[t] of the monomials η θ0n and θ0n θi , respectively. Therefore
deg0,1−ei (G) ≤
Y
j6=i
dj deg(V )
h0,1 (G) ≤
,
r+1
Y
j=1
r+1
X
hℓ
dj h(V ) + deg(V )
.
dℓ
ℓ=1
The statement follows from these estimates, together with (3.10) and (3.11).
For V = Ank(t) , the above result gives the bounds
(3.12)
degyi (E) ≤
Y
j6=i
dj
,
degt (E) ≤
n+1
XY
ℓ=1
j6=ℓ
dj hℓ .
It can be shown that, for generic polynomials of x-degree dj and t-degree hj , the Newton polytope of E coincides with the subset of Rr+1 × Rp cut out by these inequalities.
Example 3.13. Let d1 , d2 ∈ N be coprime integers and gi ∈ k[t] univariate polynomials of degree hi , i = 1, 2. Set
xd1
q1
=
−1 ,
p1
(x + 1)d1 g1
q2
(x + 1)d2
=
− 1.
p2
xd2 g2
The implicit equation of the closure of the image of the map
q
q2
1
(x), (x)
A1k(t) 99K A2k(t) , x 7−→
p1
p2
is given by the polynomial E = g1 d2 g2 d1 (y1 + 1)d2 (y2 + 1)d1 − 1 ∈ k[t, y1 , y2 ]. Theorem 3.9 is optimal in this case, since degy1 (E) = d2 , degy2 (E) = d1 and degt (E) =
d1 h2 + d2 h1 . Moreover, we can check that the Newton polytope of E coincides with
the set cut out by the inequalities in (3.12).
The previous results can be extended to polynomials depending on groups of parameters. For the sequel, we will need the multiparametric version of Theorem 3.1 that
we state below. Let tl = {tl,1 , . . . , tl,pl } be a group of variables for 1 ≤ l ≤ m and
set t = {t1 , . . . , tm }. For each 1 ≤ l ≤ m, write kl = k(t1 , . . . , tl−1 , tl+1 , . . . , tm )
and observe that kl (tl ) = k(t). Hence, for a given projective k(t)-variety V , we can
consider its height with respect to the base ring kl [tl ]. We denote this height htl (V ).
56
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Corollary 3.14. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and q1 , . . . , qr+1 ∈
k[t][x] \ k[t] such that the map
ψ : V −→ Ar+1
x 7−→ (q1 (x), . . . , qr+1 (x))
P
is generically finite onto its image. Let E = a∈Nr+1 αa y a ∈ k[t][y1 , . . . , yr+1 ] be
a primitive and squarefree polynomial defining Im(ψ). Set dj = degx (qj ), hl,j =
degtl (qj ) for 1 ≤ j ≤ r + 1, 1 ≤ l ≤ m and write d = (d1 , . . . , dr+1 ), hl =
(hl,1 , . . . , hl,r+1 ). Then, for all a ∈ supp(E),
r+1
Y
dj deg(V ),
• hd, ai ≤
j=1
r+1
r+1
Y
X
hl,ℓ
dj
• degtl (αa ) + hhl , ai ≤
htl (V ) + deg(V )
for 1 ≤ l ≤ m.
dℓ
,
j=1
ℓ=1
Proof. Observe that E is primitive and squarefree as an element of kl [tl ][y]. The result
then follows from Theorem 3.1 applied to the field kl and the group of variables tl .
3.2. The rational case. We now turn our attention to the problem of estimating the
size of the implicit equation for a rational map defined over Q. It will be convenient
to consider a more general situation where the input polynomials depend on groups
of parameters. As in the end of the previous section, we set tl = {tl,1 , . . . , tl,pl } for
1 ≤ l ≤ m and t = {t1 , . . . , tm }.
Theorem 3.15. Let V ⊂ AnQ be a Q-variety of pure dimension r and q1 , . . . , qr+1 ∈
Z[t][x] \ Z[t] such that the map
ψ : VQ(t) −→ Ar+1
Q(t)
x 7−→ (q1 (x), . . . , qr+1 (x))
P
a ∈ Z[t][y , . . . , y
is generically finite onto its image. Let E =
1
r+1 ]
a∈Nr+1 αa y
be a primitive and squarefree polynomial defining Im(ψ) Set dj = degx (qj ), δl,j =
degtl (qj ), hj = h(qj ) for 1 ≤ j ≤ r + 1, 1 ≤ l ≤ m and write d = (d1 , . . . , dr+1 ),
δ l = (δl,1 , . . . , δl,r+1 ), h = (h1 , . . . , hr+1 ). Then, for all a ∈ supp(E),
• hd, ai ≤
r+1
Y
j=1
,
dj deg(V ),
r+1
X
δl,ℓ
dj deg(V )
for 1 ≤ l ≤ m,
dℓ
j=1
ℓ=1
r+1
Y
b
• h(αa ) + hh, ai ≤
dj h(V ) + deg(V ) log(r + 2)
• degtl (αa ) + hδ l , ai ≤
r+1
Y
j=1
m
r+1
X
X
1
δl,ℓ log(pl + 1)
hℓ + log #supp(qℓ ) + 2 +
.
+
dℓ
l=1
ℓ=1
We need the following lemma for the proof of this result. It is the analogue for Q of
Lemma 3.2 and can be proved using the same arguments and Proposition 2.67 instead
of 2.24.
Lemma 3.16. Let V ⊂ AnQ be a Q-variety of pure dimension r and l ∈ N. Then
deg(V × AlQ ) = deg(V )
,
b
h(V × AlQ ) = b
h(V ).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
57
Proof of Theorem 3.15. By Proposition 2.13(2), the height htl of VQ(t) with respect to
the base ring Q(t1 , . . . , tl−1 , tl+1 , . . . , tm )[tl ] is zero. Then, the bounds for the degrees
are a direct consequence of Corollary 3.14. For the height, the proof follows closely
the lines of that of Theorem 3.1 with Z instead of k[t]. We will avoid repeating the
same verifications when they follow mutatis mutandis the parametric case.
Let w1 , . . . , wr+1 be a group of variables and consider the polynomial
P = E(w1 + H1 y1 , . . . , wr+1 + Hr+1 yr+1 ) ∈ Z[t, w, y],
with Hj = ehj ∈ N. It verifies
maxa {h(αa ) + hh, ai} = h(P (0, y)).
(3.17)
Applying successively lemmas 2.32(2) and 2.31(4),
h(P (0, y)) ≤ m(P (0, y)) + degy (P (0, y)) log(r + 2) +
(3.18)
≤ m(P ) + degy (E) log(r + 2) +
m
X
m
X
degtl (P (0, y)) log(pl + 1)
l=1
degtl (E) log(pl + 1).
l=1
The polynomial P is primitive and squarefree. Hence, it gives a defining equation for
, where p := (p1 , . . . , pm ). Considering the
the Q-hypersurface V (P ) ⊂ ApQ × A2r+2
Q
standard inclusion of this affine space into PpQ × P2r+2
, Proposition 2.53(2) implies
Q
m(P ) = b
hp,2r+2 (V (P )).
(3.19)
Consider the variety
G(d, h) = (Ap × V × A2r+2 ) ∩
r+1
\
j=1
V (wj + Hj yj − qj ) ⊂ Ap × An+2r+2 .
As in the proof of Theorem 3.1, we can verify that it is of pure dimension |p| + 2r + 1
and that the projection Ap × An+2r+2 → Ap × A2r+2 induces a generically finite map
π : G(d, h) → V (P ). Considering mixed heights with respect to the standard inclusion
Ap × An+2r+2 ֒→ Pp × Pn+2r+2 , Proposition 2.64 implies
b
hp,2r+2 (V (P )) ≤ b
hp,2r+2 (G(d, h)).
(3.20)
The closure of Ap × V × A2r+2 is Pp × ι(V × A2r+2 ), where ι denotes the standard
inclusion An+2r+2 ֒→ Pn+2r+2 . We will consider classes in the extended Chow ring
A∗ (Pp × Pn+2r+2 ; Z) = R[η, θ, ζ]/(η 2 , θ1p1 +1 , . . . , θlpl +1 , ζ n+2r+3 ).
With this notation, propositions 2.66(2) and 2.51(3) together with Lemma 3.16 imply
[Ap × V × A2r+2 ]Z = [Ap ]Z ⊗ [V × A2r+2 ]Z = [V ]Z = b
h(V )ηζ n−r−1 + deg(V )ζ n−r .
Write gj = wj + Hj yj − qj ∈ Z[t, x, w, y] and consider the class associated to its
sup-norm as defined in (2.57):
[gj ]sup = log ||gj ||sup η +
m
X
l=1
degtl (gj )θl + degx,w,y (gj )ζ = log ||gj ||sup η +
m
X
l=1
δl,j θl + dj ζ.
58
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
The divisors defined by these polynomials intersect Ap ×V ×A2r+2 properly. Applying
recursively Corollary 2.61,
[G(d, h)]Z ≤ [Ap × V × A2r+2 ]Z
r+1
Y
[gj ]sup
j=1
= b
h(V )ηζ n−r−1 + deg(V )ζ n−r
Y
r+1
j=1
log ||gj ||sup η +
m
X
l=1
δl,j θl + dj ζ .
The mixed height b
hp,2r+2 (G(d, h)) is the coefficient of η ζ n in [G(d, h)]Z . Furthermore,
log ||gℓ ||sup ≤ hℓ +log #supp(qℓ )+2 by Lemma 2.32(1). The above inequality implies
r+1
r+1
X
Y
hℓ + log #supp(qℓ ) + 2
b
b
h(V ) + deg(V )
.
dj
hp,2r+2 (G(d, h)) ≤
dℓ
j=1
ℓ=1
The statement follows from this inequality together with (3.17), (3.18), (3.19), (3.20),
and the already considered bounds for the partial degrees of E.
Theorem 0.4 in the introduction follows from the case m = 0 of this result and the
inequality #supp(qℓ ) + 2 ≤ (n + 3)dℓ .
Remark 3.21. Theorem 0.4 can be regarded as a first bound for the extended Newton
polytope of the implicit equation, defined as the convex envelope of the subset
{(a, λ) : a ∈ supp(E), 0 ≤ λ ≤ h(αa )} ⊂ Rr+1 × R.
Indeed, it would be interesting to have a better understanding of this “arithmetic”
polytope in terms of finer invariants of the input polynomials qj like for instance their
extended Newton polytope, instead of just their degree and height.
Example 3.22. Let d1 , d2 ∈ N and H1 , H2 ∈ N be two pairs of coprime integers and
set qj = Hj xdj ∈ Z[x] for j = 1, 2. The implicit equation of the image of the map
A1Q −→ A2Q
,
x 7−→ (q1 (x), q2 (x))
is given by the polynomial E = H2d1 y1d2 − H1d2 y2d1 ∈ Z[y1 , y2 ]. Then, max(a1 ,a2 ) {d1 a1 +
d2 a2 } = d1 d2 and
max(a1 ,a2 ) {log |coeff (a1 ,a2 ) (E)| + h1 a1 + h2 a2 } = d1 h2 + d2 h1 ,
with hj = log(Hj ). Hence, the bounds in Theorem 0.4 are optimal in this example,
up to a term of size O(d1 d2 ).
Using a deformation argument, we can extend Theorem 3.15 to the case when the map
is not generically finite. For simplicity, we will only state this result for polynomials
not depending on parameters.
Corollary 3.23. Let V ⊂ AnQ be a Q-variety of pure dimension r and q1 , . . . , qr+1 ∈
Z[x] \ Z. Set dj = deg(qj ) and hj = h(qj ) for 1 ≤ j ≤ r + 1 and write d =
(d1 , . . . , dr+1 ), h = (h1 , . . . , hr+1 ). Then there exists
X
E=
αa y a ∈ Z[y1 , . . . , yr+1 ] \ {0}
a∈Nr+1
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
59
satisfying E(q1 , . . . , qr+1 ) = 0 on V and such that, for all a ∈ supp(E),
• hd, ai ≤
r+1
Y
j=1
dj deg(V ),
• h(αa ) + hh, ai ≤
r+1
Y
dj
j=1
r+1
X
hℓ
b
+ (r + 2) log(2n + 8) .
h(V ) + deg(V )
dℓ
ℓ=1
Proof. We will follow the arguments in the proof of Theorem 3.1 adapted to this
situation. Choose r + 1 variables xi1 , . . . , xir+1 in the group x so that the linear
projection V → Ar+1
Q , x 7→ (xi1 , . . . , xir+1 ) is generically finite onto its image. Adding
a further variable z, consider the map
x 7−→ q1 (x) + z xi1 , . . . , qr+1 (x) + z xir+1 ,
VQ(z) −→ Ar+1
Q(z) ,
e = P αa y a ∈ Z[z][y] be a primitive
which is generically finite onto its image. Let E
a
and squarefree polynomial defining the closure of the image of this map. The polynomials qj (x) + z xij have x-degree dj , z-degree 1 and height hj . Theorem 3.15 applied
Qr+1
to this case gives hd, ai ≤
j=1 dj deg(V ) and
h(αa ) + hh, ai ≤
r+1
Y
j=1
dj
b
h(V ) + deg(V ) log(r + 2)
r+1
X
1
+
hℓ + log #supp(qℓ ) + 3 + log(2)
.
dℓ
ℓ=1
e y) ∈ Z[y] gives a non-trivial relation of algebraic deThe polynomial E := E(0,
e The
pendence for the qj ’s and it satisfies the same degree and height bounds as E.
d
statement then follows from the inequality #supp(qℓ )+3 ≤ (n+1) ℓ +3 ≤ (n+4)dℓ .
Next result gives an upper bound for the Mahler measure and a fortiori, for the height
of the implicit equation for a rational map.
Theorem 3.24. Let V ⊂ AnQ be a Q-variety of pure dimension r and
q1
qr+1
r+1
, x 7→
ψ : V 99K AQ
(x), . . . ,
(x)
p1
pr+1
a rational map, generically finite onto its image, defined by polynomials qj , pj ∈ Z[x]
such that qj /pj ∈
/ Q. Let E ∈ Z[y] be a primitive and squarefree polynomial defining Im(ψ). Set dj = max{degx (qj ), degx (pj )} and hj = log(||qj ||1 + ||pj ||1 ) for
1 ≤ j ≤ r + 1. Then
Y
• degyi (E) ≤
dj deg(V ) for 1 ≤ i ≤ r + 1,
j6=i
• m(E) ≤
r+1
Y
j=1
dj
r+1
X
hℓ
b
.
h(V ) + deg(V )
dℓ
ℓ=1
Proof. The bounds for the partial degrees of E follow from Theorem 3.9. Thus, we
and
only have to prove the upper bound for the Mahler measure. Let V (E) ⊂ Ar+1
Q
60
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
denote the closure of the image of ψ and of the graph of ψ, respectively.
G ⊂ V × Ar+1
Q
We consider mixed heights of G and V (E) with respect to the inclusions
V × Ar+1 ֒→ Pn × (P1 )r+1
,
Ar+1 ֒→ (P1 )r+1 .
Propositions 2.53 and 2.64 imply that
m(E) = b
h1 (V (E)) ≤ b
h0,1 (G).
(3.25)
2 ) be the class of the closure of G in the
Let [G]Z ∈ R[η, θ0 , θ]/(η 2 , θ0n+1 , θ12 , . . . , θr+1
extended Chow ring of Pn × (P1 )r+1 . We will bound it by applying Corollary 2.61
recursively to V × Ar+1 and gj := yj pj − qj for j = 1, . . . , r + 1. Propositions 2.66(2)
and 2.51(3) imply that [V × Ar+1 ]Z = b
h(V )ηθ0n−r−1 + deg(V )θ0n−r . Besides, ||pj yj −
qj ||sup ≤ ||pj ||1 + ||qj ||1 and so [pj yj − qj ]sup ≤ hj η + dj θ0 + θj . Hence,
[G]Z ≤ [V ×A
r+1
]Z
r+1
Y
[pj yj −qj ]sup
j=1
≤ b
h(V )ηθ0n−r−1 +deg(V )θ0n−r
Y
r+1
hj η+dj θ0 +θj ).
j=1
h0,1 (G) is the coefficient of the monomial η θ0n in [G]Z . The above
The mixed height b
inequality then implies
b
h0,1 (G) ≤
r+1
Y
dj
j=1
b
h(V ) + deg(V )
r+1 Y
X
ℓ=1
j6=ℓ
dj hℓ .
The statement follows from this inequality and (3.25).
For V = AnQ , the above result gives the bounds
degyi (E) ≤
Y
dj
,
j6=i
m(E) ≤
n+1
XY
ℓ=1
j6=ℓ
dj hℓ
for the degree and the Mahler measure of Im(ψ). Using Lemma 2.32(2), we can bound
the height of this polynomial by
h(E) ≤
n+1
XY
ℓ=1
j6=ℓ
dj (hℓ + log(2)).
Example 3.26. Let d1 , d2 , H1 , H2 ≥ 1 such that d1 , d2 are coprime and set
q1
xd1
=
p1
H1 (x + 1)d1
,
q2
(x + 1)d2
=
.
p2
H2 xd2
The implicit equation of the closure of the image of
q2
q1
1
2
(x), (x)
AQ 99K AQ , x 7−→
p1
p2
is defined by E = H1 d2 H2 d1 y1d2 y2d1 −1 ∈ Z[y1 , y2 ]. We have max{deg(pj ), deg(qj )} = dj
while log(||q1 ||1 + ||p1 ||1 ) = log(1 + H1 2d1 ) and log(||q2 ||1 + ||p2 ||1 ) = log(2d2 + H2 ).
Hence, Theorem 3.24 gives
m(E) ≤ d2 log(1 + H1 2d1 ) + d1 log(2d2 + H2 ).
Indeed, m(E) = d1 log(H2 ) + d2 log(H1 ) and so the obtained bound is optimal up a
term of size O(d1 d2 ).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
61
4. Arithmetic Nullstellensätze
4.1. An effective approach to the Nullstellensatz. Recall the statement of the
weak Nullstellensatz over a variety:
Let V ⊂ AnK be a K-variety and f1 , . . . , fs ∈ K[x1 , . . . , xn ] without
common zeros in V (K). Then there exist g1 , . . . , gs ∈ K[x1 , . . . , xn ]
such that 1 = g1 f1 + · · · + gs fs on V .
We recall Jelonek’s approach in [Jel05] to produce bounds for the degree of the gi ’s
and explain how it can be adapted to obtain bounds for their height. Set r = dim(V ).
We assume without loss of generality that s ≤ r + 1, otherwise we reduce to this case
by taking generic linear combinations of the input polynomials. Consider the regular
map
ϕ : V × A1K −→V × AsK
(4.1)
,
(x, z) 7−→ (x, zf1 (x), . . . , zfs (x)) .
Since the fj ’s have no common zeros in V (K), the subsets {(x, z1 , . . . , zs ) : fj (x) 6= 0}
form an open covering of V × AsK . This implies that
Im(ϕ) = (x, z1 , . . . , zs ) ∈ V × AsK : fi (x)zj = fj (x)zi for all i, j .
In particular, Im(ϕ) is a subvariety of V × AsK . Furthermore, let g1 , . . . , gs be polynomials such that 1 = g1 f1 + · · · + gs fs . The map Im(ϕ) → V × A1K defined as
(x, z1 , . . . , zs ) 7→ (x, g1 (x)z1 + · · · + gs (x)zs ) is a left inverse of ϕ. Hence, ϕ induces
an isomorphism V × A1K → Im(ϕ).
Assume there exists a finite linear projection π : Im(ϕ) → Ar+1
K . Such a map exists if
the field K is sufficiently big (for instance, if it is infinite). In this case, we can assume
without loss of generality that π is given by a (r + 1) × (n + s)-matrix in reduced
triangular form. Hence, there are linear forms ℓi = γi,1 x1 + · · · + γi,n xn ∈ K[x],
1 ≤ i ≤ r + 1, such that this projection writes down as
π : Im(ϕ)→Ar+1
K , (x, z1 , . . . , zs ) 7→ z1 + ℓ1 (x), . . . , zs + ℓs (x), ℓs+1 (x), . . . , ℓr+1 (x) .
is a finite map or equivalently, the
Thus, the composition π ◦ ϕ : V × A1K →Ar+1
K
∗
inclusion of algebras (π ◦ ϕ) : K[y1 , . . . , yr+1 ] ֒→ K[x, z]/I(V × A1K ) is an integral
extension. This map writes down, for (x, z) ∈ V × A1K , as
π ◦ ϕ(x, z) = (zf1 (x) + ℓ1 (x), . . . , zfs (x) + ℓs (x), ℓs+1 (x), . . . , ℓr+1 (x)).
Up to a non-zero scalar in K, the minimal polynomial of z over K[y] is of the form
E = zδ +
δ
X
X
j=1 a∈Nr+1
Therefore,
δ
(4.2) z +
δ X
X
j=1 a
αa,j y a z δ−j ∈ K[y, z].
a
α
s+1
r+1 δ−j
αa,j (zf1 + ℓ1 )a1 · · · (zfs + ℓs )as ℓs+1
· · · ℓr+1
z
= 0 on V × A1
and so all the coefficients in the expansion of (4.2) with respect to the variable z
vanish identically on V . We derive a Bézout identity from the coefficient of z δ as
follows: for each 1 ≤ j ≤ δ and a ∈ Nr+1 , the coefficient of z j in the expansion of
as+1
αr+1
· · · ℓr+1
writes down as
(zf1 + ℓ1 )a1 · · · (zfs + ℓs )as ℓs+1
X
gb f1b1 · · · fsbs
b
62
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
with gb ∈ K[x] and b ∈ Ns such that |b| = j ≥ 1. Hence each term in this sum is a
multiple of some fj and, regrouping terms, we can write the coefficient of z δ in the
expansion of (4.2) as 1 − g1 f1 − · · · − gs fs with gi ∈ K[x].
Remark 4.3. The argument above is not a proof of the Nullstellensatz since it relies
on the a priori existence of a Bézout identity: this is used to prove that the map ϕ
in (4.1) is an isomorphism onto its image.
In our treatment of the arithmetic Nullstellensatz, we will use the previous construction for the general linear forms
(4.4)
ℓi := ui,1 x1 + · · · + ui,n xn ∈ K[ui ][x]
for 1 ≤ i ≤ r + 1,
where ui = {ui,j }1≤j≤n is a group of auxiliary variables. This is a valid choice,
since the associated linear projection π : Im(ϕ)K(u) → Ar+1
K(u) is a finite map, see for
instance [Lan93, Ch. 7, Thm. 2.2].
Assume that K is the field of fractions of a factorial ring A. Let E ∈ A[u][y, z] be the
minimal polynomial of z with respect to the map π ◦ ϕ, primitive with respect to the
ring A[u]. We expand it as
E = α0,0 z δ +
(4.5)
δ
X
X
αa,j y a z δ−j
j=1 a∈Nr+1
with αa,j ∈ A[u] such that α0,0 6= 0. For 1 ≤ i ≤ s, set
r+1
δ XX
X
Y ak a −b
bi−1 bi −1
b1
k
k
(4.6)
gi = −
e
αa,j
f1 · · · fi−1 fi
ℓ
∈ A[u][x],
bk k
a
j=1
k=1
b
the sums being indexed by all a, b ∈ Nr+1 such that (a, j) ∈ supp(E) and |b| = j,
bk ≤ ak for 1 ≤ k ≤ i, bk = 0 for i + 1 ≤ k ≤ r + 1 and bi ≥ 1. Using (4.2), we can
verify that α0,0 − ge1 f1 − · · · − ges fs is the coefficient of z δ in the expansion of
E(zf1 + ℓ1 , . . . , zfs + ℓs , ℓs+1 , . . . , ℓr+1 , z).
Hence,
(4.7)
α0,0 = ge1 f1 + · · · + ges fs
on VK(u) .
We then extract a Bézout identity on V by considering the coefficient of any monomial
in u in (4.7) appearing in the monomial expansion of α0,0 .
We need the following lemma relating minimal polynomials with the implicitization
problem.
Lemma 4.8. Let V ⊂ AnK be a K-variety of pure dimension r and q1 , . . . , qr+1 ∈
K[x, z] such that the map
V × A1K −→ Ar+1
K
,
(x, z) 7−→ (q1 (x, z), . . . , qr+1 (x, z))
is finite. Let E ∈ K[y][z] \ {0} be the minimal polynomial of z with respect to this
map. Then the map
ψ : VK(z) −→ Ar+1
K(z)
,
x 7−→ (q1 (x), . . . , qr+1 (x))
is generically finite onto its image. Furthermore, E ∈ K[z][y] is a squarefree polynomial, primitive with respect to the ring K[z], defining Im(Ψ).
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
63
Proof. We reduce without loss of generality to the case of an irreducible K-variety V .
Consider the map
1
ρ : V × A1K −→ Ar+1
K × AK
,
(x, z) 7−→ (q1 (x, z), . . . , qr+1 (x, z), z).
Its image is the irreducible K-hypersurface defined by E. In algebraic terms, the
kernel of ρ∗ : K[y, z] → K[x, z]/I(V × A1K ) is the principal ideal generated by E.
Furthermore, E ∈
/ K[z] as, otherwise, this would imply that the qj ’s are constant on
V × A1K , which is not possible because of the finiteness assumption.
We have K[x, z]/I(V × A1K ) = K[z] ⊗K K[x]/I(V ). Hence, E is also a generator of
the kernel of ψ ∗ : K(z)[y] → K(z)[x]/I(VK(z) ). The image of ψ is a hypersurface
of Ar+1
K(z) , since E does not lie in K[z]. Hence, this map is generically finite onto
its image. Furthermore, E is an irreducible polynomial defining this hypersurface,
primitive with respect to the ring K[z].
Consider the polynomials qj ∈ A[u][x, z] defined by
(
zfj + ℓj , for 1 ≤ j ≤ s,
(4.9)
qj =
ℓj ,
for s + 1 ≤ j ≤ r + 1.
In our setting, π ◦ ϕ : VK(u) × A1K(u) −→Ar+1
K(u) is finite. Hence, Lemma 4.8 implies
that the polynomial E in (4.5) is a squarefree polynomial, primitive with respect to
the ring A[u, z], defining the closure of the image of the generically finite map
(4.10)
ψ : VK(u,z) −→ Ar+1
K(u,z)
,
x 7−→ (q1 (x), . . . , qr+1 (x)).
Thus, we can produce bounds for its size by using a suitable version of Perron’s
theorem. In turn, this will allow us to bound the size of the polynomials in the
associated Bézout identity. We will see the details in the next sections.
4.2. Parametric Nullstellensätze. We now apply the previous construction together with the parametric Perron’s theorem to produce different Nullstellensätze
for polynomials with coefficients depending on groups of parameters. Let k be a field.
Consider groups of variables x = {x1 , . . . , xn } and tl = {tl,1 , . . . , tl,pl }, 1 ≤ l ≤ m.
Set t = {t1 , . . . , tm } and, for each 1 ≤ l ≤ m, write kl = k(t1 , . . . , tl−1 , tl+1 , . . . , tm ).
Recall that kl (tl ) = k(t) and that for a projective k(t)-variety V , we denote by htl (V )
its height with respect to the base ring kl [tl ].
Theorem 4.11. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and f1 , . . . , fs ∈
k[t, x] \ k[t] a family of s ≤ r + 1 polynomials without common zeros in V (k(t)). Set
dj = degx (fj ) and hl,j = degtl (fj ) for 1 ≤ j ≤ s and 1 ≤ l ≤ m. Then there exist
α ∈ k[t] \ {0} and g1 , . . . , gs ∈ k[t, x] such that
α = g1 f1 + · · · + gs fs
(4.12)
on V
with
• degx (gi fi ) ≤
s
Y
j=1
dj deg(V ),
• degtl (α), degtl (gi fi ) ≤
s
Y
j=1
dj
htl (V ) + deg(V )
s
X
hl,ℓ
ℓ=1
dℓ
for 1 ≤ l ≤ m.
64
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof. We apply the construction explained in §4.1 to the ring A := k[t], the variety V
and the polynomials fj . We will freely use the notations introduced in that section.
As a result of that construction, we obtain α ∈ A \ {0} and gi ∈ A[x] satisfying the
Bézout identity (4.12). These scalar and polynomials are obtained as the coefficients
of a monomial in the variables u in the Bézout identity on VK(u) in (4.7). Hence,
gi fi ).
(4.13) degx (gi fi ) ≤ degx (e
gi fi ) , degtl (α) ≤ degtl (α0,0 ) , degtl (gi fi ) ≤ degtl (e
Let E ∈ A[u][y, z] be the polynomial in (4.5) and set d = (d1 , . . . , ds , 1, . . . , 1) ∈ Nr+1 .
From the definition of e
gi in (4.6) and using that, by hypothesis, di ≥ 1 for all i,
(4.14)
degx (e
gi fi ) ≤ maxj,a,b
r+1
Y ae
bi−1 bi
b1
ae −be
degx αa,j
ℓ
f1 · · · fi−1 fi
be e
e=1
i
r+1
o
nX
X
dℓ bℓ ≤ maxa hd, ai,
≤ maxa,b
(ae − be ) +
e=1
ℓ=1
Nr+1
with 1 ≤ j ≤ degz (E), a ∈
in the support of E with respect to the variables y
r+1
and b ∈ N
satisfying |b| = j, be ≤ ae for 1 ≤ e ≤ i, be = 0 for i + 1 ≤ e ≤ r + 1 and
bi ≥ 1. For 1 ≤ l ≤ m we set hl = (hl,1 , . . . , hl,s , 0, . . . , 0) ∈ Nr+1 . Using Lemma 2.1,
we obtain similarly
r+1
Y ae
bi−1 bi
b1
ae −be
gi fi ) ≤ maxj,a,b degtl αa,j
degtl (e
ℓ
f1 · · · fi−1 fi
(4.15)
be e
e=1
i
n
o
X
≤ maxj,a,b degtl (αa,j ) +
hl,ℓ bℓ
ℓ=1
≤ maxa degtl (αa ) + hhl , ai .
By Lemma 4.8, E is a primitive and squarefree polynomial in k[t, u, z][y] defining the
closure of the image of the map ψ in (4.10). Hence, we can apply Corollary 3.14 to
bound the partial degrees of this polynomial. We have degx (qj ) = dj , degtl (qj ) =
hl,j for 1 ≤ j ≤ s while degx (qj ) = 1, degtl (qj ) = 0 for s + 1 ≤ j ≤ r + 1. By
Proposition 2.13(2), deg(Vk(t,u,z) ) = deg(V ) and htl (Vk(t,u,z) ) = htl (V ). Therefore,
for all a ∈ supp(E),
Y
Y
s
s
s
X
hl,ℓ
.
dj
dj deg(V ) , degtl (αa )+hhl , ai ≤
hd, ai ≤
htl (V )+deg(V )
dℓ
j=1
j=1
ℓ=1
The statement follows from this inequality together with (4.13), (4.14) and (4.15).
Theorem 0.5 in the introduction corresponds to the case of polynomials depending on
one group of parameters. It follows readily from the above result and the fact that a
variety defined over k(t) can be identified with a k(t)-variety, see Remark 1.1.
Example 4.16. Consider the following variant of a classical example due to Masser
and Philippon: let t be a variable, d1 , . . . , dn ≥ 1, h ≥ 0 and set
d
n−1
f1 = xd11 , f2 = x1 xdn2 −1 − xd22 , . . . , fn−1 = xn−2 xdnn−1 −1 − xn−1
,
fn = xn−1 xndn −1 − th ∈ k[t, x1 , . . . , xn ].
n
It is a system of n polynomials without common zeros in k(t) . We have degx (fj ) = dj
for all j while degt (fj ) = 0 for 1 ≤ j ≤ n − 1 and degt (fn ) = h. Theorem 4.11 implies
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
65
that there exists a Bézout identity α = g1 f1 + · · · + gn fn with
(4.17)
degx (gi fi ) ≤ d1 · · · dn
,
deg(α), deg t (gi fi ) ≤ d1 · · · dn−1 h.
To obtain a lower bound, consider a further variable u and specialize any such Bézout
identity at xi = γi with
γ1 = td2 ···dn−1 h ud2 ···dn −1 , . . . , γn−2 = tdn−1 h udn−1 dn −1 , γn−1 = th udn −1 , γn = 1/u.
We obtain that α = g1 (γ1 , . . . , γn−1 , 1/u) td1 ···dn−1 h ud1 ···dn −d1 . From this, we deduce
the lower bounds
degx (g1 f1 ) ≥ d1 · · · dn
,
deg(α) ≥ d1 · · · dn−1 h.
Hence, both bounds in (4.17) are optimal in this case.
Example 4.18. Let V ⊂ Ank be a k-variety of pure dimension r. For 1 ≤ j ≤ r + 1,
let dj ≥ 1 and consider the general n-variate polynomial of degree dj
X
Fj =
uj,a xa ∈ k[uj ][x1 , . . . , xn ],
|a|≤dj
where uj = {uj,a }|a|≤dj . Set u = {u1 , . . . , ur+1 }. It is not difficult to verify that the
Fj ’s have no common zeros in Vk(u) (k(u)). Hence, we can apply Theorem 4.11 to the
Fj ’s as a system of polynomials with coefficients depending on the groups of parameters u1 , . . . , ur+1 . We have that deg(Vk(u) ) = deg(V ) and, by Proposition 2.13(2),
hul (Vk(u) ) = 0. Besides, degx (Fj ) = dj and degul (Fj ) is equal to 1 if l = j and to 0
otherwise. We deduce that there exist α ∈ k[u] \ {0} and gi ∈ k[u][x1 , . . . , xn ] such
that α = g1 F1 + · · · + gr+1 Fr+1 on Vk(u) with
(4.19) degx (gi Fi ) ≤
r+1
Y
j=1
dj deg(V ) ,
degul (α), degul (gi Fi ) ≤
Y
j6=l
dj deg(V ).
Using Lemma 1.34 and Corollary 1.33, we can verify that the elimination ideal
I(V ) k[u][x1 , . . . , xn ] + (F1 , . . . , Fr+1 ) ∩ k[u]
is generated by the resultant Resd1 ,...,dr+1 (V ) of the closure of V in Pnk . This is a
multihomogeneous polynomial of partial degrees
Y
degul (Resd1 ,...,dr+1 (V )) =
dj deg(V ).
j6=l
Hence, α is a multiple of this resultant and, comparing degrees, we see that α and
Resd1 ,...,dr+1 (V ) coincide up to a factor in k× . This implies that the bound for the ul degrees in (4.19) is optimal. Observe that the bound for the x-degree is not optimal,
at least when V = Ank . In this case, it can be shown that there exist gi ’s satisfying
Pn+1
the same bound for the ul -degree and such that degx (gi fi ) ≤
j=1 dj − n.
The following result is a partial extension of Theorem 4.11 to an arbitrary number of
polynomials. For simplicity, we only state it for polynomials depending on one group
of parameters t = {t1 , . . . , tp }. In this setting, we loose track of the contribution of
the t-degrees of most of the individual input polynomials. However, it is possible to
differentiate the contribution of one of them, which will be important in the proof of
the strong parametric Nullstellensatz.
66
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Corollary 4.20. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and f1 , . . . , fs ∈
k[t, x] \ k[t] without common zeros in V (k(t)) such that V (fs ) intersects V properly.
Set dj = degx (fj ) for 1 ≤ j ≤ s. Assume that d1 ≥ · · · ≥ ds−1 and that ds is arbitrary.
Set also h = max1≤j≤s−1 degt (fj ) and hs = degt (fs ). Then there exist α ∈ k[t] \ {0}
and g1 , . . . , gs ∈ k[t, x] such that α = g1 f1 + · · · + gs fs on V with
min{s−1,r}
Y
• degx (gi fi ) ≤ ds
dj deg(V ),
j=1
min{s−1,r}
h
min{s−1,r}
Y
X
h
s
• deg(α), degt (gi fi ) ≤ ds
dj h(V ) + deg(V )
.
+
ds
dℓ
j=1
ℓ=1
Proof. If s ≤ r + 1, the result follows from the case m = 1 in Theorem 4.11. In the
case s ≥ r + 2, we can reduce to r + 1 polynomials by taking linear combinations as
follows: let v = {vj,i }1≤j≤r,1≤i≤s−r−1 be a group of variables and set K = k(v). Write
f j = fj + vj,1 fr+1 + · · · + vj,s−r−1 fs−1
for 1 ≤ j ≤ r,
f r+1 = fs .
The hypothesis that V (fs ) and V intersect properly implies that the polynomials above
do not have common zeros in VK(t) (K(t)). We have that degx (f j ) = dj and degt (f j ) =
h for 1 ≤ j ≤ r while degx (f r+1 ) = ds and degt (f r+1 ) = hs . By Theorem 4.11, there
exist α ∈ k[v, t] \ {0} and g i ∈ k[v, t][x] such that α = g1 f 1 + · · · + g r+1 f r+1 on VK(t)
with
r
Y
degx (g i f i ) ≤ ds
dj deg(V ),
j=1
r
r
h
Y
X
h
s
dj h(V ) + deg(V )
.
+
degt (α), degt (g i f i ) ≤ ds
ds
dℓ
j=1
ℓ=1
Unfolding the linear combinations in the above identity and taking a non-zero coefficient with respect to the variables v, we extract a Bézout identity α = g1 f1 + · · ·+ gs fs
on V satisfying the same degree bounds.
Remark 4.21. The previous results by Smietanski are for polynomials depending
on at most two parameters without common zeros in the affine space. For instance,
let f1 , . . . , fs be polynomials in k[t1 , t2 ][x1 , . . . , xn ] \ k[t1 , t2 ] without common zeros
in An (k(t1 , t2 )). Set dj = deg(fj ), h = maxj degt1 ,t2 (fj ), and suppose that d2 ≥
· · · ≥ ds ≥ d1 . Set also ν = min{n + 1, s}. In [Smi93], it is shown that there exist
α ∈ k[t1 , t2 ] \ {0} and gi ∈ k[t1 , t2 ][x1 , . . . , xn ] such that α = g1 f1 + · · · + gs fs with
ν
Y
dj ,
• degx (gi fi ) ≤ 3
j=1
• degt1 ,t2 (α), degt1 ,t2 (gi fi ) ≤
ν
Y
ν−1
(dj + h) + 3
j=1
ν
Y
j=1
ν
X
1
dj
h.
dℓ
ℓ=1
We deduce from Corollary 4.20 the following parametric version of the strong effective
Nullstellensatz.
Theorem 4.22. Let V ⊂ Ank(t) be a k(t)-variety of pure dimension r and g, f1 , . . . , fs ∈
k[t, x] such that g vanishes on the set of common zeros of f1 , . . . , fs in V (k(t))
and degx (f1 ) ≥ · · · ≥ degx (fs ) ≥ 1. Set dj = degx (fj ) for 1 ≤ j ≤ s, h =
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
67
max1≤j≤s degt (fj ), d0 = max{1, degx (g)} and h0 = degt (g). Then there exist µ ∈ N,
α ∈ k[t] \ {0} and g1 , . . . , gs ∈ k[t, x] such that
α gµ = g1 f1 + · · · + gs fs
on V
with
min{s,r+1}
Y
•µ≤2
dj deg(V ),
j=1
min{s,r+1}
Y
• degx (gi fi ) ≤ 4
dj deg(V ),
j=0
min{s,r+1}
3h
min{s,r+1}
Y
X
h
0
• deg(α), degt (gi fi ) ≤ 2
dj h(V ) + deg(V )
.
+
2d0
dℓ
j=0
ℓ=1
Proof. Set W = V × A1k(t) , let y be an additional variable and consider
(4.23)
1 − y d0 g, f1 , . . . , fs ∈ k[t, x, y].
These polynomials have no common zeros on W (k(t)) and V (1 − y d0 g) intersects W
properly. We have that dim(W ) = r + 1, deg(W ) = deg(V ) and h(W ) = h(V ),
thanks to Lemma 3.2. Besides, degx,y (fj ) = dj and degt (fj ) ≤ h for 1 ≤ j ≤ s,
while degx,y (1 − y d0 g) = 2d0 and degt (1 − y d0 g) = h0 . By Corollary 4.20, there exists
α ∈ k[t] \ {0} and g i ∈ k[t][x, y] such that
(4.24)
α = g 0 (1 − y d0 g) + g 1 f1 + · · · + g s fs
on W
with
min{s,r+1}
Y
dj deg(V ),
degx,y (g i fi ) ≤ 2
j=0
min{s,r+1}
h
min{s,r+1}
Y
X
h
0
dj h(V ) + deg(V )
.
+
deg(α), degt (g i fi ) ≤ 2
2d0
dℓ
j=0
ℓ=1
The input system (4.23) lies in the subring k[t, x, y d0 ] of k[t, x, y] and we can suppose
without loss of generality that the Bézout identity (4.24) lies in this subring. Let
gi (t, x, y d0 ). Then
gi ∈ k[t, x, y] such that gi (t, x, y) = b
b
(4.25)
α=b
g0 1 − y g + gb1 f1 + · · · + gbs fs .
Specializing y at 1/g(t, x) in the above identity and multiplying by a suitable denominator, we obtain an identity of the form α gµ = g1 f1 +· · ·+gs fs with µ = maxl degy (b
gl )
maxl degy (b
gl )
and gi = g
gbi (x, 1/g). Therefore,
(4.26)
µ = maxl degy (b
gl ) ≤ maxl
n deg
x,y (g l )
d0
o
min{s,r+1}
Y
dj deg(V ).
≤2
j=1
68
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
gi fi ) ≤ degt (g i fi ). Therefore
Besides, degx,y (b
gi fi ) ≤ degx,y (g i fi ) and degt (b
min{s,r+1}
Y
degx (gi fi ) ≤ degx (b
gi fi ) + degx (g)maxl degy (b
gl ) ≤ 4
dj deg(V ),
j=0
degt (gi fi ) ≤ degt (b
gi fi ) + degt (g)maxl degy (b
gl )
min{s,r+1}
3h
min{s,r+1}
Y
X
h
0
dj h(V ) + deg(V )
≤2
,
+
2d0
dℓ
j=0
as stated.
ℓ=1
Remark 4.27. In the previous argument, the use of a differentiated version of the
effective weak Nullstellensatz was crucial. Otherwise, the obtained bounds for the
Noether exponent µ would have depended on the degree of g, as for instance in [Bro87],
and the bounds for the height of the gi ’s would have been considerably worse.
4.3. Nullstellensätze over Z. In this section we present different arithmetic Nullstellensätze over Z for polynomials depending on groups of parameters. As before, let
tl = {tl,1 , . . . , tl,pl } be a group of variables for 1 ≤ l ≤ m and set t = {t1 , . . . , tm }.
Write also x = {x1 , . . . , xn }.
Theorem 4.28. Let V ⊂ AnQ be a Q-variety of pure dimension r and f1 , . . . , fs ∈
Z[t, x] \ Z[t] a family of s ≤ r + 1 polynomials without common zeros in VQ(t) (Q(t)).
Set dj = degx (fj ), δl,j = degtl (fj ) and hj = h(fj ) for 1 ≤ j ≤ s and 1 ≤ l ≤ m. Then
there exist α ∈ Z[t] \ {0} and g1 , . . . , gs ∈ Z[t, x] such that
α = g1 f1 + · · · + gs fs
(4.29)
on VQ(t)
with
s
Y
• degx (gi fi ) ≤
dj deg(V ),
j=1
s
X
δl,ℓ
dj deg(V )
for 1 ≤ l ≤ m,
dℓ
j=1
ℓ=1
s
Y
b
dj h(V ) + deg(V ) (3r + 7) log(n + 3)
• h(α), h(gi ) + h(fi ) ≤
• degtl (α), degtl (gi fi ) ≤
+
s
Y
j=1
s
X
ℓ=1
m
X
1
hℓ + log (#supp(fℓ )) + 2
δl,ℓ log(pl + 1)
.
dℓ
l=1
Proof. We apply again the construction explained in §4.1 to the ring A := Z[t1 , . . . , tm ],
the variety V and the polynomials fj . As a result, we obtain α ∈ A× and polynomials gi satisfying the Bézout identity (4.29). The bounds for their x-degree and
tl -degree follow from Theorem 4.11, since deg(VQ(t) ) = deg(V ) and htl (VQ(t) ) = 0 by
Proposition 2.13(2). So, we only have to bound the height of these polynomials.
We will use the notations introduced in §4.1. Let ui denote the group of coefficients of
the general linear forms ℓi = ui,1 x1 + · · · + ui,n xn in (4.4) and set u = {u1 , . . . , ur+1 }.
Consider the minimal polynomial E ∈ Z[t, u][y, z] of z with respect to the map π ◦ ϕ,
primitive with respect to Z[t, u]. We expand it in two different ways as follows:
E = α0,0 z δ +
δ
X
X
j=1 a∈Nr+1
αa,j y a z δ−j =
X
a∈Nr+1
αa y a
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
69
Pδ
δ−k ∈ Z[t, u, z]. Set
with αa,j ∈ Z[t, u] such that α0,0 6= 0 and αa =
k=0 αa,k z
supp(E) = {a : αa 6= 0} and recall that, for 1 ≤ i ≤ s,
r+1
δ
X
X X
Y ak a −b
bi−1 bi −1
b1
k
k
gi = −
e
αa,j
f1 · · · fi−1 fi
ℓ
∈ Z[t, u][x],
bk k
j=1 a∈supp(E) b
k=1
the last sum being indexed by all b ∈ Nr+1 such that |b| = j, bk ≤ ak for 1 ≤ k ≤ i,
bk = 0 for i + 1 ≤ k ≤ r + 1 and bi ≥ 1. The constant α and polynomials gi are
obtained as the coefficient of some monomial in u in α0,0 and e
gi , respectively. Hence,
h(α) ≤ h(α0,0 ) ,
(4.30)
h(gi ) + h(fi ) ≤ h(e
gi ) + h(fi ).
We have
X r+1
Y ak
j,a,b k=1
bk
≤
X r+1
Y ak
X
a∈supp(E) b≤a k=1
bk
=
X
a∈supp(E)
2|a| ≤ #supp(E) 2deg y (E) .
Hence, by Lemma 2.37(1),
r+1
Y a −b
bi−1 bi −1
b1
k
k
(4.31) h(e
gi ) + h(fi ) ≤ maxa,j h αa,j
f1 · · · fi−1 fi
ℓk
k=1
+ h(fi ) + log #supp(E) 2deg y (E) .
Set d = (d1 , . . . , ds , 1, . . . , 1), δ l = (δl,1 , . . . , δl,s , 0, . . . , 0) for 1 ≤ l ≤ m and h =
(h1 , . . . , hs , 0, . . . , 0). Using Lemma 2.37(2),
r+1
Y a −b b
bi−1 bi −1
1
k
k
(4.32) h αa,j
ℓk
f1 · · · fi−1 fi
≤ h(αa,j ) + |a| log(n)
k=1
+ hh − ei , ai + hd, ai log(n + 1) +
since log ||ℓk ||1 = log(n), log ||fj ||1 ≤ h(fj ) + dj log(n + 1) +
b ≤ a.
m
X
hδ l , ai log(pl + 1),
l=1
Pm
l=1 δl,j
log(pl + 1) and
By Lemma 4.8, E ∈ Z[t, u, z][y] is a primitive and squarefree polynomial defining
the closure of the image of the map ψ in (4.10). Hence, we can apply Theorem 3.15
to bound its partial degrees and height. We have degx (qj ) = dj , degtl (qj ) = δl,j and
h(qj ) = hj for 1 ≤ j ≤ s while degx (qj ) = 1, degtl (qj ) = h(qj ) = 0 for s+1 ≤ j ≤ r+1.
The partial
degree of qj in the group of (r + 1)n + 1 variables u ∪ {z} is equal to 1. Set
Q
D = sj=1 dj . Then, a direct application of Theorem 3.15 gives, for all a ∈ supp(E),
(4.33) hd, ai ≤ D deg(V ) ,
degtl (αa )+hδ l , ai ≤ D deg(V )
and
s
X
δl,ℓ
ℓ=1
dℓ
for 1 ≤ l ≤ m
s
X
1
b ) + deg(V ) log(r + 2) +
hℓ + log (#supp(fℓ ) + n + 2)
h(αa ) + hh, ai ≤ D h(V
dℓ
ℓ=1
+
m
X
l=1
δl,ℓ log(pl + 1) + (r + 1 − s) log(n + 2) + (r + 1) log((r + 1)n + 2) .
70
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Using the inequalities log (#supp(fℓ ) + n + 2) ≤ log (#supp(fℓ )) + log(n + 3) and
log((r + 1)n + 2) ≤ 2 log(n + 1) together with
log(r + 2) + s log(n + 3) + (r + 1 − s) log(n + 2) + 2(r + 1) log(n + 1) ≤ (3r + 4) log(n + 3),
we deduce
b
(4.34) h(αa ) + hh, ai ≤ D h(V ) + deg(V ) (3r + 4) log(n + 3)
m
s
X
X
1
δl,ℓ log(pl + 1)
.
hℓ + log (#supp(fℓ )) +
+
dℓ
l=1
ℓ=1
Moreover, we get from (4.33) that
log #supp(E) 2deg y (E) + |a| log(n) + hd, ai log(n + 1) ≤ 3D deg(V ) log(n + 2).
The statement follows from (4.30), (4.31), (4.32), (4.34) and this inequality.
Theorem 0.1 in the introduction follows from the case m = 0 in the previous result,
noticing that for a polynomial f ∈ Z[x] of degree d it holds #(supp(f )) ≤ (n + 1)d .
Example 4.35. Let d1 , . . . , dn , H ≥ 1 and set
f1 = xd11 , f2 = x1 xdn2 −1 − xd22 ,
d
n−1
. . . , fn−1 = xn−2 xdnn−1 −1 − xn−1
, fn = xn−1 xndn −1 − H ∈ Z[x1 , . . . , xn ].
n
It is a system of polynomials without common zeros in Q . Theorem 4.28 implies that
there is a Bézout identity α = g1 f1 + · · · + gn fn with deg(gi fi ) ≤ d1 · · · dn and
(4.36)
log(α), h(gi ) + h(fi ) ≤ d1 · · · dn−1 log(H) + (4n + 7) log(n + 2)d1 · · · dn .
On the other hand, let u be an additional variable and consider the specialization of
any such identity at xi = γi with
γ1 = H d2 ···dn−1 ud2 ···dn −1 , . . . , γn−1 = H udn −1 , γn = 1/u.
We obtain α = g1 (γ1 , . . . , γn−1 , 1/u) H d1 ···dn−1 ud1 ···dn −d1 . From this, we deduce the
lower bounds
degxn (g1 f1 ) ≥ d1 · · · dn
,
log(α) ≥ d1 · · · dn−1 log(H).
Hence, the height bound in (4.36) is optimal up to a term of size O(n log(n) d1 · · · dn ).
We next analyze Example 4.18 from the point of view of heights.
Example 4.37. Let V ⊂ AnQ be a Q-variety of pure dimension r. For 1 ≤ j ≤ r + 1
and dj ≥ 1, consider again the general n-variate polynomial of degree dj
X
Fj =
uj,a xa ∈ Z[uj ][x1 , . . . , xn ].
|a|≤dj
It follows from Example 4.18 that there exist λ ∈ Z \ {0} and gi ∈ Z[u][x1 , . . . , xn ]
such that
λ Resd1 ,...,dr+1 (V ) = g1 F1 + · · · + gr+1 Fr+1 on V
satisfying the degree bounds in (4.19). For the height, we have that degx (Fj ) = dj ,
h(Fj ) = 0 and log(#supp(Fj )) ≤ dj log(n + 1). Furthermore, δl,j = degul (Fj ) equals 1
if l = j and 0 otherwise. Theorem 4.28 then implies
r+1
Y
h(Resd1 ,...,dr+1 (V )), h(gi ) ≤
dℓ b
h(V ) + (6r + 10) log(n + 3) deg(V ) .
ℓ=1
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
71
The following result is a partial extension of Theorem 4.28 to an arbitrary number of
polynomials. For simplicity, we state it for polynomials not depending on parameters.
Corollary 4.38. Let V ⊂ AnQ be a Q-variety of pure dimension r and f1 , . . . , fs ∈
Z[x] \ Z without common zeros in V (Q) such that V (fs ) intersects V properly. Set
dj = deg(fj ) for 1 ≤ j ≤ s. Assume that d1 ≥ · · · ≥ ds−1 and that ds is arbitrary.
Set also h = max1≤j≤s−1h(fj ) and hs = h(fs ). Then there exist α ∈ Z \ {0} and
g1 , . . . , gs ∈ Z[x] such that α = g1 f1 + · · · + gs fs on V with
min{s−1,r}
Y
• deg (gi fi ) ≤ ds
j=1
• h(α), h(gi ) + h(fi ) ≤ ds
dj deg(V ),
min{s−1,r}
Y
j=1
min{s−1,r}
X
h
hs
b
+
dj h(V ) + deg(V )
ds
dℓ
ℓ=1
+ (6r + 9) log(n + 3) + 3r log(max{1, s − r}) .
Proof. The proof follows closely the lines of that of Corollary 4.20. If s ≤ r + 1, the
result follows from the case m = 0 in Theorem 4.11. Hence, we only have to consider
the case s ≥ r + 2. Set
f j = fj + vj,1 fr+1 + · · · + vj,s−r−1fs−1
for 1 ≤ j ≤ r,
f r+1 = fs
for a group v = {vj,i }j,i of p := r(s − r − 1) variables. It is a system of polynomials
without common zeros in VQ(v) (Q(v)). We have that degx (f j ) = dj , degv (f j ) = 1
and h(f j ) ≤ h for 1 ≤ j ≤ r while degx (f r+1 ) = ds , degv (f r+1 ) = 0 and h(f s ) = hs .
By Theorem 4.28, there are α ∈ Z[v] \ {0}Qand g i ∈ Z[v, x] such that α = g 1 f 1 + · · · +
gr+1 f r+1 on VQ(v) with degx (g i f i ) ≤ (ds rj=1 dj ) deg(V ) and
r
1
Y
hs + log(#supp(fs ))
h(α), h(g i ) + h(f i ) ≤ ds
dj b
h(V ) + deg(V )
ds
j=1
r
X
1
+
h + log(#supp(f ℓ )) + 2 log(p + 1) + (3r + 7) log(n + 3) .
dℓ
ℓ=1
We have that log(p + 1) ≤ log(n + 2) + log(s − r) and #supp(f ℓ ) ≤ (s − r)(n + 1)dℓ
for 1 ≤ ℓ ≤ r while #supp(f r+1 ) ≤ (n + 1)ds . Therefore,
r
r
h
Y
X
h
s
(4.39) h(α), h(g i ) + h(f i ) ≤ ds
dj b
h(V ) + deg(V )
+
+ 3r log(s − r)
ds
dℓ
j=1
ℓ=1
+ (6r + 8) log(n + 3) .
P
gs = g r+1 , then
Set gei = gi for 1 ≤ i ≤ r, gei = rk=1 vk,i−r g k for r + 1 ≤ i ≤ s − 1 and e
(4.40)
ge1 f1 + · · · + ges fs = g 1 f 1 + · · · + g r+1 f r+1 = α
with h(e
gi )+h(fi ) ≤ maxk {h(g k )+h(f k )}+log(r) for all i. By taking the coefficients of
a suitable monomial in v, we extract from (4.40) a Bézout identity α = g1 f1 +· · ·+gs fs
on V . By (4.39), these polynomials satisfy the stated degree and height bounds.
We finally prove the arithmetic strong Nullstellensatz presented in the introduction.
72
CARLOS D’ANDREA, TERESA KRICK, AND MARTÍN SOMBRA
Proof of Theorem 0.2. We use the same notations of the proof of Theorem 4.22 for k =
Q and p = 0. The proof of this corollary already gives the stated bounds for the degree
of the gi ’s and the exponent µ. Hence, it only remains to bound the height in the
identity αgµ = g1 f1 + · · · + gs fs on V .
Corollary 4.38 applied to W = V × A1 and 1 − y d0 g, f1 , . . . , fs and Lemma 3.16 imply
h min{s,r+1}
X
h
0
b
+
(4.41) h(α), h(g i ) + h(fi ) ≤ 2 d0 D h(V ) + deg(V )
2d0
dℓ
ℓ=1
+ (6r + 15) log(n + 4) + 3(r + 1) log(max{1, s − r}) ,
Qmin{s,r+1}
with D := j=1
dj . From this, we deduce the bound for the height of α. Let gbi
P
P
be as in (4.25). If we write gbi = a,j αa,j xa y j , then gi = a,j αa,j xa gmaxl {deg y (bgl )}−j .
Using Lemma 2.37 we deduce that
h(gi ) ≤ maxa,j {h(αa,j xa gmaxl {degy (bgl )}−j )} + log(#supp(b
gi ))
≤ h(b
gi ) + maxl {degy (b
gl )}(h0 + log(n + 1)d0 ) + log(#supp(b
gi )).
gi ) ≤ (n + 2)deg x,y (bgi ) . Moreover, degy (b
gi ) ≤
Observe that h(b
gi ) ≤ h(g i ) and #supp(b
2D deg(V ) and degx,y (b
gi ) ≤ 2d0 D deg(V ), as shown in the proof of Theorem 4.22.
Therefore, h(gi ) + h(fi ) is bounded above by
h(g i ) + h(fi ) + 2D deg(V )(h0 + log(n + 1)d0 ) + 2d0 D deg(V ) log(n + 2)).
The stated bound for h(gi ) + h(fi ) follows from this and (4.41).
References
[Asc04]
[BY91]
[BY99]
[BS10]
[BBK09]
[Bro01]
[Bro87]
[BPS11]
[CW37]
[DKS10]
[DP99]
[Ful84]
[GKZ94]
[Har77]
[Jel05]
M. Aschenbrenner, Ideal membership in polynomial rings over the integers, J. Amer. Math.
Soc. 17 (2004) 407-441.
C. A. Berenstein, A. Yger, Effective Bézout identities in Q [x1 , . . . , xn ], Acta Math. 166
(1991) 69-120.
C. A. Berenstein, A. Yger, Residue calculus and effective Nullstellensatz, Amer. J. Math.
121 (1999) 723-796.
Yu. F. Bilu, M. Strambi, Quantitative Riemann existence theorem over a number field,
Acta Arith. 145 (2010) 319-339.
E. Bombieri, J. Bourgain, S. V. Konyagin, Roots of polynomials in subgroups of F∗p and
applications to congruences, Int. Math. Res. Not. 5 (2009) 802-834.
W. D. Brownawell, The Hilbert Nullstellensatz, inequalities for polynomials, and algebraic
independence, Chapter 16 of Introduction to algebraic independence theory, Lecture Notes
in Math. 1752 (2001) 239-248.
W. D. Brownawell, Bounds for the degrees in the Nullstellensatz, Ann. of Math. 126 (1987)
577-591.
J. I. Burgos Gil, P. Philippon, M. Sombra, Arithmetic geometry of toric varieties. Metrics,
measures and heights, 141 pp..
W.-L. Chow, B. L. van der Waerden, Zur algebraische Geometrie. IX, Math. Ann. 113
(1937) 692-704.
X. Dahan, A. Kadri, E. Schost, Bit-size estimates for triangular sets in positive dimension,
to appear in J. Complexity, 37 pp..
S. David, P. Philippon, Minorations des hauteurs normalisées des sous-variétés des tores,
Ann. Scuola Norm. Sup. Pisa Cl. Sci. 28 (1999) 489-543.
W. Fulton, Intersection theory, Springer–Verlag, 1984.
I. M. Gelfand, M. M. Kapranov, A. V. Zelevinsky, Discriminants, resultants, and multidimensional determinants, Birkhäuser, 1994.
R. Hartshorne, Algebraic geometry, Springer, 1977.
Z. Jelonek, On the effective Nullstellensatz, Invent. Math. 162 (2005) 1-17.
HEIGHTS OF VARIETIES AND ARITHMETIC NULLSTELLENSÄTZE
73
[Jou83]
J.-P. Jouanolou, Théorèmes de Bertini et applications, Progress in Math. 42,
Birkhäuser, 1983.
[Koi96]
P. Koiran, Hilbert’s Nullstellensatz is in the polynomial hierarchy, J. Complexity 12 (1996)
273-286.
[KT08]
A. Kresch, Yu. Tschinkel, Effectivity of Brauer-Manin obstructions, Adv. Math. 218 (2008)
1-27.
[KP96]
T. Krick, L. M. Pardo, A computational method for Diophantine approximation, Progress
in Math. 143 (1996) 193-253.
[KPS01] T. Krick, L. M. Pardo, M. Sombra, Sharp estimates for the arithmetic Nullstellensatz,
Duke Math. J. 109 (2001) 521-598.
[Lan83]
S. Lang, Fundamentals of Diophantine geometry, Springer-Verlag, 1983.
[Lan93]
S. Lang, Algebra, Third Edition, Addison-Wesley, 1993.
[Mac1902] F. Macaulay, Some formulæ in elimination, Proc. London Math. Soc. 1 (1902) 3-27.
[Lel94]
P. Lelong, Mesure de Mahler et calcul des constantes universelles pour les polynômes de
n variables, Math. Ann. 299 (1994) 673-695.
[Mai00]
V. Maillot, Géométrie d’Arakelov des variétés toriques et fibrés en droites intégrables,
Mém. Soc. Math. France 80, 2000.
[PS93]
P. Pedersen, B. Sturmfels, Product formulas for resultants and Chow forms, Math. Z. 214
(1993) 377-396.
[Per27]
O. Perron, Algebra I (Die Grundlagen), de Gruyter, 1927.
[Phi86]
P. Philippon, Critères pour l’indépendance algébrique, Inst. Hautes Etudes Sci. Publ.
Math. 64 (1986) 5-52.
[Phi90]
P. Philippon, Dénominateurs dans le théorème des zeros de Hilbert, Acta Arith. 58 (1990)
1-25.
[Phi91]
P. Philippon, Sur des hauteurs alternatives. I, Math. Ann. 289 (1991) 355-283.
[Phi95]
P. Philippon, Sur des hauteurs alternatives. III, J. Math. Pures Appl. 74 (1995) 345-365.
[PS08]
P. Philippon, M. Sombra, Hauteur normalisée des variétés toriques projectives, J. Inst.
Math. Jussieu 7 (2008) 327-373.
[Rem01a] G. Rémond, Elimination multihomogène, Chapter 5 of Introduction to algebraic independence theory, Lecture Notes in Math. 1752 (2001) 53-81.
[Rem01b] G. Rémond, Géométrie diophantienne multiprojective, Chapter 7 of Introduction to algebraic independence theory, Lecture Notes in Math. 1752 (2001) 95-131.
[Rem10] G. Rémond, Nombre de points rationnels des courbes, Proc. London Math. Soc 101 (2010)
759-794.
[Smi93]
F. Smietanski, A parametrized Nullstellensatz, in Computational algebraic geometry
(Nice, 1992), Progr. Math. 109 (1993) 287-300.
[Smy81]
C. Smyth, A Kronecker-type theorem for complex polynomials in several variables, Canad.
Math. Bull. 24 (1981) 447-452.
[Som04]
M. Sombra, The height of the mixed sparse resultant, Amer. J. Math. 126 (2004) 12531260.
[Tei90]
B. Teissier, Résultats récents d’algèbre commutative effective, Astérisque 189-190 (1990),
107-131, Séminaire Bourbaki, Vol. 1989/90, exp. no. 718.
Departament d’Àlgebra i Geometria, Universitat de Barcelona. Gran Via 585, 08007
Barcelona, Spain
E-mail address: [email protected]
URL: http://atlas.mat.ub.es/personals/dandrea/
Departamento de Matemática, Facultad de Ciencias Exactas y Naturales, Universidad
de Buenos Aires and IMAS, CONICET. Ciudad Universitaria, 1428 Buenos Aires, Argentina
E-mail address: [email protected]
URL: http://mate.dm.uba.ar/~ krick/
ICREA and Departament d’Àlgebra i Geometria, Universitat de Barcelona. Gran Via 585,
08007 Barcelona, Spain
E-mail address: [email protected]
URL: http://atlas.mat.ub.es/personals/sombra/
| 0math.AC
|
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE
CLASSES
KAMRAN DIVAANI-AAZAR, HOSSEIN FARIDIAN AND MASSOUD TOUSI
arXiv:1701.07727v1 [math.AC] 26 Jan 2017
Abstract. Given a Serre class S of modules, we compare the containment of the
Koszul homology, Ext modules, Tor modules, local homology, and local cohomology
in S up to a given bound s ≥ 0. As some applications, we give a full characterization of noetherian local homology modules. Further, we establish a comprehensive
vanishing result which readily leads to the formerly known descriptions of the numerical invariants width and depth in terms of Koszul homology, local homology, and
local cohomology. Also, we immediately recover a few renowned vanishing criteria
scattered about the literature.
Contents
1. Introduction
1
2. Prerequisites
3
3. Containment in Serre Classes
6
4. Noetherianness and Artinianness
13
5. Vanishing Results
15
References
21
1. Introduction
Throughout this paper, R denotes a commutative noetherian ring with identity, and
M(R) designates the category of R-modules.
The theory of local cohomology was invented and introduced by A. Grothendieck
about six decades ago and has since been tremendously delved and developed by various
algebraic geometers and commutative algebraists.
There exists a dual theory, called the theory of local homology, which was initiated
by Matlis [Mat2] and [Mat1] in 1974, and its investigation was continued by Simon in
[Si1] and [Si2]. After the ground-breaking achievements of Greenlees and May [GM],
and Alonso Tarrı́o, Jeremı́as López and Lipman [AJL], a new era in the study of this
theory has undoubtedly begun; see e.g. [Sc], [CN1], [CN2], [Fr], [Ri], [MD], [HD] and
[DFT]. However, the theory of local homology has not been addressed as it merits and
2010 Mathematics Subject Classification. 13D07; 13D45; 13C05.
Key words and phrases. Koszul homology module; local cohomology module; local homology module;
Serre class.
The research of the first author is supported by grants from IPM (No. 95130212).
1
2
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
numerous counterparts of the results in local cohomology are yet to be explored and
unearthed in local homology.
One interesting problem in local cohomology is the investigation of the artinian property. This is very well accomplished in [Me2, Theorem 5.5]. The dual problem, i.e. the
noetherian property of local homology will be, among other things, probed carefully in
this paper.
An excursion among the results [AM1, Theorem 2.10], [AM2, Theorem 2.9], [BA,
Corollary 3.1], [BKN, Lemma 2.1], [DM, Proposition 1 and Corollary 1], [HK, Lemma
4.2], [Me2, Theorem 2.1], [WW, Propositions 7.1, 7.2 and 7.4], and [Yo, Lemma 1.2]
reveals an in-depth connection between local cohomology, Ext modules, Tor modules,
and Koszul homology in terms of their containment in a Serre class of modules. The
purpose of this paper is to bring the local homology into play and uncover the true
connection between all these homology and cohomology theories, and consequently, to
illuminate and enhance the aforementioned results.
To shed some light on this revelation, we observe that given elements a = a1 , . . . , an ∈
R and an R-module M , the Koszul homology Hi (a; M ) = 0 for every i < 0 or i > n,
and further
Hn (a; M ) ∼
= 0 :M (a) ∼
= Ext0R R/(a), M
and
H0 (a; M ) ∼
= M/(a)M ∼
= TorR
0 R/(a), M .
These isomorphisms suggest the existence of an intimate connection between the
Koszul homology on the one hand, and Ext and Tor modules on the other hand. The
connection seems to manifest as in the following casual diagram for any given integer
s ≥ 0:
Ext0R R/(a), M , . . . , ExtsR R/(a), M
{z
}
|
l
z
}|
{
Hn (a; M ), . . . , Hn−s (a; M ), . . . , Hs (a; M ), . . . , H0 (a; M )
{z
}
|
l
z
}|
{
R
R/(a),
M
,
.
.
.
,
Tor
M
R/(a),
TorR
0
s
As the above diagram depicts, the Koszul homology acts as a connecting device in
attaching the Ext modules to Tor modules in an elegant way. One such connection, one
may surmise, could be the containment in some Serre class of modules that we are about
to study in the following sections. More specifically, we prove the following as our main
results; see Definitions 2.1, 2.3, 2.5, 2.7, and Theorems 3.1 and 3.3.
Theorem 1.1. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following three conditions are equivalent
for any given s ≥ 0:
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
3
(i) Hi (a; M ) ∈ SP (R) for every 0 ≤ i ≤ s.
(ii) TorR
i (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every 0 ≤ i ≤ s.
(iii) TorR
i (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
If in addition, P satisfies the condition Da , then the above three conditions are equivalent
to the following condition:
b a for every 0 ≤ i ≤ s.
(iv) Hia (M ) ∈ SP R
And dually:
Theorem 1.2. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following three conditions are equivalent
for any given s ≥ 0:
(i) Hn−i (a; M ) ∈ SP (R) for every 0 ≤ i ≤ s.
(ii) ExtiR (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every 0 ≤ i ≤ s.
(iii) ExtiR (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
If in addition, P satisfies the condition Ca , then the above three conditions are equivalent
to the following condition:
(iv) Hai (M ) ∈ SP (R) for every 0 ≤ i ≤ s.
Having these results at our disposal, by specializing when s = n and the Serre property is noetherianness, artinianness, or being zero, we draw a handful of fruitful conclusions on noetherianness, artinianness, and vanishing of the five foregoing homology and
cohomology theories.
2. Prerequisites
First and foremost, we recall the notion of a Serre class of modules.
Definition 2.1. A class S of R-modules is said to be a Serre class if, given any short
exact sequence
0 → M ′ → M → M ′′ → 0
of R-modules, we have M ∈ S if and only if M ′ , M ′′ ∈ S.
The following example showcases the stereotypical instances of Serre classes.
Example 2.2. Let a be an ideal of R. Then the following classes of modules are Serre
classes:
(i) The zero class.
(ii) The class of all noetherian R-modules.
(iii) The class of all artinian R-modules.
4
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
(iv) The class of all minimax R-modules.
(v) The class of all minimax and a-cofinite R-modules; see [Me1, Corollary 4.4].
(vi) The class of all weakly Laskerian R-modules.
(v) The class of all Matlis reflexive R-modules.
We will be concerned with a change of rings when dealing with Serre classes. Therefore, we adopt the following notion of Serre property to exclude any possible ambiguity
in the statement of the results.
Definition 2.3. A property P concerning modules is said to be a Serre property if
SP (R) := M ∈ M(R) M satisfies the property P
is a Serre class for every ring R.
Given a Serre property, there naturally arise the corresponding notions of depth and
width.
Definition 2.4. Let a be an ideal of R, and M an R-module. Let P be a Serre property.
Then:
(i) Define the P-depth of M with respect to a to be
n
o
P-depthR (a, M ) := inf i ≥ 0 ExtiR (R/a, M ) ∈
/ SP (R) .
(ii) Define the P-width of M with respect to a to be
n
o
P-widthR (a, M ) := inf i ≥ 0 TorR
/ SP (R) .
i (R/a, M ) ∈
It is clear that upon letting P be the Serre property of being zero, we recover the
classical notions of depth and width.
We next remind the definition of local cohomology and local homology functors. Let
a be an ideal of R. We let
n
o
Γa (M ) := x ∈ M at x = 0 for some t ≥ 1
for an R-module M , and Γa (f ) := f |Γa (M) for an R-homomorphism f : M → N . This
provides us with the so-called a-torsion functor Γa (−) on the category of R-modules.
The ith local cohomology functor with respect to a is defined to be
Hai (−) := Ri Γa (−)
for every i ≥ 0. In addition, the cohomological dimension of a with respect to M is
n
o
cd(a, M ) := sup i ≥ 0 Hai (M ) 6= 0 .
Recall that an R-module M is said to be a-torsion if M = Γa (M ). It is well-known
ba -module structure in such a way
that any a-torsion R-module M enjoys a natural R
ba -submodules coincide. Besides, we have
that the lattices of its R-submodules and R
b a ⊗R M ∼
ba -modules.
R
= M both as R-modules and R
Likewise, we let
ca = lim(M/an M )
Λa (M ) := M
←
−
n
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
5
for an R-module M , and Λa (f ) := fb for an R-homomorphism f : M → N . This provides
us with the so-called a-adic completion functor Λa (−) on the category of R-modules. The
ith local homology functor with respect to a is defined to be
Hia (−) := Li Λa (−)
for every i ≥ 0. Besides, the homological dimension of a with respect to M is
hd(a, M ) := sup i ≥ 0 Hia (M ) 6= 0 .
The next conditions are also required to be imposed on the Serre classes when we
intend to bring the local homology and local cohomology into the picture.
Definition 2.5. Let P be a Serre property, and a an ideal of R. We say that P satisfies
the condition Da if the following statements hold:
(i) If R is a-adically complete and M/aM ∈ SP (R) for some R-module M , then
H0a (M ) ∈ SP (R).
ba .
(ii) For any a-torsion R-module M , we have M ∈ SP (R) if and only if M ∈ SP R
Example 2.6. Let a be an ideal of R. Then we have:
(i) The Serre property of being zero satisfies the condition Da . Use [Si1, Lemma
5.1 (iii)].
(ii) The Serre property of being noetherian satisfies the condition Da . See [DFT,
Lemma 2.5].
Melkersson [AM3, Definition 2.1] defines the condition Ca for Serre classes. Adopting
his definition, we have:
Definition 2.7. Let P be a Serre property, and a an ideal of R. We say that P satisfies
the condition Ca if the containment (0 :M a) ∈ SP (R) for some R-module M , implies
that Γa (M ) ∈ SP (R).
Example 2.8. Let a be an ideal of R. Then we have:
(i) The Serre property of being zero satisfies the condition Ca . Indeed, if M is an
R-module such that (0 :M a) = 0, then it can be easily seen that Γa (M ) = 0.
(ii) The Serre property of being artinian satisfies the condition Ca . This follows from
the Melkersson’s Criterion [Me2, Theorem 1.3].
(iii) A Serre class which is closed under taking direct limits is called a torsion theory.
There exists a whole bunch of torsion theories, e.g. given any R-module L, the
class
TL := M ∈ M(R) SuppR (M ) ⊆ SuppR (L)
is a torsion theory. One can easily see that any torsion theory satisfies the
condition Ca .
6
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
We finally describe the Koszul homology briefly. The Koszul complex K R (a) on an
element a ∈ R is the R-complex
a
K R (a) := Cone(R −
→ R),
and the Koszul complex K R (a) on a sequence of elements a = a1 , . . . , an ∈ R is the
R-complex
K R (a) := K R (a1 ) ⊗R · · · ⊗R K R (an ).
It is easy to see that K R (a) is a complex of finitely generated free R-modules concentrated in degrees n, . . . , 0. Given any R-module M , there is an isomorphism of
R-complexes
K R (a) ⊗R M ∼
= Σn HomR K R (a), M ,
which is sometimes referred to as the self-duality property of the Koszul complex. Accordingly, we feel free to define the Koszul homology of the sequence a with coefficients
in M , by setting
Hi (a; M ) := Hi K R (a) ⊗R M ∼
= Hi−n HomR K R (a), M
for every i ≥ 0.
3. Containment in Serre Classes
In this section, we study the containment of Koszul homology, Ext modules, Tor
modules, local homology, and local cohomology in Serre classes.
In the proof of Theorem 3.1 below, we use the straightforward observation that given
elements a = a1 , . . . , an ∈ R, a finitely generated R-module N , and an R-module M , if
M belongs to a Serre class S, then Hi (a; M ) ∈ S, ExtiR (N, M ) ∈ S, and TorR
i (N, M ) ∈ S
for every i ≥ 0. For a proof refer to [AT, Lemma 2.1].
Theorem 3.1. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following three conditions are equivalent
for any given s ≥ 0:
(i) Hi (a; M ) ∈ SP (R) for every 0 ≤ i ≤ s.
(ii) TorR
i (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every 0 ≤ i ≤ s.
(iii) TorR
i (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
If in addition, P satisfies the condition Da , then the above three conditions are equivalent
to the following condition:
b a for every 0 ≤ i ≤ s.
(iv) Hia (M ) ∈ SP R
Proof. (i) ⇒ (iii): Set N = R/a. Let F be a free resolution of R/a consisting of finitely
generated R-modules. Then the R-complex F ⊗R M is isomorphic to an R-complex of
the form
∂
∂
2
1
· · · → M s2 −→
M s1 −→
M s0 → 0.
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
7
We note that
∼
∼
TorR
0 (R/a, M ) = coker ∂1 = H0 (a; M ) ∈ SP (R),
(3.1.1)
by the assumption. If s = 0, then we are done. Suppose that s ≥ 1. The short exact
sequence
0 → im ∂1 → M s0 → coker ∂1 → 0,
induces the exact sequence
Hi+1 (a; coker ∂1 ) → Hi (a; im ∂1 ) → Hi (a; M s0 ).
The assumption together with the display (3.1.1) imply that the two lateral terms of
the above exact sequence belong to SP (R), so Hi (a; im ∂1 ) ∈ SP (R) for every 0 ≤ i ≤ s.
The short exact sequence
0 → ker ∂1 → M s1 → im ∂1 → 0,
yields the exact sequence
Hi+1 (a; im ∂1 ) → Hi (a; ker ∂1 ) → Hi (a; M s1 ).
Therefore, Hi (a; ker ∂1 ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. As s ≥ 1, we see that
H0 (a; ker ∂1 ) ∈ SP (R). The short exact sequence
0 → im ∂2 → ker ∂1 → TorR
1 (R/a, M ) → 0,
(3.1.2)
implies the exact sequence
(R/a,
M
)
→ 0.
H0 (a; ker ∂1 ) → H0 a; TorR
1
R
Therefore, H0 a; TorR
1 (R/a, M ) ∈ SP (R). But a Tor1 (R/a, M ) = 0, so
R
∼
TorR
1 (R/a, M ) = H0 a; Tor1 (R/a, M ) ∈ SP (R).
If s = 1, then we are done. Suppose that s ≥ 2. The short exact sequence (3.1.2) induces
the exact sequence
Hi+1 a; TorR
1 (R/a, M ) → Hi (a; im ∂2 ) → Hi (a; ker ∂1 ) .
It follows that Hi (a; im ∂2 ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. The short exact sequence
0 → ker ∂2 → M s2 → im ∂2 → 0,
yields the exact sequence
Hi+1 (a; im ∂2 ) → Hi (a; ker ∂2 ) → Hi (a; M s2 ).
Thus Hi (a; ker ∂2 ) ∈ SP (R) for every 0 ≤ i ≤ s−2. As s ≥ 2, we see that H0 (a; ker ∂2 ) ∈
SP (R). The short exact sequence
0 → im ∂3 → ker ∂2 → TorR
2 R/a, M → 0,
yields the exact sequence
H0 (a; ker ∂2 ) → H0 a; TorR
2 (R/a, M ) → 0.
8
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
As before, we conclude that
R
∼
TorR
2 (R/a, M ) = H0 a; Tor2 (R/a, M ) ∈ SP (R).
If s = 2, then we are done. Proceeding in this manner, we see that TorR
i (R/a, M ) ∈
SP (R) for every 0 ≤ i ≤ s.
(iii) ⇒ (ii): Let L be a finitely generated R-module with SuppR (L) ⊆ V (a). By
induction on s, we show that TorR
i (L, M ) ∈ SP (R) for every 0 ≤ i ≤ s. Let s = 0. Then
N ⊗R M ∈ S(R). Using Gruson’s Filtration Lemma [Va, Theorem 4.1], there is a finite
filtration
0 = L0 ⊆ L1 ⊆ · · · ⊆ Lt−1 ⊆ Lt = L,
such that Lj /Lj−1 is isomorphic to a quotient of a finite direct sum of copies of N for
every 1 ≤ j ≤ t. Given any 1 ≤ j ≤ t, the exact sequence
0 → Kj → N rj → Lj /Lj−1 → 0,
(3.1.3)
induces the exact sequence
N rj ⊗R M → (Lj /Lj−1 ) ⊗R M → 0.
It follows that (Lj /Lj−1 ) ⊗R M ∈ SP (R). Now the short exact sequence
0 → Lj−1 → Lj → Lj /Lj−1 → 0,
(3.1.4)
yields the exact sequence
Lj−1 ⊗R M → Lj ⊗R M → (Lj /Lj−1 ) ⊗R M.
A successive use of the above exact sequence, letting j = 1, . . . , t, implies that L ⊗R M ∈
SP (R).
Now let s ≥ 1, and suppose that the results holds true for s − 1. The induction
hypothesis implies that TorR
i (L, M ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. Hence, it suffices
to show that TorR
s (L, M ) ∈ SP (R). Given any 1 ≤ j ≤ t, the short exact sequence (3.1.3)
induces the exact sequence
R
R
rj
TorR
s (N , M ) → Tors Lj /Lj−1 , M → Tors−1 (Kj , M ).
(Kj , M ) ∈ SP (R), so from the above exact
The induction hypothesis shows that TorR
s−1
R
sequence we get that Tors Lj /Lj−1 , M ∈ SP (R). Now, the short exact sequence (3.1.4)
yields the exact sequence
R
R
TorR
s (Lj−1 , M ) → Tors (Lj , M ) → Tors Lj /Lj−1 , M .
A successive use of the above exact sequence, letting j = 1, . . . , t, implies that
TorR
s (L, M ) ∈ SP (R).
(ii) ⇒ (i): The hypothesis implies that TorR
i (R/a, M ) ∈ SP (R) for every 0 ≤ i ≤ s.
The R-complex K R (a) ⊗R M is isomorphic to an R-complex of the form
∂
∂
∂
2
n
1
M t1 −→
M tn−1 → · · · → M t2 −→
M → 0.
0 → M −→
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
9
We have
H0 (a; M ) ∼
= TorR
= coker ∂1 ∼
0 (R/a, M ) ∈ SP (R),
(3.1.5)
by the assumption. If s = 0, then we are done. Suppose that s ≥ 1. The short exact
sequence
0 → im ∂1 → M → coker ∂1 → 0,
induces the exact sequence
R
R
TorR
i+1 (R/a, coker ∂1 ) → Tori (R/a, im ∂1 ) → Tori (R/a, M ).
The assumption along with the display (3.1.5) imply that the two lateral terms of the
above exact sequence belong to SP (R), so TorR
i (R/a, im ∂1 ) ∈ SP (R) for every 0 ≤ i ≤ s.
The short exact sequence
0 → ker ∂1 → M t1 → im ∂1 → 0,
yields the exact sequence
R
R
t1
TorR
i+1 (R/a, im ∂1 ) → Tori (R/a, ker ∂1 ) → Tori (R/a, M ).
Therefore, TorR
i (R/a, ker ∂1 ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. As s ≥ 1, we see that
(R/a) ⊗R ker ∂1 ∈ SP (R). The short exact sequence
0 → im ∂2 → ker ∂1 → H1 (a; M ) → 0,
(3.1.6)
implies the exact sequence
(R/a) ⊗R ker ∂1 → (R/a) ⊗R H1 (a; M ) → 0.
Therefore, (R/a) ⊗R H1 (a; M ) ∈ SP (R). But aH1 (a; M ) = 0, so
H1 (a; M ) ∼
= (R/a) ⊗R H1 (a; M ) ∈ SP (R).
If s = 1, then we are done. Suppose that s ≥ 2. The short exact sequence (3.1.6) induces
the exact sequence
R
R
TorR
i+1 R/a, H1 (a; M ) → Tori (R/a, im ∂2 ) → Tori (R/a, ker ∂1 ).
It follows that TorR
i (R/a, im ∂2 ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. The short exact
sequence
0 → ker ∂2 → M t2 → im ∂2 → 0,
yields the exact sequence
R
R
t2
TorR
i+1 R/a, im ∂2 → Tori (R/a, ker ∂2 ) → Tori (R/a, M ).
Thus TorR
i (R/a, ker ∂2 ) ∈ SP (R) for every 0 ≤ i ≤ s − 2. As s ≥ 2, we see that
(R/a) ⊗R ker ∂2 ∈ SP (R). The short exact sequence
0 → im ∂3 → ker ∂2 → H2 (a; M ) → 0,
implies the exact sequence
(R/a) ⊗R ker ∂2 → (R/a) ⊗R H2 (a; M ) → 0.
10
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
As before, we conclude that
H2 (a; M ) ∼
= (R/a) ⊗R H2 (a; M ) ∈ SP (R).
If s = 2, then we are done. Proceeding in this manner, we infer that Hi (a; M ) ∈ SP (R)
for every 0 ≤ i ≤ s.
Now, assume that the Serre property P satisfies the condition Da . We first note that
ba
since TorR
i (R/a, M ) is an a-torsion R-module, it has an R -module structure such that
ba
R
R
a
a ba
a
∼
∼
b
b
b
R
/a
R
,
R
⊗
M
,
Tor
R
⊗
Tor
(R/a,
M
)
(R/a,
M
)
TorR
=
=
R
R
i
i
i
ba -modules. Moreover, by [Si2, Lemma 2.3], we
for every i ≥ 0 both as R-modules and R
have
ba
Hia (M ) ∼
= HiaR
b a ⊗R M ,
R
ba -modules. With this preparation we prove:
for every i ≥ 0 both as R-modules and R
(iv) ⇒ (iii): We let N = R/a, and show that TorR
i (R/a, M ) ∈ SP (R) for every
0 ≤ i ≤ s. First suppose that R is a-adically complete. By [DFT, Lemma 2.6], there is
a first quadrant spectral sequence
R
a
2
(3.1.7)
Ep,q
= TorR
p R/a, Hq (M ) ⇒ Torp+q (R/a, M ).
p
2
The hypothesis implies that Ep,q
∈ SP (R) for every p ≥ 0 and 0 ≤ q ≤ s. Let 0 ≤ i ≤ s.
There is a finite filtration
0 = U −1 ⊆ U 0 ⊆ · · · ⊆ U i = TorR
i (R/a, M ),
∞
∞
2
for every 0 ≤ p ≤ i. Since Ep,i−p
is a subquotient of Ep,i−p
such that U p /U p−1 ∼
= Ep,i−p
and 0 ≤ i − p ≤ s, we infer that
∞
∈ SP (R)
U p /U p−1 ∼
= Ep,i−p
for every 0 ≤ p ≤ i. A successive use of the short exact sequence
0 → U p−1 → U p → U p /U p−1 → 0,
by letting p = 0, . . . , i, implies that TorR
i (R/a, M ) ∈ SP (R).
Now, consider the general case. Since
ba b a
ba ,
⊗R M ∼
HiaR R
= Hia (M ) ∈ SP R
for every 0 ≤ i ≤ s, the special case implies that
ba
R
∼
ba ,
b a /aR
ba, R
ba ⊗R M ∈ SP R
R
TorR
i (R/a, M ) = Tori
for every 0 ≤ i ≤ s. However, TorR
i (R/a, M ) is a-torsion, so the condition Da implies
that TorR
i (R/a, M ) ∈ SP (R) for every 0 ≤ i ≤ s.
(ii) ⇒ (iv): It follows from the hypothesis that TorR
i (R/a, M ) ∈ SP (R) for every
0 ≤ i ≤ s. First suppose that R is a-adically complete. We argue by induction on s. Let
s = 0. Then M/aM ∈ SP (R), whence H0a (M ) ∈ SP (R) by the condition Da .
Now let s ≥ 1, and suppose that the result holds true for s − 1. The induction
hypothesis implies that Hia (M ) ∈ SP (R) for every 0 ≤ i ≤ s − 1. Hence, it suffices to
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
11
show that Hsa (M ) ∈ SP (R). Consider the spectral sequence (3.1.7). By the hypothesis,
2
Ep,q
∈ SP (R) for every p ∈ Z and 0 ≤ q ≤ s − 1. There is a finite filtration
0 = U −1 ⊆ U 0 ⊆ · · · ⊆ U s = TorR
s (R/a, M ),
∞
for every 0 ≤ p ≤ s. As TorR
such that U p /U p−1 ∼
= Ep,s−p
s (R/a, M ) ∈ SP (R), we
conclude that
E∞ ∼
= U 0 ∈ SP (R).
= U 0 /U −1 ∼
0,s
Let r ≥ 2, and consider the differentials
drr,s−r+1
dr0,s
r
r
r
−−→ E−r,s+r−1
= 0.
−−−−−→ E0,s
Er,s−r+1
r
2
Since s − r + 1 ≤ s − 1 and Er,s−r+1
is a subquotient of Er,s−r+1
, the hypothesis implies
r
that Er,s−r+1
∈ SP (R), and consequently im drr,s−r+1 ∈ SP (R) for every r ≥ 2. We thus
obtain
r+1 ∼
r
/ im drr,s−r+1 ,
E0,s
= ker dr0,s / im drr,s−r+1 = E0,s
and consequently a short exact sequence
(3.1.8)
r+1
r
0 → im drr,s−r+1 → E0,s
→ E0,s
→ 0.
r+1
∞ ∼
There is an integer r0 ≥ 2, such that E0,s
for every r ≥ r0 . It follows that
= E0,s
r0 +1
r0
E0,s ∈ SP (R). Now, the short exact sequence (3.1.8) implies that E0,s
∈ SP (R).
Using the short exact sequence (3.1.8) inductively, we conclude that
2
Hsa (M )/aHsa (M ) ∼
∈ SP (R).
= E0,s
Therefore, by [DFT, Lemma 2.4] and the condition Da , we get
Hsa (M ) ∼
= H0a Hsa (M ) ∈ SP (R).
Now, consider the general case. Since TorR
i (R/a, M ) is an a-torsion R-module such
that TorR
i (R/a, M ) ∈ SP (R) for every 0 ≤ i ≤ s, we deduce that
ba
R
a
a ba
∼
b
b
ba ,
TorR
R
/a
R
,
R
⊗
M
Tor
(R/a,
M
)
∈
S
=
R
P R
i
i
for every 0 ≤ i ≤ s. But the special case yields that
ba
ba ,
ba ⊗R M ∈ SP R
Hia (M ) ∼
= HiaR R
for every 0 ≤ i ≤ s.
The following special case may be of independent interest.
Corollary 3.2. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property satisfying the condition Da . Then the following
conditions are equivalent:
(i) M/aM ∈ SP (R).
ba .
(ii) H0a (M ) ∈ SP R
ba .
ca ∈ SP R
(iii) M
12
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
Proof. (i) ⇔ (ii): Follows from Theorem 3.1 upon letting s = 0.
ca is
(ii) ⇒ (iii): By [Si1, Lemma 5.1 (i)], the natural homomorphism H0a (M ) → M
surjective. Thus the result follows.
ba . However, by
ca /aM
ca ∈ SP R
ba , we see that M
ca ∈ SP R
(iii) ⇒ (i): Since M
ba .
ca /aM
ca ∼
[Si1, Theorem 1.1], we have M
= M/aM . It follows that M/aM ∈ SP R
But M/aM is a-torsion, so by the condition Da , we have M/aM ∈ SP (R).
Using similar arguments, we can prove the dual result to Theorem 3.1. It is worth
noting that the following result is proved in [AM3, Theorem 2.9]. However, the condition
Ca is assumed to be satisfied in all four statements, but here we only require that the
condition Ca is satisfied for the last statement. Moreover, the techniques used there are
quite different than those used here.
Theorem 3.3. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following three conditions are equivalent
for any given s ≥ 0:
(i) Hn−i (a; M ) ∈ SP (R) for every 0 ≤ i ≤ s.
(ii) ExtiR (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every 0 ≤ i ≤ s.
(iii) ExtiR (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
If in addition, P satisfies the condition Ca , then the above three conditions are equivalent
to the following condition:
(iv) Hai (M ) ∈ SP (R) for every 0 ≤ i ≤ s.
Proof. The proof is similar to that of Theorem 3.1. However, instead of the spectral
sequence (3.1.7), we use the third quadrant spectral sequence
−p−q
2
= Ext−p
R/a, Ha−q (M ) ⇒ ExtR
(R/a, M ).
Ep,q
R
p
If we let the integer s exhaust the whole nonzero range of Koszul homology, i.e. s = n,
then we can effectively combine Theorems 3.1 and 3.3 to obtain the following result which
in turn generalizes [BA, Corollary 3.1].
Corollary 3.4. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following conditions are equivalent:
(i) Hi (a; M ) ∈ SP (R) for every 0 ≤ i ≤ n.
(ii) TorR
i (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every i ≥ 0.
(iii) TorR
i (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ n.
(iv) ExtiR (N, M ) ∈ SP (R) for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every i ≥ 0.
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
13
(v) ExtiR (N, M ) ∈ SP (R) for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ n.
Proof. (i) ⇔ (ii) and (i) ⇔ (iv): Since Hi (a; M ) = 0 for every i > n, these equivalences
follow from Theorems 3.1 and 3.2, respectively.
(i) ⇔ (iii): Follows from Theorem 3.1.
(i) ⇔ (v): Follows from Theorem 3.3.
The following corollaries describe the numerical invariants P-depth and P-width in
terms of Koszul homology, local homology, and local cohomology.
Corollary 3.5. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Let P be a Serre property. Then the following assertions hold:
(i) P-depthR (a, M ) = inf i ≥ 0 Hn−i (a; M ) ∈
/ SP (R) .
/ SP (R) .
(ii) P-widthR (a, M ) = inf i ≥ 0 Hi (a; M ) ∈
(iii) We have P-depthR (a, M ) < ∞ if and only if P-widthR (a, M ) < ∞. Moreover
in this case, we have
sup i ≥ 0 Hi (a; M ) ∈
/ SP (R) + P-depthR (a, M ) = n.
Proof. (i) and (ii): Follows from Theorems 3.1 and 3.3.
(iii): The first assertion follows from Corollary 3.4. For the second assertion, note
that
/ SP (R) = n − sup i ≥ 0 Hi (a; M ) ∈
/ SP (R) .
inf i ≥ 0 Hn−i (a; M ) ∈
Corollary 3.6. Let a be an ideal of R, and M an R-module. Let P be a Serre property.
Then the following assertions hold:
(i) If P satisfies the condition Ca , then
n
o
P-depthR (a, M ) = inf i ≥ 0 Hai (M ) ∈
/ SP (R) .
(ii) If P satisfies the condition Da , then
a
ba
.
P-widthR (a, M ) = inf i ≥ 0 Hi (M ) ∈
/ SP R
(iii) If P satisfies both conditions Ca and Da , and P-depthR (a, M ) < ∞, then
P-depthR (a, M ) + P-widthR (a, M ) ≤ ara(a).
Proof. (i) and (ii): Follows from Theorems 3.1 and 3.3.
(iii): Clear by (i), (ii), and Corollary 3.5 (iii).
4. Noetherianness and Artinianness
In this section, we apply the results of Section 3 to obtain some characterizations
of noetherian local homology modules and artinian local cohomology modules. The
following result together with Corollary 4.4 generalize [WW, Propositions 7.1, 7.2 and
7.4] when applied to modules.
14
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
Proposition 4.1. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions are equivalent for any given s ≥ 0:
(i) Hi (a; M ) is a finitely generated R-module for every 0 ≤ i ≤ s.
(ii) TorR
i (N, M ) is a finitely generated R-module for every finitely generated Rmodule N with SuppR (N ) ⊆ V (a), and for every 0 ≤ i ≤ s.
(iii) TorR
i (N, M ) is a finitely generated R-module for some finitely generated Rmodule N with SuppR (N ) = V (a), and for every 0 ≤ i ≤ s.
ba -module for every 0 ≤ i ≤ s.
(iv) Hia (M ) is a finitely generated R
Proof. Obvious in view of Example 2.6 (ii) and Theorem 3.1.
The following corollary provides a characterization of noetherian local homology modules in its full generality.
Corollary 4.2. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Then the following assertions are equivalent:
ba -module for every i ≥ 0.
(i) Hia (M ) is a finitely generated R
(ii) Hi (a; M ) is a finitely generated R-module for every 0 ≤ i ≤ n.
(iii) Hi (a; M ) is a finitely generated R-module for every 0 ≤ i ≤ hd(a, M ).
(iv) TorR
i (N, M ) is a finitely generated R-module for every finitely generated Rmodule N with SuppR (N ) ⊆ V (a), and for every i ≥ 0.
(v) TorR
i (N, M ) is a finitely generated R-module for some finitely generated Rmodule N with SuppR (N ) = V (a), and for every 0 ≤ i ≤ hd(a, M ).
(vi) ExtiR (N, M ) is a finitely generated R-module for every finitely generated Rmodule N with SuppR (N ) ⊆ V (a), and for every i ≥ 0.
(vii) ExtiR (N, M ) is a finitely generated R-module for some finitely generated Rmodule N with SuppR (N ) = V (a), and for every 0 ≤ i ≤ n.
Proof. (ii) ⇔ (iv) ⇔ (vi) ⇔ (vii): Follows from Corollary 3.4.
(iii) ⇔ (v): Follows from Proposition 4.1 upon setting s = hd(a, M ).
(i) ⇔ (iii): Since Hia (M ) = 0 for every i > hd(a, M ), the result follows from Proposition 4.1.
(i) ⇔ (iv): Follows from Proposition 4.1.
One should note that a slightly weaker version of the following result has been proved
in [Me2, Theorem 5.5] by using a different method.
Proposition 4.3. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions are equivalent for any given s ≥ 0:
(i) Hn−i (a; M ) is an artinian R-module for every 0 ≤ i ≤ s.
(ii) ExtiR (N, M ) is an artinian R-module for every finitely generated R-module N
with SuppR (N ) ⊆ V (a), and for every 0 ≤ i ≤ s.
(iii) ExtiR (N, M ) is an artinian R-module for some finitely generated R-module N
with SuppR (N ) = V (a), and for every 0 ≤ i ≤ s.
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
15
(iv) Hai (M ) is an artinian R-module for every 0 ≤ i ≤ s.
Proof. Obvious in view of Example 2.8 (ii) and Theorem 3.3.
The following corollary provides a characterization of artinian local cohomology modules in its full generality.
Corollary 4.4. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Then the following assertions are equivalent:
(i) Hai (M ) is an artinian R-module for every i ≥ 0.
(ii) Hi (a; M ) is an artinian R-module for every 0 ≤ i ≤ n.
(iii) Hn−i (a; M ) is an artinian R-module for every 0 ≤ i ≤ cd(a, M ).
(iv) ExtiR (N, M ) is an artinian R-module for every finitely generated R-module N
with SuppR (N ) ⊆ V (a), and for every i ≥ 0.
(v) ExtiR (N, M ) is an artinian R-module for some finitely generated R-module N
with SuppR (N ) = V (a), and for every 0 ≤ i ≤ cd(a, M ).
(vi) TorR
i (N, M ) is an artinian R-module for every finitely generated R-module N
with SuppR (N ) ⊆ V (a), and for every i ≥ 0.
(vii) TorR
i (N, M ) is an artinian R-module for some finitely generated R-module N
with SuppR (N ) = V (a), and for every 0 ≤ i ≤ n.
Proof. (ii) ⇔ (iv) ⇔ (vi) ⇔ (vii): Follows from Corollary 3.4.
(iii) ⇔ (v): Follows from Proposition 4.3 upon setting s = cd(a, M ).
(i) ⇔ (iii): Since Hai (M ) = 0 for every i > cd(a, M ), the result follows from Proposition 4.3.
(i) ⇔ (iv): Follows from Proposition 4.3.
5. Vanishing Results
In this section, we treat the vanishing results. Since the property of being zero is a
Serre property that satisfies the condition Da , we obtain the following result.
Proposition 5.1. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions are equivalent for any given s ≥ 0:
(i) Hi (a; M ) = 0 for every 0 ≤ i ≤ s.
(ii) TorR
i (N, M ) = 0 for every R-module N with SuppR (N ) ⊆ V (a), and for every
0 ≤ i ≤ s.
(iii) TorR
i (N, M ) = 0 for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
(iv) Hia (M ) = 0 for every 0 ≤ i ≤ s.
Proof. Immediate from Theorem 3.1. For part (ii), note that every module is a direct
limit of its finitely generated submodules and Tor functor commutes with direct limits.
16
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
The following result is proved in [St, Theorem 4.4] using a different method, but here
it is an immediate by-product of Proposition 5.1.
Corollary 5.2. Let M be an R-module, and N a finitely generated R-module. Then the
following conditions are equivalent for any given s ≥ 0:
(i) TorR
i (N, M ) = 0 for every 0 ≤ i ≤ s.
R
(ii) Tori R/ annR (N ), M = 0 for every 0 ≤ i ≤ s.
Proof. Immediate from Proposition 5.1.
We observe that Corollary 5.3 below generalizes [Va, Corollary 4.3], which states that
if N is a faithful finitely generated R-module, then N ⊗R M = 0 if and only if M = 0.
Corollary 5.3. Let M be an R-module, and N a finitely generated R-module. Then the
following conditions are equivalent:
(i) M ⊗R N = 0.
(ii) M = annR (N )M .
In particular, we have
SuppR (M ⊗R N ) = SuppR M/ annR (N )M .
Proof. For the equivalence of (i) and (ii), let s = 0 in Corollary 5.2. For the second part,
note that given any p ∈ Spec(R), we have annR (N )p = annRp (Np ), and so (M ⊗R N )p =
0 if and only if M/ annR (N )M p = 0.
The support formula in Corollary 5.3 generalizes the well-known formula
SuppR (M ⊗R N ) = SuppR (M ) ∩ SuppR (N ),
which holds whenever M and N are both assumed to be finitely generated.
Using Corollary 3.6, we intend to obtain two somewhat known descriptions of the
numerical invariant widthR (a, M ) in terms of the Koszul homology and local homology.
However, we need some generalizations of the five-term exact sequences. To the best of
our knowledge, the only place where one may find such generalizations is [Ro, Corollaries
10.32 and 10.34]. But, the statements there are not correct and no proof is presented.
Hence due to lack of a suitable reference, we deem it appropriate to include the correct
statements with proofs for the convenience of the reader.
2
Lemma 5.4. Let Ep,q
⇒ Hp+q be a spectral sequence. Then the following assertions
p
hold:
2
2
(i) If Ep,q
⇒ Hp+q is first quadrant and there is an integer n ≥ 1 such that Ep,q
=0
p
for every q ≤ n − 2, then there is a five-term exact sequence
2
2
2
Hn+1 → E2,n−1
→ E0,n
→ Hn → E1,n−1
→ 0.
2
2
(ii) If Ep,q
⇒ Hp+q is third quadrant and there is an integer n ≥ 1 such that Ep,q
=0
p
for every q ≥ 2 − n, then there is a five-term exact sequence
2
2
2
0 → E−1,1−n
→ H−n → E0,−n
→ E−2,1−n
→ H−n−1 .
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
17
Proof. (i): Consider the following homomorphisms
d20,n
d22,n−1
d24,n−2
2
2
2
2
−−−→ E−2,n+1
= 0.
−−−−→ E0,n
−−−−→ E2,n−1
0 = E4,n−2
We thus have
3
∼
E2,n−1
= ker d22,n−1 / im d24,n−2 ∼
= ker d22,n−1 ,
and
3 ∼
E0,n
= ker d20,n / im d22,n−1 = coker d22,n−1 .
Let r ≥ 3. Consider the following homomorphisms
drr+2,n−r
dr2,n−1
r
r
r
0 = Er+2,n−r
−−−−−−→ E2,n−1
−−−−→ E2−r,n+r−2
= 0.
We thus have
r+1 ∼
r
E2,n−1
.
= ker dr2,n−1 / im drr+2,n−r ∼
= E2,n−1
Therefore,
∞
4
3
∼
∼
.
ker d22,n−1 ∼
= ··· ∼
= E2,n−1
= E2,n−1
= E2,n−1
Further, consider the following homomorphisms
dr0,n
drr,n−r+1
r
r
r
−−−→ E−r,n+r−1
= 0.
−−−−−−→ E0,n
0 = Er,n−r+1
We thus have
r+1 ∼
r
.
E0,n
= E0,n
= ker dr0,n / im drr,n−r+1 ∼
Therefore,
∞
4 ∼
3 ∼
.
coker d22,n−1 ∼
= ··· ∼
= E0,n
= E0,n
= E0,n
Hence, we get the following exact sequence
(5.4.1)
d22,n−1
∞
2
2
∞
0 → E2,n−1
→ E2,n−1
−−−−→ E0,n
→ E0,n
→ 0.
There is a finite filtration
0 = U −1 ⊆ U 0 ⊆ · · · ⊆ U n+1 = Hn+1 ,
∞
∼
such that Ep,n+1−p
= U p /U p−1 for every 0 ≤ p ≤ n + 1. If 3 ≤ p ≤ n + 1, then
∞
n + 1 − p ≤ n − 2, so Ep,n+1−p
= 0. It follows that
U 2 = U 3 = · · · = U n+1 .
∞
∼
Now since E2,n−1
= U 2 /U 1 , we get an exact sequence
∞
Hn+1 → E2,n−1
→ 0.
Splicing this exact sequence to the exact sequence (5.4.1), we get the following exact
sequence
(5.4.2)
d22,n−1
2
2
∞
Hn+1 → E2,n−1
−−−−→ E0,n
→ E0,n
→ 0.
On the other hand, there is a finite filtration
0 = V −1 ⊆ V 0 ⊆ · · · ⊆ V n = Hn ,
18
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
∞
∼
such that Ep,n−p
= V p /V p−1 for every 0 ≤ p ≤ n. If 2 ≤ p ≤ n, then n − p ≤ n − 2, so
∞
Ep,n−p
= 0. It follows that
V 1 = V 2 = · · · = V n.
∞
∞ ∼
∼
As E0,n
= V 1 /V 0 , we get the short exact sequence
= V 0 /V −1 = V 0 and E1,n−1
∞
∞
0 → E0,n
→ Hn → E1,n−1
→ 0.
Splicing this short exact sequence to the exact sequence (5.4.2), yields the exact sequence
d22,n−1
2
2
∞
Hn+1 → E2,n−1
−−−−→ E0,n
→ Hn → E1,n−1
→ 0.
(5.4.3)
Let r ≥ 2, and consider the following homomorphisms
drr+1,n−r
dr1,n−1
r
r
r
0 = Er+1,n−r
−−−−−−→ E1,n−1
−−−−→ E1−r,n+r−2
= 0.
We thus have
r+1 ∼
r
.
E1,n−1
= E1,n−1
= ker dr1,n−1 / im drr+1,n−r ∼
Therefore,
∞
3
2
∼
∼
.
E1,n−1
= ··· ∼
= E1,n−1
= E1,n−1
Thus from the exact sequence (5.4.3), we get the desired exact sequence
d22,n−1
2
2
2
Hn+1 → E2,n−1
−−−−→ E0,n
→ Hn → E1,n−1
→ 0.
(ii): Similar to (i).
For the next result, we need to recall the notion of coassociated prime ideals. Given
an R-module M , a prime ideal p ∈ Spec(R) is said to be a coassociated prime ideal
of M if p = annR (M/N ) for some submodule N of M such that M/N is an artinian
R-module. The set of coassociated prime ideals of M is denoted by CoassR (M ).
Corollary 5.5. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an Rmodule. Then the following assertions hold:
(i) widthR (a, M ) = inf i ≥ 0 Hi (a; M ) 6= 0 = inf i ≥ 0 Hia (M ) 6= 0 .
a
(ii) TorR
(R/a, M ) ∼
(M ).
= (R/a) ⊗R Hwidth
R (a,M)
widthR (a,M)
a
n
(M ) ∼
TorR
(iii) Λa Hwidth
= lim
widthR (a,M) (R/a , M ).
R (a,M)
←
−
n
a
(iv) CoassR TorR
widthR (a,M) (R/a, M ) = CoassR HwidthR (a,M) (M ) ∩ V (a).
Proof. (i): Follows immediately from Corollaries 3.5 and 3.6.
(ii): Consider the first quadrant spectral sequence
R
2
a
Ep,q
= TorR
p R/a, Hq (M ) ⇒ Torp+q (R/a, M ).
p
Let n = widthR (a, M ). Then by (i),
gives the exact sequence
2
Ep,q
= 0 for every q ≤ n − 1. Now, Lemma 5.4 (i)
R
a
TorR
(R/a,
M
)
→
Tor
R/a,
H
(M
)
→ (R/a) ⊗R Hna (M ) →
n+1
2
n−1
R
a
TorR
n (R/a, M ) → Tor1 R/a, Hn−1 (M ) → 0.
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
But
so
19
R
a
a
R/a,
H
(M
)
,
R/a,
H
(M
)
=
0
=
Tor
TorR
n−1
n−1
1
2
(R/a) ⊗R Hna (M ) ∼
= TorR
n (R/a, M ).
n
(iii): Using (ii) and the facts that Hia (M ) ∼
= Hia (M ) and widthR (an , M ) =
widthR (a, M ) for all n, i ≥ 0, we get:
n a
a
a
(M
)
(M
)/a
H
(M
)
=
lim
H
Λa Hwidth
widthR (a,M)
widthR (a,M)
R (a,M)
←−
n
n
∼
TorR
= lim
widthR (a,M) (R/a , M ).
←−
n
(iv): Using (ii) and [Ya, Theorem 1.21], we have
a
CoassR TorR
(M
)
(R/a)
⊗
H
(R/a,
M
)
=
Coass
R
R
widthR (a,M)
widthR (a,M)
a
= SuppR (R/a) ∩ CoassR Hwidth
(M
)
(a,M)
R
a
= V (a) ∩ CoassR HwidthR (a,M) (M ) .
Note that part (iii) of Corollary 5.5 is proved in [Si2, Proposition 2.5] by deploying
a different method. On the other hand, in parallel with Corollary 5.10 (iii) below, one
may wonder if intersecting
with V(a) in part (iv) of Corollary 5.5 is redundant. In other
a
words, CoassR HwidthR (a,M) (M ) may be contained in V (a). However, the following
example shows that this is not the case in general.
Example 5.6. Let R := Q[X, Y ](X,Y ) and m := (X, Y )R. Then widthR (m, R) = 0 and
bm ∼
H0m (R) ∼
= Q[[X, Y ]].
=R
For each n ∈ Z, let pn := (X − nY )R. Then it is easy to see that R/pn ∼
= Q[Y ](Y ) , and
so it is not a complete local ring. By [Z, Beispiel 2.4],
bm
CoassR H0m (R) = CoassR R
= {m} ∪ p ∈ Spec R R/p is not a complete local ring .
Hence CoassR H0m (R) is not a finite set, while CoassR H0m (R) ∩ V (m) = {m}. In
particular, CoassR H0m (R) * V (m).
Since the property of being zero is a Serre property that satisfies the condition Ca ,
we obtain the following result.
Proposition 5.7. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions are equivalent for any given s ≥ 0:
(i) Hn−i (a; M ) = 0 for every 0 ≤ i ≤ s.
(ii) ExtiR (N, M ) = 0 for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every 0 ≤ i ≤ s.
20
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
(iii) ExtiR (N, M ) = 0 for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ s.
(iv) Hai (M ) = 0 for every 0 ≤ i ≤ s.
Proof. Immediate from Theorem 3.3.
The following result is proved in [St, Theorem 3.2] using a different method, but here
it is an immediate by-product of Proposition 5.7.
Corollary 5.8. Let M be an R-module, and N a finitely generated R-module. Then the
following conditions are equivalent for any given s ≥ 0:
(i) ExtiR (N, M ) = 0 for every 0 ≤ i ≤ s.
(ii) ExtiR R/ annR (N ), M = 0 for every 0 ≤ i ≤ s.
Proof. Immediate from Proposition 5.7.
The following special case may be of independent interest.
Corollary 5.9. Let M be an R-module, and N a finitely generated R-module. Then the
following conditions are equivalent:
(i) HomR (N, M ) = 0.
(ii) 0 :M annR (N ) = 0.
Proof. Let s = 0 in Corollary 5.8.
It is easy to deduce from Corollary 5.9 that given a finitely generated R-module N ,
we have HomR (N, M ) 6= 0 if and only if there are elements x ∈ N and 0 6= y ∈ M with
annR (x) ⊆ annR (y), which is known as the Hom Vanishing Lemma in [C, Page 11].
We state the dual result to Corollary 5.5 for the sake of integrity and completeness.
Parts (ii) and (iii) of Corollary 5.10 below are stated in [Mar, Proposition 1.1], and it is
only mentioned that part (ii) can be deduced from a spectral sequence. In addition, a
proof of this result is offered in [Me1, Corollary 2.3] by using different techniques.
Corollary 5.10. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions hold:
n
o
(i) depthR (a, M ) = inf i ≥ 0 Hn−i (a; M ) 6= 0 = inf i ≥ 0 Hai (M ) 6= 0 .
depthR (a,M)
depth (a,M)
(M ) .
(ii) ExtR R
(R/a, M ) ∼
= HomR R/a, Ha
depthR (a,M)
depth (a,M)
(M ) .
(R/a, M ) = AssR Ha
(iii) AssR ExtR R
Proof. (i): Follows immediately from Corollaries 3.5 and 3.6.
(ii): Consider the third quadrant spectral sequence
−p−q
2
R/a, Ha−q (M ) ⇒ ExtR
(R/a, M ),
Ep,q
= Ext−p
R
p
and use Lemma 5.4 (ii).
(iii): Follows from (ii).
Finally, we present the following comprehensive vanishing result.
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
21
Corollary 5.11. Let a = (a1 , . . . , an ) be an ideal of R, a = a1 , . . . , an , and M an
R-module. Then the following assertions are equivalent:
(i) Hi (a; M ) = 0 for every 0 ≤ i ≤ n.
(ii) Hi (a; M ) = 0 for every 0 ≤ i ≤ hd(a, M ).
(iii) Hi (a; M ) = 0 for every n − cd(a, M ) ≤ i ≤ n.
(iv) TorR
i (N, M ) = 0 for every R-module N with SuppR (N ) ⊆ V (a), and for every
i ≥ 0.
(v) TorR
i (N, M ) = 0 for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ hd(a, M ).
(vi) ExtiR (N, M ) = 0 for every finitely generated R-module N with SuppR (N ) ⊆
V (a), and for every i ≥ 0.
(vii) ExtiR (N, M ) = 0 for some finitely generated R-module N with SuppR (N ) =
V (a), and for every 0 ≤ i ≤ cd(a, M ).
(viii) Hia (M ) = 0 for every i ≥ 0.
(ix) Hai (M ) = 0 for every i ≥ 0.
Proof. Follows from Propositions 5.1 and 5.7.
The following corollary is proved in [Si2, Corollary 1.7]. However, it is an immediate
consequence of the results thus far obtained.
Corollary 5.12. Let a be an ideal of R, and M an R-module. Then depthR (a, M ) < ∞
if and only if widthR (a, M ) < ∞. Moreover in this case, we have
depthR (a, M ) + widthR (a, M ) ≤ ara(a).
Proof. Clear by Corollary 3.6 (iii).
References
[AJL] L. Alonso Tarrı́o, A. Jeremı́as López and J. Lipman, Local homology and cohomology on schemes,
Ann. Sci. École Norm. Sup., (4), 30(1), (1997), 1-39.
[AM1] M. Aghapournahr and L. Melkersson, A natural map in local cohomology, Arkiv för Matematik,
48(2), (2010), 243-251.
[AM2] M. Aghapournahr and L. Melkersson, Cofiniteness and coassociated primes of local cohomology
modules, Math. Scand., 105(2), (2009), 161-170.
[AM3] M. Aghapournahr and L. Melkersson, Local cohomology and Serre subcategories, J. Algebra,
320(3), (2008), 1275-1287.
[AT] M. Asgharzadeh and M. Tousi, A unified approach to local cohomology modules using Serre classes,
Canad. Math. Bull., 53(4), (2010), 577-586.
[BA] K. Bahmanpour and M. Aghapournahr, A note on cofinite modules, Comm. Algebra, 44(9),
(2016), 3683-3691.
[BKN] K. Bahmanpour, I. Khalili and R. Naghipour, Cofiniteness of torsion functors of cofinite modules, Colloquium Mathematicum, 136(2), (2014), 221-230.
[C] L.W. Christesen, Gorenstein dimensions, Lecture Note in Mathematics, 1747, 2000.
[CN1] N.T. Cuong and T.T. Nam, A local homology theory for linearly compact modules, J. Algebra,
319(11), (2008), 4712-4737.
22
K. DIVAANI-AAZAR, H. FARIDIAN AND M. TOUSI
[CN2] N.T. Cuong and T.T. Nam, The I-adic completion and local homology for Artinian modules,
Math. Proc. Cambridge Philos. Soc., 131(1), (2001), 61-72.
[DFT] K. Divaani-Aazar, H. Faridian and M. Tousi, Local homology, finiteness of Tor modules and
cofiniteness, preprint.
[DM] D. Delfino and T. Marley, Cofinite modules and local cohomology, J. Pure Appl. Algebra, 121(1),
(1997), 45-52.
[Fr] A. Frankild, Vanishing of local homology, Math. Z., 244(3), (2003), 615-630.
[GM] J.P.C. Greenlees and J.P. May, Derived functors of I-adic completion and local homology, J.
Algebra, 149(2), (1992), 438-453.
[HD] M. Hatamkhani and K. Divaani-Aazar, On the vanishing of local homology modules, Glasg. Math.
J., 55(2), (2013), 457-464.
[HK] G. Huneke and J. Koh, Cofiniteness and vanishing of local cohomology modules, Math. Proc.
Camb. Phil. Soc., 110(3), (1991), 421-429.
[Mar] T. Marley, The associated primes of local cohomology modules over rings of small dimension,
Manuscripta Math., 104(4), (2001), 519-525.
[Mat1] E. Matlis, The higher properties of R-sequences, J. Algebra, 50(1), (1978), 77-112.
[Mat2] E. Matlis, The Koszul complex and duality, Comm. Algebra, 1, (1974), 87-144.
[Me1] L. Melkersson, Modules cofinite with respect to an ideal, J. Algebra, 285(2), (2005), 649-668.
[Me2] L. Melkersson, On asymptotic stability for sets of prime ideals connected with the powers of an
ideal, Math. Proc. Cambridge Philos. Soc., 107(2), (1990), 267-271.
[MD] F. Mohammadi Aghjeh Mashhad and K. Divaani-Aazar, Local homology and Gorenstein flat
modules, J. Algebra Appl., 11(2), (2012), 1250022, 8 pp.
[Ri] A.S. Richardson, Co-localization, co-support and local homology, Rocky Mountain J. Math., 36(5),
(2006), 1679-1703.
[Ro] J.J. Rotman, An introduction to homological algebra, Universitext, Springer, New York, second
edition, 2009.
[Sc] P. Schenzel, Proregular sequences, local cohomology, and completion, Math. Scand., 92(2), (2003),
161-180.
[Si1] A.M. Simon, Some homological properties of complete modules, Math. Proc. Camb. Phil. Soc.,
108(2), (1990), 231-246.
[Si2] A.M. Simon, Adic-completion and some dual homological results, Publ. Mat. Camb., 36(2B),
(1992), 965-979.
[St] M. Stokes, Some dual homological results for modules over commutative rings, J. Pure Appl. Algebra, 65(2), (1990), 153-162.
[Va] W. Vasconcelos, Devisor theory in module categories, North-Holland, Amsterdam, 1974.
[WW] S. Sather-Wagstaff and R. Wicklein, Support and adic finiteness for complexes, Comm. Algebra,
to appear.
[Ya] S. Yassemi, Coassociated primes, Comm. Algebra, 23(4), (1995), 1473-1498.
[Yo] K.-I. Yoshida, Cofiniteness of local cohomology modules for ideals of dimension one, Nagoya Math.
J., 147, (1997), 179-191.
[Z] H. Zöschinger, Der Krullsche Durchschnittssatz für kleine Untermoduln, Arch. Math. (Basel),
62(4), (1994), 292-299.
K. Divaani-Aazar, Department of Mathematics, Alzahra University, Vanak, Post Code
19834, Tehran, Iran-and-School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5746, Tehran, Iran.
E-mail address: [email protected]
LOCAL HOMOLOGY, KOSZUL HOMOLOGY AND SERRE CLASSES
23
H. Faridian, Department of Mathematics, Shahid Beheshti University, G. C., Evin, Tehran,
Iran, Zip Code 1983963113.
E-mail address: [email protected]
M. Tousi, Department of Mathematics, Shahid Beheshti University, G.C., Tehran, Iran,
P.O. Box 19395-5746, Tehran, Iran.
E-mail address: [email protected]
| 0math.AC
|
Coding against deletions in oblivious and online models∗
arXiv:1612.06335v4 [cs.IT] 26 Jul 2017
Venkatesan Guruswami†
Ray Li‡
Carnegie Mellon University
Pittsburgh, PA 15213
Abstract
We consider binary error correcting codes when errors are deletions. A basic challenge concerning
deletion codes is determining p0(adv) , the zero-rate threshold of adversarial deletions, defined to be the
supremum of all p for which there exists a code family with rate bounded away from 0 capable of correcting
a fraction p of adversarial deletions. A recent construction of deletion-correcting codes [3] shows that
√
p0(adv) ≥ 2 − 1, and the trivial upper bound, p0(adv) ≤ 12 , is the best known. Perhaps surprisingly, we do
not know whether or not p0(adv) = 1/2.
In this work, to gain further insight into deletion codes, we explore two related error models: oblivious
deletions and online deletions, which are in between random and adversarial deletions in power. In the
oblivious model, the channel can inflict an arbitrary pattern of pn deletions, picked without knowledge
of the codeword. We prove the existence of binary codes of positive rate that can correct any fraction
(obl iv)
p < 1 of oblivious deletions, establishing that the associated zero-rate threshold p0
equals 1.
For online deletions, where the channel decides whether to delete bit x i based only on knowledge
of bits x 1x 2 . . . x i , define the deterministic zero-rate threshold for online deletions p0(on,d ) to be the
supremum of p for which there exist deterministic codes against an online channel causing pn deletions
with low average probability of error. That is, the probability that a randomly chosen codeword is
decoded incorrectly is small. We prove p0(adv) = 21 if and only if p0(on,d ) = 21 .
∗ Research
supported in part by NSF grant CCF-1422045.
[email protected]
‡ Email: [email protected]
† Email:
Contents
1 Introduction
1.1 Our contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Other related work on deletion codes and oblivious/online noise models . . . . . . . . . . .
1
2
3
2 Outline of oblivious deletion codes
4
3 Outline of connection between online and adversarial deletions
8
4 Preliminaries
9
5 Deletion code decoding p-fraction of oblivious deletions
5.1 Overview of proof . . . . . . . . . . . . . . . . . . .
5.2 Construction . . . . . . . . . . . . . . . . . . . . . .
5.3 Analyzing construction behavior . . . . . . . . . . .
5.4 Technical combinatorial lemmas . . . . . . . . . . .
5.5 Proof of Construction (Theorem 2.2) . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10
10
11
11
22
25
6 Relating zero-rate threshold of online and adversarial deletions
26
(on,d)
(adv)
1
1
= 2 . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.1 Proof that p0
= 2 if and only if p0
6.2 Extending to stochastic codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7 Conclusion
30
7.1 Decoding deletions vs. insertions and deletions . . . . . . . . . . . . . . . . . . . . . . . . 30
7.2 Open problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
8 Acknowledgements
31
A Relating average-case and worst-case miscommunication probability
33
B Comparison of oblivious deletions with oblivious bit flips
33
C Proofs of Lemma 5.19 and Lemma 5.20
36
1 Introduction
Tremendous progress has been made over the decades, including in recent years, on the problem of designing
error-correcting codes that can recover from bit (or symbol) errors and erasures, for both probabilistic and
adversarial noise models. The problem of correcting closely related insertion and deletion errors has also
been studied since the work of Levenshtein in the 60s [24], but has not seen commensurate progress. The
difficulty is that in the presence of deletions, the location of the missing symbol is not known to the decoder.
As a result, the sender and receiver lose valuable synchronization information on the location of various
symbols, which renders many of the standard techniques for constructing either inapplicable or at least very
tricky and messy to apply. To quote from Mitzenmacher’s survey [26]: “[C]hannels with synchronization
errors, including both insertions and deletions as well as more general timing errors, are simply not adequately
understood by current theory. Given the near-complete knowledge we have [for] channels with erasures and
errors . . . our lack of understanding about channels with synchronization errors is truly remarkable."
In particular, for both the random and adversarial noise models, there are big gaps in our knowledge
of the power and limitations of codes against insertions and deletions. For concreteness, let us focus our
discussion on the model when there are only deletions, which is already complicated enough, and captures
most of the chief difficulties associated with coding against synchronization errors. Note that over large
alphabets, in particular those which can grow with the code length, one can add the coordinate position as a
header to the codeword symbol, and reduce the deletion model to the simpler erasure model, at the expense
of a negligible decrease in rate (due to the addition of the header). However, for fixed alphabets, deletions
seem much harder to cope with than erasures. We focus on the most interesting case of binary codes in this
work.
In order to set the context and motivation for this work, let us review the situation for adversarial and
random deletions in turn, before turning to the contributions of this paper. In each setting we want a code
C ⊆ {0, 1}n consisting of binary codewords of length n that has good rate (where rate R means |C | = 2 Rn ,
and we would like R to be as large as possible, and certainly bounded away from 0 as n grows). We also
have a deletion fraction p ∈ (0, 1), roughly denoting, as a fraction of n, the number of bits the adversary can
delete. In adversarial deletions, the adversary is allowed to delete up to pn bits with full knowledge of the
code and the transmitted codeword. A code C is decodable against pn adversarial deletions if and only if, for
any two distinct codewords c and c ′ in C, it is impossible to apply pn (possibly different) deletions to c and
c ′ and obtain the same result. This is easily seen to be equivalent to the condition that the longest common
subsequence between any two codewords of C has less than (1 − p)n bits. By a lemma originally due to
Levenshtein [24], this condition also ensures that C is capable of correcting any combination of adversarial
insertions and deletions totaling pn in number. We state this lemma formally as it will be used later.
Lemma 1.1. Let LCS(C) be the maximum length of the longest common subsequence between two distinct
codewords in C. The following are equivalent: (1) LCS(C) ≤ n − t − 1, (2) C decodable under up to
t deletions, (3) C decodable under up to t insertions, and (4) C decodable under up to t insertions and
deletions.
When p ≥ 21 , the adversary can delete n/2 bits that includes either all the 0’s or all the 1’s in the codeword,
resulting in just two possible received sequences. Therefore, it is impossible to correct an adversarial deletion
fraction of 12 with rate bounded away from 0. Rather remarkably, we do not know if this trivial limit can
be approached. Denote the zero-rate threshold, p0(adv) , to be the supremum of all p for which there exist
families of codes C ⊆ {0, 1}n of size 2Ωp (n) decodable against pn adversarial deletions. In other words,
p0(adv) is the threshold at which the capacity of adversarial deletions is zero. An outstanding question in the
area is whether p0(adv) = 21 . Or, is p0(adv) < 12 , so that, for p0(adv) < p < 21 , any code C ⊆ {0, 1}n that is
decodable against pn deletions must have size at most 2o(n) ? This was explicitly raised as a key open problem
1
in [25]. Upper bounds on the asymptotic rate function in terms of the deletion fraction were obtained in [21],
improving in some respects Levenshtein’s bounds [25]. New upper bounds on code size for a fixed number
of deletions that improve over [24] were obtained in [7].
Turning to constructions of binary deletion codes, it was shown in [20] that randomly chosen codes
of small enough rate R > 0 can correctly decode against pn adversarial deletions when p ≤ 0.17. Even
non-constructively, this remained the best achievability result (in terms of correctable deletion fraction) until
recently. Bukh, Guruswami, and Håstad [2, 3] improved this and showed that√there are binary codes of rate
bounded away from 0 decodable against pn adversarial deletions for any p < 2 − 1; furthermore, they also
gave an efficient √
construction of such codes along with an efficient deletion correcting algorithm. Closing
the gap between 2 − 1 and 21 for the maximum fraction of correctable deletions, and determining whether
or not p0(adv) = 21 , remain fascinating open problems.
The other heavily studied model of deletion errors has been random deletions. Here the basic noise
model is the binary deletion channel, which deletes each bit independently with probability p. For random
deletions, the zero-rate threshold is known to be 1. In fact, Drinea and Mitzenmacher [11, 12] proved that
the capacity of random deletion codes is at least (1 − p)/9, which is within a constant factor of the upper
bound of 1 − p, the capacity of the simpler binary erasure channel (where the location of the missing bits are
known at the decoder). Recently, explicit codes of rate c(1 −p) for universal constant c > 0 efficiently coding
against random deletions were constructed in [14]. Rahmati and Duman [27] prove that the capacity is at
most 0.4143(1 − p) for p ≥ 0.65. For large deletion probabilities, this is, to our knowledge, the best known
upper bound. For small deletion probabilities, Diggavi and Grossglauser [10] establish that the capacity of
the deletion channel for p ≤ 21 is at least 1 − h(p). Kalai et al. [18] proved this lower bound is tight as p → 0,
and Kanoria and Montanari [19] determined a series expansion that can be used to determine the capacity
exactly for small deletion probabilities.
1.1 Our contributions
With the above context, we now turn to our results in this work. Our first result concerns a natural model that
bridges between the adversarial and random deletion models, namely oblivious deletions. Here we assume
that an arbitrary subset of pn locations of the codeword can be deleted, but these positions must be picked
without knowledge of the codeword. The oblivious model is well-motivated in settings where the noise may
be mercurial and caused by hard to model physical phenomena, but not by an adversary.
If the code is deterministic, tackling oblivious deletions is clearly equivalent to recovering from worstcase deletions. We therefore allow the encoder to be randomized, and require that for every message
m and every deletion pattern τ , most encodings of m can be decoded from the deletion pattern τ . The
randomness used at the encoding is private to the encoder and is not needed at the decoder, which we
require to be deterministic. Note that the oblivious model generalizes random deletions, as deleting each bit
independently with probability p is oblivious to the actual codeword, and with high probability one has ≈ pn
deletions. Of course, any code which is decodable against pn adversarial deletions is decodable also against
pn oblivious deletions, even without any randomization in the encoding. Perhaps surprisingly, we prove that
in the oblivious model, the limit of p ≤ 12 does not apply, and in fact one can correct a deletion fraction p
approaching 1. This generalizes the result that one can correct a fraction p → 1 of random deletions.
Theorem 1.2 (Oblivious). For every p < 1 there exists a family of constant rate stochastic codes that
corrects against pn oblivious deletions with high probability. More precisely, for every p < 1, there exists
R > 0 and a code family with a randomized encoder Enc : {0, 1} Rn → {0, 1}n and (deterministic) decoder
Dec : {0, 1}(1−p)n → {0, 1} Rn ∪ {⊥} such that for all deletion patterns τ with pn deletions and all messages
m ∈ {0, 1} Rn ,
Pr[Dec(τ (Enc(m))) , m] ≤ on (1) ,
2
where the probability is over the randomness of the encoder (which is private to the encoder and not known
to the decoder).1
The above result is implied by deterministic codes C decodable from arbitrary pn deletions under averageerror criterion; i.e., there is a decoding function Dec : {0, 1}(1−p)n → C such that for every deletion pattern
τ with pn deletions, for most codewords c ∈ C, Dec(τ (c)) = c. We note that Theorem 1.2 is an existential
result, and the codes it guarantees are not explicitly specified. The decoder looks for a codeword which
contains the received bit string as a subsequence, and outputting it if there is a unique such codeword.
We also prove an impossibility result for online deletions. An online channel OnAdv consists of n
functions {OnAdvi : i ∈ [n]} such that OnAdvi : X i × Y i−1 → Y, where X = {0, 1} and Y = {h0i, h1i, hi}
is a set of strings and hi denotes the empty string, satisfies OnAdv(x 1 , . . . , x i , y1 , . . . , yi−1 ) ∈ {hx i i, hi}.
The resulting string received by the output is the concatenation of the outputs of OnAdv1 , . . . , OnAdvn .
More concisely, an online channel chooses whether to delete the ith bit ci of a codeword c based on only the
first i bits c 1 , . . . , ci . Note that all online adversaries are valid adversaries in the omniscient adversary case,
so all codes correcting against pn adversarial deletions can also correct against pn online deletions. Notice
that an online channel can delete all the 0s or 1s in the string, so as in the case of adversarial deletions, it
is impossible to communicate with positive rate when the deletion fraction is at least 12 . However, as the
omniscient adversary has more information than the online channel, it could be the case that one can code
against a fraction of deletions approaching 12 for the online channel, whereas this is not possible against an
omniscient adversary. Our next result rules out this possibility for deterministic codes against the online
channel — if the zero-rate threshold p0(adv) for adversarial deletions is bounded away from 21 , then so is the
deterministic zero-rate threshold for online deletions.
Theorem 1.3 (Online). For any p > 1/(3 − 2p0(adv) ) and R > 0, there exists a deterministic online channel
OnAdv such that, for sufficiently large n, for any code C of block length n and rate R > 0, and any decoder
Dec : {0, 1}(1−p)n → C, we have Prc ∈C [Dec(OnAdv(c)) , c] ≥ η for some absolute constant η > 0.
In contrapositive form, the above states that in order to construct codes against adversarial deletions with
error fraction approaching 21 , it suffices to construct deterministic codes decodable in the average case against
a fraction of online deletions approaching 12 . We note our result does not rule out the possibility that there
exist stochastic codes against a fraction of online deletions approaching 21 in the case that p0(adv) < 21 . In
§6.2, we discuss the subtleties of extending our negative result to rule out stochastic codes against a fraction
of online deletions approaching 21 .
Note that since online deletions contain, as a special case, oblivious deletions, if we insist on deterministic
codes where every codeword is correctly decoded, then the question is just as hard as adversarial deletions.
Thus, it is important to allow some error, in the form of randomized encoding (i.e. stochastic codes) and/or
some small fraction of codewords to be decoded incorrectly (i.e., an average-error criterion). This is why,
for our result to be nontrivial, the online strategy of Theorem 1.3 ensures that a constant fraction, as opposed
to a constant number, of codewords are miscommunicated.
1.2 Other related work on deletion codes and oblivious/online noise models
Our focus in this work is on binary codes. Recently, there has been much progress in the situation for
correcting adversarial deletions over large (but fixed size) alphabets. Guruswami and Wang [16] proved that
for ϵ > 0 there exist polynomial time encodable and decodable codes with alphabet size poly(1/ϵ) correcting
1 − ϵ fraction of deletions with rate Ω(ϵ 2 ), and codes with alphabet size poly(1/ϵ) correcting ϵ fraction of
√
deletions with rate 1 − Õ( ϵ). Guruswami and Li [13] extended these results for high noise and high rate
1The notation τ (y) for y ∈ {0, 1}n denotes the bit string obtained by applying deletion pattern τ to y.
3
regimes to correcting insertions and deletions. Recently, in a beautiful work, Haeupler and Shahrasbi [17]
construct, for any ϵ, δ > 0, codes with alphabet size poly(1/ϵ) and rate 1 − δ − ϵ correcting a δ fraction of
insertions and deletions.
The model of oblivious errors (such as bit flips) has has been studied in the information-theory literature
as a particular case of arbitrarily varying channels with state constraints [23] (see the related work section
of [15] for more background on this connection). In particular, for the case of bit flips, the capacity against
the model of pn oblivious bit flips (for p ≤ 12 ) equals 1 − h(p), matching the Shannon capacity of the binary
symmetric channel that flips each bit independently with probability p. (This special case was re-proved in
[22] by a different simpler random coding argument compared to the original works [5, 6].) Similarly, the
capacity against the model of pn oblivious erasures is 1 − p, matching the Shannon capacity of the binary
erasure channel. Explicit codes of rate approaching 1 − h(p) to correct pn oblivious bit flips (in the sense
of Theorem 1.2, with randomized encoding) were given in [15]. This work also considered computationally
bounded noise models, such as channels with bounded memory or with small circuits, and gave optimal rate
codes for list decoding against those models. These models are more general than oblivious errors, but still
not as pessimistic as adversarial noise.
Notice that in the case of both erasures and errors, the capacity in the oblivious and random models
were the same. It is not clear if this is the case for deletions. While we show that the zero-rate threshold of
oblivious deletions equals the zero-rate threshold of random deletions, the rate of the codes we guarantee in
Theorem 1.2 for p → 1 is significantly worse than the Ω(1 − p) lower bound known for random deletions.
Turning to online models, Chen, Jaggi, and Langberg [4], in an impressive work, determined the exact
capacities of the online erasure channel and the online bit-flip channel. A recent work studies a seemingly
slight (but in fact fundamental) restriction of the online model where the channel’s decision regarding the
i’th bit depends only on the first (i − 1) bits and is independent of the current bit [9]. They proved that in the
setting of erasures, the capacity in this restricted online model increases to match the capacity 1 − p of the
binary erasure channel, as opposed to 1 − 2p in the true online model.
To the best of our knowledge, ours is the first work to address both oblivious and online deletions. We
feel that given the large gaps in our understanding of coding against adversarial deletions, and the potential
of taking advantage of less stringent models of deletion exhibited in this work, further study of these models
seems
timely and important. In particular, for online deletions as well, the zero-rate threshold is between
√
2 − 1 and 21 . It is an interesting challenge to ascertain if one can take advantage of the online restriction, and
push some of the ideas in [3] and this work, to enable decoding a fraction of online deletions approaching 12 .
Organization of this paper. In §2 we outline our oblivious deletion construction and sketch the proof of
why it is correct. In §3 we outline our proof that the online zero-rate threshold for deterministic code, p0(on,d) ,
is 21 if and only if the adversarial zero-rate threshold, p0(adv) , is 21 . In §4 we introduce definitions and notation
for the remainder of the paper. In §5 we state and formally prove our construction on codes in the oblivious
model with p deletion fraction where p ∈ (0, 1). In §6 we prove that the deterministic zero-rate threshold of
the online deletion channel p0(on,d) is also bounded away from 12 if p0(adv) < 21 .
2 Outline of oblivious deletion codes
Instead of proving Theorem 1.2 directly, we prove a related theorem for decoding in the average case, defined
below.
Definition 2.1. We say a (deterministic) binary code C of block length N decodes pN oblivious deletions in
the average case if for every deletion pattern τ deleting pN bits, we have
|{x ∈ C : ∃y ∈ C s. t. x , y and τ (x) ≤ y}| ≤ o N (1) · |C | .
4
(1)
Theorem 2.2 (Correcting oblivious deletions for average message). Let p ∈ (0, 1). There exists a constant
R > 0 such that there exist infinitely many N for which there is a rate R code C ⊆ {0, 1} N that decodes pN
oblivious deletions in average case.
It is standard to show that for oblivious deletions, a randomized code with low probability of incorrect
decoding for every message (as promised in Theorem 1.2), and a deterministic code with low probability of
incorrect decoding for a random message, are equivalent. In particular Theorem 2.2 implies Theorem 1.2.
For completeness we provide a proof of this implication in Appendix A. Roughly, if we have a length N
code C with rate R that decodes against pN deletions in the average case, then we can group the codewords
into sets of size 20.01Rn . Then we associate every message with a set of codewords and encode the message
by randomly choosing a codeword from its set. Our decoding function simply takes a received word s and
looks for a codeword c such that s ⊑ c, i.e. c is a superstring of s. If there is exactly one such c, output the
associated codeword, otherwise output ⊥. Using the fact that C decodes pN deletions in the average case, we
can show that for all deletion patterns τ , only a few codewords do not decode correctly via unique decoding
in our new stochastic code.
We now outline our construction for Theorem 2.2. A first, naive attempt at this problem would be
to choose a random subset of 2 RN codewords in {0, 1} N . This technique, however, does not work in the
same way it does for oblivious bit-flips. See Appendix B for a discussion on the difficulties faced by this
approach. Instead, our approach to prove Theorem 2.2 uses the so-called “clean construction” from [3] with
appropriately selected parameters. The idea is to choose a concatenated code such that the inner code widely
varies in the number of runs between codewords. Specifically, we choose sufficiently large constants R and
K and set our inner code to have length L = 2R K . For i = 1, 2, . . . , K, set our inner codewords to be
i −1 i −1 L/(2R i −1 )
дi = 0R 1R
(2)
where 0k and 1k denote strings of k 0s and 1s, respectively. In this way, the number of runs between any
two codewords differs by a factor of at least R. Our outer code is a subset of [K]n , so that the total code
length is N = nL, and we concatenate the code via a function ψ : [K]∗ → {0, 1}∗ that replaces a symbol
i ∈ [K] with the string дi ∈ {0, 1}L . The outer code is chosen via a random process, detailed in the following
paragraphs; the process throws out a small subset of “bad” elements of [K]n and chooses a constant rate
code by including each remaining element independently with some small fixed probability. Our decoding
function is unique decoding. That is, given a received word s, we find a codeword c such that s ⊑ c. 2 If
such a codeword is unique, our decoder returns that output. Otherwise, the decoder returns ⊥.
The following example illustrates why the varying run length is powerful even for correct more than
0.5N deletions: Suppose n = 100, p = 0.9, R ≫ 20, our received word is s = д110 (that is, 10 copies of д1
concatenated together) and our code contains the codeword c = д2100 . Then s is a subsequence of c if and
only if we can identify each of the 10 д1 s with non-overlapping bits of c. However, since д1 contains over
20 times as many runs as д2 , each д1 must be identified with a subsequence of c spanning at least 20 inner
codewords, i.e. copies of д2 . This means the subsequence s roughly must span at least 200 inner codewords,
but c only has 100 inner codewords, contradiction. While this imbalanced run-count behavior is a key to our
argument, it is worth noting that the behavior is asymmetric. In particular, while it takes R copies of д2 to
produce д1 as a subsequence, we only need two copies of д1 to produce д2 as a subsequence.
To analyze this code, we leverage the run-count behavior of the inner codewords. This contrasts with the
adversarial setting, where the run-count property of the same code is featured less centrally in the proof of
correctness [3]. We show that for any deletion pattern τ with up to pN deletions, two random codewords X
2The relation w ⊑ w ′ means that w is a (not-necessarily consecutive) subsequence of w ′ .
5
and Y are “confusable” with exponentially small probability in N . To be precise, we have for all τ ,
Pr
[τ (ψ (X )) ⊑ ψ (Y )] < 2−Ω(n) .3
X,Y ∼U ([K ]n )
(3)
The idea for this proof can be illustrated in the case that τ deletes only entire inner codewords. If τ
deletes pn of the n inner codewords and does not touch the remaining codewords, then τ (ψ (X )) has the same
distribution over length (1 − p)N binary strings as ψ (X ′) where X ′ ∼ U ([K](1−p)n ). Thus we would like to
show
[ψ (X ′ ) ⊑ ψ (Y )] < 2−Ω(n) .
(4)
Pr
X ′ ∼U ([K ](1−p)n )
Y ∼U ([K ]n )
′
Consider trying to find ψ (X ′) as a substring of ψ (Y ) where X ′ = X 1′ X 2′ . . . X (1−p)n
This is possible if and only
′
if we match the bits of X to bits of Y greedily. However, each inner codeword дX i′ spans a large number
(R) of inner codewords of Y unless the greedy matching encounters at least one Yj such that Yj ≤ X i′ , i.e. a
higher frequency inner codeword (if Yj < X i′ , we may need to encounter two such higher frequency inner
codewords, but two is at least one). Thus, if X i′ = 1 for some i, then the number of inner codewords spanned
by дX i′ is approximately distributed as Geometric(1/K). In general, conditioned on X i′ = k, the number
of inner codewords spanned by дX i′ is approximately distributed as Geometric(k/K). Thus, the number of
inner codewords of Y spanned by a single inner codeword X i′ is
K
K
K
1
1
1
·Θ
+ ·Θ
+···+ ·Θ
= Θ(log K)
(5)
K
1
K
2
K
K
2
then the expected number of inner codewords of Y spanned by X ′ is
If we choose K so that Θ(log K) > 1−p
more than Θ(log K) · (1 − p)n > 2n, so concentration bounds tell us that the probability that ψ (X ′ ) ⊑ ψ (Y ) is
exponentially small in n.
Note that there is a slight caveat to the above argument because the numbers of inner codewords in ψ (Y )
spanned by ψ (X i′) are not independent across all i. For example, if ψ (Y ) begins with д2д1 and ψ (X ′ ) begins
with д1д1 , then the second inner codeword of ψ (Y ) has a few “leftover bits” that easily match with ψ (X ′)’s
second inner codeword. However, we can adjust for this independence by relaxing our analysis by a constant
factor.
The above addresses the case when the deletion pattern, for each inner codeword, either deletes the
codeword entirely or does not modify it at all. We now show how to extend this to general deletion patterns,
starting with the case of p < 21 .
Note that the above argument only depended on the inner codewords having widely varying runs. By
a simple counting argument, we can verify that at least a (0.5 − p) fraction of codewords have at most
(p + (0.5 − p)/2)L = (0.5 + p)L/2 deletions. Since we are in the setting where p < 21 , applying (0.5 + p)L/2
deletions to an inner codeword with r runs cannot make the number of runs less than (0.5 − p)r , as it can
delete at most (0.5 + p)/2 fraction of the runs, and deleting each run reduces the number of runs by at most
two. If we choose R ≫ 1/(0.5 − p)2 , then we can guarantee that even if we have a generic deletion pattern,
we have a constant fraction (0.5 − p) of positions for
√ which the run-count properties of all inner codewords
in those positions are preserved up to a factor of R. Thus, as the number of runs between any two inner
codewords differs by a factor of at least R, even after these
√ corruptions the ratio between the number of
runs of two “preserved” inner codewords is still at least R. Using the same argument as above and now
2
, we can conclude that even for general deletion patterns that the probability that
requiring Θ(log K) > 0.5−p
two random candidate codewords are confused is exponentially small.
3The notation U ([K]n ) denotes the uniform distribution on [K]n .
6
As a technical note, working with deletion patterns directly is messy as they encode a large amount
of information, much of which we do not need in our analysis. Furthermore, the caveat mentioned in the
clean deletion pattern case regarding the independence of number inner codewords spanned by some ψ (X i′ )
becomes more severe for general deletion patterns. This happens because in general deletion patterns,
especially later for p > 21 , the inner codewords of ψ (X ) that have preserved their run-count property might
nonetheless be shorter, so many (in particular, Θ(1/(1 − p))) inner codewords could match to single inner
codewords of ψ (Y ). To alleviate this complexity in the analysis, we introduce a technical notion called
a matching intended to approximate the subsequence relation. This notion allows us to capture only the
run-count behavior of the deletion patterns with respect to the inner codewords while also accounting for the
lack-of-independence caveat. For a deletion pattern τ , let σ be the associated deletion pattern such that σ (X )
to removes all outer codeword symbols except the ones in whose position τ preserves the run-count property
of all the inner codewords (these exist because we are still in the case p < 1/2). In our proof, we argue that
if τ (ψ (X )) is a subsequence of ψ (Y ), then σ (X ) has a matching in Y , and that the probability that σ (X ) has a
matching in Y for two codewords X and Y is exponentially small.
To extend the argument for generic deletion patterns from p < 1/2 to p < 1, we must use a “local list
decoding” idea. Note that when the number of deletion patterns exceeds 1/2, for every codeword there exists
deletion patterns that destroy all or almost all of the information in the codeword, e.g. the deletion pattern
that deletes all the 1s (or 0s) of the codeword. For this reason, codes cannot correct against more than 1/2
fraction of adversarial deletions. However, one can show that this does not happen too frequently allowing
us to correct oblivious and average case deletions. In contrast to the p < 1/2 case where we found a small,
constant fraction of inner codeword positions in which the deletion pattern of the inner codeword preserved
the run-count property for all inner codewords, we can now find a small, constant fraction of inner codeword
positions in which the deletion pattern of the inner codeword preserves the run-count property for all but a
few inner codewords. For example, even if an inner code deletion pattern deletes every other bit and thus
deletes all the information of д1 , the number of runs of д2 , д3 , . . . , дK are still preserved. We call this idea
“local list decoding” because while we cannot decode our constant fraction of inner codewords uniquely,
we can still pin down the inner codewords to a few possibilities. By extending our definition of matching
to account for a few inner codewords potentially losing their run-count behavior, we can prove, just as for
p < 1/2, that τ (ψ (X )) ⊑ ψ (Y ) implies σ (X ) has a matching in Y , and σ (X ) having a matching in Y happens
with small probability.
At this point of the proof, we have combinatorially established everything we need to prove that our
code is decodable in the average case (and thus against oblivious deletions). That is, we have shown that
a random candidate codeword in ψ ([K]n ) has an exponentially small probability of being confused with
another random candidate codeword. Given that codewords have an exponentially small probability of being
confusable with each other, it is natural to consider choosing a code by randomly selecting a subset of [K]n .
Using this construction, we might try using concentration bounds to show that, for any deletion pattern τ ,
the probability that we have more than ϵ |C | codewords (for ϵ = o(N )) that are confusable with some other
codeword is at most 2−ω(N ) , and we can union bound over the at-most-2N choices of τ . This however does
not work directly as the decodability for a given deletion pattern τ depends on the decodability of other
deletion patterns. For example, if p > 21 and we happen to choose c = ψ (11 . . . 1) = 0101 . . . 01 as a
codeword, then for any deletion pattern τ with pN deletions, c ′ ⊑ c for all candidate codewords c ′. From this
example alone, the probability of many codewords confusable with c is at least K −n and there are many more
examples of such easily disguised candidate codewords. Fortunately, we can prove that the number of easily
disguised candidate codewords is small. In particular, we show that the majority of elements of ψ ([K]n )
are not easily disguised in all deletion patterns τ . This intuitively makes sense because, as we have shown,
in any deletion pattern τ , on average, words are disguised as an exponentially small fraction of codewords,
and because the easily disguised words tend to be easily disguised in every deletion pattern τ . For example,
7
ψ (11 . . . 1K ) = 0101 . . . 01 is easily disguised for any deletion pattern τ .
After throwing out the easily disguised candidate codewords, we randomly choose a constant rate code
from the remaining candidate codewords. Careful bookkeeping confirms that with positive probability we
obtain a code that decodes pN -deletions in the average case. The bookkeeping is nontrivial, because just as
there are a handful of words like ψ (11 . . . 1) that are easily disguised as other codewords with deletions, there
are also a handful of easily confused words like ψ (KK . . . K) that can be confused with many other words
when deletions are applied to it. Furthermore, unlike easily disguised codewords, these easily confused
words vary over the different deletion patterns, so we cannot simply throw them out. However, like for easily
disguised codewords, we show the number of easily confused words is an exponentially small fraction of
the codebook size in expectation, so such words do not contribute significantly to the number of incorrectly
decoded codewords. Note the subtle difference between easily disguised and easily confused words: a single
easily disguised word like ψ (11 . . . 1) causes many candidate codewords to fail to decode under our unique
decoding, but any easily confused codeword adds at most one failed decoding.
We model managing easily disguised and easily confusable codewords via a directed graph, where,
−−→
roughly, for each deletion pattern, we consider a graph on [K]n where YX is an edge if and only if
τ (ψ (X )) ⊑ ψ (Y ). In our proof, we replace the subsequence relation with the matching relation (see §5.4).
In this graph language, the easily disguised codewords correspond to vertices with high outdegree, and the
easily confusable codewords correspond to vertices with high indegree.
Our construction illustrates the subtle nature of the oblivious deletion channel and average case errors.
These settings share much of the behavior of the adversarial deletion channel such as the fact that for p > 12 ,
every codeword has a deletion pattern destroying all of its information. Consequently, our approach tackles
the oblivious and average case errors using a combinatorial argument just as the best adversarial deletion
results do [3]. Yet, the relaxed decoding requirement allows us to exploit it to correct a fraction of deletions
approaching 1.
3 Outline of connection between online and adversarial deletions
We give a high level outline of the proof of Theorem 1.3. Recall that the goal is to prove that, assuming
p0(adv) < 12 , for some p < 21 , for every deterministic code of rate bounded away from 0, there exists a
deterministic online deletion channel applying up to pn deletions that prevents successful decoding in the
average case. That is, when a uniformly random codeword is transmitted, the probability of incorrect
decoding is bounded away from 0. Note first that it suffices to find a randomized online deletion strategy
guaranteeing average-case decoding error because, if for a given code the channel has a random strategy to
guarantee average-case decoding error, then sampling a strategy over the randomness of the channel gives a
deterministic online strategy that inflicts similar probability of incorrect decoding.
To do this, we assume that the adversarial zero-rate threshold is p0(adv) < 12 , and consider p ∗ = 1/(3 −
(adv)
(adv)
2p0
). Our choice of p ∗ satisfies p0
< p ∗ < 21 . Now we show that for any p ∗ < p < 12 , there exists a
randomized strategy that inflicts pn deletions in an online manner to a deterministic code, and guarantees that
the probability, over the randomness of a uniformly chosen codeword and the randomness of the adversary,
1
of incorrect decoding is positive (we show it is at least 10
).
We adapt the “wait-push” strategy of Bassily and Smith [1] for an online adversary against erasures. In
the “wait phase” of this strategy, the adversary observes a prefix x ∗ of ℓ bits from the transmitted word x
without erasing any bits. Then it constructs a list Lx ∗ of candidate codewords, among which is the actual
codeword chosen by the encoder. In the “push phase”, the adversary chooses a codeword x ′ randomly from
Lx ∗ , which, with positive probability, corresponds to a different message than the actually transmitted word.
The adversary then, for the last n − ℓ bits of the transmission, erases bit x i of x where x i , x i′ (x i′ the ith bit
8
of x ′ ).
Our strategy adapts the wait push strategy to deletions. In our wait phase, we observe a prefix of the
string until we know the codeword x exactly. We choose in advance a random bit and delete every instance
of that bit that we see during the wait phase. After the wait phase, suppose we have seen qn bits and deleted
rn bits, where r < q, and we also know the codeword x exactly. Using the definition of p0(adv) , we can show
that, for most choices of x, there is another codeword y such that (i) the wait phase lengths qn for x and y
are the same, (ii) their majority bits b in the wait phases are the same, (iii) the number of bits rn deleted in
the wait phases is the same, and (iv) the length-(1 − q)n suffixes x ∗ and y ∗ of x and y, respectively, have a
common subsequence s ∗ satisfying |s ∗ | > (1 − p0(adv) )|x ∗ |. We can form pairs of such codewords (x, y) in
advance, so that when we see either x or y we push the suffixes x ∗ or y ∗ to s ∗ , so that the received word in
both cases is b r n s ∗ . For these typical choices of x, the total number of required deletions to obtain b r n s ∗ is at
(adv)
most qn/2 + p0
(1 − q)n. If q is bounded away from 1 (in our case, less than (1 − p)n), then we can bound
the number of deletions away from 21 (in our case, less than p ∗ = 1/(3 − 2p0(adv) )). On the other hand, if q
is large (larger than (1 − p)n), then we can simply observe the first (1 − p)n bits and delete the rest. Since
we can only choose one strategy to run, we choose one of these two strategies at random. This gives our
modified wait push strategy.
4 Preliminaries
General Notation. For a boolean statement P, let 1[P] be 1 if P is true and 0 otherwise.
Throughout the paper, log x refers to the base-2 logarithm.
We use interval notation [a, b] = {a, a + 1, . . . , b} to denote intervals of integers, and we use [a] = [1, a] =
{1, 2, . . . , a}.
For a set S and an integer a, let Sa denote the family subsets of S of size a. Let U (S) denote the uniform
distribution on S.
Words. A word is a sequence of symbols from some alphabet. We denote string concatenation of two
words w and w ′ with ww ′. We denote w k = ww · · · w where there are k concatenated copies of w. We also
Î
denote a concatenation of a sequence of words as w 1w 2 · · · w k = ki=1 w i . We denote words from binary
alphabets with lowercase letters c, s, w and words from non-binary alphabets with capital letters X , Y , Z .
A subsequence of a word w is a word obtained by removing some (possibly none) of the symbols in w.
A subword or interval of a word w is a contiguous subsequence of characters from w. We identify
intervals of words with intervals of integers corresponding to the indices of the subsequence. For example,
the interval {1, 2, . . . , |w |} = [1, |w |] is identified with the entire word w.
Let w ′ ⊑ w denote “w ′ is a subsequence of w”.
Define a run of a word w to be a maximal single-symbol subword. That is, a subword w ′ in w consisting
of a single symbol such that any longer subword containing w ′ has at least two different symbols. Note the
runs of a word partition the word. For example, 110001 has 3 runs: one run of 0s and two runs of 1s.
Deletion Patterns. A deletion pattern is a function τ that removes a fixed subset of symbols from words
of a fixed length. Let D(n, m) denote the set of deletion patterns τ that operate on length n words and apply
n
exactly m deletions. For example τ : x 1x 2x 3 7→ x 1x 3 is a member of D(3, 1). Let D(n) = ∪m=0
D(n, m).
We identify each deletion pattern τ ∈ D(n, m) with a size m subset of [n] corresponding to the deleted
bits. We often use sets to describe deletion patterns when the length of the codeword is understood. For
example [n] refers to the single element of D(n, n). Accordingly, let ⊆ be a partial order on deletion patterns
corresponding to set inclusion, and let |τ | denote the number of bits deleted by τ . As such, we have τ ⊆ τ ′
implies |τ | ≤ |τ ′ |.
9
For a word w and τ ∈ D(|w |), let τ (w) and w \ τ both denote the result of applying τ to w. We use
the second notation when we identify sets with deletion patterns, as in the above paragraph where the set
elements correspond to the deleted positions.
In a concatenated code with outer code length n and inner code length L, we can identify a deletion pattern
on the entire codeword τ ∈ D(nL) as the “concatenation” of n deletion patterns τ1 ⌢ · · · ⌢ τn , one for each
inner codeword. To be precise, for all τ ∈ D(nL), there exists τi ∈ D(L) such that τ = ∪ni=1 {j+(i−1)L : j ∈ τi },
and we denote this by τ = τ1 ⌢ · · · ⌢ τn . Using this notation, we refer to τi as an inner code deletion
pattern.
Graphs. In a (directed or undirected) graph G, let V (G) and E(G) denote the vertex set and edge set of G
respectively. For a subset W ⊆ V (G) of the vertices, let G ↾ W denote the subgraph induced by W . We use
G ↾ W instead of the more standard G |W to avoid too many subscripts. For a vertex v ∈ V (G), let degG (v)
denote the degree of v in G when G is undirected, and let indegG (v), outdegG (v) denote the indegree and
outdegree of v, respectively, in G when G is directed. We drop the subscript of G in deg, indeg and outdeg
notations when the graph G is understood.
Concentration Bounds. We use the following forms of Chernoff bound.
Lemma 4.1 (Chernoff). Let A1 , . . . , An be independent random variables taking values in [0, 1]. Let
Í
A = ni=1 Ai and δ ∈ [0, 1]. Then
Pr[A ≤ (1 − δ ) E[A]] ≤ exp −δ 2 E[A]/2 .
(6)
Furthermore, if A1 , . . . , An are Bernoulli random variables, then
Pr[A ≥ (1 + δ ) E[A]] ≤
eδ
(1 + δ )( 1 + δ )
E[A]
(7)
We also use the submartingale form of Azuma’s Inequality.
Lemma 4.2 (Azuma). Let c be a constant. Let X 1 , X 2 , . . . be a submartingale such that |X i − X i−1 | ≤ c for
all i. Then for all positive reals t, we have
−t 2
Pr[X k − X 1 ≤ −t] ≤ exp
.
(8)
2c 2 (k − 1)
5 Deletion code decoding p-fraction of oblivious deletions
5.1 Overview of proof
In §2 we gave a high level overview. We now begin with a brief snapshot of the proof structure and how it is
organized.
We present our general code construction in §5.2. The construction uses the “clean construction” in [3]
and an outer code that we choose randomly. We analyze properties of the concatenated construction in §5.3.
We begin by extracting the useful behavior of deletion patterns with respect to our codewords. This deletion
pattern analysis culminates in Lemma 5.7, allowing us to define the signature (Definition 5.8) of a deletion
pattern. The key result of §5.3 is Proposition 5.18. It states that for any deletion pattern τ , the probability
that two random candidate codewords c, c ′ are confusable (in the sense that τ (c) ⊑ c ′ ) is exponentially small
in the code length. However, because working with deletion patterns directly is messy, the proposition is
written in the language of matchings, a technical notion defined in Definition 5.11. In short, because the inner
10
codewords are nicely behaved, we do not need to know the exact details of the behavior of a given deletion
pattern τ , but rather only need certain properties of it, given by its signature. We thus define a matching to
approximate the subsequence relation using only the signature of τ , so that “σ (X ) is matchable in Y ” (where
σ is the outer code deletion pattern given by τ ’s signature) holds roughly when “τ (ψ (X )) ⊑ ψ (Y )” holds.
Combinatorially, Proposition 5.18 allows us to finish the proof. As stated in §2, for our outer code we
consider [K]n minus a small set of easily disguised candidate codewords. The notion of a easily disguised
candidate codeword is well defined by Lemma 5.23. For a sufficiently small constant γ , we randomly choose
a size 2γ n outer code over the remaining outer codewords. In §5.4, we use the graph language described in
§2 and prove Lemma 5.27, which roughly states that in a sparse directed graph, if we randomly sample a
small subset of vertices, the induced subgraph is also sparse (for some appropriate definition of sparse) with
high probability. Finally, in §5.5, we piece together these results, showing that Lemma 5.27 guarantees, with
positive probability, that our random code combinatorially decodes against pN deletions in the average case.
5.2 Construction
Let p ∈ (0, 1), and let λ = λ(p) be the smallest integer such that (1 + p)/2 < 1 − 2−λ . For our argument any
λ such that p < 1 − 2−λ suffices. In particular, for p < 12 , we can choose λ = 1, slightly simplifying the
argument as described in §2. However, we choose λ to be the smallest λ such that (1 + p)/2 < 1 − 2−λ to
ensure a clean formula for the rate.
Let δ be such that p = 1 − 2−λ − δ . Let n be a positive integer. With hindsight, choose
K = 2 ⌈2
λ+5 /δ ⌉
,
R = 4K 4 ,
L = 2R K ,
N = nL.
(9)
Note that R is even. For the remainder of this section, the variables p, λ, δ, K, R and L are fixed.
In this way we have 1 − 2−λ − √1 − δ2 > p. For i = 1, . . . , K, let дi be the length L word
R
i −1 i −1 L/(2R i −1 )
дi = 0R 1R
.
(10)
Consider the encodingψ : [K]∗ → {0, 1}∗ whereψ (X 1 · · · X k ) = дX 1 дX 2 · · · дX k . We construct a concatenated
code where the outer code is a length n code over [K] and the inner code is {д1 , д2 , . . . , дK } ⊆ {0, 1}L . For
the outer code, we choose a random code Cout where each codeword is chosen uniformly at random from
[K]n minus a small undesirable subset that we specify later. We choose our decoding function to be unique
decoding. That is, our decoder iterates over all codewords c ∈ C and checks if the received word s is a
subsequence of c. If it is a subsequence of exactly one c, the decoder returns that c, otherwise it fails. While
this decoder is not optimal in terms of the fraction of correctly decoded codewords (it could try to break
ties instead of just giving up), it is enough for this proof. Furthermore, since we are showing combinatorial
decodability, we do not need the decoder to be efficient.
If, for some p ′ > p, a code can decode p ′nL average case or oblivious deletions, then it can decode pnL
deletions. Thus, we can decrease δ until δn is an even integer, so may assume without loss of generality that
δn is an even integer.
5.3 Analyzing construction behavior
Definition 5.1. A дi -segment is an interval in [L] corresponding to a run of дi . Note that the дi -segments
partition [L] and are of the form [1 + aRi−1 , (a + 1)Ri−1 ] for a ∈ {0, . . . , L/Ri−1 − 1}.
11
Note that дi has 2R K +1−i runs. In particular, the number of runs greatly varies between inner codewords.
This property makes the concatenated construction powerful because it is difficult to find common subsequences of different inner codewords. The following definition allows us to reason about the inner codewords
in terms of their run counts.
√
Definition 5.2. We say an inner code deletion pattern σ preserves дi if σ (дi ) has at least 2R K +1−i / R =
1
2R K + 2 −i runs. Otherwise we say σ corrupts дi .
We start with a basic but useful fact about deletion patterns and runs.
Lemma 5.3. Suppose w is a word with r runs I1 , . . . , Ir and τ is a deletion pattern such that τ (w) has r ′
runs. We can think of these runs Ik as subsets of consecutive indices in {1, . . . , |w |}. Then the number of
runs Ik of w completely deleted by τ , i.e. satisfying I ⊆ τ when τ is thought of as a subset of {1, . . . , |w |}, is
′
at least r −r
2 .
Proof. Deleting any run reduces the number of runs in a word by at most 2, and τ reduces the number of
runs by r − r ′ , so the claim follows.
The next lemma establishes the usefulness of widely varying runs in our construction. It says that even
when an inner code deletion pattern has a large number of deletions, most of the inner codewords still look
the same in terms of the number of runs. The intuition for the lemma is as follows. Consider the extreme
example of an inner code deletion pattern σ that “completely corrupts” the inner codewords д1 , . . . , дλ . That
is, σ deletes all the zeros of each of д1 , . . . , дλ . Since σ deletes all the zeros of дλ , it must delete every other
run of R λ−1 bits, thus deleting L/2 bits. Applying these deletions alone to дλ−1 leaves it with half as many
runs of length exactly R λ−2 . However, since σ also deletes all zeros of дλ−1 , it must delete every other run of
R λ−2 bits of the remaining L/2 bits, thus deleting L/4 more bits. Similarly, since σ deletes all zeros of дλ−2 ,
it must delete an additional L/8 bits. Continuing this logic, we have σ must delete L(1 − 2−λ ) bits total. This
tells us that if an inner code deletion pattern completely corrupts the inner codewords д1 , . . . , дλ , it needs
L(1 − 2−λ ) deletions. This logic works even if we chose to corrupt any subset of λ inner codewords other
than {д1 , . . . , дλ }. One can imagine that corrupting (according to Definition 5.2) inner codewords is almost
as hard as completely corrupting them, so adding some slack gives the lemma.
Lemma 5.4. If σ is an inner code deletion pattern with |σ | ≤ L 1 − 21λ − √1 , then σ preserves all but at
R
most λ − 1 choices of дi .
Proof. Suppose λ is a positive
integer such
that σ corrupts дi for λ different values of i, say i 1 > i 2 > · · · > i λ .
1
1
We wish to show |σ | > L 1 − 2λ − √ .
R
Recall that a дi segment is an interval of the form [1 + aRi−1 , (a + 1)Ri−1 ]. Inductively define the
collections of intervals I1 , . . . , Iλ and the sets of indices I1 , . . . , I λ as follows. For 1 ≤ a ≤ λ, set
)
(
a−1
Ø
Ø
Ib
and Ia =
J.
(11)
Ia = J : J is дia -segment, J ⊆ σ \
b=1
J ∈Ia
Intuitively, I1 as the set of runs in дi1 that are entirely deleted by σ , and I1 is the set of those deleted indices.
Then, I2 is the set of runs of дi2 deleted by σ but not already accounted for by I1 , and I2 is the set of bits in
the runs of I2 . We can interpret I3 , . . . , Iλ and I3 , . . . , I λ similarly. By construction, I1 , . . . , I λ are disjoint,
Íλ
and their union is a subset of σ (thought of as a subset of [L]), so b=1
|Ib | ≤ |σ |. It thus suffices to prove
1
1
|Ib | > L 1 − λ − √ .
2
R
b=1
λ
Õ
12
(12)
′
Note that for any j < j ′, every дj ′ -segment is the disjoint union of R j −j many дj -segments. We thus have [L]
is the disjoint union of дia -segments and Ib is also the disjoint union of дia -segments when b < a (and thus
a−1 I is the disjoint union of д -segments. Furthermore, as R is even, each I covers
ib > i a ). Hence, [L] \ ∪b=1
ia
b
b
a−1 I alternate between segments corresponding
an even number of дia -segments, so the segments of [L] \ ∪b=1
b
a−1 I
to runs of 0s in дia and segments corresponding to runs of 1s in дia . It follows that all runs of дia \ ∪b=1
b
i
−1
a−1
a
have length exactly R
, so the number of runs in the string дia \ ∪b=1 Ib is
L
Ria −1
−
a−1
Õ
b=1
Rib −ia |Ib |.
(13)
By construction, the only дia -segments that are deleted√by σ are the intervals covered by I1 ∪ · · · ∪ Ia . Since
σ corrupts дia , we know σ (дia ) has less than L/(Ria −1 R) runs. By Lemma 5.3, we have
!
!
a−1
Õ
L
1
L
Rib −ia |Ib | −
−
|Ia | >
(14)
√ .
2 Ria −1
Ria −1 R
b=1
Simplifying and using |Ib | = Rib −1 |Ib | for all b, we obtain
|Ia | > L
a−1
1
1Õ
1
|Ib |.
− √ −
2 2 R
2
b=1
(15)
From here it is easy to verify by induction that, for all 1 ≤ a ≤ λ, we have
a
Õ
1
1
|Ib | > L 1 − a 1 − √ .
2
R
(16)
b=1
Indeed, (15) for a = 1 provides the base case, and if we know (16) for some a − 1, then by (15) we have
a−1
1
1
1Õ
|Ib | > L
|Ib |
− √ +
2 2 R
2
b=1
b=1
1
1
L
1
1
> L
− √ +
1 − a−1 1 − √
2 2 R
2
2
R
1
1
= L 1− a 1− √ ,
2
R
a
Õ
(17)
completing the induction. The induction proves (12), from which we have
1
1
|Ib | > L 1 − λ − √ ,
|σ | ≥
2
R
b=1
λ
Õ
(18)
as desired.
Lemma 5.4 motivates the following definition.
Definition 5.5. We say an inner code deletion pattern |σ | is ℓ-admissible if |σ | ≤ L(1 − 1/2ℓ+1 −
√1 ).
R
If, for some ℓ, σ is ℓ-admissible, then Lemma 5.4 tells us σ corrupts at most ℓ different дi . However,
note that ℓ-admissibility is stronger that corrupting at most ℓ different дi as ℓ-admissibility gives a stronger
upper bound on the number of deletions in σ , which is necessary in Lemma 5.12.
13
Lemma 5.6. Let δ > 0. Let τ = τ1 ⌢ · · · ⌢ τn be a deletion pattern with at most (1 −
deletions. There are at least δn indices i such that τi is (λ − 1)-admissible.
Proof. By a simple counting argument, we have |τi | > L(1 −
1
2λ
−
√1 )
R
1
2λ
−
√1
R
− δ2 )N
for at most
n · L 1 − 21λ − √1 − δ2
δ /2
R
≤ n (1 − δ )
≤ n 1−
1 − 2−λ
L 1 − 21λ − √1
(19)
R
values of i. For the remaining at least δn values of i, we have τi is (λ − 1)-admissible.
The following corollary allows us to reduce our analysis of a deletion pattern τ to analyzing positions
where τ ’s inner code deletion pattern is (λ − 1)-admissible. We effectively assume that our deletion pattern
completely deletes all inner codewords with a non-admissible index.
Lemma 5.7. Let τ ∈ D(nL, pnL). There exists τ ′ ∈ D(δnL), σ ∈ D(n, (1 − δ )n) and sets S 1 , . . . , S δ n such
that
1. |S i | = λ − 1 for all i,
2. for all X ∈ [K]n , we have τ ′(ψ (σ (X ))) ⊑ τ (ψ (X )), and
3. when we write τ ′ = τ1′ ⌢ · · · ⌢ τδ′ n as the concatenation of δn inner code deletion patterns, we have,
for all i and all j < S i , that τi′ ∈ D(L) preserves дj .
Proof. Let τ = τ1 ⌢ · · · ⌢ τn . By Lemma 5.6, there exist δn indices ℓ1 < · · · < ℓδ n such that, for
i = 1, . . . , δn, τ ℓi is (λ − 1)-admissible. Choose σ ∈ D(n, (1 − δ )n) via σ (X 1 . . . X n ) = X ℓ1 X ℓ2 . . . X ℓδ n , and
choose τ ′ ∈ D(δnL) via τ ′ = τℓ1 ⌢ · · · ⌢ τℓδ n . We have τ ′ ◦ ψ ◦ σ (X ) ⊑ τ ◦ ψ (X ) for all X ∈ [K]n because
τ ′ ◦ψ ◦ σ (X ) is simply the result of deleting the remaining bits in inner codewords of non-admissible indices
in τ ◦ ψ (X ). By construction, each τℓi ∈ D(L) is (λ − 1)-admissible, so we can choose S 1 , . . . , S δ n by setting
S ℓ to be the set that τℓi corrupts. Note that some S i may have size less than λ − 1, but we can arbitrarily add
elements of [K] to S i until it has λ − 1 elements. This is okay as item 3 in the corollary statement remains
true if we add elements to Si .
In our analysis, for a deletion pattern τ = τ1 ⌢ · · · ⌢ τn , we only care about the behavior of a given inner
code deletion pattern τi as far as the set Si of inner codewords дj that it corrupts; instead of considering all
possible deletion patterns τ ∈ D(nL), it suffices to only consider all possible σ , S 1 , . . . , Sδ n . This motivates
the following definition.
Definition 5.8. The signature of a deletion pattern τ is (σ , S 1 , S 2 , . . . , S δ n ), where σ , S 1 , . . . , S δ n are given
by Lemma 5.7. If the choice of σ , S 1 , . . . , Sδ n satisfying the conditions of Lemma 5.7 are not unique, then
choose one such collection of σ , S 1 , . . . , Sδ n arbitrarily and assign this collection to be the signature of τ .
Below we define the matchability relation ≺. Definition 5.11 allows us to worry only about the signature
of a deletion pattern τ rather than τ itself. Intuitively, we can think of the matchability relation ≺ as
an approximation of the subsequence relation ⊑. Proposition 5.12 establishes this relationship formally.
Specifically, it states that if τ is a deletion pattern with signature (σ , S 1 , . . . , S δ n ), then for X , Y ∈ [K]n , we
have σ (X ) ⊑ Y implies that σ (X ) has a matching in Y with appropriate parameters. This means that if
we want to show there are few incorrectly decoded codewords in a given code, it suffices to show that few
codewords have an appropriately parameterized matching in some other codeword.
14
We first define type-A and type-B pairs of indices (i, j) ∈ {1, . . . , |X |} × {1, . . . , |Y |}. Intuitively, pairs
(i, j) are type-B only if τi (ψ (X i )) has many more runs than ψ (Yj ), i.e. it is difficult to find (contiguous)
subwords of τi (ψ (X i )) as subsequences of ψ (Yj ).
Definition 5.9. Let X , Y ∈ [K]∗ be words over the alphabet [K] and let S 1 , . . . , S |X | be subsets of K. Given
a pair (i, j) ∈ {1, . . . , |X |} × {1, . . . , |Y |}, we say (i, j) is type-A with respect to X , Y , S 1 , . . . , S |X | (or simply
type-A if the parameters are understood) if X i ∈ S i or X i ≥ Yj . Call a pair (i, j) type-B with respect to
X , Y , S 1 , . . . , S |X | otherwise.
Definition 5.10. Let X , Y ∈ [K]∗ be words over the alphabet [K], let S 1 , . . . , S |X | be subsets of K, and let
s and t be positive integers. The following algorithm constructs the (s, t, S 1 , . . . , S |X | ) matching of X with
Y . Begin with a pair (a, b) = (1, 1). The first and second coordinates correspond to indices of the strings X
and Y , respectively. Define an A-move to be incrementing the first coordinate, a, by 1, and a B-move to be
incrementing of the second coordinate, b, by 1.
1. If a = |X | or b = |Y |, stop.
2. Do one of the following
(a) If the last s moves were A-moves, make a B-move.
(b) Else if the last t moves were B-moves, make an A-move.
(c) Else if (a, b) is type-A, make an A-move.
(d) Else, (a, b) must be type-B, in which case make a B-move.
3. Repeat from step 1.
Note that at the end of this algorithm, exactly one of a = |X | and b = |Y | is true. We say this matching is a
success if we ended with a = |X |, otherwise it is a failure.
Definition 5.11. Note also that the matching is uniquely determined by X , Y , S 1 , . . . , S |X | , s, t. If this matching
is a success, we say X is (s, t, S 1 , . . . , S |X | )-matachable (or has a (s, t, S 1 , . . . , S |X | )-matching) in Y , denoted
X ≺(s,t,S 1, ...,S |X | ) Y .
(20)
Proposition 5.12. Let S 1 , . . . , S δ n be subsets of [K] of size exactly λ − 1. Let τ = τ1 ⌢ · · · ⌢ τδ n ∈ D(nL)
be a deletion pattern such that for all i, τi ∈ D(L) is (λ − 1)-admissible and in particular preserves дj for all
j < S i . Suppose we have X ∈ [K]δ n and Y ∈ [K]n such that τ (ψ (X )) ⊑ ψ (Y ) (recall ⊑ is the subsequence
relation). Then X ≺(2λ , √R,S 1, ...,S δ n ) Y .
√
Proof. Let s = 2λ , t = R. Run the matching algorithm defined above to obtain a matching of X and Y .
Let M be the set of all (a, b) reached by some step of the algorithm. We wish to show this matching is a
success, i.e. that there exists some b such that (|X |, b) ∈ M, or, equivalently, there does not exist a such that
(a, |Y |) ∈ M.
Since τ (ψ (X )) ⊑ ψ (Y ), we can find τ (ψ (X )) as a subsequence of ψ (Y ) by greedily matching the bits of
τ (ψ (X )) with the bits of ψ (Y ). Let N be the set of (i, k) such that some bit of τi (ψ (X i )) is matched with some
bit in ψ (Yk ). We first establish some basic facts about M, N .
Fact 5.13.
1. (|X |, |Y |) < M.
15
Y
X
β0 κ0 βf κf
a
...
Y
...
X
b
...
...
ι0 α0 ιf αf
β 0 (α 0 (β 0 (α 0 (b))))
Y
X
β 0 (α 0 (b))
b
...
...
α 0 (β 0 (α 0 (b)))
α 0 (b)
Figure 1: Illustrations of α 0 (b), α f (b), β 0 (a), β f (a), ι 0 (b), ι f (b), κ 0 (a), κ f (a)
2. M, N ⊆ {1, . . . , |X |} × {1, . . . , |Y |}.
3. For all a ∗ ∈ {1, . . . , |X |}, b ∗ ∈ {1, . . . , |Y |}, we have {b : (a ∗ , b) ∈ M} and {a : (a, b ∗ ) ∈ M} are
intervals of consecutive integers of lengths at most t + 1 and s + 1, respectively.
4. For all i ∗ ∈ {1, . . . , |X |}, k ∗ ∈ {1, . . . , |Y |}, we have {k : (i ∗ , k) ∈ N } and {i : (i, k ∗ ) ∈ N } are
intervals of consecutive integers.
5. Let ≤ M be a relation on M such that (a, b) ≤ M (a ′, b ′ ) iff a ≤ a ′ and b ≤ b ′ . Then M is totally
ordered under ≤ M . As such, we can define next M (a, b) and prev M (a, b) to be the next larger and
next smaller element after (a, b) under ≤ M , respectively. Then next M (a, b) ∈ {(a + 1, b), (a, b + 1)}
and prev M (a, b) ∈ {(a − 1, b), (a, b − 1)}.
6. Let ≤ N be a relation on N such that (i, k) ≤ N (i ′, k ′) iff i ≤ i ′ and k ≤ k ′. Then N is totally ordered
under ≤ N . As such, we can define next N (i, k) and prev N (i, k) to be the next larger and next smaller
element after (i, k) under ≤ N , respectively. Then next N (i, k) ∈ {(i + 1, k), (i, k + 1), (i + 1, k + 1)} and
prev N (i, k) ∈ {(i − 1, k), (i, k − 1), (i − 1, k − 1)}.
7. If (a, b), (a ′ , b ′ ) ∈ M, we never have both a < a ′ and b ′ < b.
8. If (i, k), (i ′ , k ′) ∈ N , we never have both i < i ′ and k ′ < k.
For a ∈ {1, . . . , |X |} and b ∈ {1, . . . , |Y |}, define
α 0 (b) = min{a : (a, b) ∈ M}
α f (b) = max{a : (a, b) ∈ M}
ι 0 (b) = min{i : (i, b) ∈ N }
ι f (b) = max{i : (i, b) ∈ N }
β 0 (a) = min{b : (a, b) ∈ M}
β f (a) = max{b : (a, b) ∈ M}
κ 0 (a) = min{k : (a, k) ∈ N }
κ f (a) = max{k : (a, k) ∈ N }.
(21)
See Figure 1 for illustrations of the behavior of these eight functions. We first establish a few facts about the
notation α 0 , β 0 , . . . that are helpful for developing intuition and are also useful later. These proofs are more
involved than Fact 5.13 and are provided.
√
Lemma 5.14.
1. For all a ∈ {1, . . . , |X |}, if κ f (a) − κ 0 (a) < R, then there exists b ′ ∈ [κ 0 (a), κ f (a)]
such that (a, b ′ ) is type-A.
√
2. For all a ∈ {1, . . . , |X |} we have β f (a) − β 0 (a) ≤ R and for all β 0 (a) ≤ b ′ < β f (a) we have (a, b ′ ) is
√
type-B Furthermore, if β f (a) − β 0 (a) < R, then (a, β f (a)) is type-A.
3. For all b ∈ {1, . . . , |Y |}, we have ι f (b) − ι 0 (b) ≤ 2λ .
4. For all b ∈ {1, . . . , |Y |}, we have (i ′, b) is type-A for all i ′ ∈ [ι 0 (b) + 1, ι f (b) − 1].
16
Y
β0 κ0 κf βf
X
...
...
a
Figure 2: Step 2, Assume FSOC β 0 (a) ≤ κ 0 (a) ≤ κ f (a) < β f (a)
5. For all b ∈ {1, . . . , |Y |}, we have α f (b) − α 0 (b) ≤ 2λ and for all a ′ ∈ [α 0 (b), α f (b) − 1] we have (a ′, b)
is type-A. Furthermore, if α f (b) − α 0 (b) < 2λ , then (α f (b), b) is type-B.
Proof. Parts 2 and 5 follow from Definition 5.10.
For part 1, suppose for contradiction that (a, b ′ ) is type-B for all κ 0 (a) ≤ b ′ ≤ κ f (a). Then X a < S a
1
and X a < Yb ′ for all such b ′ . This means τa (ψ (X a )) has at least 2R K + 2 −X a runs while ψ (Yb ′ ) has at most
2R K +1−(X a +1) runs for κ 0 (a) ≤ b ′ ≤ κ f (a). On the other hand, we have
τa (ψ (X a )) ⊑ ψ Yκ0 (a) . . . Yκf (a) .
(22)
√
√
1
As κ f (a) − κ 0 (a) < R this means the right side of (22) has less than R · 2R K −X a = 2R K + 2 −X a runs while
the left side has at least that many runs, a contradiction.
Î ι (b)−1
For part 3, suppose for contradiction that ι f (b)−ι 0 (b)−1 ≥ 2λ . Sinceψ (Yb ) contains if′=ι (b)+1 τi ′ (ψ (X i ′ ))
0
as a strict subsequence (ψ (Yb ) additionally contains at least one bit from each of τι0 (ψ (X ι0 )) and τι f (ψ (X ι f ))),
we have
ι f −1
L
© Õ
ª
|τi ′ (ψ (X i ′ ))| ® + 2 ≤ |ψ (Yb )| = L,
L + 2 ≤ (ι f (b) − ι 0 (b) − 1) · λ + 2 ≤
2
′
«i =ι0 +1
¬
(23)
a contradiction.
For part 4, suppose for contradiction that (i ′, b) is type-B for some ι 0 (b) < i ′ < ι f (b). Thus X i ′ < S i ′
1
and X i ′ < Yb . In particular, τi ′ (ψ (X i ′ )) has at least 2R K + 2 −X i ′ runs, which is more than the at-most-2R K −X i ′
runs of ψ (Yb ). However, ι 0 (b) < i ′ < ι f (b), so (i ′, b) ∈ N and in particular τi ′ (ψ (X i ′ )) ⊑ ψ (Yb ), which is a
contradiction. Note that i ′ = ι 0 (b) and i ′ = ι f (b) do not guarantee a contradiction because for such i ′, some
bits of τi ′ (ψ (X i ′ )) might be matched with other inner codewords in ψ (Y ).
The following definition of proper indices is introduced for convenience. Intuitively, indices of Y are
Y -proper if the bit-matching N consumes corresponding indices of X “slower” than in the algorithmic
matching M, and indices of X are X -proper if the bit-matching N consumes corresponding indices of Y
“faster” than in the algorithmic matching M.
Definition 5.15. We say an index in a ∈ {1, . . . , |X |} is X -proper if
β 0 (a) ≤ κ 0 (a),
β f (a) ≤ κ f (a)
(24)
and we say an index b ∈ {1, . . . , |Y |} is Y -proper if
ι 0 (b) ≤ α 0 (b),
17
ι f (b) ≤ α f (b).
(25)
b
Y
X
...
Y
...
X
ι0 α0 αf ιf
(a) Assume FSOC ι 0 (b) ≤ α 0 (b) ≤ α f (b) < ι f (b)
β 0(α 0 )
κ 0(α 0 )
b = κ f(α 0 )
β f(α 0 )
...
...
ι0 = α0 = αf
ιf
(b) ι 0 (b) = α 0 (b) = α f (b) < ι f (b)
Figure 3: Step 3
Y
X
κ f(α 0 (b))
b
b = κ f(α 0 (b))
...
Y
...
X
α 0 (b) ι 0 (b)
(a) b < κ f (α 0 (b))
...
...
α 0 (b) ι 0 (b)
(b) b = κ f (α 0 (b))
Figure 4: Step 4
Claim 5.16. For all a ∈ {1, . . . , |X |} and all b ∈ {1, . . . , |Y |}, a is X -proper and b is Y -proper.
Remark 5.17. First we illustrate how the Claim 5.16 implies Proposition 5.12. Suppose for contradiction that
Proposition 5.12 is false. Then there exists some a such that (a, |Y |) ∈ M and for all b we have (|X |, b) < M.
In particular, α f (b) < |X | for all b ∈ {1, . . . , |Y |}. By the claim, ι f (|Y |) ≤ α f (|Y |) < |X |, implying that no
bits from τ |X | (ψ (X |X | )) are matched with bits of ψ (Y ), a contradiction of τ (ψ (X )) ⊑ ψ (Y ).
Proof. Step 1. First, note that, as (1, 1) ∈ M, N , we have α 0 (1) = β 0 (1) = ι 0 (1) = κ 0 (1) = 1.
Step 2. Next, we show that for all a, if β 0 (a) ≤ κ 0 (a) then a is X -proper. That is, we show β f (a) ≤ κ f (a).
Suppose for contradiction we have κ f (a) < β f (a) so that β 0 (a) ≤ κ 0 (a) ≤ κ f (a) < β f (a) (see Figure 2). By
√
Lemma 5.14 part 2, we have β f (a)−β 0 (a) ≤ R and (a, b ′ ) is type-B for allb ′ ∈ [β 0 (a), β f (a)−1]. In particular,
√
(a, b ′ ) is type-B for all b ′ ∈ [κ 0 (a), κ f (a)]. As κ f (a) − κ 0 (a) < β f (a) − β 0 (a), we have κ f (a) − κ 0 (a) < R,
so we can apply Lemma 5.14 part 1 to obtain that (a, b ′ ) is type-A for some b ′ ∈ [κ 0 (a), κ f (a)]. This is a
contradiction as all such b ′ must be type-B.
Step 3. Next, we show that for all b, if ι 0 (b) ≤ α 0 (b) and α 0 (b) is X -proper, then b is Y -proper. That is, we
show ι f (b) ≤ α f (b). Suppose for contradiction that α f (b) < ι f (b) so that ι 0 (b) ≤ α 0 (b) ≤ α f (b) < ι f (b) (see
Figure 3a). We have α f − α 0 < ι f − ι 0 ≤ 2λ by Lemma 5.14 part 3. Thus, by Lemma 5.14 part 5, (α f (b), b) is
type-B. By Lemma 5.14 part 4, (i ′, b) is type-A for i ′ ∈ [ι 0 (b) + 1, ι f (b) − 1]. Since α f (b) ∈ [ι 0 (b), ι f (b) − 1]
we must have α f (b) = ι 0 (b), so ι 0 (b) = α 0 (b) = α f (b) (See Figure 3b). By definition of α f (b), we have
(α (b))
≥ b + 1. However, since we assumed α f < ι f , we have
next M (α f (b), b) = (α f (b), b + 1) so β f f
next N (α f (b), b) = (α f (b) + 1, b), so κ f (α f (b)) = b. Thus
β f (α 0 (b)) = β f (α f (b)) ≥ b + 1 > b = κ f (α f (b)) = κ f (α 0 (b)).
(26)
On the other hand, β f(α 0 (b)) ≤ κ f(α 0 (b)) by assumption that α 0 (b) is X -proper, which is a contradiction. This
covers all possible cases, completing Step 3.
Step 4. We prove that, for all b ∈ {1, . . . , |Y |}, if α 0 (b) is X -proper, then b is Y -proper. By Step 3, it
suffices to prove ι 0 (b) ≤ α 0 (b). Suppose for contradiction that α 0 (b) < ι 0 (b). Since (α 0 (b), b) ∈ M, we have
18
b ≤ β f (α 0 (b)). By assumption, α 0 (b) is X -proper, so β f (α 0 (b)) ≤ κ f (α 0 (b)), which means b ≤ κ f (α 0 (b)).
If b < κ f(α 0 ) , then we have (α 0 (b), κ f (α 0 (b))) ∈ N . However, (ι 0 (b), b) ∈ N , contradicting Fact 5.13 part
8. Thus, κ f (α 0 (b)) = b. But then (α 0 (b), b) = (α 0 (b), κ f (α 0 (b))) ∈ N with α 0 (b) < ι 0 (b), contradicting the
minimality of ι 0 (b).
Step 5. By the same argument as Step 4, we have that, for all a ∈ {1, . . . , |X |}, if β 0 (a) is Y -proper, then
a is X -proper.
Step 6. We prove by strong induction that for pairs (a, b), ordered by < M , we have a is X -proper and b
is Y -proper. Combining Steps 1,2, and 3, we have
β 0 (1) ≤ κ 0 (1),
β f (1) ≤ κ f (1),
ι 0 (1) ≤ α 0 (1),
ι f (1) ≤ α f (1),
(27)
where the first and third inequalities are actually equalities arising from Step 1, the second inequality is
established by Step 2, and the fourth is established by Step 3. Thus, 1 is both X -proper and Y -proper.
Now suppose we have some pair (a, b) ∈ M with (1, 1) < M (a, b) and (22) has been established for all
smaller pairs. If prev M (a, b) = (a − 1, b), then by the inductive hypothesis, we have b is Y -proper. However,
as (a − 1, b) ∈ M, we have (a, b − 1) < M, so we have β 0 (a) = b. Thus β 0 (a) is Y -proper, so a is X -proper
by Step 5. Similarly, if prev M (a, b) = (a, b − 1), then by the inductive hypothesis, we have a is X -proper.
Thus α 0 (b) = a is X -proper, so b is Y -proper by Step 4. This completes the proof of Claim 5.16, proving
Proposition 5.12.
The following proposition is the key result of this subsection. Following our approach, it should be
possible to prove this proposition for any choice of S 1 , . . . , S δ n , not just [λ − 1], . . . , [λ − 1]. However,
because of Lemma 5.23, which tells us that [λ − 1], . . . , [λ − 1] is the “worst” possible choice of S 1 , . . . , S δ n ,
it suffices to prove the proposition as stated.
Proposition 5.18. There exists a constant β > 0 such that for any fixed deletion pattern σ ∈ D(n, (1 − δ )n)
we have
(28)
Pr n [σ (X ) ≺(2λ , √R,[λ−1], ...,[λ−1]) Y ] < 2−βn .
X,Y ∼U ([K ] )
K
Proof. Let s = 2λ . With hindsight, let β = log
16R . It suffices to prove
i
h
X ≺(2λ , √R,[λ−1], ...,[λ−1]) Y < 2−βn .
Pr
X ∼U ([K ]δ n )
Y ∼U ([K ]n )
(29)
This suffices because for any σ , the distribution of σ (X ) for X ∼ U ([K]n ) is the same as X ∼ U ([K]δ n ).
Let X 1 , . . . , X δ n , Y1 , Y2 , . . . be independently chosen from [K]. Let X√= X 1 . . . X δ n , Y = Y1 . . . Yn , and
Y∞ = Y1Y2 . . . so that X ∼ U ([K]δ n ) and Y ∼ U ([K]n ). Construct a (2λ , R, S 1 , . . . , Sδ n )-matching of X in
Y∞ . Note that, as Y∞ is an infinite random string, M succeeds almost surely, i.e. ends in (|X |, b) for some
integer b.
Let M be the set of all reached states (a, b) in the matching, and let α f (b) = max{a : (a, b) ∈ M} and
β f (a) = max{b : (a, b) ∈ M} as in Proposition 5.12. Let A0 = 1, B 0 = 1. For i ≥ 1, set Ai = α f (Bi−1 )
and Bi = β f (Ai−1 ). As Ai − Ai−1 ≤ s for all i ≥ 1, we have Ai is well defined for i ≤ δn/s − 1. By the
definition of matching, M ′ succeeds if and only if M succeeds and the final position (|X |, β f (|X |)) satisfies
β f (|X |) < |Y |. It thus suffices to prove
Pr B ⌊δ n/s ⌋−1 < |Y | < 2−βn .
19
(30)
Y
X
B0
Ω(log K )
B1
B2
B3
...
...
A0
≤ 2λ
A1
A2
A3
Figure 5: Ai ’s and Bi ’s behavior
The key idea of this proof is that the Bi ’s grow much faster than the Ai ’s, so that the Bi ’s “run out of
indices in Y ” faster than the Ai ’s “run out of indices in X ”, even though Y is longer than X . In particular, by
definition of a matching, we have Ai+1 − Ai ≤ s = 2λ , but, on the other hand, we show that Bi+1 − Bi is, in
expectation, Ω(log K) (see Figure 5).
We have two technical lemmas. The proofs are straightforward, and we include them in Appendix C for
completeness.
Lemma 5.19. Let J be chosen uniformly from [K]. Let D be a random
√ variable that is 1 if J ∈ [λ − 1] and,
conditioned on a fixed J ≥ λ, is distributed as min(Geometric(J /K), R). Then E[D] ≥ (log K)/4.
Lemma 5.20. Let λ ′ ∈ [λ, K] and let J be chosen uniformly from {λ, λ + 1, . √
. . , λ ′ }. Let D be the random
variable that, conditioned on a fixed J , is distributed as min(Geometric(J /K), R). Then E[D] ≥ (log K)/4.
Claim 5.21. Let i ≥ 1. For any fixed A0 , . . . , Ai , B 0 , . . . , Bi , X 1 , . . . , X Ai , Y1 , . . . , YBi , we have
log K
E Bi+1 − Bi |A1 , . . . , Ai , B 1 , . . . , Bi , X 1 , . . . , X Ai , Y1 , . . . , YBi >
.
4
(31)
Proof. It suffices to prove that if we additionally condition on Ai+1 − Ai < s, then the expectation is at least
(log K)/4, and that the same is true if we condition on Ai+1 − Ai = s.
√
First, note that, for 1 ≤ b < R, we have Bi+1 = Bi + b if and only if (Ai+1 , Bi + j) is √
type-B for
j ∈ {1, . . . , Bi + b − 1} and (Ai+1 , Bi + b) is type-A. If no such b exists, we have Bi+1 − Bi = R. Thus,
conditioned on fixed A0 , . . .√, Ai+1 , B 0 , . . . , Bi , X 1 , . . . , X Ai +1 , Y1 , . . . , YBi , we have Bi+1 − Bi is distributed as
min(Geometric(X Ai +1 /K), R).
Suppose we condition on Ai+1 − Ai = s. This is equivalent to saying (Ai + ℓ, Bi ) is type-A for
ℓ = 1, . . . , s − 1. However, this assertion depends only on X Ai , X Ai +1 , . . . , X Ai +s−1 , which are independent of
X Ai +1 , so we have X Ai +1 is still distributed uniformly on√[K]. If X Ai +1 ∈ [λ − 1], then Bi+1 − Bi = 1, otherwise
Bi+1 − Bi is distributed as min(Geometric(X Ai +1 /K), R), so by Lemma 5.19 on D = Bi+1 − Bi , we have
E[Bi+1 − Bi |Ai+1 − Ai = s, A1 , . . . , Ai , B 1 , . . . , Bi , X 1 , . . . , X Ai , Y1 , . . . , YBi ] >
log K
.
4
(32)
Suppose we condition on Ai+1 − Ai = s ′ < s. This is equivalent to saying (Ai + ℓ, Bi ) is type-A for
ℓ = 1, . . . , s ′ − 1 and (Ai + s ′, Bi ) is type-B. This implies X Ai +1 ≥ λ (i.e. X Ai +1 < [λ − 1]) and X Ai +1 < YBi ,
Since YBi is fixed and, without any conditioning, X Ai +1 is distributed uniformly in [K], we have X Ai +1
is distributed uniformly on√[λ, . . . , YBi ]. By the previous argument, we have Bi+1 − Bi is distributed as
min(Geometric(X Ai +1 /K), R), so by Lemma 5.20 on λ ′ = YBi and D = Bi+1 − Bi , we have
E[Bi+1 − Bi |Ai+1 − Ai = s ′ < s, A1 , . . . , Ai , B 1 , . . . , Bi , X 1 , . . . , X Ai , Y1 , . . . , YBi ] >
log K
.
4
(33)
20
Corollary 5.22. We have, for any fixed B 1 , . . . , Bi ,
E[Bi+1 − Bi |B 1, . . . , Bi ] >
1
log K.
4
(34)
1
4
log K so that
δn
1
α(k − 1) > log K · λ − 1 > 2n.
4
2
Continuing the proof of Proposition 5.18, let α =
(35)
Define the random variable Bi′ := Bi − i · α. As Bi′ and Bi uniquely determine each other, we have, by
Corollary 5.22,
′
E[Bi+1
− Bi′ |B 1′ , . . . , Bi′] = E[Bi+1 − Bi |B 1, . . . , Bi ] − α > 0.
(36)
√
′
− Bi′ | < R, so by Azuma’s Inequality (Lemma 4.2), we have
Thus, Bi′ form a submartingale with |Bi+1
!
2
2
(α(k
−
1)/2)
n log K
α (k − 1)
α(k
−
1)
′
′
≤ exp −
< exp −
. (37)
= exp −
Pr Bk − B 1 ≤ −
√
2
8R
16R
2(k − 1) · ( R)2
Combining with (35) gives
Pr[Bk ≤ n] ≤
Pr[Bk′
− B 1′
≤ −n] ≤ Pr
Bk′
We conclude
Pr
[X
X ∼U ([K ]δ n )
Y ∼U ([K ]n )
≺(√R,2λ ,[λ−1], ...,[λ−1])
log K
) < 2−
As exp(− n 16R
n log K
16R
Y ] ≤ Pr[mδ n
− B 1′
α(k − 1)
≤−
2
n log K
< exp −
.
16R
(38)
(39)
n log K
≤ n] ≤ Pr[Bk ≤ n] < exp −
.
16R
, we have (29) is true, completing the proof of Proposition 5.18.
To conclude this section, we formalize the intuition that the “worst possible deletion pattern”, that is,
the deletion pattern that makes decoding most difficult in some sense, is the deletion pattern such that each
(λ − 1)-admissible inner deletion pattern corrupts д1 , . . . , дλ−1 . This fact is intuitive, as it is hard, for example,
to find д1 as a subsequence of a random ψ (Y ) because д1 has many runs. Thus if our deletion pattern τ
corrupts the inner codewords with the most runs, there is a greater probability that over random X , Y we have
τ (ψ (X )) is a subsequence of ψ (Y ). However, note that, to make a clean assertion, we continue to argue using
the matching relation rather than subsequence relation.
Lemma 5.23. Let r, s ∈ N and S 1 , . . . , S δ n be subsets of [K] of size exactly λ − 1. Then for all Y ∈ [K]n we
have
o
n
o
n
(40)
# X ∈ [K]δ n : X ≺(2λ , √R,S 1, ...,S δ n ) Y ≤ # X ∈ [K]δ n : X ≺(2λ , √R,[λ−1], ...,[λ−1]) Y
Proof. We start with a claim.
Claim 5.24. For every set A ⊆ [K] with size exactly λ − 1, there exists a bijection h A : [K] → [K] such that
h A (x) ∈ [λ − 1] for x ∈ A and h A (x) ≥ x for x < A.
Proof. Pair each element x ∈ A \ [λ − 1] with an element y ∈ [λ − 1] \ A arbitrarily and for each pair (x, y)
set h A (x) = y, h A (y) = x. Note this always gives x > y. This is possible as A \ [λ − 1] and [λ − 1] \ A have
the same size. Then set h A (x) = x for all other x. It is easy to check this function satisfies h A (x) ≥ x for
x < A.
21
Fix Y ∈ [K]n . For all i, let hi : S i → [λ − 1] be a bijection such that hi (x) ∈ [λ − 1] for x ∈ S i and
f (x) ≥ x for all other x. This exists by Claim 5.24. Let h : [K]δ n → [K]δ n be such that h(X 1 , . . . , X δ n ) =
h 1 (X 1 )h 2 (X 2 ) · · · hδ n (X δ n ). Since each of h 1 , . . . , hδ n are bijections, h is a bijection as well.
Let X be such that X ≺(2λ , √R,S 1, ...,S δ n ) Y . We claim h(X ) ≺(2λ , √R,[λ−1], ...,[λ−1]) Y . Let M be a
√
√
(2λ , R, S 1 , . . . , S δ n )-matching of X in Y and let M ′ be a (2λ , R, [λ − 1], . . . , [λ − 1])-matching of h(X ) in
Y . We know M is a successful matching, and we wish to show M ′ is a successful matching.
If (a, b) is type-A with respect to X , Y , S 1 , . . . , S |X | , then Yb ≤ X a or X a ∈ S a , so, by definition of h,
either h(X a ) ∈ [λ − 1] or Yb ≤ X a ≤ h(X a ), so (a, b) is type-A with respect to h(X ), Y , [λ − 1], . . . , [λ − 1].
Let ℓ ≥ 0 be an integer and let (a, b) ∈ M and (a ′, b ′ ) ∈ M ′ be the state of the matchings after the ℓth step
of the matchings of X in Y and h(X ) in Y , respectively. It is easy to see by induction that a ≤ a ′ and b ′ ≤ b;
if a < a ′ then after one move we still have a ≤ a ′, and if a = a ′ and Ms next move is an A-move, then
M ′s next move must also be an A-move. Since M succeeds, we have (|X |, b) ∈ M for some b < |Y |, so we
conclude that (|X |, b ′ ) ∈ M ′ for some b ′ ≤ b < |Y |. Thus M ′ succeeds as well.
Since h is a bijection such that
o
n
o
n
⊆ X ∈ [K]δ n : X ≺(2λ , √R,[λ−1], ...,[λ−1]) Y ,
(41)
h X ∈ [K]δ n : X ≺(2λ , √R,S 1, ...,S δ n ) Y
we have
o
n
o
n
# X ∈ [K]δ n : X ≺(2λ , √R,S 1, ...,S δ n ) Y ≤ # X ∈ [K]δ n : X ≺(2λ , √R,[λ−1], ...,[λ−1]) Y
as desired.
(42)
5.4 Technical combinatorial lemmas
Lemma 5.25. Let n be a positive integer and suppose we have 0 < β < 1. Suppose that γ = β/3, M = 2γ n ,
and ϵ ≥ 2−(γ −o(1))n/2 . Suppose G is a graph on N = K (1−o(1))n vertices such that each vertex has degree
at most d = N r for some r = r (n) = 2−(β −o(1))n . Suppose Cout is chosen as a random subset of V (G)
by including each vertex of V (G) in Cout with probability M/N . Then for sufficiently large n, we have
PrCout [|E(G ↾ Cout )| > ϵM] < 2−ω(n) .
Remark 5.26. Essentially this lemma is saying that when the edge density is extremely small, around 2−βn ,
then for all but an extremely small set of choices for Cout , Cout is extremely sparse.
Proof. Let E = E(G). We know E satisfies |E | ≤ dN /2 < N 2r .
Enumerate the edges 1, . . . , |E | arbitrarily. Let Y1 , . . . , Y |E | be Bernoulli random variables denoting
whether the ith edge is in Cout . Let Z = Y1 + · · · + Y |E | . We would like to show
Pr [Z > ϵM] < 2−n .
Cout
(43)
We do this by computing a sufficiently large moment of Z and using
Pr [Z > ϵM] ≤
Cout
E[Z k ]
.
(ϵM)k
(44)
For a tuple of (not necessarily distinct) edges (e1 , . . . , ek ), denote V (e1 , . . . , ek ) to be the set of vertices
on at least one of the edges e1 , . . . , ek . Alternatively, we say V (e1 , . . . , ek ) is the set of vertices covered by
edges e1 , . . . , ek . Note that for all e1 , . . . , ek , we have 2 ≤ V (e1 , . . . , ek ) ≤ 2k. Let Pk, ℓ = #{(e1 , . . . , ek ) :
|V (e1 , . . . , ek )| = ℓ} denote the number of ordered tuples of edges from E that cover exactly ℓ vertices.
22
Claim.
Pk, ℓ < N ℓ · r ℓ/2 · ℓ 2k+ℓ−1 · k k
(45)
Proof. We bound Pk, ℓ by first bounding the number of sets (unordered tuples) of edges covering exactly ℓ
vertices and then multiply by k!. We first compute the number of ways to construct a forest of trees covering
ℓ vertices using only edges from E. We do this by casework on the number of connected components. Let
1 ≤ c ≤ ⌊ℓ/2⌋ be an integer. To construct a forest with c connected components, we first choose c disjoint
edges e1 , . . . , ec . This can be done in at most |E | c ways. We have |V (e1 , . . . , ec )| = 2c. For c + 1 ≤ i ≤ ℓ − c,
choose a vertex v ∈ V (e1 , . . . , ei−1 ) and an edge vw such that w < V (e1 , . . . , ei−1 ). By construction, for
c + 1 ≤ i ≤ ℓ − c, we have |V (e1 , . . . , ei )| = i + c. Thus the i th edge can be added in at most (i + c − 1)d ways.
The i = ℓ − c th edge completes a forest covering ℓ vertices. Recalling that |E | < N 2r and d = N r , we have
that the total number of ways to construct a forest in this fashion is at most
|E | c · 2cd · (2c + 1)d · · · (ℓ − 1)d ≤ |E | c · d ℓ−2c · ℓ ℓ−2c < N ℓ · r ℓ−c · ℓ ℓ−2c .
(46)
We have used ℓ − c edges thus far. There are k − ℓ + c remaining edges. As |V (e1 , . . . , ek )| = ℓ, these
remaining edges must connect one of the 2ℓ pairs of vertices in V (e1 , . . . , ek−ℓ+c ). As c ≤ ⌊ℓ/2⌋ < ℓ, we
k−ℓ+c
have k − ℓ + c < k. The remaining edges can thus be chosen in at most 2ℓ
< ℓ 2k ways. Using (46) and
multiplying by k! we have
ℓ/2
Õ
N ℓ · r ℓ−c · ℓ ℓ−2c · ℓ 2k < k k · (ℓ/2) · N ℓ · r ℓ/2 · ℓ ℓ−2 · ℓ 2k < N ℓ · r ℓ/2 · ℓ 2k+ℓ−1 · k k . (47)
c=1
√
Note that M r ≪ 1. With this claim, we have
Pk, ℓ < k! ·
E[Z k ] = E[(Y1 + · · · + Y |E | )k ]
Õ
E[Y1 · · · Yk ]
=
(e1, ...,ek )∈ {1, ..., |E | }k
=
2k
Õ
Õ
ℓ=2 |V (e1, ...,ek ) |=ℓ
=
2k
Õ
Õ
ℓ=2 |V (e1, ...,ek ) |=ℓ
=
<
<
ℓ
2k
Õ
M
ℓ=2
2k
Õ
ℓ=2
2k
Õ
ℓ=2
N
M
N
ℓ
E[Ye1 · · · Yek ]
M
N
ℓ
· Pk, ℓ
· N ℓ · r ℓ/2 · ℓ 2k+ℓ−1 · k k
ℓ 2k+ℓ−1 · k k
< 2k · (2k)4k−1 · k k = 24k · k 5k .
Finally, choosing k = log n, we have
as desired.
E[Z k ]
24k k 5k
Pr[Z ≥ ϵM] ≤
<
<
(ϵM)k
ϵk Mk
16k 5
ϵ2γ n
k
≤
(48)
16k 5
2γ n/2
k
< 2−ω(n)
(49)
23
Lemma 5.27. Let n be a positive integer, and suppose 0 < β < 1 and γ = β/4. Suppose M = 2γ n and
ϵ ≥ 2−γ n/2 . Suppose G is a directed graph on N = K n(1−o(1)) vertices such that each vertex has outdegree at
most N r for some r = 2−(β −o(1))n . Choose a subset Cout ⊆ V (G) at random by including each vertex of V (G)
in Cout with probability M/N so that E[|Cout |] = M. Then, for sufficiently large n, we have
h n
o
i
−−→
(50)
Pr # X ∈ Cout : ∃Y ∈ Cout s. t. YX ∈ E(G) > ϵ |Cout | < 2−ω(n) .
Cout
Proof. As, by assumption, Y has outdegree at most N · r for all Y ∈ V (G), the average indegree of G is at
most N · r . Thus at most ϵ/8 fraction of all words in V (G) have indegree larger than ϵ8 · N · r . Call this set
of vertices W . We have
8
ϵ
E # X ∈ Cout : indegG (X ) > · N · r
= E[|Cout ∩ W |] ≤ M.
(51)
Cout
ϵ
8
Since |Cout ∩ W | is the sum of i.i.d Bernoulli random variables, we have by Lemma 4.1
h
1 ϵ
ϵ
ϵ i
8
Pr # X ∈ Cout : indegG (X ) > · N · r > M = Pr |Cout ∩ W | > M < e − 2 · 8 ·M < 2−ω(n) .
Cout
ϵ
4
4
(52)
−−→
−−→
Now consider the undirected graph H on V (G) such that XY ∈ E(H ) if XY ∈ E(G) and Y < W or YX ∈ E(G)
and X < W . For every vertex v, the in-edges of v in G correspond to edges in H only if the indegree is at
most ϵ8 · N · r , and the outdegree of v in G is always at most N r . Therefore the degree of every vertex in H
is at most r ′ N where r ′ = ϵ8 + 1 · r < 2−(β −γ −o(1))n .
We are including each vertex of V (G) (and thus each vertex of V (H )) in Cout independently with
probability M/N . Let ϵ ′ = ϵ/4, β ′ = 34 β, γ ′ = γ so that γ ′ = β ′/3, ϵ ≥ 2−(γ +o(1))n/2 , M = 2γ n , and
′
r ′ = 2−(β −o(1))n . By Lemma 5.25 for β ′, γ ′, M, ϵ ′, and r ′, and H , we have for sufficiently large n that
ϵ
Pr E(H ↾ Cout ) > M < 2−ω(n) .
(53)
Cout
4
Also, by Chernoff bounds, we have
3
3
−Ω(M )
=⇒ Pr |Cout | < M < 2−ω(n) .
(54)
Pr |Cout | < M < 2
Cout
Cout
4
4
Note that if the number of X such that indegG↾Cout (X ) > 0 at least ϵ |Cout |, that is,
n
o
−−→
# X ∈ Cout : ∃Y ∈ Cout s. t. YX ∈ E(G) > ϵ |Cout |,
(55)
then one of the following must be false. .
1. |Cout | ≥ 34 M
2. #{X ∈ Cout : degH ↾Cout (X ) > 0} ≤ ϵ2 M
3. |Cout ∩ W | ≤ ϵ4 M
Indeed, suppose to the contrary all of these were true. The number of vertices in Cout ∩ W with positive
indegree in G ↾ Cout is at most |Cout ∩W | ≤ ϵ4 M. If a vertex X ∈ Cout \W has positive indegree in G ↾ Cout
−−→
then there exists Y ∈ Cout such that YX ∈ E(G), so XY ∈ E(H ) by definition of H and thus degH ↾Cout (X ) > 0.
Thus the number of vertices in Cout \W with positive indegree is at most ϵ2 M by property 2. Hence the total
number of vertices in Cout with positive indegree in G ↾ Cout is at most 34 ϵM ≤ ϵ |Cout |.
Each of items 1, 2, and 3 is false with probability 2−ω(n) by (52), (53), and (54), so the probability any
of them occur is at most 2−ω(n) , as desired.
24
5.5 Proof of Construction (Theorem 2.2)
Proof of Theorem 2.2. We would like to show there exists a code C and an ϵ = o N (1) such that, for any
deletion pattern τ deleting pN bits,
#{x ∈ C : ∃y ∈ C s. t. x , y and τ (x) ≤ y} ≤ ϵ |C |.
For Y ∈ [K]n , define
f (Y ) =
Pr
Z ∼U ([K ]δ n )
h
i
Z ≺(2λ , √R,[λ−1], ...,[λ−1]) Y .
(56)
(57)
Let σ ∈ D(n, (1 − δ )n) be a deletion pattern for our outer code [K]n . Let S 1 , . . . , S δ n be subsets of [K]
−−→
∗
be the graph on the vertex set [K]n such that YX is an edge if and only
of size exactly λ − 1. Let G σ,S
1, ...,S δ n
if σ (X ) ≺(2λ , √R,S 1, ...,S δ n ) Y . Note that for all σ ∈ D(n, (1 − δ )n) and all Y ∈ [K]n we have, by Lemma 5.23,
#{X ∈ [K]n : σ (X ) ≺(2λ , √R,S 1, ...,S δ n ) Y } = K (1−δ )n · #{Z ∈ [K]δ n : Z ≺(2λ , √R,S 1, ...,S δ n ) Y }
≤ K (1−δ )n · #{Z ∈ [K]δ n : Z ≺(2λ , √R,[λ−1], ...,[λ−1]) Y }
= K n · f (Y ).
(58)
∗
.
In the graph language, this means every Y ∈ [K]n has outdegree at most K n · f (Y ) in every G σ,S
1, ...,S δ n
We remove all Y with large f (Y ). Let β = log K/16R. By Proposition 5.18, we know the average
value of f (Y ) is 2−βn . Hence at most K n /2 such Y satisfy f (Y ) ≥ 2 · 2−βn . These are the easily disguised
candidate codewords mentioned in §2. There is thus a set W ⊆ [K]n such that |W | = K n /2 and each
Y ∈ W satisfies f (Y ) < 2 · 2−βn = 2−(β −o(1))n . For all σ ∈ D(n, (1 − δ )n), S 1 , . . . , S δ n , consider the subgraph
∗
↾ W . By construction, for all σ , S 1 , . . . , S δ n , the outdegree of every vertex of
G σ,S 1, ...,S δ n := G σ,S
1, ...,S δ n
n
G σ,S 1, ...,S δ n is at most (K /2) · r for some r = 2−(β −o(1))n .
Let γ = β/4. Let M = 2γ n and ϵ = 2−γ n/2 . Choose a subset Cout ⊆ W by including each vertex of W in
Cout independently with probability M/|W | so that E[|Cout |] = M.
By Lemma 5.27 for N = K n /2, β, γ , M, ϵ, r, and G σ,S 1, ...,S δ n , we have that, for all σ and sufficiently large
n,
h
i
−−→
Pr #{X ∈ Cout : ∃Y ∈ Cout s. t. YX ∈ E(G σ,S 1, ...,S δ n )} > ϵ |Cout | < 2−ω(n) .
(59)
Cout
This is equivalently
h
i
Pr #{X ∈ Cout : ∃Y ∈ Cout s. t. σ (X ) ≺(2λ , √R,S 1, ...,S δ n ) Y } > ϵ |Cout | < 2−ω(n) .
Cout
(60)
K δn
Union bounding over the at-most-2n possible values of σ and λ−1
values of S 1 , . . . , Sδ n gives
h
i
Pr ∃σ , S 1 , . . . , S δ n s. t. # X ∈ Cout : ∃Y ∈ Cout s. t. σ (X ) ≺(2λ , √R,S 1, ...,S σ n ) Y > ϵ |Cout |
Cout
K
< 2 ·
λ−1
n
δn
· 2−ω(n)
< 2n · 2nδ (λ−1) log K · 2−ω(n) = 2−ω(n) .
(61)
Additionally, with probability approaching 1, |Cout | ≥ 34 M. Thus, there exists a Cout with |Cout | ≥ 43 M
[K ]
such that the following holds for all σ ∈ D(n, (1 − δ )n) and all S 1 , . . . , S δ n ∈ λ−1
: at most ϵ |Cout | of the
√
λ
codewords X are (2 , R, S 1 , . . . , Sδ n ) matchable in some other codeword Y ∈ Cout .
25
By Lemma 5.7, there exists σ ∈ D(n, (1 − δ )n) and τ ′ ∈ D(δnL) such that the following holds: If we
write τ ′ = τ1′ ⌢ · · · ⌢ τδ′ n then each τi′ is (λ − 1)-admissible and preserves all дj for all j not in some
size-λ − 1 set Si , and furthermore we have τ ′(ψ (σ (X ))) ⊑ √τ (ψ (X )) for all X ∈ [K]n . By choice of Cout , for
at least (1 − ϵ)|Cout | choices of X ∈ Cout , σ (X ) is not (2λ , R, S 1 , . . . , S δ n ) matchable in all Y ∈ Cout . Thus,
for these X , we have τ ′(ψ (σ (X ))) @ ψ (Y ) for all Y ∈ Cout by Lemma 5.12. Since τ ′(ψ (σ (X ))) ⊑ τ (ψ (X ))
for all X ∈ [K]n , we have for these (1 − ϵ)|Cout | choices of X that τ (ψ (X )) @ ψ (Y ) for all Y ∈ Cout . Thus,
choosing C = ψ (Cout ) gives our desired average deletion code.
Recall λ is the smallest integer such that (1 +p)/2 < 1 − 2−λ and δ = 1 − 2−λ −p so that 2λ = Θ(1/(1 −p))
λ+5
2
and δ = Θ(1 − p). Recall K = 2 ⌈2 /δ ⌉ and R = 4K 4 so that K = 2Θ(1/(1−p) ) . The rate of the outer code is
log 2γ n /n log K and the rate of the inner code is log K/L. The total rate is thus
R =
2
log K
−2Θ(1/(1−p) )
.
=
2
48R · 2R K
This completes the proof of Theorem 2.2.
(62)
6 Relating zero-rate threshold of online and adversarial deletions
6.1 Proof that p0(adv) =
1
2
if and only if p0(on,d) =
1
2
Recall that the zero-rate threshold for adversarial deletions, p0(adv) , is the supremum of all p such that there
exists a family of code C ⊆ {0, 1}n with |C | = 2Ω(n) satisfying LCS(C) > (1 − p)n. Additionally, recall that
the deterministic zero-rate threshold for online deletions, p0(on,d) , is the supremum of p such that there exist
families of deterministic codes with rate bounded away from 0 that can correct against pn online deletions
in the average case (when a uniformly random codeword is transmitted). That is, when a uniformly random
codeword is transmitted, the probability of a decoding error, over the choice of the codeword and possibly the
randomness of the online channel strategy, is o(1) in the block length of the code. Since an online strategy
can guess the minority bit in the codeword and delete all its occurrences, adversary can always delete the
majority bit, we trivially have
1
(63)
p0(adv) ≤ p0(on,d) ≤ .
2
(adv)
is based on the code construction in [3] and
Recall that the currently best known lower bound on p0
√
(adv)
implies p0
≥ 2 − 1. An outstanding question in this area is whether p0(adv) = 1/2. Our main result in
this section ties this question to the corresponding question for online deletions. Namely, we have
Theorem 6.1. If p0(adv) =
1
2
if and only if p0(on,d) = 21 .
By virtue of (63), Theorem 6.1 follows if we show that p0(adv) < 21 implies p0(on,d) < 21 . Our result is
a negative result against deterministic codes, and our argument does not immediately extend to stochastic
codes. While we expect such a result to exist, we elaborate §6.2 on the current limitations of our approach
to proving a negative result against stochastic codes.
The idea of the proof of Theorem 6.1 is based on a suitable adaptation of the “wait-push” idea of [1].
We wait qn bits until we know the codeword, then push so that it is confused with another codeword. In
the first wait phase, we delete the minority bit for a budget of qn/2 deletions, and while we push we need at
most (1 − q)np0(adv) deletions. If we assume that p0(adv) < 21 , we end up with a total of αn deletions for some
α < 12 . The exact details are more subtle, but this is the rough idea.
From the definition of p0(adv) , we have the following fact.
26
Fact 6.2. For any R > 0 and p > p0(adv) , there exists n 0 such that for all n ≥ n 0 , for any C ⊆ {0, 1}n with
|C | ≥ 2 Rn , there exist two strings x, y ∈ C such that LCS(x, y) > (1 − p)n.
Corollary 6.3. For any R > 0 and p > p0(adv) , there exists n 0 such that for all n ≥ n 0 , for any C ⊆ {0, 1}n
with |C | ≥ 2 Rn , there exists C ′ ⊆ C such that |C ′ | ≥ |C | − 2 Rn/2 , and all elements of C ′ are confusable with
another element of C. That is, for all x ∈ C ′ there exists y ∈ C such that x , y and LCS(x, y) > (1 − p)n.
Proof. We can apply Fact 6.2 for R ′ = R/2. Start with C ′ = ∅. While |C \ C ′ | > 2 Rn/2 , find an x that is
confusable with some y in C and add it to C ′. This is possible by Fact 6.2.
The following is the precise statement of our result relating the zero-rate thresholds for online and
adversarial deletions.
Theorem 6.4. We have
p0(on,d) ≤
1
3 − 2p0(adv)
(64)
.
In particular, if p0(adv) < 21 , then p0(on,d) < 12 .
Remark 6.5. If it is the case that the bound in [3] is tight and p0(adv) =
gives p0(on,d) ≤ 0.4605.
√
2 − 1 ≈ 0.414, then Theorem 6.4
The rest of this section is devoted to proving Theorem 6.4. We start with the following helpful definition.
Definition 6.6. Given a code C and a codeword x ∈ C passing through an online deletion channel, define the
wait phase and the push phase for a codeword as follows. Let q x n be the smallest index such that x 1 . . . x qx n
uniquely determines the codeword. The wait phase refers to the time until Adv receives bit q x n, but not
acted on it (chosen whether to transmit or delete it). Thus in the wait phase Adv receives bits 1, . . . , q x n and
has acted on bits 1, . . . , q x n − 1. The push phase is the time after the end of the wait phase. We say the wait
length of x is q x n and the push length is (1 − q x )n.
For a codeword x, let q x denote the relative wait length (wait length divided by n), and let r x,b denote
number b-bits that appear in x 1x 2 . . . x qx n , so that r x,0 +r x,1 = q x . Let bx ∈ {0, 1} denote the bit that appears
more frequently in x 1 . . . x qx n (break ties arbitrarily).
Proof of Theorem 6.4. Since p0(on,d) ≤ 12 , (64) holds if p0(adv) = 21 , so assume p0(adv) < 12 .
Let p be such that
1
1
.
>p >
(adv)
2
3 − 2p
(65)
0
We show that, for large enough n, when Adv gets pn online deletions, he can, with constant probability
independent of n, guarantee that, if the encoder sends x ∈ C, the decoder receives a string s that is a substring
of y ∈ C where x , y.
Call a pair of codewords x, y online-confusable if q x = qy ≤ 1 − p, r x,0 = ry,0 , bx = by , and
LCS(x[q x n :], y[q x n :]) > (1 − q x )(1 − p0(adv) )n.
(66)
For an online-confusable pair x, y, let s ∗ (x, y) denote a common subsequence of x[q x n :] and y[q x n :] with
(adv)
)n, and let s(x, y) denote the string (bx )r x,b n s ∗ (x, y). Note that s(x, y) is a
length at least (1 − q x )(1 − p0
27
common substring of x and y and, by choice of p in (65) has length at least
|s(x, y)| = r x,b n + (1 − q x )(1 − p0(adv) )n
qx n
≥
+ (1 − q x )(1 − p0(adv) )n
2
1
(adv)
(adv)
− p0
qx n
= 1 − p0
−
2
1
(adv)
(adv)
− p0
(1 − p) n
≥ 1 − p0
−
2
> (1 − p)n ,
(67)
where the last step follows from the bound on p in (65).
We now claim we can find many disjoint online-confusable pairs. For fixed q ∗ , r ∗ , b ∗ , let
C(q ∗ , r ∗ , b ∗ ) := {x ∈ C : q x = q ∗, r x,b ∗ = r ∗ , bx = b ∗ }.
(68)
∗
By Lemma 6.2 on the set {x[q ∗n :] : x ∈ C(q ∗, r ∗ , b ∗ )}, among any 2 Rn/2 (in fact, any 2 R(1−q )n/2 ) codewords
in C(q ∗ , r ∗ , b ∗ ), there exists some two codewords x and y in C(q ∗ , r ∗ , b ∗ ) such that LCS(x[q ∗n :], y[q ∗n :
]) > (1 − q ∗ )(1 − p0(adv) )n. Thus, among any 2 Rn/2 codewords in C(q ∗ , r ∗ , b ∗ ), there are some two that are
online-confusable. Thus for every q ∗ , r ∗ , b ∗ , we can construct disjoint pairs of codewords (x, y) that are
online-confusable, so that all but 2 Rn/2 of the elements of C(q ∗ , r ∗ , b ∗ ) are paired. For (x, y) in a pair, call x
and y partners. Thus, all but 2n 2 · 2 Rn/2 of the elements of C have a partner. For sufficiently large n, this
means that at least 0.99|C | elements have a partner.
The online channel choose to run one of the following strategies, each with probability 12 .
1. Strategy 1
(a) Choose a bit b uniformly from {0, 1}.
(b) During the wait phase, delete every x i = 1 − b. After the wait phase, we know the codeword x
and have transmitted r x,b n copies of the bit b.
(c) If x < C(q x , r x,b , b) (i.e. x ∈ C(q x , r x,1−b , 1 − b)) or x has no partner codeword, transmit the
remaining bits (i.e. give up).
∗ by the canonical common subsequence
(d) If x ∈ C(q x , r x,b , b) and x has a partner y, then let s xy
∗ | ≥ (1 − q )(1 − p (adv) )n. Since we know that the remaining bits sent by the
of x ∗ , y ∗ with |s xy
x
0
encoder is x ∗ , we can delete bits so that the decoder receives s ∗ .
(e) If the online channel reaches pn deletions at any point, transmit the remaining bits.
2. Strategy 2
(a) Transmit the first (1 − p)n bits.
(b) Delete the last pn bits.
The idea behind the strategy is that either the wait phase is short, in which case strategy 1 requires
fewer than n/2 deletions, or the wait phase is long in which case strategy 2 gives less than n/2 deletions.
In particular with probability 12 , we employ a strategy (either strategy 1 or strategy 2) that, with constant
probability, needs at most min(q x n, q x n/2 + (1 − q x )p0(adv) n) deletions, and we check this is at most pn.
28
Suppose for codeword x ∈ C, the wait length is q x n. Condition on the fact that x has a partner y with
∗ of x ∗ and y ∗ with |s ∗ | ≥ (1 − q )(1 − p (adv) )n. By the
q x = qy and canonical common subsequence s xy
x
xy
0
above argument, this assumption holds with probability more than 0.99.
If q x > 1 − p, then knowing x 1 . . . x (1−p)n does not uniquely determine the codeword x, so there exists
another codeword y ∈ C with the same prefix y1 . . . y(1−p)n . Then when the encoder sends each of x and
y, the online channel applying strategy 2 transmits the same string x 1 . . . x (1−p)n = y1 . . . y(1−p)n . Since we
assume the choice of codeword in C is uniformly at random, x and y are equally likely to be transmitted, so
the decoder fails in this case with probability at least 12 .
If q x ≤ 1 − p, then with probability 12 over the choice of b, we have x ∈ C(q x , r x,b , b), in which case
strategy 1 reaches step (d). Since x ∈ C(q x , r x,b , b), we have x 1 . . . x qx n has majority bit b, so step (b) deletes
∗ | ≥ (1 − q )(1 − p (adv) )n, step (d) deletes |x ∗ | − |s ∗ | ≤ (1 − q )p (adv) n bits.
at most q x n/2 bits. Since |s xy
x
x 0
xy
0
The total number of bits required by strategy 1 is thus
qx n
1
1
+ (1 − q x )p0(adv) n = p0(adv) +
− p0(adv) q x n < p0(adv) +
− p0(adv) (1 − p) n < pn
(69)
2
2
2
∗ for both x and y being sent
by our choice of p. Because x and y are partners, the encoder receives b qx n s xy
through the online channel. Again, as each of x and y is equally likely, the decoder fails with probability at
least 21 .
We have shown that with probability at least 0.99 over the choice of codeword, either strategy 1 chooses
the correct value of b with probability 12 , and thus causes decoding error with probability at least 14 , or strategy
2 causes decoding error with probability at least 12 . Adv chooses the “correct” strategy with probability 21 ,
1
so with total probably at least 0.99 · 21 · min( 14 , 21 ) > 10
, there is a decoding error, as desired.
6.2 Extending to stochastic codes
It is natural to wonder whether the above conditional negative result on deterministic coding against online
deletions with deletion fraction approaching 1/2 can be extended to stochastic codes. Indeed, [1] and [8],
provide tight upper bounds for the capacities of the online erasure and substitution channels, respectively.
Both works provide online channels that cause constant probability decoding errors for both deterministic
and stochastic codes. Both use a two-stage strategy. In [1], the authors use a “wait-push” strategy to construct
an online erasure adversary that, for some transmitted word x, erases none of the first qn bits, for suitable
0 < q < 1, then suitably chooses a random codeword x ′ , and erases all bits x i for which i > qn and x i , x i′.
In [8], the authors use a “babble-push” strategy to construct an online substitution adversary that, for some
transmitted word x, randomly flips q ′n of the first qn bits for some 0 < q ′ < q < 1, then suitably chooses a
random codeword x ′, and with probability 1/2 flips each bit x i for which i > qn and x i , x i′ .
A key difference between our wait-push strategy and the strategies of [1] and [8] is that they do not require
the online adversary to know the transmitted word exactly - they simply need a guarantee that, with constant
probability, the codeword x ′ that they are “pushing to” (i) has low Hamming distance from the transmitted
word and (ii) is encoded by a different message. However, our wait-push strategy requires us to wait until
we know the transmitted codeword exactly, so that we can identify the LCS between the transmitted word x
and the word x ′ we are pushing to. In the case that the length of the wait phase, qn, is large (say, close to
n), our online channel against deterministic codes can still guarantee constant probability of error by, with
probability 1/2, choosing a different strategy that simply deletes the last pn bits. However, this patch of
choosing a different strategy doesn’t work in the stochastic case, since a stochastic code can have a message
be identified using a small fraction of the bits while individual codewords need many bits to be identified.
In such a code, the wait-push strategy must wait a long time until it knows the codeword exactly and is
thus expensive, while the “delete the last pn bits” strategy obscures only the codeword but not the message.
29
Stochastic codes with this description exist in the literature. One example is the mega-subcode construction
of [4], which encodes a message independently into t different stochastic codes and concatenates the results:
in this construction, the message can be identified in the first n/t bits, but an online channel must see at least
n(t − 1)/t bits to identify the codeword.
In spite of these obstacles to obtaining a negative result for stochastic codes, our result for the deterministic
case is still strong evidence that the zero rate thresholds for online and adversarial deletions are either both
or neither 1/2. In the oblivious error model, we simultaneously constructed deterministic codes under an
average error criterion and stochastic codes under a worst case error criterion for p approaching 1, suggesting
that for the online setting the zero-rate threshold and deterministic zero-rate thresholds are the same. As a
caution, we note that this is not always the case [9]: for online erasure channels with one bit of delay, the
capacity of the channel for deterministic codes under average error criterion is 1 − 2p, while for stochastic
codes it is 1 − p.
7 Conclusion
7.1 Decoding deletions vs. insertions and deletions
A lemma due to Levenshtein [24] states that a code C can decode against pn adversarial deletions if and only
if it can decode against pn adversarial insertions and deletions.
In the oblivious error model, decoding insertions vs insertions and deletions are not the same. In
decoding against oblivious deletions, we not only need to worry about whether two codewords are confusable
by insertions and deletions, but also the number of ways in which they are confusable. In particular, when
the fraction of oblivious deletions is greater than 12 , it is inevitable that there are codewords c and c ′ that are
confusable, but we need the number of deletion patterns τ that confuse them, i.e. with τ (c) ⊑ c ′ or τ (c ′ ) ⊑ c,
to be small.
It may be possible to extend our oblivious deletion construction to oblivious insertions and deletions by
tweaking the parameters. This was the case for bridging the gap between efficiently decoding adversarial
deletions [16] and efficiently decoding adversarial insertions and deletions [13]. Our analysis for deletions
analyzes runs of 0s and 1s, but perhaps it would be possible to analyze insertions and deletions by relaxing
the definition of a run to be “a subword with at least 0.9 density of 0s or 1s.” For now, however, we leave the
case of correcting a fraction of oblivious insertions and deletions approaching 1 as an open problem.
7.2 Open problems
These remain a number of open questions concerning deletion codes. Our work brings to the fore a few
more.
(adv)
1. What is the zero-rate threshold p0
for adversarial deletions?
2. We showed the existence of codes that decode against pn oblivious deletions for any p < 1. Can we
modify the proof to show existence of codes decoding against a combination of pn oblivious insertions
and deletions for every p < 1?
3. Can we adapt our existence proof of codes for oblivious deletions to obtain explicit codes that are
constructable, encodable, and decodable in polynomial time?
4. For erasures, the capacity of the random error channel and the capacity of the oblivious error channel
are both 1 − p. For bit flips, the random and oblivious error channels also have the same capacity at
1 − h(p). Can we construct codes that decode oblivious deletions with rate approaching the capacity of
30
the random deletion channel (note this capacity itself is not known)? On the other end, could we upper
bound the best possible rate for correcting a fraction p of oblivious deletions by o(1 − p) as p → 1?
Such an upper bound would fundamentally distinguish the behavior of the deletions from errors and
erasures.
5. Can one find deterministic codes correcting p fraction of online deletions when p approaches 21 ? By
Theorem 6.1, this would imply p0(adv) = 1/2.
6. Given the discussion in §6.2, can we extend our results against online deletions to stochastic codes,
showing that if p0(adv) < 1/2 then there do not exist stochastic codes correcting a fraction of online
deletions approaching 1/2?
8 Acknowledgements
The authors would like to thank Joshua Brakensiek and Boris Bukh for helpful discussions. The authors
would also like to thank anonymous reviewers for helpful feedback on earlier versions of this paper.
References
[1] Raef Bassily and Adam D. Smith. Causal erasure channels. In Proceedings of the Twenty-Fifth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7,
2014, pages 1844–1857, 2014.
[2] Boris Bukh and Venkatesan Guruswami. An improved bound on the fraction of correctable deletions.
In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2016, Arlington, VA, USA, January 10-12, 2016, pages 1893–1901, 2016.
[3] Boris Bukh, Venkatesan Guruswami, and Johan Håstad. An improved bound on the fraction of
correctable deletions. IEEE Trans. Information Theory, 63(1):93–103, 2017.
[4] Zitan Chen, Sidharth Jaggi, and Michael Langberg. A characterization of the capacity of online
(causal) binary channels. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory
of Computing, STOC ’15, pages 287–296, New York, NY, USA, 2015. ACM.
[5] Imre Csiszár and Prakash Narayan. Arbitrarily varying channels with constrained inputs and states.
IEEE Transactions on Information Theory, 34(1):27–34, 1988.
[6] Imre Csiszár and Prakash Narayan. Capacity and decoding rules for classes of arbitrarily varying
channels. IEEE Transactions on Information Theory, 35(4):752–769, 1989.
[7] Daniel Cullina and Negar Kiyavash. An improvement to Levenshtein’s upper bound on the cardinality
of deletion correcting codes. IEEE Trans. Information Theory, 60(7):3862–3870, 2014.
[8] Bikash Kumar Dey, Sidharth Jaggi, Michael Langberg, and Anand D. Sarwate. Upper bounds on the
capacity of binary channels with causal adversaries. IEEE Trans. Information Theory, 59(6):3753–
3763, 2013.
[9] Bikash Kumar Dey, Sidharth Jaggi, Michael Langberg, and Anand D. Sarwate. A bit of delay is sufficient
and stochastic encoding is necessary to overcome online adversarial erasures. In IEEE International
Symposium on Information Theory, ISIT 2016, Barcelona, Spain, July 10-15, 2016, pages 880–884,
2016.
31
[10] Suhas Diggavi and Matthias Grossglauser. On transmission over deletion channels. In Proceedings
of the 39th Annual Allerton Conference on Communication, Control, and Computing, pages 573–582,
2001.
[11] Eleni Drinea and Michael Mitzenmacher. On lower bounds for the capacity of deletion channels. IEEE
Trans. Information Theory, 52(10):4648–4657, 2006.
[12] Eleni Drinea and Michael Mitzenmacher. Improved lower bounds for the capacity of i.i.d. deletion and
duplication channels. IEEE Trans. Information Theory, 53(8):2693–2714, 2007.
[13] Venkatesan Guruswami and Ray Li. Efficiently decodable insertion/deletion codes for high-noise and
high-rate regimes. In IEEE International Symposium on Information Theory, ISIT 2016, Barcelona,
Spain, July 10-15, 2016, pages 620–624, 2016.
[14] Venkatesan Guruswami and Ray Li. Efficiently decodable codes for the binary deletion channel. CoRR,
abs/1705.01963, 2017.
[15] Venkatesan Guruswami and Adam D. Smith. Optimal rate code constructions for computationally
simple channels. J. ACM, 63(4):35:1–35:37, 2016.
[16] Venkatesan Guruswami and Carol Wang. Deletion codes in the high-noise and high-rate regimes. IEEE
Trans. Information Theory, 63(4):1961–1970, 2017.
[17] Bernhard Haeupler and Amirbehshad Shahrasbi. Synchronization strings: codes for insertions and
deletions approaching the singleton bound. In Proceedings of the 49th Annual ACM SIGACT Symposium
on Theory of Computing, STOC 2017, Montreal, QC, Canada, June 19-23, 2017, pages 33–46, 2017.
[18] Adam Kalai, Michael Mitzenmacher, and Madhu Sudan. Tight asymptotic bounds for the deletion
channel with small deletion probabilities. In IEEE International Symposium on Information Theory,
ISIT 2010, June 13-18, 2010, Austin, Texas, USA, Proceedings, pages 997–1001, 2010.
[19] Yashodhan Kanoria and Andrea Montanari. Optimal coding for the binary deletion channel with small
deletion probability. IEEE Trans. Information Theory, 59(10):6192–6219, 2013.
[20] Ian A. Kash, Michael Mitzenmacher, Justin Thaler, and Jonathan Ullman. On the zero-error capacity
threshold for deletion channels. In Information Theory and Applications Workshop, ITA 2011, San
Diego, California, USA, February 6-11, 2011, pages 285–289, 2011.
[21] Ankur A. Kulkarni and Negar Kiyavash. Nonasymptotic upper bounds for deletion correcting codes.
IEEE Trans. Information Theory, 59(8):5115–5130, 2013.
[22] Michael Langberg. Oblivious communication channels and their capacity. IEEE Trans. Information
Theory, 54(1):424–429, 2008.
[23] Amos Lapidoth and Prakash Narayan. Reliable communication under channel uncertainty. IEEE Trans.
Information Theory, 44(6):2148–2177, 1998.
[24] Vladimir I. Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals.
Dokl. Akad. Nauk, 163(4):845–848, 1965 (Russian). English translation in Soviet Physics Doklady,
10(8):707-710, 1966.
[25] Vladimir I. Levenshtein. Bounds for deletion/insertion correcting codes. In Proceedings of the IEEE
International Symposium on Information Theory, 2002.
32
[26] Michael Mitzenmacher. A survey of results for deletion channels and related synchronization channels.
Probability Surveys, 6:1–33, 2009.
[27] Mojtaba Rahmati and Tolga M. Duman. An improvement of the deletion channel capacity upper
bound. In 51st Annual Allerton Conference on Communication, Control, and Computing, Allerton
2013, Allerton Park & Retreat Center, Monticello, IL, USA, October 2-4, 2013, pages 1221–1225,
2013.
A
Relating average-case and worst-case miscommunication probability
In this appendix, we prove the following lemma, which shows that constructing (deterministic) codes
decodable against oblivious deletions in the average case (over random messages, in the sense of Theorem 2.2),
suffices to construct randomized codes with low decoding error probability for every message, as guaranteed
by Theorem 1.2.
Lemma A.1. Let p ∈ (0, 1). Suppose we have a family of codes C of length n and rate R such that, for any
deletion pattern τ with at most pn deletions, at most ϵ = on (1) of the codewords are decoded incorrectly.
Then, for any δ < 21 , provided n is sufficiently large, we can find a family of stochastic codes C ′ of length n
and rate R(1 − δ ) − o(1) that corrects pn oblivious deletions with probability 1 − 3ϵ.
Proof. For set of codewords C ∗ , a pair (c, τ ) (where c ∈ C ∗ and τ is a deletion pattern) is said to be C ∗ -bad if
τ (c) ⊆ c ′ for some c ′ ∈ C ∗ \ {c}, and C ∗ -good otherwise.
Let M = |C | = 2 Rn . For any τ , let Aτ denote the set of c such that (c, τ ) is C-bad. Then |Aτ | ≤ ϵ |C | by
assumption that C decodes against pn deletions
in the average case. Let d = ⌈h(p)/R⌉ be a constant over
n
d
varying n, so that there are at most M > pn choices of τ .
Let t = M δ , and N = ⌊M 1−δ /2⌋. We construct C ′ iteratively. For 1 ≤ k ≤ N , chose a random subset of
k−1 E to form E . With these sets of codewords, we associate each message
exactly t codewords from C \ ∪i=1
i
k
mi with a set Ei . We encode each message mi by choosing uniformly at random one of t codewords in some
set Ei .
Note that C ′ ⊆ C and |C ′ | = N t = M/2. It is easy to see the rate of C ′ is R(1 − δ ) − o(1).
We claim that, with positive probability over the choice of C ′, C ′ corrects pn oblivious deletions with
probability 1 − 3ϵ. Define Bτ to be all c such that (c, τ ) is C ′-bad. As Bτ ⊆ Aτ , we have |Bτ | ≤ ϵM.
We wish to show that the probability there exists an Bτ and a message mi with encoding set Ei such that
|Bτ ∩ Ei | ≥ 3ϵ |Ei | is less than 1. Fix a τ . We show that the probability |Bτ ∩ Ei | ≥ 3ϵ |Ei | is tiny. When
we chose Ei , we can imagine we picked the t elements of Ei one after the other. Each one was chosen from
at least M/2 codewords, so each codeword lands in |Bτ ∩ Ei | with probability at most ϵM/(M/2) = 2ϵ. By
Chernoff bounds, Pr [|Bτ ∩ Ei | ≥ 3ϵ |Ei |] ≤ e −t /3 . By union bound over all Ei and all Bτ , we have
Pr [∃Ei ∃τ : |Bτ ∩ Ei | ≥ 3ϵ |Ei |] ≤ N · M d · e −t /3 < M d+1 /e M
This is less than 1 for n >
1 − 3ϵ, as desired.
2
δR
δ /3
(70)
log(3(d + 1)), so our code corrects pn oblivious deletions with probability
B Comparison of oblivious deletions with oblivious bit flips
First, for completeness, we outline the proof that that there exist codes against oblivious bit flips of rate
matching the capacity of the binary symmetric channel. After this, we show that it is not possible to follow
33
the same approach to construct codes against oblivious deletions.
Theorem B.1 ([22, 5]). For every p ∈ [0, 1/2), and every R < 1 − h(p), there exists a binary code family of
rate R that can correct a fraction p of oblivious bit flips. 4
Proof. We consider a stochastic code C where each message m is mapped uniformly into a set E(m) of size
t. Choose t = n and ϵ = 1/log n. This code decodes successfully if, for all m,
Ø
′
Bpn (c ) ≥ 1 − ϵ
Pr c + e ∈
c ∈E(m)
c ∈C\E(m)
(71)
where Bpn (x) ⊆ {0, 1}n is the set of all words within Hamming distance pn of x.
Choose C to have 2 Rn codewords at random. Fix a message m, the rest of the codeword C \ E(m), and
the error vector e. We have
Pr
c∼{0,1}n
Rn
h(p)n
Ø
′
c ∈
≤ 1 |C | · |Bpn (0)| ≈ 2 · 2
B
(c
+
e)
,
pn
2n
2n
c
∈C\E(m)
(72)
which is 2−Ω(n) if R < 1 − h(p). In this case
Ø
′
Bpn (c ) < 1 − ϵ
Pr Pr c + e <
E(m) c ∈E(m)
c ∈C\E(m)
Ø
= Pr Pr c <
Bpn (c ′ + e) < 1 − ϵ
E(m) c ∈E(m)
c ∈C\E(m)
Ø
Bpn (c ′ + e) > ϵt
= Pr E(m) ∩
E(m)
c ∈C\E(m)
2
≤ 2−Ω(ϵ nt ) = 2−Ω̃(n ) .
(73)
Thus, applying a union bound over all m and error vectors e, we get
Ø
2
2
′
Bpn (c ) > ϵ ≤ 2 Rn · 2n · 2−Ω̃(n ) ≤ 2−Ω̃(n ) .
Pr ∃m : Pr c + e ∈
C
c ∈E(m)
c ∈C\E(m)
(74)
Thus when R < 1 −h(p), we can construct a rate R stochastic code correcting w.h.p. p oblivious bit flips.
For oblivious deletions the above technique does not work. This is established by the following lemma.
Lemma B.2. Let C be a random length n rate R stochastic binary code such that each message is encoded
uniformly at random in one of t codewords. If p ≥ 0.4 and 0 < ϵ < 1, then for any message m and deletion
pattern τ we have
Pr
C
Pr [∃c ′ ∈ C \ E(m) : τ (c) ⊑ c ′] > ϵ
c ∈E(m)
> 2−h(p)n
(75)
where D k (s) = {s ′ : |s ′ | = |s | − k, s ′ ≤ s}.
4Note that 1 − h(p) is an upper bound on the rate as the capacity of the binary symmetric channel with crossover probability p is
1 − h(p) and the oblivious channel can always just choose the error vector randomly.
34
In plain English, this lemma says that for a fixed message and deletion pattern, the probability that
message decodes incorrectly too many times (> ϵ fraction of the time) is too large (> 2−h(p)n ). In short, the
reason the lemma is true is because the probability the entire code C contains a “really bad” word (e.g. a
word with at least 0.91n runs of 1s) is too large. The following remark shows us, using Lemma B.2, why we
cannot follow the same randomized approach as Theorem B.1.
Remark B.3. Following the approach of Theorem B.1, we would like to conclude, using union bound,
′
′
Pr ∀m∀τ Pr [∃c ∈ C \ E(m) : τ (c) ⊑ c ] > ϵ
C
c ∈E(m)
ÕÕ
′
′
≤
Pr Pr [∃c ∈ C \ E(m) : τ (c) ⊑ c ] > ϵ
m
≤2
Rn
τ
C
c ∈E(m)
n
Pr
pn C
′
′
Pr [∃c ∈ C \ E(m) : τ (c) ⊑ c ] > ϵ
c ∈E(m)
< 2 Rn 2h(p)n 2−h(p)n−Ω(n) = 2−Ω(n)
(76)
Unfortunately the last inequality (indicated in red) is false by Lemma B.2.
Proof of Lemma B.2. It suffices to prove for p = 0.4. Indeed, as p increases,
Ø
′
Pr τ (c) ∈
Dpn (c )
c ∈E(m)
c ′ ∈C\E(m)
(77)
increases as τ (c) contains fewer symbols.
If p = 0.4, then τ (c) has 0.6n characters. Note that if c is uniform on {0, 1}n , then τ (c) is uniform on
{0, 1}0.6n . It is easy to check that for a uniformly chosen 0.6n string, the probability it is a subsequence
of the length 0.91n string a 0.91n = 0101 . . . 01 is 1 − 2−Ω(n) : the position in the longest string increases by
1.5 each time, so in expectation the string spans 0.9n characters of a 0.91n , so it is not a subsequence with
exponentially small probability by Chernoff bounds.
There are 20.09n length n strings that start with a 0.91n . Call this set of strings A. Thus, the probability that
C contains an element of A is at least 2−0.91n since that is the probability the first element is an element of A.
Conditioned on C containing an element of A, the probability that, for some c ∈ E(m), τ (c) is a subsequence
of some element of A is at least the probability that c is a subsequence of a 0.91n , which is 1 − 2−Ω(n) . Thus,
conditioned on C containing an element of A, we have
E [#{c : ∃c ′ ∈ C \ E(m) s. t. τ (c) ≤ c ′ }] = t(1 − 2−Ω(n) )
(78)
Conditioned on C being fixed, consider the indicator random variables X i for whether each ci ∈ E(m) satisfies
Í
τ (ci ) is a subsequence of some c ′ ∈ C \ E(m). The X 1 , . . . , X t are i.i.d, so the probability X i /t > ϵ is
approximately 1, (for sure, it is 1 − 2−Ω(n) ).
Thus we conclude
Pr[#{c : ∃c ′ ∈ C \E(m) s. t. τ (c) ≤ c ′ } > ϵt] ≥ (1−2−Ω(n) ) Pr[C ∩A , ∅] > 2(0.91−o(1))n > 2−H (0.4)n (79)
C
C
as H (0.4) ≈ 0.97.
Intuitively B.2 should be true as deletion codes behave poorly when chosen completely randomly. In
the adversarial deletion channel, random codes correct only 0.17n deletions [20], while the best known
constructions correct 0.41n deletions [3].
35
C
Proofs of Lemma 5.19 and Lemma 5.20
√
Lemma C.1. For any j ∈ [1, K], if Z is a random variable distributed as min(Geometric(j/K), R − 1),
K
− 1.
then E[Z ] > 2j
Proof. We have
√R−2
√
j
j
j
j
j
E[D] =
·1+
1−
·2+ ··· +
1−
· ( R − 1)
K
K
K
K
K
√R−1
√R−1
1 − 1 − Kj
√
j
=
− ( R − 1) 1 −
j/K
K
√
R−1
√R−1
√
1 − 1 − K1
K
1
≥
>
− ( R − 1) 1 −
−1
j/K
K
2j
(80)
The last inequality follows from recalling R = 4K 4 and twice applying
√
K
K2
√
K − 1 R−1
1
K −1
2
< 2K 2 ·
< 2K ·
< 1,
R−1
K
K
2
which is true since K > 8.
(81)
Lemma (Lemma 5.19). Let J be chosen uniformly from [K]. Let D be a random
√ variable that is 1 if J ∈ [λ −1]
and, conditioned on a fixed J ≥ λ, is distributed as min(Geometric(J /K), R). Then E[D] ≥ (log K)/4.
Proof. Applying Lemma C.1,
λ−1
K
K
Õ
Õ
1Õ
1
1 K
E[D] =
E[D |J = j] >
·1+
−1
K j=1
K
K 2j
j=1
≥ −1 +
K
Õ
j=λ
1
2j
j=λ
>
1
1
(ln K − ln λ − 1) − 1 > log K.
2
4
(82)
Lemma (Lemma 5.20). Let λ ′ ∈ [λ, K] and let J be chosen uniformly from {λ, λ + 1, . . . , λ ′ }.
√ Let D be
the random variable that, conditioned on a fixed J , is distributed as min(Geometric(J /K), R). Then
E[D] ≥ (log K)/4.
Proof. Applying Lemma C.1,
λ′
λ′
Õ
Õ
1
K
1
E[D |J = j] > ′
−1
E[D] = ′
λ −λ+1
λ −λ+1
2j
j=λ
1
>
K −λ+1
K
Õ
j=λ
K
−1
2j
36
> −1 +
K
Õ
j=λ
j=λ
1
2j
>
1
log K.
4
| 8cs.DS
|
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
arXiv:1403.0679v1 [math.AG] 4 Mar 2014
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
A BSTRACT. We provide explicit combinatorial descriptions of the primary components of codimension two lattice basis ideals. As an application, we compute the set of parameters for which a
bivariate Horn system of hypergeometric differential equations is holonomic.
1. I NTRODUCTION
Let k be a field. A binomial is an element of a polynomial ring krxs “ krx1 , . . . , xn s with at most
two terms; a binomial ideal is one whose generators can be chosen binomial. The zero sets of
binomial ideals are unions of toric varieties, which makes binomial ideals important in algebraic
geometry, and is one reason that combinatorial methods are very effective for studying them. Beyond their intrinsic interest, binomial ideals arise naturally in various contexts, such as combinatorial game theory, algebraic statistics and dynamics of mass action kinetics, see [M11, DSS, SS10],
for references and details. These ideals are also very important in the study of hypergeometric
differential equations, as is shown in [DMM10b, BMW13].
Binomial ideals are highly structured. Assuming that the field k is algebraically closed, Eisenbud
and Sturmfels [ES98] have shown that the associated primes and primary components of a binomial ideal can be chosen binomial. In addition, when k is of characteristic zero, a combinatorial
description of the primary components of a binomial ideal was provided by Dickenstein, Matusevich and Miller in [DMM10a]. This description involves graphs whose vertices are the lattice
points in certain cones in Rn , and are thus difficult to visualize when the number of variables is
high.
If I is a binomial ideal, there exists a (multi)grading of the polynomial ring that makes I a homogeneous ideal. In fact, it is often the case that a binomial ideal is given together with a specified
grading. Depending on their behavior with respect to a given grading, the associated primes and
primary components of a binomial ideal are called toral or Andean (see Definition 3.5). In general,
the toral primary components of a binomial ideal are more easily understood combinatorially than
the Andean ones, as their graphs can actually be drawn in much lower dimension than the number
of variables.
In this article, we study the primary decomposition of codimension two lattice basis ideals, which
have a prescribed grading (see Definition 2.11, Convention 3.1 and the paragraph before Definition 5.4). In this case, it is known that the graphs controlling their toral primary components have
vertices in N2 , regardless of the dimension of the ambient polynomial ring. Our goal here is to
understand the combinatorics of the Andean primary components of a codimension two lattice
basis ideal; this is achieved in Theorem 5.1. To prove this result, we construct an infinite family
2010 Mathematics Subject Classification. Primary: 13F99, 52B20 ; Secondary: 33C70, 20M25.
The authors were partially supported by NSF Grant DMS 1001763.
1
2
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
of graphs with vertices in N2 , that nevertheless contain enough information to produce the needed
primary components.
As was mentioned before, binomial ideals, and in particular, their primary decompositions, play
a key role in the study of hypergeometric functions. Using our new understanding of the Andean
components of a codimension two lattice basis ideal, we are able to compute the Andean arrangement (Definition 5.4) of a system of hypergeometric differential equations in two variables, which
consists of all the parameters for which such a system possesses a finite dimensional solution space
(Theorem 5.8).
Acknowledgements. We are grateful to Christine Berkesch Zamaere, Thomas Kahle, Ezra Miller,
and Christopher O’Neill for interesting conversations while we were working on this project, and
for comments on a previous version of this article.
2. G RAPHS
ASSOCIATED TO MATRICES
Throughout this article, N denotes the monoid t0, 1, 2, . . . u.
One of the main aims in [DMM10a] is to give a combinatorial description of the primary components of binomial ideals. This description involves graphs whose vertices belong to submonoids
of Zn , and more precisely, the connected components of those graphs (see, for instance, Theorem 3.2). In this section, we study graphs arising from 2 ˆ 2 integer matrices, whose vertices are
elements of N2 . These graphs are simpler than those introduced in [DMM10a], but it turns out that
they are sufficient to control the primary components of the special kind of binomial ideal we are
interested in.
We collect our results on graphs here, since they require no algebraic preliminaries, and may be
of independent combinatorial interest. In Sections 4 and 5, we use these graphs to compute the
primary components of codimension two lattice basis ideals.
Definition 2.1. Let Q be a subset of Zn and let M be an n ˆ m integer matrix. We define a graph
GQ pMq whose vertices are the elements of Q, and where two vertices u, v P Q are connected by an
edge if and only if u ´ v or v ´ u is a column of M. A connected component of GQ pMq is called
infinite, if it contains infinitely many vertices; otherwise it is called finite. A finite (or infinite)
vertex of GQ pMq is one that belongs to a finite (or infinite) connected component. If Q “ Nn , we
omit Q from the notation, and write GpMq instead of GNn pMq.
In this article, we consider graphs GQ pMq where Q is a submonoid of Zn such as Nn or Nk ˆ Zn´k
(for instance, in Proposition 4.3), or a subset of Zn , such as tu P Nn | λ1 u1 ` ¨ ¨ ¨ ` λn un “ λ0 u,
for fixed given λ0 , . . . , λn P Q (Lemma 4.6), or tu P N2 | u1 ď ℓu, for fixed given ℓ P N
(Proposition 2.5).
We are interested in the connected components of GQ pMq, and in particular, in determining
whether these connected components are infinite or finite.
Our first result concerns the connected components of GpMq, where M is a 2 ˆ 2 nonsingular
matrix whose rows lie in non adjacent (also called opposite) open quadrants of Z2 , and Q “ N2 is
omitted from the notation.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
3
Proposition 2.2 (Lemma 6.5 in [DMS05]). Let M “ rµij si,jPt1,2u P Z2ˆ2 of rank two, and assume
that µ11 , µ12 ą 0 and µ21 , µ22 ă 0. Set
"
tu P N2 | u1 ă µ12 , u2 ă ´µ21 u
if |µ11 µ22 | ą |µ12 µ21 |,
R“
2
tu P N | u1 ă µ11 , u2 ă ´µ22 u
if |µ11 µ22 | ă |µ12 µ21 |.
Every finite connected component of GpMq contains exactly one vertex in R. In particular, the
number of finite connected components of GpMq is the cardinality of R, which is minp|µ11 µ22 |, |µ12 µ21 |q.
„
1
3
. GpMq has minp|´4|, |´6|q “ 4 finite connected components,
Example 2.3. Let M “
´2 ´4
which are shown in Figure 1.
z
•
• •
✳
✳ • • •
✳ ✳
✳ ✳ • •
w
F IGURE 1. The graph of M.
If M P Z2ˆ2 has all positive entries, then GpMq has no finite connected components. However,
in this case, it is not GpMq that carries the algebraic information we need later on, but a family of
subgraphs of GpMq, as follows.
Notation 2.4. Let M P Z2ˆ2 of rank two, all of whose entries are positive. For ℓ P N let Qℓ “
tu P N2 | u1 ď ℓu. We denote Gℓ pMq – GQℓ pMq, and call these graphs the band graphs of M.
Note that Gℓ pMq is the induced subgraph of GpMq whose vertices lie in Qℓ , and consequently if
ℓ ď ℓ1 , then Gℓ pMq is a subgraph of Gℓ1 pMq.
In order to understand the connected components of the band subgraphs of GpMq, we first analyze
a special case.
Proposition 2.5. Consider a rank two integer matrix
„
r s
M“
a b
such that r ě s ą 0, 0 ă a ď b, and gcdpr, sq “ 1. If ℓ ă r ` s ´ 1, then every connected
component of Gℓ pMq is finite. If ℓ ě r ` s ´ 1, then for every w0 P t0, 1, . . . , ℓu there exists z0 P N
such that pw0 , z0 q belongs to an infinite component of Gℓ pMq.
Before proving this result, we introduce some auxiliary notions, and provide some examples.
Given M as above, the graphs Gℓ pMq have two types of edges: those parallel to the first column
of M are called the r-edges of Gℓ pMq, and those parallel to the second column of M are called
the s-edges of Gℓ pMq. If r “ s, we could refer to these edges as a-edges and b-edges.
4
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
Definition 2.6. Let M be as in Theorem 2.5, and consider N2 with coordinates w, z. A vertex of
Gℓ pMq is called a turn if it is adjacent to both an r-edge and an s-edge of Gℓ pMq. A turn pw0 , z0 q
is called a left turn if there is a vertex adjacent to pw0 , z0 q whose w-coordinate is smaller than w0 .
Turns that are not left turns are called right turns. Intuitively, when we walk along a connected
component of Gℓ pMq in the direction that increases z, we turn left at a left turn, and right at a right
turn.
Example 2.7. Let
M“
„
4
1
7
1
The band graphs G4 pMq and G7 pMq are illustrated in Figure 2.
z
z
•
•
•
•
•
left turn
•
•
•
w
G4 pMq
•
•
•
•
G7 pMq
w
F IGURE 2. Examples of band graphs.
All of the connected components of G4 pMq and G7 pMq are finite. The minimum ℓ P N such that
Gℓ pMq has an infinite connected component is ℓ “ 10 (see Figure 3).
z
•
•
•
•
•
•
•
•
•
•
•
•
w
G10 pMq
F IGURE 3. A band graph with an infinite component.
Proof of Proposition 2.5. Write r “ sq1 ` q2 where 0 ď q2 ă s. We claim that any connected
component of Gr pMq contains at most 2q1 `2 vertices (implying that Gr pMq, and therefore Gℓ pMq
for ℓ ď r, has no infinite connected components). A connected component of Gr pMq can only
contain one r-edge since the w-coordinates of vertices in Gr pMq are bounded by r. Thus, we can
have at most two turns in such a connected component. We can connect at most q1 -many s-edges
at each turn. Including the turns, the number of vertices in a connected component of Gr pMq is at
most equal to 2q1 ` 2.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
5
We observe that a modification of the argument above shows that a connected component of Gℓ pMq
is infinite if and only if it contains infinitely many left turns.
Now consider Gr`t pMq where 0 ď t ă s. We show that Gr`t pMq has an infinite connected
component if and only if t “ s ´ 1. Note that not all components of Gr`t pMq have left turns, for
instance, the vertex pr ` t, 0q is itself a connected component, which therefore has no turns. In
what follows, we study how many left turns a connected component can have.
The ordering ą on the elements of N2 defined by pw, zq ą pw 1 , z 1 q if z ą z 1 , or z “ z 1 and w 1 ą w,
induces a total ordering on the set of left turns of a given component of Gr`t pMq.
Let C be a connected component of Gr`t pMq, and suppose that pw0 , z0 q is a left turn in C. We
wish to produce the next left turn of C according to ą, if it exists. Since pw0 , z0 q is a left turn in C,
we have pw0 ´ r, z0 ´ aq P C. This is a right turn, because Gr`t pMq cannot contain two adjacent
r-edges, as the w-coordinates of the vertices of Gr`t pMq are bounded by r ` t, and t ă s. We
attach s-edges to pw0 ´ r, z0 ´ aq, to produce a vertex pw0 ´ r, z0 ´ aq ` pqs, qbq P C, where q ą 0
is as large as possible. The integer q is produced by writing r ` t´ pw0 ´ rq “ qs ` rr ` t´ pw0 ´ rq
mod ss, where rα mod βs denotes the remainder of α upon division by β, for α, β P Z, α ą 0.
If pw0 ´ r, z0 ´ aq ` pqs, qbq is coordinatewise greater than or equal to pr, aq, then pw0 ´ r, z0 ´
aq ` pqs, qbq is a left turn of C which is greater according to ą than pw0 , z0 q. Now, z0 ´ a ě 0 and
b ě a imply that z0 ´ a ` qb ě a. Therefore, in order for pw0 ´ r, z0 ´ aq ` pqs, qbq to be a left turn,
we need r ď w0 ´ r ` qs “ r ` t ´ rr ` t ´ pw0 ´ rq mod ss, or equivalently, t ě r2r ` t ´ w0
mod ss.
Replacing pw0 , z0 q by pw0 ´r, z0 ´aq`pqs, qbq, we see that the condition needed for the existence of
a left turn which is greater according to ą than pw0 ´r, z0 ´aq`pqs, qbq is t ě r2r`t´pw0 ´r`qsq
mod ss “ r3r ` t ´ w0 mod ss.
Continuing in this manner, the existence of infinitely many left turns in C is equivalent to requiring
t ě rℓr ` t ´ w0 mod ss for all ℓ ą 0. However, since gcdpr, sq “ 1, there exists ℓ ą 0 such that
rpℓr ` t ´ w0 mod ss “ s ´ 1. Therefore, if t ă s ´ 1, C has finitely many left turns, and is finite,
and if t “ s ´ 1, C has infinitely many left turns, and is infinite.
If t ă s ´ 1, a component of Gr`t pMq either has no left turns or finitely many left turns, which
shows that Gr`t pMq has no infinite components.
Let t “ s ´ 1 and w0 P t0, . . . , r ` s ´ 1u. If w0 ě r, then for large enough z0 , pw0 , z0 q is a vertex
of both an r- and an s-edge whose other vertex has lower z-coordinate, and is therefore a left turn
in its connected component, which is thus infinite. If w0 ă r, we can choose z0 sufficiently large
such that attaching as many s-edges to pw0 , z0 q as possible yields a left turn, which implies that the
component of pw0 , z0 q is infinite.
Finally, if ℓ ą r ` s ´ 1, for each 0 ď t ď ℓ ´ pr ` s ´ 1q, Gℓ pMq contains as a subgraph the image
of Gr`s´1pMq under the translation pw, zq ÞÑ pw`t, zq. This implies that for each w0 P t0, . . . , ℓu,
there is z0 ą 0 such that pw0 , z0 q is an infinite vertex of Gℓ pMq.
In the previous statement, we assumed that the entries in the top column of M were relatively
prime. We now remove that assumption.
Theorem 2.8. Consider a rank two matrix
„
r s
M“
a b
6
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
where r, s, a, b are positive integers, r ě s, a ď b and gcdpr, sq “ d ě 1. The minimal ℓ P N
such that Gℓ pMq has an infinite connected component is ℓ “ r ` s ´ d. If 0 ď t ă d and
w0 P t0, . . . , r ` s ´ d ` tu, there exists z0 such that pw0 , z0 q is an infinite vertex of Gr`s´d`t pMq
if and only if w0 is divisible by d. If ℓ ą r ` s, for each w0 P t0, . . . , ℓu, there exists z0 such that
pw0 , z0 q is an infinite vertex of Gℓ pMq.
„
2 6
. When ℓ “ 6, the band graph G6 pMq has an infinite connected
Example 2.9. Let M “
1 2
component. However, the vertices pw, zq where w is odd are finite vertices for all z; see Figure 4.
z
✳
•
✳
✳ •
✳ • ✳ • ✳ •
• ✳ • ✳ • ✳
• ✳ • ✳
✳ w
F IGURE 4. The band graph G6 pMq.
Proof. Let M̂ be the (integer) matrix obtained from M by dividing r and s by d, so that Proposition 2.5 applies to the band graphs of M̂.
Let ℓ P N and set ℓ̂ “ tℓ{du, the integer part of ℓ{d. We show that Gℓ pMq is a disjoint union of
graphs isomorphic to Gℓ̂ pM̂ q or Gℓ̂´1 pM̂q.
Let pw0 , z0 q P N2 such that w0 ď ℓ, so that pw0 , z0 q is a vertex of Gℓ pMq. Write w0 “ ŵ0 d ` t0
where t0 is an integer with 0 ď t0 ă d.
If t0 “ 0, then pw0 , z0 q belongs to the image of the map ϕℓ,0 : Gℓ̂ pMq Ñ Gℓ pMq defined on vertices
by pw, zq ÞÑ pdw, zq. Since r and s are divisible by d, any vertex in the connected component
of Gℓ pMq that contains pw0 , z0 q also has its d-coordinate divisible by d. This implies that the
connected component of pw0 , z0 q in Gℓ pMq is the image under ϕℓ,0 of the connected component of
pw0 {d “ ŵ0 , z0 q in Gℓ̂ pM̂ q.
If t0 ą 0, consider the map ϕℓ,t0 : Gℓ̂´1 pMq Ñ Gℓ pMq defined on vertices by pw, zq ÞÑ pdw `
t0 , zq. Since r and s are divisible by d, the w-coordinates of all the vertices of Gℓ pMq connected
to pw0 , z0 q are congruent to t0 modulo d. This implies that the connected component of pw0 , z0 q in
Gℓ pMq is the image under ϕℓ,t0 of the connected component of pŵ0 , z0 q in Gℓ̂´1 pM̂ q.
Note that the images of the maps ϕℓ,i have no common vertices, and their union is Gℓ pMq. Now
use Proposition 2.5 to obtain the desired conclusions.
We now explain how to construct graphs arising from binomial ideals. Lemma 2.12 relates these
graphs and those associated to matrices in the case of lattice basis ideals.
Definition 2.10. Let P be a monoid. A binomial ideal I in the monoid ring krP s defines a graph
GP pIq whose vertices are the elements of P and whose edges are pairs pu, vq P P ˆ P such that
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
7
xu ´ ρxv P I for some ρ P k r t0u. A connected component of GP pIq is said to be infinite if
it consists of infinitely many vertices; otherwise it is called finite. A vertex of GP pIq is called an
infinite vertex if it belongs to an infinite connected component, otherwise it is called a finite vertex.
If P “ Nn , we omit P from the notation, and write G pIq instead of GP pIq.
Note that if I Ă krP s is a binomial ideal, any connected component of GP pIq is a complete graph.
We are ready to introduce our main objects of study.
Definition 2.11. If µ P Zn , define µ` , µ´ P Nn via pµ` qi “ maxpµi , 0q and pµ´ qi “ maxp´µi , 0q,
so that µ “ µ` ´ µ´ . Let M be n ˆ m integer matrix. We define a binomial ideal associated to M
as follows:
IpMq “ xxµ` ´ xµ´ | µ is a column of My Ď krx1 , . . . , xn s “ krNn s.
If M has rank m, then IpMq is called a lattice basis ideal.
Lemma 2.12. Let M be an n ˆ m integer matrix of rank m, and IpMq its corresponding lattice
basis ideal as in Definition 2.11. Let τ Ď t1, . . . , nu and let P “ Nτ ˆ Zτ̄ . Then u, v P P are
connected in GP pkrP s ¨ IpMqq if and only if they are connected in GP pMq.
Proof. Assume that u, v P P are connected in GP pMq. We show that xu ´ xv P krP s ¨ IpMq by
induction on the length of the path connecting u to v. If this path has length one, then u and v are
connected by an edge of GP pMq, meaning that u ´ v or v ´ u, say u ´ v, equals a column µ of
M. Then u ´ v “ µ “ µ` ´ µ´ , so that v ´ µ´ “ u ´ µ` — ν. Since v ´ µ´ ` µ` “ u P P
and for all i, pµ` qi and pµ´ qi are not simultaneously nonzero, we see that ν P P . But then
xu ´ xv “ xν pxµ` ´ xµ´ q P krP s ¨ IpMq, as we wished.
Now assume that u and v are connected in GP pMq by a path of length ℓ ą 1. This means that there
are vertices u “ ν p0q , ν p1q , . . . , ν pℓq “ v of GP pMq such that pν piq , ν pi`1q q is an edge of GP pMq
for i “ 0, . . . , ℓ. By inductive hypothesis, since ν p1q and v are connected in GP pMq by a path of
p1q
p1q
length ℓ ´ 1, we have xν ´ xv P krP s ¨ IpMq. But we also know xu ´ xν P krP s ¨ IpMq. We
conclude that xu ´ xv P krP s ¨ IpMq, and therefore u and v are connected in GP pIpMqq.
For the converse, we start by noting that a lattice basis ideal (and its extension to krP s) contains no
monomials. This follows, for instance, from Lemma 7.6 in [MS], which implies that the saturation
pIpMq : xx1 ¨ ¨ ¨ xn y8 q Ď krxs is not the unit ideal.
Since every connected component of GP pIpMqq is a complete graph, if u and v are connected
in GP pIpMqq, then pu, vq is an edge in GP pIpMqq. Thus, there exists nonzero ρ P k such that
xu ´ ρxv P krP s ¨ IpMq, and if µp1q , . . . , µpmq are the columns of M, we can write xu ´ ρxv “
pmq
pmq
p1q
p1q
F1 pxqpxµ` ´ xµ´ q ` ¨ ¨ ¨ ` Fm pxqpxµ` ´ xµ´ q for certain F1 , . . . , Fm P krP s. We can represent
this expression as a subgraph K of GP pMq: for every term λxν in Fi , K contains the edge pν `
piq
piq
µ` , ν ` µ´ q and its corresponding vertices. We label this edge by the coefficient λ, and we label
each vertex by the combination of the labels of the edges adjacent to it, with a positive sign if we
piq
piq
piq
piq
look at the vertex ν ` µ` of pν ` µ` , ν ` µ´ q, and a negative sign for the vertex ν ` µ´ . Thus,
the only two vertices with nonzero labels are u and v.
Let Ku be the connected component of K containing u. We wish to show that v is a vertex in Ku ,
as this implies that u and v are connected in GP pMq. But if this is not the case, we can use Ku to
piq
piq
piq
piq
form a polynomial expression with a summand λxν pxµ` ´ xpµ´ q for each edge pν ` µ` , ν ` µ´ q
8
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
labeled by λ in Ku , and this expression equals the sum over the vertices in Ku of the label of
each vertex times the corresponding monomial. Since the only vertex with a nonzero label in Ku
is u (that label is 1), then we obtain an expression for xu as a combination of the generators of
krP s ¨ IpMq. This contradicts the fact that krP s ¨ IpMq contains no monomials.
With the hypotheses and notation of the previous result, we see that we can construct GP pIpMqq
by adding edges to GP pMq until each connected component becomes a complete graph.
Given an arbitrary binomial ideal I Ď krP s, it is always possible to construct a subgraph of GP pIq
using a generating set of I, so that the underlying vertex sets of their connected components are
the same (and therefore, saturating the connected components of this subgraph with edges yields
GP pIq). For this purpose, not every generating set of I contains sufficient information. What is
needed is a generating set of I that contains all the generators of the maximal monomial ideal in I.
The statement (and proof) of this generalization are more technical, but follow along the same lines
as above. We point out that special cases of this result can be found in the literature, for example
Lemmas 1 and 2 in [MM82].
3. P RIMARY
DECOMPOSITION OF BINOMIAL IDEALS
In this section we review important facts about the primary decomposition of binomial ideals, and
especially of lattice basis ideals (Definition 2.11).
Recall that N “ t0, 1, 2, . . . u. We work in the polynomial ring krx1 , . . . , xn s “ krNn s where k is
an algebraically closed field. Unless otherwise specified, k is of characteristic zero (in Section 5,
we use k “ C).
The associated primes of lattice basis ideals were studied by Hoşten and Shapiro in [HS00]; they
show that the minimal primes of such an ideal are determined by the sign patterns of the entries of
the corresponding matrix. In this article, we study lattice basis ideals arising from n ˆ 2 integer
matrices, known as codimension two lattice basis ideals.
Convention 3.1. From now on, B “ rbij s is an n ˆ 2 integer matrix of rank 2. The rows of B are
denoted by b1 , . . . , bn , and its columns by B1 , B2 . Fix an integer pn ´ 2q ˆ n matrix A such that
AB “ 0, and whose columns span Zn´2 as a lattice.
Since B has rank two, the ideal IpBq from Definition 2.11 is a complete intersection. Therefore
all of its associated primes are minimal. By Corollary
ś 2.1 in [HS00], the set of associated primes
of IpBq consists of the associated primes of pI : p ni“1 xi q8 q and the monomial primes xxi , xj y if
bi and bj lie in opposite open quadrants of Z2 .
ś
All of the associated primes of pI : p ni“1 xi q8 q are isomorphic, by rescaling the variables, to
IA “ xxv` ´ xv´ | v P Zn , Av “ 0y, where A is as in Convention 3.1. The prime ideal IA
is called the toric ideal associated to A. It is shown in [ES98] that the primary components of
IpBq corresponding to these associated primes are the associated primes themselves (since the
characteristic of the underlying field k is zero). Thus, we now turn our attention to the primary
components of IpBq arising from monomial associated primes.
One of the main results in [ES98] is that any binomial prime ideal in krxs is of the form krxs ¨ J `
xxi | i P σy, where σ Ď t1, . . . , nu and J Ă krxj | j R σs is isomorphic to a toric ideal by rescaling
the variables. In characteristic zero, the primary component of a binomial ideal I corresponding
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
9
ś
to such an associated prime is of the form prI ` Js : r jRσ xj s8 q ` M , where M is a monomial
ideal generated by elements of krxi | i P σs. A key idea from [DMM10a] is that the graphs from
Definition 2.10 can be used to determine the monomials in the primary components of a binomial
ideal. A specific result in this vein is Theorem 3.2 below, which describes the primary component
of a binomial ideal corresponding to a monomial associated prime.
We first set up some notation. Given σ Ď t1, . . . , nu, we set σ̄ “ t1, . . . , nu r σ. Denote Nσ “
tu P Nn | ui “ 0 for i R σu, and Zσ̄ “ tu P Zn | ui “ 0 for i P σu. We consider P “ Nσ ˆ Zσ̄ as
a submonoid of Zn . The corresponding monoid ring is krP s “ krNσ ˆ Zσ̄ s “ krx˘
j | j R σsrxi |
i P σs.
Theorem 3.2 (Theorem 2.15, [DMM10a]). Let k be an algebraically closed field (of any characteristic) and I Ă krxs a binomial ideal. Let σ Ď t1, . . . , nu, and set P “ Nσ ˆ Zσ̄ . If xxi | i P σy
is a minimal prime of I, its corresponding primary component is
˙
ˆ
ź
8
` xxu | u P Nn is an infinite vertex of GP pkrP s ¨ Iqy.
I : p xj q
(3.1)
jRσ
Moreover, the only monomials in these primary components are those of the form xu such that
u P Nn is an infinite vertex of GP pkrP s ¨ Iq.
Remark 3.3. Note that the monomial ideal in (3.1) is generated by monomials xu where uj “ 0 if
j R σ. Indeed, if u P Nn is an infinite vertex of GP pkrP s ¨ Iq, so is u ´ û, where ûi “ 0 if i P σ and
ûj “ uj if j R σ. This is because monomials in the variables xj for j R σ are units in krP s.
The following useful criterion to identify the infinite components of GP pkrP s ¨ Iq is a special case
of Lemma 2.10 in [DMM10a].
Lemma 3.4. Let I Ď krx1 , . . . , xn s be a binomial ideal, σ Ď t1, . . . , nu, and P “ Nσ ˆ Zσ̄ .
A connected component of GP pkrP s ¨ Iq is infinite if and only if it contains two distinct vertices
u, v P P such that u ´ v P P .
The graphs GP pkrP s ¨ Iq used in Theorem 3.2 have vertices in P “ Nσ ˆ Zσ̄ . When n is large, we
cannot hope to compute the finite (or infinite) connected components of GP pkrP s ¨ Iq by simply
drawing the graph. In certain situations, other graphs, whose ambient lattice is lower dimensional,
contain the same information as G ` P pkrP s ¨ Iq. This happens, for instance, when xxi | i P σy
behaves well with respect to a given grading, as follows.
The matrix A in Convention 3.1 can be used to define a Zn´2 -grading of krxs, where degpxi q
is defined to be the ith column of A. The ideal IpBq is homogeneous with respect to this Agrading (as are therefore its associated primes and primary components). The associated primes
(and primary components) of an A-graded ideal are classified according to their A-graded behavior.
Definition 3.5. Let I Ď krxs be an A-graded binomial ideal, and p an associated prime of I. If the
A-graded Hilbert function of krxs{p is bounded, then p and the corresponding primary component
of I, are called toral. Otherwise, they are called Andean.
Example 3.6. Among the associated primes of the codimension two lattice basis ideal IpBq, the
only Andean ones are the monomial primes xxi , xj y such that the corresponding rows of B, bi and
bj , are linearly dependent (and in opposite open quadrants of Z2 ).
To see this, first observe that the A-graded Hilbert function of krxs{IA takes only the values zero
and one (and the same holds for the other non-monomial associated primes of IpBq). On the
10
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
other hand, whether the Hilbert function of krxs{xxi , xj y is bounded depends on the rank of the
submatrix of A indexed by t1, . . . , nu r ti, ju: boundedness is equivalent to this submatrix having
full rank, and this happens exactly when bi and bj are linearly independent.
The toral components of an A-graded binomial ideal are more accessible combinatorially than
the Andean ones in general. The following result illustrating this fact is a consequence of Theorem 4.13 from [DMM10a].
Theorem 3.7. Let I be an A-graded binomial ideal in krxs, where k is algebraically closed of
characteristic zero, and assume that p “ xxi | i P σy is a toral minimal prime of I. Define the
binomial ideal I¯ “ I ¨ krxs{xxj ´ 1 | j R σy Ă krNσ s by setting xj “ 1 for j R σ. The p-primary
component of I is:
˙
ˆ
ź
8
¯
` xxu | u P Nσ is an infinite vertex of G pIqy.
I : p xj q
jRσ
¯ has vertices in Nσ , and the cardinality of
The main feature of the previous theorem is that G pIq
σ can be much smaller than n. In the case of codimension two lattice basis ideals, Nσ “ N2 ,
regardless of the number of variables n. In fact, if xxi , xj y is a toral associated prime of IpBq, then
|b |
|b |
|b |
|b |
IpBq “ xxi i1 ´ xj j1 , xi i2 ´ xj j2 y Ă krxi , xj s, and by Lemma 2.12, G pIpBqq can be constructed
from the graph GpMq associated to the 2 ˆ 2 matrix M whose rows are bi and bj . But we have
already characterized the connected components of GpMq in Proposition 2.2, so we can describe
the corresponding primary component by applying Theorem 3.7.
When k is of characteristic zero, this yields a very satisfactory picture of the toral components
of a codimension two lattice basis ideal: the primary components corresponding to non-monomial
associated primes are isomorphic to toric ideals by rescaling the variables; the primary components
corresponding to monomial toral primes are described by Theorem 3.7 and Proposition 2.2.
Example 3.8. Let
fi
2
4
—´4 ´6 ffi
ffi
—
3ffi
— 2
ffi
—
3ffi
—´1
B“—
ffi,
—´1 ´2 ffi
— 2 ´6 ffi
ffi
—
–´8 ´12fl
´3 ´6
6
4
8 3
4 3 3
2 2 2
so that IpBq “ xx1 x3 x6 ´ x2 x4 x5 x7 x8 , x1 x3 x4 ´ x62 x25 x66 x12
7 x8 y.
»
The rows p2, 4q„and p´4,´6q are linearly independent and lie in opposite open quadrants, so we
2
4
. We can compute the monomials of the xx1 , x2 y-primary component of
consider M “
´4 ´6
IpBq by looking the graph of GpMq; see Figure 5.
By Theorem 3.7, the xx1 , x2 y-primary component of IpBq is:
pIpBq : p
ź
ℓ‰1,2
xℓ q8 q ` xx41 , x62 , x21 x22 y “ xx41 , x62 , x21 x22 , x42 x4 x5 x87 x38 ´ x21 x23 x26 y.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
11
uy
•
•
❍
•
•
•
❍
•
•
❍
ux
F IGURE 5. The graph of GpMq
4. C ODIMENSION
TWO LATTICE BASIS IDEALS IN THREE VARIABLES
In this section, we study the Andean components of codimension two lattice basis ideals in three
variables.
Notation 4.1. For this section only, and except where otherwise noted, we let B be 3 ˆ 2 matrix
of full rank 2 as follows:
»
fi
r
s
B “ –´λr ´λsfl
a
b
where r, s, a, b P Zą0 , a ď b, r ě s, gcdpr, sq “ d, and 0 ă λ “ p{q in lowest terms. We work in
the polynomial ring krx, y, zs. The lattice basis ideal associated to B is IpBq “ xxr z a ´y λr , xs z b ´
y λs y Ď krx, y, zs. We let P “ N2 ˆ Z, and work with krP s “ krz ˘ srx, ys.
Remark 4.2. For a codimension two lattice basis ideal IpBq arising from a 3 ˆ 2 matrix B as in
Notation 4.1, a vertex u “ pux , uy , uz q of G pIpBqq that lies on a hyperplane uy “ ´λux ` λℓ, for
ℓ P Q, can only be connected to other vertices on that hyperplane. This follows from Lemma 2.12
since the columns of B are parallel to the hyperplane uy “ ´λux .
For B as above, we wish to compute the primary component of IpBq corresponding to the (Andean) associated prime xx, yy. According to Theorem 3.2, we need to understand the graph arising
from the extension of IpBq to krz ˘ srx, ys “ krP s.
The following proposition shows that all the information we need about the infinite components
of GP pkrP s ¨ IpBqq is contained in the infinite components of G pIpBqq, consequently the graph
GpBq can be used to compute the primary component of IpBq corresponding to the associated
prime xx, yy.
Proposition 4.3. Let B and P “ N2 ˆ Z as in Notation 4.1. Then
tpux , uy q P N2 | Duz P Z such that pux , uy , uz q is an infinite vertex of GP pkrP s ¨ IpBqqu “
tpux , uy q P N2 | Duz P N such that pux , uy , uz q is an infinite vertex of G pIpBqqu.
12
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
Consequently, the primary component of IpBq corresponding to the associated prime xx, yy is:
pI : z 8 q ` xxux y uy | Duz P N such that pux , uy , uz q is an infinite vertex of GpBqy.
Proof. If u “ pux , uy , uz q P N3 is an infinite vertex of G pIpBqq, then it is clear that it is also an
infinite vertex of GP pkrP s ¨ IpBqq.
Let u “ pux , uy , uz q P P be an infinite vertex of of GP pkrP s ¨ IpBqq. By Lemma 3.4 there exists
v “ pvx , vy , vz q, ṽ “ pṽx , ṽy , ṽz q P N2 ˆ Z connected to u such that ṽx ě vx and ṽy ě vy . Since u
is connected to v, we can find a nonzero ρ P k such that xux y uy z uz ´ ρxvx y vy z vz P krP s ¨ IpBq, and
by clearing denominators, we can produce µ P N such that z µ pxux y uy z uz ´ ρxvx y vy z vz q P IpBq; in
particular, µ`uz and µ`vz are non negative. Thus, the vertices pux , uy , uz `µq, pvx , vy , vz `µq P N3
are connected in G pIpBqq. Enlarging µ as needed, we may assume that pux , uy , uz `µq, pvx , vy , vz `
µq, pṽx , ṽy , ṽz ` µq are coordinatewise non negative and connected in G pIpBqq.
By Remark 4.2, there exists ℓ P Q such that vy “ ´λvx ` λℓ and ṽy “ ´λṽx ` λℓ, so that
ṽy ´ vy “ ´λpṽx ´ vx q. Since λ ą 0 and ṽx ´ vx , ṽy ´ vy are non negative, we see that vx “ ṽx
and vy “ ṽy .
In conclusion, the vertices pvx , vy , vz ` µq, pvx , vy , ṽz ` µq P N3 are connected in GpIpBqq; since
either pvx , vy , vz ` µq ´ pvx , vy , ṽz ` µq or pvx , vy , ṽz ` µq ´ pvx , vy , vz ` µq belongs to N3 , we see
that these vertices belong to an infinite component of G pIpBqq by Lemma 3.4. As these vertices
are connected to pux , uy , uz ` µq, we conclude that this is an infinite vertex of G pIpBqq.
The last statement now follows from Theorem 3.2, Remark 3.3 and Lemma 2.12.
For the remainder of this section, we study the connected components of the graph GpBq for B as
in Notation 4.1. The following definition is motivated by Remark 4.2.
Definition 4.4. For a matrix B as in Notation 4.1 and ℓ P p1{pqN, set Spℓq “ Nn X tpux , uy , uz q |
uy “ ´λux ` λℓu. The graph GSpℓq pBq is called the ℓth-slice graph of B. Slice graphs are
illustrated in Figure 6.
Ť
By Remark 4.2, GpBq equals the (disjoint) union ℓPp1{pqN GSpℓq pBq as a graph.
The following result groups the slice graphs of B according to isomorphism.
Lemma 4.5. Let B as in Notation 4.1. Suppose pcx , cy , cz q P N3 is a vertex of GSpℓq pBq, where
ℓ P p1{pqN. Write cx “ qc̄x `i, cy “ pc̄y `j, where c̄x , c̄y , i, j P N and 0 ď i ă q, 0 ď j ă p. Then
GSpℓq pBq is isomorphic to the slice graph GSpℓ´i´j{λqpBq that contains pqc̄x , pc̄y , cz q as a vertex.
Consequently, in order to understand the (connected components of) all the slice graphs of B, it is
enough to understand GSpℓq pBq for ℓ P qN.
Proof. The desired isomorphism ϕij between GSpℓq pBq and GSpℓ´i´j{λq pBq is defined by
pux , uy , uz q ÞÑ pux ´ i, uy ´ j, uz q.
Note that a vertex of the form pqc̄x , pc̄y , cz q P N3 , where c̄x , c̄y P N belongs to a slice graph
GSpℓ̄q pBq where ℓ̄ P qN.
Our next step is to “straighten out” the slice graphs of B. Recall the notation of Notation 4.1.
Given ℓ P N, let φℓ : N2 Ñ Z3 be the injective function defined by pw, kq ÞÑ pqw, λpqℓ ´ qwq, kq “
pqw, ppℓ ´ wq, kq. Note that the image φℓ ptpw, kq P N2 | 0 ď w ď ℓ, 0 ď kuq is the intersection
with N3 of the hyperplane given by uy “ ´λux ` λqℓ.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
13
uz
•
•
•
•
•
GSp4q pBq
•
•
•
GSp7q pBq
•
uy
•
•
•
•
•
ux
F IGURE 6. Slice graphs for IpBq “ xx4 z ´ y 4 , x7 z ´ y 7 y.
“ s{q ‰
Lemma 4.6. Let B be as in Notation 4.1 and set M “ r{q
. Given ℓ P N, define φℓ as above.
a b
The image under φℓ of the band graph Gℓ pMq (see Notation 2.4) is the slice graph GSpqℓq pBq.
Since we have already studied the connected components of the band graphs Gℓ pMq, we are ready
to compute the primary component of IpBq associated to xx, yy.
Theorem 4.7. Let B as in Notation 4.1. The primary component of IpBq corresponding to the
associated prime xx, yy is
pIpBq : z 8 q ` xy λpr`s´dq , xd y λpr`s´2dq , x2d y λpr`s´3dq , . . . , xr`s´d y.
Proof. By Theorem 3.2, the desired component is pIpBq : z 8 q ` M , where
M “ xxux y uy | Duz P Z with u “ pux , uy , uz q an infinite vertex of GN2 ˆZ pIpBqqy.
By Proposition 4.3, we can use G pIpBqq instead of GN2 ˆZ pIpBqq in the definition of M , and by
Lemma 2.12, we can use GpBq instead of G pIpBqq.
Now combine Remark 4.2, Lemmas 4.5 and 4.6 and Theorem 2.8 to obtain the desired result.
5. L ATTICE
BASIS IDEALS AND HYPERGEOMETRIC SYSTEMS
In this section we state and prove our main result, Theorem 5.1, that gives an explicit expression
for the Andean primary components of a codimension two lattice basis ideal. We also provide an
application to the study of hypergeometric systems of differential equations.
14
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
Theorem 5.1. Let B as in Convention 3.1, and suppose that bi , bj are linearly dependent rows
of B lying in opposite open quadrants of Z2 . Without loss of generality, assume that bi1 , bi2 ą 0,
let d “ gcdpbi1 , bi2 q, and write λ “ ´bj1 {bi1 “ ´bj2 {bi2 ą 0. The primary component of IpBq
corresponding to the associated prime xxi , xj y is
ź
λpb `b ´dq
λpb `b ´2dq
λpbi1 `bi2 ´3dq
pIpBq : p
xk q8 q ` xxj i1 i2 , xdi xj i1 i2
, x2d
, . . . , xibi1 `bi2 ´d y.
i xj
k‰i,j
λpbi1 `bi2 ´dq
The only monomials in this ideal are those in xxj
λpbi1 `bi2 ´2dq
, xdi xj
, . . . , xibi1 `bi2 ´d y.
Proof.” Let σı “ ti, ju, and set P “ Nσ ˆ Zσ̄ . Choose 0 ă a ď b P Z such that the matrix
bi1 bi2
B̂ “ bj1 bj2 has rank 2. Our result follows from Theorems 4.7 and 3.2 if we show that
a
b
tpui , uj q P Nσ | Du P P an infinite vertex of GP pIpBqqu “
tpc1 , c2 q P N2 | Dc3 P Z with pc1 , c2 , c3 q an infinite vertex of GN2 ˆZ pIpB̂qq.
By Lemma 2.12, it is enough to show that
tpui , uj q P Nσ | Du P P with ui and uj an infinite vertex of GP pBqu “
tpc1 , c2 q P N2 | Dc3 P Z with pc1 , c2 , c3 q an infinite vertex of GN2 ˆZ pB̂q.
(5.1)
We show Ď; the other inclusion is similar.
Let pui , uj q P Nσ , such that there is u P Nσ ˆ Zσ̄ (whose ith and jth coordinate are ui and uj )
that is an infinite vertex of GP pBq. By Lemma 3.4, there are v, ṽ P P connected to u such that
ṽ ´ v P P . Then there is a sequence of vertices u “ µp1q , . . . , µpℓ1 q “ v, µpℓ1 `1q , . . . , µpℓ2 q “ ṽ P P
such that pµpkq , µpk`1q q is an edge in GP pBq for k “ 1, . . . ℓ2 ´ 1.
Recall that B1 and B2 are the columns of B, and denote B̂1 and B̂2 the columns of B̂. Define
$
B̂1
if µpk`1q ´ µpkq “ B1 ,
’
’
&
´B̂1
if µpk`1q ´ µpkq “ ´B1 ,
v pkq “
’
B̂2
if µpk`1q ´ µpkq “ B2 ,
’
%
´B̂2
if µpk`1q ´ µpkq “ ´B2 .
Choose any c P Z and let ν p1q “ pui , uj , cq. Set also ν pk`1q “ ν pkq `v pkq for k “ 1, . . . , ℓ2 ´1. Then
the first and second coordinates of ν pkq are equal to the ith and jth coordinates of µpkq respectively,
which implies that ν p1q , . . . , ν pℓ2 q P N2 ˆ Z. By construction, pν pkq , ν pk`1q q is an edge of GN2 ˆZ pB̂q
for k “ 1, . . . , ℓ2 ´ 1, so that in particular, ν p1q , ν pℓ1 q and ν pℓ2 q belong to the same connected
component of GN2 ˆZ pB̂q. Moreover, ṽ´v P P implies that ν pℓ2 q ´ν pℓ1 q P N2 ˆZ, so by Lemma 3.4,
ν p1q is an infinite vertex of GN2 ˆZ pB̂q, and we conclude that pui , uj q belongs to the right hand side
of (5.1).
We use Theorem 5.1 to study hypergeometric differential equations. For the remainder of this
article, we work over the field k “ C.
Definition 5.2. Let B “ pbrs q P Znˆ2 as in Convention 3.1, and let κ P Cn . We use this information to define a system of partial differential equations in two variables, called a bivariate
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
15
hypergeometric system of Horn type, and denoted HornpB, κq, as follows. (Colored symbols are
added for convenience.)
´1 ˆ
ź br1
ź
˙
Bϕpz1 , z2 q
Bϕpz1 , z2 q
` br2 z2
` κr ´ ℓ “
br1 z1
Bz
Bz
1
2
br1 ą0 ℓ“0
˙
ˆ
r1 ´1
ź ´bź
Bϕpz1 , z2 q
Bϕpz1 , z2 q
` br2 z2
` κr ´ ℓ ,
z1
br1 z1
Bz1
Bz2
br1 ă0 ℓ“0
˙
´1 ˆ
ź br2
ź
Bϕpz1 , z2 q
Bϕpz1 , z2 q
br1 z1
` br2 z2
` κr ´ ℓ “
Bz
Bz
1
2
br2 ą0 ℓ“0
ˆ
˙
r2 ´1
ź ´bź
Bϕpz1 , z2 q
Bϕpz1 , z2 q
br1 z1
z2
` br2 z2
` κr ´ ℓ .
Bz1
Bz2
b ă0 ℓ“0
r2
Example 5.3. Hypergeometric systems contain as examples many widely studied systems of differential equations. For example, the Appell system F1 :
ˆ
˙
Bϕpz1 , z2 q
Bϕpz1 , z2 q
Bϕpz1 , z2 q
z1
` z2
`γ´1 “
Bz1
Bz1
Bz2
˙ˆ
˙
ˆ
Bϕpz1 , z2 q
Bϕpz1, z2 q
Bϕpz1 , z2 q
`β
z1
` z2
`α ,
z1
Bz1
Bz1
Bz2
ˆ
˙
Bϕpz1 , z2 q
Bϕpz1 , z2 q
Bϕpz1 , z2 q
z1
` z2
`γ´1 “
Bz2
Bz1
Bz2
ˆ
˙ˆ
˙
Bϕpz1 , z2 q
Bϕpz1 , z2 q
Bϕpz1, z2 q
1
z1
z1
`β
` z2
`α ,
Bz1
Bz1
Bz2
» 1 1fi
arises from the matrix B “ –
´1 ´1
1 0fl
0 1
´1 0
0 ´1
and the vector κ “ pγ ´ 1, ´α, 0, 0, ´β, ´β 1q P C6 . It is
clear that if α “ γ ´ 1, then the space of germs of holomorphic solutions of the Appell system
near a generic nonsingular point is infinite-dimensional, as any solution of z1 pBϕpz1 , z2 q{Bz1 q `
z2 pBϕpz1 , z2 q{Bz2 q ` α “ 0 is a solution of the Appell system. The converse of this statement,
though non-obvious, is also true. In fact, the necessary and sufficient constraints on the parameters
κ for the system HornpB, κq to have a finite dimensional solution space can be determined using
the primary decomposition of the lattice basis ideal IpBq (see [DMM10b, Theorem 6.2]). The
relevant object is the Andean arrangement of the lattice basis ideal IpBq (see Definition 5.4 below).
Recall from Convention 3.1 that we have been given a matrix A P Zpn´2qˆn such that AB “ 0,
and whose columns a1 , . . . , an span Zn´2 as a lattice. The ideal IpBq is homogeneous with respect
to the Zn´2 -grading on Crxs defined by degpxi q “ ai for i “ 1, . . . , n. If V is a Zn´2 -graded
Crxs-module, we define the set of degrees of V as follows:
degpV q “ tα P Zn´2 | Vα ‰ 0u.
Definition 5.4. Let V be a Zn´2 -graded module and consider degpV q as a subset of Cn´2 . The
Zariski closure of degpV q in Cn´2 is called the set of quasidegrees of V , denoted qdegpV q.
16
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
If A and B are as in Convention 3.1, the Andean arrangement of IpBq is the union over the Andean
components C of IpBq of qdegpCrxs{C q.
The Andean arrangement of IpBq is a finite union of translates of some faces of the cone Rą 0A of
non-negative combinations of the columns of A (see [DMM10b, Lemma 6.2]), so it is, indeed, an
arrangement of affine spaces.
The importance of this notion can be seen in the following result, which is a special case of
[DMM10b, Theorem 6.3].
Theorem 5.5. Let A and B as in Convention 3.1. Assume that the span of the columns of B
contains no nonzero coordinatewise non negative element (equivalently, the cone over the columns
of A contains no lines). The Horn system HornpB, κq has a finite dimensional solution space if
and only if Aκ lies outside the Andean arrangement of IpBq.
In fact, we can state a stronger result. A left ideal J in the ring Dℓ of linear partial differential
operators with polynomial coefficients in ℓ variables is holonomic if Extj pDℓ {J, Dℓ q “ 0 whenever
j ‰ ℓ. Holonomic ideals have finite dimensional solution spaces; the converse of this statement,
while untrue in general (see [BMW13, Example 9.1]), does hold for bivariate Horn systems.
Theorem 5.6 (Corollary 9.6 in [BMW13]). Let A and B be as in Convention 3.1. Assume that the
span of the columns of B contains no nonzero coordinatewise non negative element (equivalently,
the cone over the columns of A contains no lines). The Horn system HornpB, κq is holonomic if
and only if Aκ lies outside the Andean arrangement of IpBq.
As a consequence of Theorem 5.1, we are able to explicitly compute the Andean arrangement of
IpBq for a codimension two lattice basis ideal, thus providing a concrete description of the set
of parameters for which a Horn system in two variables is holonomic (and has finite dimensional
solution space). We first compute the quasidegree set for a single Andean component of IpBq.
Proposition 5.7. Let B be as in Convention 3.1, and suppose bi , bj , λ, d are as in Theorem 5.1.
Denote by C the component of IpBq associated to xxi , xj y. Write λ “ p{q where p and q are
relatively prime integers, and assume that the vector with ith coordinate p, jth coordinate q and
all other coordinates equal to zero is the first row of A. Then
ˆ
˙ !
(
Crxs
β P Cn´2 | β1 P t0, 1, . . . , pbi1 ` pbi2 ´ 1u r
if d{q “ 1,
( tpbi1 ` pbi2 ´ pu
qdeg
“
n´2
β
P
C
|
β
P
t0,
1,
.
.
.
,
pb
`
pb
´
1u
if d{q ą 1.
1
i1
i2
C
Proof. We start by justifying the assumption that the first row of A consists of zeros except for the
ith coordinate which equals p and the jth coordinate which equals q. We have a choice of A in
Convention 3.1: any A whose rows are a basis of the left kernel of B, and whose columns span Zd
as a lattice can be used. Since the row we want is an element of the left kernel of B, and its only
two nonzero entries are relatively prime, we are allowed to use it.
As bi and bj are linearly dependent, the submatrix of A with columns ak for k ‰ i, j has rank n´ 3.
Because the first row of this submatrix consists of all zeros, we see that, fixing natural numbers
ui , uj and varying uk P N for k ‰ i, j, the points Au are Zariski dense in the set tβ P Cn´3 | β1 “
pu1 ` qu2 u.
Now note that the degree α graded piece pCrxs{C qα equals 0 for α P Zn´2 if and only if Crxsα Ď
C ; also Crxsα is spanned by monomials, and the only monomials in C are those in M “
λpb `b ´dq
λpb `b ´2dq
xxj i1 i2 , xdi xj i1 i2
, . . . , xibi1 `bi2 ´d y.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
17
Thus,
(
u
qdegpCrxs{C q “ β P Cn´3 | β1 P tpui ` quj | ui , uj P N and xui i xj j R M u .
For fixed ℓ, consider the line Lℓ “ tpw, zq | pui ` quj “ ℓu. Given ℓ P N; we wish to know wether
or not all pui , uj q P N2 X Lℓ are exponent vectors of monomials in M . By construction of M
this occurs if ℓ ě pbi1 ` qbi2 . If ℓ ă pbi1 ` qbi2 , then Lℓ always contains an exponent vector of
a monomial not in M , unless d “ q, in which case the intersection N2 X Lppbi1 `bi2 ´1q consists of
exponent vectors of monomials in M .
Theorem 5.8. Let B be as in Convention 3.1, and let
S “ tti, ju | the rows bi and bj of B are linearly dependent in opposite open quadrants u.
For each σ “ ti, ju P S , assume i ă j and let λσ “ ´bj1 {bi1 “ ´bj2 {bi2 . Write λσ “ pσ {qσ
for pσ , qσ P N relatively prime. We can find hσ P Qn´2 such that hσ A has ith coordinate pσ , jth
coordinate qσ , and all other entries equal to zero. The Andean arrangement of IpBq equals
ˇ
)
!
ď !
ˇ
t0, 1, . . . , p |b | ` pσ |bi2 | ´ 1u r tpσ |bi1 | ` pσ |bi2 | ´ pσ u if dσ {qσ “ 1,
.
β P Cn´2 ˇhσ ¨ β P t0, 1, . . . , pσσ |bi1
if dσ {qσ ą 1.
i1 | ` pσ |bi2 | ´ 1u
σ“ti,juPS
Example 5.9. We apply the previous
» 1 1 firesult to the case Appell F1 , and compute the Andean ar-
rangement of IpBq for B “ –
´1 ´1
1 0fl
0 1 .
´1 0
0 ´1
Only the first two rows of B are linearly dependent in
2
opposite open quadrants of Z , and we can choose A “
„
1
0
0
1
1
0
0
0
0
1
0
0
0
0
1
0
0
1
0
1
0
0
1
1
. The Andean arrangement of
IpBq equals
tβ P C4 | β1 P t0, 1u r t1uu “ tβ P C4 | β1 “ 0u.
By Theorems 5.5 and 5.6, the system HornpB, κq for κ “ pγ ´ 1, α, 0, 0, ´β, ´β 1q is holonomic
if and only if it has a finite dimensional solution space, if and only if β1 “ γ ´ 1 ´ α “ 0, as we
claimed in Example 5.3.
At first glance, there is no obvious relationship between the Horn system HornpB, κq and the lattice
basis ideal IpBq, so it is unclear why the Andean arrangement of IpBq influences the behavior of
HornpB, κq. The key idea is due to Gelfand, Graev, Kapranov and Zelevinsky, who introduced a
system
B |u`| ψpy1 , . . . , yn q
pu q
pu q
“
B |u´ | ψpy1 , . . . , yn q
pu q
pu q
By1 ` 1 ¨ ¨ ¨ Byn ` n
By1 ´ 1 ¨ ¨ ¨ Byn ´ n
n
ÿ
ψpy1 , . . . , yn q
“ pAκqi ψpy1 , . . . , yn q
aij
By
j
j“1
for u a column of B,
for i “ 1, . . . n ´ 2,
denoted HpB, Aκq, and noticed that if ϕpz1 , z2 q is a solution of HornpB, κq then y κϕpy B1 , y B2 q is
a solution of HpB, Aκq (recall that B1 and B2 are the columns of B). Since HpB, Aκq contains
the generators of the lattice basis ideal IpBq as differential equations, it is natural that the behavior
of IpBq should impact HpB, Aκq. We illustrate this mechanism in an example.
” 2 4ı
Example 5.10. Let B “ ´2 ´4 . We use A “ r 1 1 0 s. The Andean arrangement of IpBq equals
1 1
the quasidegree set of the component corresponding to the associated prime xx1 , x2 y. In this case,
18
ZEKIYE SAHIN ESER AND LAURA FELICIA MATUSEVICH
λ “ 1 “ p “ q, and gcdp2{q, 4{qq “ 2, so we know that Aκ “ 2 ` 4 ´ 2 “ 4 implies that
HornpB, κq is not holonomic, and has infinite dimensional solution space. Let us exhibit infinitely
many linearly independent solutions for HornpB, p1, 3, 0qq.
It is easier to produce solutions if we start by considering HpB, 4q. Intuitively, the vertices in
a connected component of G pIpBqq that lies in the ℓth slice graph of IpBq, are the monomials
appearing in a series expansion of a solution of HpB, A ¨ pλℓ, 0, 0qq “ HpB, ℓq. Since their
monomial sets are disjoint, solutions arising from different connected components are linearly
independent. Each slice graph contains infinitely many vertices, so the only way this method
produces a finite dimensional solution space is if the connected components contain infinitely
many vertices (and thus give rise to solutions that are infinite series). If the connected components
of a given slice graph are all finite, there must be infinitely many of them, and this provides an
infinite dimensional subspace of the solution space of HpB, ℓq.
More concretely, for ℓ “ 4 and any k P N, tp1, 3, kq, p3, 1, k ` 1qu is a connected component
of GSp4q pIpBqq. We propose a solution of HpB, 4q which is a linear combination of y1 y23y3k and
y13 y2 y3k`1. Solving for the coefficients, we see that fk py1 , y2, y3 q “ pk ` 1qy1y23 y3k ` y13 y2 y3k`1 is a
solution of HpB, 4q.
We know that if we can write f py1 , y2 , y3q “ y1 y23 F py12y3 {y22 , y14y3 {y24 q, and f py1 , y2, y3 q is a solution of HpB, 4q, then F pz1 , z2 q is a solution of HornpB, p1, 3, 0qq. Using the expression y3 “
py12 y3 {y22q2 {py14 y3 {y24q, we see that fk py1 , y2 , y3 q “ y1 y23 Fk py12y3 {y22 , y14 y3 {y24q, where Fk pz1 , z2 q “
pk ` 1qpz12 {z2 qk p1 ` z1 q. The functions Fk pz1 , z2 q span an infinite dimensional subspace of the
solution space of HornpB, p1, 3, 0qq. Note that in this example, there is no obvious common factor
in the equations HornpB, p1, 3, 0qq, as was the case for the Appell system F1 .
R EFERENCES
[BMW13]
[DMM10a]
[DMM10b]
[DMS05]
[DSS]
[ES98]
[HS00]
[MM82]
[M11]
[MS]
[SS10]
Christine Berkesch-Zamaere, Laura Felicia Matusevich and Uli Walther, Torus equivariant Dmodules and hypergeometric systems. Preprint, 2013. arXiv:1308:5901.
Alicia Dickenstein, Laura Felicia Matusevich and Ezra Miller, Combinatorics of binomial primary
decomposition, Math. Z. 264 (2010), no. 4, 745–763.
Alicia Dickenstein, Laura Felicia Matusevich and Ezra Miller, Binomial D-modules, Duke Math. J.
151 (2010), no. 3, 385–429.
Alicia Dickenstein, Laura Felicia Matusevich and Timur Sadikov, Bivariate hypergeometric Dmodules, Adv. Math. 196 (2005), no. 1, 78–123.
Mathias Drton, Bernd Sturmfels and Seth Sullivant, Lectures on Algebraic Statistics, Oberwolfach
Seminars, 40. Birkhauser, Basel, 2009.
David Eisenbud and Bernd Sturmfels, Binomial ideals. Duke Math. J. 84 (1996), no. 1, 1–45.
Serkan Hosten and Jay Shapiro, Primary decomposition of lattice basis ideals, Symbolic computation
in algebra, analysis, and geometry (Berkeley, CA, 1998). J. Symbolic Comput. 29 (2000), no. 4–5,
625–639.
Ernst Mayr and Albert Meyer, The complexity of the word problem for commutative semigroups and
polynomial ideals, Adv. in Math. 46 (1982), no. 3, 305–329.
Ezra Miller, Theory and applications of lattice point methods for binomial ideals, Combinatorial
aspects of commutative algebra and algebraic geometry, 99–154, Abel Symp., 6, Springer, Berlin,
2011.
Ezra Miller and Bernd Sturmfels. Combinatorial Commutative Algebra, Graduate Texts in Mathematics, 227. Springer-Verlag, New York, 2005.
Anne Shiu and Bernd Sturmfels, Siphons in chemical reaction networks Bull. Math. Biol. 72 (2010),
no. 6, 1448–1463.
PRIMARY COMPONENTS OF CODIMENSION TWO LATTICE BASIS IDEALS
D EPARTMENT OF M ATHEMATICS , T EXAS A&M U NIVERSITY , C OLLEGE S TATION , TX 77843.
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , T EXAS A&M U NIVERSITY , C OLLEGE S TATION , TX 77843.
E-mail address: [email protected]
19
| 0math.AC
|
Structural Learning for Template-free Protein
Folding
By
Feng Zhao
Submitted to:
Toyota Technological Institute at Chicago
6045 S. Kenwood Ave, Chicago, IL, 60637
For the degree of Doctor of Philosophy in Computer Science
Thesis Committee:
Jinbo Xu (Thesis Supervisor)
David McAllester
Jie Liang
Tobin Sosni
Sosnic
snick
i
Structural Learning for Template-free Protein
Folding
By
Feng Zhao
Submitted to:
Toyota Technological Institute at Chicago
6045 S. Kenwood Ave, Chicago, IL, 60637
August 2013
For the degree of Doctor of Philosophy in Computer Science
Thesis Committee:
Jinbo Xu (Thesis Supervisor) Signature:
Date: 10/25/2013
David McAllester
Signature:
Date: 10/25/2013
Jie Liang
Signature:
Date: 10/25/2013
Tobin So
Sosnic
snick
Signature:
Date: 10/25/2013
ii
iii
Abstract
The thesis is aimed to solve the template-free protein folding problem by tackling two
important components: efficient sampling in vast conformation space, and design of
knowledge-based potentials with high accuracy.
We have proposed the first-order and second-order CRF-Sampler to sample structures from
the continuous local dihedral angles space by modeling the lower and higher order conditional
dependency between neighboring dihedral angles given the primary sequence information. A
framework combining the Conditional Random Fields and the energy function is introduced
to guide the local conformation sampling using long range constraints with the energy
function.
The relationship between the sequence profile and the local dihedral angle distribution is
nonlinear. Hence we proposed the CNF-Folder to model this complex relationship by applying
a novel machine learning model Conditional Neural Fields which utilizes the structural
graphical model with the neural network. CRF-Samplers and CNF-Folder perform very well in
CASP8 and CASP9.
Further, a novel pairwise distance statistical potential (EPAD) is designed to capture the
dependency of the energy profile on the positions of the interacting amino acids as well as the
types of those amino acids, opposing the common assumption that this energy profile depends
only on the types of amino acids. EPAD has also been successfully applied in the CASP 10 Free
Modeling experiment with CNF-Folder, especially outstanding on some uncommon
structured targets.
iv
Acknowledgements
It was when I started to write this acknowledgement that I realize I owe my gratitude to many
people who have made this thesis possible.
My deepest gratitude is to my advisor, Jinbo Xu, whose guidance, support, advice and long
discussions have enabled me to gain a profound understanding throughout the PhD studies. It
was from him that I have learned the spirit of research, which I believe will be beneficiary to
my whole life, and that I will cherish forever.
I would like to thank my committee members David McAllester, Tobin Sosnick and Jie Liang
for all their support and suggestions. My heartfelt thanks are also dedicated to all the
professors from TTI-C and University of Chicago who have taught me, David McAllester, Nati
Srebro, Tamir Hazan, Raquel Urtasun, Julia Chuzhoy, Mattias Blume, Karen Livescu, Greg
Shakhnanovich, Anastasios Sidiropoulos, Yury Makarychev, Yang Shen, Laszlo Babai, Tobin
Sosnick, Paul J. Sally, and Gregory F. Lawler for their encouraging and sharing their plethora
of knowledge on various subjects and courses.
It has been a pleasure to study at TTI-C during the years. My fellow students are cordially
thanked for their friendship: Jian Peng, Karthik Sridharan, Andrew Cotter, Zhiyong Wang,
Avleen Bijral, Payman Yadollahpour, Jianzhu Ma, Sheng Wang, Joe DeBartolo, Aashish
Adhikari, Raman Arora, Qingming Tang, Taewhan Kim, Jian Yao, Hao Tang, Somaye
Hashemifar and Behnam Tavakoli Neyshabur. Thank you for your good company and
support.
Research reported in this thesis was supported by the National Science Foundation grant DBI0960390 and National Institutes of Health grant R01GM089753. The content is solely the
responsibility of the authors and does not necessarily represent the official views of the NSF or
NIH.
This work was made possible by the computing resources and facilities of TeraGrid, Beagle,
SHARCNET and Open Science Grid (OSG).
Finally, special thanks to my wife Wentao, my parents and my beloved Ayla. It was because of
them that my work becomes meaningful.
v
Contents
Abstract.................................................................................................................................................. iv
Acknowledgements .............................................................................................................................. v
Chapter 1. Introduction ........................................................................................................................... 1
1.1 Conformation Sampling ................................................................................................................. 3
1.2 Energy function ............................................................................................................................... 8
1.3 Contribution..................................................................................................................................... 9
1.4 Organization .................................................................................................................................. 10
Part 1. ....................................................................................................................................................... 12
Probabilistic and Continuous Models of Protein Conformational Space for Fragment-free and
Template-Free Modeling ....................................................................................................................... 12
Chapter 2. Continuous representation of protein conformations ................................................... 13
2.1 Cα-trace representation ................................................................................................................. 13
2.2 Distribution of bond angles ......................................................................................................... 14
2.3 Building backbone atoms............................................................................................................. 15
2.4 Mathematical Symbols ................................................................................................................. 15
Chapter 3. First-order CRF-Sampler ................................................................................................... 18
3.1 Introduction ................................................................................................................................... 18
3.2 CRF model for sequence-structure relationship ....................................................................... 19
3.3 Conformation sampling algorithm............................................................................................. 26
3.4 Experiments and Results.............................................................................................................. 27
Data set .............................................................................................................................................. 28
Label assignment and distribution parameters ........................................................................... 28
Parameter tuning.............................................................................................................................. 28
Comparison with the HMM model ............................................................................................... 29
Comparison with Xia et al ............................................................................................................... 35
Comparison with Rosetta ............................................................................................................... 37
Computational efficiency ................................................................................................................ 38
vi
3.5 Conclusion ..................................................................................................................................... 41
Chapter 4. Second-order CRF-Sampler with Energy ........................................................................ 43
4.1 Introduction ................................................................................................................................... 43
4.2 Methods .......................................................................................................................................... 43
Model parameter training ............................................................................................................... 47
Conformation sampling and resampling ..................................................................................... 48
Energy function ................................................................................................................................ 49
Energy minimization ....................................................................................................................... 51
4.3 RESULTS ........................................................................................................................................ 52
Comparison with FB5-HMM and fragment assembly ............................................................... 55
Comparison with lattice model...................................................................................................... 57
4.4 Conclusion ..................................................................................................................................... 65
Chapter 5. CNF-Folder: Modeling with Nonlinear Features........................................................... 66
5.1 Introduction ................................................................................................................................... 66
5.2 Methods .......................................................................................................................................... 66
A 2nd-order CNF model of conformation space ......................................................................... 66
Model parameter training ............................................................................................................... 70
Conformation sampling and resampling ..................................................................................... 71
Replica exchange Monte Carlo simulation................................................................................... 71
5.3 Results ............................................................................................................................................. 72
Performance on the old test set ...................................................................................................... 72
Performance on the CASP8 test set. .............................................................................................. 73
Comparison with CASP8 models. ................................................................................................. 81
5.4 Conclusion ..................................................................................................................................... 82
Part 2. ....................................................................................................................................................... 85
Statistical Potentials Using Machine Learning Methods.................................................................. 85
Chapter 6. EPAD: A position-specific pairwise distance potential ................................................ 86
6.1 Introduction ................................................................................................................................... 86
6.2 Methods and Experimental Procedures .................................................................................... 87
Estimating pairwise Cα distance distribution using probabilistic neural network (PNN). .. 88
vii
Model parameter training. .............................................................................................................. 92
Training and validation data. ......................................................................................................... 93
Estimating inter-atom distance distribution for non-Cα main chain atoms. ........................... 93
Estimating the reference state. ....................................................................................................... 93
Window size and the number of neurons in the hidden layers. ............................................... 94
6.3 Results and Discussion ................................................................................................................. 98
Distance dependence of the statistical potentials. ....................................................................... 98
Performance on decoy discrimination. ....................................................................................... 100
Performance on the 2007 Rosetta dataset. .................................................................................. 100
Performance on template-based models. ................................................................................... 101
Performance on the CASP9 models............................................................................................. 104
Performance on the Decoy ‘R’ Us dataset. ................................................................................. 104
Performance on the I-TASSER dataset. ....................................................................................... 105
Performance on the CASP5-8 dataset. ........................................................................................ 108
Is our probabilistic neural network (PNN) model over-trained? ........................................... 108
Folding Results with EPAD in the blind CASP 10 experiment ............................................... 111
6.4 Conclusion ................................................................................................................................... 116
Chapter 7. Conclusion and Future works ........................................................................................ 117
7.1 Conclusions .................................................................................................................................. 117
7.2 Future Works ............................................................................................................................... 117
References ............................................................................................................................................. 119
viii
List of Figures
Figure 1 Yearly growth of total structures in RCSB Protein Data Bank (PDB). .............................. 2
Figure 2 The growth of the SwissProt database. ................................................................................. 2
Figure 3 The 6 Ramanchandran basins: β (blue), poly-proline II, PPII (green), αR (red), αL
(magenta) and ε (grey). Color intensity reflects the ϕ, ψ occupancy. Adapted from (Colubri et
al., 2006) ..................................................................................................................................................... 5
Figure 4 Backbone ϕ, ψ -space subdivided into 36 alphabetically labeled, 60° × 60° grid regions.
Adapted from (Gong et al., 2005) .......................................................................................................... 5
Figure 5 The , , dihedral angles in one residue of the protein backbone. ω can be assumed
to be fixed at 180° (trans) or 0° (cis). .................................................................................................... 13
Figure 6 Three point sets sampled from the FB5 distribution on the unit sphere. ....................... 16
Figure 7 An example CRF model for protein conformation sampling. ......................................... 23
Figure 8 Native structures (in orange) and the best decoys of 1FC2, 1ENH, 2GB1, 2CRO, 1CTF,
and 4ICB .................................................................................................................................................. 33
Figure 9 A second-order Conditional Random Fields (CRF) model of protein conformation
space. Each backbone angle state depends on a window (size 9) of sequence profiles and
secondary structure and also the states in its local neighborhood. ................................................ 46
Figure 10 A 1st-order CNF model consists of three layers: input, output and hidden layer...... 69
Figure 11 The relationship between RMSD (y-axis) and energy (x-axis) for T0397_D1,
T0416_D2, T0476, T0482, T0496_D1 and T0510_D3. ......................................................................... 79
Figure 12 Two typical mirror images generated by the CRF method for T0416_D2.. ................. 80
Figure 13 Ranking of our CNF predictions for T0416_D2, T0476, T0496_D1 and T0510_D3. .... 84
Figure 14 Distribution of pairwise
distances in a representative set of protein structures... 89
Figure 15 An example probabilistic neural network, in which and are the sequence profile
contexts centered at the ith and jth residues, respectively. 1 and 2 are the neurons in the 1st
and 2nd hidden layers. ............................................................................................................................ 92
Figure 16 Inter-residue contact prediction accuracy of our PNN and the top 12 (human and
server) groups in CASP8 and CASP9.................................................................................................. 96
Figure 17 Distance dependence of DOPE and our potential EPAD.. ............................................. 99
Figure 18 . RaptorX-roll’s Result on targets T0734 (A) and T0740 (B). ......................................... 115
ix
List of Tables
Table 1 Some Mathematical Symbols Used in the CRF Models ..................................................... 17
Table 2 F1-Values (%) of CRF-Sampler with Respect to Window Size .......................................... 25
Table 3 Decoy Quality Comparison between the HMM Model and CRF-Sampler ..................... 31
Table 4 Decoy Quality Comparison between the HMM Model17 and CRF-Sampler ................. 32
Table 5 Secondary Structure Content of Good Decoys .................................................................... 34
Table 6 Decoy Quality Comparison Between Xia et al. (Xia et al., 2000) and CRF-Sampler ...... 36
Table 7 Percentage of correct secondary structure (Q3-value) and secondary structure scntent
of all the decoys generated by CRF-Sampler, compared with PSIPRED Predictions.................. 37
Table 8 Decoy Quality Comparison Between ROSETTA2 and CRF-Sampler ............................. 39
Table 9 Approximate Running Time in Minutes Spent by CRF-Sampler in Generating 100
Decoys...................................................................................................................................................... 40
Table 10 Quality Comparison of the Decoys Generated by the 1st-Order and 2ND-Order CRF
Models ..................................................................................................................................................... 53
Table 11 Quality Comparison of the Decoys Generated by FB5-HMM, the 1st-Order CRF, the
2nd-Order CRF, and Rosetta .................................................................................................................. 56
Table 12 Performance Comparison between CRFFolder and Skolnick's TOUCHSTONE-II .... 59
Table 13 Performance Comparison (TM-Score) between CRFFolder and Robetta on Some
CASP8 Hard Targets.............................................................................................................................. 61
Table 14 Summarized results of RAPTOR++ predictions of free modeling targets in CASP8. .. 64
Table 15 Secondary structure prediction accuracy comparison with the current top predictors.
................................................................................................................................................................... 69
Table 16 Performance of the CNF and CRF methods on the old test set. ...................................... 74
Table 17 Performance of our CNF and CRF methods on the CASP8 test set................................ 76
Table 18 Clustering result of the 12 CASP8 free-modeling targets. ............................................... 83
Table 19
−
distance prediction accuracy with respect to the window size....................... 97
Table 20 Performance of EPAD and several popular full-atom statistical potentials on the
Rosetta decoy set. ................................................................................................................................. 102
Table 21 Performance of EPAD and several popular statistical potentials on the Rosetta decoy
sets when only Cα atoms are considered. ......................................................................................... 102
x
Table 22 Performance of EPAD and several popular statistical potentials on the template-based
models.................................................................................................................................................... 103
Table 23 Performance of statistical potentials with respect to the hardness of the CASP9
targets. ................................................................................................................................................... 106
Table 24 Performance of several statistical potentials on the Decoy ‘R’ Us dataset................... 106
Table 25 Performance of several statistical potentials on the I-TASSER dataset. ....................... 107
Table 26 Performance of a variety of potentials on the selected CASP5-8 models. ................... 109
Table 27 Folding comparison with Rosetta (GDT) on CASP9 FM targets ................................... 113
Table 28 Comparison with threading method RaptorX-ZY on CASP10 FM targets ................. 114
xi
Chapter 1. Introduction
As shown in Figure 1, the number of solved 3D protein structures in RCSB Protein Data Bank
(PDB) increases to 84223 as of Aug. 28th, 2012. On the other hand, the genome sequencing
projects have led to the identification of tens of millions of protein sequences publicly available
in NCBI Non-redundant database. The high quality annotated and non-redundant protein
sequence database SwissProt contains 536789 sequences at Release 2012_07 (Figure 2). The fact
that so many protein sequences are part of the public knowledge, and that so many of them
have structures that we have not yet solved, means that we have a long way to go to
understand the way that proteins work and function within the body. Many computational
methods have been developed to predict the structure of a protein from its primary sequence,
based on the famous Anfinsen dogma that all of the information necessary for a protein to fold
to the native structure resides in the protein’s amino acid sequence(Anfinsen, 1973). These
methods can be roughly classified into two categories: template-based and template-free
modeling.
Despite considerable progress taken place over the past decade, template free modeling
remains one of the unsolved mysteries in the field of computational structural biology. The
primary challenges still remain in two areas: the vast conformation space to be searched and
limited accuracy of the current energy functions designed. The purpose of this thesis is to
explore these topics in two sections: first, protein conformation sampling, that is, the
exploration of conformational space that corresponds with a particular protein sequence;
second, Design of energy function, that is, an accurate physics-based or knowledge-based
potential to quantify interactions among residues or atoms.
1
Thousands
90
80
70
60
50
40
30
20
10
0
Yearly
Total
Figure 1 Yearly growth of total structures in RCSB Protein Data Bank (PDB).
Adapted from http://www.rcsb.org/pdb/statistics/contentGrowthChart.do?content=total
Figure 2 The growth of the SwissProt database.
Adapted from http://web.expasy.org/docs/relnotes/relstat.html
2
1.1 Conformation Sampling
Template-free modeling that comes from fragment assembly (Bowie and Eisenberg, 1994;
Claessens et al., 1989; Jones and Thirup, 1986; Levitt, 1992; Simon et al., 1991; Sippl, 1993;
Unger et al., 1989) and lattice-models (Kihara et al., 2001; Wendoloski and Salemme, 1992; Xia
et al., 2000; Zhang et al., 2003) have been studied extensively. These two common methods and
their combined use in template-free modeling have brought impressive results in CASP
(Critical Assessment of Structure Prediction) competitions (Moult, 2005; Moult et al., 2007;
Moult et al., 2005; Moult et al., 2003). The popular fragment assembly program Robetta
(Misura et al., 2006; Simons et al., 1997) has the best track records in all of the template-free
modeling programs. Both TASSER (Zhang and Skolnick, 2005a) and its derivative ZhangServer (Wu et al., 2007a) have garnered recognition for stellar results in both CASP7 and
CASP8 through the combination of threading-generated fragments, distance restraints, and
lattice modeling.
Fragment based protein structure prediction takes place in two distinct steps. First, it is
necessary to cut a protein sequence into minuscule segments and then pick out literally scores
of possible fragments that might construct each segment. Then, one uses a simulation or search
algorithm to build the protein structure. While it is important to note that the results attained
by existing template-free modeling methods are exciting, there are several areas of concern
that have not yet been resolved. One of these has to do with the small pool of proteins with
solved experimental structures within the PDB. This makes it almost impossible to put
together a collection of fragments that can match the number of ways a protein can locally
conform – particularly in the loop regions, where the possibilities grow. It is true that a new
fold may entirely consist of rarely occurring super-secondary structure motif that is found
nowhere else in the protein data bank. Second, because the conformation space that a lattice
model or fragment library defines is by nature discrete, it may even keep the original fold from
being searched since a slight change in backbone angles, even in the most minuscule degree,
can result in a totally different fold.
It is possible to make the conformational space continuous, as Bystroff et al. have shown with
HMMSTR (Bystroff et al., 2000), which is a Hidden Markov Model (HMM) model trained from
a fragment library, to generate local conformations for a given sequence. Both Colubri et al.
(Colubri et al., 2006) and Gong et al. (Gong et al., 2005) have have analyzed the skeletal
structure of protein, to see if that structure can be reconstructed – under the assumption that
the angles are limited to the degrees allowed in their original Ramachandran basins ( a native
Ramachandran basin is the region that holds the values of the native backbone ϕ, ψ torsion
angle pairs). Sosnick and coworkers (Colubri et al., 2006) divided the torsion angle space into
six distinct sections (see Figure 3) while Gong et al. (Gong et al., 2005) split the space into 36
3
distinct regions (see Figure 4). The upshot from both pieces of research is that it is possible to
reconstruct the backbone structure of many of the smaller proteins with good accuracy, if the
angles are limited to those native basins. These studies do not consist of pure ab initio folding;
instead, they show that if researchers can come up with a reasonably close guess as to each
angle's native basin, then it should also be possible to forecast with accuracy a protein's
backbone structure.
Fragment-HMM (Li et al., 2008), which is a variant of Robetta, has the ability to sample
conformations within a continuous space. However, because Hidden Markov model is
constructed from 9-mer fragments, it still has the issue of coverage that plagues many of the
other analytical machines. The TOUCHSTONE programs (Kihara et al., 2001; Zhang et al.,
2003) , which uses the lattice model, does not have the coverage issue. but its sampling of
protein conformations comes from a three-dimensional lattice that has only finite resolution.
More importantly, these conformations that get sampled might not even have a local structure
that resembles those of native proteins, because TOUCHSTONE does not sample a
conformation based upon the primary sequence of a protein. Rather, these applications take
several statistical potentials over short range in the energy function to manage the creation of a
local structure that closely resembles the desired protein(s).
It is worth mentioning that some other methods exist for trying to sample conformations of
protein in a continuous space using the probability. For each conformation, the probability is
an approximation to its overall stability. Sequence information is used to estimate those
probabilities. In response, Feldman and Hogue developed the FOLDTRAJ (Feldman and
Hogue, 2002) program, which implements a probabilistic all-atom conformation sampling
algorithm. FOLDTRAJ is tested on three small proteins 1VII, 1ENH, and 1PMC, with 100,000
decoys for each protein. The best models achieve 3.95, 5.12 and 5.95 Angstroms in term of
RMSD from the native structures. However, because FOLDTRAJ does not use the nearest
neighbor effects or the sequence profile in modeling the relationship between sequence and
structure, it is not able to generate models that match the quality of the models from the
popular fragment assembly Rosetta application.
4
Figure 3 The 6 Ramanchandran basins: β (blue), poly-proline II, PPII (green), αR (red), αL
(magenta) and ε (grey). Color intensity reflects the ϕ, ψ occupancy. Adapted from (Colubri et
al., 2006)
Figure 4 Backbone ϕ, ψ -space subdivided into 36 alphabetically labeled, 60° × 60° grid
regions. Adapted from (Gong et al., 2005)
5
Hamelryck et.al have developed two different HMM models FB5-HMM (Hamelryck et al.,
2006) and TorusDBN (Boomsma et al., 2008) to sample protein conformations from a
continuous space. These models not only capture the relationship between backbone angles
and primary sequence, but also consider the angle-dependency between two adjacent residues.
FB5-HMM uses a Hidden Markov Model to discover the native basins of the local θ, τ
pseudo backbone dihedral angles of a protein sequence by learning the local sequencestructure relationship; and TorusDBN learns the local ϕ, ψ backbone dihedral angle
dependency using a dynamic Bayesian network, which is a generalization of the Hidden
Markov Model. There are two particular advantages of these methods: first, they model the
backbone angles using a directional statistics distribution (Kent, 1982; Mardia et al., 2007;
Singh et al., 2002) so that they can be sampled from a continuous space; second, they can find
how two adjacent backbone angles are dependent on one another. The experiments shows that
by modeling the dependency between two adjacent positions, FB5-HMM can generate more
conformations that are close to the native structure than the applications that do not model the
correlation between neighboring residue positions.
There are quite a few fragment assembly methods (Kent, 1982; Mardia et al., 2007; Singh et al.,
2002; Tuffery and Derreumaux, 2005) that also exploit the dependency between the two
adjacent fragments. As a result, researchers have learned that this dependency is one of the
most important things of conformation sampling. They also demonstrated that Torus-DBN
can generate local conformations as accurately as the fragment assembly method (Boomsma et
al., 2008). However, it is important to note that these models only consider angle-dependency
among two neighboring residues. Because of this restriction, these models cannot make use of
more enriched forms of the sequence information, such as PSI-BLAST sequence profile or the
threading-generated constraints. Hence their sampling efficiency is quite limited. Furthermore,
these models have not been considered in the context of template-free modeling in the real
world yet.
Due to the expressivity of the HMM model, FB5-HMM assumed that each residue is
independent of its secondary structure type, and each backbone angle only depends on its
corresponding residue (monomer, or 1-mer) and secondary structure at the same position,
though it actually depends on at least three neighborhood residues (3-mer). Zhao et al (2008,
2009) implemented an extensible protein conformation sampler, CRF-Sampler (Zhao et al.,
2008; Zhao et al., 2009), based on a probabilistic graphical model Conditional Random Fields.
CRF-Sampler incorporates the dependence among up to 9 neighboring residues and sequence
profile information with continuous distribution into the backbone angle model.
The application CRF-Sampler is a protein conformation sampler that is extensible and has been
built using a discriminative learning method by modeling the dependence of the unobserved
variables upon observed variables in the forms of the conditional probability distribution
6
P(backbone angle | protein sequence) to predicting θ, τ from the sequence information. CRFSampler can learn over 10,000 different parameters that quantify relationships among the
backbone angles, the primary sequence, and the secondary structure. By using just the selfavoiding constraints and compactness, CRF-Sampler can simulate conformations from the
existing primary sequence and the predicted secondary structure. CRF-Sampler also exhibits
high flexibility in that a variety of model topologies and feature sets can be defined to model
the sequence-structure relationship without worrying about parameter estimation.
Experimental results show that the first-order CRF-Sampler using a small collection of features
can generate decoys with much higher quality than the FB5-HMM model.
The second-order Conditional Random Fields (CRFs) model has the capability of portraying
more complicated levels of dependency among the dihedral angles in the local sequences.
When coupled with a simple energy function, this probabilistic method compares favorably
with the fragment assembly method in the blind CASP8 evaluation, especially on alpha or
small beta proteins. When the second-order CRF-Sampler is combined with a simple energy
function with 3 terms including a distance potential, a hydrogen potential and a hydrophobic
potential, it generates superior outcomes that compare favorably with the fragment assembly
method in the blind CASP8 evaluation, especially on alpha or small beta proteins. This is the
first probabilistic method that can search conformations in a continuous space and achieves
favorable performance. In fact, this is the very first method with the ability to search within a
continuous conformation space and also generate favorable decoys. In addition, this method
also created three dimensional models that were more accurate than template-based methods
for several of the hard targets in CASP8. The second-order CRF-Sampler can also be applied to
protein loop modeling, model refinement, and even RNA tertiary structure prediction.
The second-order CRF-Sampler is used in RAPTOR++ in CASP8 to predict the 3D structures of
the hard targets. RAPTOR++ has four different modules: threading, model quality assessment,
multiple protein alignment and template-free modeling, i.e. the second-order CRF-Sampler.
RAPTOR++ first tries to connect a target protein to the templates using the first three modules.
Next, it predicts the quality of the 3D model implied by each alignment using a model quality
assessment method. If all the alignments come back with insufficient quality, RAPTOR++ uses
the second-order CRF-Sampler. CRF-Sampler is also used to refine template-based models.
The CRF models describe the relationships among input features and outputs by using linear
potential functions. The difficulty emerges because this relationship is frequently nonlinear,
complex beyond the capacity of the model. To take advantage of both the structured graphical
models and non-linear classifiers such as SVM and neural networks, Zhao et al. proposed
CNF-Folder, a fragment-free approach to protein folding based on Conditional Neural Fields
(CNF) (Peng et al., 2009; Zhao et al., 2010). CNF-Folder extends CRF-Sampler by adding a
middle layer between input features and outputs. The middle layer consists of a series of
7
hidden gates, each with its own parameter, and each acts as a local neuron (i.e. feature
extractor) to capture the non-linear relationship between input features and outputs. Based on
experimental results, CNF-Folder can create superior decoys on a variety of test proteins
including the CASP8 free-modeling targets when it is used with replica exchange simulation
and same energy function as used by CRF-Sampler. The most impressive results that CNFFolder can forecast include the correct fold for T0496_D1, which is one of the two CASP8
targets that have a truly new fold. For T0496, the predicted model is also considerably superior
to all of the CASP8 models.
1.2 Energy function
To express the function and structure of a protein, one often needs the ability to give a
quantifiable value to evaluate the interactions among residues or atoms. This quantifiable
value is called potential which is created either based on physics rules or extracted from
knowledge base. There are studies (Bradley et al., 2005; Skolnick, 2006) indicating that
knowledge-based statistical potentials (Li and Liang, 2007; Lu et al., 2008; Miyazawa and
Jernigan, 1985; Shen and Sali, 2006; Simons et al., 1997; Sippl, 1990; Tanaka and Scheraga, 1976;
Zhang and Zhang, 2010; Zhou and Zhou, 2002) compare favorably to physics-based potentials
(Brooks et al., 2009; Bryngelson et al., 1995; Case et al., 2005; Dill, 1985, 1997; Dobson et al.,
1998; Schuler et al., 2001; Shakhnovich, 2006) in many applications including ab initio folding
(Jones and Thirup, 1986; Kachalo et al., 2006; Kihara et al., 2001; Levitt, 1992; Simons et al., 1997;
Wu et al., 2007a; Zhao et al., 2010), docking (Zhang et al., 1997), binding (Kortemme and Baker,
2002; Laurie and Jackson, 2005), mutation study (Gilis and Rooman, 1996, 1997), decoy ranking
(Bauer and Beyer, 1994; Casari and Sippl, 1992; Gatchell et al., 2000; Hendlich et al., 1990;
Samudrala and Moult, 1998; Simons et al., 1999; Vendruscolo et al., 2000) and protein model
quality assessment (Jones and Thornton, 1996; Panchenko et al., 2000; Peng and Xu, 2010; Reva
et al., 1997; Sippl, 1993). Knowledge-based statistical potentials extract interactions from the
solved protein structures in the Protein Data Bank (PDB) (Berman et al., 2000). They are more
user-friendly than physics-based potentials in terms of user API and calculation complexity.
Many Knowledge-based statistical potentials have been created, including the popular DOPE
(Shen and Sali, 2006) and DFIRE (Zhou and Zhou, 2002). Some statistical potential quantify
local atomic interactions (e.g., torsion angle potential) while others capture non-local atomic
interactions (e.g., distance-dependent potential).
Even after extensive study though, designing protein potential with significant precision
remains highly challenging. A lot of knowledge-based statistical potentials are derived from
the inverse of the Boltzmann law. They have two primary components: reference state and
observed atomic interacting probability. The observed atomic interacting probability is usually
assumed to correlate with only atom types, and is estimated through a similar simple counting
8
method. The reference state can be derived theoretically based on physical rules or be
estimated statistically from the empirical dataset. Many potentials are distinguished in the
construction of the reference state.
Zhao et al. (2012) designed the statistical potential EPAD (Evolutionary PAirwise Distance
potential) by making use of protein evolutionary information, which has not been used by
currently popular statistical potentials (e.g., DOPE and DFIRE). EPAD is unique in that is has
different energy profiles for the same type of atom pairs, depending on their sequence
positions. Experiments confirm that this position-specific statistical potential EPAD
outperforms several popular statistical potentials in both decoy discrimination and ab initio
folding. Overall, the results suggest several implications: 1. statistical potentials that are
protein-specific and position-specific work more effectively; 2. evolutionary information
makes energy potentials perform more effectively; 3. observed probability and reference state
both make energy potentials different; 4. context-specific information improves the estimation
of observed probability.
1.3 Contribution
The thesis has several contributions to science discovery as following.
First, in protein folding, the local dihedral angles are from a continuous space. A first-order
CRF-Sampler is proposed to model the conditional dependency between two neighboring
dihedral angles given the primary sequence information.
Second, the 3D structure of protein depends on higher order relationships among the
constructing amino-acids (residues) in every neighborhood, as well as the long range
interactions among the residues at the positions that are far from each other in the primary
sequence. A framework combining the second-order Conditional Random Fields and the
energy function is introduced to guide the local conformation sampling using long range
constraints with the energy function.
Third, it is a nonlinear relationship between the sequence profile and the local dihedral angle
distribution. A novel machine learning model Conditional Neural Fields which utilizes the
structural graphical model with the neural network is applied to model this complex
relationship in CNF-Folder, a powerful fragment-free protein conformation sampler.
Fourth, the energy profiles of the pairwise distance potential in the proteins depend on the
positions of the interacting amino acids in the primary sequence as well as the types of those
amino acids opposing the common assumption that this energy profile depends only on the
9
types of amino acids. A new probabilistic neural network model is proposed to estimate the
observed probability of the pairwise distances, and a novel statistical distance potential (EPAD)
is designed for protein structure evaluation and functional studies.
The demand for using these tools already exists. We have developed software packages of the
CNF-Folder and EPAD for the users to download.
1.4 Organization
This thesis is essentially a collection of original published papers in the area of template-free
protein folding and ranking. Each chapter is written in such way that it can be read
independently.
Chapter 2 introduces the basic concepts on continuous representation of protein conformation.
The notations are used in the subsequent chapters.
Chapter 3 presents the first order Conditional Random Fields model for continuous protein
conformation sampling. With the help of the continuous representations defined in Chapter 2,
Chapter 3 presents the CRF model to learn the relationship between the sequence information
and the local backbone dihedral angles. An efficient forward back-track conformation
sampling algorithm is also presented. The performance of the first order CRF-Sampler is
evaluated against several protein simulation approaches including the most popular Rosetta
method.
Chapter 4 describes the higher dependency among the dihedral angles in the local sequences,
as well as the importance of long range interactions in protein folding. Hence we proposes a
framework combining the second order Conditional Random Fields to model the local
interactions with the energy function to add long range constraints in the simulated annealing
sampling algorithm. The second order CRF-Sampler is built upon this framework, and applied
in CASP8.
Chapter 5 introduces the nonlinear relationship between the sequence information and the
local conformations. The Conditional Neural Fields model is applied to learn this complex
relationship and explained in detail in this chapter. Moreover, a different simulation method
parallel tempering is implemented in the CNF-Folder application to give out more high
quality decoys in comparison to CRF-Samplers as well as other popular free-modeling
methods.
10
Chapter 6 briefly introduces the statistical potentials and then presents in detail the novel
Evolutionary PAirwise Distance potential (EPAD). A complete comparison of EPAD against
other knowledge-based distance potentials on all the previously used datasets is provided,
which demonstrates the superior performance of EPAD. A new large scale dataset is also
proposed to avoid the statistical potentials’ over-tuning over the small datasets.
In Appendix A, we have proposed a contact capacity potential (CCP). The experimental results
imply that the contact capacity potential helps significantly improve decoy discrimination
when combined with distance-dependent potentials. The results also support the hypothesis
that the wrapping of each residue by surrounding amino acids is guided by maximizing the
local static electric field, so that the core of the residue is protected against the water
molecules.
11
Part 1.
Probabilistic and Continuous Models of
Protein Conformational Space for Fragmentfree and Template-Free Modeling
12
Chapter 2. Continuous representation of
protein conformations
Evaluating a full-atom energy function takes a great deal of time. However, a residue-level
energy function simply lacks the accuracy of an atom-level energy function. In this thesis, we
use a continuous yet simplified representation of a protein model. We only consider the main
chain of C-beta atoms in the folding simulations discussed in the subsequent chapters.
Figure 5 The , , dihedral angles in one residue of the protein backbone. can be assumed
to be fixed at 180° (trans) or 0° (cis).
2.1 Cα-trace representation
A protein backbone conformation can be described by angle triples , , , as well as a set of
bond lengths. is the dihedral angle around N−Cα bond and ψ is the dihedral angle around
Cα−C bond. Because it is possible to approximate ω and the bond lengths as constants, we can
represent a protein backbone as a set of , angles. With the exception of both the two
terminal residues of a protein chain, each residue has a pair of and angles. These ,
angles give us what we need to calculate the coordinates for all the nonhydrogen atoms of a
protein backbone. However, for some proteins, even if we have all of their native ϕ and ψ
angles, we cannot accurately rebuild their backbone conformations because of slight variation
of other angles.
13
This article utilizes a different way to represent a protein backbone instead of the ,
representation. Because the virtual bond length between two adjacent Cα atoms can be
approximated as a constant, i.e., 3.8 Å, the protein backbones can be represented as a set of
pseudo angles , (Levitt, 1976). One rare exception is when the second residue is cis proline,
the virtual bond length is approximately 3.2 Å. In this representation, all the other atoms are
ignored except the Cα atoms. For any position i in a backbone, is defined as the pseudo-bond
angle formed by the Cα atoms at positions i−1, i and i+1; τ is a pseudo dihedral angle formed by
the Cα atoms at positions i−2, i−1 i and i+1. Given the coordinates of the Cα atoms at positions
i−2, i−1, and i, the coordinates of the Cα atom at position i+1 can be calculated from , at
position i.
Therefore, for a protein with N amino acid residues, given the positions of the first three Cα
and N−2 , pairs, we can build the Cα trace of a protein. The relative positions of the first
three Cα atoms are determined by the θ angle of the second residue. Using , representation,
only the coordinates of the Cα atoms can be recovered. The coordinates of other backbone
atoms and Cβ atom can be built using programs such as MaxSprout (Holm and Sander, 1991),
BBQ (Gront et al., 2007), and SABBAC (Maupetit et al., 2006), which can build backbone
positions with RMSD less than 0.5 Å.
2.2 Distribution of bond angles
The preferred conformations of an amino acid in the protein backbone can be described as a
probabilistic distribution of the , angle pair. Each , corresponds to a unit vector in the
three-dimensional space (i.e., a point on a unit sphere surface). We can use the 5-parameter
Fisher-Bingham (FB5) distribution (Hamelryck et al., 2006; Kent, 1982) to model the probability
distributions over unit vectors. FB5 is the analogue on the unit sphere of the bivariate normal
distribution with an unconstrained covariance matrix. The probability density function of the
FB5 distribution is given by
! =
1
exp)$*+ . ! + %*. . !. − */ . !. 0
#$, %
where u is a unit vector variable and c(κ,β) is a normalizing constant (Kent, 1982). The
parameters $> 0 and %0 < 2% ≤ $ determine the concentration of the distribution and the
ellipticity of the contours of equal probability, respectively. The higher the $ and % parameters,
the more concentrated and elliptical the distribution is, respectively. The three vectors *+, *.,
and */ are the mean direction, the major and minor axes, respectively. The latter two vectors
determine the orientation of the equal probability contours on the sphere, while the first vector
determines the common center of the contours.
14
We cluster the whole space of (θ, τ) into 100 groups, each of which can be described by an FB5
distribution. We calculate the (θ, τ) distribution for each group from a set of ∼3000 nonredundant proteins with high-resolution x-ray structures using KentEstimator (Hamelryck et
al., 2006). More detailed description of how to calculate the FB5 distributions is shown in
Chapter 3 (Zhao et al., 2008). Once we know the distribution of (θ, τ) at one residue, we can
sample a pair of real-valued (θ, τ) angles in a probabilistic way and thus, explore protein
conformations in a continuous space.
2.3 Building backbone atoms
Using (θ, τ) representation, only the coordinates of the Cα atoms can be built. To use an atomlevel energy function, we also need to build the coordinates of other atoms. Given a Cα trace,
there are many methods that can build the coordinates for the main chain and Cβ atoms (Gront
et al., 2007; Holm and Sander, 1991; Maupetit et al., 2006). To save computing time, we want a
method that is both accurate and efficient. We choose to use a method similar to BBQ (Gront et
al., 2007). The original BBQ method can only build coordinates for the backbone N, C, and O
atoms. We extend the method to build coordinates for the Cβ atom. Experimental results (data
not shown) indicate that RMSD of this method is approximately 0.5 Å supposing the native Cαtrace is available. This level of accuracy is good enough for our folding simulation.
To employ the KMB hydrogen-bonding energy (Morozov et al., 2004) for β-containing proteins,
we also need to build the backbone hydrogen atoms. We use a quick and dirty method to build
coordinates for the hydrogen atom HN (Tooze and Branden, 1999)). Let Ni denote the position
of the main chain N atom in the same residue as the HN atom. Let Ni Ci−1 denote the
normalized bond vector from the N atom to the C atom in the previous residue. Let Ni Cα
denote the normalized bond vector from the N atom to the Cα atom in the same residue. Then
the position of the hydrogen atom HN can be estimated by 67 −
89 :9;< =89 :>
|89 :9;< =89 :> |
. The average
RMSD of this method is approximately 0.2Å (data not shown) supposing the native
coordinates of other main chain atoms are available.
2.4 Mathematical Symbols
Unless specifically clarified, we will use the mathematical symbols for the rest chapters in in
Part 1 as listed in Table 1. In the context of the Conditional Random Fields (CRFs) and
Conditional Neural Fields (CNFs) models we use, the primary sequence (or sequence profile)
and predicted secondary structure are viewed as observations; the backbone angles and their
FB5 distributions are treated as hidden states or labels.
15
Figure 6 Three point sets sampled from the FB5 distribution on the unit sphere.
The three node values are typical representatives of coil (blue), α-helix (red), and β-strand
(green). The arrows points to the mean directions of the three point sets. Adapted from
(Hamelryck et al., 2006).
16
Table 1 Some Mathematical Symbols Used in the CRF Models
Symbols
X
Xi
Xi(x)
M
Mi
Mi(aa)
H
Λ
Annotations
The PSIPRED-predicted secondary structure likelihood scores. A matrix with 3×N
elements where N is the number of residues in a protein.
The predicted likelihood of three secondary structure types at position i. It is a
vector of three values, indicating the likelihood of helix, beta and loop,
respectively.
The predicted likelihood of secondary structure type x at position i.
The position-specific frequency matrix with 20×N entries, each being the
occurring frequency of one amino acid at a given position.
A vector of 20 elements, denoting the occurring frequency of 20 amino acids at
position i.
The occurring frequency of amino acid aa at position i.
= @ℎ+ , ℎ. , … , ℎ+CC D, the set of 100 backbone angle states, each representing an
FB5 distribution.
Λ = FG+ , G. , … , GH I, the set of model parameters used by CRF/CNF model.
17
Chapter 3. First-order CRF-Sampler
3.1 Introduction
The FB5-HMM and TorusDBN models developed by Hamelryck et.al (Boomsma et al., 2008;
Hamelryck et al., 2006) sample the protein conformations from a continuous space. These
models are able to capture the relationship between backbone angles and primary sequence,
and the angle-dependency between two adjacent residues. FB5-HMM uses a Hidden Markov
Model to discover the native basins of the local θ, τ pseudo backbone dihedral angles of a
protein sequence by learning the local sequence-structure relationship; and TorusDBN learns
the local ϕ, ψ backbone dihedral angle dependency using a dynamic Bayesian network,
which is a generalization of the Hidden Markov Model. There are two particular advantages of
these methods: first, they model the backbone angles using a directional statistics distribution
(Kent, 1982; Mardia et al., 2007; Singh et al., 2002) so that they can be sampled from a
continuous space; second, they can find how two adjacent backbone angles are dependent on
one another.
Although experimental result shows that the sampler based on Hidden Markov Model(HMM)
method is promising in generating decoys with good quality, the model has a couple of
assumptions that may be relaxed.
The 1st assumption is that the residue at any position is independent of its type of secondary
structure. This would appear to be against the fact that one type of amino acid may prefer
some type of secondary structure to the others. The 2nd assumption is that the hidden state (i.e.,
the distribution of backbone angles) at position i is only dependent on the type of amino acid
residue and the type of secondary structure at that particular position. This contradicts with
the finding in (Jha et al., 2005) that the angles at any particular position i depends on at least
the three residues at positions i−1, i and i+1.
These assumptions are required in (Hamelryck et al., 2006) because the parameters in the
HMM model are estimated by maximizing the joint probability JK, L, of a primary
sequence A, secondary structure X and hidden states S (i.e., angles). For a reasonable estimate
of the joint probabilityJK, L, , these assumptions are necessary because of the sparsity in the
training data, as well as for the purpose to avoid overfitting. In FB5-HMM (Hamelryck et al.,
18
2006), JK, L, is estimated using Law of total probability ∑N JK|J |JL|J
where H is a possible hidden node sequence.
Due to the sparsity in the training data, it is also very difficult to incorporate more complexity
into the HMM model without elevating the risk of overfitting, which restricts the expressive
power of the HMM model. The problem is not merely limited to the HMM model, it is actually
a common problem of all the generative learning methods (e.g., HMM) that relies on
optimizing the joint probability of states and observations, where the primary sequence and
secondary structure are both observations. As an alternative, discriminative learning methods
such as CRFs (Conditional Random Fields) can be more expressive and at the same time
keeping the risk of overfitting under control. Discriminative learning is different from
generative learning in that the former one optimizes the conditional probability of states on the
observations while the latter one the joint probability of states and observations.
This chapter presents the first-order CRF-Sampler, an extensible and fully automatic
framework, for effective protein conformation sampling, based on a probabilistic graphical
model Conditional Random Fields (Lafferty et al., 2001; Sha and Pereira, 2003). Similar to the
HMM model, CRF-Sampler samples backbone angles from a continuous space using sequence
and secondary structure information. CRF-Sampler also models the dependency between the
angles at two adjacent positions. CRF-Sampler differs from the HMM model in the following
aspects. First, CRF-Sampler is more expressive than the HMM model. The backbone angles at
position i can depend on residues and secondary structures at many positions instead of only
one. In CRF-Sampler, a sophisticated model topology and feature set can be defined to
describe the dependency between sequence and structure without worrying about learning of
model parameters. Different from the HMM model, in which the model complexity (hence risk
of overfitting) roughly equals to the number of parameters in the model, the effective
complexity of CRF-Sampler is regularized by a Gaussian prior of its parameters, allowing the
user to achieve a balance between model complexity and expressivity. Second, CRF-Sampler
does not assume that primary sequence is independent of secondary structure in determining
backbone angles. Instead, CRF-Sampler can automatically learn the relative importance of
primary sequence and secondary structure. Finally, CRF-Sampler can easily incorporate
sequence profile (i.e., position-specific frequency matrix) and predicted secondary structure
likelihood scores into the model to further improve sampling performance. Our experimental
results demonstrate that, using only compactness and self-avoiding constraints, CRF-Sampler
can quickly generate more native-like conformations than the HMM model and best decoys
closer to their natives.
3.2 CRF model for sequence-structure relationship
19
Conditional random fields (CRFs) are probabilistic graphical models that have been
extensively used in modeling sequence data. Please refer to (Lafferty et al., 2001) and (Sha and
Pereira, 2003) for a complete description of CRFs. Here, we describe how to predict the
backbone angles of a protein from its primary sequence and secondary structure using CRFs.
In this context, the primary sequence and secondary structure of a protein are called
observations and its backbone angles are hidden states or labels.
Let O = @O+ , O. , … , O8 D denote an observation sequence of length N where oi is an observed
object. Each observed object can be a residue or a secondary structure type or their
combination. Let = @ℎ+ , ℎ. , … , ℎP D be a finite set of labels (also called states), each
representing a distribution of backbone angles. Let = @ + , . , … , 8 | 7 ∈ D be a sequence of
labels corresponding to the observation o. As opposed to the HMM model defining a joint
probability of the label sequence S and the observation o, our CRF model defines the
conditional probability of s given o as follows.
JR |O =
exp∑8
7T+ S , O,
UO
3.1
where Λ=(λ1,λ2,…λp) is the model parameter and UO = ∑XY exp∑8
7T+ S ′, O, is a
normalization factor summing over all the possible label sequences for a given observation
sequence. F(S,o,i) is the sum of the CRF features at sequence position i:
S , O, = Z G[ \[
where \[
[
7]+ , 7
7]+ , 7
3.2
+ Z G^ _^ O, 7
^
and _^ O, 7 are called edge and label feature functions, respectively.
The edge and label functions are defined as
\[
and
7]+ , 7
=`
7]+
= ℎ+ a`
_^ O, 7 = `b^ O, a`
7
7
3.3
= ℎ. a
3.4
= ℎa
where Si = h indicates that the label (or state) at position i is h. And b^ O, is a logical context
predicate indicating whether or not the context of the observation sequence o at position i
holds a particular property or fact of empirical data. [f] is equal to 1 if the logical expression f is
true, and zero otherwise. Note that we can also define the edge feature function \[ 7]+ , 7 as
`b^ O, a` 7]+ = ℎ+ a` 7 = ℎ. a , to capture relationship between two adjacent labels and
20
observations. By expanding Equation 3.1 using Equation 3.2, 3.3, and 3.4 and merging the
same items, the conditional probability can also be reformulated as follows.
JR |O =
exp∑8
exp∑[ G[ [ , O
7T+ S , O,
=
∑XY exp∑[ G[ [ ′, O
UO
3.5
where [ , O represents the occurring times of the kth feature in a pair of label sequence s and
observation sequence o and the model parameter λk is the weight of this feature. Here, the
parameter λk does not correspond to the log probability of an event (as in the HMM model)
Instead, it is a real-valued weight that either raises or lowers the “probability mass” of s
relative to other possible label sequences. The parameter λk can be negative, positive, or zero.
The CRF model is more expressive than the HMM model. First, we do not have to interpret the
parameter λk as the log probability of an event. Second, CRFs do not have to assume that the
observation object at one position is independent of other objects. That is, for any νl(o,Si), the
label at position i can depend on many observed objects in the observation sequence or even
the whole observation sequence. In addition, Si can also depend on any nonlinear combination
of several observed objects. Therefore, the CRF model can accommodate complex feature sets
that may be difficult to incorporate within a generative HMM model. The underlying reason is
that CRFs only optimize the conditional probability PΛ(S|o) instead of joint probability PΛ(S,o),
avoiding calculating the generative probability of the observation sequence.
Model parameter estimation
Given a set of observation sequences and their corresponding label sequences (oi, Si), CRFs
train its parameter Λ = FG+ , G. , … , GH I by maximizing the conditional log-likelihood L of the
data:
e = Z log R i |Oi −
i
1
‖G‖+
2j .
3.6
This kind of training is also called discriminative training or conditional training. Different
from the generative training in the HMM model, discriminative training directly optimizes the
predictive ability of the model while ignoring the generative probability of the observation.
The e+ norm in the last term in Equation 3.6 is a regularization item to deal with the sparsity in
the training data. When the complexity of the model is high (i.e., the model has many features
and parameters) and the training data is sparse, overfitting may occur and it is possible that
many models can fit the training data. To prevent this, we place a Gaussian prior,
exp m.no ∑[ G[ p, on the model parameter to choose the model with a “small” parameter. This
+
21
regularization can improve the generalization capability of the model in both theory and
practice.
The objective function in Equation 3.6 is convex and hence theoretically a globally optimal
solution can be found using any efficient gradient-based optimization technique. There is no
analytical solution to the above equation for a real-world application. Quasi-Newton methods
such as L-BFGS (Liu and Nocedal, 1989) can be used to solve the above equation and usually
can converge to a good solution within a couple of hundred iterations. The log-likelihood
gradient component of λk is
qe
=Z
qG[
7
[
7
, j 7 − Z Z R ) rO7 0
s
X
[)
, O7 0 −
G[
j.
3.7
The first two items on the right of the above equation is the difference between the empirical
and the model expected values of feature count Ck. The expected value ∑X R |O [ , O for a
given o can be computed using a simple dynamic programming algorithm if the model only
has edge feature functions defined in Equation 3.3.
Model topology
As illustrated in Figure 7, we use a CRF model to capture the relationship between a protein
sequence and its (pseudo) backbone angles. Let Si denote the label at position i. Each label
represents a distribution of backbone angles in a protein position. We use the Sine model
(Mardia et al., 2007; Singh et al., 2002) to describe the distribution of the , angles and the
FB5 mode (Kent, 1982) for the (θ, τ) angles, respectively. Each label depends on a window of
residues in the primary sequence, their secondary structure types, and any nonlinear
combinations of them. There is also interdependence between two adjacent labels. The CRF
model is not necessary a linear-chain graph. It can be easily extended to model the long-range
relationship between two positions. For example, if distance restraints are available from NMR
or threading programs, then we can add some edges in the CRF model to capture the longrange interactions between two nonadjacent residues. In the CRF model, we do not assume
that the residues in primary sequence are independent of each other and that primary
sequence is independent of secondary structure. Our CRF model can easily capture this kind
of interdependence in its conditional probability R |O.
In this example, Si (i.e., the label at position i) depends on the residues and secondary structure
types at positions i−2, i−1, i, i+1, and i+2 and any nonlinear combinations of them. There is also
interdependence between two adjacent labels. This CRF model can also be extended to
incorporate long-range interdependence between two labels.
22
Figure 7 An example CRF model for protein conformation sampling.
23
Model features
CRF-Sampler uses two different types of feature functions. At each position i, CRF-Sampler
uses e(Si−1,Si) = [Si−1 = h1] [Si = h2] as its edge feature function. That is, currently, we only consider
the first order dependence between labels. We are also investigating the second order
dependence between labels. We use the following label feature functions to model the
relationship among primary sequence, secondary structure, and backbone angles. Meanwhile,
w is half of the window size.
1. ν1,j(o,Si) = [Ai+j = a] [Si = h]. This feature set describes the interdependence between the
label at position i and the residue at position i+j, where − w ≤ j ≤ w. A feature in this set is
identified by a triple (a, h, j).
2. ν2,j(o,Si) = [Xi+j = x] [Si = h]. This feature set describes the interdependence between the
label at position i and the secondary structure type at position i + j, where −w ≤ j ≤ w. A
feature in this set is identified by a triple (x, h, j).
3. ν3,j(o,Si) = [Ai+j = a] [Xi+j = x] [Si = h]. This feature set describes the interdependence among
the label at position i, the residue at position i + j, and the secondary structure type at
position i + j where −w ≤ j ≤ w. A feature in this set represents a nonlinear combination
of secondary structure and primary sequence and is identified by a quadruple (a, x, h, j).
We use a window size 9 (i.e., w = 4), which is slightly better than a window size 5 when
predicted secondary structure information is used as input of CRF-Sampler (see Table 2). In
total there are more than ten thousand features in CRF-Sampler quantifying protein sequencestructure relationship.
Extension to continuous-valued observations
We can also extend CRF-Sampler to make use of sequence profile and predicted secondary
structure likelihood scores to improve sampling performance. In this case, for each protein, the
observation is not two strings any more but consists of two matrices. One is the positionspecific frequency matrix containing 20 × N entries (where N is the number of residues in a
protein); each element in this matrix is the occurring frequency of one amino acid at a given
position. The other matrix is the PSIPRED-predicted secondary structure likelihood matrix
containing 3 × N elements; each element is the predicted likelihood of one secondary structure
type at a specific position. To use this kind of continuous-valued observations, we extend label
feature functions as follows. In defining νl,j(o, Si) (l = 1,2,3), instead of assigning [Ai+j = a] to a
binary value (i.e., 0 or 1), we assign [Ai+j = a] to the frequency of amino acid a appearing at
position i + j. Similarly, we assign [Xi+j =x] to the PSIPRED-predicted likelihood of secondary
type x at position i + j.
24
Table 2 F1-Values (%) of CRF-Sampler with Respect to Window Size
Window
size
1
5
9
a
10.01
14.08
14.76
b
10.02
17.5
18.58
c
15.3
19.65
19.81
d
17.34
20.49
20.89
F1-value is an even combination of precision (p) and recall (r) and defined as 2pr/(p + r). a,
trained and tested using primary sequence; b, trained and tested using PSI-BLAST sequence
profile; c, trained and tested using primary sequence and PSIPRED-predicted secondary
structure; d, trained and tested using PSI-BLAST sequence profile and PSIPRED-predicted
secondary structure confidence scores.
25
3.3 Conformation sampling algorithm
Sample one conformation for the whole protein
Given a CRF model and its parameters, we used a forward-backward sampling algorithm to
generate protein conformations. The algorithm is an extension of the sampling algorithm
described for the HMM model (Hamelryck et al., 2006). The major difference is that our
sampling algorithm needs to deal with many more sophisticated features and we also need to
transform likelihood to probability in sampling. Let νl(i, h) denote a label feature function
associated with position i and label h. For a given position i and a label h, we recursively define
and calculate u , ℎ from N-terminal to C-terminal as follows.
u0, ℎ = exp vZ G^ _^ 0, ℎw
^
u , ℎ = exp vZ G^ _^ , ℎw Z u) − 1, ℎx0\ yz{,z
{
|
^
where λl is the trained parameter for the label feature νl(,) and G|{,| is the trained parameter for
the edge feature \)ℎx, ℎ0. After G(N − 1, h) (N is the protein size) is calculated, we can sample a
conformation from C-terminal to N-terminal. First, we sample the label h for the last position
according to probability
ℎ =
u6 − 1, ℎ
∑|Y u , ℎ′
Then we sample the label ℎx for position i according to probability
)ℎx0 =
u) , ℎx0\ yz{,z
∑|Y u , ℎ′\ yz},z
assuming that the sampled label at position i + 1 is h. Note that each label corresponds to a
distribution of backbone angles. Based on the sampled labels, we can sample the two backbone
angles for each position and build a backbone conformation for the protein.
Resample a small segment of the backbone conformation
Given a backbone conformation, we generate the next conformation by resampling a small
segment of the protein. First, we randomly sample the starting position of the segment and its
length. The length is uniformly sampled from 1 to 15. We resample the labels of the segment
26
including positions i, i + 1, …, j, conditioned on the current labels at positions i − 1 and j + 1.
Suppose that the labels at positions i − 1 and j + 1 are h1 and h2, respectively. We calculate
u̅ , ℎ ≤ ≤ from position i to j as follows.
u̅ , ℎ = exp vZ G^ _^ , ℎw \ yz<,z
^
u̅ , ℎ = exp vZ G^ _^ , ℎw Z u̅ ) − 1, ℎx0\ yz{,z
{
|
^
After calculating G(k, h) for all the k between i and j, we can sample the labels for the segment
from j to i. At position j−1, we sample a label h according to probability
ℎ =
u̅ − 1, ℎ\ yz,zo
∑| u̅ − 1, ℎ\ yz,zo
For any position k (i ≤ k ≤ j − 1), we sample a label h according to probability
u̅ , ℎ\ yz,z{
ℎ =
∑| u̅ , ℎ\ yz,z{
supposing ℎ is the sampled label at position k+1. After resampling the labels of this segment,
we can resample the angles of this segment and then rebuild the backbone conformation.
Folding simulation
Since the focus of this chapter is the protein conformation sampling algorithm, we use only
compactness and self-avoiding constraints to drive conformation search during the folding
simulation process. We start with sampling the whole backbone conformation of a given
protein and then optimize its conformation by minimizing the radius of gyration. Given a
conformation, we generate its next potential conformation by resampling the local
conformation of a small segment. If this potential conformation has no serious steric clashes
among atoms, then we compare its radius with that of current conformation. If this potential
conformation has a smaller radius, then we accept this conformation, otherwise reject it. This
process is terminated if no better conformations can be found within 1000 consecutive
resamplings. There is a steric clash if the distance between two Cα atoms is less than 4 Å.
3.4 Experiments and Results
27
Data set
The first-order CRF-Sampler is tested on the following proteins: 1FC2, 1ENH, 2GB1, 2CRO,
1CTF, 4ICB, 1AA2, 1BEO, 1DKT, 1FCA, 1FGP, 1JER, 1NKL, 1PGB, 1SRO, 1TRL, T0052 (PDB
code: 2EZM), T0056 (1JWE), T0059 (1D3B), T0061 (1BG8), T0064 (1B0N), and T0074 (1EH2). The
first six proteins have been studied in (Simons et al., 1997) and (Hamelryck et al., 2006); and
the last 18 in (Xia et al., 2000); the last six proteins are also CASP3 targets. We obtained a set of
non-redundant protein structures using the PISCES server (Wang and Dunbrack, 2003) as our
training data. Each protein in this set has resolution at least 2.0 Å, R factor no bigger than 0.25
and at least 30 residues. Any two proteins in this set share no more than 30% sequence identity.
To avoid overlap between the training data and the test proteins, we removed the following
proteins from our training data:(1) the proteins sharing at least 25% sequence identity with our
test proteins; (2) the proteins in the same fold class as our test proteins according to the SCOP
classification (Murzin et al., 1995); and (3) the proteins having a TM-score ≥ 0.5 with our test
proteins in case some recently released proteins do not have a SCOP ID. According to (Zhang
and Skolnick, 2005b), if the TM-score of two protein structures is smaller than 0.5, then a
threading program such as PROSPECTOR_3 (Skolnick et al., 2004) cannot identify their
similarity relationship with high confidence.
Label assignment and distribution parameters
To train CRF-Sampler, we also need to assign a label to each position in a protein. In this part,
we only tested our algorithm on the (θ, τ) representation of a protein backbone conformation.
There can be various methods to assign a label to a protein position. For example, we can
cluster all the (θ, τ) angles into dozens of groups; each group corresponds to a label. Here, we
just simply use the five-residue fragment libraries developed by Kolodny et al.(Kolodny et al.,
2002) since these libraries have already been carefully designed. The library containing 100
five-residue fragments is used as the set of hidden labels; each label corresponds to a cluster in
the fragment library. We calculated the (θ, τ) distribution for each cluster from the training
proteins using the KentEstimator program enclosed in Mocapy(Hamelryck et al., 2006). Only
the angles of the middle residue in a fragment are used to calculate the angle distribution
parameters. We also tested other four-residue and five-residue fragment libraries developed
by Kolodny et al. and it turns out that the five-residue fragment library with 100 clusters yields
the best performance.
Parameter tuning
We randomly divided the training proteins into five sets of same size and then used them for
five-fold cross validation. We trained CRF-Sampler using several different regularization
factors [i.e., σ2 in Equation 3.7]: 50, 100, 200, 400, and 800 and choose the one with the best F128
value. F1-value is a widely used measurement of the prediction capability of a machine
learning model in the machine learning community. F1-value is an even combination of
precision (p) and recall (r) and defined as 2/ + . The higher the F1-value is, the better.
When both PSI-BLAST sequence profile and PSIPRED-predicted secondary structure
likelihood scores are used and a window size 9 is used to define the model features, the
average F1-values for regularization factors 50, 100, 200, 400, and 800 are 20.82%, 20.89%,
20.83%, 20.71%, and 20.56%, respectively. In fact, there is no big difference among these
regularization factors in terms of F1-value. However, we prefer to choose a small
regularization factor 100 to control the model complexity. The regularization factor is the only
parameter that we need to tune manually. All the other model parameters (i.e., weights for
features) can be estimated automatically in training.
In addition, we also tested the performance of our algorithm with respect to window size in
defining model features. As shown in Table 2, our experimental results indicate that when 100
labels are used in CRF-Sampler, a window size 5 can yield a much higher F1-value than a
window size 1. Increasing the window size to 9 can improve the F1-value, but the
improvement is small when PSIPRED-predicted secondary structure is used in CRF-Sampler.
This may be because the predicted secondary structure also contains partial information of
neighbor residues. In our remaining experiments, we used window size 9 to define model
features for CRF-Sampler.
Comparison with the HMM model
It is not easy to fairly compare two protein conformation sampling algorithms. Many ab initio
folding programs use a sophisticated energy function to drive conformation search and it is
hard to evaluate performance of their conformation sampling algorithms alone, without
considering their energy functions. The focus of this chapter lies in only protein conformation
sampling algorithm. To evaluate the sampling algorithm, we drive conformation search by
minimizing the radius of gyration instead of a well-designed energy function. Here we
compare CRF-Sampler mainly with the HMM method described in (Hamelryck et al., 2006),
which is also a protein conformation sampling algorithm and drives conformation search by
minimizing the radius instead of an energy function. The major difference between CRFSampler and the HMM model is that the former generates conformations using a CRF model
while the latter uses an HMM model. We tested CRF-Sampler on six proteins studied in
(Hamelryck et al., 2006) and compared the quality of the decoys generated by CRF-Sampler
with those by the HMM model.
Since for most proteins without known structures, we cannot obtain their true secondary
structures, here we compare the HMM model and CRF-Sampler using only PSIPREDpredicted secondary structure and sequence information as their inputs. As shown in Table 3,
29
when only sequence information used, CRF-Sampler can generate decoys with much higher
quality than the HMM model. The only exception is on 1ENH where the HMM model has a
comparable performance with CRF-Sampler when only primary sequence is used. When only
primary sequence is used, the number of good decoys (RMSD ≤ 6 Å) generated by CRFSampler is 2 ∼ 4 times of that by the HMM model. This difference comes from the fact that in
sampling the angles at one position, CRF-Sampler can directly take into consideration the
effects of its neighbor residues. When predicted secondary structure is used, as shown in Table
4, CRF-Sampler is much better than the HMM model in generating good decoys and the best
decoys. Among the six test proteins, CRF-Sampler is slightly worse than the HMM model on
the Calbidin protein (PDB code: 4ICB) when predicted secondary structure type is used. For
the other five proteins, CRF-Sampler can generate many more good decoys. CRF-Sampler can
also generate the best decoys with much smaller RMSDs, although only 20,000 decoys are
generated by CRF-Sampler for each test protein while 100,000 decoys by the HMM model for
each test protein. By using PSI-BLAST sequence profile and predicted secondary structure
likelihood scores as input, CRF-Sampler can achieve overall performance better than using
primary sequence and predicted secondary structure types. This indicates the importance of
using continuous-valued observations in a conformation sampling algorithm. By contrast, it is
not easy for the HMM model to incorporate these kind of continuous-valued observations as
input. Figure 8 visualizes the native structures and the best decoys (generated by CRF-Sampler)
of the six test proteins.
Table 5 lists the percentage of correct secondary structure (i.e., Q3-value) of all the good
decoys generated by CRF-Sampler for each test protein. A software P-SEA(Labesse et al., 1997)
is used to calculate the secondary structure of a decoy. As shown in this table, even with
primary sequence only, CRF-Sampler can generate decoys with pretty good Q3-values, better
than the HMM model (see Table II in (Hamelryck et al., 2006)). This confirms that CRFSampler can capture well the relationship between a sequence stretch and its local
conformation.
30
Table 3 Decoy Quality Comparison between the HMM Model and CRF-Sampler
Note: Only sequence information is used in both training and testing. In total, 100,000 decoys
are generated by the HMM model.
Column “L” lists the length of the test proteins; column “α,β” lists the number of α-helices and
β-strands of the test proteins; columns “Good” and “Best” list the percentage of good decoys
(with RMSD ≤ 6 Å) and the RMSD of the best decoy, respectively.
Trained and tested using primary sequence and the results are taken from (Hamelryck et al., 2006).
Trained and tested using primary sequence and 40,000 decoys are generated.
c
Trained and tested using PSI-BLAST sequence profile. Only 20,000 decoys are generated.
a
b
Test proteins
Name, PDB code
Protein A, 1FC2
Homeodomain,
1ENH
Protein G, 2GB1
Cro repressor,
2CRO
Protein L7/L12,
1CTF
Calbidin, 4ICB
L
α,β
43 2,0
HMMa
Good
Best
(%)
(Å)
9.59
2.7
CRF-Samplerbb
Good
Best
(%)
(Å)
20.9
2.08
CRF-Samplerc
Good
Best
(%)
(Å)
24.8
2.09
54 2,0
56 1,4
6.6
0.04
2.5
4.9
6.23
0.16
2.68
4.67
14
10.1
1.98
3.36
65 5,0
0.46
3.9
1.94
4.05
13.3
2.37
68 3,1
76 4,0
0.01
0.09
5.4
4.3
0.04
0.17
4.94
4.57
0.15
0.42
4.49
4.72
31
Table 4 Decoy Quality Comparison between the HMM Model17 and CRF-Sampler
Note: Both sequence and secondary structure information are used. In total, 100,000 decoys are
generated by the HMM model while only 20,000 decoys by each CRF-Sampler.
Please refer to Table 3 for details of the Column definitions.
Trained using true secondary structure and primary sequence while tested using predicted
secondary structure (by PSIPRED(McGuffin et al., 2000)) and primary sequence and the results
are taken from (Hamelryck et al., 2006).
b
Trained and tested using predicted secondary structure (by PSIPRED) and primary sequence.
c
Trained and tested using predicted secondary structure likelihood scores (by PSIPRED) and
PSI-BLAST sequence profile.
a
Test proteins
Name, PDB code
Protein A, 1FC2
Homeodomain,
1ENH
Protein G, 2GB1
Cro repressor, 2CRO
Protein L7/L12, 1CTF
Calbidin, 4ICB
L
4
3
5
4
5
6
6
5
6
8
7
6
α,
β
HMMa
Good
(%)
Best
(Å)
CRF-Samplerb
Good
Best
(%)
(Å)
CRF-Samplerc
Good
Best
(%)
(Å)
2,0
17.1
2.6
26.8
2.13
49.1
1.94
2,0
12.2
3.8
16.7
2.29
22.4
2.32
1,4
0
5.9
26.4
3.05
23.3
2.91
5,0
1.1
4.1
18.3
2.76
16.8
2.79
3,1
0.35
4.1
3
4.04
2.4
3.7
4,0
0.38
4.5
0.24
4.45
0.51
4.63
32
Figure 8 Native structures (in orange) and the best decoys of 1FC2, 1ENH, 2GB1, 2CRO, 1CTF,
and 4ICB
33
Table 5 Secondary Structure Content of Good Decoys
Note:
a Percentage of correct secondary structure (Q3-value) and secondary structure content of good
decoys (RMSD ≤ 6 Å) generated using primary sequence.
b
Q3-value and secondary structure content of good decoys generated using PSI-BLAST
sequence profile.
c
Q3-value and secondary structure content of good decoys generated using primary sequence
and PSIPRED-predicted predicted secondary structure.
d Q3-value and secondary structure content of good decoys generated using PSI-BLAST
sequence profile and PSIPRED-predicted secondary structure likelihood scores.
CRF-Samplerb
CRF-Samplera
Protein Q3
H
E
C
Q3
H
E
C
1FC2
66.5
61.3
1.8
36.9
75.2
49.4
1.9
48.7
1ENH 79.3
69.2
2.9
27.9
79.2
64.9
7.1
28
2GB1
65.2
29.5
24.9
45.7
61
24.1
27.6
48.2
2CRO
80
71.6
2.3
26.2
85.2
66
1.8
32.2
1CTF
67.1
50.3
9.5
40.1
62.6
41.8
19.7
38.5
4ICB
65
59.6
2.9
37.5
66.6
59
5
36
Protein
1FC2
1ENH
2GB1
2CRO
1CTF
4ICB
Q3
85.5
86
72
87.1
77.5
65.7
CRF-Samplerc
H
E
55.1
1
64.3
2.9
24.8
32.9
69.3
1.5
56.9
10.6
62
3.81
C
43.9
32.8
42.3
29.1
32.5
34.2
34
Q3
80.7
85.1
71.5
85.9
77.2
67.1
CRF-Samplerd
H
E
61.9
0.2
67.3
3
25.9
29.7
67.5
1.6
59.5
8.6
63.1
2.5
C
37.9
29.7
44.4
31
31.9
34.4
Comparison with Xia et al
In their paper, Xia et al. (Xia et al., 2000) have developed a hierarchical method to generate
decoys. This method first exhaustively enumerates all the possible conformations on a lattice
model for a given protein sequence and then builds conformations with increasing detail. At
each step, this method chooses a good subset of conformations using hydrophobic
compactness constraint and empirical energy functions such as RAPDF(Samudrala and Moult,
1998) and Shell(Park et al., 1997) and finally generate 10,000 or 40,000 decoys for a protein
sequence. Table 6 lists the RMSD ranges of all the decoys generated by this method and CRFSampler for 18 test proteins. As shown in this table, CRF-Sampler can generate decoys with
smaller RMSD on all the test proteins with less than 100 residues. CRF-Sampler is significantly
better than this hierarchical method on 1CTF, 1NKL, 1PGB, 1SRO, 1TRL-A, T0052, T0059 and
T0074 in terms of the best decoys. For those test proteins with more than 100 residues, CRFSampler is worse than the method of Xia et al. on two proteins (1AA2 and T0056) and better
one on protein (T0064), and has comparable performance on 1JER. This may indicate that we
need to improve CRF-Sampler further to search conformation space more effectively for a
protein with more than 100 residues. We also used the Wilcoxon signed-rank test(Wilcoxon,
1945), a nonparametric alternative to the paired Student's t-test, to calculate the significance
level at which CRF-Sampler is better than the method of Xia et al. Since we only had the RMSD
ranges of the decoys generated by Xia et al., we only considered the best decoys in calculating
the statistical test. For the first 12 proteins in Table 6, we used the best decoys in a set of
randomly chosen CRF-Sampler 10,000 decoys since Xia et al. only generated 10,000 final decoys
for these proteins. For the last six proteins, we used their best decoys listed in Table 6. When
the absolute RMSD difference between the best decoys is used to calculate the statistical test,
CRF-Sampler is better than the method of Xia et al. at significance level 0.01(In fact the
significance level is very close to 0.005). When the relative RMSD difference is used, CRFSampler is better than the method of Xia et al. at significance level 0.005. Finally, CRF-Sampler
tends to generate decoys with larger RMSD variance because CRF-Sampler does not use any
empirical energy functions to filter those bad conformations.
CRF-Sampler generated 20,000 decoys for each test protein using PSI-BLAST sequence profile
and predicted secondary structure likelihood scores. Xia et al. (Xia et al., 2000) conducted a
complete enumeration on a lattice model for each test protein and then generated 10,000
decoys for each of the first 12 proteins and 40,000 decoys for each of the six CASP3 targets,
respectively, using predicted secondary structure and empirical energy functions as filters.
35
Table 6 Decoy Quality Comparison Between Xia et al. (Xia et al., 2000) and CRF-Sampler
Test proteins
PDB code L
1aa2
108
1beo
98
1ctf
68
1dktA
72
1fca
55
1fgp
67
1jer
110
1nkl
78
1pgb
56
1sro
76
1trlA
62
4icb
76
T0052
98
T0056
114
T0059
71
T0061
76
T0064
103
T0074
98
Xia et al.
Class All RMSD range
α
6.18–15.28
α
6.96–15.94
α + β 5.45–13.54
β
6.68–14.79
β
5.09–12.06
β
7.80–14.40
β
9.55–17.53
α
5.26–14.23
α + β 5.60–13.30
β
7.30–15.42
α
5.30–13.16
α
4.74–13.28
β
10.6–16.3
α
6.2–17.8
β
7.4–15.7
α
6.0–14.0
α
8.0–18.8
α
6.3–16.5
CRF-Sampler
All RMSD range
7.34–17.06
6.41–16.95
3.70–13.37
6.14–15.51
4.98–12.90
7.39–15.20
9.63–19.64
3.63–13.76
3.15–12.75
6.22–15.70
3.53–12.43
4.63–13.93
7.58–19.17
7.77–18.17
6.29–15.54
5.35–14.84
7.23–18.85
4.85–15.72
We also calculated the secondary structure content of all the decoys using P-SEA (Labesse et
al., 1997) and compared CRF-Sampler with PSIPRED(McGuffin et al., 2000) in terms of Q3value. As shown in Table 7, CRF-Sampler can generate decoys with reasonable level of
secondary structure accuracy. The average Q3-value of all the decoys generated by CRFSampler is 70.3% while the average PSIPRED Q3-value of these test proteins is 74.4%. The
Wilcoxon signed-rank test indicates that in terms of Q3-value, CRF-Sampler is worse than
PSIPRED at significance level 0.025.
36
Table 7 Percentage of correct secondary structure (Q3-value) and secondary structure scntent
of all the decoys generated by CRF-Sampler, compared with PSIPRED Predictions
Test proteins
CRF-Sampler
PSIPRED
PDB code L
Class Q3
H
E
C
Q3
H
E
C
1aa2
1beo
1ctf
1dktA
1fca
1fgp
1jer
1nkl
1pgb
1sro
1trlA
4icb
T0052
T0056
T0059
T0061
T0064
T0074
α
α
α+β
β
β
β
β
α
α+β
β
α
α
β
α
β
α
α
α
53.9
51.3
59.8
16.3
4.7
4.5
14.1
75.3
25.6
7.7
70
61.8
6
71.2
10.4
42
66.2
53.6
5.4
9.9
8.6
28.4
16.8
28.5
34.7
0.4
27.8
28.8
0.4
3.5
36.1
1.4
41.3
13.7
2.7
3.2
40.7
38.7
31.6
55.3
78.5
66.9
51.2
24.3
46.7
63.5
29.6
34.8
57.9
27.5
48.3
44.3
31
43.2
88
65.3
77.9
50
83.6
64.3
74.3
92.3
80.4
56.6
83.9
67.1
54.5
81.6
72.2
71.1
94.2
83.2
56.5
42.9
55.9
15.3
0
0
8.3
73.1
21.4
10.5
67.7
60.5
8.9
74.6
8.3
40.8
67
55.8
1.9
6.1
13.2
30.6
16.4
35.7
37.6
0
50
43.4
0
2.6
42.6
0
52.8
14.5
0
0
41.7
51
30.9
54.2
83.6
64.2
54.1
26.9
28.6
46.1
32.3
36.8
48.5
25.4
38.9
44.7
33
44.2
108
98
68
72
55
67
110
78
56
76
62
76
98
114
71
76
103
98
78.5
68.4
76.9
53.7
71.9
57.3
63.2
87.7
67.3
59.2
80.9
67.2
54.1
83.3
64.9
59.9
87.2
84.9
The average Q3-value of all the decoys is 70.3% while the average PSIPRED Q3-value is 74.4%.
CRF-Sampler is worse than PSIPRED at significance level 0.025.
Comparison with Rosetta
Here, we compare CRF-Sampler with the well-known fragment-assembly-based program
Rosetta (Simons et al., 1997). Rosetta uses multiple sequence alignment information to choose
25 fragments for each sequence segment of nine residues and then assembles them into decoys
using a time-consuming simulated annealing procedure. Rosetta drives conformation search
using a well-developed energy function and generates a few hundred decoys, while CRFSampler generates 20,000 decoys without using any energy function. Although this
comparison is interesting, we want to point out it is also unfair to both CRF-Sampler and
Rosetta. On one hand, CRF-Sampler does not use an energy function to drive conformation
37
search. On the other hand, the result for Rosetta is taken from an article published in 1997,
which may not be the performance of the state-of-the-art Rosetta. As indicated in Table 8, by
quickly generating a large number of decoys, CRF-Sampler can obtain decoys with much
smaller RMSD than Rosetta. However, it is also not surprising that Rosetta can generate higher
percentage of good decoys for five out of six test proteins by using a time-consuming energy
minimization procedure.
Computational efficiency
Although we have not optimized the C++ code of CRF-Sampler, CRF-Sampler can quickly
generate a decoy within seconds for a test protein. Table 9 the approximate running time in
minutes spent by CRF-Sampler generating 100 decoys for each test protein, on a single 2.2
GHz CPU. As indicated in this table, CRF-Sampler can generate decoys for these test proteins
very quickly. It takes CRF-Sampler approximately 1 h to generate 100 decoys for protein G
(2GB1) and no more than ten minutes for 1FC2. It does not increase the running time of CRFSampler by using more information as input such as PSI-BLAST sequence profile and
secondary structure likelihood scores. Instead, using them tend to reduce the running time of
CRF-Sampler, maybe because of the reduction in the entropy of conformation search space.
38
Table 8 Decoy Quality Comparison Between ROSETTA2 and CRF-Sampler
Note: Only sequence information is used in both training and testing. ROSETTA generates a
few hundred decoys using energy function optimization while CRF-Sampler generates 20,000
decoys by minimizing radius of gyration.
Refer to Table 3 for details of the Column definitions.
Multiple sequence alignment information is used and the results are taken from (Simons et al.,
1997).
bTrained and tested using primary sequence.
c
Trained and tested using PSI-BLAST sequence profile.
a
Test proteins
α,
β
ROSETTAa
CRF-Samplerb
CRF-Samplerc
Good
(%)
Best(Å)
Good
(%)
Best
(Å)
Good
(%)
Best
(Å)
Name, PDB code
L
Protein A, 1FC2
43 2,0
95
3.3
20.9
2.08
24.8
2.09
Homeodomain,
1ENH
54 2,0
47
2.7
6.23
2.68
14
1.98
Protein G, 2GB1
56 1,4
0
6.3
0.16
4.67
10.1
3.36
Cro repressor, 2CRO
65 5,0
18
4.2
1.94
4.05
13.3
2.37
Protein L7/L12,
1CTF
68 3,1
6
5.3
0.04
4.94
0.15
4.49
Calbidin, 4ICB
76 4,0
17
4.7
0.17
4.57
0.42
4.72
39
Table 9 Approximate Running Time in Minutes Spent by CRF-Sampler in Generating 100
Decoys
Note: The features are used in both the training and testing data set. The unit of time in the
grids is Minute.
Features
1FC2 1ENH 2GB1 2CRO 1CTF 4ICB
primary sequence only
7
13
45.5 29.5
46
37.5
PSI-BLAST sequence profile only
8
12
63
78
43.5
6
7.5
67.5 7.5
16
29.5
4
8
56
13
27.5
14.5
PSIPRED-predicted secondary structure
and
primary sequence
PSIPRED-predicted secondary structure confidence scores
and
PSI-BLAST sequence profile
40
7
3.5 Conclusion
This chapter presented an extensible and fully automatic framework CRF-Sampler that can be
used to effectively sample conformations of a protein from its sequence information and
predicted secondary structure. CRF-Sampler uses thousands of parameters to quantify the
relationship among backbone angles, primary sequence and secondary structure without
worrying about risk of overfitting. Experimental results demonstrate that CRF-Sampler is
more effective in sampling conformations than the HMM model(Hamelryck et al., 2006). CRFSampler is quite flexible. Using CRF-Sampler, the user only needs to choose a set of
appropriate features describing the relationship between protein sequence and structure. CRFSampler can take care of the remaining tasks such as parameter estimation and conformation
sampling.
The first order CRF-Sampler only takes into consideration the dependency between two
adjacent positions. It may explore the conformation space of medium-sized proteins (≥ 100
residues) more effectively by incorporating the interdependency among more residues. For
example, given the labels at positions i and i + 1, there are on average only 10 possible labels at
position i + 2. Given the labels at positions i and i + 2, there are on average only 16 possible
labels at position i + 4. If this kind of constraint information is incorporated into CRF-Sampler,
it can greatly reduce the entropy of conformation search space and is very likely to scale CRFSampler up to proteins with more than 100 residues.
In this chapter, we drove the conformation optimization by minimizing the radius of gyration.
Our next step is to couple CRF-Sampler with a good energy function such as DOPE (Shen and
Sali, 2006) and DFIRE (Zhou and Zhou, 2002), to do real protein structure prediction. The
decoys generated by CRF-Sampler can also be used to benchmark energy functions since CRFSampler does not employ any energy function to generate these decoys and thus no energybias is introduced into these decoys.
We can also incorporate other predicted information such as solvent accessibility and contact
capacity (i.e., the number of contacts for a residue) into CRF-Sampler, which may improve
sampling performance. In addition, if some distance restraints can be obtained from NMR data
or comparative modeling, then it is also possible to extend CRF-Sampler to incorporate longrange interdependency. CRF-Sampler can also be extended to make use of other NMR data
sources such as chemical shifts (Neal et al., 2006) and residue dipolar coupling (RDC) data
(Meiler et al., 2000).
We will address the above issues in the discussion of the second-order CRF-Sampler in the
next chapter. Since the experimental results indicate that using predicted secondary structure
we can dramatically improve the sampling performance of CRF-Sampler, compared to using
41
PSI-BLAST sequence profile only, we will employ those features in the second order CRFSampler.
42
Chapter 4. Second-order CRF-Sampler with
Energy
4.1 Introduction
In Chapter 3, we have described a protein conformation sampling algorithm based on the 1storder conditional random fields (CRF) and directional statistics. The CRF model is a
generalization of the HMM models and is much more powerful. Our CRF model can
accurately describe the complex sequence-angle relationship and estimate the probability of a
conformation, by incorporating various sequence and structure features and directly taking
into consideration the nearest neighbor effects. We have shown that by using the 1st-order CRF
model, we can sample conformations with better quality than Hamelryck et al.’s FB5-HMM
(Zhao et al., 2008). All these studies have demonstrated that it is promising to do template-free
modeling without using discrete representations of protein conformational space.
This chapter presents the first template-free modeling method that can search conformations
in a continuous space and at the same time achieves performance comparable to the popular
fragment assembly methods. This method differs from our previous work discussed in chapter
3 (Zhao et al., 2008) and FB5-HMM (Hamelryck et al., 2006) in that the latter two only describe
a method for conformation sampling in a continuous space, but did not demonstrate that this
sampling technique actually lead to a template-free modeling method with comparable
performance as the fragment assembly method. By contrast, in this chapter we describe a 2ndorder CRF model of protein conformational space and show that with a simple energy
function, the 2nd-order CRF model works well for template-free modeling. We will show that it
is necessary to use the 2nd-order model instead of the 1st-order model described in our previous
work since the former can dramatically improve sampling efficiency over the latter, which
makes the 2nd-order model feasible for real-world template-free modeling. Blindly tested in the
CASP8 evaluation, our CRF method compares favorably with the Robetta server (Misura et al.,
2006; Simons et al., 1997), especially on alpha and small beta proteins. Our method also
generated 3D models better than template-based methods for a couple of CAP8 hard targets.
4.2 Methods
A 2nd-order CRF model of protein conformation space
We have described a 1st-order CRF model for protein conformation sampling in Chapter 3.
Here we extend our 1st-order CRF model to a 2nd-order model to more accurately capture local
sequence-angle relationship.
43
Given a protein with solved structure, we can calculate its backbone angles at each position
and determine one of the 100 groups (i.e., states or labels) in which the angles at each position
belong. Each group is described by an FB5 distribution. Let S = @ + , . , … , +CC D 7 ∈ denote
such a sequence of states/labels (i.e., FB5 distributions) for this protein. We also denote the
sequence profile of this protein as M and its secondary structure as X. As shown in Figure 9,
our CRF model defines the conditional probability of S given M and X as follows.
JR |, L =
exp∑8
7T+ S , , L,
U, L
4.1
where Λ = )G+ , G. , … , GH 0 is the model parameter and
8
U, L = Z exp vZ S , , L, w
7T+
is a normalization factor summing over all the possible labels for the given M and X.
Comparing to Equation 3.1, we will use both the sequence profile X and predicted secondary
structure M as our observation features, since we have learned from the results in Chapter 3
that this combination contains information for generating higher percentage of good decoys.
F(S, M, X, i) consists of two edge features and two label features at position i. It is given by
S , , L, = \+
7]+ , 7
7=
+ Z _. )
T^]
+ \.
7]+ , 7 , 7=+
7]+ , 7 ,
,L 0
7=
+ Z _+ ) 7 , , L 0
T^]
4.2
where e1(si−1, si) and e2(si−1, si, si+1) are the 1st-order and 2nd-order edge feature functions,
respectively. v1(si, Mj, Xj) and v2(si−1, si, Mj, Xj) are the 1st-order and 2nd-order label feature
functions, respectively. If we remove e2(si−1, si, si+1) and v2(si−1, si, Mj, Xj), then we can get a 1storder CRF model.
The two edge functions model local conformation dependency, given by
\+
\.
7]+ , 7
= GℎY , ℎ′′`
7]+ , 7 , 7=+
7]+
= ℎ′a`
= GℎY , ℎYY , ℎ′′′`
7
7]+
= ℎ′′a
= ℎ′a`
7
= ℎ′′a`
44
7=+
= ℎ′′′a
4.3
4.4
Meanwhile, [si = h] is an indicator function, which is equal to 1 if the state at position i is ℎ ∈ ,
otherwise 0; GℎY , ℎYY is a model parameter identified by two states h′ and h″; and GℎY , ℎYY , ℎ′′′
is a model parameter identified by three states. The two label feature functions are given by
_+ ) 7 , , L 0 = Z Z G − 1, , , ℎL `
X
_. )
7]+ , 7 ,
+ Z G − , , ℎ `
,L 0
7
= ℎa
7
= ℎa + Z G − , , ℎL `
X
= Z Z G − , , , ℎY , ℎ′′L `
X
+ Z G − 1, , ℎY , ℎ′′L `
X
7]+
+ Z G − 1, , ℎY , ℎ′′ `
= ℎ′a`
7]+
7
7]+
= ℎY a`
= ℎ′′a
= ℎ′a`
7
= ℎ′′a
7
7
= ℎa
4.5
= ℎYY a
4.6
The label feature functions model the dependency of backbone angles on protein sequence
profiles and predicted secondary structure. Equation 4.5 and 4.6 indicate that not only the state
(i.e., angle distribution) itself but also the state transition depend on sequence profiles and
predicted secondary structure. As shown in the third and fourth items in the right hand side of
Equation 4.2, the state (or state transition) at one position depends on sequence profile and
secondary structure in a window of width 2w + 1 where w is set to 4 in our experiments. It will
slightly improve sampling performance by setting the window size larger. Since secondary
structure is predicted from sequence profiles, the former is not independent of the latter.
Therefore, we need to consider the correlation between sequence profiles and predicted
secondary structure, as shown in the first items of the right hand sides of Equation 4.5 and 4.6.
The model parameters for the label features are identified by one or two states, secondary
structure type, amino acid identity, and the relative position of the observations.
The 2nd-order CRF model has millions of features, each of which corresponds to a model
parameter to be trained. Once this model is trained, we can use it to sample protein
conformations in a continuous space. Coupled with an energy function and a folding
simulation method, we can also use it for template-free modeling.
45
Figure 9 A second-order Conditional Random Fields (CRF) model of protein conformation
space. Each backbone angle state depends on a window (size 9) of sequence profiles and
secondary structure and also the states in its local neighborhood.
46
Model parameter training
Given a set of m proteins with sequence profile Mi, predicted secondary structure Xi and
corresponding backbone angles 7 = 1,2, … , , our CRF model trains its parameter
= FG+ , G. , … , GH Iby maximizing the conditional log-likelihood L of the data:
e = Z log i |i , L i −
i
1
‖‖+
2j .
4.7
The second item in Equation 4.7 is a regularization factor to deal with the sparsity in the
training data. When the complexity of the model is high (i.e., the model has many features and
parameters) and the training data is sparse, overfitting may occur, and it is possible that many
models can fit the training data. Our 2nd-order CRF model has around one million of
parameters;, we place a Gaussian prior exp m−
+
.no
∑[ G[ p on the model parameter to choose the
model with a “small” parameter in order to avoid overfitting. This regularization can improve
the generalization capability of the model in both theory and practice (Vapnik, 1998). This kind
of training is also called discriminative training or conditional training. Different from the
generative training in the FB5-HMM model, discriminative training directly optimizes the
predictive ability of the model while ignoring the generative probability of the observation.
The objective function in Equation 4.7 is convex and hence theoretically a globally optimal
solution can be found using any efficient gradient-based optimization technique. There is no
analytical solution to the above equation for a real-world application. Quasi-Newton methods
such as L-BFGS (Liu and Nocedal, 1989) can be used to solve the above equation and usually
can converge to a good solution within a couple of hundred iterations. For a detailed
description of how to train a CRF model, see elsewhere (Lafferty et al., 2001; Sha and Pereira,
2003). We revised the FlexCRFs program (Phan et al., 2005) to train our CRF model, and it
takes approximately 24 hours to train a single model on a cluster of 150 2GHz CPUs.
We used a set of ∼3000 non-redundant proteins to train the parameters in our CRF model. Any
two proteins in the training set share no more than 30% sequence identity, and the resolution
of a training protein is at least 2.0Å. To avoid overlap between the training data and the test
proteins (i.e., the proteins in Table 10 - 12), we removed the following proteins from our
training set: (1) the proteins sharing at least 25% sequence identity with our test proteins; (2)
the proteins in the same fold class as our test proteins according to the SCOP classification; and
(3) the proteins having a TM-score 0.5 with our test proteins in case some recently released
proteins do not have a SCOP ID. If the TM-score of two protein structures is smaller than 0.5,
then a threading program such as PROSPECTOR_3 cannot identify their similarity
relationship with high confidence.
47
The training set is randomly divided into five sets of same size and then used for five-fold
cross validation. We trained our CRF model using three different regularization factors (i.e., σ2
in Equation 4.7): 25, 125, and 625, and chose the one with the best F1-value. F1-value is a
widely used measurement of the prediction capability of a machine learning model. F1-value
is an even combination of precision p and recall r and calculated as
.H
H=
. The higher the F1-
value is, the better the CRF model. The average F1-values for regularization factors 25, 125,
and 625 are 21.73%, 21.55%, and 22.03%, respectively. In terms of F1-value, the difference
among these regularization factors is small. Therefore, we choose a small regularization factor
25 to control the model complexity since a model with lower complexity usually generalizes
better to the test data. The regularization factor is the only parameter that we need to tune
manually. All the other model parameters (i.e., weights for features) can be estimated
automatically in the training process.
Conformation sampling and resampling
Initial conformation sampling
Once the CRF model is trained, we can sample a protein conformation or resample the local
conformation of a segment by probability using a forward-backward algorithm. We first
sample labels (i.e., angle distribution) by probability estimated from our CRF model and then
sample real-valued angles from the labels. Let V(i, Si, Si+1) denote the sum of all the edge
features associated with edge (si, si+1) and all the label features associated with labels St, Si+1 and
(Si, Si+1). Let G(i, St, Si+1) denote the marginal probability of a label pair (Si, Si+1). We can
recursively calculate G(i, St, Si+1) from N-terminal to C-terminal as follows.
u , 7 ,
7=+
u0,
C, +
= \ C,X ,X<
= \ 7,X9 ,X9< Z u − 1,
X9;<
7]+ , 7 \
yX9;< ,X9 ,X9<
where λ (si−1, si, si+1) can be interpreted as state transition log-likelihood. Once G(N−1, sN− 1, sN) is
calculated where N is the protein size, we can sample a conformation from C-terminal to Nterminal. First, we sample a label pair (sN−1, sN) for the last two positions by probability
u6 − 1, 8]+ , 8
∑X;< ,X u6 − 1, 8]+ ,
Then we sample the label si for position i by probability
48
8
.
u , 7 , 7=+ \ yX9 ,X9< ,X9o
,
∑X9 u , 7 , 7=+ \ yX9 ,X9< ,X9o
supposing that the sampled labels at position i+1 and i+2 are si+1 and Si+2, respectively.
Conformation resampling
The algorithm for resampling the local conformation of a randomly chosen segment is similar.
We first randomly determine a segment for which we are going to resample backbone angles.
Then we resample the angles for this segment using a forward-backward algorithm similar to
the initial conformation sampling algorithm. The major difference is that in this scenario we
calculate G(i, Si, Si+1) for a segment conditioning on the labels of the two residues flanking this
segment at the left and when do resampling we also have to consider the two residues
flanking this segment at the right.
Biased sampling
Our sampling method works well in alpha regions but not in beta regions. We decided that we
should do more frequent sampling in beta and loop regions than in alpha regions. This is
because both beta and loop regions are more varied than alpha regions. By sampling in the
beta and loop regions more frequently, we can generate decoys with better quality. We achieve
this goal by empirically assigning different weights to each position depending on its
predicted secondary structure type. The weights for alpha, beta and loop regions are 1, 5, and
3 respectively. These weights are empirically determined using a simple enumeration method
on the test proteins in Table 10. To determine which segment with angles to be resampled, we
first uniformly sample the segment length l between 1 and 15. Then we sample the starting
position of this segment using biased sampling. We calculate the weight of a segment as the
sum of the weights of all the positions in this segment. Then we randomly sample a segment
with the length l by probability proportional to the weight of this segment.
Biased sampling is employed only when we do folding simulations using the energy function
described in this work. In the case that the energy function is not used, we still use uniform
sampling.
Energy function
The energy function we used for folding simulation consists of three items: DOPE, KMBhbond,
and ESP. The weight factors combining these three energy items are trained on the proteins in
Table 10 using grid search in a progressive way. First, we fix the weight factor of DOPE to 1
and determine the weight factor for ESP by minimizing the average RMSDs of generated
49
decoys. Then we fix the weight factors of both DOPE and ESP and determine the weight factor
for KMBhbond using the same way.
DOPE
DOPE is a full-atom, distance-dependent pairwise statistical potential originally designed by
Shen and Sali and then improved by the Sosnick group (Fitzgerald et al., 2007; Shen and Sali,
2006). DOPE performs as well or better than many other statistical potentials and force fields
in differentiating a native structure from decoys. The statistical potential in DOPE
distinguishes the amino acid identity and atomic identity of two interacting particles. In our
folding simulation, we only build coordinates for main chain and Cβ atoms, so only the
statistical potentials related to main-chain and Cβ atoms are used to calculate the energy of a
conformation. We denote this revised DOPE as DOPE-Cβ. According to (Fitzgerald et al., 2007),
DOPE-Cβ is highly correlated with the full-atom DOPE. DOPE-Cβ also performs favorably in
applications to intra-basin protein folding (Colubri et al., 2006).
Hydrogen bonding
KMBhbond is a statistical potential for hydrogen bonding developed by the Baker group
(Morozov et al., 2004). It depends on the distance between the geometric centers of the N–H
bond vector and the C=O bond vector, the bond angle between the N–H bond vector and the
hydrogen bond, the bond angle between the C=O bond vector and the hydrogen bond, and the
dihedral angle about the acceptor-acceptor base bond. The three angles describe the relative
orientation of the bond vectors in the hydrogen bond.
ESP
ESP is an approximation to the Ooi-Scheraga solvent-accessible surface area (SASA) potential
(Li et al., 2008). Since our conformation representation does not contain side-chain atoms,
which are necessary for the calculation of the solvent-accessible surface area potential, we
employ a simple ESP that assigns each residue with an environmental energy score. ESP is a
function of the protein size and the number of Cα atoms contained within an 8.5-Å sphere
centered on the residue's Cα atom (Fernandez et al., 2002). Explicitly, the ESP statistical
potential has the form given by
ESP, = − ln
J|,
J|
where n is the number of Cα atoms in an 8.5-Å sphere centered on the Cα atom of the residue, R
is the radius of gyration of the protein, aa is the amino acid identity of the residue, P(n|R) is
the number of Cα atoms in an 8.5-Å sphere for a given protein radius regardless of amino acid
50
identity, and P(n|R,aa) is the number of Cα atoms in an 8.5-Å sphere for a given protein radius
and amino acid identity. We calculate ESP(aa, n) from a set of ∼3000 non-redundant
experimental structures chosen by the PISCES server(Wang and Dunbrack, 2003). Each protein
in this set has resolution at least 2.0 Å, R factor no bigger than 0.25, and at least 30 residues.
Any two proteins in this set share no more than 30% sequence identity.
To parameterize the ESP potential, we need to discretize the radius of gyration R, which
ranges from 7Å to 39Å in our training set. We tested the following three discretization schemes:
(1) R is discretized into 65 bins with equal width 0.5Å; (2) R is discretized into 33 bins with
equal width 1Å; and (3) R is first discretized into 33 bins with equal width 1Å. Then we merge
[7, 9), [34, 36) and [37, 39] into a single bin, respectively, to guarantee sufficient statistics for
these intervals. We calculated the Pearson correlation coefficient between the resultant ESP
energies and TM-score of the decoys. The third scheme yields the best correlation and thus is
used in our energy function.
Energy minimization
We employ a simulated annealing (SA) algorithm to minimize the energy function for a given
protein. The SA routine is based on the algorithm proposed by Aarts and Korst in 1991 (Aarts
and Korst, 1991). We start with sampling an initial conformation and then search for a better
one by minimizing the energy function. Given a conformation, we propose a new
conformation by resampling the local conformation of a randomly-chosen small segment using
the CRF model. The new conformation is rejected if there are serious steric clashes among
atoms; otherwise, it is accepted with probability min)1, ]/i 0 where ΔE is the energy
increment and t is the annealing temperature.
The initial annealing temperature is chosen so that at the beginning of the annealing process
an energy increase is accepted with a given probability p0 (=0.8). The initial temperature t0 is
determined by C = −
where ΔE is the average energy increase. To determine ΔE, we
H
first conduct a series of trial conformation samplings and accept all the generated
conformations. Then we estimate ΔE by calculating the average energy increase observed in
our trial samplings.
During the folding simulation process, we decrease the annealing temperature gradually using
an exponential cooling schedule. The temperature is updated by tk+1 = 0.9tk. At each annealing
temperature, the number of sampled conformations is set to (100+N) where N is the number of
residues in the protein. This number is set to achieve thermal equilibrium. The termination of
the SA process is triggered when any of the following two conditions is satisfied: (1) either the
temperature is low enough such that almost no energy increase is accepted and the annealing
51
process is trapped at local minima; or (2) or the number of conformations generated in a single
simulation process reaches a threshold (say 10,000).
4.3 RESULTS
The 2nd-order CRF model is much better than the 1st-order model
We compare our 2nd-order CRF model with the 1st-order model described in Chapter 3 to see
how much improvement we can achieve by considering the interdependency among three
adjacent residues. To exclude the impact of an energy function in this comparison, we guide
conformation search using only compactness and self-avoiding constraints but not an energy
function (see Chapter 3 for more details). In total, we tested our models on a set of 22 proteins
with different structure properties. We generated ∼20,000 decoys for each test protein using
each CRF model and then calculated the percentage of decoys with RMSD smaller than a given
threshold, as shown in Table 10.
In terms of the best decoys, the 2nd-order model is better on 13 out of 22 test proteins and
worse on seven proteins. The best decoys may be generated by chance, so they cannot be
reliably used to evaluate the performance of the two CRF models. We further examine their
difference in terms of the percentage of decoys with RMSD smaller than a given threshold. A
general trend we observed is that when the proteins under consideration are not big (<100
residues), the 2nd-order model generally outperforms the 1st-order model by a large margin.
The only exception is 4icbA. The performance difference between these two CRF models is
small on relatively large proteins such as 1aa2, 1jer, and T056. This may be because a large
protein tends to have a large conformation space and neither CRF model can search a very
large conformation space efficiently. The reason that the 2nd-order model performs worse on
4icb is because there is a cis-proline in 4icb, and the length of the virtual Cα-bond ending at this
proline is approximately 3.2Å instead of our assumption 3.8Å. Therefore, the more accurately
can our CRF models predict the backbone angles, the more the decoys deviate from the native
structure of 4icb. It is not very difficult to resolve this issue since from PSI-BLAST sequence
profile we can predict with accuracy 92% if a residue is a cis-proline or not (data not shown).
This comparison result indicates that we can dramatically improve sampling efficiency by
using the 2nd-order CRF model.
52
Table 10 Quality Comparison of the Decoys Generated by the 1st-Order and 2ND-Order CRF
Models
Note: Columns 1–3 list the PDB code, protein size and the type of the test proteins. Columns
“best” list the RMSD (Å) of the best decoys; the other columns list the percentage of decoys
with RMSD smaller than a given threshold. “O1” and “O2” denote the 1st-order and the 2ndorder CRF models, respectively. In total, ∼20,000 decoys are generated for each protein
without an energy function.
1aa2
Size C
108 α
1beo
98
1ctfA
68
1dktA
72
1enhA 54
1fc2C
43
1fca
55
1fgp
67
1jer
110
1nkl
78
1pgb
56
1sro
76
1trlA
62
2croA
65
O1
O2
α O1
O2
αβ O1
O2
β O1
O2
α O1
O2
α O1
O2
β O1
O2
β O1
O2
β O1
O2
α O1
O2
αβ O1
O2
β O1
O2
α O1
O2
α O1
Best
7.34
7.31
6.42
5.84
3.7
3.67
6.15
5.07
2.32
2.21
1.94
2.28
4.99
4.96
7.4
5.94
9.64
10.2
3.64
3.06
3.15
2.6
6.22
5.39
3.53
3.72
2.8
≤6Å
0
0
0
0.0096
2.41
6.62
0
0.121
22.4
69.3
49.1
85.4
0.145
0.207
0
0.0048
0
0
5.91
20.3
22.3
63.2
0
0.0193
13.5
34.1
16.8
≤7Å
0
0
0.02
0.048
7.18
22.3
0.1
1.48
32.4
72
64.1
91.7
1.3
2.65
0
0.043
0
0
14.1
30.1
45
85.8
0.07
0.289
25.3
45.3
31.4
53
≤8Å
0.035
0.0145
0.2
0.385
16.2
47.1
0.87
5.94
44.7
77.7
85
97.3
6
13.5
0.035
0.582
0
0
25.1
44.2
65.1
93.9
0.525
1.54
38.5
53.5
47.9
≤9Å
0.245
0.116
0.99
1.37
31.4
64.7
3.81
16.2
61.2
87.1
97.6
99.8
19.9
36.3
0.46
4.25
0
0
45.2
65.3
81
97.7
2.75
5.09
57.4
68
63.3
≤10Å
1.05
0.97
3.27
4.5
51.7
78.9
12.5
34.3
85.4
97.4
99.7
100
49.7
68.2
4.21
18.2
0.005
0
66.6
84
93
99.5
10
14.5
85.1
94.8
79.9
≤11Å
4.27
4.99
9.7
12.5
77.2
94.8
33.4
59.3
98.5
99.9
100
100
85.5
94
20.2
48.2
0.12
0.185
86.8
96.7
98.6
99.9
27.1
32.9
99.1
100
93
≤12Å
13.5
17.7
23
29.3
95.9
99.5
62.8
82.6
100
100
100
100
99.3
99.8
54.2
81.8
0.91
1.11
97.7
99.8
99.9
100
54.7
60
99.9
100
99.6
2gb1A 56
β
4icbA
76
α
T052
98
β
T056
114
α
T059
71
β
T061
76
α
T064
103
α
T074
98
α
O2
O1
O2
O1
O2
O1
O2
O1
O2
O1
O2
O1
O2
O1
O2
O1
O2
2.58
2.91
2.04
4.63
4.4
7.58
8.37
7.78
7.57
6.3
6.21
5.36
6.04
7.23
7.47
4.86
4.22
35.4
23.3
65.2
0.515
0.125
0
0
0
0
0
0
0.01
0
0
0
0.015
0.098
52.9
45.6
86.1
2.65
0.6
0
0
0
0
0.01
0.025
0.37
0.282
0
0
0.235
0.835
54
67.4
65.8
93.8
7.64
2.85
0.01
0
0.0198
0.005
0.135
0.421
2.89
4.73
0.035
0.032
1.23
3.56
79.3
81.8
97.8
16.8
12.3
0.035
0.025
0.084
0.094
1.14
3.85
10.7
19.9
0.2
0.412
4
9.62
88.9
93.2
99.5
33.6
31
0.135
0.296
0.51
1.07
7.2
17.4
27
40.5
0.91
1.53
10.4
18.7
95.6
98.8
100
59.1
58.9
0.8
1.78
2.26
3.83
26.8
45.1
50.9
62.7
2.79
3.62
22
30.4
99.9
99.9
100
84.3
88.5
3.52
6.58
7.07
8.38
61.1
77.3
77.6
82.6
7.3
9.05
41
49.3
Comparison with FB5-HMM and fragment assembly
We further compare our two CRF models with the FB5-HMM model in (Hamelryck et al.,
2006), as shown in Table 11. Here we compare FB5-HMM and our CRF models using
PSIPRED-predicted secondary structure and sequence information as their input. For each test
protein, FB5-HMM generates 100,000 decoys, and we generated only ∼20,000 decoys. As
shown in Table 11, our CRF models can generate decoys with significantly better quality than
FB5-HMM on five out of six proteins tested in the FB5-HMM article. The only exception is 4icb,
which has been explained in the above section. The result of FB5-HMM in Table 11 is taken
from (Hamelryck et al., 2006). The significant improvement of the 2nd-order CRF models over
the FB5-HMM model lies in that in estimating the probability of the angles at one residue, the
2nd-order CRF model can directly take into consideration the effects of its neighbor residues.
Our 2nd-order CRF also models the relationship among three adjacent residues. By contrast,
FB5-HMM only takes into consideration the relationship between two adjacent residues.
Furthermore, FB5-HMM does not consider the effects of the neighbor residues when estimate
the probability of angles at one residue.
We also compare our 2nd-order CRF model with the fragment assembly method without using
energy function. We revised the Rosetta code to do conformation optimization using the
compactness and self-avoiding constraints instead of the Rosetta energy function. As shown in
Table 11, the advantage of our 1st-order CRF model over Rosetta is not obvious. However, our
2nd-order model can generate a much larger percentage of good decoys than Rosetta for five
out of six proteins. The protein 4icb is an exception, which has been explained in previous
sections. This comparison result further indicates that it is essential to use the 2nd-order model
instead of the 1st-order model for template-free modeling. In terms of the quality of the best
decoys, Rosetta is slightly better. One of the major differences between these two methods is
that our CRF model uses a more simplified representation of protein conformation than
Rosetta. That is, we use the pseudo backbone angles to represent a protein conformation while
Rosetta uses the true backbone angles (i.e., phi/psi). The phi/psi representation has almost
twice the degree of freedom as that of the pseudo backbone angle representation. This may
explain why our method tends to generate more decoys with RMSD smaller than 6 Å and the
best decoys generated by Rosetta tend to have smaller RMSD.
55
Table 11 Quality Comparison of the Decoys Generated by FB5-HMM, the 1st-Order CRF, the
2nd-Order CRF, and Rosetta
For each protein, 100,000 decoys are generated by FB5-HMM, while only ∼20,000 decoys by
each CRF model and Rosetta. No energy function is used in this comparison. Columns 1–3 list
name and PDB code, size and number of α-helices, and β-strands of the test proteins. Columns
“Good” and “Best” list the percentage of good decoys (with RMSD ≤ 6 Å) and the RMSD of the
best decoys, respectively.
Test proteins
FB5-HMM 1st-order CRF 2nd-order CRF Rosetta
PDB
L α, β Good Best Good Best Good Best
Good Best
1FC2
43 2,0 17.1
2.6
49.1
1.94 85.4
2.28
36.5
2.72
1ENH 54 2,0 12.2
3.8
22.4
2.32 69.3
2.21
44.8
1.23
2GB1
56 1,4 0
5.9
23.3
2.91 65.2
2.04
5.82
2.26
2CRO
65 5,0 1.1
4.1
16.8
2.79 35.4
2.58
17.2
2.38
1CTF
68 3,1 0.35
4.1
2.4
3.7
3.67
2.35
1.3
4ICB
76 4,0 0.38
4.5
0.51
4.63 0.125 4.4
4.51
3.9
56
6.62
Comparison with lattice model
By combining a simple energy function and our 2nd-order CRF model, we build a program,
denoted as CRFFolder, for template-free modeling. We compare CRFFolder with
TOUCHSTONE-II, a representative lattice model method developed by Skolnick group.
TOUCHSTONE-II is an excellent template-free modeling program, and its two derivatives
TASSER (Zhang and Skolnick, 2007) and I-TASSER (Wu et al., 2007a) perform very well in
both CASP7 and CASP8. We do not compare CRFFolder with the two derivatives because both
TASSER and I-TASSER use threading-generated constraints to guide conformation search,
while CRFFolder does not. Due to the limitations of computational power, we tested
CRFFolder on a set of 15 test proteins with various structure properties, which were also tested
by TOUCHSTONE-II. These test proteins have very different secondary structures and sizes
ranging from 47 to 157. We generated approximately 3000 decoys for each alpha protein, 7000
decoys for each alpha-beta protein, and 10,000 decoys for each beta protein. By contrast,
TOUCHSTONE-II used a complex energy function consisting of 21 items and generated 24,000
decoys for each test protein (Zhang et al., 2003). As shown in Table 12, CRFFolder performs
much better than TOUCHSTONE-II on all the alpha proteins except one. CRFFolder also has
comparable performance on beta and alpha-beta proteins. On larger proteins, CRFFolder is
slightly worse than TOUCHSTONE-II. This may be because the replica exchange Monte Carlo
algorithm used by TOUCHSTONE-II for energy minimization is better than the simulated
annealing algorithm used in CRFFolder. Note that since two programs use very different
clustering methods, it is not easy to compare these two programs fairly. TOUCHSTONE-II
used a program SCAR to do decoy clustering while we use MaxCluster
(http://www.sbg.bio.ic.ac.uk/maxcluster/). For the purpose of comparison, we also show the
RMSD of the best decoys and the average RMSDs of the top 1% and 2% decoys generated by
CRFFolder.
Performance in the blind CASP8 evaluation
We tested the performance of our method by participating in the blind CASP8 evaluation with
RAPTOR++, a new protein structure prediction method. Our 2nd-order CRF model was
trained before CASP8 started (in May 2008), so it is unlikely for us to overfit our model for the
CASP8 targets.
Multi-domain proteins
Some of the CASP target proteins are large and contain multiple domains. For those targets,
we parse them into several possible domains by searching through the Pfam database (Finn et
al., 2008) using HMMER (Eddy, 1998; Krogh et al., 1994). If one target protein can be aligned to
a single template, then domain parsing is skipped. In the case that there is a big chunk of the
target not aligned to any top templates, we will treat this unaligned chunk as a single target
57
and do protein modeling separately. Except the last several CASP8 targets, the models for
multiple domains are not assembled into a single coordinate system.
Comparison with Robetta
We first examine the performance of our method by comparing it with Baker's Robetta server
on some CASP8 hard targets, on which both Robetta and CRFFolder did template-free
modeling before their experimental structures were released. These hard targets have no good
templates in the PDB. It is unclear how many decoys Robetta generated for each target, but the
top five models generated by Robetta for each target are available for download from the
official website of CASP8 (http://predictioncenter.org/download_area/CASP8/server_predictions/).
Using our template-free modeling program CRFFolder, we generated ∼7000 decoys for each
target and then chose the top five models. Note that the first models chosen by CRFFolder are
not exactly the same as our CASP8 submissions since we submitted template-based models for
some of these targets.
Table 13 compares CRFFolder and Robetta in terms of the quality of the first-ranked models.
The model quality is evaluated by a program TM-score (Zhang and Skolnick, 2005b), which
generates a real number between 0 and 1 to indicate the quality of a structure model. Roughly,
the higher the TM-score is, the better the model quality. Note that in this table the domain
definition
of
T0510_D3
is
from
Zhang's
CASP8
assessment
page
(http://zhanglab.ccmb.med.umich.edu/casp8/), while others are from Robetta CASP8 web site.
58
Table 12 Performance Comparison between CRFFolder and Skolnick's TOUCHSTONE-II
Note: Columns 1–3 list the PDB code, size, and type of the test proteins. Column “Best Cluster”
lists the RMSDs of the representative decoys of the best clusters. Column “best” lists the
RMSDs of the best decoys. The first number in parentheses denotes the rank of the best cluster
and the second is the total number of clusters. Columns “1%” and “2%” list the average
RMSDs of the top 1% and 2% decoys, respectively. The results of TOUCHSTONE-II are from
(Zhang et al., 2003).
TOUCHSTONE
Target
Size Class
BestCluster
CRFFolder
BestCluster Best
1%
2%
1bw6A
56
α
4.79 (2/3)
3.82 (3/3) 2.75
3.38 3.54
1lea
72
α
5.69 (5/5)
4.10 (5/7) 3.41
4.19 4.48
2af8
86
α
11.07 (5/6)
8.9 (12/19) 7.07
8.53 8.97
256bA
106
α
3.61 (2/3)
2.75 (6/11)
3.45
1sra
151
α
2.5
3.7
10.71 (3/12) 13.95 (17/25) 10.82
13.8 14.24
6.30 (1/25) 5.55 (42/67) 4.34
5.2 5.47
1gpt
47 α β
1kp6A
79 α β
10.01 (8/14)
7.99 (1/7) 6.29
7.51 7.81
1poh
85 α β
9.10 (5/9)
8.84 (5/10) 7.49
8.7 9.04
1npsA
88 α β
6.89 (33/34) 9.91 (41/57) 7.87
9.19 9.66
1t1dA
100 α β
8.96 (7/13) 9.22 (10/13) 6.51
9.51 9.94
1msi
66
β
7.72 (19/28) 7.77 (12/15) 6.24
7.55 7.89
1hoe
74
β
9.39 (5/13) 9.87 (16/35) 7.96
10 10.37
1ezgA
82
β
11.03 (40/44) 10.42 (42/66) 9.66
10.4 10.62
1sfp
111
β
7.48 (2/18) 11.07 (5/11) 9.32
11.1 11.59
1b2pA
119
β
12.52 (31/56) 10.01 (18/25) 8.76
10.9 11.32
59
As shown in Table 13, overall CRFFolder is better than Robetta by ∼8%. Compared to the
Robetta server, our method performs very well on mainly alpha proteins, e.g., T0460,
T0496_D1 and T0496_D2. This could be expected since our CRF model can capture well the
local sequence-structure relationship and alpha helices are stabilized by local interactions
between neighbor residues. Our method also works well on small, mainly beta proteins. For
example, our method is better than Robetta on two small beta proteins T0480 and T0510_D3.
(Please note that T0480 is evaluated without removing the disordered regions at the two ends.
If the disorder regions are removed, CRFFolder is still better than Robetta by about 12 in GDT
score.) However, our method does not perform very well on some relatively large proteins
(>100 residues) with a few beta strands, e.g., T0482 and T0513_D2. This is probably because
our CRF method can only model local sequence-structure relationship while a beta sheet is
stabilized by non-local hydrogen bonding. For a small beta protein, our method can search
more thoroughly the conformation space by sampling in a continuous space and potentially do
better. However, for a large beta-containing protein, the search space is too big to be explored
in a continuous space. Another possible reason is that our energy function is not as good as the
Robetta energy function in guiding the formation of beta-sheets.
It is also worth noting that compared to Robetta, our method did better on the first domain of
T0496, a mainly-alpha protein with120 residues. According to the study in (Shi et al., 2009),
this domain target is one of the only two CASP8 targets with really new folds. Our method did
as well as Robetta on another target with new fold (i.e., T0397_D1).
Comparison with template-based methods
Our program CRFFolder can also generate 3D models better than template-based methods for
a couple of hard CASP8 targets. According to the CASP8 official assessment, if only the firstranked models are evaluated, CRFFolder produced the best model among all the CASP8
human and server groups for T0510_D3, a small alpha/beta protein with 43 residues
(http://predictioncenter.org/casp8/results.cgi). T0510_D3 is treated as a free-modeling target by
CASP8 while Grishin et al. classified it as a fold recognition target (Shi et al., 2009). We also
examined all the template-based models generated by our threading methods in CASP8 for
this target. The best template-based model has TM-score of 0.339.
60
Table 13 Performance Comparison (TM-Score) between CRFFolder and Robetta on Some
CASP8 Hard Targets
Target
Size
Class
Robetta CRFFolder
T0397_D1 70
αβ
0.25
0.258
T0460
111
αβ
0.262
0.308
T0465
157
αβ
0.243
0.253
T0466
128
β
0.326
0.217
T0467
97
β
0.303
0.364
T0468
109
αβ
0.253
0.308
T0476
108
αβ
0.279
0.25
T0480
55
β
0.208
0.307
T0482
120
αβ
0.352
0.223
T0484
62
α
0.253
0.249
T0495_D2 65
αβ
0.312
0.436
T0496_D1 110
αβ
0.235
0.293
T0496_D2 68
α
0.291
0.5
T0510_D3 43
αβ
0.147
0.352
T0513_D2 77
αβ
0.581
0.367
T0514
αβ
0.283
0.277
0.286
0.31
145
Average
61
CRFFolder also produced one of the best models for T0496_D1, better than other templatebased models. Both CASP8 and Grishin et al. classified T0496_D1 as a free-modeling target
(Shi et al., 2009). In fact, the first-ranked model we submitted for T0496_D1 is much worse than
the best decoy we generated for this target. Among the ∼7000 decoys we generated, there are
around 18% decoys with TM-score better than the first-ranked model. The best decoy has TMscore 0.475 and RMSD to native 6.592Å. By contrast, the first-ranked template-free model has
TM-score 0.293 and RMSD 11.457Å. The best template-based model generated by our
threading methods in CASP8 for this has TM-score 0.251, and RMSD 15.372Å.
We also examined all the template-based models generated by our threading methods in
CASP8 for T0397_D1, another target with really new fold (Shi et al., 2009). There are only six
template-based models with TM-score higher than the first-ranked template-free models
generated by CRFFolder. The best template-based model generated by us in CASP8 has a TMscore of 0.338, while the best template-free model generated by CRFFolder has a TM-score of
0.364. There are 6.6% decoys generated by CRFFolder have have a TM-score better than our
first-ranked template-free model.
Table 14 summarizes the results of RAPTOR++ in CASP8 free modeling targets. CASP8
defined 164 effective domains and classified them into three categories while Grishin et al
(http://prodata.swmed.edu/CASP8/evaluation/DomainsAll.First.html) defined 146 domains and
classified them into five categories. As shown in Columns 3-5, the best models generated by
RAPTOR++ for FM targets are much better than the first models. This indicates that we still
need to improve our model selection method for FM targets. As shown in Columns 5-7, for FM
targets, the best models submitted by all CASP8 servers are much better than the best
generated by RAPTOR++. This means that in addition to improve model selection, we also
need to further improve our model generation method for FM targets. We can have similar
observations when Grishin’s domain definition and classification is used.
Our template-free modeling method samples protein conformations in a continuous space
without using fragments in the PDB. Our method aims to overcome two major issues with
current popular fragment assembly and lattice model methods. These two methods may
exclude native structure from search space by sampling in a discrete space since a small
change in a backbone angle can result in a totally different fold or by assembling a protein
structure using even medium-sized fragments since a “new fold” seems to be composed of
rarely occurring super-secondary structure motifs (Andras Fisher, CASP8 talk).
Compared to the Robetta server (see Table 13), our method performs very well on mainlyalpha proteins, e.g., T0460, T0496_D1 and T0496_D2. This is not surprising since our CRF
model can capture well the local sequence-structure relationship. Our method also works well
on small mainly-beta proteins. For example, our method is better than Robetta on T0480 and
T0510_D3. However, our method does not fare well on a relatively large protein (>100
62
residues) with a few beta strands, e.g., T0482 and T0513_D2. This is probably because our CRF
method can only model local sequence-structure relationship while a beta sheet is stabilized by
non-local hydrogen bonding. Although sampling in a continuous space, our method can still
efficiently search the conformation space of a small beta protein. However, for a large protein
with a few beta sheets, the search space is too big to be explored by our continuous
conformation sampling algorithm. It is also worth to note that compared to Robetta, our
method works well on T0397_D1 and T0496_D1, which, according to Nick Grishin, are the
only two CASP8 targets with really new folds.
63
Table 14 Summarized results of RAPTOR++ predictions of free modeling targets in CASP8.
Note: The upper half table contains the results of 13 CASP8 official domains and the lower half
contains
the
results
of
35
domains
by
Grishin’s
definition
(http://prodata.swmed.edu/CASP8/evaluation/).
R1: GDT-TS score sum of the first-ranked models by RAPTOR.
RB: GDT-TS score sum of the best models submitted by RAPTOR.
RBAll: GDT-TS score sum of the best models generated by RAPTOR.
S1: GDT-TS score sum of the best first models submitted by all servers.
SB: GDT-TS score sum of the best models submitted by all servers.
FR: Top 10 First-GDT_TS avg ≥ 30
FM: Top 10 First-GDT_TS avg < 30
Category(#)
CASP8 Official FM (13)
Grishin’s
FR (30)
Definition
FM (5)
R1
393.20
997.41
106.46
64
RB
459.24
1125.73
117.58
RBAll
511.74
1242.65
131.59
S1
SB
591.34 646.47
1386.72 1456.00
157.11 168.15
4.4 Conclusion
This chapter has presented a probabilistic and continuous model of protein conformational
space for template-free modeling. By using the 2nd-order CRF model and directional statistics,
we can accurately describe protein sequence-angle relationship and explore a continuous
conformation space by probability, without worrying about that the native fold is excluded
from our conformation space. This method overcomes the following limitations of the
fragment assembly method: (1) fragment assembly samples conformations in a discrete space;
and (2) fragment assembly is not really template free since it still uses short fragments (e.g., 9mer) extracted from the PDB. Both restrictions may cause loss of prediction accuracy.
Even though we use a simple energy function to guide conformation search, our probabilistic
model enables us to do template-free modeling as well as two well-developed programs
TOUCHSTONE-II and Robetta. Both of them have been developed for many years and have
well-tuned and sophisticated energy functions. Our template-free modeling is much better
than TOUCHSTONE-II on alpha proteins and has similar performance on mainly beta proteins.
Blindly tested on some CASP8 hard targets, our method is also better than the Robetta server
on quite a few (mainly alpha and small beta) proteins but worse on some relatively large betacontaining proteins. Finally, our method also generated the 3D models for a couple of CASP8
targets (i.e., one mainly-alpha target T0496_D1 and one small alpha/beta target T0510_D3)
better than template-based methods. The good performance on alpha proteins indicates that
our 2nd-order CRF model can capture well the local sequence-structure relationship for alpha
proteins. The good performance on small beta proteins indicates that by sampling in a
continuous space we can explore the conformational space of small beta proteins more
thoroughly. To improve the performance of our template-free modeling on relatively large
beta-containing proteins, we need to further improve our probabilistic model of beta regions
and develop a better hydrogen-bonding energy item for the formation of beta sheets.
65
Chapter 5. CNF-Folder: Modeling with
Nonlinear Features
5.1 Introduction
This chapter presents a new probabilistic graphical model Conditional Neural Fields (CNF) for
ab initio protein folding. Please see the work of (Peng et al., 2009) for a detailed exposition.
CNF is similar to but much more powerful than CRF on modeling the sequential data in that
CNF can naturally model the nonlinear relationship between input and output while CRF
cannot do so. Thus, CNF can model better the sophisticated relationship between backbone
angles, sequence profile and predicted secondary structure, estimate the probability
distribution of backbone angles more accurately and sample protein conformations more
efficiently.
In addition, this work also differs from those shown in Chapter 3&4 in that (1) We developed a
Replica Exchange Monte Carlo (REMC) method for folding simulation instead of using a
simulated annealing (SA) method; (2) This work will use the position-specific scoring matrix
(PSSM) generated by PSI-BLAST as the input of our CNF model instead of using the positionspecific frequency matrix (PSFM) generated by PSI-BLAST as the input. The REMC method
enables us to minimize energy function to a lower level and thus possibly produce better
decoys. Also, it has been proved that PSSM contains more information than PSFM for
structure prediction such as secondary structure prediction. We did not use PSSM with CRF
because CRF cannot easily take PSSM as input. By contrast, we can easily feed PSSM into our
CNF model. We will show that our new method is much more effective than our previous
method and can dramatically improve sampling efficiency and we can generate much better
decoys than before on a variety of test proteins.
5.2 Methods
A 2nd-order CNF model of conformation space
The Conditional Random Fields (CRF) methods developed earlier for protein conformation
sampling use a linear combination of input features (i.e., PSI-BLAST sequence profile and
predicted secondary structure) to estimate the probability distribution of backbone angles.
This kind of linear parameterization implicitly assumes that all the features are linearly
independent, which contradicts with the fact that some input features are highly correlated.
66
For example, the predicted secondary structure is correlated with sequence profiles since the
former is usually predicted from the latter using tools such as PSIPRED (Jones, 1999). To
model the correlation between predicted secondary structure and sequence profiles, an easy
way is to explicitly enumerate all the possible combinations of secondary structure type and
amino acid identity in the linear CRF model. In fact, we can always combine some basic
features to form a complex feature. However, explicitly defining complex features may
introduce a number of serious issues. First, it will result in a combinatorial explosion in the
number of complex features, and hence, in the model complexity. It is challenging to train a
model with a huge number of parameters without overfitting. Second, explicit enumeration
may miss some important complex features. For example, the CRF model presented in
Chapter 3&4 (Zhao et al., 2008; Zhao et al., 2009) does not accurately model the correlation
among sequence information at several adjacent positions. Finally, explicit enumeration of
complex features may also introduce a large number of unnecessary features, which will
increase the running time of probability estimation.
Instead of explicitly enumerating all the possible nonlinear combinations of the basic sequence
and structure features, we can use a better graphical model to implicitly account for the
nonlinear relationship between sequence and structure. Very recently, we have developed a
new probabilistic graphical model Conditional Neural Fields (CNF) (Peng et al., 2009), which
can implicitly model nonlinear relationship between input and output. As shown in Figure 10,
CNF consists of at least three layers: one or more hidden layers, input (i.e., sequence profile
and secondary structure) and output (i.e., backbone angles) while CRF consists of only two
layers: input and output. The relationship between the backbone angles and the hidden layer
is still linear. However, the hidden layer uses some gate functions to nonlinearly transform the
+
input features into complex features. Here we use u b =
as the gate function
+= ¡¢]£
¤
where θ is the parameter vector and x a feature vector. CNF can also be viewed as the seamless
integration of CRF and neural networks (NN). The neurons in the hidden layer will
automatically extract nonlinear relationship among input features. Therefore, without explicit
enumeration, CNF can directly model nonlinear relationship between input and output. The
training of a CNF model is similar to that of a CRF, but more complicated.
We have tested this CNF model for protein secondary structure (SS) prediction from sequence
profiles. Table 15 compares the performance of various machine learning methods for SS
prediction. The results are averaged on a 7-fold cross-validation on the CB513 data set, except
that SPINE uses 10-fold cross-validation. As shown in Table 15, by using only one hidden
layer to model nonlinear relationship between output and input, CNF achieves almost 10%
relative improvement over CRF. CNF also outperforms other methods including SVMpro
(Hua and Sun, 2001), SVMpsi (Kim and Park, 2003), YASSPP (Karypis, 2006), PSIPRED (Jones,
1999), SPINE (Dor and Zhou, 2007) and TreeCRFpsi (Dietterich et al., 2004). The linear CRF is
the worst since it does not model nonlinear relationship between secondary structure and
sequence profile. This result indicates that we can indeed benefit from modeling nonlinear
67
sequence-structure relationship. We expect that using CNF, we are able to more accurately
model sequence-angle relationship and thus, to sample conformations more efficiently.
In the context of CNF, the PSI-BLAST sequence profile (i.e., position-specific scoring matrix)
and predicted secondary structure are viewed as observations; the backbone angles and their
FB5 distributions are treated as hidden states or labels. Similar to what we have discussed in
Chapter 4, let H denote the 100 groups (i.e., states or labels) generated from clustering of the
backbone angles. Each group is described by an FB5 distribution. Given a protein with solved
structure, we calculate its backbone angles at each position and determine one of the 100
groups (i.e., states or labels) to which the angles at each position belong. Let
= @ + , . , … , 8 D 7 ∈ denote such a sequence of states/labels (i.e., FB5 distributions) for
this protein. We also denote the sequence profile of this protein as M and its secondary
structure as X. As shown in Figure 10, our CNF model defines the conditional probability of S
given M and X in the same form as in Equation (4.1) as follows.
exp)∑8
7T+ S , , L, 0
J |, L =
U, L
where = )G+ , G. , … , GH , 0 is the model parameter and U, L = ∑X exp∑8
7T+ S , , L, is a
normalization factor summing over all the possible labels for the given M and X. S , , L,
consists of two edge feature functions and one label feature function at position i. It is given by
¥¦, §, ¨, © = ª« ¦©]« , ¦© + ª¬ ¦©]« , ¦© , ¦©=«
©=¯
+ Z )¦©]« , ¦© , §® , ¨® 0
®T©]¯
°. «
where \+ 7]+ , 7 and \. 7]+ , 7 , 7=+ are the 1st-order and 2nd-order edge feature functions,
respectively, and _) 7]+ , 7 , , L 0 is the label feature function. The edge functions describe the
interdependency between two or three neighboring labels. CNF is different from CRF in the
label feature function. In CRF, the label feature function is defined as a linear combination of
features. In CNF, there is an extra hidden layer between the input and output, which consists
of K gate functions (see Figure 10). The K gate functions extract a K -dimensional implicit
nonlinear representation of input features. Therefore, CNF can be viewed as a CRF with its
inputs being K homogeneous hidden feature-extractors at each position.
68
Figure 10 A 1st-order CNF model consists of three layers: input, output and hidden layer.
A 2nd-order model is similar but not shown for the purpose of simplicity. By contrast, a CRF
model consists of only input and output.
Table 15 Secondary structure prediction accuracy comparison with the current top predictors.
Note : Q3 denotes the Percentage of correct secondary structure;
Methods
CRF
TreeCRFpsi
SVMpro
SVMpsi
Q3 (%)
72.3
77.6
73.5
76.6
69
Methods
CNF
YASSPP
PSIPRED
SPINE
Q3 (%)
80.1
77.8
76.0
76.8
The label feature function of CNF is defined as follows.
_
7]+ , 7 , , L
´
= Z ±X9;<,X9 ,² u³ ) , L, 0
²T+
5.2
That is, the label feature function is a linear combination of K gate functions G. In the above
definition, w is the parameter vector and f is a vector of basic features at position i. In our
current implementation, f contains 23×9 (=207) elements, corresponding to the sequence profile
and secondary structure information in a window of size 9 centered at position i. We use
PSIPRED to predict the secondary structure of a protein from its sequence profile. PSIPRED
generates likelihood score of three secondary structure types for each residue, which is used as
the input of our CNF model.
Similar to CRF, we use the maximum likelihood method to train the model parameters such
that J |, L is maximized. That is, we maximize the occurring probability of a set of ~3000
non-redundant high-resolution protein structures. Although both the output and hidden
layers contain model parameters, all the parameters can be learned together by gradient-based
optimization. We use LBFGS (Liu and Nocedal, 1989) as the optimization routine to search for
the optimal model parameters. Since CNF contains a hidden layer of gate functions G, the loglikelihood function is not convex any more. Therefore, it is very likely that we can only obtain
a local optimal solution of the model parameters. To achieve a good solution, we run the
training algorithm several times and use the solution with the best objective function as the
final solution of the model. See (Peng et al., 2009) for a detailed description of training CNF.
Model parameter training
To do a fair comparison between our previous CRF model and this CNF model, we used
exactly same data to train both CRF and CNF models. That is, we use a set of ~3000 nonredundant proteins to train the parameters in our CNF and CRF models. Any two proteins in
the training set share no more than 30% sequence identity and the resolution of a training
protein is at least 2.0Å. To avoid overlap between the training data and the test proteins, we
removed the following proteins from our training set: 1) the proteins sharing at least 25%
sequence identity with our test proteins; 2) the proteins in the same fold class as our test
proteins according to the SCOP classification; and 3) the proteins having a TM-score (Zhang
and Skolnick, 2007) at least 0.5 with our test proteins. Finally, the training data was prepared
before CASP8 started. Therefore, we can use our CRF/CNF models to test the CASP8 freemodeling targets without worrying about bias.
The training set is randomly divided into five sets of same size and then used for 5-fold cross
validation. To train a CNF model, we shall determine the number of gate functions at the
70
hidden layer. In addition, since the CNF model contains a very large number of model
parameters, to avoid overfitting, we shall also control the model complexity. We achieve this
by regularizing the L2 -norm of the model parameters using a regularization factor. We trained
our CNF model by enumerating the number of gate functions (50, 100, 200, and 300) and
different regularization factors: 25, 50 100, and 200 to see which one yields the best F1-value.
F1-value is widely-used to measure the prediction capability of a machine learning model. F1value is an even combination of precision p and recall r and defined as 2pr/ + . The
higher the F1-value is, the better the CNF model. Our CNF model achieves the best F1-value
(23.44%) when 200 gate functions are used with regularization factor 50. By contrast, the best
F1-value achieved by our previous CRF method is 22.0%. The F1-value improvement achieved
by CNF over CRF seems not to be very big, partially because in total 100 labels are used in our
models. Later we will show that CNF can do conformation sampling much better than CRF.
Conformation sampling and resampling
Using the trained CNF model, we can sample the whole conformation of a protein or propose
a new conformation from an existing one by resampling the local conformation of a segment.
This procedure is very similar to the conformation sampling algorithm in our CRF method
described in Chapter 3&4. That is, we can use the forward-backward algorithm to first sample
labels (i.e., angle distribution) by probability estimated from our CNF model and then sample
real-valued angles from the labels. See Chapter 3 for a detailed description of the algorithm.
Replica exchange Monte Carlo simulation
The energy function we used for folding simulation consists of three items: DOPE (a pairwise
statistical potential) (Fitzgerald et al., 2007; Shen and Sali, 2006), KMBhbond (hydrogenbonding energy) (Morozov et al., 2004) and ESP (a simplified solvent accessibility potential)
(Fernandez et al., 2002). We use the weight factors previously trained for the CRF model for
these three energy items. Therefore, the energy function is not biased towards our CNF
method. The weight factor for DOPE is always fixed to 1, so only two weight factors shall be
determined. See Chapter 4 for a detailed description of weight determination.
Previously we employ a simulated annealing (SA) algorithm to minimize energy function,
based upon the algorithm proposed by Aarts and Korst (Aarts and Korst, 1991). In this work,
we employ a Replica Exchange Monte Carlo (REMC) method (Earl and Deem, 2005; Swendsen
and Wang, 1986) to minimize energy function. By using REMC, we can minimize energy
function to lower values and thus produce better decoys for most of our test proteins. Our
REMC method employs 20 replicas and the highest temperature is set to 100. The temperature
for replica i (i=1, 2, …, 20) is set to 5i. We have also tested other temperature assignment, but
have not seen much difference in terms of folding performance. Each replica consists of 24,000
time steps. At each time step a new conformation is proposed and then accepted with
71
probability min ¶1, exp m−
·9
p¸ where Δº is the energy difference between the new and old
conformations and Ti is the temperature for this replica. The conformations between two
neighboring replicas are exchanged every 30 time steps. Therefore, in total 800 conformation
exchange events will happen between two neighboring replicas during the whole folding
simulation. It will make our simulation process very inefficient if we yield only the decoy with
the lowest energy at the end of the folding simulation. To generate more decoys from a single
folding simulation, we output the final decoy of each replica as long as it has an energy value
within 15% of the lowest energy we can achieve. Experimental results indicate that on average,
each folding simulation can generate ~10 decoys.
5.3 Results
In Chapter 4, we have demonstrated that our CRF method compares favorably with the
popular fragment-based Robetta server in the CASP8 blind prediction, in this chapter we
will focus on the comparison between our CNF and CRF methods and show that our new
method is indeed superior over our previous method.
We test our new method using two datasets and compare it with our previous method. These
two datasets were used to evaluate our previous method before. The first dataset consists of 22
proteins: 1aa2, 1beo, 1ctfA, 1dktA, 1enhA, 1fc2C, 1fca, 1fgp, 1jer, 1nkl, 1pgb, 1sro, 1trlA, 2croA,
2gb1A, 4icbA, T052, T056, T059, T061, T064 and T074. These proteins have very different
secondary structure type and their sizes range from 40 to 120 residues. Some proteins (e.g.,
T052, T056, T059, T061, T064 and T074) in this dataset are very old CASP targets. Therefore, we
denote this dataset as “old testset”. The second dataset contains 12 CASP8 free-modeling
targets: T0397_D1, T0405_D1, T0405_D2, T0416, T0443_D1, T0443_D2, T0465, T0476, T0482,
T0496_D1, T0510_D3 and T0513_D2. These proteins are called free-modeling targets because a
structurally similar template cannot be identified for them using a template-based method. We
denote this dataset as “CASP8 testset”. To avoid bias, we removed all the proteins similar to
the first dataset from our training set (see section Model parameter training). Since the training
set was constructed before CASP8 started, there is no overlap between our training data and
the CASP8 testset.
Performance on the old test set
As shown in Table 16, we evaluate our CNF and CRF methods in terms of their capability of
generating good decoys. We run both methods on each test protein and generate similar
number of decoys (5000-10,000). Each decoy is compared to its native structure and RMSD to
the native is calculated for this decoy. Then we rank all the decoys of one test protein in an
ascending order by RMSD. Finally we calculate the average RMSD of the top 1%, 2%, 5% and
10% decoys, respectively. We do not compare these two methods using the best decoys
because they may be generated by chance and usually the more decoys are generated, the
72
better the best decoys will be. In terms of the average RMSD of the top 5% or 10% decoys, our
CNF method outperforms the CRF method on all test proteins except 1ctfA, 1dktA, 1fc2C and
1fgp. The CNF method reduces the average RMSD of top 10% decoys by at least 1Å for many
proteins such as 1aa2, 1beo, 1fca, 1pgb, 1sro, 2gb1A, 4icbA, T052, T056, T059, T061 and T064.
Furthermore, our CNF method dramatically reduces the average RMSD of top 10% decoys for
some proteins. For example, our CNF method reduces the average RMSD of top 10% decoys
for 4icbA from 8.0Å to 5.2Å, for T056 from 11.1Å to 7.2Å and for T061 from 7.6Å to 5.6Å. Even
for some test proteins (e.g., 1enhA, 1pgb and 2gb1A) on which the CRF method has already
performed well, our CNF method still improves a lot.
Performance on the CASP8 test set.
To further compare our CRF and CNF methods, we also evaluate them on the 12 CASP8 freemodeling (FM) targets. During the CASP8 competition, structurally similar templates cannot
be identified for these targets. Similarly, we evaluate both methods in terms of the average
RMSD of the top 1%, 2%, 5% and 10% decoys, respectively. Compared to CRF, our CNF
method does not significantly worsen the decoy quality of any of the 12 CASP8 targets.
Instead, our CNF method outperforms the CRF method on 10 of the 12 targets and yields
slightly worse performance on another two targets: T0397_D1 and T0482. In particular, our
CNF method reduces the average RMSD of the top 10% decoys by at least 1Å for the following
seven targets: T0405_D1, T0405_D2, T0416_D2, T0443_D2, T0476, T0496_D1, and T0510_D3.
Our CNF method reduces the average RMSD of top 10% decoys for T0510_D3 from 9.1 Å to 6.3
Å and for T0496_D1 from 10.1 Å to 8.1 Å. Even for T0416_D2, a target on which our CRF
method performed well, our CNF method improves the average RMSD of the top 10% decoys
by 1Å. We have also examined the average TM-score/GDT-TS of the top 10% decoys, on
average our CNF method is better than the CRF method by ~10% (data not shown due to space
limitation).
73
Table 16 Performance of the CNF and CRF methods on the old test set.
Note: Column “s/t” lists the size and secondary structure content of the test proteins. Column
“M” indicates methods. “N” and “R” represent the CNF and CRF methods, respectively.
Column “x%” lists the average RMSD (Å) of the decoys among the top x% of the generated
decoys. Column “best” lists the RMSD of the best decoys.
s/t
108
1aa2
5α
98
1beo
5α
68
1ctfA
3α3β
72
1dktA
4β
54
1enhA
3α
43
1fc2C
2α
55
1fca
4β
67
1fgp
6β
110
1jer
2α6β
78
1nkl
5α
56
1pgb
1α4β
76
1sro
6β
62
1trlA
6α
2croA
2gb1A
4icbA
M
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
best
6.0
7.1
5.5
5.6
3.6
3.3
4.5
4.5
1.5
2.1
2.0
2.1
3.2
5.0
6.4
6.6
9.6
10.0
1.8
2.3
1.4
2.2
4.2
5.1
3.2
3.9
1%
7.0
9.0
6.1
7.2
4.5
3.9
5.1
5.0
2.0
2.6
2.3
2.3
3.9
5.6
7.5
7.3
10.8
11.5
2.5
2.8
1.9
3.0
5.2
6.4
3.6
4.2
2%
7.6
9.4
6.5
7.8
4.8
4.1
5.5
5.3
2.1
2.7
2.4
2.3
4.2
5.8
8.0
7.6
11.1
11.9
2.6
2.9
2.0
3.2
5.9
6.9
3.7
4.4
5%
8.5
10.0
7.4
8.7
5.4
4.6
6.2
5.9
2.3
2.9
2.5
2.4
4.6
6.2
8.6
8.1
11.6
12.4
2.8
3.2
2.3
3.5
6.7
7.7
3.9
4.5
10%
9.2
10.4
8.3
9.3
6.1
5.2
6.9
6.6
2.4
3.0
2.6
2.4
5.0
6.4
9.1
8.6
12.1
12.8
3.0
3.4
2.6
3.7
7.4
8.4
4.1
4.7
65
N
1.8
2.2
2.3
2.4
2.5
5α
R
2.2
2.5
2.6
2.7
2.8
56
N
1α4β R
76
N
1.7
1.9
4.1
1.9
3.1
4.8
2.0
3.3
4.9
2.3
3.6
5.1
2.6
3.8
5.2
74
T052
T056
T059
T061
T064
T074
4α
98
8β
114
6α
71
7β
76
4α
103
8α
98
4α
R
N
R
N
R
N
R
N
R
N
R
N
R
5.3
7.6
8.6
4.1
7.9
5.7
6.9
2.8
5.9
6.5
5.9
3.7
5.0
6.1
8.1
9.6
4.9
9.4
6.9
8.4
3.4
6.6
7.0
7.1
5.0
6.0
75
6.5
8.5
10.0
5.3
9.7
7.3
8.7
3.7
6.8
7.2
7.5
5.4
6.4
7.3
9.1
10.7
6.1
10.3
7.7
9.2
4.6
7.2
7.5
8.2
5.9
6.7
8.0
9.6
11.3
7.2
11.1
8.1
9.6
5.6
7.6
7.9
8.9
6.3
6.9
Table 17 Performance of our CNF and CRF methods on the CASP8 test set.
Column “s/t” lists the size and secondary structure content of the test proteins. Column “M”
indicates methods. “N” and “R” represent the CNF and CRF methods, respectively. Column
“x%” lists the average RMSD (Å) of the decoys among the top x% of the generated decoys.
Column “best” lists the RMSD of the best decoys.
T0397_D1
T0405_D1
T0405_D2
T0416_D2
T0443_D1
T0443_D2
T0465
T0476
T0482
T0496_D1
T0510_D3
T0513_D2
s/t
70
7β
80
4α
112
3α6β
57
4α
86
6α
114
2α8β
157
5α8β
108
4α6β
120
3α5β
110
3α6β
M
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
N
R
best
6.4
7.0
5.0
5.7
7.1
8.5
1.4
1.6
4.8
5.6
9.3
10.4
11.0
10.2
5.3
5.9
10.7
8.8
5.7
6.3
3.0
1%
8.2
8.0
5.4
6.6
9.0
10.1
1.9
2.6
6.0
7.1
10.6
11.9
11.8
12.2
6.3
7.8
11.9
10.9
6.2
8.2
4.0
2%
8.5
8.3
5.5
6.8
9.5
10.5
2.1
2.8
6.4
7.7
10.9
12.3
12.2
12.7
6.8
8.2
12.2
11.5
6.6
8.7
4.5
5%
9.0
8.9
5.7
7.1
10.1
11.0
2.3
3.3
7.2
8.3
11.5
12.9
12.9
13.4
7.4
8.7
12.8
12.3
7.3
9.5
5.3
10%
9.4
9.4
5.9
7.4
10.5
11.5
2.6
3.6
7.9
8.7
11.9
13.4
13.5
13.9
8.0
9.3
13.2
13.0
8.1
10.1
6.3
44
N
1α3β R
4.7
7.2
7.7
8.6
9.1
77
N
2α4β R
7.5
8.0
8.4
9.3
8.7
9.6
9.1 9.5
10.0 10.4
76
We have also examined the relationship between RMSD and energy. Due to space limitation,
here we only visualize the RMSD-energy relationship for several typical targets: T0397_D1,
T0416_D2, T0476, T0482, T0496_D1 and T0510_D3, as shown in Figure 11. Note that in the
figure, we normalize the energy of a decoy by the mean and standard deviation calculated
from the energies of all the decoys of one target. By energy normalization, we can clearly see
the energy difference between the decoys generated by the CNF/CRF methods. Figure 11
clearly demonstrates that our CNF method can generate decoys with much lower energy than
the CRF method. However, decoys with lower energy might not have better quality if the
correlation between RMSD and energy is very weak. For example, our CNF method can
generate decoys for T0397_D1 and T0482 with much lower energy, but cannot improve decoy
quality for them. To improve the decoy quality for T0397_D1 and T0482, we have to improve
the energy function. By contrast, the correlation between RMSD and energy is positive for
T0416_D2, T0476, T0496_D1 and T0510_D3. Therefore, we can improve decoys quality for
these four targets by generating decoys with lower energy.
Our CNF method dramatically improves the decoy quality on T0416_D2 over the CRF method,
as shown in Figure 11(b). The underlying reason is that our CNF method can estimate the
backbone angle probability more accurately. Around half of the decoys generated by the CRF
method for T0416_D2 are the mirror images of the other half. These mirror images are
introduced by the non-native-like backbone angles around residue #31, as shown in Figure 12.
We calculated the marginal probability of the 100 angle states at these residues and found out
the native-like angle states have much higher marginal probability in the CNF model than in
the CRF model. Thus, our CNF method can sample native-like angles at these residues more
frequently than the CRF method and avoid generating a large number of mirror images. In
addition to the CNF sampling method, our energy function also helps improve the occurring
frequency of native-like angles at these residues.
77
(a) T0397_D1
(b) T0416_D2
(c) T0476
(d) T0482
78
(e) T0496_D1
(f) T0510_D3
Figure 11 The relationship between RMSD (y-axis) and energy (x-axis) for T0397_D1,
T0416_D2, T0476, T0482, T0496_D1 and T0510_D3.
The red and blue colors represent the CRF and CNF methods, respectively. See text for the
energy normalization method.
79
Figure 12 Two typical mirror images generated by the CRF method for T0416_D2. The decoys
in blue and gold represent the lower and upper regions in Figure 11(b), respectively.
80
Comparison with CASP8 models.
In order to compare our method with the CASP8 results, we cluster the decoys of the 12
CASP8 FM targets using MaxCluster (http://www.sbg.bio.ic.ac.uk/~maxcluster/index.html).
We ran MaxCluster so that for a given target, the first cluster contains ~30% of all the decoys
and the top 5 clusters in total cover ~70% of the decoys. We examine only the top 5 clusters
because CASP8 evaluated at most 5 models for a FM target. As shown in Table 18, we list the
GDT-TS of a cluster centroid, its rank among the CASP8 models and its percentile ranking
among all the decoys we generated. As shown in this table, our method did pretty well on
T0405_D1, T0416_D2, T0443_D1, T0476, T0496_D1, T0510_D3, and T0513_D2, reasonably well
on T0397_D1, T0405_D2 and T0465 and badly on T0443_D2 and T0482. Roughly speaking our
method can do well on mainly-alpha or small beta proteins, but not well on large beta
proteins. This is expected since our CNF method can model well local sequence-structure
relationship, but cannot model long-range hydrogen bonding.
Note that we generated decoys using domain definition we decided during the CASP8 season.
Therefore, our domain definition may not be consistent with the CASP8 official definition. In
this case we calculate the GDT-TS of a model using the native structure common to our
domain definition and CASP8 definition. The GDT-TS of a model is calculated using the TMscore program and may be slightly different from the CASP8 official GDT-TS.
Specific examples
In CASP8, we did prediction using the CRF method for T0476, T0496_D1 and T0510_D3, but
not for T0416_D2 because our CRF method was not ready at the beginning of CASP8. The
server model generated by our CRF method for T0510_D3 is among the best CASP8 server
models (available at http://predictioncenter.org/casp8/results.cgi). Our CNF method further
improves predictions for these four targets over the CRF method.
T0416_D2. The first and best cluster centroids have GDT-TS 69.3 and 76.8, respectively. As
shown in Figure 13(a), the best cluster centroid is better than all the CASP8 server models. In
fact the best cluster centroid is also better than all the CASP8 human models (data not shown).
The best cluster centroid also has a small RMSD 2.7Å.
T0476. The first and best cluster centroids have GDT-TS 34.2 and 35.6, respectively. Our first
and best cluster centroids for T0476 are ranked No.4 out of 66 and No. 15 out of 287 CASP8
server models, respectively. The best human model for T0476 has GDT-TS 48.3 and RMSD
7.8Å. Our best cluster centroid also has RMSD 7.8 Å.
T0496_D1. According to Grishin group, T0496_D1 is one of the only two CASP8 targets
representing new folds (Shi et al., 2009). Our first and best cluster centroids have GDT-TS 30.5
and 49.1, respectively. As shown in Figure 13(c), the best cluster centroid is significantly better
than all the CASP8 server models. In fact the best cluster centroid is also significantly better
than all the CASP8 human models. The best CASP8 model has GDT-TS only 33.96. The
smallest RMSD among the CASP8 models with 100% coverage is 11.34Å. Our best cluster
81
centroid has a pretty good RMSD 6.2Å considering that this target has more than 100 residues.
In summary, our CNF method can predict an almost correct fold for this target.
T0510_D3. The first and best cluster centroids have GDT-TS 47.7 and 51.7, respectively. The
best cluster centroid has RMSD 6.9Å. As shown in Figure 13(d), our first cluster centroid is
better than all the #1 models submitted by the CASP8 servers. If all the 321 CASP8 models are
considered, our first cluster centroid is worse than only 3 of them (There are very few human
predictions for T0510_D3) and our best centroid is ranked No. 2.
5.4 Conclusion
This chapter has presented a new fragment-free approach to protein ab initio folding by using
a recently-invented probabilistic graphical model Conditional Neural Fields (CNF). Our
fragment-free approach can overcome some limitations of the popular fragment assembly
method. That is, this new method can sample protein conformations in a continuous space
while the fragment-based methods cannot do so. This CNF method is also better than our
previous CRF method in that 1) this method can easily model nonlinear relationship between
protein sequence and structure; and 2) we can also minimize energy function to lower values.
Experimental results indicate that our CNF method clearly outperforms the CRF method on
most of the test proteins. Previously, we have compared our CRF method with the popular
fragment-based Robetta server in the CASP8 blind prediction and shown that our CRF method
is on average better than Robetta on mainly-alpha or small beta proteins (Zhao et al., 2009).
This work further confirms our advantage on mainly-alpha or small beta proteins. Since CNF
is better than CRF in modeling nonlinear sequence-structure relationship, we are going to
incorporate more information (such as amino acid physical-chemical property profile) to our
model so that we can improve sampling efficiency further.
82
Table 18. Clustering result of the 12 CASP8 free-modeling targets.
Column “GDT” lists the GDT-TS of the first and best cluster centroids. Column “CASP8 Rank”
lists the rank of the #1 cluster centroid or the best cluster centroid among the first CASP8
server models or all the CASP8 server models, respectively. Column “Internal Rank” lists the
percentile ranking (%) of a cluster centroid among all the decoys we generated for the target.
Target
T0397_D1
T0405_D1
T0405_D2
T0416_D2
T0443_D1
T0443_D2
T0465
T0476
T0482
T0496_D1
T0510_D3
T0513_D2
First Cluster
CASP8
GDT
Rank
25.7 12/60
39.2 6/63
27.0 10/62
69.3 1/53
46.9 3/64
24.8 26/59
31.3 12/65
34.2 4/66
34.2 34/65
30.5 1/59
47.7 1/54
57.7 5/50
Best Cluster
Internal
CASP8
GDT
Rank(%)
Rank
50.6
28.6 28/262
41.6
48.4 14/285
72.3
34.6 19/280
5.4
76.8 1/242
38.2
49.2 6/253
35.3
27.9 73/252
12.6
31.3 34/286
17.5
35.6 15/287
4.3
34.2 132/279
30.3
49.1 1/266
15.7
51.7 2/244
3.8
57.7 17/225
83
Internal
Rank(%)
18.8
6.5
5.1
3.5
19.7
12.1
12.6
10.0
4.3
0.4
3.3
3.8
(a) T0416_D2
(b) T0476
(c) T0496_D1
(d) T0510_D3
Figure 13 Ranking of our CNF predictions for T0416_D2, T0476, T0496_D1 and T0510_D3.
The x-axis is percentile ranking and y-axis GDT-TS. Our first and best cluster centroids are
plotted in black and magenta lines, respectively. The #1 models submitted by the CASP8
server are ordered by their GDT-TS and their percentile ranking is displayed as a cyan curve,
so are the best models from each server but as a green curve.
84
Part 2.
Statistical Potentials Using Machine Learning
Methods
85
Chapter 6. EPAD: A position-specific pairwise
distance potential
6.1 Introduction
A lot of statistical potentials are derived from the inverse of the Boltzmann law. In the
traditional position-independent distance-dependent statistical potentials (e.g., DOPE and
DFIRE), the interaction potential of two atom types a and b can be estimated as follows.
J¾ | , Á
½¾ |a, b = −¿ À
6.1
q¾
Where k is the Boltzmann constant, T is the temperature, and ¾ represents the inter-atom
distance shell `¾, ¾ + ∆¾a. Meanwhile, J¾ |a, b is the observed probability of two atoms
interacting within the distance shell and q¾ is the reference state (i.e., the expected
probability of two non-interacting atoms within the distance shell). The reference state is used
to rule out the average and generic correlation of two atoms not due to atomic interactions.
Most statistical potentials parameterize the observed atomic interacting probability by
(residue-specific) atom types and use a simple counting method to estimate it. For example,
PÄÅÆiÇ,,È
J¾ |a, b in Eq. (6.1) is often calculated by
where #O!¾, , Á is the number of
∑É PÄÅÆiÇ,,È
observed occurrences of two atoms a and b within a distance shell [¾, ¾ + ∆¾a.The distancedependent statistical potentials developed so far mainly differ from one another in estimating
the reference state (Shen and Sali, 2006; Wu et al., 2007b; Zhang and Zhang, 2010; Zhou and
Zhou, 2002). Some (e.g., DFIRE and DOPE) use analytical methods to estimate the reference
state while others use statistical methods (e.g., KBP (Lu and Skolnick, 2001) and RAPDF
(Samudrala and Moult, 1998)). Although using different reference states, these potentials do
not have very different energy curves (see Figure 2 in the RW paper (Zhang and Zhang, 2010)
and Figure 4 in the DOPE paper (Shen and Sali, 2006)). These traditional position-independent
potentials share a couple of common properties: 1) once the atom distance and types are given,
the atomic interaction potential is fixed across all proteins and residues; and 2) the atomic
interaction potentials approach to 0 when the distance is larger than 8Å.
This chapter presents a novel protein-specific and position-specific statistical potential EPAD.
We parameterize the observed probability in EPAD by the evolutionary information and
radius of gyration of the protein under consideration, in addition to atom types. EPAD
distinguishes itself from others in that it may have different energy profiles for two atoms of
given types, depending on the protein under consideration and the sequence profile context of
the atoms (i.e., evolutionary information). Evolutionary information has been extensively used
86
in protein secondary structure prediction(Jones, 1999; Wang et al., 2011), fold recognition
(Maiorov and Crippen, 1992; Panchenko et al., 2000; Peng and Xu, 2009, 2010; Sippl and
Weitckus, 1992; Skolnick et al., 2000), protein alignment (Notredame et al., 2000; Pei et al.,
2008; Wu and Zhang, 2008b; Xu, 2005; Zhang and Skolnick, 2005b), model quality assessment
(Jones and Thornton, 1996; Panchenko et al., 2000; Peng and Xu, 2010; Reva et al., 1997; Sippl,
1993) and even protein conformation sampling (Bystroff et al., 2000; Simons et al., 1997; Zhao
et al., 2008; Zhao et al., 2010). However, evolutionary information is rarely used to design a
statistical potential suitable for ab initio protein folding. Panjkovich et al have developed a
structure-specific statistical potential using evolutionary information for the assessment of
comparative models (Panjkovich et al., 2008). Nevertheless, this potential is not positionspecific and subject to a couple of restrictions: 1) it requires the knowledge of at least one
native structure in a protein family, so it cannot be applied to ab initio folding a protein with
novel fold or to the assessment of models built from distantly-related templates; and 2) it
requires at least 50 sequence homologs for sufficient statistics. By contrast, our statistical
potential is not subject to such restrictions and thus, is more widely applicable. We term our
statistical potential as evolutionary pairwise distance-dependent potential (EPAD).
Experimental results show that our position-specific statistical potential outperforms currently
many popular ones in several decoy discrimination tests. These results imply that in addition
to reference state, the observed atomic interacting probability is also critical to statistical
potentials and can be estimated much more accurately using context-specific evolutionary
information.
6.2 Methods and Experimental Procedures
Let 7 and denote two atoms of two residues at positions and j, respectively.
Let denote the sequence profile of the protein under consideration. It is generated by
running PSI-BLAST on the NR database with at most 8 iterations and E-value 0.001. is a
position-specific scoring matrix with dimension 20 × 6 where 6 is the sequence length. Each
column in is a vector of 20 elements, containing the mutation potential to the 20 amino acids
at the corresponding sequence position. The sequence profile context of the residue at
sequence position is a 20×15 submatrix of , consisting of 15 columns − 7, − 6,…, , + 1,
…, + 7. In case that one column does not exist in (when ≤ 7 or + 7 > 6), the zero vector
is used.
Let 7 and
denote position-specific sequence profile contexts at positions i and j,
respectively. Our distance-dependent statistical potential is defined as follows.
½¾ | 7 , , 7 , , ² = −¿ À
J¾ r 7 , , 7 , , ² 0
q¾ | ²
87
6.2
where k is the Boltzmann constant and T is the temperature, q¾| ² is the reference state, and
J¾|7 , , 7 , , ² is the observed probability of two atoms 7 and interacting within a
distance shell [¾, ¾ + ̾a conditioned on atom types, residue sequence profile contexts and ²
(the estimated radius of gyration of the protein under consideration). We use ² = 2.2 6 C./Í to
estimate the radius of gyration where N is the protein sequence length. Comparing to Eq. (6.1),
our statistical potential differs from the traditional position-independent potentials (e.g.,
DOPE and DFIRE) in a couple of aspects. First, the interaction potential of two atoms is
protein-specific since it depends on the evolutionary information and radius of gyration of the
protein under consideration. Second, our potential is position-specific since it is parameterized
by sequence profile contexts in addition to atom types. We use the same reference state as
DOPE (Shen and Sali, 2006), which is a finite sphere of uniform density with appropriate
radius. That is, the reference state depends on only the size of a sample protein structure (see
the section of Estimating the reference state for more details).
We cannot use the simple counting method to calculate J¾|7 , , 7 , , ² since there are
insufficient number of solved protein structures in PDB for reliable simple counting of
sequence profile contexts 7 and . Instead, we apply a probabilistic neural network (PNN)
(Specht, 1990) to estimating J¾|7 , , 7 , , ² when both 7 and are Ï atoms. PNN will
effectively learn the sophisticated relationship between inter-atom distance and sequence
profiles and yield accurate distance probability distribution. We then estimate
J¾|7 , , 7 , , ² for non- Ï atoms conditioned upon Ï distance distribution.
Distribution of ÐÑ distances in a representative set of protein structures.
Simple statistics on the PDB25 data set indicates that nearly 90% of residue pairs have
ÐÑ distance larger than 15Å and only 1% of them have ÐÑ distance less than 4Å (see Figure 14).
Estimating pairwise ÒÓ distance distribution using probabilistic neural network
(PNN).
We discretize all the Ï − Ï distance into 13 bins (3~4Å, 4-5Å, 5-6Å, …, 14-15Å, and >15Å).
Given a protein and its kth residue pair of two residues i and j, let ¾[ denote the bin into which
the distance of the kth residue pair falls into, and b[ the position-specific feature vector, which
contains sequence profile contexts 7 and
centered at the two residues i and j under
consideration and the estimated radius of gyration of the protein under consideration.
88
Figure 14 Distribution of pairwise ÐÑ distances in a representative set of protein structures.
89
We always use ² = 2.2 6 C./Í to estimate the radius of gyration for one protein where 6 is the
protein sequence length. That is, ² is independent of any 3D models including the native
structure. We do not use ² specific to a decoy because our training set does not contain any
decoys. We do not use ² calculated from the native structures either because in the realistic
settings they are unavailable.
Let ¾[ |b[ be the probability of the distance label ¾[ conditioned on the feature vector b[ .
Meanwhile, is the model parameter vector. We estimate ¾[ |b[ as follows.
¾[ |b[ =
exp) b[ , ¾[ 0
U b[
6.3
where U b[ = ∑Ç exp) b[ , ¾0 is the partition function and b, ¾ is a two-layer neural
network. Figure 15 shows an example of the neural network with three and five neurons in the
first and second hidden layers, respectively. Each neuron represents a sigmoid function
ℎb = 1/1 + expb. Therefore, we have
Ö<
b[ , ¾[ = Z
²< T+
ÇCÔ,²< ℎ Õ
Öo
Z ²+< ,²o ℎ)< ².o , b[ >0×
²o T+
6.4
Where u+ and u. are the number of gates in the two hidden layers, <. , . > denotes the inner
product of two vectors, ².o is the weight vector of the Ø.ÙÚ neuron (also known as gate) in the
2nd layer; ²+< ,²o is the weight connecting the Ø.ÙÚ neuron in the 2nd layer to the Ø+ÙÚ neuron in the
1st layer; and ÇC9 ,²< is the weight connecting the Ø+i| neuron in the 1st layer to the label ¾[ .
In the implementation, our neural network consists of two hidden layers. The first hidden
layer (i.e., the layer connecting to the input layer) contains 100 neurons and the second hidden
layer (i.e., the layer connecting to the output layer) has 40 neurons. This neural network is
similar to what is used by the Zhou group for inter-residue contact prediction (Xue et al., 2009),
which uses 100 and 30 neurons in the two hidden layers, respectively. The Zhou group has
shown that using two hidden layers can obtain slightly better performance than using a single
hidden layer. The input layer of our network has about 600 neurons, so in total our neural
network has between 60,000 and 70,000 parameters to be trained.
Model parameter training.
We use the maximum likelihood method to train the model parameter and to determine the
window size and the number of neurons in each hidden layer, by maximizing the occurring
probability of the native Ï − Ï distance in a set of training proteins. Given a training protein
t with solved experimental structure, let Ûi denote the set of pairwise residue distances and L Ù
the set of all feature vectors. By assuming any two residue pairs to be independent of one
another, we have
90
ÝÞ
Ûi |L i = Ü ¾[i |b[i
6.5
[T+
where i is the number of residue pairs in the protein t.
Given T training proteins, we need to maximize ∏·iT+ Ûi |L i , which is equivalent to the
following optimization problem.
·
min Z − log Ûi |L i + G‖ ‖..
à
iT+
·
ÝÞ
= min Z Z)− log U b[i + b[i , ¾[i 0 + G‖ ‖..
à
iT+ 7T+
6.6
Meanwhile, Gr| |r. is a e. -norm regularization item to avoid overfitting and λ is a hyper.
parameter to be determined. This optimization problem can be solved by LBFGS(Liu and
Nocedal, 1989).
It is very challenging to solve this non-convex optimization problem due to the huge amount
of training data. We generate an initial solution randomly and then run the training algorithm
on a supercomputer for about a couple of weeks. Our training algorithm terminates when the
probability of either the training set or the validation set does not improve any more. Note that
all the model parameters are learned from the training set, but not the validation set. The
validation set, combined with the training set, is only used to determine when our training
algorithm shall terminate. Our training algorithm usually terminates after 3000 iterations. We
also reran our training algorithm starting from 9 initial solutions and did not observe explicit
performance difference among these runs.
91
Figure 15 An example probabilistic neural network, in which ¦© and ¦® are the sequence profile
contexts centered at the ith and jth residues, respectively. á«â and á¬ã are the neurons in the 1st
and 2nd hidden layers.
92
Training and validation data.
We use the PDB25 set of the PISCES server (Wang and Dunbrack, 2003) early in 2011 as the
training and validation data. Any two proteins in PDB25 share no more than 25% sequence
identity. Such a set in total includes more than 6000 proteins. We randomly chose about 5000
proteins from this PDB25 set as the training and validation proteins and also make sure that
they have no overlap (i.e., > 25% sequence identity) with the Rosetta set (Qian et al., 2007) and
the Decoy ‘R’ Us set (Samudrala and Levitt, 2000). We randomly choose 3/4 of the 5000
proteins as the training data and the remaining 1/4 as the validation data, which contain ~73
million training and ~19 million validation residue pairs, respectively. It is challenging to train
our neural network model because 1) the number of training residue pairs is huge; and 2) the
distance distribution is extremely unbalanced. As shown in Figure 14, 90% of residue pairs
have Ï distance larger than 15Å and only 1% of them have Ï distance less than 4Å. It takes a
couple of weeks to train a single neural network model using 1296 CPUs on a Cray
supercomputer.
Estimating inter-atom distance distribution for non-ÒÓ main chain atoms.
We discretize the inter-atom distance of non-Cä atoms into 26 equal-width bins, each with
0.5Å. Due to limited computation resources, instead of training neural network models for
each pair of atom types, which will take months or even a year to finish, we use a different
approach to estimate the pairwise distance probability distribution for non-Cä main chain
atoms. In particular, we calculate the inter-atom distance probability distribution for non-Cä
main chain atoms conditioned upon Ï − Ï distance probability distribution. Let
JÏÏ )¾ÏÏ | 7 , , ² 0 denote the Ï − Ï distance probability distribution for residues i and j,
which can be estimated by our probabilistic neural network. Let a and b denote the amino acid
types of the residues at i and j, respectively. For the purpose of simplicity, we use 6 and å
atoms as an example to show how to calculate the observed atomic interacting probability. Let
J¾|6, å, 7 , , ² denote the distance probability distribution for the nitrogen atom in residue
i and the oxygen atom in residue j. We calculate J¾|6, å, 7 , , ² as follows.
J¾|6, å, 7 , , ² = Z
Ç>>
,È
¾ | ¾ÏÏ JÏÏ )¾ÏÏ | 7 , , ² 0
J8æ
6.7
,È
¾ | ¾ÏÏ is the conditional distance probability distribution for atom N in amino
where J8æ
acid a and O in amino acid b when the Ï distance of these two amino acids is ¾ÏÏ . Since
,È
¾ | ¾ÏÏ is position-independent, it can be estimated by simple counting.
J8æ
Estimating the reference state.
We calculate the reference state of a distance d as the probability of two random points at the
distance d within a 3D ball of uniform density. Such a reference state has been discussed in
93
detail before (Deltheli, 1919; Garcia-Pelayo, 2005; Shen and Sali, 2006). Here we briefly
introduce it for completeness. Let ² be the radius of gyration of a protein. The estimated
radius of the 3D sphere, denoted as
/
, corresponding to such a protein is = ç ² . The
è
/
probability density ¾|² for the distance ¾ between two randomly chosen points é+ and é.
inside such a 3D sphere can be calculated as follows.
)¾|² 0 =
ëëëëì. êXí ¾é
ëëëëì+ î|é
ëëëëì+ − é
ëëëëì|
êXí ¾é
. − ¾
ëëëëì. êXí ¾é
ëëëëì+
êXí ¾é
1, é = 0
where the delta function is defined by îé = ð
.
0, Oℎ\± \
6.8
To calculate )¾ |² 0, we fix point é+ and move é. around é+ within the sphere and then
consider the spherical area formed by é. , denoted as åℎ, ¾ where ℎ is the distance between
the center of the 3D ball and point é+ . See (Garcia-Pelayo, 2005) for detailed deduction of
åℎ, ¾. We only give the final result as follows.
4ò¾ . , ¾ < − ℎ
−¾
åℎ, ¾ = ñò¾¾ + − ℎ ó1 +
ô, − ℎ ≤ ¾ ≤ + ℎ
ℎ
0, ¾ > + ℎ
Since êXí ¾é
ëëëëì. = êXí ¾é
ëëëëì+ =
)¾|² 0 =
4ò
4ò/
ó 3 ô
õöí
/
, we have
.
. ÷ ¾ℎ ℎ åℎ, ¾ =
C
3¾ . ¾ − 2. ¾ + 4
16ø
6.9
Window size and the number of neurons in the hidden layers.
The window size for a sequence profile context and the number of neurons in the hidden
layers are important hyper-parameters of our probabilistic neural network. Because it is timeconsuming to train even a single neural network model for the estimation of distance
probability distribution, we determine these hyper-parameters by training a neural network
for inter-residue contact prediction, which obtains the best performance when the window size
is 15 and the numbers of neurons in the first and second hidden layers are 40 and 100,
respectively. The window size used by us is consistent with what is used by the Zhang group
(Wu and Zhang, 2008a) and the numbers of hidden neurons are not very different from what
is used by the Zhou group (Xue et al., 2009).
A small window size or number of neurons may result in a neural network with suboptimal
performance while a very large window size or number of neurons may cause overfitting.
Because it is time-consuming to train even a single PNN model for the estimation of distance
probability distribution, we determine these hyper-parameters by training a neural network
94
for inter-residue contact prediction. This takes much less time since there are fewer labels and
we also use less training data. Our neural network obtains the best performance when the
window size is 15 and the numbers of neurons in the first and second hidden layers are 40 and
100, respectively.
Following CASP definition, there is a contact between two residues if their ú − ú distance is
no more than 8Å. A pair of residues is treated as a positive example if they are in contact;
otherwise a negative example. To compare our prediction results with CASP8 and CASP9
(human and server group) submissions, we use the PDB25 set generated by the PISCES server
before CASP8 started to train our neural network. This PDB25 set contains ~3000 proteins and
any two of them share no more than 25% sequence identity. We randomly chose ¾ of the set as
the training data and the remaining ¼ as the validation data and use PSI-BLAST sequence
profiles and PSIPRED predicted secondary structures as the input features. To avoid bias, we
also used PSIPRED, PSI-BLAST and the NR database released before CASP8 in 2008. Our
method achieves the best performance when the window size is 15 and the numbers of
neurons in the first and second hidden layers are 40 and 100, respectively.
We test our neural network models on the 13 CASP8 and 28 CASP9 contact prediction targets,
all of which are template-free modeling targets. Our method obtains an e/5 average accuracy
of 57.87% on the CASP8 targets and of 36.63% on the CASP9 targets, respectively. As shown in
Figure 16, our method exceeds all the CASP8 top groups and the 10 CASP9 server groups
(represented by blue bars). Our method works particularly well for the beta proteins and
alpha-beta proteins, obtaining an accuracy of 49.30% on the CASP9 beta proteins, significantly
better than the accuracy (25.65%) on the alpha or mainly-alpha proteins.
After determining the number of neurons in the hidden layers, we also examine the Ï −
Ï distance prediction accuracy of our PNN with respect to the window size. As shown in
Table 19, the neural network model yields almost the best performance when the window size
is 15, which is consistent what is used by the Zhang group for contact prediction(Wu and
Zhang, 2008a). The results in this table are obtained from the EPAD validation data.
95
60
60
50
50
40
40
30
30
20
20
10
10
0
0
(A)
(B)
Figure 16 Inter-residue contact prediction accuracy of our PNN and the top 12 (human and
server) groups in CASP8 and CASP9.
The accuracy of our PNN is displayed by two brown lines. The predictions of the 12 groups
are
downloaded
from
the
CASP
website
(see
http://www.predictioncenter.org/download_area/). (A) The performance on the CASP8
targets. (B) The performance on the CASP9 targets and the top 2 groups are human groups
and the the remainings are the server groups.
96
Table 19 Ï − Ï distance prediction accuracy with respect to the window size
window size 9
15
19
23
distance
precision recall precision recall precision recall precision recall
<8 Å
26.79
5.97
38.04
20.1
36.84
19.6
36.23
20.17
8-9 Å
10.23
0.01
0
0
0
0
0
0
9 -10Å
16.25
0.01
25.79
0.66
23.93
0.49
22.29
0.65
10 -11Å
15.79
0.17
48.14
9.56
49.84
9.3
50.35
9.36
11 -12Å
24.03
0.01
17.92
0.22
17.58
0.19
18.37
0.2
12 -13Å
58.13
0.87
58.59
5.38
58.76
5.38
54.58
5.63
13 -14Å
11.11
0.01
25.09
0.05
22.75
0.04
22.81
0.04
>14 Å
83.65
99.9
81.38
99.7
81.36
99.7
81.44
99.63
97
6.3 Results and Discussion
Distance dependence of the statistical potentials.
To examine the difference between our potential EPAD and the popular DOPE, we plot the
potentials as a function of inter-atom distance for two atom pairs, as shown in Figure 17.
Figure 17(A) shows the DOPE interaction potential for the atom pair ALA Cä and LEU Cä and
also the EPAD interaction potential for this pair in three different positions of protein 1gvp.
DOPE has the same energy curve for this atom pair regardless of its sequence positions. In
particular, DOPE always has a favorable potential when the distance of this pair of atoms is
between 5 and 7Å and has an interaction potential close to 0 when the distance is larger than
8.0Å. By contrast, EPAD has one unique energy curve for this atom pair for each position. The
figure legend indicates the corresponding native distances (Å) between atom ALA Cä and
atom LEU Cä at the three different sequence positions. For example, the bottom curve in
Figure 17(A) visualizes the EPAD interaction potential for the ALA Cä and LEU Cä pair with
native distance 8.31Å. This curve shows that when the distance between ALA Cä and LEU Cä
is close to the native, their EPAD interaction potential is favorable. In fact, EPAD always has a
favorable potential for these three ALA Cä and LEU Cä pairs when their distances are close to
the natives.
Figure 17(B) compares the EPAD and DOPE interaction potentials for another atom pair Cys N
and Trp O in three different proteins of 1b3a, 1bkr and 1ptq. Similar to Figure 17(A), EPAD has
different interaction potentials for the same atom pair in three different proteins while DOPE
has the same potential across all proteins. In particular, EPAD has a favorable potential when
the distance between Cys N and Trp O is close to the native. Nevertheless, DOPE has a
favorable potential when their distance is between 2 and 4Å and a potential close to 0 when
the distance is >8.0Å.
In summary, our statistical potential EPAD is significantly different from currently popular
potentials such as DOPE and DFIRE. DOPE, DFIRE, RAPDF and RW have more or less similar
energy profiles for atom pairs of the same type. The difference between EPAD and any of
DOPE, DFIRE, RAPDF and RW is much larger than that among DOPE (Shen and Sali, 2006),
DFIRE (Zhou and Zhou, 2002), RAPDF (Samudrala and Moult, 1998) and RW (Zhang and
Zhang, 2010).
98
(A)
(B)
(A) The solid curve shows the DOPE interaction potential for atom ÐÑ in ALA and atom ÐÑ in
LEU. The other 3 curves show the EPAD potentials for the same atom pair in three different
positions of protein 1gvp. The legend shows the native distances of this atom pair in these
positions. (B) The curves show the DOPE and EPAD potential for atom N in Cys and atom O
in Trp in three different proteins of 1b3a, 1bkr and 1ptq.
Figure 17 Distance dependence of DOPE and our potential EPAD.
99
Performance on decoy discrimination.
We test our backbone-based potential EPAD on several decoy sets including the Rosetta set
(Qian et al., 2007), the CASP9 models, the I-TASSER dataset (Zhang and Zhang, 2010), the
CASP5-8 dataset (Rykunov and Fiser, 2010) and the Decoy ‘R’ Us (Samudrala and Levitt, 2000)
set as well as an in-house large set of template-based models. We evaluate EPAD and several
others DOPE, DFIRE, OPUS (Lu et al., 2008) and RW (Zhang and Zhang, 2010) using five
performance metrics including the number of correctly-identified natives, the ranking of the
native structures, the Z-score of the native energy, the model quality of the first-ranked decoys
and the Pearson correlation coefficient (Pearson CC) between the energy and the model
quality. The first three metrics evaluate how well a statistical potential can differentiate natives
from decoys. The Pearson CC is more important when we want to apply the potentials to
folding a protein sequence. We evaluate the model quality of a decoy using the widely-used
GDT (Zemla, 2003; Zemla et al., 1999, 2001), which compares a decoy with its native and
generates a quality value between 0 and 100. The higher the GDT, the better quality the decoy
has.
Performance on the 2007 Rosetta dataset.
The set contains decoys generated and refined by the popular fragment assembly ab initio
folding program Rosetta (Qian et al., 2007) for 58 proteins. To evaluate our potential in a more
realistic setting, for each protein we use only the 100 low-quality decoys in the set, excluding
the high-quality decoys. The average GDT of the best decoys is about 60. As shown in Table
20, our EPAD, which currently considers only backbone atoms, correctly identifies 34 native
structures with the lowest Z-score (-2.46), while two full-atom potentials DOPE and DFIRE can
identify only 21 natives. EPAD also exceeds DFIRE and DOPE in terms of the average ranking
of the native structures (15.70 vs. 23.71 and 21.59). In terms of the average per-target Pearson
correlation coefficient (CC) between the energy and GDT, EPAD (-0.42) is significantly better
than DOPE (-0.32) and DFIRE (-0.25). EPAD also exceeds RW by all the 5 metrics.
EPAD compares favorably to OPUS-PSP, a full-atom statistical potential. OPUS can correctly
identify many more native structures than EPAD, but it has a very low Pearson CC to decoy
quality, which indicates that OPUS-PSP may not be good for ab initio folding. Since EPAD does
not contain side-chain atoms, we simply build a full-atom potential EPAD2 by linearly
combining EPAD with the side-chain component in OPUS-PSP (with equal weight). EPAD2
significantly outperforms DOPE, DFIRE, RW and OPUS-PSP by all the 5 metrics. EPAD2
greatly outperforms EPAD in correctly recognizing the native structures, which may imply
that side-chain information is very helpful for the identification of the native structures. This
trend is also observed on other datasets (e.g., I-TASSER and CASP5-8).
100
Table 21 compares the performance of several statistical potentials when only Cä atoms are
considered in scoring a decoy. Again, EPAD significantly outperforms DOPE, DFIRE and
OPUS. EPAD even performs as well as the full-atom potentials DOPE, DFIRE and RW. To
excluding the impact of the different datasets used to build EPAD and DOPE, we rebuild a
DOPE using the EPAD training data, denoted as MyDope. MyDope performs slightly worse
than DOPE, possibly because we do not fine-tune MyDope. However, EPAD performs
significantly better than both DOPE and MyDope. This indicates that EPAD outperforms
DOPE not due to the training data, but due to the novel methodology.
Performance on template-based models.
To examine the performance of EPAD on template-based models, we constructed a large set of
3600 protein pairs, denoted as InHouse, with the following properties: 1) any two proteins in a
pair share less than 25% sequence identity; 2) all protein classes (i.e., alpha, beta and alphabeta proteins) are covered by this set; 3) the protein lengths are widely distributed; 4) the
structure similarity of two proteins in a pair, measured by TM-score, ranges from 0.5 to 0.8 and
is almost uniformly distributed; and 5) within each protein pair, one protein is designated as
the target protein and the other as the template protein. Any two target proteins share less
than 25% sequence identity. We generated four different target-template alignments for each
protein pair using our in-house protein structure alignment program DeepAlign, our two
threading programs BoostThreader (Peng and Xu, 2009) and CNFpred and HHpred (Soding,
2005). CNFpred is an improved version of BoostThreader, replacing regression trees in the
latter with neural networks and making use of more protein features. Then we used
MODELLER (N. Eswar, 2006) to build four different 3D models for each target protein based
upon the four alignments, respectively. MODELLER also builds models for the unaligned
regions. To remove overlap with the training data, we constructed 4 subsets Set1, Set2, Set3
and Set4 of InHouse as follows. Set1 contains no target proteins sharing >25% sequence identity
with the EPAD training proteins. Set2 is a subset of Set1, containing no target proteins sharing
>25% sequence identity with the EPAD validation proteins. Set3 contains no target proteins
with a BLAST E-value<0.01 with the EPAD training proteins. Set4 is a subset of Set3,
containing no target proteins with a BLAST E-value<0.01 with the EPAD training and proteins.
In total, Set1, Set2, Set3 and Set4 contain 1139, 331, 965, and 266 protein pairs, respectively.
Table 22 lists the performance of several statistical potentials in identifying the 3D models with
the best GDT in the five datasets: InHouse, Set1, Set2, Set3 and Set4. As shown in Table 22,
EPAD is able to recognize many more the best template-based models than the others, which
are no better than random guess. Further, EPAD has similar performance on the 5 sets, which
confirms that our probabilistic neural network (PNN) model is not over-trained. For over 95%
of protein pairs in InHouse, the 3D models built from the structure alignments have the best
GDT. This implies that except EPAD, the other potentials are not able to differentiate structure
alignments from threading-generated alignments.
101
Table 20 Performance of EPAD and several popular full-atom statistical potentials on the
Rosetta decoy set.
Numbers in bold indicate the best performance. The per-target Pearson CC is calculated
between the energy and GDT and then the average value is reported in the table.
EPAD DOPE DFIRE OPUS RW
EPAD2
#natives identified 34
21
21
39
21
46
ranking of native
15.7
23.7
21.6
9.8
23.9
13.4
first-ranked GDT
51.6
49.7
49.4
49.7
48.5
52.4
Pearson CC
-0.42
-0.32
-0.25
-0.20
-0.32 -0.39
Z-score
-2.46
-1.61
-1.67
-3.27
-1.51 -3.28
Table 21 Performance of EPAD and several popular statistical potentials on the Rosetta decoy
sets when only Cä atoms are considered.
Numbers in bold indicate the best performance. MyDope is a re-compiled DOPE using the
EPAD training data.
EPAD DOPE DFIRE MyDope OPUS
#natives identified 33
11
12
10
6
Ranking of native
15.8
18. 7
30.7
21.7
55.3
first-ranked GDT
51.2
47.0
47.8
48.2
45.9
Pearson CC
-0.40
-0.24
-0.20
-0.21
-0.15
Z-score
-2.45
-1.51
-0.66
-1.23
0.25
102
Table 22 Performance of EPAD and several popular statistical potentials on the template-based
models.
Only EPAD is a backbone-based potential while the others are full-atom potentials. In each cell,
the numbers in and out parenthesis are the number and percentage of correctly-identified
models (i.e., models with the lowest energy and the best GDT). Bold numbers indicate the best
performance.
EPAD
DOPE
DFIRE
OPUS
RW
InHouse Set1
Set2
Set3
Set4
1903
617
178
514
143
(53%)
(54%) (54%) (53%) (54%)
900
288
(25%)
(25%) (25%) (26%) (28%)
936
286
(26%)
(25%) (26%) (26%) (28%)
900
289
(25%)
(25%) (22%) (26%) (26%)
762
248
(21%)
(22%) (21%) (23%) (22%)
82
86
73
68
103
252
253
251
218
74
74
69
60
Performance on the CASP9 models.
To further examine the performance of EPAD, we compile a test set from the CASP9 models
submitted by the top 18 servers. We exclude the CASP9 targets with many domains since some
servers do not place the models of all the domains in a single coordinate system. These 18
servers are BAKER-ROSETTASERVER (Raman et al., 2009), chunk-TASSER (Zhou et al., 2009),
chuo-fams (Kanou et al., 2009), CLEF-Server (Shao et al., 2011), FAMSD (Kanou et al., 2009),
gws (Joo et al., 2009), HHpredA (Hildebrand et al., 2009), Jiang_Assembly (Hu et al., 2011),
MULTICOM-CLUSTER (Tegge et al., 2009), MULTICOM-NOVEL (Tegge et al., 2009), Pcomb
(Larsson et al., 2008), Phyre2 (Kelley and Sternberg, 2009), pro-sp3-TASSER (Zhou and
Skolnick, 2009), QUARK (Xu et al., 2011), RaptorX (Peng and Xu, 2011), Seok-server (Lee et al.,
2010), Zhang-Server (Xu et al., 2011), ZHOU-SPARKS-X (Yang et al., 2011). We do not include
the models from RaptorX-MSA, RaptorX-Boost, HHpredB, HHpredC, MULTICOM-REFINE,
MULTICOM-CONSTRUCT and Jiang_THREADER since they are not very different from
some of the 18 servers. In summary, this CASP9 dataset contains the first models submitted by
18 servers for 92 targets. This set is very challenging for any energy potentials because the
models submitted by these top servers have similar quality especially for those not-so-hard
targets. The first-ranked models by EPAD, DOPE, DFIRE and RW have an average GDT of
58.6, 55.7, 56.0, and 57.4, respectively. The average Pearson correlation coefficient (between
GDT and energy values) for EPAD is -0.364, which is significantly better than DOPE (-0.25),
DFIRE (-0.23) and RW (-0.28). Note that RW parameters are fine-tuned using the CASP8 and
CASP9 models while EPAD, DOPE and DFIRE are independent of any decoy sets. In addition,
EPAD is only a backbone-based potential while the other three are full-atom potentials.
Table 23 shows the performance of EPAD, DOPE, DFIRE and RW with respect to the hardness
of the targets, which is judged based upon the average GDT of all the models of this target. We
divide the targets into four groups according to the average GDT: <30, 30-50, 50-70, >70. EPAD
performs very well across all difficulty levels and has particularly good correlation coefficient
for the targets with average GDT less than 30. Even for easy targets EPAD also outperforms
the others although it is believed that sequence profiles are not very effective in dealing with
easy targets. The only exception is that EPAD has a worse average GDT of the first-ranked
models than RW for the targets with average GDT between 30 and 50. This is because RW
performs exceptionally well on a single target T0576. The best model identified by RW has
GDT 53.3 while EPAD, DOPE and DFIRE can only identify a model with GDT 17.0.
Performance on the Decoy ‘R’ Us dataset.
The set is taken from http://dd.compbio.washington.edu/, containing decoys for some very
small proteins. In terms of the average rank of the native structures EPAD significantly
exceeds the others, but EPAD correctly identifies slightly fewer native structures than DOPE
and OPUS_PSP, in part because EPAD does not include side-chain atoms.
104
The results of DOPE and DFIRE in Table 24 are taken from (Shen and Sali, 2006), that of
OPUS_PSP is from (Lu et al., 2008), and that of RW is calculated by the program in (Zhang and
Zhang, 2010). As shown in the table, all the energy functions are able to correctly differentiate
the native structures from decoys for the proteins in Lattice_ssfit. EPAD identifies more native
structures for the proteins in Lmds, while DOPE and OPUS_PSP do so for 4state_reduced and
Fisa_casp3. In terms of the average rank of the native structures EPAD significantly exceeds
the others, but EPAD correctly identifies slightly fewer native structures than DOPE and
OPUS_PSP, in part because EPAD does not contain side-chain information.
Performance on the I-TASSER dataset.
This
set
contains
decoys
for
56
proteins
generated
by
I-TASSER
(http://zhanglab.ccmb.med.umich.edu/). The average TMscore of the decoys in this set ranges
from 0.346 to 0.678. EPAD outperforms DFIRE and DOPE by 5 measures. EPAD is slightly
better than RW in terms of the first-ranked TMscore and the correlation, but slightly worse
than RW in terms of the Z-score of the natives. EPAD2 (i.e., the combination of the OPUS-PSP
side-chain potential and EPAD) can obtain much better Z-score of the natives although the
correlation is slightly decreased. This is consistent with what is observed on the Rosetta set.
As shown in Table 25, our statistical potential EPAD outperforms DFIRE and DOPE no matter
which measure is used. EPAD is slightly better than RW in terms of the first-ranked TMscore
and the correlation, but slightly worse than RW in terms of the Z-score of the natives. EPAD2
(i.e., the combination of the OPUS-PSP side-chain potential and EPAD) can obtain much better
Z-score of the natives although the correlation is slightly decreased. This is consistent with
what is observed on the Rosetta set.
105
Table 23 Performance of statistical potentials with respect to the hardness of the CASP9 targets.
To save space, DOPE and DFIRE are denoted as “DP” and “DF”, respectively, and we also
omit the negative sign of the correlation coefficient. The first column indicates the hardness of
the targets, judged by the average GDT of all the models of the target.
GDT of the first-ranked models Correlation Coefficient
EPAD
DP
DF
RW
EPAD DP DF RW
27.4
0.44
<30
23.1
24.1
25.8
0.28 0.23 0.33
42.7
0.31
30-50 42.0
40.0
40.6
0.26 0.24 0.27
0.37
50-70 64.4
61.6
61.5
63.4
0.24 0.22 0.27
80.0
0.35
>70
77.1
77.1
77.1
0.26 0.26 0.26
Table 24 Performance of several statistical potentials on the Decoy ‘R’ Us dataset.
In each entry of columns 2-6, the numbers in and out parenthesis are the numbers of correctlyidentified native structures and the average rankings of the natives, respectively. Numbers in
bold indicate the best performance.
Decoy Set
EPAD
DFIRE
DOPE
OPUS_PSP RW
#targets
4state_reduced 5/1.4
6/1.4
7/1
7/1
6/1.4
7
Fisa
3/65.5
3/64.3
3/94.5
3/78.8
3/17.0
4
Fisa_casp3
2/1.3
3/1
3/1
3/1
3/1
3
Lattice_ssfit
8/1
8/1
8/1
8/1
8/1
8
Lmds
9/51.0
7/143.9 7/101.8 8/91.8
All
27/24.8 27/55.4 28/45.6 29/39.1
106
7/141.6 10
27/47.0 32
Table 25 Performance of several statistical potentials on the I-TASSER dataset.
CC-TM is the correlation coefficient between energy and the quality, measured by TMscore, of
the decoys. Numbers in bold indicate the best performance.
EPAD DFIRE DOPE RW
EPAD2
#natives identified
53
first-ranked TMscore 0.579
47
30
53
53
0.558
0.560
0.569
0.564
CC-TM
-0.532
-0.492
-0.317
-0.500
-0.512
Z-score of natives
-3.61
-3.58
-2.18
-4.42
-5.94
107
Performance on the CASP5-8 dataset.
EPAD is only second to QMEAN6, RW and RWplus in ranking the best models in the absence
of the native structures. When the native structures are included, EPAD does not perform as
well as when the native structures are not included. EPAD2 outperforms all the others in terms
of the average ranking of the best models in the absence of the native structures or the average
ranking of the native structures. EPAD2 also performs very well in terms of the number of
correctly-identified models (or native structures). These results may further indicate that sidechain information is needed for the accurate identification of the native structures.
Table 26 shows the performance of a variety of potentials on the CASP5-8 models selected by
Rykunov and Fiser (Rykunov and Fiser, 2010). As shown in this table, EPAD2, which is the
combination of our potential EPAD and the side-chain potential in OPUS, outperforms all the
others in terms of the average ranking of the best models in the absence of the native
structures or the average ranking of the native structures. EPAD2 also performs very well in
terms of the number of correctly-identified models (or native structures). EPAD is only second
to QMEAN6, RW and RWplus in ranking the best models in the absence of the native
structures. When the native structures are included, EPAD does not perform as well as when
the native structures are not included. All these results may indicate that the side-chain
component helps a lot in recognizing the native structures.
Is our probabilistic neural network (PNN) model over-trained?
Our PNN model has 60,000-70,000 parameters to be trained. A natural question to ask is if our
PNN model is biased towards some specific patterns in the training data? Can our PNN model
be generalized well to proteins of novel folds or sequences? According to our experimental
results on contact prediction (see “Window size and the number of neurons in the hidden
layers” section in 6.2 Methods and Experimental procedures), our PNN model is not overtrained. In this experiment, we used a training set built before CASP8 started, which are
unlikely to have similar folds and sequence profiles with our test set (i.e., the CASP8 and
CASP9 free modeling targets). Experimental results indicate that our PNN method compares
favorably to the best CASP8 and CASP9 server predictors, which implies that our PNN model
is not biased towards the training data.
108
Table 26 Performance of a variety of potentials on the selected CASP5-8 models.
Note: The results of EPAD and EPAD2 are shown in bold and the results of all the other
potentials are taken from (Zhang and Zhang, 2010).
a. The average rank of lowest energy decoy according to GDT score (over 143 decoy sets) in
the absence of native structures.
b. The number of sets when the best model was ranked as first, in the absence of native
structures.
c. The average rank of the lowest energy decoy in GDT when native structures are present.
d. The number of sets when the best model was ranked as first when native structures are
present.
e. Expected random values were generated by picking a wining model from the decoy sets
randomly. Average values over 1000 random trials are shown in the last row.
Scoring function
models only
native structures included
Average a Ranking b Average c
Ranking d
EPAD2
2.79
69
1.24
127
QMEAN6
2.87
85
1.71
113
RWplus
2.97
57
1.78
106
RW
3.08
51
1.71
110
EPAD
3.29
62
2.71
74
QMEANall_atom
3.59
74
1.71
119
QMEANSSE_agree
3.74
62
3.72
39
QMEANACC_agree 4.04
40
3.78
48
RF_CB_SRS_OD
4.16
61
2.08
110
RF_CB_OD
4.62
62
2.00
111
RF_HA_SRS
4.65
49
1.38
137
RF_CB_SRS
4.72
56
2.18
114
OPUS_CA
4.72
79
5.13
55
VSCOREcombined
4.79
53
2.2
117
QMEAN-pairwise
4.8
54
3.15
85
Rosetta
5.01
57
4.09
68
Dong-pair
5.01
58
6.32
4
RF_CB
5.06
52
2.46
106
VSCORE-pair
5.08
54
1.85
128
109
PROSAcombined
5.11
57
3.38
87
OPUS_PSP
5.39
54
2.99
118
RF_HA
5.44
62
2.78
112
DOPE
5.77
54
3.27
95
DFIRE
6.03
50
5.69
33
PROSA-pair
6.03
56
3.54
95
QMEAN-torsion
6.71
45
3.24
114
Shortle2006
6.85
35
1.79
129
Liang_geometric
6.88
44
2.48
114
QMEANsolvation
7.32
33
6.27
54
Shortle2005
7.73
42
3.39
109
Floudas-CM
7.75
38
7.05
42
Floudas-Ca
7.79
33
8.36
10
NAMD_1000
8.06
24
4.96
78
Melo-ANOLEA
9.62
19
5.19
86
PC2CA
9.75
19
5.06
85
Melo-NL
9.99
14
5.85
80
NAMD_1
11.91
5
10.98
24
Random e
9.72
13.9
10.1
8.3
110
Note that our PNN model for statistical potential uses exactly the same architecture (2 hidden
layers with 100 and 40 hidden neurons, respectively) as our PNN model for contact prediction.
Considering that much more training data (~73 millions of residue pairs) is used for the
derivation of our statistical potential than for contact prediction, it is less likely that our PNN
model for statistical potential is biased towards some specific patterns in the training set. The
result in Table 22 further confirms this. We use the 25% sequence identify or an E-value 0.01 as
the cutoff to exclude proteins in InHouse with similar sequences to the training set and
generate two subsets Set2 and Set4. Even if Set2 (Set4) contains some sequence profiles similar
to the training set, the similarity between the whole InHouse set and the training set is still
much larger than that between Set2 (Set4) and the training set, but the performance on the
whole InHouse set is even slightly worse than that on Set2 (Set4).
Folding Results with EPAD in the blind CASP 10 experiment
One of the most important functions of a good potential is its support in the folding process.
We tested EPAD using CNF-Folder (see Chapter 5) by replacing the DOPE term in the energy
function (see Chapter 4 for more details) with EPAD. The Energy guiding the folding
simulation is the linear combination of 4 potential terms: ½^^ = ±+ ½ûüý + ±. ½ýæû +
±/ ½þெ + ±õ ½Xû . The weight parameters (±7 ) are learned using grid search through a series of
folding experiments over a set of proteins containing all the targets from Table 18. Base on the
observation from the experiments, our CNF-Folder adopts the strategy to apply DOPE with
more weight on those low Neff targets or targets with short sequence length (≤65), while
keeping ±. to zero for other targets.
We have tested this new CNF-Folder with the new energy function on CASP9 Free modeling
targets. Table 27 compares the performance of CNF-Folder and Rosetta on CASP9 Free
modeling targets. We have downloaded the server prediction of Rosetta from CASP9 official
web
site
(http://predictioncenter.org/download_area/CASP9/server_predictions/)
and
calculated the GDT scores using TM-score (Zhang and Skolnick, 2005b). In the previous CASP
experiments, the GDT scores of our best clusters have been almost always within the top 7%
decoys. Hence, for simplicity, we do not generate clusters as we normally do in the CASP
blind experiments, but instead use the average GDT score of the top 10% decoys for each
target in the comparison. As can be seen from Table 27, with the EPAD term, the new CNFFolder performs favorably comparing to the most up to date Rosetta server.
Backed with this encouraging result, we have participated in the CASP 10 blind experiment
using the new CNF-Folder-EPAD program. The server we registered for CASP 10 FM targets
is named RaptorX-roll.
111
There are 16 template free modeling targets in CASP 10 experiment. We used a standalone
classification program to pick out the FM targets from where 10 targets have overlap with
CASP 10 official classifications. Table 28 compares the performance of our FM method
RaptorX-roll on the FM targets with RaptorX-ZY, one of the top threading servers in CASP 10.
It can be seen that RaptorX-roll yields better decoys on most of the FM targets than those from
RaptorX-ZY.
Figure 18 visualizes RaptorX-roll’s predictions on targets T0734 and T0740. The native
structures are plotted in blue and decoy structures are in pink color. The Percentage of
Residues comparison figures are also listed beside each target structure where RaptorX-roll’s
distributions are marked in pink color. The more a decoy’s distribution curve stands to the
right hand side, the higher quality a decoy structure is. T0734 had two domains and T0740 has
one domain. It can be seen that RaptorX-roll gave much better prediction than any of the other
predictors. One thing to notice is that we have given better predictions on most of the alpha
targets while mediocre on the beta proteins. It emphasizes the need of more focus on the
hydrogen bond potential in the future works.
112
Table 27 Folding comparison with Rosetta (GDT) on CASP9 FM targets
Note: The Rosetta data is obtained from CASP9 official web site. Rosetta’s Best denotes the
GDT score of the best decoy submitted by Rosetta. We list the average GDT of top 10% decoys
generated by CNF-Folder as well as the best GDT. For convenience, those low Neff targets are
listed below the other targets.
Target
Type
Neff
Len
T0531
1α3β
2.9
61
T0534_1
5α
5.6
T0534_2
5α2β
7
T0544_1
3α
T0544_2
Rosetta
Top 10%
Best
36.2
38.6
46.2
166
30.2
19.1
20.5
175
22.4
34.8
40.3
4.4
61
68.0
70.1
77.9
3α
4.3
78
46.6
55.0
60.5
T0547_3
1α8β
5.4
148
40.8
68.7
74.7
T0550_1
11β
8
159
14.7
20.0
24.1
T0553
4α
4.6
141
28.4
36.1
44.7
T0571_1
15β
7.4
168
37.0
14.7
18.8
T0571_2
8β
3.1
133
35.2
22.8
29.6
T0578
3α6β
3.3
163
23.3
23.5
29.0
T0604_1
2α6β
5.8
84
34.1
48.6
56.7
T0604_3
3α10β
4.8
205
13.2
21.7
26.9
T0616
2α3β
5.2
93
34.8
35.5
40.9
T0618
6α2β
4.5
170
35.4
31.2
38.8
T0624
7β
2.5
69
44.6
35.9
41.7
T0637
7α
7.3
135
39.6
30.1
35.6
T0561
6α
1.7
161
31.5
34.4
41.2
T0581
3α6β
1.8
112
67.9
33.8
42.9
T0608_1
5α
1.6
89
19.1
36.3
46.9
T0621
2α10β
1.2
169
18.3
15.9
19.7
T0639
4α
1.5
123
46.8
30.9
35.9
113
Best
CNF-Folder
Table 28 Comparison with threading method RaptorX-ZY on CASP10 FM targets
Note: RaptorX-ZY is one of the top threading servers in CASP 10. The scores listed are GDT
score of the best decoys submitted by each server. The CaspBest column lists the GDT scores of
the best decoys out of all the CASP10 servers.
Target
Type Neff length RaptorX-roll RaptorX-ZY CaspBest
T0653
40β
12.2
414
4.2
21.9
31.8
T0658D1 9β
3.6
185
16.1
13.1
24.7
T0666
4.3
195
26.3
23.2
33.8
T0684D2 6α4β
1.2
167
17.4
16.5
21.3
T0693D1 4α4β
4.6
100
25.3
24.0
35.0
T0734
5.5
214
20.3
13.9
21.0
4.6
117
39.8
31.0
41.5
T0737D1 5α2β
1
116
27.4
27.8
36.8
T0740
7α2β
1
165
38.9
25.3
38.9
T0741
14β
1
181
13.6
15.2
17.2
6α
8α2β
T0735D2 4α
114
(A)
(B)
Figure 18 . RaptorX-roll’s Result on targets T0734 (A) and T0740 (B).
The native structures are plotted in blue and decoy structures are in pink color. In the
Percentage of Residues comparison figures, RaptorX-roll’s distributions are marked in pink
color. The more a decoy’s distribution curve stands to the right hand side, the higher quality a
decoy structure is.
115
6.4 Conclusion
This chapter has presented a novel protein-specific and position-specific knowledge-based
statistical potential EPAD for protein structure and functional study. EPAD is unique in that it
may have different energy profiles for two atoms of given types, depending on the protein
under consideration and the sequence profile contexts of the residues containing them, while
other potentials have the same energy profile for a given atom pair across all proteins. We
achieved this by parameterizing EPAD using evolutionary information and radius of gyration
of the protein under consideration in addition to atom types, which enables us to obtain a
much more accurate statistical potential.
We also made a novel technical contribution to estimating the observed atomic interacting
probability by introducing a probabilistic neural network to calculate the inter-atom distance
probability distribution from sequence profiles and the radius of gyration. This is very
different from the simple counting method widely used to derive the position-independent
statistical potentials such as DOPE and DFIRE. The simple counting method does not work for
our potential simply because there is not enough number of solved protein structures in PDB
for reliable counting of sequence profile contexts.
Experimental results indicated that EPAD significantly outperforms several popular higherresolution full-atom potentials in several decoy discrimination tests even if only backbone
atoms are considered in EPAD. If we combine EPAD with the side-chain component in OPUSPSP, we can achieve much better decoy discrimination performance especially in the presence
of native structures. As opposed to the RW potential and many others, EPAD is not trained by
any decoys, so in principal it is not restricted to any decoy generation method. Currently
EPAD uses only 1Å resolution for the Ï − Ï distance discretization. We will further improve
EPAD by using a 0.5Å resolution, but this will take a very long time to train a neural network
model for accurate estimation of the extremely unbalanced distance probability distribution.
116
Chapter 7. Conclusion and Future works
7.1 Conclusions
The thesis is aimed to solve the template-free protein folding problem by tackling two
important components: efficient sampling in vast conformation space, and design of
knowledge-based potentials with high accuracy.
The CRF-Samplers are proposed to sample structures from the continuous local dihedral
angles space by modeling the lower and higher order conditional dependency between
neighboring dihedral angles given the primary sequence information. A framework combining
the Conditional Random Fields and the energy function is introduced to guide the local
conformation sampling using long range constraints with the energy function. In order to
model the complex nonlinear relationship between the sequence profile and the local dihedral
angle distribution, we further built the CNF-Folder by applying a novel machine learning
model Conditional Neural Fields which utilizes the structural graphical model with the neural
network. CRF-Samplers and CNF-Folder perform very well in CASP8 and CASP9.
On the other track, we have designed a novel pairwise distance statistical potential (EPAD) to
capture the dependency of the energy profile on the positions of the interacting amino acids as
well as the types of those amino acids, opposing the common assumption that this energy
profile depends only on the types of amino acids. EPAD has also been successfully applied in
the CASP 10 Free Modeling experiment with CNF-Folder, especially outstanding on some
uncommon structured targets.
7.2 Future Works
Conformation Sampling
Although the primary purpose of the free modeling folders including the CRF-Samplers and
CNF-Folder is for template-free protein folding, they can be applied to other important
applications. One direct example is to refine template-based models. Some preliminary results
on CASP blind experiments can be found in (Xu et al., 2009) and (Ma et al., 2013). The method
can also be applied to protein loop modeling, model refinement, and even RNA tertiary
structure prediction (Wang and Xu, 2011).
The free modeling folders have some immediate limitations to relax.
117
By extracting distance constraints from template-based models, the conformational space of a
target is dramatically reduced and thus we can afford to search this reduced space, which may
search conformational space more thoroughly and lead to better prediction accuracy.
They can also be further extended to model the long-range hydrogen bonding effect to
significantly enhance the performance on beta proteins.
Statistical Potentials
We may extend our statistical potential as follows.
Currently EPAD considers only backbone atoms and is also orientation-independent. We can
extend it to side-chain atoms and also make it orientation-dependent.
Second, in estimating the distance probability distribution of two positions, we use only
sequence profile contexts relevant to only these two positions. We shall also use information in
the sequence segment connecting the two residues, which contains important information in
determining the relative orientation of the two residues.
Thirdly, we may also estimate the distance probability distribution more accurately by adding
some physical constraints. For example, given any three atoms in a protein, their pairwise
distances must satisfy the triangle inequality.
Further, for any three residues which are close to one another along the primary sequence,
their Cä distances are also subject to the restriction of local atomic interaction. If we assume
that there is a contact between two residues if their Cä or Cஒ atoms are within 8Å, then the
number of contacts for any given residue is limited by a constant (~13) due to geometric
restraint. By enforcing these constraints, we shall be able to estimate the inter-atom distance
probability distribution much more accurately and thus, design a much better statistical
potential.
Other two important potentials to probe are hydrogen bonding potential and environmental
potential.
We have some preliminary results on the latter category that a contact number potential is
designed and experimental results support the hypothesis that the wrapping of each residue
by surrounding amino acids is guided by maximizing the local static electric field. We will
keep working on it in the near future.
118
References
Aarts, E., and Korst, J. (1991). Simulated Annealing and Boltzmann Machines: A Stochastic Approach to
Combinatorial Optimization and Neural Computing (Wiley).
Anfinsen, C.B. (1973). Principles that govern the folding of protein chains. Science 181, 223-230.
Bauer, A., and Beyer, A. (1994). An Improved Pair Potential to Recognize Native Protein Folds. Proteins
18, 254-261.
Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T., Weissig, H., Shindyalov, I.N., and Bourne,
P.E. (2000). The protein data bank. Nucleic Acids Res 28, 235-242.
Boomsma, W., Mardia, K.V., Taylor, C.C., Ferkinghoff-Borg, J., Krogh, A., and Hamelryck, T. (2008). A
generative, probabilistic model of local protein structure. Proceedings of the National Academy of
Sciences 105, 8932-8937.
Bowie, J.U., and Eisenberg, D. (1994). An Evolutionary Approach to Folding Small Alpha-Helical
Proteins That Uses Sequence Information and an Empirical Guiding Fitness Function. P Natl Acad Sci
USA 91, 4436-4440.
Bradley, P., Malmstrom, L., Qian, B., Schonbrun, J., Chivian, D., Kim, D.E., Meiler, K., Misura, K.M.S.,
and Baker, D. (2005). Free modeling with Rosetta in CASP6. Proteins-Structure Function and
Bioinformatics 61, 128-134.
Brooks, B.R., Brooks, C.L., Mackerell, A.D., Nilsson, L., Petrella, R.J., Roux, B., Won, Y., Archontis, G.,
Bartels, C., Boresch, S., et al. (2009). CHARMM: The Biomolecular Simulation Program. J Comput Chem
30, 1545-1614.
Bryngelson, J.D., Onuchic, J.N., Socci, N.D., and Wolynes, P.G. (1995). Funnels, Pathways, and the
Energy Landscape of Protein-Folding - a Synthesis. Proteins 21, 167-195.
Bystroff, C., Thorsson, V., and Baker, D. (2000). HMMSTR: a hidden Markov model for local sequencestructure correlations in proteins. J Mol Biol 301, 173-190.
Casari, G., and Sippl, M.J. (1992). Structure-Derived Hydrophobic Potential - Hydrophobic Potential
Derived from X-Ray Structures of Globular-Proteins Is Able to Identify Native Folds. J Mol Biol 224,
725-732.
Case, D.A., Cheatham, T.E., Darden, T., Gohlke, H., Luo, R., Merz, K.M., Onufriev, A., Simmerling, C.,
Wang, B., and Woods, R.J. (2005). The Amber biomolecular simulation programs. J Comput Chem 26,
1668-1688.
Claessens, M., Vancutsem, E., Lasters, I., and Wodak, S. (1989). Modeling the Polypeptide Backbone
with Spare Parts from Known Protein Structures. Protein Eng 2, 335-345.
Colubri, A., Jha, A.K., Shen, M.Y., Sali, A., Berry, R.S., Sosnick, T.R., and Freed, K.F. (2006). Minimalist
representations nearest neighbor effects in and the importance of protein folding simulations. J Mol
Biol 363, 535-557.
Deltheli, R. (1919). Sur la théorie des probabilités géométriques. Ann Fac Sci Univ Toulouse 11, 1-65.
119
Dietterich, T., Ashenfelter, A., and Bulatov, Y. (2004). Training conditional random fields via gradient
tree boosting. Paper presented at: In Proceedings of the 21th International Conference on Machine
Learning (ICML).
Dill, K.A. (1985). Theory for the Folding and Stability of Globular-Proteins. Biochemistry-Us 24, 15011509.
Dill, K.A. (1997). Additivity principles in biochemistry. J Biol Chem 272, 701-704.
Dobson, C.M., Sali, A., and Karplus, M. (1998). Protein folding: A perspective from theory and
experiment. Angew Chem Int Edit 37, 868-893.
Dor, O., and Zhou, Y.Q. (2007). Achieving 80% ten-fold cross-validated accuracy for secondary
structure prediction by large-scale training. Proteins-Structure Function and Bioinformatics 66, 838-845.
Earl, D.J., and Deem, M.W. (2005). Parallel tempering: Theory, applications, and new perspectives.
Physical Chemistry Chemical Physics 7, 3910-3916.
Eddy, S.R. (1998). Profile hidden Markov models. Bioinformatics 14, 755-763.
Feldman, H.J., and Hogue, C.W. (2002). Probabilistic sampling of protein conformations: new hope for
brute force? Proteins-Structure Function and Bioinformatics 46, 8-23.
Fernandez, A., Sosnick, T.R., and Colubri, A. (2002). Dynamics of hydrogen bond desolvation in
protein folding. J Mol Biol 321, 659-675.
Finn, R.D., Tate, J., Mistry, J., Coggill, P.C., Sammut, S.J., Hotz, H.R., Ceric, G., Forslund, K., Eddy, S.R.,
Sonnhammer, E.L.L., et al. (2008). The Pfam protein families database. Nucleic Acids Res 36, D281-D288.
Fitzgerald, J.E., Jha, A.K., Colubri, A., Sosnick, T.R., and Freed, K.F. (2007). Reduced C-beta statistical
potentials can outperform all-atom potentials in decoy identification. Protein Science 16, 2123-2139.
Garcia-Pelayo, R. (2005). Distribution of distance in the spheroid. J Phys a-Math Gen 38, 3475-3482.
Gatchell, D.W., Dennis, S., and Vajda, S. (2000). Discrimination of near-native protein structures from
misfolded models by empirical free energy functions. Proteins 41, 518-534.
Gilis, D., and Rooman, M. (1996). Stability changes upon mutation of solvent-accessible residues in
proteins evaluated by database-derived potentials. J Mol Biol 257, 1112-1126.
Gilis, D., and Rooman, M. (1997). Predicting protein stability changes upon mutation using databasederived potentials: Solvent accessibility determines the importance of local versus non-local
interactions along the sequence. J Mol Biol 272, 276-290.
Gong, H.P., Fleming, P.J., and Rose, G.D. (2005). Building native protein conformation from highly
approximate backbone torsion angles. P Natl Acad Sci USA 102, 16227-16232.
Gront, D., Kmiecik, S., and Kolinski, A. (2007). Backbone building from quadrilaterals: A fast and
accurate algorithm for protein backbone reconstruction from alpha carbon coordinates. J Comput
Chem 28, 1593-1597.
Hamelryck, T., Kent, J.T.T., and Krogh, A. (2006). Sampling Realistic Protein Conformations Using
Local Structural Bias. PLoS Comput Biology 2.
120
Hendlich, M., Lackner, P., Weitckus, S., Floeckner, H., Froschauer, R., Gottsbacher, K., Casari, G., and
Sippl, M.J. (1990). Identification of Native Protein Folds Amongst a Large Number of Incorrect Models the Calculation of Low-Energy Conformations from Potentials of Mean Force. J Mol Biol 216, 167-180.
Hildebrand, A., Remmert, M., Biegert, A., and Soding, J. (2009). Fast and accurate automatic structure
prediction with HHpred. Proteins-Structure Function and Bioinformatics 77 Suppl 9, 128-132.
Holm, L., and Sander, C. (1991). Database Algorithm for Generating Protein Backbone and Side-Chain
Coordinates from a C-Alpha Trace Application to Model-Building and Detection of Coordinate Errors.
J Mol Biol 218, 183-194.
Hu, Y., Dong, X., Wu, A., Cao, Y., Tian, L., and Jiang, T. (2011). Incorporation of Local Structural
Preference Potential Improves Fold Recognition. Plos One 6.
Hua, S.J., and Sun, Z.R. (2001). A novel method of protein secondary structure prediction with high
segment overlap measure: Support vector machine approach. J Mol Biol 308, 397-407.
Jha, A.K., Colubri, A., Zaman, M.H., Koide, S., Sosnick, T.R., and Freed, K.F. (2005). Helix, sheet, and
polyproline II frequencies and strong nearest neighbor effects in a restricted coil library. BiochemistryUs 44, 9691-9702.
Jones, D.T. (1999). Protein secondary structure prediction based on position-specific scoring matrices. J
Mol Biol 292, 195-202.
Jones, D.T., and Thornton, J.M. (1996). Potential energy functions for threading. Curr Opin Struc Biol 6,
210-216.
Jones, T.A., and Thirup, S. (1986). Using Known Substructures in Protein Model-Building and
Crystallography. Embo J 5, 819-822.
Joo, K., Lee, J., Seo, J.H., Lee, K., Kim, B.G., and Lee, J. (2009). All-atom chain-building by optimizing
MODELLER energy function using conformational space annealing. Proteins-Structure Function and
Bioinformatics 75, 1010-1023.
Kachalo, S., Lu, H.-M., and Liang, J. (2006). Protein folding dynamics via quantification of kinematic
energy landscape. Physical Review Letters 96, 058106.
Kanou, K., Iwadate, M., Hirata, T., Terashi, G., Umeyama, H., and Takeda-Shitaka, M. (2009). FAMSD:
A powerful protein modeling platform that combines alignment methods, homology modeling, 3D
structure quality estimation and molecular dynamics. Chem Pharm Bull (Tokyo) 57, 1335-1342.
Karypis, G. (2006). YASSPP: Better kernels and coding schemes lead to improvements in protein
secondary structure prediction. Proteins-Structure Function and Bioinformatics 64, 575-586.
Kelley, L.A., and Sternberg, M.J.E. (2009). Protein structure prediction on the Web: a case study using
the Phyre server. Nat Protoc 4, 363-371.
Kent, J.T. (1982). The Fisher-Bingham Distribution on the Sphere. J Roy Stat Soc B Met 44, 71-80.
Kihara, D., Lu, H., Kolinski, A., and Skolnick, J. (2001). TOUCHSTONE: An ab initio protein structure
prediction method that uses threading-based tertiary restraints. P Natl Acad Sci USA 98, 10125-10130.
121
Kim, H., and Park, H. (2003). Protein secondary structure prediction based on an improved support
vector machines approach. Protein Eng 16, 553-560.
Kolodny, R., Koehl, P., Guibas, L., and Levitt, M. (2002). Small libraries of protein fragments model
native protein structures accurately. J Mol Biol 323, 297-307.
Kortemme, T., and Baker, D. (2002). A simple physical model for binding energy hot spots in proteinprotein complexes. P Natl Acad Sci USA 99, 14116-14121.
Krogh, A., Brown, M., Mian, I.S., Sjolander, K., and Haussler, D. (1994). Hidden Markov-Models in
Computational Biology - Applications to Protein Modeling. J Mol Biol 235, 1501-1531.
Labesse, G., Colloc'h, N., Pothier, J., and Mornon, J.P. (1997). P-SEA: a new efficient assignment of
secondary structure from C alpha trace of proteins. Comput Appl Biosci 13, 291-295.
Lafferty, J., McCallum, A., and Pereira, F.C.N. (2001). Conditional random fields: Probabilistic models
for segmenting and labeling sequence data.
Larsson, P., Wallner, B., Lindahl, E., and Elofsson, A. (2008). Using multiple templates to improve
quality of homology models in automated homology modeling. Protein Science 17, 990-1002.
Laurie, A.T.R., and Jackson, R.M. (2005). Q-SiteFinder: an energy-based method for the prediction of
protein-ligand binding sites. Bioinformatics 21, 1908-1916.
Lee, J., Lee, D., Park, H., Coutsias, E.A., and Seok, C. (2010). Protein loop modeling by using fragment
assembly and analytical loop closure. Proteins-Structure Function and Bioinformatics 78, 3428-3436.
Levitt, M. (1976). A simplified representation of protein conformations for rapid simulation of protein
folding. J Mol Biol 104, 59-107.
Levitt, M. (1992). Accurate Modeling of Protein Conformation by Automatic Segment Matching. J Mol
Biol 226, 507-533.
Li, S.C., Bu, D., Xu, J., and Li, M. (2008). Fragment-HMM: a new approach to protein structure
prediction. Protein Science 17, 1925-1934.
Li, X., and Liang, J. (2007). Knowledge-based energy functions for computational studies of proteins. In
Computational methods for protein structure prediction and modeling (Springer), pp. 71-123.
Liu, D., and Nocedal, J. (1989). On the Limited Memory Bfgs Method for Large-Scale Optimization.
Math Program 45, 503-528.
Lu, H., and Skolnick, J. (2001). A distance-dependent atomic knowledge-based potential for improved
protein structure selection. Proteins 44, 223-232.
Lu, M., Dousis, A., and Ma, J. (2008). OPUS-PSP: An orientation-dependent statistical all-atom potential
derived from side-chain packing. J Mol Biol 376, 288-301.
Ma, J., Wang, S., Zhao, F., and Xu, J. (2013). Protein threading using context-specific alignment potential.
Bioinformatics 29, i257-i265.
Maiorov, V.N., and Crippen, G.M. (1992). Contact Potential That Recognizes the Correct Folding of
Globular-Proteins. J Mol Biol 227, 876-888.
122
Mardia, K.V., Taylor, C.C., and Subramaniam, G.K. (2007). Protein bioinformatics and mixtures of
bivariate von Mises distributions for angular data. Biometrics 63, 505-512.
Maupetit, J., Gautier, R., and Tuffery, P. (2006). SABBAC: online structural alphabet-based protein
backbone reconstruction from alpha-carbon trace. Nucleic Acids Res 34, W147-W151.
McGuffin, L.J., Bryson, K., and Jones, D.T. (2000). The PSIPRED protein structure prediction server.
Bioinformatics 16, 404-405.
Meiler, J., Peti, W., and Griesinger, C. (2000). DipoCoup: A versatile program for 3D-structure
homology comparison based on residual dipolar couplings and pseudocontact shifts. Journal of
biomolecular NMR 17, 283-294.
Misura, K.M.S., Chivian, D., Rohl, C.A., Kim, D.E., and Baker, D. (2006). Physically realistic homology
models built with ROSETTA can be more accurate than their templates. P Natl Acad Sci USA 103, 53615366.
Miyazawa, S., and Jernigan, R.L. (1985). Estimation of Effective Interresidue Contact Energies from
Protein Crystal-Structures - Quasi-Chemical Approximation. Macromolecules 18, 534-552.
Morozov, A.V., Kortemme, T., Tsemekhman, K., and Baker, D. (2004). Close agreement between the
orientation dependence of hydrogen bonds observed in protein structures and quantum mechanical
calculations. Proc Natl Acad Sci U S A 101, 6946-6951.
Moult, J. (2005). A decade of CASP: progress, bottlenecks and prognosis in protein structure prediction.
Curr Opin Struct Biol 15, 285-289.
Moult, J., Fidelis, K., Kryshtafovych, A., Rost, B., Hubbard, T., and Tramontano, A. (2007). Critical
assessment of methods of protein structure prediction-Round VII. Proteins-Structure Function and
Bioinformatics 69 Suppl 8, 3-9.
Moult, J., Fidelis, K., Rost, B., Hubbard, T., and Tramontano, A. (2005). Critical assessment of methods
of protein structure prediction (CASP)-round 6. Proteins: Structure, Function and Bioinformatics 61
Suppl 7, 3-7.
Moult, J., Fidelis, K., Zemla, A., and Hubbard, T. (2003). Critical assessment of methods of protein
structure prediction (CASP)-round V. Proteins-Structure Function and Bioinformatics 53 Suppl 6, 334339.
Murzin, A.G., Brenner, S.E., Hubbard, T., and Chothia, C. (1995). SCOP: a structural classification of
proteins database for the investigation of sequences and structures. J Mol Biol 247, 536-540.
N. Eswar, M.A.M.-R., B. Webb, M. S. Madhusudhan, D. Eramian, M. Shen, U. Pieper, A. Sali. (2006).
Comparative Protein Structure Modeling With MODELLER. Current Protocols in Bioinformatics
Supplement 15, 5.6.1-5.6.30.
Neal, S., Berjanskii, M., Zhang, H., and Wishart, D.S. (2006). Accurate prediction of protein torsion
angles using chemical shifts and sequence homology. Magnetic Resonance in Chemistry 44, S158-S167.
Notredame, C., Higgins, D.G., and Heringa, J. (2000). T-Coffee: A novel method for fast and accurate
multiple sequence alignment. J Mol Biol 302, 205-217.
123
Panchenko, A.R., Marchler-Bauer, A., and Bryant, S.H. (2000). Combination of threading potentials and
sequence profiles improves fold recognition. J Mol Biol 296, 1319-1331.
Panjkovich, A., Melo, F., and Marti-Renom, M. (2008). Evolutionary potentials: structure specific
knowledge-based potentials exploiting the evolutionary record of sequence homologs. Genome Biol 9.
Park, B.H., Huang, E.S., and Levitt, M. (1997). Factors affecting the ability of energy functions to
discriminate correct from incorrect folds. J Mol Biol 266, 831-846.
Pei, J., Kim, B., and Grishin, N. (2008). PROMALS3D: a tool for multiple protein sequence and structure
alignments. Nucleic Acids Res 36, 2295-2300.
Peng, J., Bo, l., and Xu, J. (2009). Conditional Neural Fields. Paper presented at: Advances in Neural
Information Processing Systems (NIPS) (Vancouver, Canada).
Peng, J., and Xu, J. (2009). Boosting Protein Threading Accuracy. Lect Notes Comput Sci 5541, 31.
Peng, J., and Xu, J. (2010). Low-homology protein threading. Bioinformatics 26, i294-i300.
Peng, J., and Xu, J. (2011). RaptorX: exploiting structure information for protein alignment by statistical
inference. Proteins-Structure Function and Bioinformatics 79 Suppl 10, 161-171.
Phan, X.-H., Nguyen, L.-M., and Nguyen, C.-T. (2005). FlexCRFs: Flexible Conditional Random Fields.
Qian, B., Raman, S., Das, R., Bradley, P., McCoy, A., Read, R., and Baker, D. (2007). High-resolution
structure prediction and the crystallographic phase problem. Nature 450, 259-U257.
Raman, S., Vernon, R., Thompson, J., Tyka, M., Sadreyev, R., Pei, J.M., Kim, D., Kellogg, E., DiMaio, F.,
Lange, O., et al. (2009). Structure prediction for CASP8 with all-atom refinement using Rosetta.
Proteins-Structure Function and Bioinformatics 77, 89-99.
Reva, B.A., Finkelstein, A.V., Sanner, M.F., and Olson, A.J. (1997). Residue-residue mean-force
potentials for protein structure recognition. Protein Eng 10, 865-876.
Rykunov, D., and Fiser, A. (2010). New statistical potential for quality assessment of protein models
and a survey of energy functions. BMC Bioinformatics 11, 128.
Samudrala, R., and Levitt, M. (2000). Decoys 'R' Us: A database of incorrect conformations to improve
protein structure prediction. Protein Science 9, 1399-1401.
Samudrala, R., and Moult, J. (1998). An all-atom distance-dependent conditional probability
discriminatory function for protein structure prediction. J Mol Biol 275, 895-916.
Schuler, L.D., Daura, X., and Van Gunsteren, W.F. (2001). An improved GROMOS96 force field for
aliphatic hydrocarbons in the condensed phase. J Comput Chem 22, 1205-1218.
Sha, F., and Pereira, F. (2003). Shallow parsing with conditional random fields. Paper presented at:
Proceedings of the 2003 Conference of the North American Chapter of the Association for
Computational Linguistics on Human Language Technology-Volume 1 (Association for Computational
Linguistics).
Shakhnovich, E. (2006). Protein folding thermodynamics and dynamics: Where physics, chemistry, and
biology meet. Chem Rev 106, 1559-1588.
124
Shao, M.F., Wang, S., Wang, C., Yuan, X.Y., Li, S.C., Zheng, W.M., and Bu, D.B. (2011). Incorporating
Ab Initio energy into threading approaches for protein structure prediction. BMC Bioinformatics 12.
Shen, M., and Sali, A. (2006). Statistical potential for assessment and prediction of protein structures.
Protein Science 15, 2507-2524.
Shi, S., Pei, J., Sadreyev, R., Kinch, L., Majumdar, I., Tong, J., Cheng, H., Kim, B.-H., and Grishin, N.
(2009). Analysis of casp8 targets, predictions and assessment methods, . Database 2009:bap003.
Simon, I., Glasser, L., and Scheraga, H.A. (1991). Calculation of Protein Conformation as an Assembly
of Stable Overlapping Segments: Application to Bovine Pancreatic Trypsin Inhibitor. Proceedings of
National Academy Sciences, USA 88, 3661-3665.
Simons, K.T., Kooperberg, C., Huang, E., and Baker, D. (1997). Assembly of protein tertiary structures
from fragments with similar local sequences using simulated annealing and Bayesian scoring functions.
J Mol Biol 268, 209-225.
Simons, K.T., Ruczinski, I., Kooperberg, C., Fox, B.A., Bystroff, C., and Baker, D. (1999). Improved
recognition of native-like protein structures using a combination of sequence-dependent and sequenceindependent features of proteins. Proteins 34, 82-95.
Singh, H., Hnizdo, V., and Demchuk, E. (2002). Probabilistic model for two dependent circular
variables. Biometrika 89, 719-723.
Sippl, M. (1993). Recognition of errors in three-dimensional structures of proteins. Proteins: Structure,
Function, and Bioinformatics 17, 355-362.
Sippl, M.J. (1990). Calculation of Conformational Ensembles from Potentials of Mean Force - an
Approach to the Knowledge-Based Prediction of Local Structures in Globular-Proteins. J Mol Biol 213,
859-883.
Sippl, M.J., and Weitckus, S. (1992). Detection of native-like models for amino acid sequences of
unknown three-dimensional structure in a data base of known protein conformations. ProteinsStructure Function and Bioinformatics 13, 258-271.
Skolnick, J. (2006). In quest of an empirical potential for protein structure prediction. Curr Opin Struc
Biol 16, 166-171.
Skolnick, J., Kihara, D., and Zhang, Y. (2004). Development and large scale benchmark testing of the
PROSPECTOR_3 threading algorithm. Proteins 56, 502-518.
Skolnick, J., Kolinski, A., and Ortiz, A. (2000). Derivation of protein-specific pair potentials based on
weak sequence fragment similarity. Proteins 38, 3-16.
Soding, J. (2005). Protein homology detection by HMM-HMM comparison. Bioinformatics 21, 951-960.
Specht, D.F. (1990). Probabilistic Neural Networks. Neural Networks 3, 109-118.
Swendsen, R.H., and Wang, J.S. (1986). Replica Monte-Carlo Simulation of Spin-Glasses. Physical
Review Letters 57, 2607-2609.
Tanaka, S., and Scheraga, H.A. (1976). Medium- and long-range interaction parameters between amino
acids for predicting three-dimensional structures of proteins. Macromolecules 9, 945-950.
125
Tegge, A., Wang, Z., Eickholt, J., and Cheng, J. (2009). NNcon: improved protein contact map
prediction using 2D-recursive neural networks. Nucleic Acids Res 37, W515-W518.
Tooze, J., and Branden, C. (1999). Introduction to protein structure (USA New York: Garland
Publishing, Inc).
Tuffery, P., and Derreumaux, P. (2005). Dependency between consecutive local conformations helps
assemble protein structures from secondary structures using Go potential and greedy algorithm.
Proteins-Structure Function and Bioinformatics 61, 732-740.
Unger, R., Harel, D., Wherland, S., and Sussman, J.L. (1989). A 3D building blocks approach to
analyzing and predicting structure of proteins. Proteins: Structure, Function and Genetics 5, 355-373.
Vapnik, V. (1998). Statistical learning theory (Wiley New York).
Vendruscolo, M., Najmanovich, R., and Domany, E. (2000). Can a pairwise contact potential stabilize
native protein folds against decoys obtained by threading? Proteins 38, 134-148.
Wang, G., and Dunbrack, R. (2003). PISCES: a protein sequence culling server. Bioinformatics 19, 15891591.
Wang, Z., and Xu, J. (2011). A conditional random fields method for RNA sequence–structure
relationship modeling and conformation sampling. Bioinformatics 27, i102-i110.
Wang, Z., Zhao, F., Peng, J., and Xu, J. (2011). Protein 8-class secondary structure prediction using
conditional neural fields. Proteomics 11, 3786-3792.
Wendoloski, J.J., and Salemme, F.R. (1992). Probit - a Statistical Approach to Modeling Proteins from
Partial Coordinate Data Using Substructure Libraries. Journal of Molecular Graphics 10, 124-126.
Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics bulletin 1, 80-83.
Wu, S., Skolnick, J., and Zhang, Y. (2007a). Ab initio modeling of small proteins by iterative TASSER
simulations. Bmc Biol 5.
Wu, S., and Zhang, Y. (2008a). A comprehensive assessment of sequence-based and template-based
methods for protein contact prediction. Bioinformatics 24, 924-931.
Wu, S., and Zhang, Y. (2008b). MUSTER: Improving protein sequence profile-profile alignments by
using multiple sources of structure information. Proteins-Structure Function and Bioinformatics 72,
547-556.
Wu, Y., Lu, M., Chen, M., Li, J., and Ma, J. (2007b). OPUS-Ca: A knowledge-based potential function
requiring only C alpha positions. Protein Science 16, 1449-1463.
Xia, Y., Huang, E.S., Levitt, M., and Samudrala, R. (2000). Ab initio construction of protein tertiary
structures using a hierarchical approach. J Mol Biol 300, 171-185.
Xu, D., Zhang, J., Roy, A., and Zhang, Y. (2011). Automated protein structure modeling in CASP9 by ITASSER pipeline combined with QUARK-based ab initio folding and FG-MD-based structure
refinement. Proteins-Structure Function and Bioinformatics 79 Suppl 10, 147-160.
Xu, J. (2005). Fold recognition by predicted alignment accuracy. Ieee Acm T Comput Bi 2, 157-165.
126
Xu, J., Peng, J., and Zhao, F. (2009). Template-based and free modeling by RAPTOR++ in CASP8.
Proteins: Structure, Function, and Bioinformatics 77, 133-137.
Xue, B., Faraggi, E., and Zhou, Y. (2009). Predicting residue-residue contact maps by a two-layer,
integrated neural-network method. Proteins-Structure Function and Bioinformatics 76, 176-183.
Yang, Y., Faraggi, E., Zhao, H., and Zhou, Y. (2011). Improving protein fold recognition and templatebased modeling by employing probabilistic-based matching between predicted one-dimensional
structural properties of query and corresponding native properties of templates. Bioinformatics 27,
2076-2082.
Zemla, A. (2003). LGA: a method for finding 3D similarities in protein structures. Nucleic Acids Res 31,
3370-3374.
Zemla, A., Venclovas, C., Moult, J., and Fidelis, K. (1999). Processing and analysis of CASP3 protein
structure predictions. Proteins, 22-29.
Zemla, A., Venclovas, C., Moult, J., and Fidelis, K. (2001). Processing and evaluation of predictions in
CASP4. Proteins-Structure Function and Bioinformatics, 13-21.
Zhang, C., Vasmatzis, G., Cornette, J.L., and DeLisi, C. (1997). Determination of atomic desolvation
energies from the structures of crystallized proteins. J Mol Biol 267, 707-726.
Zhang, J., and Zhang, Y. (2010). A Novel Side-Chain Orientation Dependent Potential Derived from
Random-Walk Reference State for Protein Fold Selection and Structure Prediction. Plos One 5.
Zhang, Y., Kolinski, A., and Skolnick, J. (2003). TOUCHSTONE II: a new approach to ab initio protein
structure prediction. Biophys J 85, 1145-1164.
Zhang, Y., and Skolnick, J. (2005a). The protein structure prediction problem could be solved using the
current PDB library. P Natl Acad Sci USA 102, 1029-1034.
Zhang, Y., and Skolnick, J. (2005b). TM-align: a protein structure alignment algorithm based on the TMscore. Nucleic Acids Res 33, 2302-2309.
Zhang, Y., and Skolnick, J. (2007). Scoring function for automated assessment of protein structure
template quality (vol 57, pg 702, 2004). Proteins-Structure Function and Bioinformatics 68, 1020-1020.
Zhao, F., Li, S., Sterner, B.W., and Xu, J. (2008). Discriminative learning for protein conformation
sampling. Proteins: Structure, Function, and Bioinformatics 73, 228-240.
Zhao, F., Peng, J., DeBartolo, J., Freed, K., Sosnick, T., and Xu, J. (2009). A Probabilistic Graphical Model
for Ab Initio Folding. In Research in Computational Molecular Biology (Tucson, Arizona, Springer
Berlin Heidelberg), pp. 59-73.
Zhao, F., Peng, J., and Xu, J. (2010). Fragment-free approach to protein folding using conditional neural
fields. Bioinformatics 26, i310-i317.
Zhou, H., Pandit, S.B., and Skolnick, J. (2009). Performance of the Pro-sp3-TASSER server in CASP8.
Proteins-Structure Function and Bioinformatics 77 Suppl 9, 123-127.
Zhou, H., and Skolnick, J. (2009). Protein structure prediction by pro-Sp3-TASSER. Biophys J 96, 21192127.
127
Zhou, H., and Zhou, Y. (2002). Distance-scaled, finite ideal-gas reference state improves structurederived potentials of mean force for structure selection and stability prediction. Protein Science 11,
2714-2726.
128
| 5cs.CE
|
arXiv:1607.01228v1 [math.AC] 5 Jul 2016
A CANONICAL POLYTOPAL RESOLUTION FOR
TRANSVERSAL MONOMIAL IDEALS
RAHIM ZAARE-NAHANDI
UNIVERSITY OF TEHRAN
Abstract. Let S = k[x11 , · · · , x1b1 , · · · , xn1 , · · · , xnbn ] be a polynomial ring
in m = b1 + · · · + bn variables over a field k. For all j, 1 ≤ j ≤ n, let Pj be
the prime ideal generated by variables {xj1 , · · · , xjbj } and let
X
Pj1 . . . Pjt
In,t =
1≤j1 <···<jt ≤n
be the transversal monomial ideal of degree t on P1 , · · · , Pn . We explicitly
construct a canonical polytopal Zt -graded minimal free resolution for the ideal
In,t by means of suitable gluing of polytopes.
1. Introduction
The idea to describe a resolution of a monomial ideal by means of combinatorialgeometric chain complexes was initiated by Bayer, Peeva and Sturmfels [2], and
was extended by Bayer and Sturmfels [3]. Novik, Postnikov and Sturmfels gave
a polytopal complex that supports the minimal free resolution of matroidal ideals
[17]. Further extension was made by Jöllembeck and Welker who used discrete
Morse theory to construct desired CW complexes [13]. Sinefakoupols showed that
the Eliahou-Kervaire resolution [6] of Borel-fixed monomial ideals which are Borelfixed generated in one degree are cellular supported on a union of convex polytopes
[20, Theorem 20]. Mermin and Clark, separately, constructed regular cell complexes which support the Eliahou-Kervaire resolution of a stable monomial ideal
[14], [7]. On the other hand, Valesco showed that there exist minimal free monomial resolutions which can not be supported on any CW-complex [21]. Still from a
different point of view, Fløystad investigated monomial labelings on cell complexes
which give minimal free resolution of the ideal generated by these monomials in the
case the quotient ring is Cohen-Macaulay [10]. Dochtermann and Engström gave
a cellular resolutions for the ideals of cointerval hypergraphs supported on polyhedral complexes and, they extended their construction to more general hypergraphs
[5, Theorems 4.4 and 6.1]. Further on, Engström and Norén constructed cellular
resolutions for powers of ideals of a bipartite graph [8, Proposition 4.4 and Theorem 7.2]. Okazaki and Yanagawa showed that the Eliahou-Kervaire resolution of
a Cohen-Macaulay stable monomial ideal is supported by a regular CW complex
whose underlying space is a closed ball and they improved their result to some
variants of Bore-fixed ideals [18, Theorem 3.8]. More recently, Goodarzi proved
that the Herzog-Takayama resolution [12] is cellular [9]. This is a resolution for the
2000 Mathematics Subject Classification. 13D02, 13F20.
Key words and phrases. Minimal free resolution, cellular resolution, transversal monomial
ideal.
1
2
RAHIM ZAARE-NAHANDI
class of monomial ideals with regular linear quotients. This class includes matroidal
ideals and stable monomial ideals.
In this paper we consider the class of transversal monomial ideals of degree t (see
Definition 2.4). A minimal free resolution for monomial ideals in this class was given
in [23] and [22]. However, in general, this resolution is not polytopal (see Remark
3.4). We construct a labeled polytopal cell complex which provides a canonical
Zt -graded minimal free resolution for transversal monomial ideals of degree t. Any
such monomial ideal is in fact a matroidal ideal (see Remark 2.5). Thus, by [9] its
Herzog-Takayama resolution is cellular. Nevertheless, the Herzog-Takayama resolution is obtained by consecutive applications of mapping cone constructions. In
contrast, the resolution provided here is based on decomposing any ideal in this class
and “gluing” together the polytopal cell complexes associated to each summand to
get a polytopal cell complex which gives the expected resolution (see Lemma 3.2
and Theorem 3.3). This resolution is explicit and is much simpler to compute compared to the Herzog-Takayama resolution (see Example 3.5). Furthermore, unlike
most results which prove cellularity of certain known resolutions, the cell-complex
constructed here provides the resolution.
The paper is organized is as follows. After the introduction in Section 1, in
Section 2 we provide preliminary material and notations. Section 3 is devoted
to the main result, the construction of a polytopal cell complex that supports a
graded minimal free resolution of a transversal monomial ideal. Then the squarefree Veronese ideals of fixed degree, the only Cohen-Macaulay case of this class
of ideals, is treated. In this case, the cell complex is a polytopal subdivision of a
simplex, in particular, its underlying space is homeomorphic to a closed ball. This
last result, using some “depolarization”, also follows by a result of Sinefakopoulos
[20].
2. Notations and preliminaries
We assume familiarity with basic notions of cell complexes, polytopes and monomial ideals referring to [24], [4], [15] and [11]. In particular, we have freely used
several statements and results from [20]. Let I ⊂ R = k[y1 , · · · , yn ] be a monomial
ideal in the polynomial ring over a field k and let G(I) be the unique minimal
monomial generating set of I. Let X be a regular cell complex with vertices labeled
by members of G(I). Let ǫX be an incidence function on X. Any face of X will
be labeled by mF , the least common multiple of the monomials in G(I) on the
vertices of F . If mF = y1a1 · · · ynan , then the degree aF is defined to be the exponent
vector e(mF ) = (a1 , · · · , an ). Let RF be the free R-module with
Lone generator in
degree aF . The cellular complex FX is the Zn -graded R-module ∅6=F ∈X RF with
differentials
X
mF ′
ǫ(F, F ′ )
F ,
∂(F ) =
m
F′
′
∅6=F ∈X
where ǫ(F, F ′ ) is nonzero only if F ′ is a face of F of codimension 1 and is +1 or −1
such that the resulting sequence is a chain complex (see [3]).
If the complex FX is is exact, then FX is called a cellular resolution of I. Alternatively, we say that I has a cellular resolution supported on the labeled cell
A CANONICAL POLYTOPAL RESOLUTION FOR TRANSVERSAL MONOMIAL IDEALS
3
complex X. If X is a simplicial complex or a polytopal cell complex, then FX is
called simplicial and polytopal resolution, respectively. A cellular resolution FX is
minimal if and only if any two comparable faces F ′ ⊂ F of the same degree coincide. For more on cellular resolutions and polytopal complexes we refer to [4], [15]
and [24].
The following two lemmas will be essential for our constructions.
Lemma 2.1. (The gluing lemma) [20, Lemma 6]. Let I and J be two monomial
ideals in R such that G(I + J) = G(I) ∪ G(J). Suppose that
(i) X and Y are labeled regular cell complexes in some RN that supports a minimal
free resolution FX and FY of I and J, respectively, and
(ii) X ∩ Y is a labeled regular cell complex that supports a minimal free resolution
FX∩Y of I ∩ J.
Then X ∪ Y is a labeled regular cell complex that supports a minimal free resolution
of I + J.
Remark 2.2. [20, Remark 7]. For any two monomial ideals I and J, we have
G(I + J) ⊆ G(I) ∪ G(J).
A case where equality holds is when all elements of G(I) ∪ G(J) are of the same
degree, or more generally, when G(I) ∪ G(J) is a minimal set of generators.
Lemma 2.3. [20, Lemma 8]. Let I ⊂ k[y1 , · · · , yk ] and J ⊂ k[yk+1 , · · · , yn ] be
two monomial ideals. Suppose that X and Y are labeled regular cell complexes in
some RN of dimensions k − 1 and n − k − 1, respectively, that support minimal
free resolutions FX and FY of I and J, respectively. Then the labeled cell complex
X × Y supports a minimal free resolution FX×Y of IJ.
Recall that a simplicial complex M on [n] = {1, · · · , n} is called a matroid if, for
any two distinct facets F and G of M and for any i ∈ F , there exists j ∈ G such
that (F \ {i}) ∪ {j} is a facet of M. A matroid M is said to have strong exchange
property if, for any two distinct facets F and G of M and for any i ∈ F , j ∈ G \ F ,
(F \ {i}) ∪ {j} is a facet of M. Let A = {v1 , · · · , vm } be a set of vectors in Rn . The
set of all linearly independent subsets of A of cardinality t ≤ n forms a matroid
which is called a linear matroid. A matroid is called linear if it has a representation
using a vector space over a field. A monomial ideal I ⊂ k[y1 , · · · , yn ] is said to be
a matroidal ideal if it is the facet ideal of a matroid M on [n] under some labeling
of vertices of M by the variables y1 , · · · , yn .
Definition 2.4. We fix positive integers 1 ≤ b1 ≤ · · · ≤ bn and a polynomial ring
S = k[x11 , · · · , x1b1 , · · · , xn1 , · · · , xnbn ]
in m = b1 + · · · + bn variables over a field k. Let Pj = (xj1 , · · · , xjbj ), j = 1, · · · , n.
The ideal
X
X
In,t =
Pj1 . . . Pjt =
Pj1 ∩ . . . ∩ Pjt
1≤j1 <···<jt ≤n
1≤j1 <···<jt ≤n
is called the transversal monomial ideal of degree t on P1 , · · · , Pn .
4
RAHIM ZAARE-NAHANDI
This class of ideals naturally arises in the study of generic singularities of algebraic varieties [19].
It is helpful to notice that the ideal In,t may also
t-minors of a matrix D where
0 ···
0
0 ···
x11 · · · x1b1
0 ···
0 ···
0
x
·
·
·
x
21
2b
2
D= .
.
.
.
.
.
..
..
..
..
..
..
..
..
.
.
0 ···
0
0 ···
0
0 ···
be realized as the ideal of
0
0
..
.
0
0
..
.
0 xn1
···
···
..
.
0
0
..
.
···
xnbn
.
Remark 2.5. Let {v1 , · · · , vn } be a vector space basis in Rn . For each i, 1 ≤ i ≤ n,
let λij , j = 1, · · · , bi , be distinct nonzero real numbers. Let A = {λij vi : 1 ≤ i ≤
n, 1 ≤ j ≤ bi }. In other words, A is a set of vectors linearly independent up to
proportionality. Let M be the linear matroid of all linearly independent subsets
of A of cardinality t ≤ n. Then In,t is the facet ideal of the linear matroid M,
when the vertex corresponding to λij vi is labeled by xij . In particular, In,t is a
matroidal ideal. Therefore, In,t has linear quotients [11, Theorem 12.6.2], and a
linear resolution [1] or [11, Lemma 8.2.1]. On the other hand, the only case In,t is
Cohen-Macaualay is the case of the square-free Veronese ideal of degree t, i.e., the
case b1 = · · · = bn = 1, n = m [11, Theorem 12.6.7]. In this case the corresponding
matroid M has the strong exchange property.
3. A cellular graded resolution for transversal monomial ideals
In this section we construct a polytopal cell complex that supports a graded
minimal free resolution of In,t .
We will need the construction for certain special cases.
Proposition 3.1. Let ∆(Pi ) be the (bi − 1)-simplex with vertices labeled by variables generating Pi . Then the following statements hold:
(a) The labeled cartesian product of simplexes Γn,n = ∆(P1 ) × · · · × ∆(Pt ) supports
a Zn -graded minimal free resolution of In,n = P1 . · · · .Pn .
Pn−1
(b) In,n−1 = i=1 P1 . · · · .Pi−1 .(Pi + Pi+1 ).Pi+2 . · · · .Pn , and the labeled cell complex
Γn,n−1 =
n−1
[
∆(P1 ) × · · · × ∆(Pi−1 ) × ∆(Pi + Pi+1 ) × ∆(Pi+2 ) × · · · × ∆(Pn )
i=1
supports a Zn−1 -graded minimal free resolution of In,n−1 and Γn,n−1 ⊂ Γn,n .
Proof. (a) The Koszul complex on the sequence {xi1 , · · · , xibi } is a minimal free
resolution of Pi , i = 1, · · · , n and ∆(Pi ) supports the minimal free resolution of
Pi . Hence by Lemma 2.3 the cartesian product ∆(P1 ) × · · · ∆(Pn ), as a labeled
polytopal cell complex, supports a Zn -graded minimal free resolution of In,n .
(b) We use induction on n for both claims. A generator of In,n−1 is either a
monomial with a variable in Pn or with a variable in Pn−1 + Pn . Thus we have
In,n−1 = In−1,n−2 .Pn + In−2,n−2 .(Pn−1 + Pn ).
A CANONICAL POLYTOPAL RESOLUTION FOR TRANSVERSAL MONOMIAL IDEALS
5
Replacing In−1,n−2 by induction hypothesis and In−2,n−2 by (a) the equality for
In,n−1 follows.
Now by induction hypothesis and Lemma 2.3, Γn−1,n−2 × ∆(Pn ) and Γn−2,n−2 ×
∆(Pn−1 + Pn ) provide labeled cell complexes which support Zn−1 -graded minimal
free resolutions of In−1,n−2 .Pn and In−2,n−2 .(Pn−1 + Pn ), respectively. The conditions of the gluing lema are satisfied for these two ideals and the cell complexes. In
fact, since the generators of Pi ’s form a partition of the set of variables, we have
[In−1,n−2 .Pn ] ∩ [In−2,n−2 .(Pn−1 + Pn )]
= [In−1,n−2 ∩ Pn ] ∩ [In−2,n−2 ∩ (Pn−1 + Pn )]
= In−2,n−2 .Pn .
On the other hand, replacing Γn−1,n−2 and Γn−2,n−2 by induction hypothesis and
(a), we get
[Γn−1,n−2 × ∆(Pn )] ∩ [Γn−2,n−2 × ∆(Pn−1 + Pn )] = Γn−2,n−2 × ∆(Pn ),
while the last equality follows by the inclusion Γn−2,n−2 ⊂ Γn−1,n−2 which is a
consequence of the expressions of them using induction hypothesis. Therefore,
[Γn−1,n−2 × ∆(Pn )] ∪ [Γn−2,n−2 × ∆(Pn−1 + Pn )] = Γn,n−1
is a Zn−1 -graded minimal free resolution of In,n−1 .
Finally, the inclusion Γn,n−1 ⊂ Γn,n follows by the expression of the cell complexes
in (b).
To prove the main result, we will further need the following lemma.
Lemma 3.2. For integers 1 ≤ q ≤ p ≤ n and 1 ≤ r ≤ n, let
X
Ip,q =
Pj1 . . . . .Pjq ,
1≤j1 <···<jq ≤p
and
Qr =
n
X
Pj .
j=r
Then;
(a) For all t, 2 ≤ t ≤ n, the following holds
In,t =
n−1
X
Ii,t−1 .Qi+1 .
i=t−1
(b) For all s, t ≤ s ≤ n, let
In,t (s) =
s−1
X
Ii,t−1 .Qi+1 .
i=t−1
Then for all s, t ≤ s ≤ n − 1, we have
In,t (s) ∩ [Is,t−1 .Qs+1 ] = Is−1,t−1 .Qs+1 .
(By (a), In,t (s) is the transversal monomial ideal of degree t on P1 , · · · , Ps−1 , and
Qs = Ps + · · · + Pn .)
6
RAHIM ZAARE-NAHANDI
Proof. For 1 ≤ r ≤ s ≤ n, let
Qr (s) =
s
X
Pj ,
j=r
so that in particular, Qr (n) = Qr .
To prove (a) we use induction on n. For n = t the equality is clear. Assuming the
quality for integers less than n, the inclusion ⊇ is obvious for n because each term
in the sum is contained in In,t . For the other inclusion, observe that
In,t = In−1,t + In−1,t−1 .Qn .
By induction hypothesis, we have
In−1,t =
n−2
X
Ii,t−1 .Qi+1 (n − 1).
i=t−1
Replacing this in the previous equality the inclusion ⊆ also follows by term by term
inclusion.
(b). By (a) we get
In,t (s) =
s−1
X
Ii,t−1 .Qi+1 (s) + (
s−1
X
Ii,t−1 ).Qs+1 = Is,t + Is−1,t−1 .Qs+1 .
i=t−1
i=t−1
Since the modular law holds for monomial ideals, we have
In,t (s) ∩ [Is,t−1 .Qs+1 ] = [Is,t + Is−1,t−1 .Qs+1 ] ∩ [Is,t−1 .Qs+1 ]
= Is,t ∩ [Is,t−1 .Qs+1 ] + [Is−1,t−1 .Qs+1 ] ∩ [Is,t−1 .Qs+1 ].
Since Is−1,t−1 ⊆ Is,t−1 , the resulting expression is equal to
Is,t ∩ [Is,t−1 .Qs+1 ] + [Is−1,t−1 .Qs+1 ].
But Is,t−1 .Qs+1 = Is,t−1 ∩ Qs+1 and Is,t ∩ Qs+1 = Is,t .Qs+1 because the set of
variables involved in the generators of Is,t−1 or Is,t is disjoint from the the set of
variable generators of Qs+1 . Thus by the fact that Is,t ⊆ Is,t−1 we get
Is,t ∩ [Is,t−1 .Qs+1 ] = Is,t ∩ Is,t−1 ∩ Qs+1 = Is,t ∩ Qs+1 = Is,t .Qs+1 .
Therefore, using the inclusion Is,t ⊆ Is−1,t−1 , we have
In,t (s) ∩ [Is,t−1 .Qs+1 ] = [Is,t .Qs+1 ] + [Is−1,t−1 .Qs+1 ] = Is−1,t−1 .Qs+1 ,
as required.
We may now give the main result of the paper.
Theorem 3.3. There exists a labeled regular polytopal cell complex Γn,t ⊂ Rm−t
that supports a Zt -graded minimal free resolution of In,t . Explicitly,
Γn,t =
n−1
[
[Γi,t−1 × ∆(Qi+1 )],
i=t−1
where byP
∆(Qi+1 ) we mean the simplex with vertices on the variables generating
n
Qi+1 = j=i+1 Pj . Moreover, Γs,t is connected, and as labeled polytopes, Γs,t ⊂
Γs+1,t for all s, t ≤ s ≤ n − 1.
A CANONICAL POLYTOPAL RESOLUTION FOR TRANSVERSAL MONOMIAL IDEALS
7
Proof. Observe that the cell complex Γn,t defined recursively as above, can also be
labeled recursively. Thus we first show the inclusion Γk,t ⊂ Γk+1,t as labeled cell
complexes, for all k, t ≤ k ≤ n − 1. Using the notation of the previous lemma we
have
k
k−1
[
[
[Γi,t−1 × ∆(Qi+1 (k + 1))] = Γk+1,t .
[Γi,t−1 × ∆(Qi+1 (k))] ⊂
Γk,t =
i=t−1
i=t−1
For all s, t ≤ s ≤ n, similar to In,t (s) defined in the previous lemma, let
Γn,t (s) =
s−1
[
[Γi,t−1 × ∆(Qi+1 )],
i=t−1
so that Γn,t (n) = Γn,t . Thus we have
Γn,t (s) = Γn,t (s − 1) ∪ [Γs−1,t−1 × ∆(Qs )].
Using this identity we prove by induction on s, t ≤ s ≤ n, that Γn,t (s) supports
a Zt -graded minimal free resolution of In,t (s), implying the claim for s = n. For
s = t, Γn,t (t) = Γt−1,t−1 × ∆(Qs ) which support a graded minimal free resolution
of In,t (t) = It−1,t−1 .Qs by Proposition 3.1 and Lemma 3.2. Assume that Γn,t (s)
supports a graded minimal free resolution of In,t (s) for all s, n, t with t ≤ s < s0 ≤
n. We claim
Γn,t (s0 ) = Γn,t (s0 − 1) ∪ [Γs0 −1,t−1 × ∆(Qs0 )]
supports a graded minimal free resolution of
In,t (s0 ) = In,t (s0 − 1) + [Is0 −1,t−1 .Qs0 ].
We apply the gluing lemma. By induction hypothesis each summand of Γn,t (s0 )
supports a graded minimal free resolution of the corresponding summand in In,t (s0 ).
By the previous lemma,
In,t (s0 − 1) ∩ [Is0 −1,t−1 .Qs0 ] = Is0 −2,t−1 .Qs0 .
We also have
Γn,t (s0 − 1) ∩ [Γs0 −1,t−1 × ∆(Qs0 )] = [
s[
0 −2
[Γi,t−1 × ∆(Qi+1 )] ∩ [Γs0 −1,t−1 × ∆(Qs0 )]
i=t−1
=
s[
0 −2
[(Γi,t−1 ∩ Γs0 −1,t−1 ) × (∆(Qi+1 ) ∩ ∆(Qi+1 )] = Γs0 −2,t−1 × ∆(Qs0 ),
i=t−1
where the last equality is valid by Proposition 3.1. But Γs0 −2,t−1 ×∆(Qs0 ) supports
a graded minimal free resolution of Is0 −2,t−1 .Qs0 , which is what we need for the
gluing lemma. Therefore, Γn,t (s0 ) supports a graded minimal free resolution of
In,t (s0 ) as required. The regularity of Γn,t is clear because every polytopal cell
complex is regular [24]. The connectedness of Γn,t follows from the equality
Γn,t (s0 − 1) ∩ [Γs0 −1,t−1 × ∆(Qs0 )] = Γs0 −2,t−1 × ∆(Qs0 ).
Remark 3.4. The minimal free resolution provided in [23, Theorem 3.1] and [22,
Theorem 2.1] for a transversal monomial ideal is not polytopal for 2 ≤ t ≤ n − 3.
As an example, consider I5,2 (m = n = 5). Following the construction there,
the corresponding cell complex is the same as the Morse complex which supports a
8
RAHIM ZAARE-NAHANDI
minimal free resolution of (y1 , y2 , y3 , y4 )2 ⊂ k[y1 , y2 , y3 , y4 ] illustrated in [20, Fig.
3] which is not polytopal. In fact, the two “faces” {x1 x2 , x1 x3 , x3 x5 , x2 x3 } and
{x1 x2 , x1 x3 , x3 x4 , x2 x3 } have two edges in common but their union is not a face of
either of them.
Example 3.5. Consider I4,3 with b1 = b2 = 2, b3 = b4 = 1. Then
Γ4,3 = [∆(x11 , x12 , x21 , x22 ) × ∆(x31 ) × ∆(x41 )] ∪ [∆(x11 , x12 ) × ∆(, x21 , x22 , x31 ) ×
∆(x41 )] ∪ [∆(x11 , x12 ) × ∆(, x21 , x22 ) × ∆(x31 , x41 )].
This is illustrated in Figure 1.
x11 x21 x31
x11 x21 x41
x12 x22 x31
x11 x22 x31
x22 x31 x41
x11 x31 x41
x12 x22 x41
x11 x22 x41
x12 x21 x31
x21 x31 x41
x12 x31 x41
x12 x21 x41
F igure 1
The minimal free resolution of I4,3 (ignoring the fine grading) is immediate from
Figure 1 as
0 −→ S 3 (−5) −→ S 22 (−4) −→ S 12 (−3) −→ I4,3 −→ 0.
The Herzog-Takayama resolution for I4,3 requires 11 consecutive mapping cone constructions.
This example reveals that, in general, the cell complex Γn,t is not convex, and its
underlying topological space is not homeomorphic to a closed ball.
For m = n, i.e., b1 = · · · = bn = 1, In,t is the square-free Veronese ideal of degree
t on the variables. To distinguish this case, we use the notation xi , Im,t and Γm,t
instead of xi1 , In,t and Γn,t , respectively. In this case, Γm,t enjoys the following
further features.
A CANONICAL POLYTOPAL RESOLUTION FOR TRANSVERSAL MONOMIAL IDEALS
9
Proposition 3.6. The polytopal cell complex Γm,t which supports the Zt -graded
minimal free resolution of the square-free Veronese ideal of degree t on x1 , · · · , xm ,
is a polytopal subdivision of the (m − t)-simplex
∆(x1 · · · xt , x2 · · · xt+1 , · · · , xn−t+1 · · · xm ).
In particular, Γm,t is shellable and its underlying topological space is homeomorphic
to a closed ball.
Proof. The proof is by induction on t. For t = 1, it is trivial; Γm,1 = ∆(x1 , · · · , xm ).
We also need to check the claim for t = 2. Adopting the notation of Lemma 3.2,
we have
m−1
m−1
[
[
[∆(x1 , · · · , xi ) × ∆(Qi+1 )].
[Γi,1 × ∆(Qi+1 )] =
Γm,2 =
i=1
i=1
Thus, using the notation || for the underlying topological spaces of the polytopes
in Rm−t , we have
|Γm,2 | =
m−1
[
|∆(x1 , · · · , xi )| × |∆(Qi+1 )| = |∆(x1 x2 , x2 x3 , · · · , xm−1 xm )|,
i=1
where the last equality is a natural partition of the (m − 2)-simplex on
{x1 x2 , x2 x3 , · · · , xm−1 xm }
into the union of polytopes while each polytope is the product of an (i − 1)-simplex
on {x1 , · · · , xi } with an (m − i − 1)-simplex on {xi+1 , · · · , xm }, i = 1, · · · , m − 1.
Thus, Γm,2 is a regular subdivision of the (m−2)-simplex ∆(x1 x2 , x2 x3 , · · · , xm−1 xm ).
For the induction step, by the previous theorem
Γm,t =
m−1
[
[Γi,t−1 × ∆(Qi+1 )].
i=t−1
By induction hypothesis
|Γm,t | =
m−1
[
[|∆(yt−1 , · · · , yi )| × |∆(Qi+1 )|],
i=t−1
where, yi = xi−t+2 · · · xi , i = t − 1, · · · , m − 1. Thus by the case t = 2,
|Γm,t | = |∆(yt−1 xt , yt xt+1 , · · · , ym−1 xm )| = |∆(x1 · · · xt , x2 · · · xt+1 , · · · , xm−t+1 · · · xm )|
as required.
Shellability of Γn,t follows from a general result that any regular subdivision of a
simplex is shellable [24, page 243]. The underlying topological space of Γn,t is a
closed ball since it is a regular subdivision of a simplex.
Remark 3.7. Proposition 3.6 also follows by a result of Sinefakopoulos [20, Theorem 12]. In fact, let s = n − t + 1and let J = (y1 , · · · , ys )t ⊂ k[y1 , · · · , ys ]. The
depolarization given by Nagel and Reiner (see Section 3 [16])
xi1 xi2 · · · xit 7→ yi1 yi2 −1 · · · yit −t+1 , i1 < i2 < · · · < it
is a bijection on the minimal monomial generators of In,t and J. The inverse map
is
y1j1 y2j2 · · · ysjs 7→ x1 x2 · · · xj1 xj1 +2 xj1 +3 · · · xj1 +j2 +1 · · · xj1 +···+js−1 +s · · · xj1 +···+js +s−1 ,
10
RAHIM ZAARE-NAHANDI
where j1 + · · · + js = t. The change of labeling xi1 xi2 · · · xit to yi1 yi2 −1 · · · yit −t+1
converts the polytopal cell complex Γn,t into the polytopal cell complex Pt (y1 , · · · , ys )
constructed by Sinefakopoulos which supports the resolution of J (see the proof of
[20, Theorem 12] for notations). Observe that under this change of labeling, the
subdivision structures of Γn,t on the simplex ∆(x1 · · · xt , · · · , xn−t+1 · · · xn ) converts
to that of Pt (y1 , · · · , ys ) on ∆(y1t , · · · , yst ).
Remark 3.8. For m > n the ideal In,t is not Borel-fixed (i.e., strongly stable
for characteristic zero) under certain depolarization similar to Remark 3.7. For
example, consider I2,2 with b1 = b2 = 2. Then it follows that I2,2 is not strongly
stable under the expected depolarization.
Question 3.9. Sinefakopoulos [20] has used the gluing procedure to build up
a polytopal cell complex which gives a minimal free resolution for the Borel-fixed
monomial ideals which are Borel-fixed generated in one degree. In this paper similar
procedure has been employed for the transversal monomial ideals. One expects
other monomial ideals for which the gluing construction provides a minimal free
resolution. In particular, we may ask whether one could construct a similar cell
complex to the Herzog-Takayama resolution, and if possible, when is the resulting
cell complex homeomorphic to a closed ball?
Acknowledgments
The author is grateful to Gunnar Fløystad for his comments on the first attempt
towards this paper, for bringing attention to [20] and [16], and for suggesting to
work on the cellularity of transversal monomial ideals. He thanks Afshin Goodarzi
for his useful comments on the final version of the paper. This work has been
partially supported by research grant no. 4/1/6103011 of University of Tehran.
References
1. A. Björner, Homology and shellability of matroids and geometric lattices, in Matroid Applications (ed. by N. White), Cambridge University Press, Cambridge (1992).
2. D. Bayer, I. Peeva and B. Sturmfels, Monomial resolutions, Math. Res. Lett. 5 (1998) 31–46.
3. D. Bayer and B. Sturmfels, Cellular resolutions of monomial modules, J. Reine Angew. Math.
502 (1998) 123–140.
4. W. Bruns and J. Herzog, Cohen-Macaulay Rings, Cambridge University Press, 1998.
5. A. Dochtermann and A. Engström, Cellular resolutions of cointerval ideals, Math. Z. 270
(1-2) (2012) 145–163.
6. S. Eliahou and M. Kervaire, Minimal resolutions of some monomial ideals, J. Algebra 129
(1990) 1–25.
7. T. Clark, A minimal poset resolution of stable ideals In: Progress in Commutative Algebra I:
Combinatorics and Homology, pp. 143-166, de Gruyter, Berlin, 2012.
8. A. Engström and P. Norén, Cellular resolutions of powers of monomial ideals,
arXiv:1212.2146v.1 10 Dec 2012.
9. A. Goodarzi, Cellular structure for the Herzog-Takayam resolution, J. Algebr. Comb. 41
(2015) 21–28.
10. G. Flöystad, Cellular resolutions of Cohen-Macaulay monomial ideals, J. Commut. Algebra
1 (1) (2009), 57–89.
11. J. Herzog and T. Hibi, Monomial Ideals, Grad. Texts Math. 260, Springer, New York, 2011.
12. J. Herzog and Y. Takayama, Resolutions by mapping cones, The Roos Festschrift, Vol. 2,
Homology, Hmotopt Appl. 4 (2, part 2) (2002) 277–294.
13. M Jöllenbeck and V. Welker, Minimal resolutions via algebraic dicrete Morse thory, Mm.
Aer. Math. Soc. 197 (2009).
A CANONICAL POLYTOPAL RESOLUTION FOR TRANSVERSAL MONOMIAL IDEALS 11
14. J. Mermin, The Eliahou-Kervaire resolution is cellular, J. Commut. Algebra 2 (1) (2010)
55–78.
15. E. Miller and B. Sturmfels, Combinatorial Commutative Algebra, Grad. Texts Math. 227,
Springer, New York, 2005.
16. U. Nagel and V. Reiner, Betti numbers of monomial ideals and shifted skew shapes, The
Electronic Journal of Combinatorics 16 (2) (2009), #R3.
17. I. Novik, A. Postnikov, and B. Sturmfels, Syzygies of oriented matroids, Duke Math. J. 111
(2) (2002) 287–317.
18. R. Okazaki, and K. Yanagawa, On CW complexes supporting the Eliahou-Kervaire type resolutions of Borel fixed ideals, Collect. Math. 66 (1) (2015) 125–147.
19. P. Salmon and R. Zaare-Nahandi, Ideals of minors defining generic singularities and their
Gröbner bases, Commun. Algebra 33 (8) (2005) 2725–2747.
20. A. Sinefakopoulos, On Borel fixed ideals generated in one degree, J. Algebra 319 (7) (2008)
2739–2760.
21. M. Valesco, Minimal free resolutions that are not supported by a CW-complex, J. Algebra 319
(2008) 102–114.
22. R. Zaare-Nahandi, Betti numbers of transversal monomial ideals, Algebra Colloq. 18 (Spec
1) (2011) 925–936.
23. Rahim Zaare-Nahandi and Rashid Zaare-Nahandi, The minimal free resolution of a class of
square-free monomial ideals, J. Pure Appl. Algebra 189 (2004) 263-278.
24. G. Ziegler, Lectures on Polytopes, Grad. Texts in Math., vol. 152, Springer-Verlag, New York,
1995.
| 0math.AC
|
PREDICTION-ADAPTATION-CORRECTION RECURRENT NEURAL NETWORKS
FOR LOW-RESOURCE LANGUAGE SPEECH RECOGNITION
Yu Zhang1 , Ekapol Chuangsuwanich1 , James Glass1 , Dong Yu2
1
MIT Computer Science and Artificial Intelligence Laboratory
arXiv:1510.08985v1 [cs.CL] 30 Oct 2015
{yzhang87, ekapolc, glass}@mit.edu
2
Microsoft Research
[email protected]
ABSTRACT
In this paper, we investigate the use of prediction-adaptationcorrection recurrent neural networks (PAC-RNNs) for lowresource speech recognition. A PAC-RNN is comprised of a
pair of neural networks in which a correction network uses
auxiliary information given by a prediction network to help
estimate the state probability. The information from the correction network is also used by the prediction network in a
recurrent loop. Our model outperforms other state-of-theart neural networks (DNNs, LSTMs) on IARPA-Babel tasks.
Moreover, transfer learning from a language that is similar to
the target language can help improve performance further.
Index Terms— DNN, LSTM, PAC-RNN, Multilingual
1. INTRODUCTION
The behavior of prediction, adaptation, and correction is
widely observed in human speech recognition [1]. For example, listeners may guess what you will say next and wait
to confirm their guess. They may adjust their listening effort
by predicting the speaking rate and noise condition based on
current information, or predict and adjust a letter to sound
mapping based on the talker’s pronunciations.
Previously [2], we proposed the prediction-adaptationcorrection RNN (PAC-RNN) which tries to emulate some of
these mechanisms by using two DNNs; a prediction DNN
that predicts the next phoneme, and a correction DNN that
estimates the current state probability based on both the current frame and the hypothesis from the prediction DNN. The
model showed promising results on TIMIT, but it was unclear
whether a similar gain could be achieved on larger ASR tasks
where the prediction information might already be incorporated by the language models. Here, we successfully apply
the PAC-RNN to LVCSR on several low-resource languages
that are currently being used in the IARPA-Babel program.
This work was supported in part by the Intelligence Advanced Research
Projects Activity (IARPA) via Department of Defense US Army Research
Laboratory contract number W911NF-12-C-0013. The U.S. Government is
authorized to reproduce and distribute reprints for Governmental purposes
notwithstanding any copyright annotation thereon. Disclaimer: The views
and conclusions contained herein are those of the authors and should not be
interpreted as necessarily representing the official policies or endorsements,
either expressed or implied, of IARPA, DoD/ARL, or the U.S. Government.
In addition, we study the effect of transfer learning for recurrent architectures. Recurrent networks such as LSTMs [3,
4] are known to require a large amount of training data in
order to perform well [5]. For the IARPA-Babel tasks, multiple groups have incorporated multilingual training in order
to alleviate data limitation issues. One popular approach is
multi-task training using DNNs. In a multi-task setup, a single DNN is trained to generate outputs for multiple languages
with some tied parameters. This approach has been used
for robust feature extraction via bottleneck (BN) features [6,
7, 8, 9], or for classifiers in hybrid DNN-HMM approaches
[10, 11]. In [12], Karafiát et al. found that using CMLLR
transformed BN features as inputs to a hybrid DNN could further improve ASR performance. However, we believe none of
this research has investigated recurrent networks for low resource languages in a multilingual scenario.
The work presented here is an extension of [2] based
on our multilingual framework in [7]. We first extract BN
features using multilingual networks to train different hybrid neural network architectures. Experiments show that the
LSTMs outperform DNNs, and that the PAC-RNN provides
the biggest gains for this task. Additional improvements are
observed when the models are adapted from networks trained
on languages that are most similar to the target language.
The rest of the paper is organized as follows. In Section
2, we review the PAC-RNN model, and describe an enhanced
version that incorporates an LSTM. In Section 3, we describe
our multilingual system and how it is used with the PACRNN. We explain our experiments results in Section 4.
2. PREDICTION-ADAPTATION-CORRECTION
RECURRENT NEURAL NETWORKS
2.1. Model structure and training
The PAC-RNN used in this work follows our previous work
in [2]. Fig. 1 illustrates the structure of the PAC-RNN studied
in this paper. The main components of the model are a correction DNN and a prediction DNN. The correction DNN estimates the state posterior probability pcorr (st |ot , xt ) given ot ,
the observation feature vector, and xt , the information from
the prediction DNN, at time t. The prediction DNN predicts
p
Prediction DNN
Prediction info
Delay
Softmax
Bottleneck
Sigmoid
Sigmoid
pcorr (st |ot , xt )
Softmax
Sigmoid
pred >
xt = [hpred
t T corr , . . . , ht 1 ]
Sigmoid
Context expansion
hpred
t
Sigmoid
Input features
ot
yt = hcorr
t
(lt+n |ot , yt )
Projection
layer
Auxiliary info for prediction
pred
ot
Correction DNN
Fig. 1. The Structure of the PAC-RNN-DNN
future target information. Note that since yt , the information from the correction DNN, depends on xt , the information
from the prediction DNN, and vice versa, a recurrent loop is
formed.
The information from the prediction DNN, xt , is from a
bottleneck hidden layer output value hpred
t−1 . To exploit additional previous predictions, we stack multiple hidden layer
values as
pred T
xt = [hpred
(1)
t−T corr , ..., ht−1 ] ,
where α is the interpolation weight, and is set to 0.8 in our
study unless otherwise stated, and T is the total number of
frames in the training utterance. Note that in a standard PACRNN as described here, both the correction model and prediction model are DNNs. From this point onwards we will call
this particular setup, the PAC-RNN-DNN.
where T corr is the contextual window size used by the correction DNN and is set to 10 in our study. Similarly, we can
stack multiple frames to form yt , the information from the
correction DNN, as
LSTMs have improved speech recognition accuracy on many
tasks over DNNs [3, 4, 5]. To further enhance the PAC-RNN
model, we use an LSTM to replace the DNN used in the correction model. The input of this LSTM is the acoustic feature
ot concatenated with the information from prediction model,
xt . The prediction model can also be an LSTM but we did
not observe performance gain on the experiments. To keep it
simple, we use the same DNN prediction model as [2].
corr T
yt = [hcorr
] ,
t−T pred −1 , ..., ht
(2)
where T pred is the contextual window size used by the prediction DNN and is set to 1 in our study. In addition, in the
specific example shown in Fig. 1, the hidden layer output
hcorr
is projected to a lower dimension before it is fed into
t
the prediction DNN.
To train the PAC-RNN, we need to provide supervision information to both the prediction and correction DNNs. As we
have mentioned, the correction DNN estimates the state posterior probability, and thus the state label, so that the frame
cross-entropy (CE) criterion can be used. For the prediction
DNN, we follow [2], and use the phoneme label for the prediction targets.
The PAC-RNN training problem is a multi-task learning
problem. The two training objectives can be combined into a
single one as
J=
T
X
2.2. PAC-RNN-LSTM
3. STACKED BOTTLENECK ARCHITECTURE
3.1. Stacked bottleneck (SBN) features
The BN features used in this work follow our previous work
in [13]. An SBN is a hierarchical architecture realized as a
concatenation of two DNNs, each with its own bottleneck
layer. The outputs from the BN layer in the first DNN are
used as the input features for the second DNN, whose outputs
at the BN layer are then used as the final features for standard
GMM-HMM training.
3.2. Bottleneck-CMLLR features in a hybrid system
(α∗ln pcorr (st |ot , xt )+(1−α)∗ln ppred (lt+n |ot , yt )), In [12], the authors proposed a DNN hybrid system that
used the first stage BN features with speaker adaptation (BNt=1
(3)
CMLLR). This system yielded the best results for the Babel
LID-based multilingual SBN
Purely
Multilingual
Features
Target
Language
Features
are initialized from the DNN, LSTM or PAC-RNN which is
trained using the closest language.
Random Initialization
Closest
Language
Multi‐
lingual
4. EXPERIMENTS
Closest
Language
Adapt
Target
lang
Target
lang
Second DNN
First DNN
Multi‐
lingual
Adapt
Target
lang
Target
lang
Training Progress
Fig. 2. Steps to adapt a multilingual SBN to a target language via
the closest language selected via LID.
evaluation that year. In this work, we follow a similar approach by replacing the DNN with recurrent architectures
(LSTM or PAC-RNN). The BN-CMLLR features were taken
from a network trained in a multilingual fashion and adapted
to the target language. For the DNN and PAC-RNN, these
features were stacked in context of 31(±15) frames and
downsampled by a factor of 5. Following [5], no context
expansion is used for the LSTM. The output state label is also
delayed by 5 to utilize the information from the future.
4.1. IARPA-Babel corpus
The IARPA-Babel program focuses on ASR and spoken term
detection on low-resource languages [15]. The goal of the
program is to reduce the amount of time needed to develop
ASR and spoken term detection capabilities in a new language. The data from the Babel program consists of collections of speech from a growing list of languages. The project
is on its fourth year. For this work we will consider the Full
pack (60-80 hours of training data) of the 11 languages released in the first two years as source languages, while the
languages in the third year will be the target languages [16].
Some languages also contain a mixture of microphone data
recorded at 48kHz in both train and test utterances. For the
purpose of this paper, we downsampled all the wideband data
to 8kHz and treated it the same way as the rest of the recordings. For the target languages, we will focus on the Very Limited Language Pack (VLLP) condition which includes only 3
hours of transcribed training data. This condition excludes
any use of human generated pronunciation dictionary. Unlike
in the previous two years of the program, usage of web data is
permitted for language modeling and vocabulary expansion.
3.3. Multilingual training and adaptation of SBN features
The multilingual training of the SBN follows [14] where all
the DNN targets from each language are pooled together, each
with its own softmax layer. Adapting the multilingual SBN to
a target language can be done by performing additional finetuning steps on each DNN sequentially using the data from
the target language. Our previous work [7] shows that using
just the language closest to the target language from the pool
of source languages to train the second DNN can serve as a
better initialization model than the multilingual second DNN.
The closest language can be identified from just the acoustic
data by training a Language Identification (LID) system.
A flowchart of how a LID-based multilingual system can
be trained is shown in Fig. 2. We start by adapting the first
DNN with data from the target language. Instead of using the
second multilingual DNN to initialize, we train the second
DNN from random initialization using the closest language’s
data and output targets. After the DNN converges, we then do
a final adaptation to the target language.
3.4. Multilingual training of BN-hybrid system
The input of the hybrid system (DNN, LSTM or PAC-RNN)
is the same as the second DNN in the SBN system. During the adaptation stage, the softmax is replaced by the target
language state labels (phone labels for the PAC-RNN prediction model) with random initialization while the hidden layers
4.2. Recognition system
For each language, we used tied-state triphone CD-HMMs,
with 2500 states and 18 Gaussian components per state.
Grapheme-based dictionaries were used for the target languages. Note that for IARPA-Babel languages, the difference
between phonetic and graphemic systems in WER are often
less than 1% [17, 18]. All the output targets were from CD
states. To train the multilingual SBN, we kept only the SIL
frames that appear 5 frames before and after actual speech.
This reduced the total amount of frames for the multilingual
DNN to around 520 hours. We observed no loss in accuracy
from doing so, and it also reduced the training time significantly. Discriminative training was done on the CD-HMMs
using the Minimum Bayes risk (MBR) criterion [19]. The
web data was cleaned and filtered using techniques described
in [20]. For language modeling, n-gram LMs were created from training data transcripts and the web data. The
LMs were then combined using weighted interpolation. The
vocabulary included words that appeared in the training transcripts augmented with the top 30k most frequent words from
the web. We chose 30k words by looking at the rate of OOV
reduction as we augmented the train vocabulary with frequent words from the web. We report results on the 10-hour
development set.
We consider two baseline hybrid systems: a DNN with
three 1024-unit hidden layers, and a stacked LSTM with three
layers each containing 512 cells. No gains were observed by
further increasing the model size of these baseline systems.
In the PAC-RNN model, the prediction DNN has a 2048unit hidden layer and a 80-unit bottleneck layer. For the correction model, we have two systems: a DNN with two 2048unit hidden layers or an LSTM with 1024 memory cells. The
correction model’s projection layer contains 500 units.
All models are randomly initialized without either generative or discriminative pretaining. No momentum is used
for the first epoch and a momentum of 0.9 is used for all the
subsequent epochs. We have found that turning off the momentum for the first epoch helps to improve the performance
of the final model. To train the DNN, a learning rate of 0.1
per mini-batch is used for the first epoch. The learning rate
is increased to 1.0 at the second epoch, after which it is kept
the same until the development set training criterion no longer
improves, under which condition the learning rate is halved.
A similar schedule is used to train the LSTMs and PAC-RNNs
except that all the learning rates are reduced to 1/10 of that
used in the DNN training.
We implemented the hybrid models using the computational network toolkit (CNTK) [21]. The truncated backpropagation-through-time (BPTT) [22] is used to update the
model parameters and each utterance is segmented into multiple chunks. To speed up the training, we process multiple
utterances simultaneously as a batch. We have found that this
reduces training time and improves the quality of the final
model. In this study each BPTT segment contains 20 frames
and we process 20 utterances simultaneously. For decoding,
we fed the posteriors generated by CNTK into the Kaldi ASR
toolkit [23], which then generates the recognition results.
4.3. PAC-RNN results with BN features
Table 1 summarizes the WERs achieved with different models evaluated in this study. The first three rows are the results
from SBN systems. Both the multilingual and the closest
language systems are adapted to the target language for the
whole stacked network. For the hybrid systems, the input is
the BN features extracted from the first DNN of the adapted
multilingual SBN.
The DNN hybrid system outperforms the multilingual
SBN but is very similar to the closest language system. The
LSTM improves upon the DNN by around 1%. The PACRNN-DNN outperforms LSTM by another percent across all
languages. By simply replacing the correction model with a
single layer LSTM, we observe even further improvements.
4.4. Effect of transfer learning on recurrent architectures
In this subsection we investigate the effect of the multilingual
transfer learning for each model. We first use the rich resource
closest language (based on the LID prediction shown in the
table) to train DNN, LSTM and PAC-RNN models, and then
Target language
Cebuano Kurmanji Swahili
Closest language
Tagalog
Turkish
Zulu
SBN models
Monolingual
73.5
86.2
65.8
Adapted multilingual
65.0
75.5
54.9
Closest language
63.7
75.0
54.2
Hybrid models
DNN
63.9
74.9
54.0
LSTM
63.0
74.0
53.0
PACRNN-DNN
62.1
72.9
52.1
PACRNN-LSTM
60.6
72.5
51.4
Hybrid models with closest language initialization
DNN
62.7
73.1
52.4
LSTM
61.3
72.5
52.2
PAC-RNN-DNN
60.8
71.8
51.6
PAC-RNN-LSTM
59.7
71.4
50.4
Table 1. WER (%) results for each ASR system.
adapt them to the target language. The lower part of Table 1
summarizes the ASR results. As shown, the LSTM models
perform significantly better than the baseline SBN system.
Using the PAC-RNN model yields a noticeable improvement
over the LSTM. Similarly, the PAC-RNN-LSTM can further
improve the results.
5. CONCLUSION
In this paper, we explored a PAC-RNN model for lowresource language speech recognition. The results on multiple
languages demonstrated that the PAC-RNN achieves better
performance than DNNs and LSTMs. We also showed that
by replacing the correction model in the PAC-RNN with an
LSTM could further enhance the model. Moreover, the multilingual experiment results show that traditional DNN transfer
learning approaches can also be applied to the PAC-RNN architecture. Our future work includes applying the PAC-RNN
to tasks on which conventional models do not work well, and
extending it by predicting additional information such as the
speech signal, speaker, speaking rate, and noise.
Acknowledgements
The authors would like to thank everyone in the Babelon team for
feedback and support for various resources. The work uses the following language packs: Cantonese (IARPA-babel101-v0.4c), Turkish (IARPA-babel105b-v0.4), Pashto (IARPA-babel104b-v0.4aY),
Tagalog (IARPA-babel106-v0.2g) and Vietnamese (IARPA-babel107bv0.7), Assamese (IARPA-babel103b-v0.3), Lao (IARPA-babel203bv2.1a), Bengali (IARPA-babel102b-v0.4), Zulu (IARPA-babel206bv0.1e), Tamil (IARPA-babel204b-v1.1b), Cebuano (IARPA-babel301bv2.0b), Kurmanji (IARPA-babel205b-v1.0a), and Swahili (IARPAbabel202b-v1.0d-build).
6. REFERENCES
[1] F. Guenther and J. Perkell, “A neural model of speech
production and its application to studies of the role of
auditory feedback in speech,” Speech Motor Control in
Normal and Disordered Speech, 2004.
[2] Y. Zhang, D. Yu, M. Seltzer, and J. Droppo, “Speech
recognition with prediction-adaptation-correction recurrent neural networks,” in Proc. ICASSP, 2015.
[3] A. Graves, A. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” in Proc.
ICASSP, 2013.
[4] A. Graves, N. Jaitly, and A. Mohamed, “Hybrid speech
recognition with deep bidirectional LSTM,” in Proc.
ASRU, 2013.
[5] H. Sak, A. Senior, and F. Beaufays, “Long short-term
memory recurrent neural network architectures for large
scale acoustic modeling,” in Fifteenth Annual Conference of the International Speech Communication Association, 2014.
[6] Z. Tüske, P. Golik, D. Nolden, R. Schlüter, and H. Ney,
“Data augmentation, feature combination, and multilingual neural networks to improve ASR and KWS performance for low-resource languages,” in Proc. InterSpeech, 2014.
[7] E. Chuangsuwanich, Y. Zhang, and J. Glass, “Language ID-based training of multilingual stacked bottleneck features,” in Proc. InterSpeech, 2014.
[8] F. Grézl, M. Karafiát, and K. Veselý, “Adaptation of
multilingual stacked bottle-neck neural network structure for new languages,” in Proc. ICASSP, 2014.
[9] Y. Miao, H. Zhang, and F. Metze, “Distributed learning
of multilingual DNN feature extractors using GPUs,” in
Proc. InterSpeech, 2014.
[10] J. Huang, J. Li, D. Yu, L. Deng, and Y. Gong, “Crosslanguage knowledge transfer using multilingual deep
neural network with shared hidden layers,” in Proc.
ICASSP, 2013.
[11] A. Mohan and R. Rose, “Multi-lingual speech recognition with low-rank multi-task deep neural networks,” in
Proc. ICASSP, 2015.
[12] M. Karafiát, K. Veselý, I. Szőke, L. Burget, F. Grézl,
M. Hannemann, and J. Černocký, “BUT ASR system
for BABEL surprise evaluation 2014,” in Proc. SLT,
2014.
[13] Y. Zhang, E. Chuangsuwanich, and J. Glass, “Extracting
deep neural network bottleneck features using low-rank
matrix factorization,” in Proc. ICASSP, 2014.
[14] F. Grézl, M. Karafiát, and M. Janda, “Study of probabilistic and bottle-neck features in multilingual environment,” in Proc. ASRU, 2011.
[15] IARPA broad agency announcement IARPA-BAA-11-02,
2011.
[16] E. Chuangsuwanich, Y. Zhang, and J. Glass, “Multilingual data selection for training stacked bottleneck features,” in Proc. ICASSP, 2016, p. submitted.
[17] V. Le, L. Lamel, A. Messaoudi, W. Hartmann, J. Gauvain, C. Woehrling, J. Despres, and A. Roy, “Developing STT and KWS systems using limited language resources,” in Proc. InterSpeech, 2014.
[18] M.Gales, K.Knill, and A.Ragni,
“Unicode-based
graphemic systems for limited resources languages,” in
Proc. ICASSP, 2015.
[19] M. Gibson and T. Hain, “Hypothesis spaces for minimum bayes risk training in large vocabulary speech
recognition,” in Proc. InterSpeech, 2006.
[20] L. Zhang, D. Karakos, W. Hartmann, R. Hsiao,
R. Schwartz, and S. Tsakalidis, “Enhancing low resource keyword spotting with automatically retrieved
web documents,” in Proc. InterSpeech, 2015.
[21] D. Yu, A. Eversole, M. Seltzer, K. Yao, B. Guenter,
O. Kuchaiev, F. Seide, H. Wang, J. Droppo, Z. Huang,
Y. Zhang, G. Zweig, C. Rossbach, J. Currey, J. Gao,
A. May, A. Stolcke, and M. Slaney, “An introduction to
computational networks and the computational network
toolkit,” Tech. Rep. MSR, Microsoft Research, 2014,
http://cntk.codeplex.com.
[22] R. Williams and J. Peng, “An efficient gradient-based
algorithm for online training of recurrent network trajectories,” Neural Computation, vol. 2, pp. 490–501,
1990.
[23] D. Povey, A. Ghoshal, G. Boulianne, L. Burget,
O. Glembek, N. Goel, M. Hannemann, P. Motlı́ček,
Y. Qian, P. Schwarz, J. Silovský, G. Stemmer, and
K. Veselý, “The Kaldi speech recognition toolkit,” in
Proc. ASRU, 2011.
| 9cs.NE
|
Person Re-Identification by Deep Joint Learning of Multi-Loss Classification
arXiv:1705.04724v2 [cs.CV] 23 May 2017
Wei Li, Xiatian Zhu, Shaogang Gong
Queen Mary University of London, London, UK
{w.li, xiatian.zhu, s.gong}@qmul.ac.uk
Abstract
Existing person re-identification (re-id) methods
rely mostly on either localised or global feature representation alone. This ignores their joint benefit
and mutual complementary effects. In this work,
we show the advantages of jointly learning local
and global features in a Convolutional Neural Network (CNN) by aiming to discover correlated local
and global features in different context. Specifically, we formulate a method for joint learning of
local and global feature selection losses designed
to optimise person re-id when using only generic
matching metrics such as the L2 distance. We design a novel CNN architecture for Jointly Learning
Multi-Loss (JLML) of local and global discriminative feature optimisation subject concurrently to the
same re-id labelled information. Extensive comparative evaluations demonstrate the advantages
of this new JLML model for person re-id over a
wide range of state-of-the-art re-id methods on five
benchmarks (VIPeR, GRID, CUHK01, CUHK03,
Market-1501).
1
Introduction
Person re-identification (re-id) is about matching identity
classes in detected person bounding box images from nonoverlapping camera views over distributed open spaces. This
is an inherently challenging task because person visual appearance may change dramatically in different camera views
from different locations due to unknown changes in human
pose, illumination, occlusion, and background clutter [Gong
et al., 2014]. Existing person re-id studies typically focus
on either feature representation [Gray and Tao, 2008; Farenzena et al., 2010; Kviatkovsky et al., 2013; Zhao et al., 2013;
Liao et al., 2015; Matsukawa et al., 2016a; Ma et al., 2017]
or matching distance metrics [Koestinger et al., 2012; Xiong
et al., 2014; Zheng et al., 2013; Wang et al., 2014b; Paisitkriangkrai et al., 2015; Zhang et al., 2016; Wang et al., 2016b;
Wang et al., 2016c; Wang et al., 2016d; Chen et al., 2017b] or
their combination in deep learning framework [Li et al., 2014;
Ahmed et al., 2015; Wang et al., 2016a; Xiao et al., 2016;
Subramaniam et al., 2016; Chen et al., 2017a]. Regardless, the overall objective is to obtain a view- and location-
invariant (cross-domain) representation. We consider that
learning any matching distance metric is intrinsically learning a global feature transformation across domains (two disjoint camera views) therefore obtaining a “normalised” feature representation for matching.
Most re-id features are typically hand-crafted to encode
local topological and/or spatial structural information, by
different image decomposition schemes such as horizontal
stripes [Gray and Tao, 2008; Kviatkovsky et al., 2013], body
parts [Farenzena et al., 2010], and patches [Zhao et al., 2013;
Matsukawa et al., 2016a; Liao et al., 2015]. These localised features are effective for mitigating the person pose
and detection misalignment in re-id matching. More recent
deep re-id models [Xiao et al., 2016; Wang et al., 2016a;
Chen et al., 2017a; Ahmed et al., 2015] benefit from the
availability of larger scale datasets such as CUHK03 [Li et
al., 2014] and Market-1501 [Zheng et al., 2015] and from
lessons learned on other vision tasks [Krizhevsky et al., 2012;
Girshick et al., 2014]. In contrast to local hand-crafted features, deep models, in particular Convolutional Neural Networks (CNN) [LeCun et al., 1998], favour intrinsically in
learning global feature representations with a few exceptions.
They have been shown to be effective for re-id.
We consider that either local or global feature learning
alone is suboptimal. This is motivated by the human visual system that leverages both global (contextual) and local
(saliency) information concurrently [Navon, 1977; Torralba
et al., 2006]. This intuition for joint learning aims to extract correlated complementary information in different context whilst satisfying the same learning constraint1 therefore
achieving more reliable recognition. To that end, we need to
address a number of non-trivial problems: (i) the model learning behaviour in satisfying the same label constraint may be
different at the local and global levels; (ii) any complementary correlation between local and global features is unknown
and may vary among individual instances, therefore must be
learned and optimised consistently across data; (iii) People’s
appearance in public scenes is diverse in both pattens and
configurations. This makes it challenging to learn correlations between local and global features for all appearances.
This work aims to formulate a deep learning model for
1
In person re-id context, the learning constraint refers to the image person identity label supervision.
jointly optimising local and global feature selections concurrently and to improve person re-id using only generic matching metrics such as the L2 distance. We explore a deep learning approach for its potential superiority in learning from
large scale data [Xiao et al., 2016; Chen et al., 2017a]. For
the bounding box image based person re-id, we consider the
entire person in the bounding box as a global scene context
and body parts of the person as local information sources,
both are subject to the surrounding background clutter within
a bounding box, and potentially also misalignment and partial
occlusion from bounding box detection. In this setting, we
wish to discover and optimise jointly correlated complementary feature selections in the local and global representations,
both subject to the same label constraint concurrently. Whilst
the former aims to address pose/detection misalignment and
occlusion by localised fine-grained saliency information, the
latter exploits holistic coarse-grained context for more robust
global matching.
To that end, we formulate a deep two-branch CNN architecture, with one branch for learning localised feature selection (local branch) and the other for learning global feature selection (global branch). Importantly, the two branches
are not independent but synergistically correlated and jointly
learned concurrently. This is achieved by: (i) imposing interbranch interaction between the local and global branches, and
(ii) enforcing a separate learning objective loss function to
each branch for learning independent discriminative capabilities, whilst being subject to the same class label constraint.
Under such balancing between interaction and independence,
we allow both branches to be learned concurrently for maximising their joint optimal extraction and selection of different discriminative features for person re-id. We call this
model the Joint Learning Multi-Loss (JLML) CNN model.
To minimise poor learning due to inherent noise and potential
covariance, we introduce a structured feature selective and
discriminative learning mechanism into both the local and
global branches subject to a joint sparsity regularisation.
The contributions of this work are: (I) We propose the
idea of learning concurrently both local and global feature
selections for optimising feature discriminative capabilities
in different context whilst performing the same person re-id
tasks. This is currently under-studied in the person re-id literature to our best knowledge. (II) We formulate a novel Joint
Learning Multi-Loss (JLML) CNN model for not only learning both global and local discriminative features in different
context by optimising multiple classification losses on the
same person label information concurrently, but also utilising their complementary advantages jointly in coping with local misalignment and optimising holistic matching criteria for
person re-id. (III) We introduce a structured sparsity based
feature selection learning mechanism for improving multiloss joint feature learning robustness w.r.t. noise and data covariance between local and global representations. Extensive
comparative evaluations demonstrate the superiority of the
proposed JLML model over a wide range of existing state-ofthe-art re-id models on five benchmark datasets VIPeR [Gray
and Tao, 2008], GRID [Loy et al., 2009], CUHK01 [Li et al.,
2012], CUHK03 [Li et al., 2014], and Market-1501 [Zheng
et al., 2015].
2
Related Works
The proposed JLML model considers learning both local
and global feature selections jointly for optimising their correlated complementary advantages. This goes beyond existing methods mostly relying on only one level of feature
representation. Specifically, the JLML method is related
to the saliency learning based models [Zhao et al., 2013;
Wang et al., 2014a] in terms of modelling localised part importance. However, these existing methods consider only the
patch appearance statistics within individual locations but no
global feature representation learning, let alone the correlation and complementary information discovery between local
and global features as modelled by the JLML.
Whilst the more recent Spatially Constrained Similarity
(SCS) model [Chen et al., 2016] and Multi-Channel Parts
(MCP) network [Cheng et al., 2016] consider both levels
of representation, the JLML model differs significantly from
them: (i) The SCS method focuses on supervised metric
learning, whilst the JLML aims at joint discriminative feature learning and needs only generic metrics for re-id matching. Also, hand-crafted local and global features are extracted
separately in SCS without any inter-feature interaction and
correlation learning involved, as opposite to the joint learning of global and local feature selections concurrently subject
to the same supervision information in the JLML; (ii) The local and global branches of the MCP model are supervised and
optimised by a triplet ranking loss, in contrast to the proposed
multiple classification loss design (Sec. 3.2). Critically, this
one-loss model learning is likely to impose negative influence on the discriminative feature learning behaviour for both
branches due to potential over-low pre-branch independence
and over-high inter-branch correlation. This may lead to suboptimal joint learning of local and global feature selections
in model optimisation, as suggested by our evaluation in Section 4.3. (iii) In addition, the JLML is capable of performing structured feature sparsity regularisation along with the
multi-loss joint learning of local and global feature selections
for providing additional benefits (Sec. 4.3). Whilst similar in
theory to the sparsity constraint on the supervised SCS metric
learning, we perform differently sparse generic feature learning without the need for supervised metric optimisation.
In terms of loss function, the HER model [Wang et al.,
2016b] similarly does not exploit pair-wise re-id labels but
defines a single identity label per training person for regression loss (vs. the classification loss in the JLML) based reid feature embedding optimisation. Importantly, HER relies
on the pre-defined feature (mostly hand-crafted local feature)
without the capability of jointly learning global and local feature representations and discovering their correlated complementary advantages as specifically designed in JLML. Also,
the DGD [Xiao et al., 2016] model uses the classification loss
for model optimisation. However, this model considers only
the global feature representation learning of one-loss classification as opposite to the proposed joint global and local feature learning of multi-loss classification concurrently subject
to maximising the same person identity matching.
3
3.1
Model Design
Problem Definition
We assume a set of n training images I = {Ii }ni=1 with
the corresponding identity labels as Y = {yi }ni=1 . These
training images capture the visual appearance of nid (where
yi ∈ [1, · · · , nid ]) different people under non-overlapping
camera views. We formulate a Joint Learning Multi-Loss
(JLML) CNN model that aims to discover and capture concurrently complementary discriminative information about a
person image from both local and global visual features of
the image in order to optimise person re-id under significant
viewing condition changes across locations. This is in contrast to most existing re-id methods typically depending only
on either local or global features alone.
3.2
Joint Learning Multi-Loss
Local Branch
Sparsity
Pooling-256
Shared Conv 1
Local 1
FC-512
FC-IDs
Stripe 1
ID Class
Labels
Local k
Slice
Stripe k
Stripe m
𝑾𝐿
Local m
1,2
Global Branch
Sparsity
Pooling-512 FC-512
Global
𝑾𝐺
FC-IDs
risks. This is especially critical in learning person re-id models when labelled training data is limited.
(II) Multi-task independent learning subject to shared label
constraints. To maximise the learning of complementary discriminative features from local and global representations, the
remaining layers of the two branches are learned independently subject to given identity labels. That is, the JLML
model aims to learn concurrently multiple identity feature
representations for different local image regions and the entire image, all of which aim to maximise the same identity
matching both individually and collectively at the same time.
Independent multi-task learning aims to preserve both local
saliency in feature selection and global robustness in image
representation. To that end, the JLML model is designed
to perform multi-task independent learning subject to shared
identity label constraints by allocating each branch with a
separate objective loss function. By doing so, the per-branch
learning behaviour is conditioned independently on the respective feature representation. We call this branch-wise loss
formulation as the MultiLoss design.
Table 1: JLML-ResNet39. MP: Max-Pooling; AP: AveragePooling; S: Stride; SL: Slice; CA: Concatenation; G: Global; L:
Local.
Layer # Layer Output Size Global Branch
Local Branch
1
conv1
112×112
9
conv2 x
G: 56×56
L: 28×56
9
conv3 x
G: 28×28
L: 14×28
9
conv4 x
G: 14×14
L: 7×14
9
conv5 x
G: 7×7
L: 4×7
1
fc
1×1
ID Class
Labels
2,1
Figure 1: The Joint Learning Multi-Loss (JLML) CNN model architecture.
The overall design of the proposed JLML model is depicted in Figure 1. This JLML model consists of a twobranches CNN network: (1) One local branch of m streams
of an identical structure with each stream learning the most
discriminative local visual features for one of m local image regions of a person bounding box image; (2) Another
global branch responsible for learning the most discriminative global level features from the entire person image.
For concurrently optimising per-branch discriminative feature representations and discovering correlated complementary information between local and global feature selections,
a joint learning scheme that subjects both local and global
branches to the same identity label supervision is considered
with two underlying principles:
(I) Shared low-level features. We construct the global and local branches on a shared lower conv layer, in particular the
first conv layer2 , for facilitating inter-branch common learning. The intuition is that, the lower conv layers capture lowlevel features such as edges and corners which are common
to all patterns in the same images. This shared learning is
similar in spirit to multi-task learning [Argyriou et al., 2007],
where the local and global feature learning branches are two
related learning tasks. Sharing the low-level conv layer reduces the model parameter size therefore model overfitting
2
We found empirically no clear benefits from increasing the
number of shared conv layers in our implementation.
1
fc
1×1
3×3, 32, S-2
3×3 MP, S-2 SL-4, 2×2 MP, S-1
1×1, 32
1×1, 16
3×3, 32×3
3×3, 16×3
1×1, 64
1×1, 32
1×1, 64
1×1, 32
3×3, 64 ×3 3×3, 32×3
1×1, 128
1×1, 64
1×1, 128
1×1, 64
3×3, 128×3 3×3, 64 ×3
1×1, 256
1×1, 128
1×1, 256
1×1, 128
3×3, 256×3 3×3, 128×3
1×1, 512
1×1, 256
7×7
AP
4×7
h
i
h AP, CA-4
i
1×1, 512
1×1, 512
ID#
ID#
Network Construction. We adopt the Residual CNN unit
[He et al., 2016] as the JLML’s building blocks due to its
capacity for deeper model design whilst retaining a smaller
model parameter size3 . Specifically, we customise the
ResNet50 architecture in both layer and filter numbers and
design the JLML model as a 39 layers ResNet (JLMLResNet39) tailored for re-id tasks. The configuration of
JLML-ResNet39 is given in Table 1. Note that, the ReLU
3
The choice of base network is independent of our JLML model
design. Other types, e.g. GoogLeNet [Szegedy et al., 2015] or
VGG-Net [Simonyan and Zisserman, 2015], can be readily applied
in our model.
(1)
i=1
d
where WG = [wg1 , · · · , wg g ] ∈ R cg ×dg is the parameter
matrix of the global branch feature layer taking as input dg
dimensional vectors from the previous layer and outputting
cg dimensional (512-D) feature representation. Specifically,
with the `1 norm applied on the `2 norm of wgi , our aim is to
learn (tune) selectively feature dimension importance subject
to both the sparsity principle and the identity label constraint
simultaneously.
Similarly, we also enforce a local feature sparsity constraint by an exclusive group LASSO [Kong et al., 2014]:
`1,2 = kWL k1,2 =
cl X
m
X
i
kwl,j
k21
(2)
i=1 j=1
1>
wl,1
WL = · · ·
cl >
wl,1
···
···
···
1>
1
. . . 𝑤𝑙111,𝑑
𝑤𝑙1
1,1
𝑙
𝑐𝑔
𝑤𝑔
1,1
...
𝑐𝑔
𝑤𝑔
𝑤𝑙1𝑙 1,1. . . 𝑤𝑙1𝑙 1,𝑑
𝑐
1,𝑑𝑔
2,1
= 𝒘𝑔𝑖
i=1
1
. . .𝑤𝑙𝑘11,𝑑
𝑤𝑙𝑘
1,1
𝑐
...
𝑤𝑙𝑘𝑙 1,1. . . 𝑤𝑙𝑘𝑙 1,𝑑
𝑙
𝑙
𝟐
Small
𝑙
𝑐𝑙
Large
...
𝑐
𝑐
Weights
𝑑𝑔
𝑾𝐺
...
...
1
1
. . . 𝑤𝑙𝑚
𝑤𝑙𝑚
1,𝑑
1,1
𝑙
𝑾𝐿
1,2
𝑤𝑙𝑚𝑙 1,1. . . 𝑤𝑙𝑚𝑙 1,𝑑
𝑐
𝑐
𝑙
𝑚
= 𝒘𝑖𝑙𝑗
i=1 j=1
2
1
Figure 2: Group sparsity regularisations on fc layer parameter matrices (WG for the global branch and WL for the local branch) for
selectively learning feature representations. Solid and dashed rectangles denote `2 norm and `1 norm respectively.
probability ỹi of image Ii over the given identity label yi :
exp(wy>i xi )
p(ỹi = yi |Ii ) = P|n |
id
>
k=1 exp(wk xi )
(4)
where xi refers to the feature vector of Ii from the corresponding branch, and Wk the prediction function parameter
of training identity class k. The training loss on a batch of nbs
images is computed as:
nbs
1 X
l=−
log p(ỹi = yi |Ii )
nbs i=1
(5)
Combined with the group sparsity based feature selection regularisations, we have the final loss function for the global and
local branch sub-networks as:
lglobal = l + λglobal kWG k2,1 , llocal = l + λlocal kWL k1,2 (6)
where
1,𝑑𝑔
...
kwgi k2
𝑤𝑔1
...
dg
X
...
1,1
...
`2,1 = kWG k2,1 =
𝑤𝑔1
...
rectification non-linearity [Krizhevsky et al., 2012] after each
conv layer is omitted for brevity.
Feature Selection. To optimise JLML model learning robustness against noise and diverse data source, we introduce
a feature selection capability in JLML by a structure sparsity induced regularization [Kong et al., 2014; Wang et al.,
2013]. Our idea is to have a competing-to-survive mechanism in feature learning that discourages irrelevant features
whilst encourages discriminative features concurrently in different local and global context to maximise a shared identity
matching objective. To that end, we sparsify the global feature representation with a group LASSO [Wang et al., 2013]:
1>
wl,m
wl
··· = ···
cl >
wl,m
wlcl >
(3)
is the parameter matrix of the local branch feature layer with
m×dl and cl (512) as the input and output dimensions (m the
i
image stripe number). The wl,j
∈ R dl ×1 defines the parameter vector for contributing the i-th output feature dimension
from the j-th local input feature vector, j ∈ [1, 2, · · · , m]. In
particular, the `2,1 regulariser performs sparse feature selection for individual image regions as below: (1) We perform
feature selective learning at the local region level by enforci
ing the `1 norm directly on wl,j
, conceptually similar to the
group LASSO at the global level. (2) We then apply a nonsparse smooth fusion with the `2 norm to combine the efi
fects of different local features weighted by the sparse wl,j
.
(3) Lastly, we exploit the `1 norm again at the level of wlk
(k ∈ [1, 2, · · · , cl ]) to learn the local 512-D feature representation selection. Figure 2 shows our structured sparsity
regularisations for both local and global feature selections.
Loss Function. For model training, we utilise the crossentropy classification loss function for both global and local
branches so to optimise person identity classification given
training labels of multiple person classes extracted from pairwise labelled re-id dataset. Formally, we predict the posterior
where λglobal and λlocal control the balance between the identity label loss and the feature selection sparsity regularisation. We empirically set λlocal = λglobal = 5×10−4 by crossvalidation in our evaluations.
Choice of Loss Function. Our JLML model learning deploys
a classification loss function. This differs significantly from
the contrastive loss functions used by most existing deep
re-id methods designed to exploit pairwise re-id labels defined by both positive and negative pairs, such as the pairwise
verification [Varior et al., 2016; Subramaniam et al., 2016;
Ahmed et al., 2015; Li et al., 2014], triplet ranking [Cheng et
al., 2016], or both [Wang et al., 2016a; Chen et al., 2017a].
Our JLML model training does not use any labelled negative pairs inherent to all person re-id training data, and we
extract identity class labels from only positive pairs. The
motivations for our JLML classification loss based learning
are: (i) Significantly simplified training data batch construction, e.g. random sampling with no notorious tricks required,
as shown by other deep classification methods [Krizhevsky
et al., 2012]. This makes our JLML model more scalable
in real-world applications with very large training population
sizes when available. This also eliminates the undesirable
need for carefully forming pairs and/or triplets in preparing
re-id training splits, as in most existing methods, due to the
inherent imbalanced negative and positive pair size distributions. (ii) Visual psychophysical findings suggest that representations optimised for classification tasks generalise well
to novel categories [Edelman, 1998]. We consider that reid tasks are about model generalisation to unseen test identity classes given training data on independent seen identity
classes. Our JLML model learning exploits this general classification learning principle beyond the strict pair-wise relative verification loss in existing re-id models.
3.3
Model Training
We adopt the standard Stochastic Gradient Descent (SGD)
optimisation algorithm [Krizhevsky et al., 2012] to perform
the batch-wise joint learning of local and global branches.
Note that, with SGD we can naturally synchronise the optimisation processes of the two branches by constraining their
learning behaviours subject to the same identity label information at each update. This is likely to avoid representation
learning divergence between two branches and help enhance
the correlated complementary learning capability.
3.4
Re-Id by Generic Distance Metrics
Once the JLML model is learned, we obtain a 1,024-D joint
representation by concatenating the local (512-D) and global
(512-D) feature vectors (the fc layer in Table 1). For person re-id, we deploy this 1,024-D deep feature representation
using only a generic distance metric without camera-pair specific distance metric learning, e.g. L2 distance. Specifically,
given a test probe image I p from one camera view and a
set of test gallery images {Iig } from other non-overlapping
camera views: (1) We first compute their corresponding
1,024-D feature vectors by forward-feeding the images to
the trained JLML model, denoted as xp = [xpg ; xpl ] and
{xgi = [xgg ; xgl ]}. (2) We then compute L2 normalisation on
the global and local features, separately. (3) Lastly we compute the cross-camera matching distances between xp and xgi
by some generic matching metric, e.g. L2 distance. We then
rank all gallery images in ascendant order by their L2 distances to the probe image. The probabilities of true matches
of probe person images in Rank-1 and among the higher ranks
indicate the goodness of the learned JLML deep features for
person re-id tasks.
4
numbers of images per person per view captured from a supermarket, with all bounding boxes automatically detected.
These datasets present a wide range of re-id evaluation scenarios with different population sizes under different challenging viewing conditions (Table 2).
Experiments
Datasets. For evaluation, we used five benchmarking re-id
datasets, VIPeR [Gray and Tao, 2008], GRID [Loy et al.,
2009], CUHK01 [Li et al., 2012], CUHK03 [Li et al., 2014],
and Market-1501 [Zheng et al., 2015]. Figure 3 shows some
examples of person bounding box images from these datasets.
The datasets are collected by different data sampling protocols from different environments, where: (a) VIPeR has
one image per person per view, with low-resolution under severe lighting change. (b) GRID provides one image per person per view, with additional images for 775 distracting persons under very poor lighting from underground stations. (c)
CUHK01 contains two images person per view from a university campus. (d) CUHK03 consists of up to five images
per person per view, obtained by both manually labelled and
auto-detected person bounding boxes with the latter posing
a more challenging re-id task due to detection bounding box
misalignment and occlusion. (e) Market-1501 has variable
(a)
VIPeR
(b)
GRID
(c)
CUHK01
(d)
CUHK03
(e)
Market
Figure 3: Example cross-view image pairs from five re-id datasets.
Table 2: Settings of person re-id datasets. TS: Test Setting; SS:
Single-Shot; MS: Multi-Shot. SQ: Single-Query; MQ: Multi-Query.
Dataset Cams IDs Train IDs Test IDs Labelled Detected TS
VIPeR
2 632
316
316
1,264
0
SS
GRID
8 250
125
125
1,275
0
SS
CUHK01 2 971 871/485 100/486 1,942
0
SS/MS
CUHK03 6 1,467 1,367
100 14,097 14,097
SS
Market
6 1,501 751
750
0
32,668 SQ/MQ
Evaluation Protocol. We adopted the standard supervised
re-id setting to evaluate the proposed JLML model (Sec. 4.1).
The training and test data splits and testing settings of each
dataset is given in Table 2. Specifically, on VIPeR, we split
randomly the whole population (632 people) into two halves:
One for training (316) and another for testing (316). We repeated 10 trials of random people splits and used the averaged
results. On CUHK01, we considered two training/test splits:
485/486 [Liao et al., 2015] and 871/100 [Ahmed et al., 2015].
Again, we reported the results averaged over 10 random trials
for either split. On GRID, the training/test split were 125/125
with 775 distractor people included in the test gallery. We
used the benchmarking 10 people splits [Loy et al., 2009]
and the averaged performance. On CUHK03, following [Li
et al., 2014] we repeated 20 times of random 1260/100 training/test splits and reported the averaged accuracies under the
single-shot evaluation setting. On Market-1501, we used the
standard training/test split (750/751) [Zheng et al., 2015]. We
used the cumulative matching characteristic (CMC) to measure re-id accuracy on all benchmarks, except on Market1501 we also used in addition the recall measure of multiple
truth matches by mean Average Precision (mAP), i.e. first
computing the area under the Precision-Recall curve for each
probe, then calculating the mean of Average Precision over
all probes [Zheng et al., 2015].
Competitors. We compared the JLML model against 10 existing state-of-the-art methods as listed in Table 3. They range
from hand-crafted and deep learning features to domainspecific distance metric learning methods. We summarise
them into three categories: (A) Hand-crafted (feature) with
domain-specific distance learning (metric); (B) Deep learning
(feature) with domain-specific deep verification metric learning; (C) Deep learning (feature) with generic non-learning L2
distance (metric).
Implementation. We used the Caffe framework [Jia et al.,
2014] for our JLML model implementation. We started by
Table 3: Person re-id method categorisation by features and metrics. Cat: Category; DL: Deep Learning; CPSL: Camera-Pair Specific Learning; DVM: Deep Verification Metric; DVM,L2: Ensemble of DVM and L2; CHS: Fusion of Colour, HOG, SILPT features.
Cat
A
B
C
Feature
Metric
Method
Hand-Crafted
DL
CPSL Generic
XQDA [Liao et al., 2015]
LOMO
XQDA
GOG [Matsukawa et al., 2016b]
GOG
XQDA
NFST [Zhang et al., 2016]
LOMO, KCCA
NSFT
SCS [Chen et al., 2016]
CHS
SCS
DCNN+ [Ahmed et al., 2015]
DCNN+ DVM
X-Corr [Subramaniam et al., 2016]
X-Corr DVM
MTDnet [Chen et al., 2017a]
MTDnet DVM, L2
S-CNN [Varior et al., 2016]
S-CNN
L2
DGD [Xiao et al., 2016]
DGD
L2
MCP [Cheng et al., 2016]
MCP
L2
JLML (Ours)
JLML
L2
pre-training the JLML model on ImageNet (ILSVRC2012).
Subsequently, for CUHK03 or Market, we used only their
own training data for model fine-tuning, i.e. ImageNet →
CUHK03/Market; For CUHK01 or VIPeR or GRID, we pretrained JLML on CUHK03+Market (whole datasets), and
then fine-tuned on their respective training images, i.e. ImageNet → CUHK03+Market → CUHK01 / VIPeR / GRID.
All input person images were resized to 224 × 224 in pixel.
For local branch, according to a coarse body part layout we
evenly decomposed the whole shared convolutional feature
maps (i.e. the entire image) into four (m = 4) horizontal
strip-regions. We used the same parameter settings (summarised in Table 4) for pre-training and training the JLML
model on all datasets. We also adopted the stepped learning
rate policy, e.g. dropping the learning rate by a factor of 10
every 100K iterations for JLML pre-training and every 20K
iterations for JLML training. We utilised the L2 distance as
the default matching metric, unless stated otherwise.
Table 4: JLML training parameters. BLR: base learning rate; LRP:
learning rate policy; MOT: momentum; IT: iteration; BS: batch size.
Parameter BLR
LRP
MOT
IT #
BS
Pre-train
0.01 step (0.1, 100K)
0.9
300K 32
Train
0.01
step (0.1, 20K)
0.9
50K
32
4.1
Conventional Intra-Domain Re-Id Evaluations
We conducted extensively comparative evaluations on conventional supervised learning based person re-id tasks.
(I) Evaluation on CUHK03. Table 5 shows the comparisons of JLML against 8 existing methods on CUHK03. It is
evident that JLML outperforms existing methods in all categories on both labelled and detected bounding boxes, surpassing the 2nd best performers DGD and X-Corr on corresponding labelled and detected images in Rank-1 by 7.9%(83.275.3) and 8.6%(80.6-72.0) respectively. X-Corr/GOG/JLML
also suffer the least from auto-detection misalignment, indicating the robustness and competitiveness of the joint learning approach to mining complementary local and global discriminative features.
(II) Evaluation on Market-1501.
We evaluated JLML
against four existing models on Market-1501. Table 6 shows
the clear performance superiority of JLML over all state-ofthe-arts with more significant Rank-1 advantages over other
methods compared to CUHK03, giving 19.3%(85.1-65.8)
Table 5: CUHK03 evaluation. 1st /2nd best in red/blue.
Cat
A
B
C
Annotation
Rank (%)
XQDA
GOG
NSFT
DCNN+
X-Corr
MTDnet
S-CNN
DGD
JLML
R1
55.2
67.3
62.5
54.7
72.4
74.7
75.3
83.2
Labelled
R5 R10
77.1 86.8
91.0 96.0
90.0 94.8
86.5 93.9
95.5 96.0 97.5
98.0 99.4
R20
83.1
98.1
98.1
98.4
99.8
R1
46.3
65.5
54.7
44.9
72.0
68.1
80.6
Detected
R5 R10
78.9 83.5
88.4 93.7
84.7 94.8
76.0 83.5
96.0 88.1 94.6
96.9 98.7
R20
93.2
95.2
93.2
98.2
99.2
(SQ) and 13.7%(89.7-76.0) (MQ) gains over the 2nd best SCNN. This further validates the advantages of our joint learning of multi-loss classification for optimising re-id especially
when the re-id test population size increases (751 people on
Market-1501 vs. 100 people on CUHK03).
Table 6: Market-1501 evaluation. 1st /2nd best in red/blue. All person bounding box images were auto-detected.
Query Type
Single-Query
Cat
Measure (%)
R1
mAP
XQDA
43.8
22.2
A
SCS
51.9
26.3
NFST
61.0
35.6
S-CNN
65.8
39.5
C
JLML
85.1
65.5
Multi-Query
R1
mAP
54.1
28.4
71.5
46.0
76.0
48.4
89.7
74.5
(III) Evaluation on CUHK01. We compared our JLML
model with 8 state-of-the-art methods on CUHK01. Table
7 shows that JLML surpasses clearly all compared models
under both training/test splits in single- and multi-short settings. Moreover, JLML outperforms in Rank-1 (76.7%) the
best hand-crafted feature method NFST (R1 69.1%) when
the training population size is small (486 people). When the
training population size increases (871 people), JLML is even
more effective than all deep competitors in exploiting extra training classes by inducing more identity-discriminative
joint person features in distinct context. For example, JLML
gains 5.8%(87.0-81.2) more Rank-1 than the 2nd best method
X-Corr in single-shot re-id, further improved the gain of
4.8%(69.8-65.0) under the 486/485 split. These results show
consistent superiority and robustness of the proposed JLML
model over the existing methods.
(IV) Evaluation on VIPeR. We evaluated the performance
of JLML against 8 strong competitors on VIPeR, a more challenging test scenario with fewer training classes (316 people)
and lower image resolution. On this dataset, the best performers are hand-crafted feature methods (SCS and NFST)
rather than deep models. This is in contrast to the tests on
CUHK01, CUHK03, and Market-1501. This is due to (i)
the small training data insufficient for learning effectively discriminative deep models with millions of parameters; (ii) the
greater disparity to CUHK03 in camera viewing conditions
which makes knowledge transfer less effective (see Implementation). Nevertheless, the JLML model remains the best
among all deep methods with or without deep verification
metric learning. This validates the superiority and robustness
of our deep joint global and local representation learning of
Table 7: CUHK01 evaluation. 1st /2nd best in bold/typewriter.
Split
871/100 split
486/485 split
Cat
Rank (%) R1 R5 R10 R20 R1 R5 R10 R20
Single-Shot Testing Setting
A
GOG
- 57.8 79.1 86.2 92.1
DCNN+ 65.0 - 47.5 71.6 80.3 87.5
B
X-Corr 81.2 97.3 - 98.6 65.0 89.7 - 94.4
MTDnet 78.5 96.5 97.5 DGD
- 66.6 C
MCP
- 53.7 84.3 91.0 96.3
JLML 87.0 97.2 98.6 99.4 69.8 88.4 93.3 96.3
Multi-Shot Testing Setting
XQDA
- 63.2 83.9 90.0 94.2
A
GOG
- 67.3 86.9 91.8 95.9
NFST
- 69.1 86.9 91.8 95.4
C
JLML 91.2 98.4 99.2 99.8 76.7 92.6 95.6 98.1
multi-loss classification given sparse training data. We attribute this property to the JLML’s capability of mining complementary features in different context for both handling local misalignment and optimising global matching.
Table 8: VIPeR evaluation. 1st /2nd best in red/blue.
Cat
A
B
C
Rank (%)
XQDA
GOG
NFST
SCS
DCNN+
MTDnet
MCP
DGD
JLML
R1
40.0
49.7
51.1
53.5
34.8
47.5
47.8
38.6
50.2
R5
68.1
82.1
82.6
63.6
73.1
74.7
74.2
R10
80.5
88.7
90.5
91.5
75.6
82.6
84.8
84.3
R20
91.1
94.5
95.9
96.7
84.5
91.1
91.6
(V) Evaluation on GRID. We compared JLML against 4
competing methods on GRID4 . In addition to poor image
resolution, poor lighting and a small training size (125 people), GRID also has extra distractors in the testing population
therefore presenting a very challenging but realistic re-id scenario. Table 9 shows a significant superiority of JLML over
existing state-of-the-arts, with Rank-1 12.8%(37.5-24.7) better than the 2nd best method GOG, a 51.8% relative improvement. This demonstrates the unique and practically desirable advantage of JLML in handling more realistically challenging open-world re-id matching where large numbers of
distractors are usually present. It is worth pointing out that
this step-change advantage in re-id matching rate on GRID
is achieved by deep learning from only a limited number of
training identity classes with highly imbalanced images sampled from 8 distributed camera views, e.g. 25 images from
the 6th camera vs. 513 from the 5th camera. This imbalanced
sampling directly results in not only scarce pairwise training
4
The GRID dataset has not been evaluated as extensively
as others like VIPeR / CUHK01 / CUHK03, although GRID
provides a more realistic test setting with a large number of distractors in testing.
One possible reason is the
more challenging re-id setting imposed by GRID resulting
in significantly poorer matching rates by all published methods (see http://personal.ie.cuhk.edu.hk/˜ccloy/
downloads_qmul_underground_reid.html), also as verified by our evaluation in Table 9.
data but also insufficient training samples for pairwise camera views, resulting in significant degradation in re-id performance from all pairwise supervised learning based models
XQDA, GOG, SCS, and X-Corr. In contrast, JLML is designed to avoid the need for pairwise labelled information in
model learning by instead learning from multi-loss classifications. Moreover, the joint learning of multi-loss classification
benefits from concurrent local and global feature selections in
different context, resulting in more robust and accurate re-id
matching in a heterogeneous search space.
Table 9: GRID evaluation. 1st /2nd best in red/blue.
Cat
A
B
C
4.2
Rank (%)
XQDA
GOG
SCS
X-Corr
JLML
R1
16.6
24.7
24.2
19.2
37.5
R5
33.8
47.0
44.6
38.4
61.4
R10
41.8
58.4
54.1
53.6
69.4
R20
52.4
69.0
65.2
66.4
77.4
CNN Architecture Comparisons
We compared the proposed JLML-ResNet39 model with
four seminal classification CNN architectures (Alexnet
[Krizhevsky et al., 2012], VGG16 [Simonyan and Zisserman,
2015], GoogLeNet [Szegedy et al., 2015], and ResNet50 [He
et al., 2016]) in model size and complexity. Table 10 shows
that the JLML has both the 2nd smallest model size (7.2 million parameters) and the 2nd smallest FLOPs (1.54×109 ), although containing more streams (5 vs. 1 in all other CNNs)
and more layers (39, more than all except ResNet50).
Table 10: Comparisons of model size and complexity. FLOPs: the
number of FLoating-point OPerations; PN: Parameter Number.
Model
FLOPs
PN (million) Depth Stream #
AlexNet
7.25×108
58.3
7
1
VGG16
1.55×1010
134.2
16
1
ResNet50
3.80×109
23.5
50
1
GoogLeNet
1.57×109
6.0
22
1
JLML-ResNet39 1.54×109
7.2
39
5
4.3
Further Analysis and Discussions
We further examined the component effects of our JLML
model on Market-1501 in the following aspects.
(I) Complementary Benefits of Global and Local Features. We evaluated the complementary effects of our jointly
learned local and global features by comparing their individual re-id performance against that of the joint features. Table
11 shows: (i) Any of the two feature representations alone
is competitive for re-id, e.g. the local JLML feature surpasses S-CNN (Table 6) by Rank-1 13.1%(78.9-65.8) (SQ)
and 10.4%(86.4-76.0) (MQ); and by mAP 18.3%(57.8-39.5)
(SQ) and 20.0%(68.4-48.4) (MQ). (ii) A further performance
gain is obtained from the joint feature representation, yielding
further 6.2%(85.1-78.9) (SQ) and 3.3%(89.7-86.4) (MQ) in
Rank-1 increase, and 7.7%(65.5-57.8) (SQ) and 6.1%(74.568.4) (MQ) in mAP boost. These results show the complementary advantages of jointly learning the local and global
features in different context using the JLML model.
Table 11: Complementary benefits of global and local features.
Query Type
Measure (%)
JLML (Global)
JLML (Local)
JLML (joint)
Single-Query
R1
mAP
77.4
56.0
78.9
57.8
85.1
65.5
Multi-Query
R1
mAP
85.0
66.0
86.4
68.4
89.7
74.5
(II) Importance of Branch Independence. We evaluated the
importance of branch independence by comparing our MultiLoss design with a UniLoss design that merges two branches
into a single loss [Cheng et al., 2016]. Table 12 shows that
the proposed MultiLoss model significantly improves the discriminative power of global and local re-id features, e.g. with
Rank-1 increase of 9.0%(85.1-76.1) (SQ) and 6.0%(89.783.7) (MQ); and mAP improvement of 13.3%(65.5-52.2)
(SQ) and 11.7%(74.5-62.8) (MQ). This shows that branch independence plays a critical role in joint learning of multi-loss
classification for effective feature optimisation. One plausible
reason is due to the negative effect of a single loss imposed on
the learning behaviour of both branches, caused by the potential divergence in discriminative features in different context
(local and global). This is shown by the significant performance degradation of both global and local features when the
UniLoss model is imposed.
Table 12: Importance of branch independence.
Loss
UniLoss
MultiLoss
Query Type
Measure (%)
Global Feature
Local Feature
Full
Global Feature
Local Feature
Full
Single-Query
R1
mAP
58.3
31.7
46.3
26.3
76.1
52.2
77.4
56.0
78.9
57.8
85.1
65.5
Multi-Query
R1
mAP
70.4 43.2
58.0 34.0
83.7 62.8
85.0 66.0
86.4 68.4
89.7 74.5
(III) Benefits from Shared Low-Level Features. We evaluated the effects of interaction between global and local
branches introduced by the shared conv layer (common
ground) by deliberately removing it and then comparing the
re-id performance. Table 13 shows the benefits from jointly
learning low-level features in the common conv layers, e.g.
improving Rank-1 by 1.9%(85.1-83.2) / 1.4%(89.7-88.3) and
mAP by 2.4%(65.5-63.1) / 2.4%(74.5-72.1) for single-/multiquery re-id. This confirms a similar finding as in multi-task
learning study [Argyriou et al., 2007].
Table 13: Benefits from shared low-level features.
Query Type
Measure (%)
Without Shared Feature
With Shared Feature
Single-Query
R1
mAP
83.2
63.1
85.1
65.5
Multi-Query
R1
mAP
88.3
72.1
89.7
74.5
(IV) Effects of Selective Feature Learning. We evaluated
the contribution of our structured sparsity based Selective
Feature Learning (SFL) (Eq. (6)). Table 14 shows that our
SFL mechanism can bring additional re-id matching benefits, e.g. improving Rank-1 rate by 1.7%(85.1-83.4) (SQ) and
1.0%(89.7-88.7) (MQ); and mAP by 1.7%(65.5-63.8) (SQ)
and 1.6%(74.5-72.9) (MQ).
(V) Choice of Generic Matching Metrics. We evaluated the
choice of generic matching distances on person re-id using
Table 14: Effects of selective feature learning (SFL).
Query Type
Single-Query
Multi-Query
Measure (%)
R1
mAP
R1
mAP
Without SFL
83.4
63.8
88.7
72.9
With SFL
85.1
65.5
89.7
74.5
the full JLML feature. Table 15 shows that L1 and L2 generate very similar and competitive re-id matching accuracies.
This suggests the flexibility of the JLML model in adopting
generic matching metrics.
Table 15: Effects of generic matching metrics.
Query-Type
Measure (%)
L1
L2
Single-Query
R1
mAP
84.9
65.3
85.1
65.5
Multi-Query
R1
mAP
89.2
74.6
89.7
74.5
(VI) Effects of Body Parts Number. We evaluated the sensitivity of local decomposition, i.e. body parts number m. Table 16 shows that the decomposition of 4 body-parts is the optimal choice, approximately corresponding to head+shoulder,
upper-body, upper-leg and lower-leg (Figure 4).
Table 16: Effects of body parts number.
Query-Type
Measure (%)
2
4
6
8
10
Single-Query
R1
mAP
83.9
64.4
85.1
65.5
83.4
62.6
82.3
61.3
81.7
60.4
Multi-Query
R1
mAP
88.8
72.9
89.7
74.5
88.5
71.8
87.4
70.7
87.2
69.8
(VII) Complementary Effects between JLML Deep Features and Supervised Metric Learning.
We evaluated
the complementary effects of the JLML deep features and
conventional supervised metric learning (XQDA [Liao et al.,
2015], KISSME [Koestinger et al., 2012], and CRAFT [Chen
et al., 2017b]). Results from Table 17 show that: (1) Given
strong deep learning features such as JLML, additional distance metric learning does not benefit further from the same
training data. (2) Moreover, it may even suffer from some
adversary effect.
(VIII) Local Features vs. Global Features. A strength of
the local features is the capability of mitigating misalignment
and occlusion, as compared to the global features. This is
inherently learned from data by the JLML local branch. Figure 5 shows the single-query re-id results on six randomly
selected probe persons with misalignment and/or occlusion.
It is evident that the local features achieve better re-id matching ranks than the global counterparts in most cases. This
clearly demonstrates the robustness of local features against
the misalignment of and occlusion within a person bounding
box.
(IX) Feature Extraction Time Cost. The average time for
extracting JLML feature is 2.75 milliseconds per image (364
images per second) on a Nvidia Pascal P100 GPU card.
Local features:
1,2,3,8,9
Local features:
1,6,8,11,15
Local features:
1,5,34,38,39
Global features:
1,2,8,10,24
Global features:
1,5,34,38,39
Global features:
1,7,76,81,88
Local features:
2,4,7,21,57,99
Local features:
1,32,122,409,460
Local features:
1,2,7,10,13,14
Global features:
1,9,124,173,212
Global features:
1,144,163,948,960
Global features:
1,11,15,16,17
Figure 4: Visualisation of the optimal body part decomposition.
Table 17: Complementary of JLML features and metric learning.
Query-Type
Measure (%)
KISSME
XQDA
CRAFT
L2
5
Single-Query
R1
mAP
82.1
61.4
82.6
63.2
77.9
56.4
85.1
65.5
Multi-Query
R1
mAP
87.5
70.2
88.2
72.4
89.7
74.5
Conclusion
In this work, we presented a novel Joint Learning of MultiLoss (JLML) CNN model (JLML-ResNet39) for person reidentification feature learning. In contrast to existing re-id
approaches that often employ either global or local appearance features alone, the proposed model is capable of extracting and exploiting both and maximising their correlated complementary effects by learning discriminative feature representations in different context subject to multi-loss classification objectives in a unified framework. This is made possible
by the proposed JLML-ResNet39 architecture design. Moreover, we introduce a structured sparsity based feature selective learning mechanism to reduce feature redundancy and
further improve the joint feature selections. Extensive comparative evaluations on five re-id benchmark datasets were
conducted to validate the advantages of the proposed JLML
model over a wide range of the state-of-the-art methods on
both manually labelled and more challenging auto-detected
person images. We also provided component evaluations and
analysis of model performance in order to give insights on the
model design.
References
[Ahmed et al., 2015] Ejaz Ahmed, Michael Jones, and Tim K
Marks. An improved deep learning architecture for person reidentification. In CVPR, 2015.
[Argyriou et al., 2007] Andreas Argyriou, Theodoros Evgeniou,
and Massimiliano Pontil. Multi-task feature learning. In NIPS,
2007.
[Chen et al., 2016] Dapeng Chen, Zejian Yuan, Badong Chen, and
Nanning Zheng. Similarity learning with spatial constraints for
person re-identification. In CVPR, 2016.
[Chen et al., 2017a] Weihua Chen, Xiaotang Chen, Jianguo Zhang,
and Kaiqi Huang. A multi-task deep network for person reidentification. In AAAI, 2017.
[Chen et al., 2017b] Ying-Cong Chen, Xiatian Zhu, Wei-Shi
Zheng, and Jian-Huang Lai. Person re-identification by camera
correlation aware feature augmentation. IEEE TPAMI, 2017.
[Cheng et al., 2016] De Cheng, Yihong Gong, Sanping Zhou, Jinjun Wang, and Nanning Zheng. Person re-identification by multichannel parts-based cnn with improved triplet loss function. In
CVPR, 2016.
Figure 5: Comparing the gallery true match ranks of each probe
image (single-query) with occlusion and/or misalignment by the local and global features. Each probe may have multiple truth matches
in the gallery. Smaller numbers mean better ranking performances.
[Edelman, 1998] Shimon Edelman. Representation is representation of similarities. Behavioral and Brain Sciences, 21(04):449–
467, 1998.
[Farenzena et al., 2010] Michela Farenzena, Loris Bazzani,
Alessandro Perina, Vittorio Murino, and Marco Cristani. Person
re-identification by symmetry-driven accumulation of local
features. In CVPR, 2010.
[Girshick et al., 2014] Ross Girshick, Jeff Donahue, Trevor Darrell,
and Jitendra Malik. Rich feature hierarchies for accurate object
detection and semantic segmentation. In CVPR, 2014.
[Gong et al., 2014] Shaogang Gong, Marco Cristani, Shuicheng
Yan, and Chen Change Loy. Person re-identification. Springer,
January 2014.
[Gray and Tao, 2008] Douglas Gray and Hai Tao. Viewpoint invariant pedestrian recognition with an ensemble of localized features.
In ECCV, 2008.
[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and
Jian Sun. Deep residual learning for image recognition. In CVPR,
2016.
[Jia et al., 2014] Yangqing Jia, Evan Shelhamer, Jeff Donahue,
Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for
fast feature embedding. In ACM MM, 2014.
[Koestinger et al., 2012] Martin Koestinger, Martin Hirzer, Paul
Wohlhart, Peter M. Roth, and Horst Bischof. Large scale metric learning from equivalence constraints. In CVPR, 2012.
[Kong et al., 2014] Deguang Kong, Ryohei Fujimaki, Ji Liu, Feiping Nie, and Chris Ding. Exclusive feature learning on arbitrary
structures via l1,2-norm. In NIPS, 2014.
[Krizhevsky et al., 2012] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional
neural networks. In NIPS, 2012.
[Kviatkovsky et al., 2013] Igor Kviatkovsky, Amit Adam, and
Ehud Rivlin. Color invariants for person reidentification. IEEE
TPAMI, 35(7):1622–1634, 2013.
[LeCun et al., 1998] Yann LeCun, Léon Bottou, Yoshua Bengio,
and Patrick Haffner. Gradient-based learning applied to document recognition. IEEE, 86(11):2278–2324, 1998.
[Li et al., 2012] Wei Li, Rui Zhao, and Xiaogang Wang. Human
reidentification with transferred metric learning. In ACCV, 2012.
[Li et al., 2014] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang.
Deepreid: Deep filter pairing neural network for person reidentification. In CVPR, 2014.
[Liao et al., 2015] Shengcai Liao, Yang Hu, Xiangyu Zhu, and
Stan Z Li. Person re-identification by local maximal occurrence
representation and metric learning. In CVPR, 2015.
[Loy et al., 2009] Chen Change Loy, Tao Xiang, and Shaogang
Gong. Multi-camera activity correlation analysis. In CVPR,
2009.
[Ma et al., 2017] Xiaolong Ma, Xiatian Zhu, Shaogang Gong,
Xudong Xie, Jianming Hu, Kin-Man Lam, and Yisheng Zhong.
Person re-identification by unsupervised video matching. Pattern
Recognition, 65:197–210, 2017.
[Matsukawa et al., 2016a] Tetsu Matsukawa, Takahiro Okabe,
Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, 2016.
[Matsukawa et al., 2016b] Tetsu Matsukawa, Takahiro Okabe,
Einoshin Suzuki, and Yoichi Sato. Hierarchical gaussian descriptor for person re-identification. In CVPR, 2016.
[Navon, 1977] David Navon. Forest before trees: The precedence
of global features in visual perception. Cognitive Psychology,
9(3):353–383, 1977.
[Paisitkriangkrai et al., 2015] Sakrapee Paisitkriangkrai, Chunhua
Shen, and Anton van den Hengel. Learning to rank in person
re-identification with metric ensembles. In CVPR, 2015.
[Simonyan and Zisserman, 2015] Karen Simonyan and Andrew
Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015.
[Subramaniam et al., 2016] Arulkumar Subramaniam, Moitreya
Chatterjee, and Anurag Mittal. Deep neural networks with inexact matching for person re-identification. In NIPS, 2016.
[Szegedy et al., 2015] Christian Szegedy, Wei Liu, Yangqing Jia,
Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper
with convolutions. In CVPR, 2015.
[Torralba et al., 2006] Antonio Torralba, Aude Oliva, Monica S
Castelhano, and John M Henderson. Contextual guidance of eye
movements and attention in real-world scenes: the role of global
features in object search. Psychological Review, 113(4):766,
2006.
[Varior et al., 2016] Rahul Rama Varior, Mrinal Haloi, and Gang
Wang. Gated siamese convolutional neural network architecture
for human re-identification. In ECCV, 2016.
[Wang et al., 2013] Hua Wang, Feiping Nie, and Heng Huang.
Multi-view clustering and feature learning via structured sparsity.
In ICML, 2013.
[Wang et al., 2014a] H. Wang, S. Gong, and T. Xiang. Unsupervised learning of generative topic saliency for person reidentification. In British Machine Vision Conference, Nottingham, UK, September 2014.
[Wang et al., 2014b] Taiqing Wang, Shaogang Gong, Xiatian Zhu,
and Shengjin Wang. Person re-identification by video ranking. In
ECCV, 2014.
[Wang et al., 2016a] Faqiang Wang, Wangmeng Zuo, Liang Lin,
David Zhang, and Lei Zhang. Joint learning of single-image
and cross-image representations for person re-identification. In
CVPR, 2016.
[Wang et al., 2016b] Hanxiao Wang, Shaogang Gong, and Tao
Xiang. Highly efficient regression for scalable person reidentification. In BMVC, 2016.
[Wang et al., 2016c] Hanxiao Wang, Shaogang Gong, Xiatian Zhu,
and Tao Xiang. Human-in-the-loop person re-identification. In
ECCV, 2016.
[Wang et al., 2016d] Taiqing Wang, Shaogang Gong, Xiatian Zhu,
and Shengjin Wang. Person re-identification by discriminative
selection in video ranking. IEEE TPAMI, 38(12):2501–2514,
2016.
[Xiao et al., 2016] Tong Xiao, Hongsheng Li, Wanli Ouyang, and
Xiaogang Wang. Learning deep feature representations with domain guided dropout for person re-identification. In CVPR, 2016.
[Xiong et al., 2014] Fei Xiong, Mengran Gou, Octavia Camps, and
Mario Sznaier. Person re-identification using kernel-based metric
learning methods. In ECCV. Springer, 2014.
[Zhang et al., 2016] Li Zhang, Tao Xiang, and Shaogang Gong.
Learning a discriminative null space for person re-identification.
In CVPR, 2016.
[Zhao et al., 2013] Rui Zhao, Wanli Ouyang, and Xiaogang Wang.
Unsupervised salience learning for person re-identification. In
CVPR, 2013.
[Zheng et al., 2013] Wei-Shi Zheng, Shaogang Gong, and Tao Xiang. Reidentification by relative distance comparison. IEEE
TPAMI, 35(3):653–668, March 2013.
[Zheng et al., 2015] Liang Zheng, Liyue Shen, Lu Tian, Shengjin
Wang, Jingdong Wang, and Qi Tian. Scalable person reidentification: A benchmark. In ICCV, 2015.
| 2cs.AI
|
1
Model Predictive Control for Signal Temporal
Logic Specifications
arXiv:1703.09563v1 [cs.SY] 28 Mar 2017
Vasumathi Raman1 , Alexandre Donzé2 , Mehdi Maasoumy3 ,
Richard M. Murray1 , Alberto Sangiovanni-Vincentelli2 and Sanjit A. Seshia2
Abstract—We present a mathematical programming-based
method for model predictive control of cyber-physical systems
subject to signal temporal logic (STL) specifications. We describe
the use of STL to specify a wide range of properties of these
systems, including safety, response and bounded liveness. For
synthesis, we encode STL specifications as mixed integer-linear
constraints on the system variables in the optimization problem
at each step of a receding horizon control framework. We prove
correctness of our algorithms, and present experimental results
for controller synthesis for building energy and climate control.
Index Terms—formal synthesis, timed logics, model predictive
control, cyberphysical systems
I. I NTRODUCTION
Controlling a cyber-physical system (CPS) involves handling complex interactions between computing components
and their physical environment, and often necessitates hierarchies of controllers. Typically at the highest level, a supervisory controller is responsible for making high-level decisions,
while at the lowest level traditional control laws such as PID
control are used. In general, the design of these different
controllers is done mostly in isolation at each level, and
their combination is implemented ad hoc. As the complexity
of these systems grows, reasoning about the correctness of
interactions between the various layers of control becomes
increasingly challenging, begging automation.
Formal methods is the subfield of computer science concerned with verification and synthesis, i.e., automatic and
rigorous design of digital systems. It provides mathematical
formalisms for specifying behaviors and algorithms for verification and synthesis of a system against properties specified
within these formalisms. Methods for synthesis of correctby-construction discrete supervisory controllers have been
developed and successfully used for cyber-physical systems in
domains including robotics [1] and aircraft power system design [2]. However, for physical systems that require constraints
This work was supported in part by TerraSwarm, one of six centers
of STARnet, a Semiconductor Research Corporation program sponsored by
MARCO and DARPA.
1 V. Raman and R. M. Murray are with the California Institute
of Technology, Pasadena, CA 91125, USA [email protected],
[email protected]
2 A.
Donzé, A. Sangiovanni-Vincentelli and S. A. Seshia
are
with
the
Department
of
Electrical
Engineering
and
Computer Science, UC Berkeley, Berkeley, CA 94720, USA
[email protected], [email protected],
[email protected]
3 M.
Maasoumy is with C3 Energy Inc. Redwood City, CA 94063, USA
[email protected]
Manuscript received January 14, 2016.
not just on the order of events, but on the temporal distance
between them, simulation and testing is still the method of
choice for validating properties and establishing guarantees;
the exact exhaustive verification or synthesis of such systems
is in general undecidable [3].
Model predictive control (MPC) or receding horizon control
(RHC) is based on iterative, finite horizon optimization over
a model of the plant, i.e. the system to be controlled. At
any given time t, the current plant state is observed, and
an optimal control strategy computed for some finite time
horizon in the future, [t, t + H]. An online calculation is
performed to explore trajectories originating from the current
state, and an optimal control strategy computed up to time
t + H. To provide robustness with respect to modelling errors,
only the first step of the computed optimal control strategy is
implemented. The plant state is then sampled again, and new
calculations are performed on a horizon of H starting from
the new current state. While the global optimality of such a
receding horizon approach is not ensured, it tends to do well
in practice: in addition to reducing computational complexity,
it improves the system robustness with respect to exogenous
disturbances and modeling uncertainties [4]. Another reason
MPC is particularly attractive to industry is its ability to handle
constrained dynamical systems [5].
Signal temporal logic (STL) [6] was originally developed in
order to specify and monitor the expected behavior of physical
systems, including temporal constraints between events. STL
allows the specification of properties of dense-time, realvalued signals, and the automatic generation of monitors for
testing these properties on individual simulation traces. It
has since been applied to the analysis of several types of
continuous and hybrid systems, including dynamical systems
and analog circuits, where the continuous variables represent
quantities like currents and voltages in a circuit. STL has
the advantage of naturally admitting quantitative semantics
which, in addition to the yes/no answer to the satisfaction
question, provide a real number that grades the quality of the
satisfaction or violation. Such semantics have been defined for
timed logics, including metric temporal logic (MTL) [7] and
STL [8], to assess the robustness of systems to parameter or
timing variations.
In this paper, we solve the problem of control synthesis from
STL specifications, using a receding horizon approach. We
allow the user to specify desired properties of the system using
an STL formula, and synthesize control such that the system
satisfies that specification, while using a receding horizon
approach to ensure practicality and robustness. We do so by
2
where for all k > 0, tk+1 −tk = ∆t. A run of Σd is a sequence
ξ = (x0 u0 w0 )(x1 u1 w1 )(x2 u2 w2 )...
where xk = x(tk ) ∈ X is the state of the system at
index k, and for each k ∈ N, uk = u(tk ) ∈ U , wk =
w(tk ) ∈ W and xk+1 = fd (xk , uk , wk ). We assume that
given an initial state x0 ∈ X, a control input sequence uN =
u0 u1 u2 . . . uN −1 ∈ U N , and a sequence of environment inputs
wN = w0 w1 w2 . . . wN −1 ∈ W N , the resulting horizon-N run
of a system modeled by equation (1), which we denote by
ξ(x0 , uN , wN ) = (x0 u0 w0 )(x1 u1 w1 )(x2 u2 w2 )...(xN uN wN ),
is unique. In addition, we introduce a generic cost function
J(ξ(x0 , u, w)) that maps (infinite and finite) runs to R.
B. Signal Temporal Logic
ϕ ::= π µ | ¬ψ | ϕ1 ∧ ϕ2 |
We consider a continuous-time system Σ of the form
ẋt = f (xt , ut , wt )
where xt ∈ X ⊆ (Rnc × {0, 1}nl ) are the continuous and
binary/logical states, ut ∈ U ⊆ (Rmc ×{0, 1}ml ) are the (continuous and logical) control inputs, wt ∈ W ⊆ (Rec ×{0, 1}el )
are the external environment inputs (also referred to as “disturbances”). As xt includes binary components, the above
ODE may contain discontinuities corresponding to switches in
these components. In general, such hybrid systems are more
accurately modeled using differential-algebraic equations, but
we do not dwell on this point since we approximate their
behavior with a difference equation, as follows.
Given a sampling time ∆t > 0, we assume that Σ admits a
discrete-time approximation Σd of the form
x(tk+1 ) = fd (x(tk ), u(tk ), w(tk ))
(1)
ϕ | ϕ1 U [a,b] ϕ2 ,
ξ |= ϕ
(ξ, tk ) |= π µ
(ξ, tk ) |= ¬ψ
(ξ, tk ) |= ϕ ∧ ψ
(ξ, tk ) |= [a,b] ϕ
(ξ, tk ) |= ϕ U [a,b] ψ
⇔
⇔
⇔
⇔
⇔
⇔
(ξ, t0 ) |= ϕ
µ(xk , yk , uk , wk ) > 0
¬(ξ, tk ) |= ψ)
(ξ, tk ) |= ϕ ∧ (ξ, tk ) |= ψ
∃tk0 ∈ [tk +a, tk +b], (ξ, tk0 ) |= ϕ
∃tk0 ∈ [tk +a, tk +b] s.t. (ξ, tk0 ) |= ψ
∧∀tk00 ∈ [tk , tk0 ], (ξ, tk00 ) |= ϕ.
An STL formula ϕ is bounded-time if it contains no
unbounded operators; the bound of ϕ is the maximum over the
sums of all nested upper bounds on the temporal operators, and
provides a conservative maximum trajectory length required
to decide its satisfiability. For example, for [0,10] [1,6] ϕ, a
trajectory of length N such that tN ≥ 10+6 = 16 is sufficient
to determine whether the formula is satisfiable.
A. Systems
[a,b]
where π µ is an atomic predicate X × U × W → B whose truth
value is determined by the sign of a function µ : X ×U ×W →
R and ϕ1 , ϕ2 are STL formulas. The fact that a run ξ(x0 , u, w)
satisfies an STL formula ϕ is denoted by ξ |= ϕ. Informally,
ξ |= [a,b] ϕ if ϕ holds at some time step between a and b, and
ξ |= ϕ U [a,b] ψ if ϕ holds at every time step before ψ holds,
and ψ holds at some time step between a and b. Additionally,
we define [a,b] ϕ = ¬ [a,b] (¬ϕ), so that ξ |= [a,b] ϕ if ϕ
holds at all times between a and b. Formally, the validity of a
formula ϕ with respect to the run ξ is defined inductively as
follows
II. P RELIMINARIES
We consider STL formulas defined recursively according to
the following grammar:
decomposing the STL specifications into a series of formulas
over each time horizon, such that synthesizing a controller fulfilling the formula at each horizon results in satisfaction of the
global specification. Recent work on optimal control synthesis
of aircraft load management systems [9] represented STLlike specifications as time-dependent equality and inequality
constraints, yielding a mixed integer linear program (MILP).
The MILP was then solved in an MPC framework, yielding
an optimal control policy. However, the manual transformation
of specifications into equality and inequality constraints is
cumbersome and problem-specific. As a key contribution, this
paper presents two automatically-generated MILP encodings
for such STL specifications.
Our main contribution is a pair of bounded model checkingstyle encodings [10] for STL specifications as MILP constraints on a cyber-physical system. We show how these encodings can be used to generate open-loop control signals that
satisfy finite and infinite horizon STL properties and, moreover, to generate signals that maximize quantitative (robust)
satisfaction. We provide a fragment of STL, denoted SNNSTL, such that, under reasonable assumptions on the system
dynamics, the problem of synthesizing an open-loop control
sequence such that the system satisfies a provided specification is a linear program (LP), and therefore polynomial-time
solvable. We also demonstrate how our MILP formulation of
the STL synthesis problem can be used in an MPC framework
to compute feasible and optimal controllers for cyber-physical
systems under timed specifications. We present experimental
results comparing both encodings, and two case studies: one on
a thermal model of an heating, ventilation and air-conditioning
(HVAC) system, and another in the context of regulation
services in a micro-grid. These case studies were previously
reported in [11], [12]. We show how the MPC schemes in
these examples can be framed in terms of synthesis from an
STL specification, and present simulation results to illustrate
the effectiveness of our methodology.
Remark 1. Here we have defined a semantics for STL over
discrete-time signals, which is formally equivalent to the
simpler linear temporal logic (LTL), once time and predicates
are abstracted into steps and Boolean variables, respectively.
There are several advantages of still using STL over LTL,
though. First, STL allows us to explicitly use real time in
our specifications instead of abstract integer indices, which
improves the readability relative to the original system’s behaviors. Second, although in the rest of this paper we focus
on the control of the discrete-time system Σd , our goal is to
use the resulting controller for the control of the continuous
3
system Σ. Hence the specifications should be independent
from the sampling time ∆t. Finally, note that the relationship
between the continuous-time and discrete-time semantics of
STL, depending on discretization error and sampling time, is
beyond the scope of this paper. The interested reader can refer
to [13] for further discussion on this topic.
Note that we assume that the state of the plant is fully
observable, and the environment inputs are known in advance.
Problem 2 (closed-loop). Given a horizon 0 < L < N , for
all 0 ≤ k ≤ N − L, compute u∗k = uL∗
k , the first element of
L∗ L∗
L∗
the sequence uL∗
=
u
u
.
.
.
u
k
k
k+1
k+L−1 satisfying
uL∗
k =
C. Quantitative semantics for STL
argmin J(xk , uL
k , wk , ϕ)
L
uL
K ∈U
s.t. ξ(xk , uL
k , wk ) |= ϕ
Quantitative or robust semantics for STL are defined by
providing a real-valued function ρϕ of signal ξ and time t
such that ρϕ (ξ, t) > 0 ⇒ (ξ, t) |= ϕ. We define one such
function recursively, as follows:
The closed-loop formulation corresponds to a model predictive control scheme, where the reference disturbance can
change at each iteration k.
πµ
In Sections IV and V, we present both an open-loop solution
ρ (ξ, tk )
= µ(xk , yk , uk , wk )
to Problem 1, and a solution to Problem 2 for a large class
ρ¬ψ (ξ, tk )
= −ρψ (ξ, tk )
of STL formulas. In the absence of an objective function
ρϕ1 ∧ϕ2 (ξ, tk )
= min(ρϕ1 (ξ, tk ), ρϕ2 (ξ, tk ))
ϕ1 ∨ϕ2
J on runs of the system, we maximize the robustness of
(ξ, tk )
= max(ρϕ1 (ξ, tk ), ρϕ2 (ξ, tk ))
ρ
the generated runs with respect to ϕ. A key component of
ψ
ψ
[a,b]
(ξ, tk ) = maxtk0 ∈[t+a,t+b] ρ (ξ, tk0 )
ρ
our solution is encoding the STL specifications as MILP
ϕ1 U [a,b] ϕ2
ϕ2
ρ
(ξ, tk ) = maxtk0 ∈[t+a,t+b] (min(ρ (ξ, tk0 ),
constraints, which can be combined with MILP constraints
ϕ1
mintk00 ∈[tk ,tk0 ] ρ (ξ, tk00 ))
representing the system dynamics to efficiently solve the
µ
Note that if µ(xk , yk , uk , wk ) = 0, neither ρπ (ξ, tk ) > 0 resulting state-constrained optimization problem.
µ
nor ρπ (ξ, tk ) > 0. Therefore, (ξ, t) |= ϕ 6⇒ ρϕ (ξ, t) > 0. To
µ
simplify notation, we denote ρπ by ρµ for the remainder of
IV. O PEN - LOOP C ONTROLLER S YNTHESIS
the paper. The robustness of satisfaction for an arbitrary STL
formula is computed recursively from the above semantics by
To solve Problem 1, we extend the bounded model checking
propagating the values of the functions associated with each
encoding of [14] from finite, discrete systems to dynamical
operand using min and max operators corresponding to the
systems using mixed-integer programming instead of SAT. Our
various STL operators. For example, the robust satisfaction of
presentation and notation below follow that of [15]. For openπ µ1 where µ1 (x) = x − 3 > 0 at time 0 is ρµ1 (ξ, 0) = x0 − 3.
loop controller synthesis, we will search for a trajectory of
The robust satisfaction of µ1 ∧ µ2 is the minimum of ρµ1
length N that satisfies ϕ. To admit STL formulas describing
and ρµ2 . Temporal operators are treated as conjunctions and
infinite runs, we parametrize an infinite sequence of states
disjunctions along the time axis: since we deal with discrete
using a finite sequence with a loop. Imposing this lassotime, the robustness of satisfaction of ϕ = [0,2.1] µ1 is
shaped structure renders our synthesis procedure conservative
ϕ
µ1
for infinite-state systems, in the sense that a solution may exist
ρ (x, 0) = mintk ∈[0,2.1] ρ (x, tk )
that is not found when imposing such a structure. For finite= min{x0 − 3, x1 − 3, . . . , xK − 3},
state systems, the lasso shape is without loss of generality but
where 0 ≤ t0 < t1 < . . . < tK ≤ 2.1 < tK+1 .
we must still find an appropriate trajectory length N .
The robustness score ρϕ (ξ, t) can be interpreted as how
much ξ satisfies ϕ. Its absolute value can be viewed as the
ϕ, Σd , J
signed distance of ξ from the set of trajectories satisfying or
violating ϕ, in the space of projections with respect to the
function µ that define the predicates of ϕ [7].
III. P ROBLEM S TATEMENT
We now formally state the STL control synthesis problem
and its model predictive control formulation. Given an STL
formula ϕ, a cost function of the form J(x0 , u, w, ϕ) ∈ R, an
initial state x0 ∈ X , a horizon L and a reference disturbance
signal w ∈ W N , we formulate two problems: open-loop and
closed-loop synthesis. The two scenarios are depicted as block
diagrams in Fig. 1 and Fig. 2.
Synthesis
x0 ,
w0 , . . . , wN −1
u∗ =
u∗0 u∗1 . . . u∗N −1
Plant Σ
x, y
x0 , w
Fig. 1. Open-loop problem formulation: given an STL formula φ, the discretetime plant model Σd , the cost function J, the goal is to generate a sequence of
control inputs u∗ over a horizon of N time steps. The problem is additionally
parametrized by the the initial state x0 and disturbance vector w.
Problem 1 (open-loop). Compute u∗ = u∗0 u∗1 . . . u∗N −1 where
u∗ =
argmin J(x0 , u, w, ϕ)
u∈U N
s.t. ξ(x0 , u, w) |= ϕ
The encoding of Problem 1 as an MILP consists of system
constraints, loop constraints and STL constraints, as defined
below.
4
C. Boolean encoding of STL constraints
Given a formula ϕ, we introduce a variable ztϕ , whose value
is tied to a set of mixed integer linear constraints required for
the satisfaction of ϕ at position t in the state sequence of
horizon N . In other words, ztϕ has an associated set of MILP
constraints such that ztϕ = 1 if and only if ϕ holds at position
t. We recursively generate the MILP constraints corresponding
to z0ϕ – the value of this variable determines whether a formula
ϕ holds in the initial state.
1) Predicates: The predicates are represented by constraints on system state variables. For each predicate µ ∈ P ,
we introduce binary variables ztµ ∈ {0, 1} for times t =
0, 1, ..., N . The following constraints enforce that ztµ = 1 if
and only if µ(xt ) > 0:
µ(xt )
≤
−µ(xt ) ≤
Mt ztµ − t
Mt (1 − ztµ ) − t
ϕi
Conjunction: ztψ = ∧m
i=1 zti
ztψ ≤ ztϕii , i = 1, ..., m,
Pm
ztψ ≥ 1 − m + i=1 ztϕii
ztψ ≥ ztϕii , i = 1, ..., m,
Pm
ztψ ≤ i=1 ztϕii
ϕi
Disjunction: ztψ = ∨m
i=1 zti
Given a formula ψ containing a Boolean operation, we add
new continuous variables ztψ ∈ [0, 1], and set ztψ = ¬ztµ ,
ϕi
ψ
ϕi
m
m
ztψ = ∧m
i=1 zti , and zt = ∨i=1 zti for ψ = ¬µ, ψ = ∧i=1 ϕi
m
and ψ = ∨i=1 ϕi , respectively. These constraints enforce that
ztψ = 1 if ψ holds at time t and ztψ = 0 otherwise.
3) Temporal constraints: We first present encodings for
the and
operators. We will use these encodings to define
the encoding for the U[a,b] operator.
Always: ψ = [a,b] ϕ
N
Let aN
t = min(t + a, N ) and bt = min(t + b, N )
W
Vb̂N
bN
N
ϕ
ψ
ϕ
j
t
Define zt = ∧i=a
lj ∧ i=â
N zi ∧ (
N zi )
j=1
t
j
The logical operation ∧ on the variables ziϕ here is as defined
in Section IV-C2. Intuitively, this encoding enforces that the
formula ϕ is satisfied at every time step on the interval [a, b]
relative to time step t.
Eventually: ψ =
[a,b]
ϕ
WN
Wb̂N
bN
ϕ
ϕ
j
t
Define ztψ = ∨i=a
lj ∧ i=â
N zi ∧ (
N zi ) This encoding
j=1
t
j
enforces that the formula ϕ is satisfied at some time step on
the interval [a, b] relative to time step t.
Until: ψ = ϕ1 U[a,b] ϕ2
The bounded until operator U[a,b] can be defined in terms of
the unbounded U (inherited from LTL) as follows [18]:
ϕ1 U[a,b] ϕ2 = [0,a] ϕ1 ∧
[a,b]
ϕ2 ∧
As described above, to synthesize open-loop control for
unbounded (infinite-horizon) specifications, we parametrize
the trajectory as a lasso, i.e. constrain it to contain a loop. This
loop encoding is again inspired by the basic idea of bounded
model checking [10], which is to consider only a finite prefix
of a path when looking for a solution to an existential model
checking problem. A crucial observation is that although the
considered path prefix is finite, it can still represent an infinite
path if there is a loop back from the last state to any of the
previous states.
To enforce the existence of a loop in the finite system
trajectory, we introduce N binary variables l1 , ..., lN , which
determine where the loop forms. These are constrained such
that only one can be high at a time, and if lj = 1, then
xj−1 = xN . The following constraints enforce these requirements:
PN
•
j=1 lj = 1
• xN ≤ xj−1 + Mj (1 − lj ), j = 1, ..., N ,
• xN ≥ xj−1 + Mj (1 − lj ), j = 1, ..., N ,
where Mj are sufficiently large positive numbers, picked based
on X .
B. Loop constraints for trajectory parametrization
The first component of the set of constraints is provided by
the system model. Our approach applies to any system that
yields to a MILP formulation for model predictive control
over horizon N . The system constraints encode valid finite
(horizon-N ) trajectories for a system with form (1) – these
constraints hold if and only if the trajectory x(x0 , uN ) satisfies
(1) for t = 0, 1, ..., N . Note that this is quite general, and
accommodates any system for which the resulting constraints
and objectives form a mixed integer-linear program. An example is the smart grid regulation control system presented in
[16]. Other useful examples include mixed logical dynamical
systems such as those presented in [17]. Other cost functions
and system dynamics can also be included by using appropriate solvers.
where Mt are sufficiently large positive numbers, and t
are sufficiently small positive numbers that serve to bound
µ(xt ) away from 0. This encoding restricts the set of STL
formulas that can be encoded using our approach to those
over linear predicates, but admits arbitrary STL formulas over
such predicates.
2) Boolean operations on MILP variables: As described
in Section IV-C1, each predicate µ has an associated binary
variable ztµ which equals 1 if µ holds at time t, and 0
otherwise. In fact, by the recursive definition of our MILP
constraints on STL formulas, each operand ϕ in a Boolean
operation has a corresponding variable ztϕ which is 1 if ϕ holds
at t and 0 otherwise. Here we define Boolean operations on
these variables: these are the building blocks of our recursive
encoding. The definitions in this subsection are consistent with
those in [15].
Logical operations on variables ztψ ∈ [0, 1] are defined as
follows:
ztψ = 1 − ztϕ
Negation: ztψ = ¬ztϕ
A. Constraints on system evolution
[a,a] (ϕ1
U ϕ2 )
We will use the encoding of the unbounded U from
[10]. When encoding over infinite trajectories, this requires
5
an auxiliary encoding that prevents the pitfalls of circular
reasoning on the finite parametrization of the infinite
sequences. The interested reader is referred to [10] for
the details of the encoding. The auxiliary encoding of the
unbounded until is
hh ϕ1(U ϕ2 iit =
ztϕ2 ∨ (ztϕ1 ∧ hh ϕ1 U ϕ2 iit+1 ), t = 1, ..., N − 1
ϕ2
zN
.
With this definition in place, we define
ztϕ1
U ϕ2
ϕ1 U
= ztϕ2 ∨ (ztϕ1 ∧ zt+1
ϕ2
)
for t = 1, ..., N − 1, and
ϕ1
zN
U ϕ2
ϕ2
ϕ1
= zN
∨ (zN
∧(
N
_
(lj ∧ hh ϕ1 U ϕ2 iij ))).
j=1
[0,a] ϕ1
ϕ1 U[a,b] ϕ2
zt
∧ zt
= zt
[a,b]
ϕ2
Given this encoding of the unbounded until and the encodings
of [a,b] and [a,b] above, we can encode
∧ zt
[a,a] (ϕ1
U ϕ2 )
.
ztϕ
By induction on the structure of STL formulas ϕ,
= 1
if and only if ϕ holds on the system at time t. With this
motivation, given a specification ϕ, we add a final constraint:
z0ϕ = 1.
(2)
For a bounded horizon formula, the union of the STL
constraints, loop constraints and system constraints gives the
MILP encoding of Problem 1; this enables checking feasibility
of this set of constraints and finding a solution using an MILP
solver. Given an objective function on runs of the system, this
approach also enables finding the optimal open-loop trajectory
that satisfies the STL specification. Algorithm 1 reviews the
procedure for solving Problem 1.
Algorithm 1 Algorithm for Problem 1
1: procedure OPEN LOOP(f, x0 , w, N, ϕ, J)
2:
LOOP CONSTRAINTS ← Sec. IV-B
3:
SYSTEM CONSTRAINTS ← Sec. IV-A
4:
STL CONSTRAINTS ← Sec. IV-C2 OR Sec. IV-D
5:
u∗ ← argmin
J(x0 , u, w, ϕ)
u∈U N
s.t.
LOOP CONSTRAINTS
SYSTEM CONSTRAINTS
STL CONSTRAINTS
Return u∗
6: end procedure
D. Quantitative Encoding
The robustness of satisfaction of the STL specification, as
defined in II-C, provides a natural objective for the MILP
defined in Section IV-C, either in the absence of, or as
a complement to domain-specific objectives on runs of the
system. The robustness can be computed recursively on the
structure of the formula in conjunction with the generation of
constraints. Moreover, since max and min operations can be
expressed in an MILP formulation using additional binary variables, this does not add complexity to the encoding, although
the additional variables do make it more computationally
expensive in practice.
In this section, we sketch the MILP encoding of the
predicates and Boolean operators using the quantitative semantics; the encoding of the temporal operators builds on
these encodings, as in Section IV-C. Given a formula ϕ,
we introduce a variable rtϕ , and an associated set of MILP
constraints such that rtϕ > 0 if and only if ϕ holds at position
t. We recursively generate the MILP constraints, such that
r0ϕ determines whether a formula ϕ holds in the initial state.
Additionally, we enforce rtϕ = ρϕ (x, t).
For each predicate µ ∈ P , we now introduce variables rtµ
for time indices t = 0, 1, ..., N , and set rtµ = µ(xt ). To define
rtψ , where ψ is a Boolean formula, we inductively assume that
each operand ϕ has a corresponding variable rtϕ = ρϕ (x, t).
Then the Boolean operations are defined as:
rtψ = −rtφ
Negation: rtψ = ¬rtφ
ϕi
Conjunction: rtψ = ∧m
i=1 rti
m
X
i
pϕ
ti = 1
i=1
rtψ ≤ rtϕii , i = 1, ..., m
ψ
i
rtϕii − (1 − pϕ
ti )M ≤ rt
(3)
(4)
≤
rtϕii
+ M (1 −
i
pϕ
ti )
(5)
i
pϕ
ti
where we introduce new binary variables
for i = 1, ..., m,
and M is a sufficiently large positive number. Then equation
(3) enforces that there is one and only one j ∈ {1, ..., m}
ϕ
such that ptij = 1, equation (4) ensures that rtψ is smaller
ϕ
than all rtϕii , and equation (5) enforces that rtψ = rtjj if and
ϕj
only if ptj = 1. Together, these constraints enforce that rtψ =
mini (rtϕii ).
ϕi
Disjunction: ψ = ∨m
i=1 rti is encoded similarly to conjunc-
tion, replacing (4) with rtψ ≥ rtϕii , i = 1, ..., m. Using a similar
reasoning to that above, this enforces rtψ = maxi (rtϕii ).
The encoding for bounded temporal operators is defined as
in Section IV-C; robustness for the unbounded until is defined
using sup and inf instead of max and min, but these are
equivalent on our finite trajectory representation with discrete
time. By induction on the structure of STL formulas ϕ, this
construction yields rtϕ > 0 if and only if ϕ is satisfied at time
t. Therefore, we can replace the constraints over ztϕ in Section
IV-C by these constraints that compute the value of rtϕ , and
instead of (2), add the constraint r0ϕ > 0.
Since we consider only the discrete time semantics of
STL in this work, the Boolean encoding in Section IV-C
could be achieved by converting each formula to LTL, and
using existing encodings such as that in [15]. However, the
robustness-based encoding we presented in this section has no
natural analog for LTL. The advantage of this encoding is that
it allows us to maximize the value of r0ϕ , obtaining a trajectory
that maximizes robustness of satisfaction. Additionally, an
encoding based on robustness has the advantage of allowing
the STL constraints to be softened or hardened as necessary.
6
For example, if the original problem is infeasible, we can
allow ρϕ
0 > − for some > 0, thereby easily modifying
the problem to allow a limited violation of the STL property.
The disadvantage is that it is more expensive to compute,
due the the additional binary variables introduced during each
Boolean operation. Additionally, including robustness as an
objective makes the cost function inherently non-convex, with
potentially many local minima, and harder to optimize. On the
other hand, the robustness constraints are more easily relaxed,
allowing us to use a simpler cost function, which can make
the problem more tractable.
E. Complexity
In general, our synthesis algorithm has the same complexity as MILPs, which are NP-hard, hence computationally
challenging when the dimensions of the problem grow. It
is nevertheless appropriate to characterize the computational
costs of our encoding and approach in terms of the number
of variables and constraints in the resulting MILP. In practice,
one measure of problem size is the number of binary variables
required to indicate the satisfaction of the predicates µ. This
depends directly on the number of predicates used in the STL
formula ϕ.
For the Boolean encoding, if P is the set of predicates
used in the formula, then O(N · |P |) binary variables are
introduced. In addition, continuous variables are introduced
during the MILP encoding of the STL formula. The number
of continuous variables used is O(N · |ϕ|), where |ϕ| is the
length (i.e. the number of operators) of the formula.
For the robustness-based encoding, O(N · |P |) continuous
variables are introduced (one per predicate per time step).
In addition, binary variables are introduced during the MILP
encoding of each operator in the STL formula. The number
of binary variables used is thus O(N · |ϕ|), where |ϕ| is the
number of operators of the formula.
Our synthesis algorithm also has polynomial runtime for the
following fragment of STL.
Definition 1 (SNN-STL). Safe Negation-Normal STL (SNNSTL) is the fragment of STL generated by the recursive
grammar
ϕ ::= π µ | ¬π µ | ϕ1 ∧ ϕ2 | [a,b] ϕ
SNN-STL has the following properties:
• All negations appear only on atomic propositions (pushed
down to the leaf nodes of the formula abstract syntax
tree).
• The only temporal operators are (with unbounded and
bounded intervals).
• Only conjunctions are allowed, no disjunctions.
Such specifications are expressive enough to enforce, e.g.,
safety specifications in environments where the system state
is confined to a conjunction of polyhedra.
Let
OPEN LOOP NO STL(f, x0 , N, ϕ, J)
denote the procedure that is identical to Algorithm 1, except that the optimization problem in Step 5 is solved with
STL CONSTRAINTS = ∅.
ϕ, Σd , J
Controller
u∗ =
u∗0 u∗1 . . . u∗N −1
x0
Synthesis
w0 w1 . . . wN −1
u∗0
Plant Σ
x1
x0
w0
Fig. 2. Closed-loop (MPC) problem formulation. As in the open loop scenario,
a sequence of control inputs is synthesized from the specifications, dynamics
and cost function. However, at each time step, only the first computed input
is used by the plant.
Theorem 1 (Polynomial-time Synthesis for SNN-STL). Suppose that ϕ is in SNN-LTL and has linear predicates.
Then if OPEN LOOP NO STL(f, x0 , w, N, ϕ, J) is convex, so
is OPEN LOOP(f, x0 , N, ϕ, J).
Proof. The proof proceeds by induction on the structure
of the formula, showing that the constraints added by
STL CONSTRAINTS for each operator restrict the solution to
a convex set. First note that since the predicates π µ are linear,
negation of predicates preserves convexity, since ¬π µ is also
linear. Because the intersection of convex sets is convex, the
conjunction of a set of convex constraints is also convex.
Finally, since the operator is implemented in terms of
conjunctions, the constraints imposed by also preserve
convexity of the resulting optimization problem.
Informally, Theorem 1 states that our encoding of SNN-STL
constraints into an MILP preserves convexity in the resulting
optimization problem. The resulting optimization problem is
therefore encodable as an LP, i.e. without the use of integer
variables.
Corollary 1. Algorithm 1 is polynomial-time for SNN-STL
specifications ϕ.
V. M ODEL P REDICTIVE C ONTROL S YNTHESIS
In this section, we will describe a solution to Problem 2
by adding STL constraints to an MPC problem formulation.
At each step t of the MPC computation, we will search for
a finite trajectory of fixed horizon length H, such that the
accumulated trajectory satisfies ϕ.
A. Synthesis for bounded-time STL formulas
The length of the horizon H is chosen to be at least the
bound of formula ϕ. At time step 0, we will synthesize control
uH,0 using the open-loop formulation in Section IV, including
7
the STL constraints on the length-H trajectory, but without the
loop constraints. We will then execute only the first time step
u0H,0 . At the next step of the MPC, we will solve for uH,1 ,
while constraining the previous values of x0 , u0 in the MILP,
and the STL constraints on the trajectory up to time H. In this
manner, we will keep track of the history of states in order to
ensure that the formula is satisfied over the length-H prefix
of the trajectory, while solving for uH,t at every time step t.
Algorithm 2 MPC Algorithm for Problem 2
1: procedure MPC(f, x0 , φ = ϕM P C , J)
2:
Let M be a large positive constant.
3:
Let H be the bound of ϕM P C .
4:
Set P0 = 0 and Pi = −M ∀0 < i ≤ H.
5:
wt ← PREDICT W(0).
6:
Compute u0 = u00 , u01 , ...., u02H−1 as:
B. Extension to unbounded formulas
7:
8:
9:
10:
11:
For certain types of unbounded formulas, we can stitch
together trajectories of length H using a receding horizon
approach, to produce an infinite computation that satisfies the
STL formula. An example of this is safety properties, i.e.
ϕ = (ϕM P C ) for bounded STL formulas ϕM P C . For such
formulas, at each step of the MPC computation, we will search
for a finite trajectory of horizon length H (determined from
ϕM P C as above) that satisfies ϕM P C .
We now describe this approach in more detail. At each step
t of the receding horizon control computation, we will employ
the open-loop approach in Section IV to find a finite trajectory
of fixed horizon length H, such that the trajectory accumulated
over time satisfies ϕ. Given a specification ϕ = ϕM P C ,
where ϕM P C is a bounded-time formula with bound H. In
this case, we can stitch together trajectories of length H
using a receding horizon approach to produce an infinite
computation that satisfies the STL formula. At each step of the
receding horizon computation, we search for a finite trajectory
of horizon length 2H, keeping track of the past values and
robustness constraints necessary to determine satisfaction of
ϕ at every time step in the trajectory. Note that we omit the
loop constraints in this approach, because at each step we
search for a finite trajectory, rather than an infinite trajectory
with a finite parametrization.
First we define a procedure
OPEN LOOP∗ (f, x0 , w, N, ϕM P C , J, PH , utold )
that takes additional inputs P = {P0 , P1 , ..., PH−1 } and
utold = u0 , u1 , ..., ut−1 , and is identical to Algorithm 1, except
that the optimization problem posed in Step 5 is solved without
the loop constraints, and with the added constraints:
ρϕ (f (x0 , u, w), i) > Pi ∀i ∈ [0, H − 1]
u[i...t] = utold
We then define a receding horizon control procedure as in
Algorithm 2. At each step, we are optimizing over a horizon
of 2H. We assume available a method PREDICT W(t) for
predicting the sequence of 2H environment inputs starting at
time step t.
Algorithm 2 has two phases, a transient phase (Lines 410) and a stationary phase (Lines 11-14). The transient phase
applies until an initial control sequence of length H has been
computed, and the stationary phase follows. In the transient
phase, the number of stored previous inputs (utold ) as well as
the number of time steps at which formula ϕM P C is enforced
(i.e. time steps for which Pi = 0) grows by one at each
iteration, until they both attain a maximum of H at iteration H.
u0 ← OPEN LOOP∗ (f, x0 , w0 , 2H, [0,H] ϕM P C , J, PH , ∅)
for t=1; t¡=H;t=t+1 do
Set utold = u00 , u11 , u22 , ..., ut−1
t−1 .
Set Pi = 0 for 0 ≤ i ≤ t, Pi = −M ∀t < i ≤ H.
wt ← PREDICT W(t).
Compute ut = ut0 , ut1 , ...., ut2H−1 as:
ut ← OPEN LOOP∗ (f, xt , wt , 2H, [0,H] ϕM P C , J, PH , utold )
12:
13:
14:
15:
16:
end for
while True do
t−1
t−1
t−1
.
Set utold = ut−1
1 , u2 , u3 , ..., ut
Set Pi = 0 for 0 ≤ i ≤ H.
wt ← PREDICT W(t).
ut ← OPEN LOOP∗ (f, xt , wt , 2H, [0,H] ϕM P C , J, PH , utold )
17:
18:
end while
end procedure
Every following iteration uses a window of size H for stored
previous inputs, and sets all Pi = 0. The size-H window
of previously-computed inputs advances forward one step in
time at each iteration after step H. In this manner, we keep
a record of the previously computed inputs required to ensure
satisfaction of ϕM P C up to H time steps in the past.
We now show that if Algorithm 2 does not terminate,
then the resulting infinite sequence of control inputs enforces
satisfaction of the specification φ = ϕM P C .
Theorem 2. Let φ = ϕM P C , and assume that u∗ is
an infinite sequence of control inputs generated by setting
u∗ [t] = ut0 , where ut = ut0 ut1 ...ut2H−1 is the control input
sequence of length 2H generated by Algorithm 2 at time t.
Then f (x0 , u∗ , w) |= ϕ.
Proof. Since H is the bound of ϕM P C , the satisfaction of
ϕM P C at time t is established by the control inputs u∗ [t :
t + H − 1]. At time t + H,
ut+H
old
=
=
=
=
ut+H
, ut+H
, ut+H
, ..., ut+H
0
1
2
t+H−1
t+H−1
t+H−1
u1
, u2
, ut+H−1
, ..., ut+H−1
3
t+H
t+1
t+2
t+H−1
utt , ut+1 , ut+2 , ..., ut+H−1
u∗ [t : t + H − 1],
and so all the inputs required to determine satisfaction of ϕ
at time t have been fixed. Moreover, if ut+H is successfully
computed, then by the correctness of Algorithm 1, ut+H
old has
H
∗
the property that f (xt , ut+H
old , w ) |= ϕM P C . Since u [t : t +
t+H
∗
H
H −1] = uold , we see that f (xt , u [t : t+h], w ) |= ϕM P C .
It follows that f (x0 , u∗ , w) |= ϕM P C .
We have therefore shown how a control input can be
8
synthesized for infinite sequences satisfying ϕ, by repeatedly
synthesizing control for sequences of length 2H. A similar
approach applies for formulas ϕM P C and ϕM P C U ψM P C ,
where ϕM P C , ψM P C are bounded-time.
Note that we assumed that PREDICT W(t) returns an exact
prediction of the disturbance signal over the next 2H time
steps. The correctness of our approach relies on this assumption. An interesting direction of future work is to relax this
requirement, demanding only an uncertain prediction of the
disturbance signal.
Remark 2. The control objective for MPC is usually to steer
the state to the origin or to an equilibrium state. Questions that
arise include those of ensuring feasibility at each time step,
closed-loop stability and near-optimal performance [19]. There
is a mature theory of stability for MPC, where the essential
ingredients are terminal costs, terminal constraint sets, and
local stabilizing controller that ensure closed-loop stability [5].
In this work, our control objective is not closed-loop stability, but satisfaction of an STL formula. We achieve this, as
detailed above, through choice of a sufficiently large prediction
horizon H. This can be compared with the manner in which
automatic satisfaction of a terminal constraint is sometimes
attained by prior choice of a sufficiently large horizon.
VI. E XPERIMENTAL C OMPARISON OF E NCODINGS
We implemented the Boolean and robust encodings using
the tools Breach [20] and YALMIP [21], and now present
results obtained with the following formulas:
ϕ1
ϕ2
ϕ3
ϕ4
#constraints
B
R
154
488
364
897
244 1282
574 1330
YALMIP
B
1.71
1.94
1.84
2.29
Time (s)
R
2.04
2.69
3.15
3.37
Solver time (s)
B
R
0.0070 0.0085
0.0115 0.0229
0.0064 0.1356
0.2167
238.6
TABLE I
B OOLEAN (B) VS ROBUST (R) ENCODINGS . YALMIP TIME REPRESENTS
THE TIME TAKEN BY THE TOOL YALMIP IN ORDER TO GENERATE THE
MILP AND S OLVER TIME IS THE TIME TAKEN BY THE SOLVER G UROBI TO
ACTUALLY SOLVE IT.
the Boolean encoding. The reason is hard to pinpoint without
a more thorough investigation, but we can already note that
solving a MILP is NP-hard, and while solvers use sophisticated
heuristics to mitigate this complexity, instances for which these
heuristics fail are bound to appear.
VII. C ASE S TUDY: B UILDING C LIMATE C ONTROL
A. Mathematical Model of a Building
Next we consider the problem of controlling building indoor
climate, using the model proposed by Maasoumy et al [22]. In
this section we present a summary of the building’s thermal
model.
As shown in Fig. 3, the building is modeled as a resistorcapacitor circuit with n nodes, m of which are rooms and
the remaining n − m are walls. We denote the temperature of
room ri by Tri . The wall and temperature of the wall between
rooms i and j are denoted by wi,j and Twi,j , respectively.
The temperature of wall wi,j and room ri are governed by the
following equations:
ϕ1
ϕ2
ϕ3
ϕ4
(1)
= [0,0.1] xt > 0.1
(1)
(2)
•
= [0,0.1] (xt > 0.1) ∧ [0,0.1] (xt < −0.5)
(1)
•
= [0,0.5] [0,0.1] (xt > 0.1)
(1)
(2)
•
= [0,0.2] (xt > 0.1 ∧ ( [0,0.1] (xt > 0.1)
(3)
∧ [0,0.1] (xt > 0.1)))
In this study, we used the trivial system x = u, where x
(1) (2) (3)
is a 3-dimensional signal (i.e. xt = xt xt xt ), so that
no constraint is generatedPfor the system dynamics, and the
N
cost function J(x, u) = k=1 kutk k1 . Note that the output
of this procedure for a formula ϕ is a signal of minimal
norm which satisfies ϕ when using the Boolean encoding and
which satisfies ϕ with a specified robustness ρϕ (x) = 0.1
for the robust encoding. For each formula we computed the
Boolean and robust encodings for an horizon N = 30 and
sampling time τ = 0.025s and report the number of constraints
generated by each encoding, the time to create the resulting
MILP with YALMIP and the time to solve it using the solver
Gurobi.1 All experiments were run on a laptop with an Intel
Core i7 2.3 GHz processor and 16 GB of memory.
A first observation is that for both encodings, most of the
time is spent creating the MILP, while solving it is done
in a fraction of a second. Also, while the robust encoding
generates 3 to 5 times more constraints, the computational
time to create and solve the corresponding MILPs is hardly
twice more. The exception is solving the MILP for ϕ4 , which
takes significantly more time for the robust encoding than for
•
Formula
1 http://www.gurobi.com/
w
Ci,j
X
dTwi,j
=
dt
k∈Nwi,j
Trk − Twi,j
+ ri,j αi,j Awi,j Qradi,j
Ri,jk
(6)
X Tk − Tr
dTri
i
Cir
=
+ ṁri ca (Tsi − Tri )+
dt
Ri,ki
k∈Nri
wi τwi Awini Qradi + Q̇inti ,
(7)
w
where Ci,j
, αi,j and Awi,j are heat capacity, a radiative heat
absorption coefficient, and the area of wi,j , respectively. Ri,jk
is the total thermal resistance between the centerline of wall
(i, j) and the side of the wall on which node k is located.
Qradi,j is the radiative heat flux density on wi,j . Nwi,j is the
set of all neighboring nodes to wi,j . ri,j is a wall identifier,
which equals 0 for internal walls and 1 for peripheral walls,
where either i or j is the outside node. Tri , Cir and ṁri are
the temperature, heat capacity and air mass flow into room i,
respectively. ca is the specific heat capacity of air, and Tsi is
the temperature of the supply air to room i. wi is a window
identifier, which equals 0 if none of the walls surrounding
room i have windows, and 1 if at least one of them does. τwi
is the transmissivity of the glass of window i, Awini is the
total area of the windows on walls surrounding room i, Qradi
is the radiative heat flux density per unit area radiated to room
i, and Q̇inti is the internal heat generation in room i. Nri is
9
Room temperature [oC]
23
Simulation
Measurement
22
21
20
19
18
12:00 am
6:00 am
12:00 pm
6:00 pm
12:00 am
6:00 am
12:00 pm
Time of day [hr]
6:00 pm
12:00 am
2
o
Disturbance [ C/hr]
3
1
0
−1
12:00 am
Fig. 4. Simulated temperature, measured temperature and unmodelled dynamics of a thermal zone in Bancroft library on UC Berkeley campus.
Fig. 3. Resistor-capacitor representation of a typical room with a window.
the set of neighboring room nodes for room i. Further details
on this thermal model can be found in [22].
The heat transfer equations for each wall and room yield
the system dynamics:
ẋt = f (xt , ut , wt ).
Here xt ∈ Rn is the state vector representing the temperature
of the nodes in the thermal network, and ut ∈ Rlm is the
input vector representing the air mass flow rate and discharge
air temperature of conditioned air into each thermal zone (with
l being the number of inputs to each thermal zone, e.g. two
for air mass flow and supply air temperature). The HVAC
system of the building considered for this study operates with
a constant supply air temperature, while air mass flow is the
time varying control input. Hence, in the following simulations
we consider supply air temperature constant and treat air mass
flow as the control signal. Vector wt stores the estimated
disturbance values, aggregating various unmodelled dynamics
such as Tout , Q̇int and Qrad , and can be estimated using
historical data [23]. yt ∈ Rm is the output vector, representing
the temperature of the thermal zones. The building model
was trained using historical data, and the result of the system
identification is shown in Fig. 4.
B. MPC for Building Climate Control
We consider the problem of controlling the above building’s
HVAC system using an MPC scheme. We adopt the MPC
formulation proposed by Maasoumy et al. [24], with the
objective of minimizing the total energy cost (in dollar value).
τ and H denote the length of each time slot and the prediction
horizon (in number of time slots) of the MPC, respectively.
Assume that the system dynamics are also discretized with
a sampling time of τ . Here we consider τ = 0.5 hr and
H = 24. At each time t, the predictive controller solves an
optimal control problem to compute ~ut =P[ut , . . . , ut+H−1 ],
H−1
and minimizes the cumulative norm of ut : k=0 kut+k k. We
assume known an occupancy function occt which is equal to
1 when the room is occupied and to 0 otherwise. The purpose
of the MPC is to maintain a comfort temperature given by
T comf whenever the room is occupied while minimizing the
cost of heating. This problem can be expressed as follows:
min
~
ut
H−1
X
kut+k k s.t.
k=0
xt+k+1 = f (xt+k , ut+k , wt+k ),
xt |= ϕ with ϕ = [0,H] ((occt > 0) ⇒ (Tt > Ttcomf )
ut+k ∈ Ut+k , k = 0, ..., H − 1
The STL formula was encoded using the robust MILP encoding and results are presented in Fig. 5. Again we observed
that creating the MILP structure was longer than solving an
instance of it (4.1s versus 0.15s). However, by using a proper
parametrization of the problem in YALMIP, the creation of the
MILP structure can be done once offline and reused online for
each step of the MPC, which makes the approach promising
and potentially applicable even for real-time applications.
VIII. C ASE S TUDY II: R EGULATION C ONTROL FOR
S MART G RID
A. Mathematical Model
The second case study we consider is the n-areas smart
grid model presented in [16] and depicted in Fig. 6. The
interconnection of power system components, including a
governor, turbine and generator in each area is shown in the
block diagram in Fig. 7. In the diagram, δPC is a control
input which acts against an increase or decrease in power
demand to regulate the system frequency ω, and δPD denotes
fluctuations in power demand, modeled as an exogenous input
(disturbance). Under steady state, we have: ω = ωo and
o
o
PM = P G = P M
, where ωo , Vto , and PM
are the nominal
values for rated frequency, terminal voltage and mechanical
power input.
10
70
Temperature
Comfort Limit
69
68
67
66
6
12
1
0.5
0
18
24
Time of day (Hours)
Occupancy
6
12
18
60
24
Air Flow
Fig. 7. Block diagram of power system and its relation to governor, turbine,
generator, and the AGC signal for each control area. More details on the
power grid model can be found in [16].
40
where νij is the transmission line stiffness coefficient, and
the state variable xi8 is the integral of xi1 .
20
0
6
12
Time of day (Hours)
18
24
Fig. 5. Room temperature control with constraints based on occupancy,
expressed in STL.
Fig. 6. Power system grid with n areas. The dynamics in each area is depicted
in Fig.7.
Next, we present the mathematical model for one area i
(note that superscripts refer to the control area, and subscripts
index states in each area).
i
i
i
i
dxi1
(−D i xi1 + δPM
− δPD
− δPtie
+ δPanc
)
=
,
dt
M
i
i
i
i
i
i
i
(x3 − x2 )
(x4 − x3 )
(xi5 − xi4 )
dx2
dx3
dx4
=
,
=
,
=
,
dt
T7i
dt
T6i
dt
T5i
i
dxi5
(PGV
− xi5 )
=
,
dt
T4i
dxi7
dt
=
(−xi7
+
i
δPC
T1i
−
dxi6
(xi7 − xi6 )
=
,
dt
T3i
xi1 /Ri )
,
dxi8
dt
(8a)
(8b)
(8c)
i
= x1
x(tk+1 ) = Ax(tk ) + B1 uanc (tk ) + B2 w(tk ).
(10)
n T
1
. . . δPanc
] are the ancillary inputs,
Where uanc = [δPanc
and the exogenous inputs (i.e. disturbances or variations in
1
n T
demands) are denoted by w = [δPD
. . . δPD
] . We propose
controller synthesis for the ancillary services, complementing
the primary control of AGC.
(8d)
i
i
i
where δPM
and PGV
are given by δPM
= K1i xi5 + K3i xi4 +
i
i i
i i
K5 x3 + K7 x2 , and PGV = (1 − T2 /T3 )xi6 + (T2 /T3 )xi7 . D
is the damping coefficient, M is the machine inertia constant,
R is the speed regulation constant, Ti ’s are time constants
for power system components, and Ki ’s are fractions of total
mechanical power outputs associated with different operating
i
parts of the turbine. δPtie
represents power transfer from area
i to other areas. In equation (8), the first state represents the
i
frequency increment, xi1 = δωi . It can be shown that Ptie
can
be obtained from
n
X
ij
δPtie
=
νij (xi8 − xj8 ),
(9)
j=1
The classical automatic generation control (AGC) implements a simple PI control to regulate the grid frequency. In a
multi-area power system, in addition to regulating frequency
within each area, the auxiliary control should maintain the
net interchange power with neighboring areas at scheduled
values [25]. This is generally accomplished by adding a tieline flow deviation to the frequency deviation in the auxiliary
feedback control loop. A suitable linear combination of the
frequency and tie-line deviations for area i, is known as
the Area Control Error (ACE): this measures the difference
between the scheduled and actual electrical generation within
a control area while taking frequency bias into account. The
i
ACE of area i is thus defined as ACE i = δPtie
+β i xi1 , and β i
is the bias coefficient of area i. The standard industry practice
is to set the bias β i at the so-called Area Frequency Response
Characteristic (AFRC), which is defined as β i = Di + 1/Ri .
The integral of ACE is used to construct the speed changer
position feedback control signal (δPCi ), i.e., δPCi = −K i xi9 ,
dxi
where K i is the feedback gain and dt9 = ACE i .
The resulting state space model can be discretized and
written in compact form as
B. MPC for Ancillary Services
We require that uanc be bounded and satisfies a maximum
ramp constraint, i.e., uanc ≤ uanc (tk ) ≤ uanc with uanc >
0 and |uanc (tk+1 ) − uanc (tk )| ≤ λ, for some λ > 0. At each
time step k, we thus solve the following problem:
min
Uanc (k)
s.t.
J(ACE, Uanc )
(11)
x(tk+1 ) = Ax(tk ) + B1 uanc (tk ) + B2 w(tk )
uanc ≤ uanc (tk+j ) ≤ uanc
|uanc (tk+j+1 ) − uanc (tk+j )| ≤ λ
where Uanc (k) = (uanc (tk ), uanc (tk+1 ), . . . , uanc (tk+H )) is
the vector of inputs from k to k + H and H is the prediction
11
0.3
1
ACE , τ=5
0.2
ACE1, τ=10
0.1
0
−0.1
−0.2
−0.3
−0.4
0
10
20
30
40
50
60
70
80
0.8
90
100
u1anc, τ=5
0.6
u1anc, τ=10
0.4
0.2
0
−0.2
−0.4
−0.6
0
10
20
30
40
50
60
70
80
90
100
Fig. 8. Comparison of ACE1 (Area Control Error of Control Area 1)
stabilization using ϕt in (12) with τ = 5 and τ = 10. The controller enforces
the stabilization delay in both cases.
horizon. All the constraints of problem (11) that depend on j
should hold for j = 0, 1, . . . , H − 1.
The cost function proposed in [16] minimizes the `2 norm
of the ACE signal in areas i = 1, . . . , n, by exploiting the
ancillary service available in each area, while taking into
account the system dynamics and constraints. We propose to
constrain the ACE signal to satisfy a specified set of STL
properties, while minimizing the ancillary service used by
each
Thus we defined J(ACE, Uanc ) = kUanc k`2 =
P2 area.
PH−1
2
i
i=1
j=0 (Uanc [k + j]) , and an STL formula ϕ which
says that whenever |ACEi | is larger than 0.01, it should
become less than 0.01 in less than τ s. More precisely we
used ϕ = (ϕt ) with
1
| < .01)
2
(|ACE
| < .01)
[0,τ ]
[0,τ ] (|ACE
ϕt = ¬(|ACE1 | < .01)) ⇒ (
∧ (¬(|ACE2 | < .01)) ⇒ (
(12)
We encoded this formula and added the resulting constraints
to the MPC problem as described in the previous sections, and
solved it for different values of τ . Results are shown in Fig. 8,
and demonstrate that the STL constraint is correctly enforced
in the stabilization of the ACE signal.
IX. R ELATED W ORK
Receding horizon control for temporal logic has been considered before in the context of LTL [26], where the authors
propose a reactive synthesis scheme for specifications with
GR(1) goals. The authors in [27] also propose an MPC scheme
for specifications in synthetically co-safe LTL – our approach
extends synthesis capabilities to a wider class of temporal
logic specifications. In [28], the authors consider full LTL
but use an automata-based approach, involving potentially
expensive computations of a finite state abstraction of the
system and a Buchi automaton for the specification. We
circumvent these expensive operations using a bounded model
checking (BMC) approach to synthesis. In [17], the authors
present a model predictive control scheme to stabilize mixed
logical dynamical systems on desired reference trajectories,
while fulfilling propositional logic constraints and heuristic
rules. A major contribution of this work is to extend the
constraint specification language for such systems to STL
specifications, which allow expression of complex temporal
properties including safety, liveness, and response.
Our work extends the standard BMC paradigm for finite
discrete systems [14] to STL, which accommodates continuous
systems. In BMC, discrete state sequences of a fixed length,
representing counterexamples or plans, are obtained as satisfying assignments to a Boolean satisfiability (SAT) problem.
The approach has been extended to hybrid systems, either by
computing a discrete abstraction of the system [29], [30] or by
extending SAT solvers to reason about linear inequalities [31],
[32]. Similarly, MILP encodings inspired by BMC have been
used to generate trajectories for continuous systems with LTL
specifications [15], [33], [34], and for a restricted fragment
of MTL without nested operators [35]. While we draw much
inspiration from these early efforts, ours is the first work to
consider a BMC approach to synthesis for full STL.
X. C ONCLUDING R EMARKS
The main contribution of this paper is a pair of bounded
model checking style encodings for signal temporal logic
specifications as mixed integer linear constraints. We showed
how our encodings can be used to generate control for systems
that must satisfy STL properties, and additionally to ensure
maximum robustness of satisfaction. Our formulation of the
STL synthesis problem can be used as part of existing controller synthesis frameworks to compute feasible and optimal
controllers for cyber-physical systems. We presented experimental results for controller synthesis on simplified models of
a smart micro-grid and HVAC system, and showed how the
MPC schemes in these examples can be framed in terms of
synthesis from an STL specification, with simulation results
illustrating the effectiveness of our proposed synthesis.
We have demonstrated the ability to synthesize control for
systems on both the demand and supply sides of a smart grid.
We view this as progress toward a contract-based framework
for specifying and designing components of the smart grid
and their interactions using STL specifications. Future work
includes a reactive synthesis approach to synthesizing control
inputs for systems operating in uncertain environments: we
have already demonstrated preliminary results in this direction
in [36]. We will also further explore synthesis in an MPC
framework for unbounded STL properties. As mentioned in
Section V-B, this is an easy extension of our approach for
certain types of properties. Extending this to arbitrary properties has ties to online monitoring of STL properties [37],
which is another direction of further exploration.
R EFERENCES
[1] G. E. Fainekos, A. Girard, H. Kress-Gazit, and G. J. Pappas,
“Temporal logic motion planning for dynamic robots,” Automatica,
vol. 45, no. 2, pp. 343 – 352, 2009. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S000510980800455X
12
[2] P. Nuzzo, H. Xu, N. Ozay, J. Finn, A. Sangiovanni-Vincentelli, R. Murray, A. Donze, and S. Seshia, “A contract-based methodology for aircraft
electric power system design,” Access, IEEE, vol. PP, no. 99, pp. 1–1,
2013.
[3] R. Alur, T. A. Henzinger, G. Lafferriere, and G. J. Pappas, “Discrete
abstractions of hybrid systems,” Proceedings of the IEEE, vol. 88, no. 7,
pp. 971–984, 2000.
[4] R. M. Murray, J. Hauser, A. Jadbabaie, M. B. Milam, N. Petit, W. B.
Dunbar, and R. Franz, “Online control customization via optimizationbased control,” in In Software-Enabled Control: Information Technology
for Dynamical Systems. Wiley-Interscience, 2002, pp. 149–174.
[5] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert,
“Constrained model predictive control: Stability and optimality,”
Automatica, vol. 36, no. 6, pp. 789–814, 2000. [Online]. Available:
http://dx.doi.org/10.1016/S0005-1098(99)00214-9
[6] O. Maler and D. Nickovic, “Monitoring temporal properties of continuous signals,” in FORMATS/FTRTFT, 2004, pp. 152–166.
[7] G. E. Fainekos and G. J. Pappas, “Robustness of temporal
logic specifications for continuous-time signals,” Theor. Comput.
Sci., vol. 410, no. 42, pp. 4262–4291, 2009. [Online]. Available:
http://dx.doi.org/10.1016/j.tcs.2009.06.021
[8] A. Donzé and O. Maler, “Robust satisfaction of temporal logic over
real-valued signals,” in FORMATS, 2010, pp. 92–106.
[9] M. Maasoumy, P. Nuzzo, F. Iandola, M. Kamgarpour, A. SangiovanniVincentelli, and C. Tomlin, “Optimal load management system for
aircraft electric power distribution,” in IEEE Conference on Decision
and Control (CDC), 2013.
[10] A. Biere, K. Heljanko, T. A. Junttila, T. Latvala, and V. Schuppan,
“Linear encodings of bounded LTL model checking,” Logical Methods
in Computer Science, vol. 2, no. 5, 2006.
[11] V. Raman, M. Maasoumy, and A. Donzé, “Model predictive control
from signal temporal logic specifications: A case study,” in Proceedings
of the 4th ACM SIGBED International Workshop on Design, Modeling,
and Evaluation of Cyber-Physical Systems, ser. CyPhy’14. New
York, NY, USA: ACM, 2014, pp. 52–55. [Online]. Available:
http://doi.acm.org/10.1145/2593458.2593472
[12] V. Raman, M. Maasoumy, A. Donzé, R. M. Murray, A. SangiovanniVincentelli, and S. A. Seshia, “Model predictive control with signal
temporal logic specifications,” in Proc. of the IEEE Conf. on Decision
and Control, 2014.
[13] G. E. Fainekos and G. J. Pappas, “Robust sampling for MITL
specifications,” in Formal Modeling and Analysis of Timed Systems, 5th
International Conference, FORMATS 2007, Salzburg, Austria, October
3-5, 2007, Proceedings, 2007, pp. 147–162. [Online]. Available:
http://dx.doi.org/10.1007/978-3-540-75454-1 12
[14] A. Biere, A. Cimatti, E. M. Clarke, and Y. Zhu, “Symbolic model
checking without BDDs,” in TACAS, 1999, pp. 193–207.
[15] E. M. Wolff, U. Topcu, and R. M. Murray, “Optimization-based
trajectory generation with linear temporal logic specifications,” in 2014
IEEE International Conference on Robotics and Automation, ICRA
2014, Hong Kong, China, May 31 - June 7, 2014, 2014, pp. 5319–5325.
[Online]. Available: http://dx.doi.org/10.1109/ICRA.2014.6907641
[16] M. Maasoumy, B. M. Sanandaji, A. Sangiovanni-Vincentelli, and
K. Poolla, “Model predictive control of regulation services from commercial buildings to the smart grid,” in IEEE American Control Conference (ACC), 2014.
[17] A. Bemporad and M. Morari, “Control of systems integrating logic,
dynamics, and constraints,” Automatica, vol. 35, no. 3, pp. 407–427,
1999. [Online]. Available: http://dx.doi.org/10.1016/S0005-1098(98)
00178-2
[18] A. Donzé, T. Ferrère, and O. Maler, “Efficient robust monitoring for
stl,” in CAV, 2013, pp. 264–279.
[19] M. Morari and J. Lee, “Model predictive control: Past, present and
future,” Computers & Chemical Engineering, vol. 23, no. 4, pp.
667–682, 1999. [Online]. Available: https://control.ee.ethz.ch/index.cgi?
page=publications;action=details;id=1641
[20] A. Donzé, “Breach, a toolbox for verification and parameter synthesis
of hybrid systems,” in CAV, 2010, pp. 167–170.
[21] J. Lfberg, “YALMIP : A toolbox for modeling and optimization in
MATLAB,” in Proceedings of the CACSD Conference, Taipei, Taiwan,
2004. [Online]. Available: http://users.isy.liu.se/johanl/yalmip
[22] M. Maasoumy Haghighi, “Controlling energy-efficient buildings in
the context of smart grid: A cyber physical system approach,”
Ph.D. dissertation, University of California, Berkeley, Dec 2013.
[Online]. Available: http://www.eecs.berkeley.edu/Pubs/TechRpts/2013/
EECS-2013-244.html
[23] M. Maasoumy and A. Sangiovanni-Vincentelli, “Total and peak energy
consumption minimization of building hvac systems using model predictive control,” Design Test of Computers, IEEE, vol. 29, no. 4, pp. 26
–35, aug. 2012.
[24] M. Maasoumy, M. Razmara, M. Shahbakhti, and A. SangiovanniVincentelli, “Selecting building predictive control based on model uncertainty,” in IEEE American Control Conference (ACC 2014), Portland,
USA, June 2014.
[25] H. Bevrani, Robust Power System Frequency Control, ser. Power Electronics and Power Systems. Springer, 2009.
[26] T. Wongpiromsarn, U. Topcu, and R. M. Murray, “Receding horizon
temporal logic planning,” IEEE Trans. Automat. Contr., vol. 57, no. 11,
pp. 2817–2830, 2012.
[27] E. A. Gol and M. Lazar, “Temporal logic model predictive control
for discrete-time systems,” in Proceedings of the 16th international
conference on Hybrid systems: computation and control, HSCC 2013,
April 8-11, 2013, Philadelphia, PA, USA, 2013, pp. 343–352. [Online].
Available: http://doi.acm.org/10.1145/2461328.2461379
[28] X. C. Ding, M. Lazar, and C. Belta, “LTL receding horizon control for
finite deterministic systems,” Automatica, vol. 50, no. 2, pp. 399–408,
2014. [Online]. Available: http://dx.doi.org/10.1016/j.automatica.2013.
11.030
[29] G. J. P. Nicoló Giorgetti and A. Bemporad, “Bounded model checking
of hybrid dynamical systems,” in Decision and Control, 2005 and 2005
European Control Conference. CDC-ECC ’05. 44th IEEE Conference
on, Dec 2005, pp. 672–677.
[30] S. Jha, B. A. Brady, and S. A. Seshia, “Symbolic reachability analysis
of lazy linear hybrid automata,” in Proc. 5th International Conference
on Formal Modeling and Analysis of Timed Systems (FORMATS), ser.
Lecture Notes in Computer Science, vol. 4763, 2007, pp. 241–256.
[31] G. Audemard, M. Bozzano, A. Cimatti, and R. Sebastiani, “Verifying
industrial hybrid systems with MathSAT,” Electronic Notes in Theoretical Computer Science, vol. 119, no. 2, pp. 17 – 32, 2005, proceedings
of the 2nd International Workshop on Bounded Model Checking
(BMC 2004) Bounded Model Checking 2004. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1571066105000885
[32] M. Franzle and C. Herde, “Efficient proof engines for bounded
model checking of hybrid systems,” Electronic Notes in Theoretical
Computer Science, vol. 133, no. 0, pp. 119 – 137, 2005,
proceedings of the Ninth International Workshop on Formal Methods
for Industrial Critical Systems (FMICS 2004) Formal Methods
for Industrial Critical Systems 2004. [Online]. Available: http:
//www.sciencedirect.com/science/article/pii/S1571066105050279
[33] S. Karaman, R. G. Sanfelice, and E. Frazzoli, “Optimal control of mixed
logical dynamical systems with linear temporal logic specifications,” in
Decision and Control, 2008. CDC 2008. 47th IEEE Conference on, Dec
2008, pp. 2117–2122.
[34] Y. Kwon and G. Agha, “Ltlc: Linear temporal logic for control,” in
HSCC, M. Egerstedt and B. Mishra, Eds., 2008, pp. 316–329.
[35] S. Karaman and E. Frazzoli, “Vehicle routing problem with metric
temporal logic specifications,” in Proceedings of the 47th IEEE
Conference on Decision and Control, CDC 2008, December 9-11,
2008, Cancún, México, 2008, pp. 3953–3958. [Online]. Available:
http://dx.doi.org/10.1109/CDC.2008.4739366
[36] V. Raman, A. Donzé, D. Sadigh, R. M. Murray, and S. A. Seshia,
“Reactive synthesis from signal temporal logic specifications,” in Hybrid
Systems: Computation and Control, HSCC 2015, Seattle, WA, USA, April
14-16, 2015, 2015, pp. 239–248.
[37] J. V. Deshmukh, A. Donzé, S. Ghosh, X. Jin, G. Juniwal, and S. A.
Seshia, “Robust online monitoring of signal temporal logic,” in Runtime
Verification - 6th International Conference, RV 2015 Vienna, Austria,
September 22-25, 2015. Proceedings, 2015, pp. 55–70. [Online].
Available: http://dx.doi.org/10.1007/978-3-319-23820-3 4
Vasumathi Raman received the B.A. degree in
Computer Science and Mathematics from Wellesley
College in 2007 and the M.S. and Ph.D. degrees
in Computer Science from Cornell University in
2011 and 2013, respectively. She was a postdoctoral
scholar in the Department of Computing & Mathematical Sciences at the California Institute of Technology from 2013-2015, and is currently a Senior
Scientist at the United Technologies Research Center
in Berkeley, CA. Her research explores algorithmic
methods for designing and controlling autonomous
systems, guaranteeing correctness with respect to user-defined specifications.
13
physical systems and systems biology.
Mehdi Maasoumy received his PhD in Electrical
Engineering & Computer Sciences in 2013 from UC
Berkeley, where he was advised by Prof. Alberto
Sangiovanni-Vincentelli. He received his MSc degree in 2010 from Mechanical Engineering Department at UC Berkeley with major in Controls and
minor in Optimization with Designated Emphasis
in Energy Systems and Technology (DEEST). He
graduated form Sharif University of Technology in
2008 with a BSc degree in Mechanical Engineering.
Richard M. Murray received the B.S. degree in
Electrical Engineering from California Institute of
Technology in 1985 and the M.S. and Ph.D. degrees
in Electrical Engineering and Computer Sciences
from the University of California, Berkeley, in 1988
and 1991, respectively. He is currently the Thomas
E. and Doris Everhart Professor of Control & Dynamical Systems and Bioengineering at Caltech.
Murray’s research is in the application of feedback
and control to networked systems, with applications
in biology and autonomy. Current projects include
analysis and design biomolecular feedback circuits; specification, design and
synthesis of networked control systems; and novel architectures for control
using slow computing.
Alberto Sangiovanni Vincentelli holds the Edgar
L. and Harold H. Buttner Chair of Electrical Engineering and Computer Sciences at the University of
California at Berkeley. He has been on the Faculty of
the Department since 1976. He obtained an electrical
engineering and computer science degree (“Dottore
in Ingegneria”) summa cum laude from the Politecnico di Milano, Milano, Italy in 1971. In 19801981, he spent a year as a Visiting Scientist at the
Mathematical Sciences Department of the IBM T.J.
Watson Research Center. In 1987, he was Visiting
Professor at MIT. He has held a number of visiting professor positions at
Italian Universities, including Politecnico di Torino, Università di Roma, La
Sapienza, Università di Roma, Tor Vergata, Università di Pavia, Università
di Pisa, Scuola di Sant’Anna. He was awarded the IEEE/RSE Wolfson
James Clerk Maxwell Medal “for groundbreaking contributions that have
had an exceptional impact on the development of electronics and electrical
engineering or related fields” and the Kaufman Award of the Electronic Design
Automation Council for “pioneering contributions to EDA” . He is Honorary
Professor at Politecnico di Torino and has Honorary Doctorates from the
University of Aalborg and KTH. He helped founding Cadence and Synopsys,
the two leading companies in EDA. He is the author of over 850 papers, 18
books and 2 patents in the area of design tools and methodologies, large scale
systems, embedded systems, hybrid systems and innovation.
Alexandre Donzé is a research scientist at the
University of California, Berkeley in the department
of Electrical Engineering and Computer Science. He
received his Ph.D. degree in Mathematics and Computer Science from the University of Joseph Fourier
at Grenoble in 2007. He worked as a post-doctoral
researcher at Carnegie Mellon University in 2008,
and at Verimag in Grenoble from 2009 to 2012. His
research interests are in simulation-based design and
verification techniques using formal methods, Signal
Temporal Logic (STL) with applications to cyberSanjit A. Seshia received the B.Tech. degree in
Computer Science and Engineering from the Indian
Institute of Technology, Bombay in 1998, and the
M.S. and Ph.D. degrees in Computer Science from
Carnegie Mellon University in 2000 and 2005 respectively. He is currently an Associate Professor
in the Department of Electrical Engineering and
Computer Sciences at the University of California,
Berkeley. His research interests are in dependable
computing and computational logic, with a current
focus on applying automated formal methods to embedded and cyber-physical systems, electronic design automation, computer
security, and synthetic biology. He has served as an Associate Editor of the
IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Systems. His awards and honors include a Presidential Early Career Award
for Scientists and Engineers (PECASE) from the White House, an Alfred P.
Sloan Research Fellowship, the Prof. R. Narasimhan Lecture Award, and the
School of Computer Science Distinguished Dissertation Award at Carnegie
Mellon University.
| 3cs.SY
|
arXiv:1801.04086v1 [math.RA] 12 Jan 2018
Maximal and typical nonnegative ranks of
nonnegative tensors
Toshio SUMI∗, Mitsuhiro MIYAZAKI†and Toshio SAKATA‡
Abstract
Let N1 , . . . , Nd be positive integers with N1 ≤ · · · ≤ Nd . Set
N = N1 · · · Nd−1 . We show in this paper that an integer r is a typical
nonnegative rank of nonnegative tensors of format N1 × · · · × Nd if and
only if r ≤ N and r is greater than or equals to the generic rank of
tensors over C of format N1 ×· · ·×Nd . We also show that the maximal
nonnegative rank of nonnegative tensors of format N1 × · · · × Nd is N .
Keywords: tensor rank, typical rank, nonnegative rank, nonnegative
tensor
MSC:15A69, 14P10
1
Introduction
Recently high-dimensional array data are intensively studied in connection
with data analysis. In these areas of research, high-dimensional data are
called tensors. More precisely, a d-dimensional datum aligned in an N1 ×
· · · × Nd high-dimensional box form is called an N1 × · · · × Nd tensor or a
d-tensor of format N1 × · · · × Nd .
Let S be a set. The set of N1 ×· · ·×Nd tensors with entries in S is denoted
S N1 ×···×Nd . An element T ∈ S N1 ×···×Nd is denoted by T = (ti1 ···id )1≤j≤Nj or
simply T = (ti1 ···id ). The case where d = 2, S N1 ×N2 is nothing but the set of
N1 × N2 matrices with entries in S. If S is a field, the rank of an element
M ∈ S N1 ×N2 is defined in linear algebra. The rank of M is identical with
the minimum number r such that M can be expressed as a sum of r rank 1
matrices, where we set empty sum to be zero. Further, a nonzero matrix is
rank 1 if and only if it can be expressed as a product of a nonzero column
∗
Faculty of Arts and Science, Kyushu University, Fukuoka, Japan
Department of Mathematics, Kyoto University of Education, Kyoto, Japan
‡
Emeritus professor, Kyushu University, Fukuoka, Japan
†
1
vector and a nonzero row vector. We generalize the notion of rank to the
general case by using this characterization of ranks of matrices.
Suppose S is a subset of a commutative ring with identity which is
closed under addition and multiplication and contains 0 and 1. A nonzero
element T = (ti1 ···id ) ∈ S N1 ×···×Nd is defined to be S-rank 1 if there are
Q
vj = (vj1 , . . . , vjNj )⊤ ∈ S Nj for 1 ≤ j ≤ d such that ti1 ···id = dj=1 vjij for
any (i1 , . . . , id ). For a general element T ∈ S N1 ×···×Nd , we define the S-rank
of T as the minimum number r such that T can be expressed as a sum of r
S-rank 1 tensors. The S-rank of T is denoted rankS T . We attach the prefix
“S-” since the rank depends on S. In fact, there is an element of R2×2×2 such
that rankR T = 3 and rankC T = 2.
In this paper, we consider the case where S = R≥0 := {x ∈ R | x ≥
0}. Since it is important to work in the set of nonnegative real numbers
in the statistical point of view, R≥0 -rank is studied by several researchers
1 ×···×Nd
[ARSZ, QCL]. They called an element RN
a nonnegative tensor and
≥0
called the R≥0 -rank nonnegative rank. Allman et al. [ARSZ] showed that
under mild condition, nonnegative rank of a nonnegative tensor is identical
with its R-rank if its R-rank is less than or equals to 2.
1 ×···×Nd
We consider in this paper typical nonnegative ranks of RN
(see sec≥0
tion 2 for definition). Suppose that N1 ≤ · · · ≤ Nd and set N = N1 · · · Nd−1 .
We show that r is a typical nonnegative rank of nonnegative N1 × · · · × Nd
tensors if and only if r ≤ N and r is greater than or equals to the minimal
typical R-rank. In the way of proving this fact, we also show that N is the
maximal nonnegative rank of N1 × · · · × Nd nonnegative tensors.
2
Preliminaries
In this section, we fix notation, state definitions and recall some basic facts
in order to prepare for the main argument. Let F be the real number field R
or the complex number field C and let n be a positive integer. We denote by
Fn the vector space of n-dimensional column vectors with entries in F and by
ei the element of Fn whose i-th entry is 1 and other entries are 0. By abuse
of notation, we use the same symbol ei to express an element of Fn and Fm
even if n 6= m. The meaning will be clear from the context.
Let N1 , . . . , Nd be positive integers. Since an element FN1 ⊗F · · · ⊗F FNd
is expressed uniquely as
N1
X
i1 =1
···
Nd
X
ti1 ···id ei1 ⊗ · · · ⊗ eid ,
id =1
2
we identify FN1 ×···×Nd and FN1 ⊗F · · · ⊗F FNd by the correspondence
(ti1 ···id ) ↔
N1
X
i1 =1
···
Nd
X
ti1 ···id ei1 ⊗ · · · ⊗ eid .
id =1
We introduce the Euclidean topology to FN1 ×···×Nd by identifying FN1 ×···×Nd
with FN1 ···Nd .
Let S be R≥0 := {x ∈ R | x > 0}, R or C. We interpret F = R if S = R≥0
or S = R and F = C if S = C in the rest of this section. We define
ΦS1 : S N1 × · · · × S Nd → S N1 ×···×Nd
by ΦS1 ((v11 , . . . , v1N1 )⊤ , . . . , (vd1 , . . . , vdNd )⊤ ) = (v1i1 · · · vdid )1≤ij ≤Nj . Further,
we define for r ≥ 2, ΦSr : (S N1 × · · · × S Nd )r → S N1 ×···×Nd by ΦSr (a1 , . . . , ar ) =
ΦS1 (a1 ) + · · · + ΦS1 (ar ). We define the S-rank of a nonzero element T of
S N1 ×···×Nd is r if T ∈ ImΦSr \ ImΦSr−1 , where we interpret ImΦS0 to be the set
consists of zero element only, and we define the S-rank of the zero element
of S N1 ×···×Nd to be 0. If the set of elements S N1 ×···×Nd whose S-rank is r
contains a nonempty Euclidean open subset of FN1 ×···×Nd , we say that r is a
typical S-rank of N1 × · · · × Nd tensors.
It is known that there exists a unique typical C-rank of N1 × · · · × Nd
tensors over C which is called the generic rank of N1 × · · · × Nd tensors over
C and is denoted grankC (N1 , . . . , Nd ). Further, it is known that the generic
rank of N1 × · · · × Nd tensors over C is identical with the minimal typical
R-rank of N1 × · · · × Nd tensors. See [BT, Theorem 2] or [SSM, Chapter 6].
In this paper, we consider the case where S = R≥0 . In this case an
element of S N1 ×···×Nd is called a nonnegative tensor of format N1 × · · · × Nd
or an N1 × · · · × Nd nonnegative tensor, an S-rank is called a nonnegative
rank and a typical S-rank is called a typical nonnegative rank. In the case
where d = 2, i.e. the matrix case is studied extensively as a nonnegative
matrix factorization.
1 0 1 0
0 1 0 1
Example 2.1 The matrix
1 0 0 1 has R-rank 3 and nonnegative
0 1 1 0
rank 4. In fact, the column vectors of this matrix span the convex polyhedral cone {(x1 , x2 , x3 , x4 )⊤ ∈ R≥0 | x1 + x2 − x3 − x4 = 0}. See [Sta,
Example 3.6].
3
3
Main result
In this section, we study maximal and typical nonnegative ranks of N1 ×· · ·×
Nd nonnegative tensors with N1 ≤ · · · ≤ Nd . Set N = N1 · · · Nd−1 . First we
state the following upper bound of nonnegative ranks of nonnegative tensors
of format N1 × · · · × Nd .
Lemma 3.1 Let T be a nonnegative tensor of format N1 × · · · × Nd . Then
rankR≥0 T ≤ N.
Proof Set T = (ti1 ···id ). Then
T =
N1
X
i1 =1
Nd−1
···
X
ei1 ⊗ · · · ⊗ eid−1 ⊗ (ti1 ···id−1 1 , . . . , ti1 ···id−1 Nd )⊤ .
id−1 =1
Thus, rankR≥0 T ≤ N.
Next we show the following.
Lemma 3.2 N is a typical nonnegative rank of nonnegative tensors of format N1 × · · · × Nd .
Proof Set I = {(i1 , . . . , id ) | 1 ≤ ij ≤ Nj for
P 1 ≤ j ≤ d and i1 + · · · + id ≡ 0
(mod Nd )}. Then #I = N. Set also T0 = (i1 ,...,id )∈I ei1 ⊗ · · · ⊗ eid and ǫ =
1
1 ×···×Nd
. We show that if T ∈ RN
and ||T − T0 || < ǫ, then rankR≥0 T = N,
≥0
3N
where ||T − T0 || denotes the Euclidean norm.
Assume the contrary and set r = rankR≥0 T . Then by Lemma 3.1, we
P
N
see that r < N. Further set T = rk=1 a1k ⊗ · · · ⊗ adk , where ajk ∈ R≥0j
for any 1 ≤ j ≤ d and 1 ≤ k ≤ r. Set ajk = (ajk1 , . . . , ajkNj )⊤ and for
(i1 , . . . , id ) ∈ I, we set
K(i1 ,...,id ) := {k | 1 ≤ k ≤ r, ajkl < ajkij for any 1 ≤ j ≤ d and l 6= ij .}
It is clear that if (i1 , . . . , id ) 6= (ı′1 , . . . , ı′d ), then K(i1 ,...,id ) ∩ K(ı′1 ,...,ı′d ) = ∅.
Suppose that (i1 , . . . , id ) ∈ I and k0 6∈ K(i1 ,...,id ) . Then there are j and
l with ij 6= l and ajk0 l ≥ ajk0 ij . Since (i1 , . . . , ij−1 , l, ij+1 , . . . , id ) 6∈ I by the
definition of I, we see that (i1 , . . . , ij−1 , l, ij+1 , . . . id ) entry of T0 is 0. Thus,
(i1 , . . . , ij−1, l, ij+1 , . . . id ) entry of T is less than ǫ. Therefore,
a1k0 i1 · · · ajk0 ij · · · adk0 id ≤ a1k0 i1 · · · ajk0 l · · · adk0 id < ǫ,
since ajki ≥ 0 for any j, k and i.
4
P
Set T ′ = k6∈K(i ,...,i ) a1k ⊗ · · · ⊗ adk . Then, by the above argument, we
1
d
see that the (i1 , . . . , id ) entry of T ′ is less than ǫ(r − #K(i1 ,...,id ) ) < ǫN. On
the other hand, since (i1 , . . . , id ) entry of T0 is 1, and ||T − T0 || < ǫ, we see
1
that (i1 , . . . , id ) entry of T is greater than 1 − ǫ. Since ǫ = 3N
, we see that
′
1 − ǫ > ǫN and therefore, T 6= T , i.e. K(i1 ,...,id ) 6= ∅. Thus, we see that
N = #I ≤ r since K(i1 ,...,id ) ∩ K(ı′1 ,...,ı′d ) = ∅ if (i1 , . . . , id ) 6= (ı′1 , . . . , ı′d ). This
contradicts to the assumption that r < N.
By Lemmas 3.1 and 3.2, we see the following fact.
Theorem 3.3 The maximal nonnegative rank of nonnegative tensors of format N1 × · · · × Nd is N.
Next we consider the set of typical nonnegative ranks of nonnegative
tensors of format N1 × · · · × Nd . First we recall the following fact.
Lemma 3.4 (see e.g. [SSM, Chapter 6]) An integer r is greater than
or equals to grankC (N1 , . . . , Nd ) if and only if the Jacobian of ΦCr is row
full rank for some (a1 , . . . , ar ) ∈ (CN1 × · · · × CNd )r . In particular, if
r ≥ grankC (N1 , . . . , Nd ), then there is a nonzero maximal minor of Jacobian of ΦCr as a polynomial.
Note that every Jacobian of ΦCr as a polynomial is a polynomial with coefficients in Z. We also note the following basic fact.
Lemma 3.5 Let R be an integral domain, X1 , . . . , Xn indeterminates and
A1 , . . . , An infinite subsets of R. Then for any nonzero polynomial f ∈
R[X1 , . . . , Xn ], there are a1 , . . . , an with aj ∈ Aj for 1 ≤ j ≤ n such that
f (a1 , . . . , an ) 6= 0.
The following fact is easily verified.
Lemma 3.6 Let T be a nonnegative tensor with nonnegative rank s. Suppose
T = T1 + · · · + Ts with rankR≥0 Tj = 1 for 1 ≤ j ≤ s, then
rankR≥0 (T1 + · · · + Tr ) = r
for any 1 ≤ r ≤ s.
We also note the following fact.
Lemma 3.7 Let r be a typical nonnegative rank of nonnegative tensors of
format N1 × · · · × Nd . Then r ≥ min(trankR (N1 , . . . , Nd )).
5
Proof By the definition of typical nonnegative rank, there is a nonempty
1 ×···×Nd
Euclidean open subset U of RN
such that rankR≥0 T = r for any T ∈ U.
≥0
Since rankR T ≤ rankR≥0 T , U consists of tensors whose R-rank is less than
or equals to r. Therefore, r ≥ min(trankR (N1 , . . . , Nd )).
Now we state the following.
Theorem 3.8 The set of typical nonnegative ranks of nonnegative tensors
of format N1 × · · · × Nd is {r | r ∈ Z, grankC (N1 , . . . , Nd ) ≤ r ≤ N}.
Proof Let r be a typical nonnegative rank of nonnegative tensors of format
N1 × · · · × Nd . By Lemma 3.1, we see that r ≤ N. Further, by Lemma 3.7,
r ≥ min(trankR (N1 , . . . , Nd )) = grankC (N1 , . . . , Nd ).
Conversely assume that grankC (N1 , . . . , Nd ) ≤ r ≤ N. By Lemma 3.4, we
see that there is a maximal minor f of the Jacobian of ΦCr which is nonzero as
a polynomial. Since N is a typical nonnegative rank of nonnegative tensors
of format N1 × · · · × Nd , we see that there is a Euclidean open subset O of
1 ×···×Nd
RN
such that rankR≥0 T = N for any T ∈ O.
≥0
Since
R
Nd N
N1 ×···×Nd
1
ΦN≥0 : (RN
≥0 × · · · × R≥0 ) → R≥0
is continuous and identical with
Nd r
Nd N −r
N1
1 ×···×Nd
1
(RN
→ RN
,
≥0 × · · · × R≥0 ) × (R≥0 × · · · × R≥0 )
≥0
R
R
(A1 , A2 ) 7→ Φr ≥0 (A1 ) + ΦN≥0
−r (A2 ), we see that there are open subsets U1 of
R
Nd r
Nd N −r
N1
1
(R≥0 × · · · × R≥0 ) and U2 of (RN
such that Φr ≥0 (U1 ) +
≥0 × · · · × R≥0 )
R
ΦN≥0
−r (U2 ) ⊂ O.
By Lemma 3.5, we see that there exists A1 ∈ U1 such that f (A1 ) 6= 0.
R
Therefore, by the inverse function theorem, we see that ImΦr ≥0 (U1 ) contains
R
an open neighborhood O ′ of Φr ≥0 (A1 ). On the other hand, by Lemma 3.6,
R
we see that every element of ImΦr ≥0 (U1 ) has nonnegative rank r. Thus, we
see that every element of O ′ has nonnegative rank r. Therefore, r is a typical
nonnegative rank.
Remark 3.9 The fact that the set of typical nonnegative ranks form a consecutive integers can also be proved along the same line with [BBO, Theorem
2.2].
6
References
[ARSZ] Elizabeth S. Allman, John A. Rhodes, Bernd Sturmfels and Piotr
Zwiernik, Tensors of nonnegative rank two, Linear Algebra and its
Applications 473 (2015) 37–53.
[BT]
Grigoriy Blekherman and Zach Teitler, On maximum, typical and
generic ranks. Mathematische Annalen 362 1021-1031, (2015).
[BBO] Alessandra Bernardi, Grigoriy Blekherman and Giorgio Ottaviani,
On real typical ranks, arXiv:1512.01853v1 (2015).
[QCL]
Yang Qi, Pidrre Comon and Lek-Heng Lim, Semialgebraic geometry
of nonnegative tensor rank, SIAM J. Matrix Anal. Appl. 37 (2016)
1556–1580.
[SSM]
Toshio Sakata, Toshio Sumi and Mitsuhiro Miyazaki, Algebraic and
Computational Aspects of Real Tensor Ranks, Springer, (2016)
[Sta]
Richard P. Stanley, Combinatorics and commutative algebra. Second edition. Progress in Mathematics, 41 Birkhauser Boston, Inc.,
Boston, MA, (1996)
7
| 0math.AC
|
Perfectly capturing traveling single photons of arbitrary
temporal wavepackets with a single tunable device
arXiv:1609.05643v1 [quant-ph] 19 Sep 2016
Hendra I. Nurdin∗, Matthew R. James†, and Naoki Yamamoto‡
Abstract
We derive the explicit analytical form of the time-dependent coupling parameter to an external field for perfect absorption of traveling single photon fields with
arbitrary temporal profiles by a tunable single input-output open quantum system,
which can be realized as either a single qubit or single resonator system. However, the
time-dependent coupling parameter for perfect absorption has a singularity at t = 0
and constraints on real systems prohibit a faithful physical realization of the perfect
absorber. A numerical example is included to illustrate the absorber’s performance
under practical limitations on the coupling strength.
1
Introduction
Quantum networks composed of open quantum systems as network nodes that are interconnected by traveling quantum optical fields are of much interest for quantum information
applications, including quantum communication and quantum computing; see, e.g., [1, 2]
and the references therein. Photons in optical fields are ideal particles for transmitting information between nodes in a quantum network as they can propagate through free space
and various physical media. Quantum systems at the nodes are typically matter systems
such as an atom, atomic ensembles, superconducting circuits, amongst many possibilities.
Each node can process or store classical or quantum information and exchange its information content with an optical field, either to receive information contained in the field or to
transmit information into the field. Thus an important problem for quantum networks is
the problem of state transfer between an optical field impinging on a node and the matter
system at the node. That is, the transfer of the quantum state of the optical field to the
matter and vice-versa.
∗
H. I. Nurdin is with the School of Electrical Engineering and Telecommunincations, UNSW Australia,
Sydney NSW 2052, Australia (email: [email protected]). Research was supported by the Australian
Research Council.
†
M. R. James is with the School of Engineering, The Australian National University, Canberra ACT
0200, Australia (email: [email protected]). Research was supported by the Australian Research
Council.
‡
N. Yamamoto is with the Department of Applied Physics and Physico-Informatics, Keio University,
Hiyoshi 3-14-1, Kohoku, Yokohama, Japan (email: [email protected]). Research was supported
by JSPS Grant-in-Aid No. 15K06151.
1
The problem of state transfer in a quantum network with two nodes was first considered
in the seminal work [3]. In this work, the authors consider the perfect transfer of an
arbitrary superposition state α|0i + β|1i from a qubit on one node to a qubit on the other
via a one way optical field connecting the two nodes. Each node is a cavity QED system
that can be modelled as a qubit coupled to an optical cavity which is in turn coupled to the
optical field. It was shown that perfect transfer can be achieved by suitably modulating the
Rabi frequency and phase of a Raman laser driving the qubits at each node. They employ
the quantum filtering equation or quantum trajectory equation [4, 5] for the network and
apply the dark state principle to derive differential equations characterising the required
modulation. The dark state principle states that there should be no photon counted at the
output optical field reflected from the receiving node when a photon counting measurement
is performed on that output field.
In this paper, we are interested in the problem of the perfect transfer of a traveling
single photon field into a system that can function as a perfect single photon absorber.
The state of a travelling single photon field is characterized
by a temporal wavepacket
R∞
ξ, a complex-valued function of time, that satisfies 0 |ξ(s)|2 ds = 1. The wavepacket
gives the detection probability of the single photon state, that is, the probability that the
photon will be detected (say, byR registering a click in a photo detector) in a time interval
t
[t1 , t2 ] (0 ≤ t1 < t2 ) is given by t12 |ξ(s)|2 ds. We are interested in a tunable single photon
absorber with a coupling parameter that can be modulated to absorb single photons of any
wavepacket shape.
In [6] it was shown how single photon fields with symmetric temporal profiles can be
mode-matched to be absorbed by a coupled cavity-oscillator system into a single photon
state of the oscillator, by suitably modulating the coupling parameter between the cavity
and oscillator modes. In the work [7], it was shown that an absorber can be implemented
with a cavity QED system composed of a three level atom coupled to an optical cavity.
Perfect absorption of a photon with an arbitrary temporal shape could be achieved by appropriately modulating the amplitude of a laser beam driving a Raman transition process
in the atom, under the assumption of no spontaneous emission from the atom. The modulation of the laser beam is tailored to the temporal profile of the incoming single photon field.
In this paper, we consider a different class of single photon absorbers, a single two-level
system (qubit) or, equivalently, a single resonator with a tunable coupling parameter to an
external optical field. This type of coupling has been theoretically proposed for the transfer
of state between two microwave resonators connected by a one-way travelling optical field
between the resonators [8]. Importantly, such a tunable coupling parameter has already
been demonstrated experimentally in a microwave superconducting resonator [9], where
the tunability is realized using an externally controlled variable inductance. This coupling
is analogous to a mirror with tunable transmissivity on an optical cavity. The contribution of this work is to analytically derive the exact form for the time-dependent coupling
parameter for perfect absorption of a single photon with an arbitrary wavepacket shape.
We describe the system with a QSDE and derive the optimal evolving coupling parameter
by two methods: by explicitly solving the QSDE, and by application of the zero-dynamics
principle from [10]. We note that a form of this principle had also been employed in earlier
works [6, 8, 7] without being referred to as such. That is, these works require that the
incoming field and the field reflected from the system interfere destructively, resulting in
no photon in the output field (zero output dynamics).
2
This paper is organized as follows. Section 2 introduces the notation of the paper
and gives a brief overview of quantum stochastic calculus, quantum stochastic differential
equations, single photon generators, and systems driven by a single photon field. Section
3 introduces the model of the single photon absorber of interest, analytically derives the
modulating function for the coupling parameter for perfect absorption of any single photon
field, and develops a numerical example showing the effects of practical limitations on the
absorber’s performance. Finally, Section 4 draws a conclusion for the paper.
2
Preliminaries
√
Notation. We will use ı = −1, ∗ to denote the adjoint of a linear operator as well as
the conjugate of a complex number. If A = [ajk ] then A# = [a∗jk ], and A† = (A# )> , where
(·)> denotes matrix transposition. <{A} = (A + A# )/2 and ={A} = 2ı1 (A − A# ). We
use R+ to denote the set of non-negative real numbers and L2 (R+ ; C) to denote the set of
square-integrable functions on R+ . hXi denotes the quantum expectation of an operator
X, and Tr(X) denotes the trace of X.
2.1
Quantum stochastic calculus and quantum stochastic differential equations
We will be working with open Markov quantum systems that are coupled to n continuousmode boson fields indexed by j = 1, 2, . . . , m with annihilation field operators ηj (t) satisfying the field commutation relations [ηj (t), ηk (t0 )∗ ] = δjk δ(t − t0 ) and [ηj (t), ηk (t0 )] = 0. For
our purposes, we can focus on fields
R t in a vacuum state. Let us introduce the integrated
field annihilation process Aj (t) = 0 ηj (s)ds and its adjoint process, the integrated field
Rt
creation process, A∗j (t) = 0 ηj∗ (s)ds. In the vacuum representation, their future-pointing
Itō increments dAj (t) = Aj (t + dt) − Aj (t) and dA∗j (t) = A∗j (t + dt) − A∗j (t) satisfy the
quantum Itō table [11, 12, 13]
×
dA∗k dAk
dAj δjk dt 0
dA∗j 0
0
We may also define the counting process (or gauge process)
Z t
Λjk (t) =
b∗j (r)bk (r)dr,
0
which may be included in the Itō table [11]. The additional non-trivial products of differentials are
dΛjk dA∗l = δkl dA∗j , dAj dΛkl = δjk dAl , dΛjk AΛli = δkl dΛji .
Using the processes A = (A1 , A2 , . . . , Am )> , A# = (A∗1 , A∗2 , . . . , A∗m )> and Λ = [Λjk ]j,k=1,...,m ,
one may define quantum stochastic integrals of adapted processes on the tensor product
of the system and joint boson (symmetric) Fock space of the fields. The system is the
quantum mechanical object that is being coupled to the fields, and adapted means that
at time t the process acts trivially on the portion of the boson Fock space after time t ,
see, e.g., [11, 12, 13] for details. An adapted process commutes at time t with all of the
3
future pointing differentials. The product of two adapted processes X(t) and Y (t) is again
adapted, and the increment of the product obeys the quantum Itō rule
d(X(t)Y (t)) = (dX(t))Y (t) + X(t)dY (t) + dX(t)dY (t),
Based on these quantum stochastic integrals, one may define quantum stochastic differential equations (QSDEs).
For the remainder of the paper, we will only consider a system coupled to a single
field, m = 1 and A(t) = A1 (t), A∗ (t) = A∗1 (t), Λ(t) = Λ11 (t). An important QSDE that
describes the joint unitary evolution of an open Markov system coupled to vacuum boson
fields, common in quantum optics and related fields, is the Hudson-Parthasarathy QSDE
given by
dU (t) = (−(ıH(t) + 1/2L(t)∗ L(t))dt + dA(t)∗ L(t) − L(t)∗ S(t)dA(t) + (S(t) − I)dΛ(t))U (t),
(1)
with initial condition U (0) = I. We adopt a very general setting [14] where H(t) = H(t)∗
can be a general adapted process representing the time-dependent Hamiltonian of the
system, L(t) an adapted process representing the (possibly time-dependent) coupling of
the system to the creation process A∗ (t), and S(t) a unitary adapted process, S(t)∗ S(t) =
S(t)S(t)∗ = I, representing the (possibly time-dependent) coupling of the system to the
process Λ of the field. This QSDE has a unique solution whenever S(t), L(t), H(t) are
bounded operators for all t ≥ 0 [14]. Moreover, in that case the solution is guaranteed to
be unitary. The input field A(t) after the interaction is transformed into the output field
Y (t) = U (t)∗ A(t)U (t). Let jt (X) = U (t)∗ XU (t) denote the Heisenberg picture evolution
of a system operator X, then we have the following QSDEs for the Heisenberg picture
evolution of jt (X) and Y (t) [15],
djt (X) = jt (Lt (X)) + dA(t)∗ jt (S(t)∗ [X, L(t)]) + jt ([L(t)∗ , X]S(t))dA(t)
+jt (S(t)∗ XS(t) − X)dΛ(t),
dY (t) = jt (L(t)) + jt (S(t))dA(t),
with
Lt (X) = 1/2L(t)∗ [X, L(t)] + 1/2[X, L(t)]L(t) − ı[X, H(t)].
We will denote an input-output open quantum system G with parameters S(t), L(t), H(t)
with the notation G = (S(t), L(t), H(t)) [15]. The output of system G1 = (S1 (t), L1 (t), H1 (t))
can be passed as the input to a second system G2 = (S2 (t), L2 (t), H2 (t)) forming a cascaded
system which is again an input-output open quantum system G3 , the parameters of which
is given by the series product rule /,
G3 = G2 / G1 = (S2 (t)S1 (t), L2 (t) + S2 (t)L1 (t), H1 (t) + H2 (t) + ={L2 (t)∗ S2 (t)L1 (t)}).
We note that G1 and G2 can be subsystems of one system sharing the same Hilbert space,
see [15] for details. The series product is associative, thus if Gn = (Sn (t), Ln (t), Hn (t))
then the series product Gn / Gn−1 / . . . / G1 is unambiguously defined.
4
2.2
Single photon generator
Our main interest is in systems driven by a traveling single photon field with wavepacket
ξ, defined by a continuous-mode single photon state |1ξ i on the boson Fock space, given by
Rt
Rt
|1ξ i = 0 ξ(s)η(s)∗ ds |Ωi = 0 ξ(s)dA∗ (s)|Ωi [16, 17, 18, 19], where |Ωi denotes the vacuum
state of the field.
In [20, 19] it has been shown that |1ξ i can be generated as the output of a two-level
system (qubit) that is coupled to single vacuum input field through a modulated coupling
coefficient. The generator G1 has the Hilbert space H1 = C2 , where we take as basis
vectors |0i1 = (0, 1)> and |1i1 = (1, 0)> , and is given by G1 = (I, λ(t)σ1,− , 0), where σ1,−
is the qubit’s lowering operator,
0 0
σ1,− = |0i1 h1| =
1 0
and λ(t) is a time-dependent complex coupling coefficient of the system to the field given
by
ξ(t)
λ(t) = p
,
w(t)
(2)
R∞
with w(t) = t |ξ(s)|2 ds. The coupling coefficient depends on the wavepacket ξ of the
single photon field that one wishes to generate. Let U1 (t) denote the unitary solution to
the generator’s QSDE. When the generator is initialized in the state |1i1 then we have
[20, 19],
Z t
p
ξ(s)dA∗ (s)|Ωi
U1 (t)|1i1 |Ωi = w(t)|1i1 ⊗ |Ωi + |0i1 ⊗
0
and so limt→∞ U1 (t)|1i|Ωi = |0i ⊗ |1ξ i. Thus, if a photon detection measurement is performed at the output of the generatorR at any time t ≥ 0, the probability of detecting a
t
photon in the interval [0, t] is exactly 0 |ξ(s)|2 ds, as expected.
2.3
Systems driven by a single photon
Systems driven by a traveling single photon field with wavepacket ξ can be viewed equivalently as being driven by an ancillary generator that outputs the single photon field, as
described in the preceding subsection. Suppose that a system G2 = (S2 (t), L2 (t), H2 (t))
on the Hilbert space H2 is driven by a single photon field with wavepacket ξ. Then the
dynamics of the system can be analyzed by studying the equivalent cascaded system
G2 / G1 = (S2 (t), L2 (t) + λ(t)S2 (t)∗ σ1,− , H(t) + ={λ(t)L2 (t)∗ S2 (t)σ1,− })
that is driven by a vacuum field. The cascaded system is defined on the composite Hilbert
space C2 ⊗ H2 . Note that for notational simplicity for an arbitrary operator X1 on the
generator and X2 on the system, we will often denote the ampliations X1 ⊗ I, I ⊗ X2 , and
X1 ⊗ X2 simply as X1 , X2 , and X1 X2 , etc, with the tensor product being omitted.
5
3
A tunable single photon absorber with time-dependent
coupling to an external field
We now describe our tunable photon absorber that can perfectly absorb a single photon
field with an arbitrary wavepacket. The Hilbert space of the absorber is that of a qubit,
but since only a single excitation will be involved (i.e., a single photon) it can actually be
implemented by a quantum harmonic oscillator with a tunable coupling parameter such as
the microwave system realized in [7]. We will say more about the physical realization of
the absorber near the end of this section.
We choose |0i2 = (0, 1)> and |1i2 = (1, 0)> as basis vectors for the absorber Hilbert
space H2 = C2 . Our perfect single photon absorber is then a system G2 = (I, γ(t)σ2,− , 0),
with σ2,− the qubit lowering operator on H2 . Note that the absorber Hamiltonian has been
set to 0, so we are implicitly working in a rotating frame with the respect to the absorber’s
original Hamiltonian. We only require that the original Hamiltonian commutes with σ2,−
or only rotates the latter in time as e−ıωt σ2,− for some real constant ω ( e−ıωt can then
be absorbed into γ(t)), but is otherwise arbitrary. In principle, the analysis to follow can
be adapted to accommodate more general Hamiltonians but we do not pursue this here.
The time-dependent parameter γ will be chosen to enable perfect absorption of any single
photon field, with γ dependent on the wavepacket ξ of the latter. Let G1 be the ancillary
generator for a single photon as given in Section 2.2 for a given ξ. The cascade of G1 onto
G2 is given by:
G2 / G1 = (I, λ(t)σ1,− + γ(t)σ2,− , ={γ(t)∗ λ(t)σ2,+ σ1,− }),
(3)
∗
where σj,+ = σj,−
denotes the raising operator for the system Gj . The generator G1
is initialized in the excited state |1i1 while the absorber G2 is initialized in the ground
state |0i2 . The field to which the cascade is coupled to is in the vacuum state |Ωi. We
shall show that for any wavepacket ξ(t), γ(t) can always be chosen such that in the limit
t → ∞, the joint system and field state converges to |0i1 |1i2 |Ωi. That is, a single photon
is asymptotically perfectly absorbed by G2 . We shall derive the form for γ explicitly in
two ways: (i) by explictly solving the QSDE for the cascaded system in Section 3.1, and
(ii) by application of the zero-dynamics principle in Section 3.2.
It will turn out that the coupling parameter γ will have a singularity at t = 0 for
any wavepacket ξ. In actual practical implementation this singularity must be truncated
to some finite value. We propose a truncation in Section 3.3 and apply it in a numerical
example for an exponentially decaying wavepacket to give an illustration of the effect of this
truncation on the absorber’s performance in capturing a single photon with this wavepacket
shape.
Coming back to the discussion at the beginning of this section, we note that since the
absorber can have no more than a single quanta of excitation when driven by a single
photon field, the dynamics of G2 when initialized at |0i2 can be embedded in the dynamics
of an equivalent quantum harmonic oscillator system G02 = (I, λ(t)a, 0), where a is the
annihilation operator of the oscillator, satisfying the commutation relation [a, a∗ ] = 1.
This is because H2 is isomorphic to span{|0io , |1io }, where |0io and |1io is the vacuum
and 1-photon Fock state of the oscillator, respectively, and σ2,− can be identified with the
restriction of a to span{|0io , |1io }. Indeed, this kind of embedding of the dynamics of a
6
qubit system into a quantum harmonic oscillator under single photon driving has already
been exploited in [21].
3.1
Determining γ by solving the QSDE for G2 / G1
The wavefunction of the cascaded system (generator and absorber) and the external field is
initially |1i|0i|Ωi and at time t ≥ 0 is of the form ψ(t) = (ψ1 (t), ψ2 (t), ψ3 (t), ψ4 (t))> ⊗|Ω[t i,
where ψ(t) is adapted on the joint cascaded system and Fock space, in the sense that ψ(t)
lives in the tensor product of the system and the portion of the Fock space up to time
t. That is, U (t)|1i|0i|Ωi = ψ(t) ⊗ |Ω[t i, where U (t) is a solution of the QSDE (1) with
coefficients given by (3). Here |Ωt] i and |Ω[t i denote the portion of Fock vacuum from time
0 up to time t, and from time t onwards, respectively, see [11, 12]. We have that the vector
ψ(t) satisfies the QSDE:
−(|λ(t)|2 + |γ(t)|2 )/2
0
0
0
0
−|λ(t)|2 /2
0
0
ψ(t) ⊗ |Ω[t idt
d(ψ(t) ⊗ |Ω[t i) =
∗
2
0
−γ(t) λ(t) −|γ(t)| /2 0
0
0
0
0
0
0
0 0
γ(t) 0
0 0
ψ(t) ⊗ dA(t)∗ |Ω[t i,
+
λ(t) 0
0 0
0 λ(t) γ(t) 0
with initial condition ψ(0) = |1i1 |0i2 = (0, 1, 0, 0)> . From the above equation one readily
gets that
Z t
2
2
1
ψ1 (t) = exp − /2 (|λ(s)| + |γ(s)| )ds ψ1 (0)|Ωt] i
0
= 0,
for all t ≥ 0. Using this fact and recalling from [19] (using (2)) that
sR ∞
Z t
|ξ(τ )|2 dτ
,
exp −1/2
|λ(τ )|2 dτ = Rt∞
|ξ(τ )|2 dτ
s
s
for 0 ≤ s ≤ t, we can also solve for ψ2 (t), obtaining
Z t
2
1
|λ(τ )| dτ ψ2 (0)|Ωt] i
ψ2 (t) = exp − /2
0
sZ
∞
|ξ(τ )|2 dτ |Ωt] i.
=
t
We want that as t → ∞, ψ(t) → |0i1 |1i2 |Ωi = (0, 0, |Ωi, 0)> . That is, G2 completely
absorbs the single photon emitted by G1 , so that the output field state goes to the vacuum.
From the equation for ψ4 (t) we have that
Z t
(λ(s)ψ2 (s)dA(s)∗ + γ(t)ψ3 (s)dA(s)∗ )|Ωt] i,
ψ4 (t) = ψ4 (0)|Ωt] i +
0
Z t
=
(ξ(s)dA(s)∗ + γ(s)ψ3 (s)dA(s)∗ )|Ωt] i,
0
7
which we want to go to 0 as t → ∞. Also, since
dψ3 (t) = −1/2|γ(t)|2 ψ3 (t)dt − γ(t)∗ λ(t)ψ2 (t)dt,
we can solve this to give:
Z t
2
1
ψ3 (t) = exp − /2
|γ(s)| ds ψ3 (0)|Ωt] i
0
Z t
Z t
2
exp −1/2
|γ(τ )| dτ γ(s)∗ λ(s)ψ2 (s)ds)|Ωt] i,
−
0
s
Z t
Z t
2
exp −1/2
|γ(τ )| dτ γ(s)∗ λ(s)ψ2 (s)ds)|Ωt] i,
= −
0
s
the second line following from the fact that ψ3 (0) = 0. The expressions for ψ4 (t) and ψ3 (t)
just given, and the form of λ(t) for the generator, suggest the following as an educated
guess for γ(t) to achieve limt→∞ ψ4 (t) = 0 and limt→∞ ψ3 (t) = |Ωi,
γ(t) = −eıφ0 qR
t
0
ξ(t)
,
|ξ(s)|2 ds
for some arbitrary real constant φ0 . We will verify below that this form of the modulation
function γ(t) indeed achieves perfect absorption. Note that although γ is singular at t = 0,
γ(0) = −∞, it is square integrable on [s, t] for any 0 < s ≤ t ≤ ∞. Indeed, by direct
integration,
Z t
Z τ
Z t
2
2
2
|ξ(y)| dy dτ,
|γ(τ )| dτ =
|ξ(τ )| /
s
0
s
Z t Z τ
2
=
d ln
|ξ(y)| dy ,
s
0
!
Rt
2
|ξ(y)|
dy
= ln R 0s
.
|ξ(y)|2 dy
0
Moreover, from this it follows immediately that
rR
s
|ξ(y)|2 dy
Z t
R0t
s>0
2
0 |ξ(y)| dy
exp −1/2
|γ(τ )|2 dτ
=
,
0
s = 0, t > 0
s
1
s=t=0
which is thus well-defined for any 0 ≤ s ≤ t ≤ ∞. Substituting the expression for γ(t) and
solving for ψ3 (t) gives
(
Rt
e−ıφ0 √R t 1 2
|ξ(s)|2 ds|Ωt] i t > 0
0
|ξ(τ
)|
dτ
ψ3 (t) =
.
0
0
t=0
8
Notice that limt→∞ ψ3 (t) = |Ωi, as desired. Substituting all of the above into the expression
for ψ4 (t) gives,
Z t
(λ(s)ψ2 (s)dA(s)∗ + γ(s)ψ3 (s)dA(s)∗ )|Ωt] i,
ψ4 (t) =
Z0 t
=
(ξ(s)dA(s)∗ − ξ(s)dA(s)∗ )|Ωt] i,
0
= 0,
again exactly as expected. Thus, weq
conclude that a two-level system of the form G =
Rt
(I, γ(t)σ− , 0) with γ(t) = −eıφ0 ξ(t)/ 0 |ξ(τ )|2 dτ can perfectly absorb a single photon
with an arbitrary wavepacket ξ(t).
3.2
Determining γ by application of the zero-dynamics principle
We now show that the modulation signal γ(t) can also be found by application of the
zero-dynamics principle proposed in [10]. Recall that the cascaded generator and absorber
has the coupling operator L(t) = λ(t)σ1,− + γ(t)σ2,− and the Hamiltonian is H(t) =
={γ(t)∗ λ(t)σ2,+ σ1,− }. The initial state of the composite generator and absorber system is
|1i1 |0i2 . The formulation of the zero-dynamics
R t principle given in [10] is in the Heisenberg
picture. It states that the output field Y (t) = 0 js (L(s))ds+A(t) should be a vacuum field.
This is necessary as no photon should be present in the output field if perfect absorption
is taking place. This means that one must have [22],
Z ∞
Z ∞
∗
∗
2
1
exp ı
(u(t) dY (t) + u(t)dY (t) )
= exp − /2
|u(s)| ds ,
0
0
for any u ∈ L2 (R+ ; C), where the right hand side is the characteristic function of the
vacuum state of a bosonic field. Evaluating the left hand side is in general difficult, if
not intractable, except in special instances like the linear quantum networks analysed
in [10]. However, the zero-dynamics principle can still be applied, but by looking at it
from another equivalent viewpoint. Zero-dynamics corresponds to the state of the output
field always being vacuum at all times t ≥ 0, U (t)|1i1 |0i2 |Ωi = |ϕ(t)i|Ωi for some pure
state vector |ϕ(t)i on H1 ⊗ H2 for all t ≥ 0. This is similar in spirit to the form of
the principle employed in [8, 7]. Thus we may state the ansatz that the wavefunction is
|ϕ(t)i = α(t)|1i1 |0i2 + β(t)|0i1 |1i2 for t ≥ 0 for some scalar complex functions α(t) and
β(t) that will be sought, satisfying |α(t)|2 + |β(t)|2 = 1. We also have the initial condition
α0 = 1 and β0 = 0. Substituting this ansatz into the joint QSDE of the system and field
gives:
d|ϕ(t)i ⊗ |Ωi = (−(ıH(t) + 1/2L(t)∗ L(t))dt + L(t)dA(t)∗ − L(t)∗ dA(t)) |ϕ(t)i ⊗ |Ωi.
Since the input to the generator-absorber is vacuum, for zero-dynamics to hold the term
L(t)dA(t)∗ in the QSDE should be 0 for all times (otherwise dA∗ (t) will create a photon
in the field after time t and the field will no longer be in the vacuum state), meaning that
L(t)|ϕ(t)i = 0 for all t ≥ 0. This gives the condition
(λ(t)α(t) + β(t)γ(t))|0i1 |0i2 = 0, ∀t ≥ 0
9
and thus,
λ(t)α(t) + β(t)γ(t) = 0, ∀t ≥ 0.
(4)
Moreover, noting that L(t)∗ dA(t)|ϕ(t)i|Ωi = 0 (since dA(t) annihilates the portion of the
vacuum after time t), the QSDE reduces to the deterministic equations,
d|ϕ(t)i = −ıH(t)|ϕ(t)idt,
α̇(t)|1i1 |0i2 + β̇(t)|0i1 |1i2 = 1/2γ(t)λ(t)∗ β(t)|1i1 |0i2 − 1/2γ(t)∗ λ(t)α(t)|0i1 |1i2
Thus we arrive at the following set of differential equations,
α̇(t) = γ(t)λ(t)∗ β(t),
β̇(t) = −γ(t)∗ λ(t)α(t),
λ(t)α(t) + β(t)γ(t) = 0
Using the constraint (4), subtituting β(t)γ(t) = −λ(t)α(t) gives,
α̇(t) = −1/2|λ(t)|2 α(t),
β̇(t) = 1/2|γ(t)|2 β(t),
λ(t)α(t) + β(t)γ(t) = 0
The solution for α with initial condition α0 = 1 is:
sZ
α(t) = e−
Rt
0
|λ(s)|2 ds
∞
|ξ(s)|2 ds,
α0 =
t
Rt
But we also have that |α(t)|2 + |β(t)|2 = 1, so |β(t)|2 = 0 |ξ(s)|2 ds. It follows from the
constraint (4) that
|ξ(t)|
.
|γ(t)| = |λ(t)α(t)/β(t)| = qR
t
2 ds
|ξ(s)|
0
So, we arrive at
sZ
β(t) =
0
t
|ξ(s)|2 ds e−iφ(t) , γ(t) = − qR
ξ(t)
t
0
eiφ(t) .
|ξ(s)|2 ds
for some arbitrary real-valued differentiable function φ(t). However, to satisfy the differential equation for β(t), φ(t) must in fact be a real constant, say, φ0 . Whence,
sZ
t
ξ(t)
β(t) =
|ξ(s)|2 ds e−iφ0 , γ(t) = − qR
eiφ0 .
t
0
|ξ(s)|2 ds
0
Thus we recover the form of the modulating function γ(t) that was obtained in Section
3.1 by explicitly solving the QSDE, indicating the power of the zero-dynamics principle
for state-transfer problems such as this. Since one does not need to solve a QSDE, this
approach would be more widely applicable. Moreover, even if the QSDE is explicitly
solvable, it allows the differential equations characterizing γ(t) to be derived thus avoiding
having to make an educated guess about γ(t).
10
3.3
Effect of imperfect realization of γ
We have seen that the analytical form of γ for perfect absorption has a singularity that
grows as O(t−1/2 ) as t → 0 for any wavepacket ξ. This poses a practical challenge as
an infinite coupling magnitude cannot be physically realized. Moreover, in the potential
implementation of the absorber using a single mode resonator, the coupling parameter
should be much smaller than the resonator’s free spectral range as to not excite higher
frequency modes of the resonator. A sub-optimal implementation would be to truncate
the magnitude of γ(t) for small values of t to a finite value. However, this will mean that
the absorber will no longer perfectly absorb a single photon. In this section, we numerically
evaluate the effect of bounding the magnitude of γ(t) for small t on a particular example.
Let nj = σj,+ σj,− denote the number operator for the generator (j = 1) and absorber
(j = 2). To ease the notation, in the following we will often not explicitly write the
time dependence on operators, with the time dependence being implicitly understood. We
will be interested in the evolution of the mean number operator for the absorber, hn2 i,
in the Heisenberg picture, as this gives the probability of excitation of a single photon
in the absorber. In the perfect absorption case we have already discussed, we have that
limt→∞ hn2 i = 1. Since the combined generator-absorber is driven by a vacuum field, the
evolution of the mean hX1 X2 i is given by the ordinary differential equation (the general
expression for Lt (X1 X2 ) is derived in the Appendix),
d
hX1 X2 i = hLt (X1 X2 )i,
dt
= γ(t)∗ λ(t) X1 σ1,− [σ2,+ , X2 ]
+1/2|λ(t)|2 (σ1,+ [X1 , σ1,− ] + [σ1,+ , X1 ]σ1,− )X2
+λ(t)∗ γ(t) σ1,+ X1 [X2 , σ2,− ]
+1/2|γ(t)|2 (σ2,+ [X2 , σ2,− ] + [σ2,+ , X2 ]σ2,− )X1 .
When X2 = n2 and X1 = 1, we get that:
d
hn2 i = hLt (n2 )i,
dt
= γ(t)∗ λ(t) σ1,− [σ2,+ , n2 ] + λ(t)∗ γ(t) σ1,+ [n2 , σ2,− ]
+1/2|γ(t)|2 (σ2,+ [n2 , σ2,− ] + [σ2,+ , n2 ]σ2,− ) ,
= −|γ(t)|2 hn2 i − γ(t)∗ λ(t) σ1,+ σ2,−
∗
− γ(t)λ(t)∗ σ1,+ σ2,− .
(5)
(6)
Let σ2,z = [σ2,+ , σ2,− ]. Then when X2 = σ2,− and X1 = σ1,+ the equation is:
d
hσ1,+ σ2,− i = hL(σ1,+ σ2,− )i,
dt
= −1/2(|λ(t)|2 + |γ(t)|2 )hσ1,+ σ2,− i + γ(t)∗ λ(t) n1 σ2,z i.
(7)
When X1 = n1 and X2 = σ2,z , the equation is:
d
hn1 σ2,z i = hLt (n1 σ2,z )i,
dt
= −|λ(t)|2 hn1 σ2,z i + 2|γ(t)|2 n1 n2 i.
11
(8)
Finally, when X1 = n1 and X2 = n2 the equation is:
d
hn1 n2 i = hLt (n1 n2 )i,
dt
= −(|λ(t)|2 + |γ(t)|2 )hn1 n2 i.
(9)
The initial conditions are hn2 (0)i = 0, hσ1,+ (0)σ2,− (0)i = 0, hn1 (0)σ2,z (0)i = −1, and
hn1 (0)n2 (0)i = 0. From the last initial condition, we can solve for (9) to get that
hn1 (t)n2 (t)i = 0 for all t ≥ 0. Thus there are actually only three coupled equations
that can have a non-trivial solution. The remaining initial conditions can then be used to
sequentially solve equations (8), (7), and (6) (in that order), yielding the explicit solutions.
However, here we are interested in evaluating the effect of approximating the function γ(t),
which is singular at t = 0, with a function γT (t) given by
− √R Tξ(T ) 2 , t ≤ T,
0 |ξ(s)| ds
,
γT (t) =
− √R t ξ(t) 2 , t > T
0
|ξ(s)| ds
where T is a real parameter taking a value in the open interval (0, ∞). By construction,
γT (t) is continuous for all t ≥ 0. To evaluate the probability of the absorber being excited
to the single photon state |1i2 when γ(t) is replaced by γT (t) for some chosen value of T ,
we can solve for (8) to obtain
Z ∞
hn1 (t)σ2,z (t)i = −
|ξ(τ )|2 dτ,
t
and numerically integrate Eqs. (6) and (7), with the given initial conditions.
Let us consider the case√where the input wavepacket ξ(t) is a decaying exponential
function of the form ξ(t) = ce−ct/2 for some positive real constant c. This is the form of
the wavepacket that would be produced at the output of an optical cavity that is initialized
in the 1-photon Fock state. Let us take c = 7.2 × 107 (this is a value that can be realized
in table-top quantum optical experiments, see, e.g., [23]) and allow the absorber to run up
to time t1 = 10/c = 1.3889 × 10−7 . Fig. 1 shows the evolution of hn2 (t)i for T = 0.001t1 ,
T = 0.01t1 , and T = 0.1t1 . It can be seen that for larger T (wider truncation) the
excitation probabillity of the absorber is lower for all t ≥ 0, as can be expected. The steadystate excitation probability is approximately 0.9957, 0.9575, and 0.6037 for T = 0.001t1 ,
T = 0.01t1 , and T = 0.1t1 , respectively.
4
Conclusion
In this work we have considered a single photon absorber with a tunable coupling parameter
to an external travelling single photon field. We analytically derived the exact form of the
time-dependent coupling parameter for perfect absorption of a single photon field of any
temporal wavepacket shape. The ideal modulation function has a singularity at t = 0 which
cannot be attained in real devices, therefore it is approximated with a continuous function
that is truncated to a finite value at t = 0. In a numerical example, we illustrate the effect
of this truncation on the ability of the absorber to perfectly absorb a single photon for a
particular truncation scheme applied to an exponentially decaying wavepacket.
12
Evolution of the probability of single photon excitation in the absorber
1
0.9
0.8
0.7
< n2(t) >
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
1
1.2
Time (s)
1.4
−7
x 10
Figure 1: Time evolution of the excitation probability hn2 i. with the absorber initalized in
the state |0i2 , from t = 0 to t = t1 = 1.3889 × 10−7 , for T = 0.001t1 (top solid black curve),
T = 0.01t1 (middle dashed blue curve), and T = 0.1t1 (bottom dash dotted red curve).
Appendix: General expression for Lt(X1X2)
LL(t) (X1 X2 ) =
=
=
−ı[X1 X2 , H(t)] =
=
=
=
1/2L(t)∗ [X
∗
1
1 X2 , L(t)] + /2[L(t) , X1 X2 ]L(t),
∗
1/2(λ(t)σ
1,+ + γ(t) σ2,+ )[X1 X2 , λ(t)σ1,− + γ(t)σ2,− ]
+1/2[λ(t)∗ σ1,+ + γ(t)∗ σ2,+ , X1 X2 ](λ(t)σ1,− + γ(t)σ2,− )
∗
1/2(λ(t)∗ σ
1,+ + γ(t) σ2,+ )(λ(t)[X1 , σ1,− ]X2 + γ(t)X1 [X2 , σ2,− ])
+1/2(λ(t)∗ [σ1,+ , X1 ] + γ(t)∗ σ2,+ X1 [σ2,+ , X2 ](λ(t)σ1,− + γ(t)σ2,− ).
−ı[X1 X2 , ={γ(t)∗ λ(t)σ2,+ σ1,− }],
−ı[X1 X2 , 1/2ı(γ(t)∗ λ(t)σ2,+ σ1,− − γ(t)λ(t)∗ σ2,− σ1,+ )],
[X1 X2 , 1/2(γ(t)λ(t)∗ σ2,− σ1,+ − γ(t)∗ λ(t)σ2,+ σ1,− )],
∗
1/2γ(t)λ(t)∗ [X X , σ
1
1 2
2,− σ1,+ ] − /2γ(t) λ(t)[X1 X2 , σ2,+ σ1,− ].
13
Combining yields,
Lt (X1 X2 ) = −ı[X1 X2 , H(t)] + LL(t) (X1 X2 ),
= 1/2γ(t)∗ λ(t)[X1 , σ1,− ]σ2,+ X2 + 1/2γ(t)∗ λ(t)X1 [σ2,+ , X2 ]σ1,−
−1/2γ(t)∗ λ(t)(X1 σ1,− X2 σ2,+ − σ1,− X1 σ2,+ X2 ),
+1/2|λ(t)|2 (σ1,+ [X1 , σ1,− ]X2 + [σ1,+ , X1 ]σ1,− X2 ]
+1/2λ(t)∗ γ(t)σ1,+ X1 [X2 , σ2,− ] + 1/2λ(t)∗ γ(t)[σ1,+ , X1 ]X2 σ2,−
+1/2λ(t)∗ γ(t)(X1 X2 σ2,− σ1,+ − σ1,+ σ2,− X1 X2 ),
+1/2|γ(t)|2 (σ2,+ [X2 , σ2,− ]X1 + X1 [σ2,+ , X2 ]σ2,− ),
= γ(t)∗ λ(t)X1 σ1,− [σ2,+ , X2 ] + 1/2|λ(t)|2 (σ1,+ [X1 , σ1,− ] + [σ1,+ , X1 ]σ1,− )X2
+λ(t)∗ γ(t)σ1,+ X1 [X2 , σ2,− ] + 1/2|γ(t)|2 (σ2,+ [X2 , σ2,− ] + [σ2,+ , X2 ]σ2,− )X1 .
References
[1] L.-M. Duan, M. D. Lukin, J. I. Cirac, and P. Zoller, “Long-distance quantum communication with atomic ensembles and linear optics,” Nature, vol. 414, pp. 413–418,
2001.
[2] H. J. Kimble, “The quantum internet,” Nature, vol. 453, pp. 1023–1030, 2008.
[3] J. I. Cirac, P. Zoller, H. J. Kimble, and H. Mabuchi, “Quantum state transfer and
entanglement distribution among distant nodes in a quantum network,” Phys. Rev.
Lett., vol. 78, no. 3221, 1997.
[4] L. Bouten, R. van Handel, and M. R. James, “An introduction to quantum filtering,”
SIAM J. Control Optim., vol. 46, pp. 2199–2241, 2007.
[5] H. Carmichael, An Open Systems Approach to Quantum Optics.
1993.
Berlin: Springer,
[6] Q. Y. He, M. D. Reid, and P. D. Drummond, “Digital quantum memories with symmetric pulses,” Opt. Express, vol. 17, no. 12, pp. 9662–9668, 2009.
[7] J. Dilley, P. Nisbet-Jones, B. W. Shore, and A. Kuhn, “Single-photon absorption in
coupled atom-cavity systems,” Phys. Rev. A, vol. 85, no. 023834, 2012.
[8] A. N. Korotkov, “Flying microwave qubits with nearly perfect transfer efficiency,”
Phys. Rev. B, vol. 84, no. 014510, 2011.
[9] Y. Yin, Y. Chen, D. Sank, P. J. J. O’Malley, T. C. White, R. Barends, J. Kelly,
E. Lucero, M. Mariantoni, A. Megrant, C. Neill, A. Vainsencher, J. Wenner, A. N. .
Korotkov, A. N. Cleland, and J. M. Martinis, “Catch and release of microwave photon
states,” Phys. Rev. Lett., vol. 110, no. 107001, 2013.
[10] N. Yamamoto and M. R. James, “Zero-dynamics principle for perfect quantum memory in linear networks,” New J. Phys., vol. 16, no. 073032, pp. 1–30, 2014.
14
[11] R. L. Hudson and K. R. Parthasarathy, “Quantum Ito’s formula and stochastic evolution,” Commun. Math. Phys., vol. 93, pp. 301–323, 1984.
[12] K. Parthasarathy, An Introduction to Quantum Stochastic Calculus.
Birkhauser, 1992.
[13] P.-A. Meyer, Quantum Probability for Probabilists, 2nd ed.
Springer-Verlag, 1995.
Berlin:
Berlin-Heidelberg:
[14] L. Bouten and R. van Handel, “On the separation principle of quantum control,” in
Quantum Stochastics and Information: Statistics, Filtering and Control (University
of Nottingham, UK, 15 - 22 July 2006), V. P. Belavkin and M. Guta, Eds. Singapore:
World Scientific, 2008, pp. 206–238.
[15] J. Gough and M. R. James, “The series product and its application to quantum
feedforward and feedback networks,” IEEE Trans. Autom. Control, vol. 54, no. 11,
pp. 2530–2544, 2009.
[16] R. Loudon, The Quantum Theory of Light, 3rd ed. Oxford University Press, 2000.
[17] G. J. Milburn, “Coherent control of single photon states,” EPJ ST, vol. 159, no. 1,
pp. 113–117, 2008.
[18] J. E. Gough, M. R. James, and H. I. Nurdin, “Quantum filtering for systems driven by
fields in single photon states and superposition of coherent states using non-markovian
embeddings,” Quantum Inf Process, vol. 12, pp. 1469–1499, 2013.
[19] J. Gough, M. R. James, H. I. Nurdin, and J. Combes, “Quantum filtering for systems
driven by fields in single-photon states or superposition of coherent states,” Phys. Rev.
A, vol. 86, p. 043819, 2012.
[20] J. E. Gough, M. R. James, and H. I. Nurdin, “Quantum master equation and filter
for systems driven by fields in a single photon state,” in Proceedings of the 50th IEEE
Conference on Decision and Control (CDC), 2011, pp. 5570–5576.
[21] Y. Pan, G. Zhang, and M. R. James, “Analysis and control of quantum finite-level
systems driven by single-photon input states,” Automatica, vol. 69, pp. 18–23, 2016.
[22] J. E. Gough, M. R. James, and H. I. Nurdin, “Squeezing components in linear quantum
feedback networks,” Phys. Rev. A, vol. 81, pp. 023 804–1– 023 804–15, 2010.
[23] S. Iida, M. Yukawa, H. Yonezawa, N. Yamamoto, and A. Furusawa, “Experimental
demonstration of coherent feedback control on optical field squeezing,” IEEE Trans.
Autom. Control, vol. 57, no. 8, pp. 2045–2050, 2012.
15
| 3cs.SY
|
Tree-to-tree Neural Networks for Program Translation
Xinyun Chen 1 Chang Liu 1 Dawn Song 1
arXiv:1802.03691v1 [cs.AI] 11 Feb 2018
Abstract
Program translation is an important tool to migrate legacy code in one language into an ecosystem built in a different language. In this work,
we are the first to consider employing deep neural networks toward tackling this problem. We
observe that program translation is a modular procedure, in which a sub-tree of the source tree is
translated into the corresponding target sub-tree
at each step. To capture this intuition, we design a
tree-to-tree neural network as an encoder-decoder
architecture to translate a source tree into a target
one. Meanwhile, we develop an attention mechanism for the tree-to-tree model, so that when
the decoder expands one non-terminal in the target tree, the attention mechanism locates the corresponding sub-tree in the source tree to guide
the expansion of the decoder. We evaluate the
program translation capability of our tree-to-tree
model against several state-of-the-art approaches.
Compared against other neural translation models,
we observe that our approach is consistently better
than the baselines with a margin of up to 15 points.
Further, our approach can improve the previous
state-of-the-art program translation approaches
by a margin of 20 points on the translation of
real-world projects.
1. Introduction
Programs are the main tool for building computer applications, the IT industry, and the digital world. Various
programming languages have been invented to facilitate programmers to develop programs for different applications.
At the same time, the variety of different programming languages also introduces a burden when programmers want to
combine programs written in different languages together.
Therefore, there is a tremendous need to enable program
translation between different programming languages.
Nowadays, to translate programs between different program1
University of California, Berkeley. Correspondence to:
Xinyun Chen <[email protected]>.
ming languages, typically programmers would manually
investigate the correspondence between the grammars of the
two languages, then develop a rule-based translator. However, this process can be inefficient and error-prone. In this
work, we make the first attempt to examine whether we can
leverage deep neural networks to build a program translator
automatically.
Intuitively, the program translation problem in its format
is similar to a natural language translation problem. Some
previous work propose to adapt phrase-based statistical machine translation (SMT) for code migration (Nguyen et al.,
2013; Karaivanov et al., 2014; Nguyen et al., 2015). Recently, neural network approaches, such as sequence-tosequence-based models (Bahdanau et al., 2015; Cho et al.,
2015; Eriguchi et al., 2016; He et al., 2016; Vaswani et al.,
2017), have achieved the state-of-the-art performance on
machine translation. In this work, we study neural machine translation methods to handle the program translation
problem. However, a big challenge making a sequenceto-sequence-based model ineffective is that, unlike natural
languages, programming languages have rigorous grammars
and are not tolerant to typos and grammatical mistakes. It
has been demonstrated that it is very hard for an RNN-based
sequence generator to generate syntactically correct programs when the lengths grow large (Karpathy et al., 2015).
In this work, we observe that the main issue of an RNN
that makes it hard to produce syntactically correct programs
is that it entangles two sub-tasks together: (1) learning the
grammar; and (2) aligning the sequence with the grammar. When these two tasks can be handled separately, the
performance can typically boost. For example, (Dong &
Lapata, 2016) employs a tree-based decoder to separate the
two tasks. In particular, the decoder in (Dong & Lapata,
2016) leverages the tree structural information to (1) generate the nodes at the same depth of the parse tree using
an LSTM decoder; and (2) expand a non-terminal and generate its children in the parse tree. Such an approach has
been demonstrated to achieve the state-of-the-art results on
several semantic parsing tasks.
Inspired by this observation, we hypothesize that the structural information in both input and output parse trees can be
leveraged to enable such a separation. Inspired by this intuition, we propose tree-to-tree neural networks to combine
Tree-to-tree Neural Networks for Program Translation
both a tree encoder and a tree decoder. In particular, we
observe that in the program translation problem, both input
and output programs have their parse trees. In addition, a
cross-language compiler typically follows a modular procedure to translate the individual sub-components in the source
tree into their corresponding target ones, and then compose
them to form the final target tree. Therefore, we design the
workflow of a tree-to-tree neural network to align with this
procedure: when the decoder expands a non-terminal, it
locates the corresponding sub-tree in the source tree using
an attention mechanism, and uses the information of the
sub-tree to guide the non-terminal expansion. In particular, a tree encoder is helpful in this scenario, since it can
aggregate all information of a sub-tree to the embedding
of its root, so that the embedding can be used to guide the
non-terminal expansion of the target tree.
We follow the above intuition to design the tree-to-tree translation model. Existing works (Socher et al., 2011b; Kusner
et al., 2017) have proposed tree-based autoencoder architectures. However, in these models, the decoder can only
access to a single hidden vector representing the source
tree, thus they are not performant in the translation task.
In contrast, we employ an attention mechanism for the decoder to access the source tree, so that our approach can
significantly improve the translation performance (i.e., from
0% to > 90%). To the best of our knowledge, this is the
first tree-to-tree neural network architecture proposed for
translation tasks in the literature.
To test our hypothesis, we develop two novel program translation tasks, and employ a Java to C# benchmark used by
existing program translation works (Nguyen et al., 2015;
2013). First, we compare our approach against several
neural network approaches on our proposed two tasks. Experimental results demonstrate that our tree-to-tree model
outperforms other state-of-the-art neural networks on the
program translation tasks, and yields a margin of up to 5%
on the token accuracy and up to 15% on the program accuracy. Further, we compare our approach with previous
program translation approaches on the Java to C# benchmark, and the results show that our tree-to-tree model outperforms previous state-of-the-art by a large margin of 20%
on program accuracy. These results demonstrate that our
tree-to-tree model is promising toward tackling the program
translation problem. Meanwhile, we believe that our proposed tree-to-tree neural network could also be adapted to
other tree-to-tree tasks, and we consider it as future work.
2. Program Translation Problem
In this work, we consider the problem of translating a program in one language into another. One approach is to
model the problem as a machine translation problem between two languages, and thus numerous neural machine
translation approaches can be applied.
For the program translation problem, however, a unique
property is that each input program unambiguously corresponds to a unique parse tree. Thus, rather than modeling
the input program as a sequence of tokens, we can consider
the problem as translating a source tree into a target tree.
Note that most modern programming languages are accompanied with a well-developed parser, so we can assume that
the parse trees of both the source and the target programs
can be easily obtained.
The main challenge of the problem in our consideration is
that the cross-compiler for translating programs typically
does not exist. Therefore, even if we assume the existence
of parsers for both the source and the target languages, the
translation problem itself is still non-trivial. We formally
define the problem as follows.
Definition 1 (Program translation). Given two programming languages Ls and Lt , each being a set of instances
(pk , Tk ), where pk is a program, and Tk is its corresponding
parse tree. We assume that there exists a translation oracle
π, which maps instances in Ls to instances in Lt . Given a
dataset of instance pairs (is , it ) such that is ∈ Ls , it ∈ Lt
and π(is ) = it , our problem is to learn a function F that
maps each is ∈ Ls into it = π(is ).
Note that this problem definition does not restrict on whether
F knows only the program pk or only the parse tree Tk .
Thus it can leverage both information to finish the translation
task. Also, in the target language, since a program and its
parse tree uniquely matches each other, F only needs to
predict one of them to be correct.
In this definition, Ls and Lt are the source and target languages respectively, and we use the oracle π to model the
goal of the translation. Notice that different tasks can be
modeled with different choices of π. For example, besides
program translation between two languages, we can also
model the compiler optimization by setting Ls = Lt and
choosing π that maps an un-optimized program into an optimized one.
In this work, we focus on the problem setting that we
have a set of paired source and target programs to learn
the translator. Note that all existing program translation
works (Karaivanov et al., 2014; Nguyen et al., 2015; 2013)
also study the problem under such an assumption. When
such an alignment is lacking, the program translation problem is more challenging. Several techniques for NMT have
been proposed to handle this issue, such as dual learning (He
et al., 2016), which have the potential to be extended for the
program translation task. We leave these more challenging
problem setups as future work.
Tree-to-tree Neural Networks for Program Translation
CoffeeScript Program: x=1 if y==0
Parse Tree
Program
If
IfStatement
Block
Op===
Value
Identifier
Literal
Number
Literal
y
0
3.2. Tree-to-tree Neural Network
BinaryExpression
Assign
Value
Identifier
Value
Value
Identifier
Literal
Number
Literal
x
1
program translation model.
JavaScript Program: if (y === 0) { x = 1; }
Parse Tree
Block
===
BlockStatement
ExpressionStatement
Literal
y
AssignExpression
0
Identifier
x
=
Literal
1
Figure 1. Translating a CoffeeScript program into JavaScript. The
sub-component in the CoffeeScript program and its corresponding
translation in JavaScript are highlighted.
3. Tree-to-tree Neural Network
In this section, we present our design of the tree-to-tree
neural network. We first motivate the design, and then
present the details.
3.1. Program Translation as a Tree-to-tree Translation
Problem
Figure 1 presents an example of translation from CoffeeScript to JavaScript. We observe several interesting properties of the program translation problem. First, the translation
can be modular. The figure highlights a sub-component in
the source tree corresponding to x=1 and its translation in
the target tree corresponding to x=1;. This correspondence
is independent of other parts of the program. Consider when
the program grows longer and this statement may repetitively occur multiple times, it may be hard for a sequenceto-sequence model to capture the correspondence based on
only token sequences without structural information.
Second, such a correspondence makes it a natural solution
to locate the referenced sub-tree in the source tree when
expanding a non-terminal in the target tree into a sub-tree.
Inspired by this intuition, we design the tree-to-tree neural network with attention, so that when decoding a nonterminal into a (sub-)tree, the model can employ an attention
mechanism to locate the referenced source sub-tree. The
soft-attention mechanism makes the model differentiable,
so that it can be trained end-to-end.
Third, although the corresponding sub-trees are analogues
to each other, they differ in two aspects. On the one hand,
the non-terminal nodes are different. For example, the nonterminal IdentifierLiteral in CoffeeScript is called
Identifier in JavaScript. Some non-terminals, such as
Value, may not even have a correspondence. On the other
hand, the topology structures of the corresponding sub-trees
can be different. For example, the Assign node in the
source tree has two children, while its correspondence in
the target tree has three: an additional = is inserted in the
middle. These differences can pose a challenge to build a
A tree-to-tree neural network follows an encoder-decoder
framework to encode the source tree into an embedding,
and decode the embedding into the target tree. To capture
the intuition of the modular translation process, the decoder
employs an attention mechanism to locate the corresponding source sub-tree when expanding the non-terminal. We
illustrate the workflow of a tree-to-tree model in Figure 2.
In the following, we present each component of the model.
Converting a tree into a binary one. Note that the input
and output trees may contain multiple branches. Although
we can design tree-encoders and tree-decoders to handle
trees with arbitrary number of branches, we observe that
encoder and decoder for binary trees can be more effective.
Thus, the first step is to convert both the source tree and the
target tree into a binary tree. To this end, we employ the
Left-Child Right-Sibling representation for this conversion.
Binary tree encoder. The encoder employs a TreeLSTM ((Tai et al., 2015)) to compute embeddings for both
the entire tree and each sub-tree. In particular, consider a
node N that is attached with a token ts in its one-hot encoding representation, and it has two children NL and NR ,
which are its left child and right child respectively. The
encoder recursively computes the embedding for N from
the bottom up.
Assume that the left child and the right child maintain the
LSTM state (hL , cL ) and (hR , cR ) respectively. Then the
LSTM state (h, c) of N is computed as
(h, c) = LSTM(([hL ; hR ], [cL ; cR ]), x)
(1)
where [a; b] denotes the concatenation of a and b. Note that
a node may lack one or both of its children. In this case, the
encoder sets the LSTM state of the missing child to be zero.
Binary tree decoder. The decoder generates the target tree
starting from a single root node. The decoder first copies
the LSTM state (h, c) of the root of the source tree, and
attaches it to the root node of the target tree. Then the
decoder maintains a queue of all nodes to be expanded,
and recursively expands each of them. In each iteration,
the decoder pops one node from the queue, and expands
it. In the following, we call the node being expanded the
expanding node.
First, the decoder will predict the token of expanding node.
To this end, the decoder computes the embedding et of
the expanding node N , and then feeds it into a softmax
regression network for prediction:
tt = argmax softmax(W et )
(2)
Tree-to-tree Neural Networks for Program Translation
Source Tree
Block
If
Target Tree
ℎ1
ℎ2
ℎ3 ′
ℎ3
Block
Op===
ℎ4
ℎ8
Number
Literal
ℎ6
y
ℎ4 ′
To Expand
Expanding
Attention map: 𝑤𝑖 ∝ exp(ℎ𝑖𝑇 𝑊0 ℎ5′ )
ℎ12
Value
Value
ℎ15
ℎ13
Identifier
Literal
Number
Literal
ℎ16
ℎ14
x
1
ℎ17
ℎ9
0
IfStatement
BinaryExpression
ℎ10
Assign ℎ11
Value
ℎ5
Identifier
Literal
Program
ℎ2 ′
ℎ5 ′
ℎ7
Value
ℎ1 ′
Source embedding: 𝑒𝑠 = ∑17
𝑖=1 𝑤𝑖 ℎ𝑖 = ℎ1 ; … ; ℎ17 𝑤
Combined embedding: 𝑒𝑡 = tanh(𝑊1 𝑒𝑠 + 𝑊2 ℎ5 ′)
Predicting the node: node = argmax 𝐬𝐨𝐟𝐭𝐦𝐚𝐱(𝑊et )
Figure 2. Tree-to-tree workflow: The arrows indicate the computation flow of the encoder-decoder architecture. Blue solid arrows indicate
the flow from/to the left child, while orange dashed arrows are for the right child. The black dotted arrow from the source tree root to the
target tree root indicates that the LSTM state is copied. The green box denotes the expanding node, and the grey one denotes the node to
be expanded in a queue. The sub-tree of the source tree corresponding to the expanding node is highlighted in yellow. The right corner
lists the formulas to predict the token for the expanding node.
Here, W is a trainable matrix of size Vt × d, where Vt is
the vocabulary size of the outputs and d is the embedding
dimension. Note that et is computed using the attention
mechanism, which we will explain later.
Each token tt is a non-terminal, a terminal, or a special
hEOSi token. If tt = hEOSi, then the decoder finishes
expanding this node. Otherwise, the decoder generates
one new node as the left child and another new node as
the right child of the expanding one. Assume that (h0 , c0 ),
(h00 , c00 ) are the LSTM states of its left child and right child
respectively, then they are computed as:
(h0 , c0 ) = LSTML ((h, c), Btt )
00
00
(h , c ) = LSTMR ((h, c), Btt )
(3)
(4)
probability:
P (Ns is the source sub-tree corresponding to Nt |Nt )
where Nt is the expanding node. We denote this probability
as P (Ns |Nt ), and we compute it as
P (Ns |Nt ) ∝ exp(hTs W0 ht )
where W0 is a trainable matrix of size d × d. To leverage the
information from the source tree, the decoder can sample a
source node Ns following P (Ns |Nt ) or simply predict the
most-likely source node Ns = argmaxNs P (Ns |Nt ), and
then get its hidden hs as es . This embedding can then be
combined with h, the hidden state of the expanding node, to
compute et as follows:
et = tanh(W1 [h; es ])
Here, B is a trainable word embedding matrix of size d × Vt .
Note that the generation of the left child and right child use
two different sets of parameters for LSTML and LSTMR
respectively. These new children are pushed into the queue
of all nodes to be expanded.
The target tree generation process terminates when the
queue is empty.
Attention mechanism to locate the source sub-tree. Now
we consider how to compute et . One straightforward approach is to compute et as h, which is the hidden state
attached to the expanding node. However, in doing so, the
embedding will soon forget the information about the source
tree when generating deep nodes in the target tree, and thus
the model yields a very poor performance.
To make better use of the information of the source tree, our
tree-to-tree model employs an attention mechanism to locate
the source sub-tree corresponding to the sub-tree rooted at
the expanding node. Specifically, we compute the following
(5)
(6)
where W1 is a trainable matrix of size d × 2d.
However, this approach suffers a big issue that the entire network is no longer end-to-end differentiable unless explicit
supervision is given to train P (Ns |Nt ). Unfortunately, this
supervision is unavailable. We may rely on reinforcement
learning algorithms to mitigate this issue, but doing so is
expensive and may not be effective.
We solve this problem by changing the above hard-attention
scheme to a soft-attention mechanism. Note that in Equation (6), we only need a continuous embedding vector es
rather than the discrete choice of Ns . Therefore, we can
approximate es as the expectation of the hidden state value
across all Ns conditioning on Nt . In fact, we have
X
es = E[hNs |Nt ] =
hNs · P (Ns |Nt )
(7)
Ns
Note that this final formula (7) coincides with the standard
attention formalism. Combining all equations from (1) to
Tree-to-tree Neural Networks for Program Translation
(7), the entire tree-to-tree neural network is fully differentiable and can be trained end-to-end.
Parent attention feeding. In the above approach, the attention vectors et are computed independently to each other,
since once et is used for predicting the token tt , et is no
longer used for further predictions. However, intuitively, the
attention decisions for the prediction of each node should
be related to each other. For example, for a non-terminal
node Nt in the target tree, suppose that it is related to Ns
in the source tree, then it is very likely that the attention
weights of its children should focus on the descendants of
Ns . Therefore, when predicting the attention vector of a
node, the model should leverage the attention information
of its parent as well.
Following this intuition, we propose a parent attention feeding mechanism, so that the attention vector of the expanding
node is taken into account when predicting the attention
vectors of its children. Formally, besides the embedding of
the token tt , we modify the inputs to LSTML and LSTMR
of the decoder in Equations (3) and (4) as below:
(h0 , c0 ) = LSTML ((h, c), [Btt ; et ])
00
00
(h , c ) = LSTMR ((h, c), [Btt ; et ])
(8)
(9)
Notice that these formulas in their formats coincide with the
input-feeding method for sequential neural networks (Luong
et al., 2015), but their meanings are different. For sequential
models, the input attention vector belongs to the previous
token, while here it belongs to the parent node. In our
evaluation, we will show that such a parent attention feeding
mechanism significantly improves the performance of our
tree-to-tree model.
4. Evaluation
In this section, we evaluate our tree-to-tree neural network
with several baseline approaches on the program translation
task. To do so, we first describe three benchmark datasets in
Section 4.1 for evaluating different aspects. Then we evaluate our tree-to-tree model with several baseline approaches,
including the state-of-the-art neural network approaches and
program translation approaches. In the following, we start
with presenting the details of the benchmark datasets and
models, and then present the results.
4.1. Datasets
To evaluate different approaches for the program translation
problem, we employ three tasks: (1) a synthetic translation
task from an imperative language to a functional language;
(2) translation between CoffeeScript and JavaScript, which
Source program
Target program
for i=1; i<10; i+1 do letrec f i =
if x>1 then
if i<10 then
y=1
let = if x>1 then
else
let y=1 in ()
y=2
else let y=2 in ()
endfor
in f i+1
else ()
in f 1
Figure 3. An example of the translation for the synthetic task.
are both full-fledged languages; and (3) translation of realworld projects from Java to C#, which has been used as a
benchmark in the literature.
For the synthetic task, we design an imperative source language and a functional target language. Such a design makes
the source and target languages use different programming
paradigms, so that the translation can be challenging. Figure 3 illustrates an example of the translation, which demonstrates that a for-loop is translated into a recursive function.
We manually implement a translator, which is used to acquire the ground truth. The formal grammars of the two
languages and the implementation of the translator can be
found in Appendix D.1 and Appendix E respectively .
For the CoffeeScript-JavaScript task, the programming
paradigms of the two languages are identical. CoffeeScript
employs a Python-style succinct syntax, while JavaScript
employs a C-style verbose syntax. To control the program
lengths of the training and test data, we develop a pCFGbased program generator and a subset of the core CoffeeScript grammar. We also limit the set of variables and
literals to restrict the vocabulary size. We rely on the CoffeeScript compiler to generate the corresponding ground truth
JavaScript programs. The grammar used to generate the
programs in our experiments can be found in Appendix D.2 .
In doing so, we obtain a set of CoffeeScript-JavaScript pairs,
and thus we can build a CoffeeScript-to-JavaScript dataset,
and a JavaScript-to-CoffeeScript dataset by exchanging the
source and the target.
For both of the above tasks, we randomly generate 100,000
pairs of source and target programs for training, 10,000 pairs
as the development set, and 10,000 pairs for testing. More
statistics of the above datasets can be found in Appendix B .
To build the Java to C# dataset, we employ the same approach as in (Nguyen et al., 2015) to crawl several opensource projects, which have both a Java and a C# implementation. Same as in (Nguyen et al., 2015), we pair the methods
in Java and C# based on their file names and method names.
The statistics of the dataset is summarized in Table 2. Due
to the change of the versions of these projects, the concrete
dataset in our evaluation may differ from (Nguyen et al.,
Tree-to-tree Neural Networks for Program Translation
T→T
Tree2tree
T→T
(-PF)
T→T
(-Attn)
Seq2seq
P→P
SYN-S
SYN-L
99.99%
99.60%
99.95%
96.68%
55.60%
34.48%
99.75%
68.31%
SYN-S
SYN-L
99.76%
97.50%
98.61%
57.42%
0%
0%
97.92%
12.19%
P→T
T→P
Token accuracy
99.59% 99.90%
45.28% 67.37%
Program accuracy
97.35% 98.38%
0%
9.19%
Seq2tree
Tree2seq
T→T
P→T
T→T
T→P
T→T
99.73%
35.01%
99.70%
96.95%
99.51%
97.41%
99.88%
97.08%
99.82%
95.88%
98.18%
0%
96.14%
67.34%
98.01%
68.11%
98.51%
91.35%
98.36%
87.84%
Table 1. Token accuracy and program accuracy of different approaches for the synthetic task.
Project
Lucene (luc)
POI (poi)
Itext (ite)
JGit (jgi)
JTS (jts)
Antlr (ant)
Total
# of matched methods
5,516
3,153
3,079
2,780
2,003
465
16,996
Table 2. Statistics of the Java to C# dataset.
2015). We apply ten-fold validation for matched method
pairs in each project, as in (Nguyen et al., 2015).
4.2. Metrics
In our evaluation, we measure the following two metrics:
• Token accuracy. Given a set of source programs, we
run the evaluated model to translate them into target
programs, and calculate the percentage of the tokens
that can match the ground truth.
• Program accuracy. Given a set of source programs,
we run the evaluated model to translate them into target
programs, and calculate the percentage of the predicted
programs that can entirely match the ground truth.
Note that the program accuracy is an underestimation of
the true accuracy based on semantic equivalence, and this
metric has been used in (Nguyen et al., 2015). This metric
is more meaningful than other previously proposed metrics,
such as syntax-correctness and dependency-graph-accuracy,
which are not directly comparable to semantic equivalence.
In addition, token accuracy can provide more insight into
the performance of different models.
4.3. Model Details
We evaluate our tree-to-tree model against a sequence-tosequence model (Bahdanau et al., 2015; Vinyals et al., 2015),
a sequence-to-tree model (Dong & Lapata, 2016), and a treeto-sequence model (Eriguchi et al., 2016). Note that for a
sequence-to-sequence model, there can be four variants to
handle different input-output formats. For example, given
a program, we can simply tokenize it into a sequence of
tokens. We call this format as raw program, denoted as
P. We can also use the parser to parse the program into a
parse tree, and then serialize the parse tree as a sequence
of tokens. Our serialization of a tree follows its depthfirst traversal order, which is the same as (Vinyals et al.,
2015). We call this format as parse tree, denoted as T. For
both input and output formats, we can choose either P or
T. Similarly, for a sequence-to-tree model, we have two
variants based on its input format being either P or T, and
for a tree-to-sequence model, we have two variants based on
its output format being either P or T. Note that the sequenceto-tree model generates a tree as output, and thus requires
its output format to be T (unserialized). For the same reason,
the tree-to-sequence model requires its input format to be
T (unserialized), and our tree-to-tree only has one form.
Therefore, we have 9 different models in our evaluation.
The hyper-parameters used in different models can be found
in Appendix A . The baseline models have employed their
own input-feeding or parent-feeding method that is analogous to our parent attention feeding mechanism.
4.4. Results on the Synthetic Task
We create two datasets for the synthetic task: one with an
average length of 20 (SYN-S) and the other with an average
length of 50 (SYN-L). Here, the length of a program indicates
the number of tokens in the source program.
We present the results in Table 1. We can observe that our
tree2tree model outperforms all baseline models on both
datasets. Especially, on the dataset with longer programs,
the program accuracy significantly outperforms all seq2seq
models by a large margin, i.e., > 85%. Its margin over
a seq2tree model can also reach around 30 points. These
results demonstrate that tree2tree model is more capable of
learning the program translation relationship.
Notice that increasing the length of the programs indeed
makes the problem harder: the program accuracy drops for
all models. However, we observe that the decrease of our
Tree-to-tree Neural Networks for Program Translation
T→T
Tree2tree
T→T
(-PF)
T→T
(-Attn)
CJ-AS
CJ-BS
CJ-AL
CJ-BL
99.57%
99.75%
97.15%
95.60%
98.80%
99.67%
71.52%
78.61%
0.09%
0%
0%
0%
JC-AS
JC-BS
JC-AL
JC-BL
87.75%
86.37%
78.59%
75.62%
85.11%
80.35%
54.93%
44.40%
0.09%
0%
0%
0%
Seq2seq
P→P
P→T
Seq2tree
T→P
T→T
CoffeeScript to JavaScript translation
90.51% 79.82% 92.73% 89.13%
97.44% 16.26% 98.05% 93.89%
21.04%
0%
0%
0%
19.26% 9.98% 25.35% 42.08%
JavaScript to CoffeeScript translation
83.07% 86.13% 73.88% 86.31%
80.49% 85.94% 69.77% 85.28%
77.10% 77.30% 65.52% 75.70%
73.14% 73.96% 61.92% 74.51%
Tree2seq
P→T
T→T
T→P
T→T
86.52%
91.97%
80.82%
76.12%
88.50%
88.22%
78.60%
76.21%
96.96%
96.83%
82.55%
83.61%
92.18%
78.77%
46.94%
26.83%
86.86%
85.06%
77.11%
74.34%
86.99%
84.25%
77.59%
71.56%
71.61%
66.82%
60.75%
57.09%
86.53%
85.31%
75.75%
73.86%
Table 3. Program accuracy of different approaches for translation between CoffeeScript and JavaScript. The token accuracy can be found
in Appendix C .
tree2tree model on the program accuracy is only around 2
points, which is much smaller than the decrease of other
models. This shows that the tree2tree model is significantly
better than other baselines at handling longer inputs.
In addition, we perform ablation study to compare the full
tree2tree model with (1) tree2tree without parent attention
feeding (T→T (-PF)) and (2) tree2tree without attention
(T→T (-Attn)). We observe that the full tree2tree model
significantly outperforms the other alternatives. In particular,
on SYN-L, the full tree2tree’s program accuracy is 40 points
higher than the tree2tree model without parent attention
feeding.
More importantly, we observe that the program accuracy of
tree2tree model without the attention mechanism is always
0%. Note that such a model is similar to a tree-to-tree
autoencoder architecture. This result shows that our novel
architecture can significantly outperform previous tree-totree-like architectures on the program translation task.
4.5. Results on the CoffeeScript-JavaScript Task
We now move on to the CoffeeScript-JavaScript task. We
create several datasets named as XY-ZW: X and Y (C or J)
indicate the source and target language respectively; Z (A
or B) indicates the vocabulary; and W (S or L) indicates
the program length. In particular, vocabulary A uses only
{x,y} as variable names and {0,1} as literals; vocabulary
B uses all alphabetical characters as variable names, and
all single digits as literals. S indicates that the CoffeeScript
programs contain 10 tokens on average; and L for 20.
The program accuracy results are presented in Table 3, and
the token accuracy results can be found in Appendix A .
Most of the observations from our synthetic task remain: our
tree2tree model outperforms all baseline models; all models
perform worse on longer inputs; both the attention and the
parent attention feeding mechanisms boost the performance
of our tree2tree model significantly.
Tree2tree
Lucene
POI
Itext
JGit
JTS
Antlr
72.8%
72.2%
67.5%
68.7%
68.2%
31.9% (58.3%)
J2C# 1pSMT
mppSMT
Reported in (Nguyen et al., 2015)
21.5% 21.6%
40.0%
18.9% 34.6%
48.2%
25.1% 24.4%
40.6%
10.7% 23.0%
48.5%
11.7% 18.5%
26.3%
10.0% 11.5%
49.1%
Table 4. Program accuracy on the Java to C# translation. In the
parentheses, we present the program accuracy that can be achieved
by increasing the training set.
In addition, we observe that for the translation from
JavaScript to CoffeeScript, the improvements of the
tree2tree model over the baselines are much smaller than
for CoffeeScript to JavaScript translation. We attribute this
to the fact that the target programs are much shorter. For
example, for a CoffeeScript program with 20 tokens, its corresponding JavaScript program may contain more than 300
tokens. Thus, the model needs to predict much less tokens
for a CoffeeScript program than a JavaScript program, so
that even seq2seq models can achieve a reasonably good
accuracy. However, still, we can observe that our tree2tree
model outperforms all baselines.
4.6. Results on Real-world Projects
We now compare our approach with three state-of-theart program translation approaches, i.e., J2C# (jav),
1pSMT (Nguyen et al., 2013), and mppSMT (Nguyen et al.,
2015), on the real-world benchmark from Java to C#. Here,
J2C# is a rule-based system, 1pSMT directly applies the
phrase-based SMT on sequential programs, and mppSMT
is a multi-phase phrase-based SMT approach that leverages
both the raw programs and their parse trees. Our approach
handles the out-of-vocabulary problem by canonicalizing
all identifiers and literals. For example, the first identifier in
a program (and its subsequent references) is renamed as v1,
the second as v2, etc. In this way, the dataset contains only
Tree-to-tree Neural Networks for Program Translation
a small set of terminals.
The results are summarized in Table 4. For previous approaches, we report the results from (Nguyen et al., 2015).
We can observe that our tree2tree approach can significantly
outperform the previous state-of-the-art on all projects except Antlr. The improvements range from 20.2% to 41.9%.
On Antlr, the tree2tree model performs worse. We attribute
this to the fact that Antlr contains too few data samples for
training. We test our hypothesis by constructing another
training and validation set from all other 5 projects, and
test our model on the entire Antlr. We observe that our
tree2tree model can achieve a test accuracy of 58.3%, which
is 9 points higher than the state-of-the-art. Therefore, we
conclude that our approach can significantly outperform
previous program translation approaches when there are
sufficient training data.
5. Related Work
Statistical approaches for program translation. Some
recent work have applied statistical machine translation
techniques to program translation (Allamanis et al., 2017;
Karaivanov et al., 2014; Nguyen et al., 2015; 2013; 2016).
For example, several works propose to adapt phrase-based
statistical machine translation models and leverage grammatical structures of programming languages for code migration (Karaivanov et al., 2014; Nguyen et al., 2015;
2013). (Nguyen et al., 2016) proposes to use Word2Vec
representation for APIs in libraries used in different programming languages, then learns a transformation matrix
for API mapping. On the contrary, our work is the first to
employ deep learning techniques for program translation.
Neural networks with tree structures. Recently, various
neural networks with tree structures have been proposed
to employ the structured information of the data (Dong &
Lapata, 2016; Rabinovich et al., 2017; Parisotto et al., 2017;
Yin & Neubig, 2017; Alvarez-Melis & Jaakkola, 2017; Tai
et al., 2015; Zhu et al., 2015; Socher et al., 2011a; Eriguchi
et al., 2016; Zhang et al., 2016; Socher et al., 2011b; Kusner et al., 2017). In these work, different tree-structured
encoders are proposed for embedding the input data, and
different tree-structured decoders are proposed for predicting the output trees. In particular, in (Socher et al., 2011b;
Kusner et al., 2017), they propose tree-structured autoencoders to learn vector representations of trees, and show
better performance on tree reconstruction and other tasks
such as sentiment analysis. In this work, we are the first to
design the tree-to-tree neural network for translation tasks.
Neural networks for parsing. Other work study using
neural networks to generate parse trees from input-output
examples (Dong & Lapata, 2016; Vinyals et al., 2015; Aha-
roni & Goldberg, 2017; Rabinovich et al., 2017; Yin &
Neubig, 2017; Alvarez-Melis & Jaakkola, 2017; Dyer et al.,
2016; Chen et al., 2018; 2016). (Dong & Lapata, 2016)
proposes a seq2tree model that allows the decoder RNN to
generate the output tree recursively in a top-down fashion.
This approach achieves the state-of-the-art results on several
semantic parsing tasks. Some other work incorporate the
knowledge of the grammar into the architecture design (Yin
& Neubig, 2017; Rabinovich et al., 2017) to achieve better
performance on specific tasks. However, these approaches
are hard to generalize to other tasks. Again, none of them is
designed for program translation or proposes a tree-to-tree
architecture.
Neural networks for code generation. A recent line of research study using neural networks for code generation (Balog et al., 2017; Devlin et al., 2017; Parisotto et al., 2017;
Ling et al., 2016; Rabinovich et al., 2017; Yin & Neubig,
2017). In (Ling et al., 2016; Rabinovich et al., 2017; Yin &
Neubig, 2017), they study generating code in a DSL from
inputs in natural language or in another DSL. However, their
designs require additional manual efforts to adapt to new
DSLs in consideration. In our work, we consider the treeto-tree model as a generic approach that can be applied to
any grammar.
6. Conclusion and Future Work
In this work, we are the first to consider neural network
approaches for the program translation problem, and are
the first to propose a tree-to-tree neural network that combines both a tree-RNN encoder and a tree-RNN decoder for
translation tasks. Extensive evaluation demonstrates that
our tree-to-tree neural network outperforms several stateof-the-art models. This renders our tree-to-tree model as
a promising tool toward tackling the program translation
problem. In addition, we believe that our proposed treeto-tree neural network has potential to generalize to other
tree-to-tree tasks, and we consider it as future work.
At the same time, we observe many challenges on program
translation that existing techniques are not capable of handling. For example, the models are hard to generalize to
programs longer than the training ones; it is unclear how
to handle an infinite vocabulary set that may be employed
in real-world applications; further, the training requires a
dataset of aligned input-output pairs, which may be lacking
in practice. We consider all these problems as important future work in the research agenda toward solving the program
translation problem.
Tree-to-tree Neural Networks for Program Translation
Acknowledgement
This material is in part based upon work supported by
the National Science Foundation under Grant No. TWC1409915, Berkeley DeepDrive, and DARPA STAC under
Grant No. FA8750-15-2-0104. Any opinions, findings, and
conclusions or recommendations expressed in this material
are those of the author(s) and do not necessarily reflect the
views of the National Science Foundation.
References
Antlr. https://github.com/antlr/.
Itext. http://sourceforge.net/projects/itext/.
Java2csharp. http://sourceforge.net/projects/j2cstranslator/.
Jgit. https://github.com/eclipse/jgit/.
Jts. http://sourceforge.net/projects/jts-topo-suite/.
Lucene. http://lucene.apache.org/.
Poi. http://poi.apache.org/.
Aharoni, Roee and Goldberg, Yoav. Towards string-to-tree
neural machine translation. In ACL, 2017.
Allamanis, Miltiadis, Barr, Earl T, Devanbu, Premkumar,
and Sutton, Charles. A survey of machine learning for big
code and naturalness. arXiv preprint arXiv:1709.06182,
2017.
Alvarez-Melis, David and Jaakkola, Tommi S. Treestructured decoding with doubly-recurrent neural networks. In ICLR, 2017.
Bahdanau, Dzmitry, Cho, Kyunghyun, and Bengio, Yoshua.
Neural machine translation by jointly learning to align
and translate. In ICLR, 2015.
Balog, Matej, Gaunt, Alexander L, Brockschmidt, Marc,
Nowozin, Sebastian, and Tarlow, Daniel. Deepcoder:
Learning to write programs. In ICLR, 2017.
Chen, Xinyun, Liu, Chang, Shin, Eui Chul, Song, Dawn,
and Chen, Mingcheng. Latent attention for if-then program synthesis. In Advances in Neural Information Processing Systems, pp. 4574–4582, 2016.
Chen, Xinyun, Liu, Chang, and Song, Dawn. Towards synthesizing complex programs from input-output examples.
In ICLR, 2018.
Cho, Sébastien Jean Kyunghyun, Memisevic, Roland, and
Bengio, Yoshua. On using very large target vocabulary
for neural machine translation. In ACL, 2015.
Devlin, Jacob, Uesato, Jonathan, Bhupatiraju, Surya, Singh,
Rishabh, Mohamed, Abdel-rahman, and Kohli, Pushmeet.
Robustfill: Neural program learning under noisy I/O.
arXiv preprint arXiv:1703.07469, 2017.
Dong, Li and Lapata, Mirella. Language to logical form
with neural attention. In ACL, 2016.
Dyer, Chris, Kuncoro, Adhiguna, Ballesteros, Miguel, and
Smith, Noah A. Recurrent neural network grammars. In
NAACL, 2016.
Eriguchi, Akiko, Hashimoto, Kazuma, and Tsuruoka, Yoshimasa. Tree-to-sequence attentional neural machine translation. In ACL, 2016.
He, Di, Xia, Yingce, Qin, Tao, Wang, Liwei, Yu, Nenghai, Liu, Tieyan, and Ma, Wei-Ying. Dual learning for
machine translation. In Advances in Neural Information
Processing Systems, pp. 820–828, 2016.
Karaivanov, Svetoslav, Raychev, Veselin, and Vechev, Martin. Phrase-based statistical translation of programming
languages. In Proceedings of the 2014 ACM International
Symposium on New Ideas, New Paradigms, and Reflections on Programming & Software, pp. 173–184. ACM,
2014.
Karpathy, Andrej, Johnson, Justin, and Fei-Fei, Li. Visualizing and understanding recurrent networks. arXiv preprint
arXiv:1506.02078, 2015.
Kusner, Matt J, Paige, Brooks, and Hernández-Lobato,
José Miguel. Grammar variational autoencoder. arXiv
preprint arXiv:1703.01925, 2017.
Ling, Wang, Grefenstette, Edward, Hermann, Karl Moritz,
Kočiskỳ, Tomáš, Senior, Andrew, Wang, Fumin, and
Blunsom, Phil. Latent predictor networks for code generation. In ACL, 2016.
Luong, Thang, Pham, Hieu, and Manning, Christopher D.
Effective approaches to attention-based neural machine
translation. In Proceedings of the 2015 Conference on
Empirical Methods in Natural Language Processing, pp.
1412–1421, 2015.
Nguyen, Anh Tuan, Nguyen, Tung Thanh, and Nguyen,
Tien N. Lexical statistical machine translation for language migration. In Proceedings of the 2013 9th Joint
Meeting on Foundations of Software Engineering, pp.
651–654. ACM, 2013.
Nguyen, Anh Tuan, Nguyen, Tung Thanh, and Nguyen,
Tien N. Divide-and-conquer approach for multi-phase
statistical migration for source code (t). In Automated
Software Engineering (ASE), 2015 30th IEEE/ACM International Conference on, pp. 585–596. IEEE, 2015.
Tree-to-tree Neural Networks for Program Translation
Nguyen, Trong Duc, Nguyen, Anh Tuan, and Nguyen,
Tien N. Mapping api elements for code migration with
vector representations. In Software Engineering Companion (ICSE-C), IEEE/ACM International Conference on,
pp. 756–758. IEEE, 2016.
Parisotto, Emilio, Mohamed, Abdel-rahman, Singh,
Rishabh, Li, Lihong, Zhou, Dengyong, and Kohli, Pushmeet. Neuro-symbolic program synthesis. In ICLR, 2017.
Rabinovich, Maxim, Stern, Mitchell, and Klein, Dan. Abstract syntax networks for code generation and semantic
parsing. In Proceedings of the 55th Annual Meeting of
the Association for Computational Linguistics (Volume 1:
Long Papers), volume 1, pp. 1139–1149, 2017.
Socher, Richard, Lin, Cliff C, Manning, Chris, and Ng,
Andrew Y. Parsing natural scenes and natural language
with recursive neural networks. In Proceedings of the 28th
international conference on machine learning (ICML-11),
pp. 129–136, 2011a.
Socher, Richard, Pennington, Jeffrey, Huang, Eric H, Ng,
Andrew Y, and Manning, Christopher D. Semi-supervised
recursive autoencoders for predicting sentiment distributions. In Proceedings of the conference on empirical
methods in natural language processing, pp. 151–161.
Association for Computational Linguistics, 2011b.
Tai, Kai Sheng, Socher, Richard, and Manning, Christopher D. Improved semantic representations from treestructured long short-term memory networks. In Proceedings of the Annual Meeting of the Association for
Computational Linguistics, 2015.
Vaswani, Ashish, Shazeer, Noam, Parmar, Niki, Uszkoreit,
Jakob, Jones, Llion, Gomez, Aidan N, Kaiser, Lukasz,
and Polosukhin, Illia. Attention is all you need. arXiv
preprint arXiv:1706.03762, 2017.
Vinyals, Oriol, Kaiser, Łukasz, Koo, Terry, Petrov, Slav,
Sutskever, Ilya, and Hinton, Geoffrey. Grammar as a
foreign language. In NIPS, 2015.
Yin, Pengcheng and Neubig, Graham. A syntactic neural model for general-purpose code generation. In ACL,
2017.
Zhang, Xingxing, Lu, Liang, and Lapata, Mirella. Top-down
tree long short-term memory networks. In Proceedings
of NAACL-HLT, pp. 310–320, 2016.
Zhu, Xiaodan, Sobihani, Parinaz, and Guo, Hongyu. Long
short-term memory over recursive structures. In International Conference on Machine Learning, pp. 1604–1612,
2015.
Tree-to-tree Neural Networks for Program Translation
A. Hyper-parameters of Neural Network
Models
We present the hyper-parameters of different neural networks in Table 5. These hyper-parameters are chosen to
achieve the best validation accuracy through a grid search.
B. More Statistics of Our Proposed Datasets
We present more detailed statistics of the datasets for the
synthetic task and the CoffeeScript-JavaScript task in Table 6.
C. More Results on CoffeeScript-JavaScript
Task
<Expr>
Table 7 shows the token accuracy of different approaches
for the translation between CoffeeScript and JavaScript.
D. Grammars for Generating the Programs
in the Evaluation
<Cmp>
D.1. Grammars for the Synthetic Task
The grammar specifications of the source language (FOR
language) and the target language (LAMBDA language)
used in the synthetic task are provided in Figure 4 and
Figure 5 respectively.
D.2. Grammar for the CoffeeScript-JavaScript Task
The grammar used to generate the CoffeeScript-JavaScript
dataset, which is a subset of the core CoffeeScript grammar,
is provided in Figure 6.
E. Python Implementation of the Translator
for the Synthetic Task
The python source code to implement the translator from a
FOR program to a LAMBDA program in the synthetic task
is provided in Figure 7.
<Assign>
<If>
<For>
<Single>
<Seq>
<Statement>
::=
|
|
|
|
|
<Var>
<Const>
<Expr> + <Var>
<Expr> + <Const>
<Expr> − <Var>
<Expr> − <Const>
::= <Expr> == <Expr>
|
<Expr> > <Expr>
|
<Expr> < <Expr>
::= <Var> = <Expr>
::= if <Cmp> then <statement>
else <statement> endif
::= for <Var> = <Expr> ;
<Cmp> ; <Expr> do
<Statement> endfor
::= <Assign> | <If> | <For>
::= <Single> ; <Single>
|
<Seq> ; <Single>
::= <Seq> | <Single>
Figure 4. Grammar for the source language FOR in the synthetic
task.
Tree-to-tree Neural Networks for Program Translation
Batch size
Number of RNN layers
Encoder RNN cell
Decoder RNN cell
Initial learning rate
Learning rate decay schedule
Hidden state size
Embedding size
Dropout rate
Gradient clip threshold
Weights initialization
Seq2seq
100
3
LSTM
Seq2tree
20
1
LSTM
Tree2seq
Tree2tree
100
100
1
1
Tree LSTM
Tree LSTM
LSTM
0.005
Decay the learning rate by a factor of 0.8× when the
validation loss does not decrease for 500 mini-batches
256
256
0.5
5.0
Uniformly random from [-0.1, 0.1]
Table 5. Hyper-parameters chosen for each neural network model.
Average input length (P)
Minimal output length (P)
Maximal output length (P)
Average output length (P)
Minimal input length (T)
Maximal input length (T)
Average input length (T)
Minimal output length (T)
Maximal output length (T)
Average output length (T)
CJ-(A/B)S
10
23
151
44
34
61
48
38
251
71
CJ-(A/B)L
20
33
311
69
69
111
85
73
531
129
SYN-S
20
22
44
30
40
56
49
41
82
55
SYN-L
50
46
96
71
100
134
111
90
177
133
Table 6. Statistics of the datasets used for the synthetic task and the CoffeeScript-JavaScript task.
<Unit>
::=
<Expr>
::= <Var>
|
<Expr> + <Var>
|
<Expr> − <Var>
<Cmp>
<Term>
<LetTerm>
<IfTerm>
()
::= <Expr> == <Expr>
|
<Expr> > <Expr>
|
<Expr> < <Expr>
::= <LetTerm> | <Expr>
|
<Unit> | <IfTerm>
::= let <Var> = <Term> in <Term>
|
letrec <Var> <Var> = <Term>
in <Term>
::= if <Cmp> then <Term>
else <Term>
Figure 5. Grammar for the target language LAMBDA in the synthetic task.
Tree-to-tree Neural Networks for Program Translation
T→T
Tree2tree
T→T
(-PF)
T→T
(-Attn)
CJ-AS
CJ-BS
CJ-AL
CJ-BL
99.97%
99.98%
99.37%
99.36%
99.97%
99.98%
98.16%
99.27%
56.21%
47.54%
32.99%
31.80%
JC-AS
JC-BS
JC-AL
JC-BL
99.14%
98.84%
96.95%
96.48%
98.81%
98.18%
92.65%
92.49%
65.42%
55.22%
42.23%
39.89%
Seq2seq
P→P
P→T
T→P
Seq2tree
T→T
CoffeeScript to JavaScript translation
93.51% 92.30% 95.46% 95.05%
99.08% 87.51% 99.11% 96.14%
85.84% 25.65% 19.13% 36.18%
80.22% 63.49% 87.27% 79.85%
JavaScript to CoffeeScript translation
88.44% 96.27% 88.46% 98.34%
86.85% 97.92% 85.98% 98.09%
88.09% 95.94% 87.19% 95.04%
87.31% 94.12% 85.70% 96.24%
Tree2seq
P→T
T→T
T→P
T→T
93.29%
98.31%
95.64%
94.09%
95.94%
98.09%
94.74%
94.64%
98.96%
99.27%
94.18%
93.85%
98.09%
98.10%
84.71%
78.07%
98.20%
96.93%
93.51%
94.79%
99.06%
98.84%
96.59%
96.33%
86.93%
84.81%
84.57%
83.03%
98.36%
97.94%
94.63%
94.68%
Table 7. Token accuracy of different approaches for translation between CoffeeScript and JavaScript.
<Expr>
<Var>
<Const>
<Expr> + <Var>
<Expr> + <Const>
<Expr> * <Var>
<Expr> * <Const>
<Expr> == <Var>
<Expr> == <Const>
<Var> = <Expr>
<Expr>
<Simple> if <Expr>
<IfShort> if <Expr>
<Simple> while <Expr>
<WhileShort> while <Expr>
<Simple> | <IfShort> | <WhileShort>
<ShortStatement>
if <Expr> <br> <indent+> <Block> <indent->
while <Expr> <br> <indent+> <Block> <indent->
if <Expr> <br> <indent+> <Block> <indent-> <br>
else <br> <indent+> <Block> <indent->
|
if <Expr> then <ShortStatement> else <ShortStatement>
::= <Statement>
|
<Block> <br> <Statement>
::=
|
|
|
|
|
|
|
<Simple> ::=
|
<IfShort> ::=
|
<WhileShort> ::=
|
<ShortStatement> ::=
<Statement> ::=
|
|
|
<Block>
Figure 6. A subset of the CoffeeScript grammar used to generate the CoffeeScript-JavaScript dataset. Here, <br> denotes the newline
character.
Tree-to-tree Neural Networks for Program Translation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
def t r a n s l a t e f r o m f o r ( self , a s t ) :
i f t y p e ( a s t ) == t y p e ( [ ] ) :
i f a s t [ 0 ] == ’<SEQ>’ :
t1 = s e l f . t r a n s l a t e f r o m f o r ( ast [1])
t2 = s e l f . t r a n s l a t e f r o m f o r ( ast [2])
i f t 1 [ 0 ] == ’<LET>’ and t 1 [ −1] == ’<UNIT>’ :
t 1 [ −1] = t 2
return t1
else :
r e t u r n [ ’<LET>’ , ’ b l a n k ’ , t 1 , t 2 ]
e l i f a s t [ 0 ] == ’<IF>’ :
cmp = a s t [ 1 ]
t1 = s e l f . t r a n s l a t e f r o m f o r ( ast [2])
t2 = s e l f . t r a n s l a t e f r o m f o r ( ast [3])
r e t u r n [ ’<IF>’ , cmp , t 1 , t 2 ]
e l i f a s t [ 0 ] == ’<FOR>’ :
var = s e l f . t r a n s l a t e f r o m f o r ( a s t [ 1 ] )
init = self . translate from for ( ast [2])
cmp = s e l f . t r a n s l a t e f r o m f o r ( a s t [ 3 ] )
inc = s e l f . t r a n s l a t e f r o m f o r ( ast [4])
body = s e l f . t r a n s l a t e f r o m f o r ( a s t [ 5 ] )
t b = [ ’<LET>’ , ’ b l a n k ’ , body , [ ’<APP>’ , ’ f u n c ’ , i n c ] ]
f u n c b o d y = [ ’<IF>’ , cmp , t b , ’<UNIT>’ ]
t r a n s l a t e = [ ’<LETREC>’ , ’ f u n c ’ , v a r , f u n c b o d y , [ ’<APP>’ , ’ f u n c ’ , i n c ] ]
return translate
e l i f a s t [ 0 ] == ’<ASSIGN>’ :
r e t u r n [ ’<LET>’ , a s t [ 1 ] , a s t [ 2 ] , ’<UNIT>’ ]
e l i f a s t [ 0 ] == ’<Expr>’ :
return ast
e l i f a s t [ 0 ] == ’<Op+> ’ :
return ast
e l i f a s t [ 0 ] == ’<Op−>’ :
return ast
e l i f a s t [ 0 ] == ’<CMP>’ :
return ast
else :
return ast
Figure 7. The Python code to translate a FOR program into a LAMBDA program.
| 2cs.AI
|
1
Deep Recurrent Neural Networks for TimeSeries Prediction
Sharat C Prasad, Member, IEEE, and Piyush Prasad
Abstract— Ability of deep networks to extract high level
features and of recurrent networks to perform time-series
inference have been studied. In view of universality of one hidden
layer network at approximating functions under weak
constraints, the benefit of multiple layers is to enlarge the space
of dynamical systems approximated or, given the space, reduce
the number of units required for a certain error. Traditionally
shallow networks with manually engineered features are used,
back-propagation extent is limited to one and attempt to choose a
large number of hidden units to satisfy the Markov condition is
made. In case of Markov models, it has been shown that many
systems need to be modeled as higher order. In the present work,
we present deep recurrent networks with longer backpropagation through time extent as a solution to modeling
systems that are high order and to predicting ahead. We study
epileptic seizure suppression electro-stimulator. Extraction of
manually engineered complex features and prediction employing
them has not allowed small low-power implementations as, to
avoid possibility of surgery, extraction of any features that may
be required has to be included. In this solution, a recurrent
neural network performs both feature extraction and prediction.
We prove analytically that adding hidden layers or increasing
backpropagation extent increases the rate of decrease of
approximation error. A Dynamic Programming (DP) training
procedure employing matrix operations is derived. DP and use of
matrix operations makes the procedure efficient particularly
when using data-parallel computing. The simulation studies show
the geometry of the parameter space, that the network learns the
temporal structure, that parameters converge while model output
displays same dynamic behavior as the system and greater than
.99 Average Detection Rate on all real seizure data tried.
Index Terms— Recurrent Neural Networks, Multi-layer
Neural Networks, Artificial Neural Network, Neural Prosthesis
T
I.
INTRODUCTION
IME-SERIES are found in many areas including
communication, health and finance. Large-scale computer
and communication networks generate time-varying metric
that characterize the state of the networks. Audio signal
arriving at a microphone or radio-frequency signal arriving at
a receive antenna are series of values. In the area of health
Electrocardiograph, electroencephalogram, etc all are or
generate time-series data. End-of-day face-values of different
Date first submitted 8/15/2014.
Sharat C Prasad is Best3 Corp, Saratoga, CA 95070 (e-mail:
[email protected]).
Piyush Prasad is with the Dept of Biomedical Engineering, University of
Washington, ST Louis, MO 65030 (e-mail:[email protected]).
financial instruments are time-series. In all these cases, it is of
interest to discern patterns, detect behavior that differs from
ordinary and predict future behavior.
Many applications have understandably stringent
requirements. We pick implantable epileptic seizure
suppression electro-stimulator driven by automatic seizure
prediction as such an application and to provide the context
for a study of time-series prediction. Epileptic seizure is an
abnormality characterized by brief and episodic neuronal
synchronous discharges with dramatically increased
amplitude.
Electroencephalography (EEG) provides measure of cortical
activity with millisecond temporal resolution. Signal
representations and processing of EEG recordings has been
considered since early days of EEG. Recordings are band-pass
filtered (e.g. with a pass band of 1-70Hz) and digitized (e.g. at
200 samples/second and 12-bit resolution).
The current gold standard for determining onset times is the
trained eye of an encephalographer. Expert encephalographers
look for rhythmic activity. Brain activity during a seizure
displays a great deal of variation from patient to patient and
even from one incident to the next for the same patient.
Implantable neural prosthetic that deliver electrical
stimulation on-demand have emerged as a viable and preferred
treatment for a portion of the population of seizure patients.
Up until recently for reasons of feasibility of implementation,
systems normally take the form of bed-side equipment
communicating with implanted electrodes.
Size is a consideration for any implantable device.
Furthermore both on account of power availability and of
damage to surrounding tissue by the heat produced by the
implanted device, power consumption is a consideration.
Detection algorithms employ characteristics of or features
extracted from EEG signal for detection.
Both feature computation and prediction modules contribute
to the overall size and power requirement. Feature
computation and prediction employ algorithms required to
achieve their objective. Each feature computation may
compute a distinct feature and only this feature.
Savings result if same computation element is reused for
sub-phases of the same computation phase (e.g. different
feature computations). Greater savings result if same
computation element is reused across phases (e.g. feature
computation and prediction).
Given the variability in signature in EEG of seizure even for
the same patient, a rich variety of features has been designed
to assist with the task of seizure detection [17][9]. Given the
2
consideration of surgery, it is desirable that the implant
include support for any feature computation that is or may
become necessary1. Support for comprehensive set of features
is made complicated by the possibility that the optimal set of
features may differ for the same patient from one stage of
treatment to a following stage.
Given the vector of inputs
, vector of features
is
computed. System is modeled as having internal state
and
transition function
and
measurement function
such that
such that
In defining the internal state
, attempt is made to satisfy the
.
As is usual to the type of problem at hand, the parameters of
the prediction algorithm may need to be set by a training
phase. Also, as elsewhere, there are two options to training offline and online. A consideration with offline trained
algorithms is that as parameters change, the offline training
needs to be repeated (or also continuously run in parallel) and
the detection algorithm needs to be configured afresh with the
new parameters.
Systems employing online training can be considered to
continuously incrementally train themselves and therefore able
to adapt.
We will refer to frameworks where given the sequence of
U[
k+1 ]
and
= { u1 , u 2,…, u k+1}
observations
[k]
= { z1 , z 2,…, z k } where each
is
,
is a
known – the measurement - function but
s are not
€
observable,
notation
stands for estimate of
given
Z
€
and
, we first predict
is available, compute
{xˆ
k+1| k+1
and then, when
, xˆ k| k+1,…, xˆ k−q+2| k+1
} as predict-
correct frameworks. The ability of predict-correct frameworks
to “go back” and compute an improved estimate
than
€
computed earlier offers the possibility that
occurrence of an easier to detect event at instant
can be
used to learn correct mapping when
were either not
available or were not informative.
i=M
The problem is one of computing parameters w = { w i } i=1
of a parameterized function
such that when it is applied to
the training input data X = { x j }
from values Zˆ = zˆ j = f w x j
{
( )}
j=N
j=1
j=N
j=1
and supplied output
values Z = { z j }
€
some metric computed
€
returned by the function
j=N
j=1
is optimized.
For example, the problem can be expressed as
€
€
1 Another application
area with this characteristic is Deep Space
Exploration. Once a explorer is launched with certain, preferably learning,
computation engines, these computation engines have to learn to solve the
problem even if the characteristics of the problems evolve as explorer travels
deeper and deeper into outer space.
(1)
Here
is a trade-off constant that trades-off training error
minimization for regularization.
As expressed above the problem is called ridge-regression
and has a simple closed form solution
.
, are defined.
Markov condition
inputs
.
(2)
The above solution represents an off-line training method and,
as introduced above has following assumptions:
•
is linearly dependent on
,
•
s are i.i.d.,
•
components of
•
given training pairs (
are not linearly correlated,
,
) are noise free and
objective and regularization are as in Ridge
Regression.
We know each of these to not hold for Epileptic Seizure
prediction. Adaptability and therefore online training is
desirable in seizure prediction methods. Only a highly
complex feature can hope to achieve linear relation to
prediction output. Individual feature vectors in a sequence of
feature vectors are not independent. Training data is usually
significantly noisy. An objective such as maximizing average
detection rate is not maximizing linear combination of
distances.
When components of
are not completely orthogonal,
methods exist to orthogonalize them [13].
When
s are not i.i.d., 'Time Series' methods consider
that
instant'
is value at 'time instant'
,
is value at 'time
and so on. These methods take covariance
explicitly into account.
Assumption of linear relation of
to
can be
eliminated by employing a Neural Network.
Feedback interconnections within Recurrent Neural
Networks impart them the ability to retain information and
enable them to perform inference tasks such as Epileptic
Seizure Prediction involving time series such as
Electroencephalographic data of brain activity.
Deep Neural Networks can automatically extract relevant
high-level characteristics making it unnecessary to have to
manually engineer characteristic extraction and avoiding loss
of performance from characteristics being not matched to task
at hand. However, most importantly, automatic feature
extraction, addresses the need for adaptability whereby
features suited to a patient are used.
However both, the network being deep and network
incorporating feedback, add to the difficulty of training the
network. Approaches considered for training deep networks
include combining and following unsupervised pre-training
with supervised training [4]. At core of both of these training
algorithms is usually the backpropagation gradient descent.
3
Among the earliest methods considered for training
Recurrent Networks is the Back-Propagation Through Time
(BPTT) method [24] that can be considered the analogue of
the Back-Propagation method for Feed-Forward Networks.
However the BPTT method had been believed to be prone to
getting caught in many shallow local minima on the error
hyper-surface in the parameter hyper-space. Traditionally, for
local minima avoidance, unit activation and/or weight update
are made stochastic and the activation or the weight are
assigned, rather than the exact computed value, a value drawn
from the applicable distribution with the computed values as
moments of the distribution.
Furthermore limitations stem from approximating the nonlinear dynamics of systems as (locally) linear for purposes of
deriving the moments of the posterior state distribution and
from approximating the state distributions as Gaussian.
It is to be noted that greater the non-linearity greater are the
errors introduced by the linear approximation. In case when
process and measurement functions are deep neural networks,
if the transformation from external inputs thru the collection of
layers to final outputs is considered, the transformation
represents a large degree of non-linearity.
Computing the derivatives is tedious and is considered
another shortcoming of methods requiring derivatives.
It is needed to address both the eventual and the speed of
convergence. The model of a system, that when supplied with
some input with stationary statistics, is producing the same
time varying output as the system, cannot be considered
converged if one-step updates are non-zero.
The main contribution of the present work are (i) Deep
Recurrent Neural Network as a unified solution for automatic
extraction of relevant features, automatic construction of
internal state and prediction for time-series data, (ii)
theoretical derivation of the broader space of dynamical
systems enabled by both more than 1 hidden layer and backpropagation through time by greater than 1 instant (iii) the
Back-Propagation Thru Time-and-Space training method as a
solution to hard problem of training deep recurrent networks,
(iv) the dynamic programming formulation of the weight
computation and (v) small, low-power and adaptive epileptic
seizure prediction solution.
In outline, this report presents in Section 2, Deep Recurrent
Neural Networks and their Back-propagation Thru Time and
Space Training. Section 3 discusses universality of Recurrent
Neural Networks and the bound on error in approximation by
RNN. Section 5 presents simulation study of training of and
prediction using a Deep Recurrent Neural Network for
automatically extracted features.
Electroencephalography (EEG) provides measure of cortical
activity with millisecond temporal resolution. It is a record of
electrical potential generated by the cerebral cortex nerve cells
(Figure 1). When epileptic brain activity occurs locally in a
region of the brain and it is seen only in a few channels of
EEG and, when it happens globally in the brain, it is seen in
all channels of EEG.
An example setup for EEG recording [21] may include
Ag/AgCl disc electrodes placed in accordance with one of the
electrode placement systems (e.g. the 10-20 international
electrode placement system [16]). Electrodes may be placed
on the scalp, or may be leads that are implanted in the
cranium. These may be implanted in or on the cortex at the
seizure onset zones. An example is the sensor-neurostimulator device from NeuroPace [23] that includes an
implantable battery-powered neuro-stimulator and quadripolar
depth or strip leads that are implanted in the cranium and are
used for both sensing and stimulation.
Signal representations and processing of EEG recordings
has been considered since early days of EEG. Recordings are
band-pass filtered (e.g. with a pass band of 1-70Hz) and
digitized (e.g. at 200 samples/second and 12-bit resolution).
Representations based on Fourier Transform have been
popular. Four characteristic bands - delta (<4 Hz), theta (4-8
Hz), alpha (8-13 Hz) and beta (13-10 Hz) are identified in the
Fourier Transform.
The current gold standard for determining onset times is the
trained eye of an encephalographer. Expert encephalographers
look for rhythmic activity. Epileptiform discharge may be seen
in the form of spikes repeating at a frequency of 3-Hz and
wave complex in absence seizure (Figure 2). Phase
synchronization and changes in it, are believed to accompany
seizures. Reported measures representative of seizure also
include spectral power, synchronization or amplitude deviates
from normal bounds.
II. TIME-SERIES AND PREDICTION
A. Epileptic Seizure
Epileptic seizure is an abnormality characterized by brief
and episodic neuronal synchronous discharges with
dramatically increased amplitude. Seizures are generated by
abnormal synchronization of neurons which is unforeseeable
for the patients.
Figure 1: Normal brain activity.
Brain activity during a seizure displays a great deal of
variation from patient to patient and even from one incident to
the next for the same patient. A patient may exhibit low
voltage fast gamma activity, may have a beta frequency burst
4
just prior to high frequency activity, may have onset activity in
alpha bands characterized by rhythmic round sinusoidal
waves, may have sharp waves in the delta-theta range
followed by high amplitude spiking activity, may have one of
these onset activities or voltage gamma activity preceded by
spike-and-wave activity or may have activity from the above
entire range preceded by semi rhythmic slow waves [17].
There are two benefits to the Time Series analysis of EEG
data advocated here - (1) offline to determine whether pre-ictal
states exist enabling long-term advance prediction of seizure
While wireless communication between implanted electrode
and external control unit is possible to simplify the implanted
electronics, this may limit fine-grained analysis as the possible
data rates may remain of the order of 100s of Kbps and
therefor too low for some time to come [25]. While new
wireless powering technology is coming into being, the rate of
energy transferred may remain low for some time to come
[18]. Hence size and power consumption need to be
considered along side detection efficiency.
Detection algorithms employ characteristics of or features
extracted from EEG signal for detection. Feature may be
computed on the raw time-domain or frequency or another
transform domain representation. For example, sequence of
samples can be divided into time windows, frequency
transform can be computed for each time window to obtain
spectral power in a few different frequency bands, the vector
Figure 2: Brain activity during seizure.
and (2) online employing information gained from (1) for
preventative therapy in the form of electro-stimulation that
have been seen to be able to suppress or limit the seizure if
applied in a timely manner. Electro-stimulation is believed to
reset brain dynamic from pre-ictal to inter-ictal state.
Implantable neural prosthetic that deliver electrical
stimulation on-demand have emerged as a viable and preferred
treatment for a portion of the population of seizure patients.
While timely application of electro-stimulation is seen to
suppress seizures, extended or more than necessary electrostimulation is seen to damage neuronal tissue and have such
adverse consequences as requiring increasing amounts of
stimulation to have effect and stimulation induced suppression
of neuronal excitability. Up until recently for reasons of
feasibility of implementation, systems normally take the form
of bed-side equipment communicating with implanted
electrodes. In comparison self-contained neural prosthetic will
not require the patient to be bed-bound. Responsive neurostimulation, by virtue of being timely and limiting the amount,
may also significantly increase the effectiveness of therapy.
The advent of very low-power battery-operated electronics has
made possible implantable electronics with closed-loop
working of detection and stimulation.
Traditionally detection algorithms have emphasized
detection efficacy and have disregarded power consumption
and size. Size is a consideration for any implantable device.
Furthermore both on account of power availability and of
damage to surrounding tissue by the heat produced by the
implanted device, power consumption is a consideration.
Figure 3: Organization of detection sub-system.
of spectral powers can be a feature vector and a sequence of
such vectors may be analyzed by a detection algorithm. Figure
3 shows the overall organization of the detection sub-system
comprising of feature computation and prediction modules.
Both feature computation and prediction modules contribute
to the overall size and power requirement. Feature
computation and prediction employ algorithms required to
achieve their objective. Each feature computation may
compute a distinct feature and only this feature.
In above, we gave examples of characteristics expert
encephalographers look for in the EEG. Some of these
characteristics were described in terms of frequency domain
representation. To make use of such characteristic, frequency
domain representation needs to be computed. Efficient
frequency transform computation requires specialized
algorithms and space, time and power. If realized as circuits,
the circuits may not serve to compute a different
representation.
Time-domain features avoid the cost of transform
computation but require specialized algorithms.
One noteworthy aspect of the transform computation
5
algorithms, that we will use to argue in favor of our thesis, is
that their regular structure enables easy to visualize trading-off
of space (number of computation engines, amount of memory)
for time and remain acceptable as long overall latency remains
within limits. To save space and power, computation may be
performed iteratively, reusing the same computation element.
Savings result if same computation element is reused for subphases of the same computation phase (e.g. different feature
computations). Greater savings result if same computation
element is reused across phases (e.g. feature computation and
prediction).
Given the variability in signature in EEG of seizure even for
the same patient, a rich variety of features has been designed
to assist with the task of seizure detection [17][9]. There have
been studies of efficacy of different detection features. Given
the consideration of surgery, it is desirable that the implant
include support for any feature computation that is or may
become necessary2. Support for comprehensive set of features
is made complicated by the possibility that the optimal set of
features may differ for the same patient from one stage of
treatment to a following stage.
Advent of small low-power computer-on-chip [20] makes
them one option to address the feature computation
requirement. However, as numerous studies have shown, size
and power requirement of general purpose programmable
solutions always exceed custom optimized solutions.
There have also been studies of attempts [9] to improve
performance by using detection features in combination.
Latter has shown that when all of a set of detection features
are used best performance results. However the benefit
€ from
using more than a certain number of features may not justify
their incremental cost when size and power constraints exist
[17].
B. Time-series prediction
Given the vector of inputs
, vector of features
is
computed. System is modeled as having internal state
and
transition function
and
measurement function
such that
such that
In defining the internal state
, are defined.
, attempt is made to satisfy the
Markov condition
.
As is usual to the type of problem at hand, the parameters of
the prediction algorithm may need to be set by a training
phase. Also, as elsewhere, there are two options to training offline and online. An offline (batch) training algorithm is
given feature vector
and target
pairs
{( x , z )}
i
i
i= N
.
i=1
The training goes over all the data to determine detection
parameters. A consideration with offline trained algorithms is
that as parameters change, the offline training needs to be
€
2 Another application area with this characteristic is Deep Space
Exploration. Once a explorer is launched with certain, preferably learning,
computation engines, these computation engines have to learn to solve the
problem even if the characteristics of the problems evolve as explorer travels
deeper and deeper into outer space.
repeated (or also continuously run in parallel) and the
detection algorithm needs to be configured afresh with the
new parameters.
On the other hand, an online (incremental) algorithm
initializes its parameters to certain initial values, is given
either just one feature vector
or one feature-vector-target
pair
. An online algorithm only produces a prediction
if only given a feature vector or also updates its parameters
if given a feature-vector-target pair
. To start with, an
on-line trained algorithm may produce poor predictions. As
the data characteristics are learnt, predictions improve.
Systems employing online training can be considered to
continuously incrementally train themselves and therefore able
to adapt. One issue with online training is that for supervised
training, target
corresponding to each feature-vector
is
required. A second issue is that, on one hand, recent featurevector-target pairs may cause invariant characteristics to be
over-written by transient characteristics and on the other hand,
recent pairs may not cause transient characteristics to be
overwritten by new transient characteristics.
We will refer to frameworks where given the sequence of
inputs
U [ k+1] = {u1,u2 ,…,uk+1}
and
Z [ k ] = {z1, z 2 ,…, z k } where each
observations
is
,
is a known – the measurement - function but
€observable, notation
and
s are not
stands for estimate of
, we first predict
is available, compute
{xˆ
given
and then, when
}
, xˆ k| k+1,…, xˆ k −q+2| k+1
k+1| k+1
as
predict-correct frameworks. The ability of predict-correct
frameworks to “go back” and compute an improved estimate
than
computed earlier offers the
€ occurrence of an easier to detect event at
possibility that
instant
can be used to learn correct mapping when
were either not available or were not informative.
Metrics called specificity (
Average Detection Rate (
(
), sensitivity (
) and
) are used. Sensitivity is
)
defined as SEN ≡ Y + / Y + + N − where
true positives and
is the count of
is the count of false negatives.
Specificity is defined as
€
the count of true negatives and
positives. Finally
where
is
is the count of false
.
The problem is one of computing parameters
of a parameterized function
such that when it is applied to
the training input data U = { u j }
from values Zˆ = zˆ j = f w x j
{
( )}
j=N
€
returned by the function
j=1
and supplied output
values Z = { z j }
€
€
some metric computed
j=1
j=N
j=N
j=1
is optimized.
€
€
€
€
6
For example, it may be desired that sum of squared errors is
minimized. Then the problem is, find
where
(3)
Having computed
prediction output. Individual feature vectors in a sequence of
feature vectors are not independent. Training data is usually
significantly noisy. Objectives such as maximizing average
detection rate are not maximizing linear combination of
distances.
, prediction is made as
.
For the problem under consideration,
, and a
problem encountered is that of over-fitting wherein a
is
determined that well minimizes the sum of squared errors for
the training data (by learning too well the characteristics
present) but then performs poorly on other valid data (that
may exhibit a broader set of characteristics). To counter this, a
regularization term may be added to the minimization and the
problem may be written as
argmin ⎛
ˆ=
w
∑⎜⎝ f w u j − z j
w
( ( )
) + λg(w)⎞⎟⎠
2
If it is assumed that the relation of output
input is linear, i.e.
(4)
to corresponding
m
z j = ∑ uij w i = uTj w j ,
(5)
i=1
s are drawn from independent identical distributions (
does not depend on any
,
and
for all
drawn from the same distribution), components of
, are
are not
Figure 4: Prediction with manually engineered features
linearly correlated and lastly the regularization function
simply takes the sum of square of individual elements, i.e.
, the problem can be expressed as
2
argmin ⎛
⎞
ˆ=
w
⎜ f w u j − z j + λg( w)⎟
∑
⎝
⎠
w
argmin
T
=
(Uw − z) (Uw − z) + λwT w .
w
( ( ) )
(
)
(6)
Here
is a trade-off constant that trades-off training error
minimization for regularization.
As expressed above the problem is called ridge-regression
and has a simple closed form solution
.
(7)
The above solution represents an off-line training method and,
as introduced above has following assumptions:
•
is linearly dependent on
,
•
s are i.i.d.,
•
components of
•
given training pairs (
are not linearly correlated,
,
) are noise free and
objective and regularization are as in Ridge
Regression.
We know each of these to not hold for Epileptic Seizure
prediction. Adaptability and therefore online training is
desirable in seizure prediction methods. Only a highly
complex feature can hope to achieve linear relation to
•
Figure 5: Metrics SPC, SEN and ADR versus illustrative manually
engineered features on a sample of real data shows SPC, SEN and ADR with
respect to four manually engineered features [17] – Energy, Hjorth Variance,
Coastline and RMS Amplitude.
Complex objectives rarely have solution in closed form. As
long as the objective is differentiable, we can employ a
member of the Gradient Descent family of algorithms for
solution.
Convex
functions,
characterized
by
are able to express a wider set
of objectives. Carefully designed members of the Gradient
Descent family can optimize convex objectives, can
accommodate samples that are not independent [7] and even
overcome noise and certain forms of non-convexity by taking
advantage of the structure in the objective [15].
7
When components of
are not completely orthogonal,
methods exist to orthogonalize them [13]. Most regression
packages detect when some components are linearly correlated
and either suggest components to eliminate or set respective
coefficients to zero.
Ignoring for the moment that
s are not i.i.d. and
employing manually engineered features, we can use
organization shown in Figure 3 to perform prediction. The
organization uses a single neuron. The weights of the neuron
can be trained using a member of the Gradient Descent family
of algorithms. This organization does not compute any further
higher layer features beyond those manually engineered and
makes predictions that are non-linear but not highly non-linear
functions of the features.
Figure 5 shows SPC, SEN and ADR with respect to four
manually engineered features 0 – Energy, Hjorth Variance,
Coastline and RMS Amplitude. The non-smooth and gradientdiscontinuous nature of the plot is from, among others,
presence of noise and use of discrete 'counts' of 'events' – a 0 if
did not occur and a 1 if it did - in the definition of the SPC and
SEN.
When
s are not i.i.d., 'Time Series' methods consider
that
instant'
is value at 'time instant'
,
is value at 'time
and so on. These methods take covariance
explicitly into account.
Assumption of linear relation of
to
can be
eliminated by employing a Neural Network. Often inference
requires computing an unknown and, possibly highly, nonlinear function of some characteristics of the data. Neural
Networks can efficiently find nonlinear mappings.
Feedback interconnections within Recurrent Neural
Networks impart them the ability to retain information and
enable them to perform inference tasks such as Epileptic
Seizure Prediction involving time series such as
Electroencephalographic data of brain activity.
Deep Neural Networks can automatically extract relevant
high-level characteristics making it unnecessary to have to
manually engineer characteristic extraction and avoiding loss
of performance from characteristics being not matched to task
at hand. However, most importantly, automatic feature
extraction, addresses the need for adaptability whereby
features suited to a patient are used. Furthermore, as the
treatment progresses, the features used can also evolve to suit.
However both, the network being deep and network
incorporating feedback, add to the difficulty of training the
network. Approaches considered for training deep networks
include combining and following unsupervised pre-training
with supervised training [4]. The unsupervised pre-training
tries to minimize energy as low-energy states correspond to
parameters attaining values that capture structures present in
data. The following training then is intended to further evolve
the parameters to capture the desired input-output mapping. At
core of both of these training algorithms is usually the
backpropagation gradient descent.
Figure 6: Automatic High-Level Feature Extraction Using Deep Recurrent
Neural Network
Among the earliest methods considered for training
Recurrent Networks is the Back-Propagation Through Time
method [24] that can be considered the analogue of the Back
Propagation method for Feed-Forward Networks. However the
BPTT method had been believed to be prone to getting caught
in many shallow local minima on the error hyper-surface in
the parameter hyper-space. Traditionally, for local minima
avoidance, unit activation and/or weight update are made
stochastic and the activation or the weight are assigned, rather
than the exact computed value, a value drawn from the
applicable distribution with the computed values as the
moments of the distribution.
It was first noted in late 80s that the Neural Networks are
not unlike non-linear physical systems and that there exists a
well-developed theory of estimating hidden state and
estimating parameters of linear systems from noisy
observations and somewhat less developed similar theories for
nonlinear systems. A number of methods adhering to PredictCorrect framework (e.g. the celebrated Kalman Filter) have
been developed over time that are optimal in the sense of
Minimum Mean Squared Error for state only, parameter only
and joint state and parameter estimation problems for linear
systems. Optimal estimation is no longer computationally
tractable when the system is nonlinear.
The realization that Recurrent Neural Networks are nonlinear systems led to use of Extended Kalman Filter, popularly
used for non-linear system identification, parameter
estimation, etc for training Recurrent Networks. The Extended
Kalman Filter has a set of well known limitations. The
limitations stem from approximating the non-linear dynamics
of systems as (locally) linear for purposes of deriving the
moments of the posterior state distribution and from
approximating the state distributions as Gaussian.
€
8
Time-series prediction has been addressed by Statistics,
Dynamical Systems, Machine Learning, Signal Processing
and Automatic Control communities. The need to address
larger and more complex systems has led beyond BackPropagation Through Time and first order term (in Taylor
expansion) only Extended Kalman Filter [11] to including
second order terms [14], avoiding explicit use of Jacobian and
Hessian [14], Quadrature [1] and Cubature [1] forms, the
Unscented [12] form, the marginalized Sigma Points [19]
form, the Particle Bayesian [22] form, the Variational
Bayesian [22] form and the Gaussian Process [3] form. While
an improvement results, a conceptual and sometimes also a
computational complexity cost has to be paid.
It is to be noted that greater the non-linearity greater are the
errors introduced by the linear approximation. In case when
process and measurement functions are deep neural networks,
if the transformation from external inputs thru the collection of
layers to final outputs is considered, the transformation
represents a large degree of non-linearity.
Computing the derivatives is tedious and is considered
another shortcoming of methods requiring derivatives.
It is needed to address both the eventual and the speed of
convergence. The model of a system, that when supplied with
some input with stationary statistics, is producing the same
time varying output as the system, cannot be considered
converged if one-step updates are non-zero.
C. One-step and long-term time-series prediction
The problem comprises of being given an input time-series
instead
be
approximations
obtained
as
a
sequence
of
,
,
, etc.
In the context of training multi-layer recurrent neural
networks, where weight update is computed as
, this fact can be taken advantage of by
including in the state, in addition to prior outputs of hidden
layers, the current output of visible and hidden layers.
E. High order Markov Processes
Given a process that receives input
produces output
at instant
and
, the process may employ internal state
to remember as much of past history as it needs to so that
it can compute output
considering only
and
.
Expressed another way,
and system output is considered independent of past input
given the current internal state. Such processes are called firstorder. However this may require, not unlike carefully
manually engineering features, carefully manually engineering
the internal state that system must remember. A solution is to
require system to remember history going back
instants.
Then
and
the process is called of order
.
and noisy observations of corresponding output
time-series
where observations may not be
available at each instant of time
and the required
result being
. As an example, the input time-series
may be intra-cranial electrode sensed electric potential, the
output time-series values may only be available for certain ,
may be whether the patient was experiencing an epileptic
seizure3 and the required result is whether the patient is about
to experience an epileptic seizure instants from now.
D. Extrapolating highly non-linear functions
Consider a function
a
sequence
of
y = f ( x ) = e a1 e
functions:
a2x+b2
+b1 specified as
y = f1 ( y1 ) = e a1 y1 +b1
and
a 2 x+b 2
.
y1 = f 2 ( x ) = e
We are given
€ value
of
at
and it is desired to
extrapolate value at €
.
A linear approximation can be done and we can write
.
It is easy to show that a more accurate approximation can
3 Note that in an offline training situation, the EEG data may be
examined by an expert encephalographer and intervals corresponding to
seizure determined. In an online training context, the parameters may be
initialized to values determined offline and predictions used only after
stability is reached.
III. DEEP RECURRENT NEURAL NETWORKS
Neural Networks comprise of one or more inputs, one or
more neurons (referred to in this work as units), one or more
outputs, directed or undirected interconnections among units, a
weight associated with each undirected connection of or each
directed connection into a unit (from an input or another unit)
and each unit computing its activation by applying its
activation function to weighted sum of inputs.
Feed-Forward Neural Networks are arranged in layers with
units in a layer receiving inputs from only units in previous
layer (or outside in case of the input layer). Feed-Forward
networks process data presented at an instant independent of
data presented at previous instants and so are unable to
perform inference on time series data. In a Recurrent Network,
some of the interconnections are from units in a layer through
a delay to units in the same or an earlier layer.
As both stacking of recurrent layers and back-propagating
have the effect of extracting information from input or from
information extracted from input to consider alongside input,
it is a question as to whether (i) either is more effective than
the other and (ii) having both together adds to the
approximation ability of the network.
Each unit in a layer associates a weight to and has as inputs
a fixed bias, output of all units in the previous layer (external
inputs in case of the visible layer) and, in case of the recurrent
layer, its own outputs delayed by a time instant. Output of a
€
€
€
€
9
unit
is the result of applying a nonlinear function
weighted sum of inputs
:
to the
.
The network has a visible layer,
hidden layers and an
output layer of units (neurons) with, respectively,
and
inputs
units in the layer. The network receives
,
and produces outputs
Furthermore,
is the
the visible layer,
the unit
is the input
,
) in
is
is the weight from unit
€,
is
(xˆ ( ) ) . Final outputs of
k
i
the network are also considered collected into a vector. This
vector at instant is denoted as
.
A significant component of any method making use of the
€ computations
derivatives are the derivative
( )
∂( xˆ ( ) )
∂( zˆ ( ) )
( )
(H ) = ∂ xˆ ( )
( )
(F )
( k)
ij
=
∂ xˆ ( k)
i
k
(12)
j
k
ij
i
k
(13)
j
to unit
of the node
l, j
of the same is denoted as
k
.
is the vector with elements
y h ( − k) is the output at (past) instant
hidden layer
.
in the hidden layer ,
in the output layer,
the output of unit
,
-th unit (or alternatively unit
is the unit
As a convention,
,
,
element
€
. Notionally, associated with the unit
in
in
hidden layer
are
memories that retain the previous
outputs y ( − B ) through y ( −1) of the unit.
is referred to as
€
h l, j
€
h l, j
the extent of backpropagation.
Inputs to each unit in any layer include a fixed bias 1. We
use the same symbol as used to denote the output signal vector
of a layer, €
but with a tilde accent, to denote a signal vector
with 1 (corresponding to the fixed bias) as an additional and
the first element and remaining elements same as the original
signal vector. Then
y v = Wv,u y˜ u
(8)
yh
NL
= WhNL , v y˜ v + Wh
NL
( −1)
, h NL
( −1)
h NL
NL
(9)
NL
y h ( −1) = Wh ( −1) ,v ( −1) y˜ u( −1) + W
NL
y h ( −1)
,hN( −2)
L
y h ( −2)
(10)
NL
Reader may notice that we have chosen to qualify, rather
than the symbol denoting the output signal of a node, the
symbol denoting the node with the superscripts
denoting
value of “the node at past instant ” rather than “value at past
instant ” of the node. This is deliberate and is to suggest that
we consider the network to have been unfolded back and there
to exist
copies, one corresponding to each of
instants
spanning the extent of the backpropagation, of the complete
“sub-network” comprising of the visible and each of the
hidden layers. As a result, the notation allows not only vectors
to be distinguished from the vector
, but
also weight matrices
distinguished from
Wv ( −1) , u( −1) ,, Wv ( − B ) , u( − B )
to be
. If this distinction was not made (10)
above will be
y h ( −1) = Wh ( −1) , v ( −1) y˜ v ( −1) + Wh ( −1) , h ( −2) y h ( −2)
NL
N€
NL
NL
NL
L
(11)
denoted as
and
Outputs of all the units as well as the weights associated
with the inputs to the units are considered collected into a
joint-state-and-parameter-vector. This vector at instant
is
, estimate of the vector is denoted as
Figure 7: Back-Propagation Through Time and Space in Deep Recurrent
Neural Networks
Visualization of dependencies considered in the derivative
computation is eased by considering the nodes making up the
network to be arranged as columns of stacks of layers of linear
array of nodes (Figure 7). The input layer is at the bottom, the
hidden layers in the middle and the output layer at the top. The
rightmost column corresponds to current instant and the
columns to left to successively previous instants. A notion of
distance can be associated with any pair of units and of a unit
and a weight. A unit in the (present-instant, output-layer) is at
a distance of 1 from any unit in the (present-instant, hidden
layer), at a distance 2 from any unit in (present instant, visible
layer), also at a distance 2 from (previous instant, hidden
layer) and so on. Similarly the instant
weight
at a distance 1 from (instant
, hidden layer
) unit
is
,
10
distance 2 from any unit in the (present instant, hidden layer
), also distance 2 from any unit in the (instant
, layer
), distance 3 from any unit in the (present instant, layer
), also distance 3 from any unit in (instant
, layer
), etc.
Output at instant
depends on parameters (weights)
€ and
state at . The state at instant
in turn depends on weights
and state at
. Once weights have been updated at instant €
, a new output at
can be computed. To compute this
output, the state at
is required. However the state at
itself can be recomputed using the updated weights and
€
the state at
and so on.
Given the error in output at , part of the error can be
attributed to errors in weights and part to error in
€ the
(estimated) state at . The error in state at
in turn depends
€
in part on errors in weights and part on error in state at
.
If the mapping from input to output is invariant, the weights
at instants
and
are the same. However if the
mapping itself is evolving then the two sets of weights are not
the same.
€
Once the weights corresponding to present and to each past
instant are made different and the present error is computed,
weights at instant
can be overwritten with weights
€ at
instant
and present instant weights can be updated to next
€
instant weights. Alternatively, the back-propagation itself can
be used to compute weights at present as well as each retained
past instant.
The mapping
may be time-invariant,
may need to be learnt and, once learnt, does not change.
€
Alternatively the mapping may continuously evolve.
€
Computing only next instant weights by updating present
instant weights and overwriting instant
weights with
instant
weights is consistent with models that allow
€ the
mapping to be learnt and then assume them to remain
constant. Computing all the weights using back-propagation is
€
consistent with models where mapping may continuously
evolve.
A. Universality of single hidden layer feed-forward network,
bounds on approximation error and multilayer networks
depends on the norm of the Fourier Transform of the function
being approximated
2
2
c f ' = 2rC f = 2r ∫ |ω|| F˜ ( dω )|
(14)
(
(
)
)
where F˜ ( w) Fourier distribution of function
,
i
θ
w
)
such that f ( x) = e iwx F˜ ( dw) , F˜ ( dw) = e ( F ( dw) ,
,
is
∫
the magnitude distribution,
the phase at frequency
and w the vector of frequencies (ω1 ,, ω d ) .
€ Let
€ of functions
be the class
on
for which
C f = ∫ |w|| F˜ ( dw)| < ∞ and
the class of functions
on
€
d
R for which C f = ∫ |w|| F˜ ( dw)| < C .
Then, for every choice of fixed basis functions
[2]
()
€
C ⎛ 1 ⎞
sup s( f,span( h1, h2 ,, hn )) ≥ κ ⎜ ⎟
d ⎝ n ⎠
f ∈ ΓC
where
s( f,G ) = inf s( f, g)
1
d
(15)
(16)
g ∈G
s( f, g) =
2
∫ ( f ( x) − g( x)) dx
(17)
Let,
be a bounded set in R d which contains
the set of functions
on R d
f ( x ) = f ( 0) + e iw.x − 1 F˜ ( dw) holds for
∫(
€
iθ w
F˜ ( dw) = e ( ) F ( dw) ,
)
,
for which
and some
€
the set of functions
Rd
on
for which
(18)
∫ |ω| F ( dω) ≤ C > 0
B
where
|w| B = sup{ w.x} ,
€
(19)
x∈B
then for every function
measure
, for every probability
, for every sigmoidal function
, for every
there exists a linear combination of functions
,
of the form
n
f n ( x) = ∑ c kφ ( a k x +bk ) +c 0
k=1
such that [2]
2
2
Single hidden layer feed-forward networks using the
f n ( x ) − f ( x)) µ ( dx) < c f ' / n = 2rC f / n
(20)
(
∫
monotone cosine squashers are capable of embedding as a
€
special case a Fourier network that yields a Fourier series where Br = x ∈ R d | x| ≤ r .
approximation to a given function at its outputs [8] and thus
possesses all the approximation properties of Fourier €series
Hence adding a sigmoidal layer improves the rate from
representations.
1/ d
to
. Will adding another layer improve further?
A feed-forward network with a single hidden layer of € (1/ n)
continuous sigmoidal nonlinear squashers (in addition to linear By how much?
In a one linear layer network we take a weighted sum of
visible and output layers) are complete in the space of
inputs
and add in a bias
where for
functions on
that have continuous and uniformly formed
€
derivatives up through order
[10].
ease of explanation we number the elements of
and
For a network with a hidden layer with
units, the starting at .
approximation error is bounded by
[2] where
Equation (7) bounds below the approximation error of one
(
{
}
)
11
linear layer network.
In a 1 hidden layer network, we take
distinct copies of
one linear layer networks, feed the output of each into a
nonlinearity, take a weighted sum of the outputs of the
nonlinearities and add in a bias
) (
((
⎛
⎞
= b .⎜( f ( x )),, ( f ( x))⎟ +b .
⎝
⎠
f n1 ( x) = bT . φ a1T x + a10 ,, φ a Tn x + a n0
))
T
n 0 ,1
n 0 ,n
(21)
0
Equation (8) bounds below the rate of error. Note that here
we are concerned with inherent representability using multi€ in
layer neural networks, are not concerned with error
determination of weights and assume that values of weights
minimizing error, if they exist, can be found.
€
€
€
2
∫ ( f (x) − f (x)) µ (dx) .
Consider
f n 2 ( x) is the function
n2
computed by the two hidden layers network. The two hidden
€
layers network has (logically)
instances of the one hidden
€
€
€ network as sub-networks.
layer
Let
be the function
computed by the two hidden layers network when it achieves
minimum error and f˜ ( x),, f˜ ( x) be the functions
(
n1 ,1
n1 , n
)
needed to be computed by its one hidden layer sub-networks
when the two hidden layers network achieves minimum error.
(
€
€
€
€
€
€
)
Let f n1 ,1 ( x),,
€ f n1 , n ( x) be the actual functions computed by
€
the one hidden layer sub-networks. Similarly let
f˜n 0 ,1 ( x),, f˜n 0 , n ( x) be the functions needed to be computed
€
by their one linear layer sub-networks when the one hidden
layer network achieves minimum error relative to
f˜ ( x),, f˜ ( x) and f ( x),, f ( x) be the actual
(
(
n1 ,1
n1 , n
)
(
n 0 ,1
)
n0 , n
functions computed by the one linear layer sub-networks.
Let us denote
⎛
ψ n i ( x) = ⎜φ f n i ,1€
( x) ,, φ f n i , n ( x) ⎞⎟⎠
⎝
€
⎛
⎞
ψ˜ n i ( x) = ⎜φ f˜n i ,1 ( x) ,, φ f˜n i , n ( x) ⎟
⎝
⎠
€
⎛⎜
⎞⎟
+
+
+
ψ n i ( x) = φ f n i ,1 ( x) ,, φ f n i , n ( x)
⎝
⎠
€
⎛⎜
⎞
−
−
−
ψ n i ( x) = φ f n i ,1 ( x) ,, φ f n i , n ( x) ⎟
⎝
⎠
€
Then
€ (22)
f ∗2 ( x) = ψ˜ 1 ( x).c +c 0 ,
(
(
(
(
)
)
)
)
(
(
(
(
)
)
)
)
n
€
f n1 ( x) = ψ n 0 ( x).b+b0 ,
€
∫ ( f ( x) − f˜ ( x)) µ ( dx) < c ' / n
€ (23)
2
€
)
2
/n.
What if the errors of functions computed by the
2
€
2rC f / n
subnetworks
are
smaller
e.g.
versus
€ (1/ d )
κ (C / d )(1/ n) ?
)
Note function computed is weighted sum of
€
f 1 ( x) = ψ n 0 ( x).b+b0
terms
(26)
n
For squared minimum error
⎛⎜ ∗
⎞2
f 1 ( x ) − f ( x ) ⎟ =
inf
ψ n 0 ( x).b+b0 − f ( x)
⎝ n
⎠
f 0 ,, f 0 ,b ,b0
n ,1
n ,n
(
(
)
= ψ˜ n 0 ( x ).b+b0 − f ( x)
)
2
2
(27)
While there can be cases of small2 input errors and errors
adding, such that a large output error results. However in
expectation and averaged over large number of points, smaller
input (sub-network) errors will result in small output
(network) errors.
If
2
∫ ψ˜ 0 (x).b+b0 − f ( x) dµ ( x)
(
=
)
n
∫ (ψ (x).b' +b'
inf
,, fn0,n ,b' ,b0'
n0
f 0
n ,1
0
)
2
(28)
− f ( x ) dµ ( x )
and
)
n
€
(
bounded above by 2rC f
n2
(
+b0
T
T
2
∫ ( f (x) − f ( x)) φ (dx)
subnetworks manage to achieve error
1
n ,k
1
f
n ,k
and
d
(
= 2rC f
)
2
/nd
€ (24)
€
()
1
C ⎛ 1 ⎞ d
(25)
sup s f˜n 0 , f n 0 ≥ κ ⎜ ⎟ .
d ⎝ n ⎠
f ∈ΓC
Hence in spite of the one linear layer subnetworks having
(
)
error bounded below by
, 1 hidden layer
(29)
then
2
∫ (ψ (x) .b+b − f (x)) dµ ( x)
< ∫ (ψ ( x ) .b+b − f ( x )) dµ ( x )
−
n0
0
€
€
(30)
0
Consider
2
2
(ψ (x) .b+b − f (x)) − (ψ ( x) .b+b − f ( x))
= (ψ ( x ) .b+b − f ( x )) − (ψ˜ ( x) .b+b − f ( x))
−(ψ ( x) .b+b − f ( x)) + (ψ˜ ( x) .b+b − f ( x))
= ((ψ ( x) +ψ˜ ( x )) .b+ 2b − 2 f ( x)) ×((ψ ( x) − ψ˜
−((ψ ( x) +ψ˜ ( x)) .b+ 2b − 2 f ( x )) ×((ψ ( x) − ψ˜
+
n0
−
n0
0
0
2
+
n0
2
n0
0
0
2
−
n0
2
n0
0
+
n0
n0
−
n0
n0
(31)
0
0
+
n0
n0
( x)) .b)
0
−
n0
n0
( x)) .b)
>0
(32)
€ each (inner) term of both factors
The last step follows from
of the first (outer) term of€the above expression being greater
than or equal to the corresponding (inner) term of the second
(outer) term.
Hence (11) follows.
Intuitively
difference
⎛⎜
⎞
∫ span⎝φ f n L −1 ,1 (x) ,, φ f n L −1 , n (x) ⎟⎠ dµ (x)
(
∫ span( f
€
2
+
n0
n L −1 ,1
) (
( x),, f n
L −1
,n
between
)
( x)) dµ ( x)
over
and
the
entire
12
hypersphere
function
( 0,1)
f ( x) of
d
can be understood by considering a
which remains equal to some value with
large magnitude
nearly everywhere but dips steeply down
to € at some number of (related to ) points. A linear
€function can only descend and then ascend back up at rate
determined by the linear coefficients and will produce large
error. Note nonlinearity such as
is not convex.
However
convex/concave
is. Furthermore, nonlinearity may still be
between
−( n − 2) π < x < −( n − 1) π
−( n −1) π < x < −nπ .
€
€
and
(vector of weights for the hidden layer unit at instant
also
, .. ,
and we compute
, outputs of hidden layer unit
through
but also
at instants
. When the weights are updated, not only
, .. ,
y H,i,k−1 are recomputed.
instant
history.
) but
, .. ,
are updated and
and
, .. ,
, to be computed at
now make use of recomputed weights and
€
IV. BACK-PROPAGATION THRU TIME AND SPACE
For Back-Propagation Thru Time and Space Gradient
Descent Training, weight
is updated according to
where
is the error,
is the
output of the single output network at the instant when input
has been applied at its input and
is a training inputoutput pair.
Visualization of dependencies considered in the derivative
computation is eased by considering the nodes making up the
neural network to be arranged as columns of stacks of layers
of linear array of nodes (Figure 7). The output layer is at the
top, the hidden layers in the middle and the input layer at the
bottom. The rightmost column corresponds to current time and
the columns to left to successively previous times.
A notion of distance can be associated with any pair of units
and any pair of a unit and a weight. A unit in the (presentinstant, output-layer) is at a distance of 1 from any unit in the
(present-instant, hidden layer
), at a distance 2 from any
unit in (present instant, visible layer), also at a distance 2 from
(previous instant, hidden layer
) and so on. Similarly the
present instant weight
is at a distance 1 from (present
instant, visible layer) unit
, distance 2 from any unit in the
(present instant, hidden layer
Figure 8: Minimum error (over all output layer weights) versus hidden layer
weights for input weights (-1.5, -1.5)
The ability of predict-correct frameworks to “go back” and
compute an improved estimate
than the
offers the possibility that occurrence of an easier to detect
event at instant
can be used to learn a mapping that
produces a better
than when
alone was
available. This is especially relevant to seizure onset
prediction as determining that a seizure is about to occur is
harder for both humans and computers than determining that a
seizure is in progress.
), distance 3 from any unit in
the (present instant, hidden layer
), distance 3 from any
unit in (previous instant, hidden layer
), etc. Hence the
distances here are partially over space and partially over time.
Details of computation of the derivatives are in the
Appendix. These suggest a dynamic programming method for
computing the derivatives comprising of traversing the graph
from inputs to outputs, computing derivative of unit outputs
wrt distance 1 away unit outputs, distance 1 away weights and
distance 2 away weights. Derivatives wrt unit outputs at
distance greater than 1 away and weights at distance greater
than 2 away are computed by taking weighted sums over
derivatives of distance 1 away unit outputs.
A. Adaptation and Autonomous Self Improvement
BPTT implies that at instant
, we not only have
Figure 9: Minimum error (over all output layer weights) versus hidden layer
weights for input weights (-1.5, 1.5)
13
V. EXPERIMENTAL RESULTS
The first set of experiments explored the geometry of the
parameter space first with (hidden layers 1, back-propagation
extent 1), (hidden layers 1, back-propagation extent 2) and
finally with (hidden layers 2, back-propagation extent 1). The
visible, hidden and output layers each had one unit. Units in
visible and output layers had one input from outside or the
previous layer and one bias input. The unit in the hidden layer
had one input from the visible layer, one bias input and one
input that was its own output delayed by one instant. There
was one input to and one output from the network. Hence
there are in total 7 weights in the first, 12 in the second and 10
in the last case. Each weight corresponded to one dimension.
A point in the first parameter space was specified by 7
coordinates. Each coordinate was restricted to one of a set of
discrete values. Range spanned by all of the coordinates was
restricted to a subrange of [-2, 2). For different coordinates,
different number of discrete values evenly distributed across
the range were visited during the systematic exploration.
Figure 12: Error
versus
layer
weights
hidden
and input layer
Figure
11: output
Predicted
versus
noisewith
corrupted
actual
weights corresponding to minimum error
Figure 13: Predicted versus noise corrupted actual
Figure 10: Minimum error (over all output layer weights) versus hidden layer
weights for input weights (1.5, -1.5)
For the first two coordinates, corresponding to the weights
associated with the visible layer unit, typically a small number
(e.g. 2), for the next 3 coordinates corresponding to the hidden
layer, a medium number (e.g. 8) and finally for the last two
coordinates corresponding to the output layer a larger (e.g 32)
discrete values were visited. Example number of values
mentioned above resulted in
different points
in the parameter space to be visited.
For each of the 64 vectors of values of two of the three
hidden layers weights – corresponding to the recurrent input
and input from visible layer - minimum error was determined
separately for each of 4 values of the weight pair
corresponding to weights associated with the visible layer unit.
These are plotted in Figures 8 through 11.
Next with the vector of visible and the hidden layer weights
set to the vector corresponding to minimum error determined
above, the error at the 1024 output layer weight pairs is plotted
in Figure 13.
Figures 8 through 11 have a bearing on the question as to
whether with visible and hidden layer weights initialized
randomly and output layer weights learnt using gradientdescent-backpropagation-through-time suffices to minimize
error to arbitrarily small values. These present empirical
evidence that it is not so. Arbitrary initialization of 'lower'
layers and only learning weights of higher layers lower bounds
the error, such that smaller error cannot be achieved.
Figure 12 has a bearing on the question whether with 'lower'
layer weights set correctly, if the values of the error
corresponding to different values of output layer weights are
considered to lie on a manifold, what does this manifold look
like. As the output layer comprises of sigmoidal units as well,
the manifold is non-convex.
Simulation study was conducted with synthetic and real
data. The synthetic data was the Sine and Mackey Glass
function generated series. In case of the sine function
generated series, the run was conducted upto
.
However during
, prediction was performed but
measurement update was not performed. We see that (Figure
11) predictions match the true data well. Even in case of the
14
Mackey Glass function generated series, which is not simply
periodic, predictions match the true data (Figure 14).
At each instant, input data is received, output is estimated
and, if actual output is made available, the parameters are
updated. It can be argued that with input with stationary
statistics, if parameters do not attain stable values, correct
output being estimated does not mean that input to output
mapping has been learnt. As the plot of evolution of weights
(Figure 14) shows, stable values are attained and the inputoutput mapping is truly being learnt.
The second set of simulation experiments are with real
Electroencephalogram data collected at the Freiburg university
[26]. There are 6 channels in all the recordings because of use
of 6 electrodes. Each recording has been examined by an
expert encephalographer and occurrences of seizure identified.
Each recording covers a nearly continuous length of time that
included several Ictal as well as inter-Ictal intervals. To
shorten the run-times, middle of the inter-seizure intervals
were shortened (e.g. to a unit variance Gaussian random
number with a mean 262144 samples rather than original
intervals exceeding 921600 samples) but intervals leading
upto and the seizure intervals were included unmodified. For
purposes of the simulation experiment, for each recording,
measurement updates were performed through first several
seizures (“training” seizures), measurement updates were
stopped sometime before the occurrence of last few seizures
(“test” seizures) and seizure prediction was compared to
determination made by the expert encephalographer. Table
presents summary of these experiments.
These set of experiments were repeated for prediction ahead
into future by different amounts of time. The three sets
presented in the table corresponded to prediction ahead by 1,
1025 and 2049 instants. High ADRs were achieved for each of
these cases. To study whether, in spite of quantitatively similar
performance for prediction ahead into future by different
amounts of time, whether performance differed qualitatively.
Figure 16, Figure 17 and Figure 18 show close-up of when in
network determination recordings transition from no seizure to
seizure when predicting ahead by 1, 1025 and 2049 instants.
Figure 15: Evolution of weights
TABLE 1: SIMULATION EXPERIMENTS WITH EPILEPTIC SEIZURE DATA
EXPERIMENT
RECORD
RECORD
PREDICT
TRAIN
TEST
NAME
LENGTH
AHEAD
SAMPLES
SAMPLES
1
Pat003
2361858
1
2128306
233552
2
Pat003
2361858
1025
2128306
233552
3
Pat003
2302384
2049
2128304
174080
TABLE 2
REAL EPILEPTIC SEIZURE DATA TRAINING SEGMENTS
EXPERIMENT
TRAIN
TRAIN
SEIZURES
SAMPLES
MARKED
SEIZURE
1
3
111560
2
3
111560
3
3
111560
EXPERIMENT
1
2
3
TABLE 3
REAL EPILEPTIC SEIZURE DATA TEST RESULTS
TRUE
SPC
TRUE
SEN
POSITIVE,
NEGATIVE,
FALSE
FALSE
NEGATIVE
POSITIVE
6224, 175
0.97
227010,
1
143
6308, 91
0.97
227106,
1
47
6224175
0.97
167538,
1
143
ADR
0.986011
0.992786
0.99
The blue curve corresponds to expert determination and the
green curve to network determination. Similarly Figure 19,
Figure 20 and Figure 21 show close-up of when in network
determination recordings transition from seizure to no seizure
when predicting ahead by 1, 1025 and 2049 instants.
Figure 14: Predicted versus noise corrupted actual data for the Mackey Glass
series
15
Figure 16: 1 instant ahead prediction of seizure onset – predicted (green)
versus expert determination (blue)
Figure 18: 2048 instants ahead prediction of seizure onset – predicted
(dots) versus expert determination (solid line)
Figure 19: 1 instant ahead prediction of seizure subsiding
Figure 17: 1024 instants ahead prediction of seizure onset – predicted
(green) versus expert determination (blue)
Figure 20: 1024 instants ahead prediction of seizure subsiding
16
APPENDIX
Output of visible node
⎛
⎞
y v p = f ⎜⎜∑ w v p uq y˜ uq ⎟⎟
⎝ q
⎠
where
is set to 1,
€
(33)
is the
th input and
the weight from input to the visible node
∂y v p
= f ʹ′' ( a) a=a w v p uq
vp
∂y uq
is
. Hence
(34)
and
∂y v
€
Figure 21: 2048 instants ahead prediction of seizure subsiding
€
VI. CONCLUSION
We presented the time-series prediction problem€ and
critically discussed several aspects. Some of the aspects
pertain to applications with stringent requirements such
€ as
Epileptic Seizure onset prediction. We discussed suitability of
Recurrent Neural Networks for time series prediction. We
discuss an architecture for Recurrent Neural Networks that
grows along three dimensions – number of units in layers,
number of hidden layers and extent in number of instants of
back-propagation. We highlighted the similarity between RNN
with greater than 1 instant back-propagation extent and higher
order Markov processes. We prove analytically the fact that
additional hidden layers improve the approximation error rate.
Application of theory of back-propagation gradient descent
leads to a training method called Gradient Descent Backpropagation Through Time-and-Space. We derive a Dynamic
Programming (DP) procedure employing matrix operations for
the training. DP and use of matrix operations makes the
procedure efficient particularly when using data-parallel
libraries and on data-parallel or scalar architectures. The
simulations studies present the geometry of the parameter
space and verify using synthetic data that the network learns
the temporal structure in the data so that parameters converge
while model output displays same dynamic behavior as the
system. Simulation further showed the method to attain nearperfect Average Detection Ratio on real epileptic seizure data.
We have several related ongoing work that are
enhancements to reduce cost of software and Integrated
Circuit implementations while preserving the very good
performance on the discussed as well as other applications.
The enhancements fit number of hidden layers, number of
€ to
units in hidden layers and extent of back-propagation
problem and consider local minima avoidance in gradient
descent training. Applications include Natural Language
Processing, Automatic Speech Recognition and music
analysis.
p
∂w v p uq
= f 'ʹ′ ( a) a= a y uq .
(35)
vp
Alternatively, above can be written in matrix notation as
,
(36)
∂y v
= f ʹ′( a) a=a Wvu and
(37)
v
∂y u
∂y v
= f ʹ′( a) a=a .
(38)
v
∂Wvu
Now
is the vector of inputs to the network,
vector
is the
with a 1 as the additional and the first element
(corresponding to the bias input),
is the matrix of weights
to visible layer units from the inputs,
f ʹ′( a) a=a
(diagonal) matrix of derivatives of vector
evaluated at
and
is the
v
wrt vector
is the matrix of partial
derivatives of outputs of visible€ layer units wrt inputs.
, derivative of vector
wrt matrix
is also a
matrix (and not a 3-dimensional array) as derivative of the
output of a unit in a layer wrt weights to other units in the
same layer are 0.
Note that representing the set of derivatives as matrices is
more than just a matter of compactness. Representing as
matrices eases implementation on scalar processors using
mathematical library functions by reducing the overheads
associated with library function calls and on vector processors
such as GPUs enables gains from the data parallelism.
However in the rest of the following derivation, for ease of
understanding, we continue to use scalar notation and employ
sum over elements of vector as necessary.
Output of hidden layer
node
y h NL , j
NV
⎛ N H
⎞
= f ⎜⎜∑ w h h − y h − + ∑ w hNL , j v p y v p ⎟⎟
N , j N ,p
NL , p
p=1
⎝ p=1 L L
⎠
where
is the output of hidden layer
the weight from hidden layer
previous instant to hidden layer
instant,
is the output of the hidden layer
the previous instant,
node
node
node
(39)
,
at the
at present
node
at
the weight from the hidden layer
17
N L node j to visible node p and y v p the output of visible
propagating T instants, the derivative of final output y z i wrt
node p .
Hence
the weight associated with input q at instant −T to input layer
unit p is given by
NH
∂y hN , p €
€∂y zi
L
= f' ( a) a=a ∑ w zi hN , p
(49)
L
z i€
∂w ( − T ) ( − T )
∂w
€
p=1
( − T) ( −T)
∂y
€ h NL , j
€
€
= f 'ʹ′ ( a) a= a
∂y h −
€
NL , p
€
= f 'ʹ′ ( a) a= a
h NL , j h N−L , p
∂y hN
w€h
−
NL , j h NL , p
,
(40)
€
∂y hNL , j
∂w
h NL , j
h NL , p
and
yh −
(41)
NL , p
€
L,j
= f ʹ′' ( a) a=a
yvp .
(42)
hNL , j
∂w hN , j v p
L
From (33) and (39),
⎛ N H
NV
⎛ N I
⎞⎞
y hN , j = f ⎜∑ w h h ( −1) y h ( −1) + ∑ w hN , j v p f ⎜⎜∑ w v p uq y uq ⎟⎟⎟ (43)
L
⎜
L
⎟
NL , j NL , p
NL , p
p=1
⎝ q=1
⎠⎠
⎝ p=1
Hence
∂y hN , j
L
= f ʹ′( a) a=a
w hN , j v p f ʹ′( a) a=a y uq
L
hN L , j
vp
∂w v p uq
€
€
€
= f ʹ′' ( a) a=a
€
∂y hN
L,j
∂y uq
€
hNL , j
∂y v p
w hN , j v p
L
hNL , j
∑w
p=1
NV
= f ʹ′( a) a=a
hNL , j
∑w
p=1
v
v f ʹ′( a) a=ah
hNL , j p
NL , p
w v p uq
∂y v p
hNL , j p
(44)
∂y uq
Output of hidden layer N L −1,
⎛ N H
⎞
⎟⎟
y hN −1, j = f ⎜⎜ ∑ w h
y
+
χ
hN− −1, p hN− L −1, p
L
⎝ p=1 NL −1, j L
⎠
€
where
€
€
uq
vp
layer unit outputs and unit input weights are assigned indices
so as to facilitate computation of the derivatives.
are the current final outputs,
hidden
are the
layer
€
χ = ∑ wh
N L −1, j
p=1
those with hidden layers
(45)
(51)
NV
⎛ N I
⎞
+ ∑ w hN , p vq f ⎜⎜∑ w vqur y ur ⎟⎟
L
N L ,q
q=1
⎝ r=1
⎠
NH
q=1
€
Hence
∂y hN −1, j
€
∂w v p ur
and finally those with the input layer
hNL , p f (ξ )
ξ = ∑ wh
−
N L , p hN L ,q
y h−
(46)
.
L
= f ʹ′( a) a=a
hNL −1, j
NH
€
∑w
.
p=1
€
∂y hN
L −1,j
∂w v p ur
€
hNL −1, j hNL , p
f ʹ′( a) a=a
hNL , p
w hN
∂y vq
v
L,p q
∑w
a=ah
N L −1, j p=1
hNL −1, j hNL , p
(47)
∂w v p ur
∂y hN
NH
= f ʹ′( a)
L,p
∂w v p ur
(48)
(23) and (26) can in fact be generalized so we may compute
the derivative of a unit output or a weight wrt to a unit output
(say y x1 ) at an arbitrary distance away in terms of the
€
derivative of the same weight or output wrt to (an) output(s)
y x2 distance 1 away, the weight connecting y x2 to y x1 and
€
derivative of the original unit output or weight wrt y x2 . For
example in a network that has N L hidden layers, we are back
€
€
€
€
€
€
unit
outputs,
are the current
visible layer unit outputs. Current outputs are followed by
instant
, instant
, …, instant
outputs where
is
the extent of back-propagation. Weights follow the unit
outputs. Current weights associated with output layer units
occupy indices
,
(50)
NH
€
uq
Above suggests a dynamic programming method for
computing the derivatives comprising of traversing the graph
from inputs to outputs, computing derivative of unit outputs
wrt distance 1 away unit outputs, distance 1 away weights and
distance 2 away weights. Derivatives wrt unit outputs at
distance greater than 1 and weights at distance greater than 2
are computed by taking weighted sums over corresponding
derivatives of distance 1 away unit outputs. The distances here
are partially over space and partially over time.
In the joint-state-and-parameter-vector
, the individual
current
∂w v p uq
NV
= f ʹ′( a) a=a
vp
(52)
As with outputs, current weights are followed by instant
through instant
weights.
Before computing new current outputs, the current through
instant
outputs are copied to instant
through instant
outputs. Similarly before updating, the current through
instant
weights are copied to instant
through
instant
weights.
However significant time-efficiency is achieved by noting
that no actual copying needs to be performed. The unit outputs
and weights sections are treated as
(instants) long
circular buffers and the current instant pointer is simply
advanced one instant!
The row and column indices of matrices of derivatives of
unit outputs and weights wrt upstream unit outputs and
weights are also similarly incremented circularly when
advancing to next instant. This makes it unnecessary to
18
physically copy weights. As operations of matrix-vector
product and vector-vector inner product involves summing
over products of pairs of elements and the sum operation is
order-independent, the matrix operations are not affected. This
makes the time-complexity when using back-propagation
extent greater than 1 same as the time complexity when using
back-propagation extent equal to 1.
With N O long network outputs vector y z and the
derivatives of weights wrt network outputs arranged in a
N O × N W matrix J , the updated weights are given
by w ← w − ηJ( yˆ z − y z ) .
€
€
€
€
Thus, the space complexity is
. With
€
a constant independent of network dimensions, this is
linear in network dimension. With access to previous outputs
and weights requiring
time, the time-complexity of
computations at each instant is
under normally
valid assumptions of
.
and
ACKNOWLEDGMENT
Piyush was a summer intern at the Nano-Electronics Research
Laboratory in the Electrical Engineering Dept of Purdue
University for a part of this work and is grateful for support
and encouragement received from Prof K. Roy and Dr S.
Raghunathan.
REFERENCES
[1]
I. Arasaratnam et. al., “Discrete Time Nonlinear Filering Using GaussHermite Quadrature,” Proc. of IEEE, vol. x, no. x, pp. 953-977, May,
2007.
[2] Andrew S. Barron, “Universal Approximation Bounds for
Superpositions of a Sigmoidal Function,” IEEE Transactions on
Information Theory, vol. 39(3), pp. 930-945, 1993.
[3] Belhouari S. B. et. al., “Gaussian Process for Nonstationary Time Series
Prediction,” in Computational Statistics and Data Analysis, xth ed. City
of Publisher, Country if not USA: Elsevier, 2004, ch. x, sec. x, pp. xxx–
xxx.
[4] Bengio Y. et. al., “Deep Learning of Representations for Unsupervised
and Transfer Learning,” Journal of Machine Learning Research (Proc.
Unsupervised and Transfer Learning Challenge and Workshop), Vol.
27, 2012.
[5] R. van der Merwe and E. Wan, “Sigma-point Kalman filters for
probabilistic inference in dynamic state-space models,” in Proceedings
of the Workshop on Advances in Machine Learning, City of Conf.,
Abbrev. State (if given), 2003, pp. xxxxxx.
[6] R. van der Merwe, “Sigma-point Kalman filters for probabilistic
inference in dynamic state-space models,” Ph.D. thesis, OGI School of
Science & Engineering, Oregon Health & Science University, City of
Univ., Portland, OR, USA, April 2004.
[7] Duchi et. al. , “Ergodic Subgradient Descent,” Allerton Conference on
Communication, Control and Computing, 2011.
[8] Gallant A. R. et. al., “There exists a neural network that does not make
avoidable mistakes,” Proceedings of the Second Annual IEEE
Conference on Neural Networks, Vol 2, I.657 – I.664, 1988.
[9] Greene B R et al, “A comparison of quantitative EEG features for neonatal seizure detection,” Clinical Neurophysiology, 2008.
[10] Hornik K. et. al., “Universal Approximation of an Unknown Function
and its Derivatives using a Multilayer Feedforward Networks,” Neural
Networks, Vol 3, 550-560, 1990.
[11] A. H. Jazwinski, “Stochastic Processes and Filtering Theory,”
Academic Press, 1970.
[12] S. J. Julier, J. K. Uhlmann, and H. F. Durrant-Whyte, “A new method
for the nonlinear transformation of means and covariances in filters and
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
estimators,” IEEE Transactions on Automatic Control, vol. 45(3), pp.
477–482, March 2000.
Kimball B. F., “Note on Computation of Orthogonal Predictors,” The
Annals of Mathematical Statistics, Vol. 24, No. 2, 1953.
F. Le Dimet et. al. “Second Order Information in Data Assimilation,”
Journal of the American Metrological Society, Vol. 130, No. 3, pp 629648, March 2002.
Loh P.-L. et. al. “High-dimensional regression with noisy and missing
data: Provable guarantees with non-convexity,” available at
http://arxive.org/abs/1104.4824, 2012.
Niedermeyer E et. al., “Electroencephalography: Basic Principles,
Clinical Applications, and Related Fields,” Liipicott Williams &
Wilkins, 2004.
Raghunathan S et al, “ A hardware-algorithm co-design approach to
optimize seizure detection algorithms for implantable applications,”
Elsevier Journal of Neuroscience Methods, 2010.
Rhew H-G et al, “A Wirelessly Powered Log-based Closed-Loop Deep
Brain Stimulation SOC with Two-Way Wireless Telemetry for
Treatment of Neurological Disorders,” Symposium on VLSI Circuits,
2012.
F. Sandblom et. al., “Marginalized Sigma Point Form,” , 2011.
Sha H et al, “A micro-controller-based implantable nerve stimulator
used for rats,” IEEE Engineering in Medicine and Biology Society
Conference, 2006.
Subasi A., “Classification of EEG signal using neural network and
logistic regression,” Computer Methods and Programs in Bio-Medicine,
Vol. 78, No. 2, 2005.
Sudderth E. B. et. al. “Nonparametric Belief Propagation,”
Communications of the ACM, Vol. 53, No. 10, pp 95-103, October
2010.
Sun F. T. et. al., “Responsive Cortical stimulation for the treatment of
epilepsy,” Neurotherapeutics, Vol. 5, 2008.
Werbos P. J., “Backpropagation Thru Time: what it does and how to do
it,” Proceedings of the IEEE, Vol.78, No. 10, 1990.
Yeager D et al. “A Fully-Integrated 10.5 uW Miniaturized (0.125mm2)
Wireless Neural Sensor,” Symposium on VLSI Circuits, 2012.
Seizure Prediction Project, http://epilepsy.uni-freiburg.de/friburgseizure-prediction-project.
Sharat C Prasad (M’87) received the B.Tech. degree in
Electronics and Communication Engineering from I.I.T.
Kaharagpur, India in 1983, M.E. in Computer Science and
Automation from I.I.Sc. Bangalore India in 1987 and Ph.D. in
Electrical Engineering from University of Texas at Dallas in
2005.
He was an Intern at the Tata Institute of Fundamental
Research during the summer of 1982, worked at the Indian
Subsidiary of International Computers Limited, U.K. from
1983 until 1985, at Texas Instruments from 1987 until 1997, at
Cisco Systems from 1998 until 2006, at Google Inc from 2007
until 2012 and is with Best3 Corp since 2012.
Dr Prasad is a member of IEEE and ACM. He has authored
or co-authored one book, over 5 journal, over 10 conference
and over 30 patents.
Piyush Prasad was born in Plano, Texas, in 1995. He is
expecting to receive his B.S. degree in Biomedical
Engineering from Washington University, St Louis, MO in
2017.
He was a summer intern at the Nano-electronics laboratory
of Dept of Electrical Engineering at Purdue University.
Mr Prasad was a recipient of several awards at a national
Robotics league and is the co-author of a patent.
19
| 9cs.NE
|
A Two-Step Distribution System State Estimator
with Grid Constraints and Mixed Measurements
Miguel Picallo∗† , Adolfo Anta∗ and Ara Panosyan∗
∗ General
arXiv:1703.10815v1 [cs.SY] 31 Mar 2017
Electric Global Research
Munich, Germany
Email: miguel.picallocruz|anta|[email protected]
Abstract—In this work we study the problem of State Estimation (SE) in large-scale, 3-phase coupled, unbalanced distribution
systems. More specifically, we address the problem of including
mixed real-time measurements, synchronized and unsynchronized, from phasor measurement units and smart meters, into
existing solutions. We propose a computationally efficient twostep method to update a prior solution using the measurements,
while taking into account physical constraint caused by buses
with no loads. We test the method on a benchmark test feeder
to illustrate the effectiveness of the approach.
I. I NTRODUCTION
The operation of power networks requires accurate monitoring of its state: bus voltages, line currents, consumption and
generation. This is specially relevant in transmission networks
where volatile and distributed generation and consumption
cause bidirectional power flow and voltage drops. Normally,
State Estimation (SE) consists in estimating the bus voltage
phasors and then computing the currents and loads using
the power flow equations derived from the structure of the
network, represented by the admittance matrix. SE is typically
performed by taking several measurements and solving a
weighted least squares problem using an iterative approach
like Newton-Raphson [1]–[4].
However, SE has not yet been studied and implemented
deeply in distribution networks, since they used to have a
simple radial structure with a single source bus connected
to the grid and injecting power, being the point of common
coupling (PCC). Nowadays, this is changing, since the increasing number of sources of distributed generation like PVpanels, electrical vehicles, etc. inject power in the network,
causing bidirectional power flow. Consequently, SE becomes
necessary in distribution networks. One of the major difficulties for its implementation lies in the structural difference
between transmission and distribution networks: in the latter,
unbalanced loads and lower X/R ratios cause coupled phases.
Therefore, the power flow equations need to be solved for the
3 phases simultaneously, and fast methods like fast decoupled
power flow [5] cannot be used. Another limitation is the lack
of sufficient real-time measurement units in distribution networks: despite Phasor Measurement Units (PMU) and smart
meters are being introduced, their high cost prevent installing
the required number of sensors to make the system fully
This project has received funding from the European Union’s Horizon 2020
research and innovation programme under the Marie Skłodowska-Curie grant
agreement No 675318 (INCITE).
Bart De Schutter†
† Delft
Center for Systems and Control
Delft, The Netherlands
Email: [email protected]
observable. Some methods have been proposed to perform
SE by exploiting the network structure and nature of the
measurements, such as branch-current-based SE [6]–[8]. Other
methods use the few measurements available to improve the
SE accuracy of previous methods based on load assumptions
only [9]–[11]. Yet, most of this work is limited to transmission
networks, small-scale distribution networks, or single-phase
networks. Some recent work extends it more general networks,
but assumes measurement availability at every bus [12].
Our main contributions can be summarized as follows:
First, we include equality constraints more efficiently by a
dimension reduction. Then, we extend some recently proposed
methods that split the problem [9], to large-scale, 3-phase
coupled, unbalanced and constrained distribution systems, with
linear as well as nonlinear real-time measurements considering
different sources of measurement noises.
The rest of the paper is structured as follows: Section II
presents some background about power networks. Section III
discusses the different types of measurements. Section IV
presents the standard methodologies for general SE. Section V
proposes the linear dimension reduction method and extends
recent methodologies. Section VI defines the test case simulation and shows its results. Finally, Section VII presents the
conclusions and proposed future work.
II. D ISTRIBUTION S YSTEM M ODEL
A distribution system can be modeled as a graph G =
(V, E, W) with nodes V = {1, ..., Nbus }, edges E = {(vi , vj ) |
vi , vj ∈ V} and edge weights W. The nodes represent the
buses in the network, where power is injected or consumed, the
edges the electrical connections in the grid, and their weights
express the impedance between the buses, which is determined
by the length and type of the line cables.
In 3-phase networks buses may have up to 3 phases, so that
the voltage Vi at bus i lives in CPi , where Pi is the number of
its phases. Moreover, Vsource ∈ C3 are the known voltages at
2π T
the source bus with reference angles Vsource = [0 −2π
3
3 ] .
The state of the network is then typically represented with
the vector of the remaining bus voltages V ∈ CN where N
depends on the number of buses and phases per bus.
The impedance of the connection between buses i and j is a
square symmetric matrix Zi,j ∈ CPi,j ×Pi,j where Pi,j is the
number of phases in the connection i → j. Since phases are
coupled, Zi,j has off-diagonal terms different from 0. These
impedance matrices are used to compute the admittance matrix
Y ∈ C(N +3)×(N +3) , which defines the structure of the system.
Assuming lines with no shunt admittance, the element of Y
corresponding to the connection between phase l of bus i and
phase k of j can be expressed as:
(
− (Zi,j1 )l,k
if i 6= j
P
(Y )il ,jk =
(1)
1
if i = j
m6=i (Zi,m )l,k
where il is the index of phase l of bus i, similarly for jk ,
(·)i,j denotes the i, j element of a matrix, and (Y )il ,jk = 0
if no connection exists between these phases of these buses.
Using the admittance matrix Y , the formulas of the power
flow equations to compute the current I and the load S at all
buses are expressed as:
Isource
Vsource
Y Y
Vsource
=Y
= a b
I
V
Yc Yd
V
(2)
¯
S = diag(I)V
¯ denotes the complex conjugate, diag(·) represents
where (·)
the diagonal operator, converting a vector into a diagonal
matrix. Separating Y in blocks according to the indices of the
source bus Vsource , see (2), the voltage V for the non-source
buses can be rewritten as:
V = Yd−1 I − Yd−1 Yc Vsource
(3)
= Yd−1 (diag(V̄ ))−1 S̄ − Yd−1 Yc Vsource
so that the voltage under no load (I = 0, S = 0) is:
V0 = −Yd−1 Yc Vsource
(4)
III. M EASUREMENTS
Several different sources of information are available to
solve the SE problem:
1) Pseudo-measurements: Load estimations for every hour,
based on predictions and known installed load capacity
at every bus, and represented by Spseudo . Since they are
estimations rather than actual measurements, their noise is
modeled with a relative large standard deviation (a typical
value is σpseudo ≈ 50%):
S = Spseudo + <{Spseudo }ωpseudo,< + j={Spseudo }ωpseudo,=
with ωpseudo,< , ωpseudo,= ∼ N (0, σpseudo Id,N ), where
<{·}, ={·} denote the real and imaginary part, and Id,N
the identity matrix with size N × N . The noise covariance
of these measurements is then:
ΣSpseudo = E[(S − Spseudo )(S − Spseudo )∗ ]
(5)
2
2
= σpseudo
diag( Spseudo )
where E[·] denotes expectation, (·)∗ complex conjugate
transpose, |·| element-wise magnitudes of a complex vector
and (·)2 element-wise square.
2) Virtual measurements: Buses with no loads connected,
acting only as a connection [13]. They can be modeled
as physical constraints for the voltage states. Let ε =
{i, · · · , j} be the set with their indices, then:
(S)ε = 0, (I)ε = 0
(6)
where (·)ε denotes the elements at indices in ε.
3) Real-time measurements: Voltages and currents measurements from PMU, smart meters, or conventional remote
terminal units. There are two kinds: GPS-synchronized
measuring magnitude and phase angle, and unsynchronized
for only magnitude. We model the noises with a low
standard deviation for magnitude and angle, σmag ≈ 1%
and σang ≈ 0.01rad respectively, according to the IEEE
standard for PMU [14].
Synchronized measurements can be expressed using a linear approximation (see Appendix) with magnitude and angle noise caused by the measurements and by an imperfect
synchronization. For a number of NmeasL measurements
zmeasL ∈ CNmeasL we have:
zmeasL ≈ CmeasL V + diag(CmeasL V )(ωmag + jωang ) (7)
where ωmag , ωang ∼ N (0, σmeas Id,NmeasL ) using σmeas =
σang = σmag (since they are similar [14]), and CmeasL is
the matrix mapping state voltages to measurements. For
element j of CmeasL V measuring at phase l of bus i:
(C
V ) = (CmeasL )j,• V =
measL j
for a voltage measurement
Vil
(8)
(Y )il ,• V
for a current measurement
(Y
)
(V
−
V
)
for
a
branch-current
i
→
m
il ,ml
il
ml
measurement
where (·)j,• denotes row j.
Unsynchronized measurements have a nonlinear relation
CmeasNL (V ) with the voltages V . For NmeasNL measurements zmeasNL ∈ RNmeasNL :
zmeasNL = CmeasNL (V ) + diag(CmeasNL (V ))ωmeasNL (9)
with ωmeasNL ∼ N (0, σmeas Id,NmeasNL ). For element j of
CmeasNL (V ) measuring at phase l of bus i:
(CmeasNL (V ))j =
for a voltage measurement
Vil
for a current measurement
(Y )il ,• V
(Y
)
(V
−
V
)
for a branch-current i → m
i
,m
i
m
l
l
l
l
measurement
(10)
Since the measurement noises are small, their covariance
matrices can be approximated using the measurements:
2
2
2
2
diag(|zmeasL | )
diag(|CmeasL V | ) ≈ 2σmeas
ΣmeasL = 2σmeas
2
2
2
diag(zmeasNL
)
diag(CmeasNL (V )2 ) ≈ σmeas
ΣmeasNL = σmeas
IV. S TANDARD M ETHODOLOGY FOR S TATE E STIMATION
In a power network, SE consists in estimating the state of
the network, represented by the bus voltages vector V . The
standard methodology for SE computes the maximum likelihood estimation by solving a constrained nonlinear weighted
least squares problem with all measurements and estimations:
(Spseudo )εc , zmeasL , zmeasNL , their noise covariance and the system constraints (Spseudo )ε = 0, where εc denotes the complimentary of ε. Typically this is solved using the NewthonRaphson method [1]. However, since this requires the use of
gradients and the power flow formulas in (2) are not holomorphic, i.e. complex differentiable, the problem is solved in real
T
variables using a polar representation Vpolar = [|V | ∠V T ]T :
min z − h(Vpolar ) W −1 s.t. g(Vpolar ) = 0
(11)
Vpolar
2
kxkA =
where
x∗ Ax is the norm with respect to the real
positive-definite matrix of weights A, z is the vector of mea-
surements in rectangular coordinates and h(·) is a nonlinear
function mapping Vpolar to measurements:
<{Sεc (V (Vpolar ))}
<{Spseudo,εc }
={Sεc (V (Vpolar ))}
={Spseudo,εc }
z=
<{zmeasL } , h(Vpolar ) = <{(CmeasL V (Vpolar ))}
={(C
={zmeasL }
measL V (Vpolar ))}
zmeasNL
CmeasNL (V (Vpolar ))
where S(V ) represents S as a function of V , V (Vpolar ) is
defined similarly. The function g(·) indicates the null loads:
g(Vpolar ) = [<{[Sε (V (Vpolar ))}T ={[Sε (V (Vpolar ))}T ]T
and the weight matrix W −1 is the inverse of the measurement
noises in rectangular variables
!
2
c
<{(S
)
}
pseudo ε
2
diag
0
0
σpseudo
={(Spseudo )εc }2
W =
0
Σrect,measL
0
0
0
ΣmeasNL
where 0 denotes a matrix with 0 entries, and
2
Σrect,measL ≈ σmeas
"
#
2
diag(|zmeasL | )
2diag(<{zmeasL }={zmeasL })
·
2
2diag(<{zmeasL }={zmeasL })
diag(|zmeasL | )
This problem (11) is typically solved iteratively using Lagrangian multipliers λ [15], so that the size of the optimization
T
variables is given by [Vpolar
λ]T ∈ R2N +2|ε| .
V. P ROPOSED M ETHODS FOR S TATE E STIMATION
In this section we propose new SE methods that significantly
improve the standard methodology in (11). First, we present a
method that simplifies the problem and reduces its complexity,
while obtaining an equal solution; then, we propose another
method that significantly reduces the computation cost while
achieving approximately the same accuracy.
A. Weighted least squares in subspace
Previous methods either do not consider constraints, or
do it by introducing Lagrange multipliers as in (11), which
increases the complexity of the algorithm. Here we propose
an alternative method to include the constraints using a dimension reduction through a linear transformation that produces
feasible solutions. The space of feasible solutions is:
{V | Iε (V ) = (Yd )ε V + (Yc )ε Vsource = 0}
(12)
= {V | ∃x ∈ CN −|ε| s.t. V = F x + Vp }
where Vp satisfies Iε (Vp ) = 0, and F is such that:
(Yd )ε F = 0
(13)
Therefore, V0 is chosen for Vp , and F is built using an
orthonormal basis of the kernel of (Yd )ε e.g. by computing
the Singular Value Decomposition of (Yd )ε = USV ∗ , and
taking the columns of V with null singular value. Then,
F = ker((Yd )ε ) and F ∗ F = Id , where ker(·) denotes the
kernel subspace. If we consider the rectangular representation
of voltages Vrect = [<{V }T ={V }T ]T instead of the polar
representation Vpolar , the subspace equations can still be represented linearly in real variables:
{Vrect | <{Iε (Vrect )} = 0, ={Iε (Vrect )} = 0}
(
)
<{(Yd )ε } −={(Yd )ε }
<{(Yc )ε Vsource }
V +
=0
= Vrect
={(Yd )ε } <{(Yd )ε } rect
={(Yc )ε Vsource }
= {Vrect | ∃x̃ ∈ C2N −2|ε| s.t. Vrect = F̃ x̃ + Ṽp }
(14)
where now
!
<{V0 }
<{(Yd )ε } −={(Yd )ε }
Ṽp =
, F̃ = ker
(15)
={V0 }
={(Yd )ε } <{(Yd )ε }
Then the problem in (11) becomes:
Vrect = F̃ arg min z − h̃(F̃ x̃ + Ṽp )
x
W −1
+ Ṽp
(16)
where h̃(·) is h(·) in rectangular coordinates, using V (Vrect )
instead of V (Vpolar ), so that h̃(Vrect ) = h(Vpolar ). This again
can be solved using the Newton-Raphson method:
x̃k+1 = x̃k + ∆x̃k
with
∆x̃k = (HkT W −1 Hk )−1 HkT W −1 (z − h̃(F̃ x̃k + Ṽp ))
Hk = ∇x̃ h̃(F̃ x̃ + Ṽp ) |x̃k = ∇Vrect h̃(Vrect ) |Vrect,k F̃
Note that now no Lagrangian multipliers are required. We
eliminate the equality constraints by embedding the solution
into a smaller subspace of feasible solutions. This simplifies
the method and reduces its computational cost, since we
reduce the size of the optimization variables x̃ ∈ R2N −2|ε| .
B. Power flow solution plus optimal linear update
The methods (11) and (16) are iterative, and thus can
become computationally expensive as the size of the network
N increases. This may be a limitation in large distribution
systems when fast calculations are required every few seconds.
Therefore, we propose an alternative method splitting the
problem into two parts as in [9]: first, we use only the
constraints and the load estimates Spseudo available beforehand
to solve the power flow problem offline and to obtain a prior
solution defined as Vprior ; then, we update in real-time the
solution using the measurements zmeasL , zmeasNL to compute a
better estimate, represented by Vpost . We also prove that several
alternatives for the update are equal under certain conditions.
1) Obtaining a prior solution:
a) The prior estimate can be computed using the previous
method in (16) without real-time measurements, so that
z = [<{Spseudo,εc }T ={Spseudo,εc }T ]T and h̃(Vrect ) =
[<{Sεc (V (Vrect ))}T ={Sεc (V (Vrect ))}T ]T . This method
will estimate real and imaginary parts Vrect,prior , the estimation error covariance of which can be approximated by:
−1
Σrect,prior ≈ F̃ (HkT Wrect
Hk )−1 F̃ T
(17)
The covariance matrix in complex variables is:
Σprior = Σrect,prior,<< + Σrect,prior,==
(18)
−jΣrect,prior,<= + jΣTrect,prior,<=
where Σrect,prior,<< , Σrect,prior,== indicate the covariance of
real and imaginary parts respectively, and Σrect,prior,<= the
covariance between real and imaginary parts.
b) An alternative approach using complex variables would be
the method proposed in [9]. This method can be extended
to a 3-phase coupled, unbalanced system using the voltage
V0 under no loads defined in (4). There, V0 is used as
initial value in this iterative method, where the iterations
are inspired by (3), i.e. for the k-iteration we have:
Vk+1 = Yd−1 (diag(V̄k ))−1 S̄ + V0
(19)
The estimation error covariance can be approximated by
the covariance after the first iteration, which corresponds
to the linear approximation in [16]:
Σprior ≈ Yd−1 (diag(V̄0 ))−1 ΣSpseudo (diag(V0 ))−1 (Yd−1 )∗
Remark 1. The estimated current at any iteration k is:
Ik+1 = (diag(V̄k ))−1 S̄
(20)
If the loads satisfy (S)ε = 0, then (Ik+1 )ε = 0. This means
that there exists a vector x ∈ CN−|ε| s.t. Vk = F x + V0
and thus Vk is feasible ∀k w.r.t. (12). Note that we use
F ∈ CN ×(N −|ε|) instead of F̃ since Vk ∈ CN .
Remark 2. This method can be extended to real variables
to get Σprior in rectangular coordinates:
<{Spseudo }
Vrect,k+1 = Bk
+ Vrect,0
(21)
={Spseudo }
with
# !!−1
"
#
"
2
<{Yd−1 } −={Yd−1 }
|Vk |
Bk =
diag
2
={Yd−1 } <{Yd−1 }
|Vk |
diag(<{Vk }) diag(={Vk })
·
diag(={Vk }) −diag(<{Vk })
and estimation error covariance:
!
2
<{S
}
pseudo
2
B0T (22)
Σrect,prior = B0 σpseudo
diag
={Spseudo }2
The advantage of approach b) w.r.t. a) is that b) can be solved
in complex variables. In both cases, since the load estimates
Spseudo are known beforehand, this part of the problem can
be computed offline, taking this iterative computational cost
outside the real-time operation of the system.
2) Updating with only synchronized measurements:
If there are not any unsynchronized measurements, the mapping from the state variables to the real-time measurements (8)
is linear and thus holomorphic in x ∈ CN −|ε| . Consequently,
the update can be solved using complex variables to obtain a
linear expression. Different alternative methods are possible,
here we present two of them and later show their equivalence:
a) If the prior solution Vprior is feasible, then there exists a
vector xprior s.t. Vprior = F xprior +V0 . It can be computed as
xprior = F ∗ (Vprior − V0 ), with estimation error covariance
F ∗ Σprior F , or xrect,prior = F̃ T (Vrect,prior − Vrect,0 ) for real
variables. Then the maximum likelihood update can be
computed using weighted least squares:
2
x − xprior
Vpost = V0 +F arg min
CmeasL (F x + V0 ) − zmeasL
x
−1
Wpost
(23)
−1
where the optimal weights Wpost
are:
∗
F Σprior F 0
Wpost =
(24)
0
ΣmeasL
Taking derivatives at (23) we obtain the solution:
∗
−1
Vpost = F ((F ∗ Σprior F )−1 + F ∗ CmeasL
Σ−1
measL CmeasL F )
∗
·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
Σ−1
measL
·(zmeasL − CmeasL V0 )) + V0
(25)
Remark 3. Considering the first order (and thus linear)
approximation of the measurement function h̃(·), in [10] is
proven that solving the weighted least squares problem in
one step (16) and in two steps: weighted least squares for
power flow plus update linear (23), produce the same exact
solutions. Consequently, we can expect a similar accuracy
between the one step and the two steps methods.
b) An alternative proposed [9] is to improve the solution using
an minimum variance linear update:
Vpost = Vprior + K zmeasL − CmeasL Vprior
(26)
Then the error covariance of estimation Vpost is:
∗
Σpost = Σprior + K(ΣmeasL + CmeasL Σprior CmeasL
)K ∗
(27)
∗
∗
−KCmeasL Σprior − Σprior CmeasL K
and the optimal gain K is computed minimizing the
expected error E[(Vpost − V )∗ (Vpost − V ) | Vprior , zmeasL ] =
E[tr(Σpost ) | Vprior , zmeasL ]:
K = arg minK E[tr(Σpost ) | Vprior , zmeasL ]
(28)
∗
∗
= Σprior CmeasL
(CmeasL Σprior CmeasL
+ ΣmeasL )−1
Proposition 1. If Vprior is a feasible solution to (12), as the
solutions in Section V-B1, then the solution (26) satisfies (12).
Proof. For Vprior : (Yd )ε Vprior +(Yc )ε Vsource = 0. Also, since the
first term in K is F , because K starts with Σprior = F Σxprior F ∗ ,
we have (Yd )ε K = 0 due to (13). Therefore, Vpost from (26)
satisfies (12):
(Yd )ε Vpost + (Yc )ε Vsource
= (Yd )ε Vprior + (Yc )ε Vsource + (Yd )ε K(zmeasL − CmeasL V )
=0+0=0
Proposition 2. If Vprior is a feasible solution to (12), then both
alternatives (26) and (25) are equal.
Proof. We start from expression (25), using Woodbury’s identity [17] on the inverse term and (28) for F we get:
∗
Vpost = F ((F ∗ Σprior F ) − (F ∗ Σprior F )F ∗ CmeasL
∗
·(CmeasL Σprior CmeasL
+ ΣmeasL )−1 CmeasL F (F ∗ Σprior F ))
∗
·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
Σ−1
measL
·(zmeasL − CmeasL V0 )) + V0
= (F (F ∗ Σprior F ) − KCmeasL F (F ∗ Σprior F ))
∗
·((F ∗ Σprior F )−1 xprior + F ∗ CmeasL
Σ−1
measL
·(zmeasL − CmeasL V0 )) + V0
= V0 + (Id − KCmeasL )F xprior
∗
+(Id − KCmeasL )F (F ∗ Σprior F )F ∗ CmeasL
Σ−1
measL
·(zmeasL − CmeasL V0 ))
(29)
Doing some manipulation to the formula of K in (28) we get:
∗
(Id − KCmeasL )Σprior CmeasL
Σ−1
(30)
measL = K
So that finally we can convert the expression in (25) to (26):
Vpost = V0 + (Id − KCmeasL )F xprior + K(zmeasL − CmeasL V0 )
= Vprior + K(zmeasL − CmeasL Vprior )
The maximum likelihood estimator of weighted least square
coincides with the minimum variance estimator of the linear
update, since we consider gaussian distributions for the noise.
3) Updating with unsynchronized measurements:
If there are nonlinear measurements, since the magnitude of a
complex number as in (10) is not an holomorphic function, the
update step needs to be solved using real variables. It becomes:
<{zmeasL − CmeasL V (Vrect,prior )}
Vrect,post = Vrect,prior + K̃ ={zmeasL − CmeasL V (Vrect,prior )}
zmeasNL − CmeasNL (V (Vrect,prior ))
(31)
Given that
<{CmeasL V (Vrect,prior )}
<{CmeasL } −={CmeasL }
=
V
={CmeasL V (Vrect,prior )}
={CmeasL } <{CmeasL } rect,prior
and defining
<{CmeasL } −={CmeasL }
CmeasLNL = ={CmeasL } <{CmeasL }
∇Vrect CmeasNL (V (Vrect )) |Vrect,prior
Σrect,measL
0
ΣmeasLNL =
0
ΣmeasNL
The first-order approximation of the estimation error covariance of this nonlinear filter is:
T
Σrect,post ≈ K̃(ΣmeasLNL + CmeasLNL Σrect,prior CmeasLNL
)K̃ T
T
+Σrect,prior − K̃CmeasLNL Σrect,prior − Σrect,prior CmeasLNL
K̃ T
(32)
where now the optimal gain K̃ is:
T
T
K̃ = Σrect,prior CmeasLNL
(CmeasLNL Σrect,prior CmeasLNL
+ΣmeasLNL )−1
Remark 4. As in Proposition 2, it can be proven that the
first-order approximation of the estimation error covariance
(32) of (31) is equal to the one approximating the estimation
error covariance of (23) including nonlinear measurements,
computed similarly as in (17).
VI. S IMULATION R ESULTS
A simulation of 24 hours with 15 min intervals is run on
a test case to compare the different methods for SE. Here we
describe the settings of this simulation and analyze the results.
A. Settings
• System: We use the 123-bus test feeder [18] available online
[19], see Fig. 1. This is a challenging example, since it is 3phase coupled, unbalanced, and larger than other examples
in the literature [7], [9].
• Measurements (see Fig. 1): Voltage measurements (red
solid, circle for phasor, square for magnitude only) are
placed at buses 79, 95, 83 and 300, current measurements
(blue dashed, circle and square) at bus 65 and 48, and branch
current measurements (blue dashed arrow) at branch 150
(after the regulator) → 149. The standard deviation used to
simulate noise in the measurements is σmeas = 0.01 [14].
Fig. 1: 123-bus test feeder with measurements location. The
network image has been taken from [19].
•
•
•
This sensors locations are chosen since they correspond to
nodes with big loads and/or at the end of a feeder.
Load Profiles: Historical data from a German local utility is
scaled so that the mean load is the base load of the 123-bus
test feeder and the relative standard deviation is σpseudo =
50% at every hour. The values of Spseudo are an average of
the bus loads for every hour.
Metric: For every simulation time step t, a normalized root
mean square error (nRMSE) is determined for the voltages
estimated by every method j:
v
u
N
1 u
1 X
2
t
nRMSEt,Vmethodj =
Vmethodj ,t,i − Vt,i
|Vbase | N i=1
where Vt,i , Vmethodj ,t,i denote the i-element at time t of
the actual voltage and the voltage estimated by method j
respectively. This method quantifies the percentage deviation
of the solutions relative to the network’s base voltage Vbase .
Nomenclature: The solutions compared are denoted as:
VWLS , VWLSNL , which correspond to the solutions of (16)
without and with nonlinear measurements respectively; Vprior
corresponding the prior solution (19) before the update; and
Vpost , VpostNL corresponding the updated solutions (23), (31).
B. Results
Fig. 2a shows how a solution with real-time measurements
like Vpost clearly outperforms the prior solution Vprior using
only load estimations. This shows that even when few measurements are available, in this case only 4 measurements in
a 123-bus network, using them can dramatically increase the
accuracy of the prediction. Moreover, Fig. 2b shows that the
two-step solutions Vpost and VpostNL perform statistically as well
as the respective weighted least squares solutions VWLS and
VWLSNL . However, the methods using weighted least squares
are computationally much more expensive than the others,
specially the one with nonlinear measurements VWLSNL , see
Fig. 3 for the comparison of different computation times.
These algorithms are coded in Python and run on an Intel
Core i7-5600U CPU at 2.60GHz with 16GB of RAM.
To summarize, the solutions including real-time measurements
improve the accuracy of the estimation, specially if adding extra nonlinear measurements. However, the two-step approach
8
0.9
7
0.8
0.7
nRMSE [%]
nRMSE [%]
6
5
4
3
2
0.6
0.5
0.4
0.3
0.2
1
0
1
where ωmag ∼ N (0, 1%|u|) (or |u|
ωmag ∼ N (0, 0.01)),
ωθ ∼ N (0, 0.01rad), according to the standards in [14]. Using
trigonometric identities for cos(θu + ωθ ) and sin(θu + ωθ );
using approximations sin(ωθ ) ≈ ωθ and cos(ωθ ) ≈ 1,
since |ωθ | 1; and neglecting second order terms of noise
O(ωθ ωmag ), then we have:
ũ ≈|u| (cos(θu ) + j sin(θu )) + ωmag (cos(θu ) + j sin(θu ))
+ωθ |u| (− sin(θu ) + j cos(θu ))
1
1
= u + ωmag |u|
u + jωθ u = u + u |u|
ωmag + jωθ
0.1
Vprior
Vpost
0.0
VWLS
VWLSNL
(a)
Vpost
VpostNL
(b)
Fig. 2: Box-plots of MAPE errors for the different solutions.
The red line indicates the median, the red square the mean.
0.35
0.30
time [s]
0.25
0.20
0.15
0.10
0.05
0.00
VWLS
VWLSNL
Vpost
VpostNL
Fig. 3: Box-plots of execution times for the different solutions.
The red line indicates the median, the red square the mean.
has a smaller computational cost, and remains small after
adding nonlinear measurements.
VII. C ONCLUSIONS
We have proposed a method for state estimation in largescale, 3-phase coupled, unbalanced and constrained distribution systems, with mixed measurements: phasor-synchronized
(linear) and magnitude-unsynchronized (nonlinear), and different sources of noise. We have shown that the proposed method
is as accurate as standard methods, but computationally more
efficient in real-time operation, since it is a non-iterative
method. The computational cost reduction is achieved by
splitting the problem into an offline and an online problem.
The iterative offline method estimates the prior solution, taking
most of the computational cost, while the online problem
updates this solution in a single step.
Future work could include adding different sources of distributed generation and investigating how they affect the
solutions; developing methodologies to compute the minimum
and optimal allocation of measurement units to guarantee a
given accuracy level; and adding dynamic state equations to
exploit the historical information provided by the sensors.
A PPENDIX
A PMU measure of a complex number u = |u| ejθu with
magnitude and angle noise can be expressed as:
ũ = (|u| + ωmag )ej(θu +ωθ )
= (|u| + ωmag )(cos(θu + ωθ ) + j sin(θu + ωθ ))
R EFERENCES
[1] A. Abur and A. G. Exposito, Power System State Estimation: Theory
and Implementation. CRC Press, 2004.
[2] Y.-F. Huang, S. Werner, J. Huang, N. Kashyap, and V. Gupta, “State
estimation in electric power grids: Meeting new challenges presented by
the requirements of the future grid,” IEEE Signal Processing Magazine,
vol. 29, no. 5, pp. 33–43, 2012.
[3] A. Monticelli, “Electric power system state estimation,” Proceedings of
the IEEE, vol. 88, no. 2, pp. 262–282, 2000.
[4] B. Hayes and M. Prodanovic, “State estimation techniques for electric
power distribution systems,” in 2014 IEEE European Modelling Symposium (EMS), 2014, pp. 303–308.
[5] B. Stott and O. Alsaç, “Fast decoupled load flow,” IEEE Transactions
on Power Apparatus and Systems, no. 3, pp. 859–869, 1974.
[6] M. E. Baran and A. W. Kelley, “State estimation for real-time monitoring
of distribution systems,” IEEE Transactions on Power Systems, vol. 9,
no. 3, pp. 1601–1609, 1994.
[7] M. B. A. Kelley, “A branch current based state estimation method for
distribution systems,” IEEE Transactions on Power Systems, vol. 10, pp.
483–491, 1995.
[8] W.-M. Lin, J.-H. Teng, and S.-J. Chen, “A highly efficient algorithm in
treating current measurements for the branch-current-based distribution
state estimation,” IEEE Transactions on Power Delivery, vol. 16, no. 3,
pp. 433–439, 2001.
[9] L. Schenato, G. Barchi, D. Macii, R. Arghandeh, K. Poolla, and
A. Von Meier, “Bayesian linear state estimation using smart meters and
pmus measurements in distribution grids,” in 2014 IEEE International
Conference on Smart Grid Communications, 2014, pp. 572–577.
[10] M. Zhou, V. A. Centeno, J. S. Thorp, and A. G. Phadke, “An alternative
for including phasor measurements in state estimators,” IEEE Transactions on Power Systems, vol. 21, no. 4, pp. 1930–1937, 2006.
[11] Y. Hu, A. Kuh, T. Yang, and A. Kavcic, “A belief propagation based
power distribution system state estimator,” IEEE Computational Intelligence Magazine, vol. 6, no. 3, pp. 36–46, 2011.
[12] S. Bolognani, R. Carli, and M. Todescato, “State estimation in power
distribution networks with poorly synchronized measurements,” in 2014
IEEE 53rd Annual Conference on Decision and Control. IEEE, 2014,
pp. 2579–2584.
[13] G. N. Korres, N. D. Hatziargyriou, and P. J. Katsikas, “State estimation in multi-microgrids,” European Transactions on Electrical Power,
vol. 21, no. 2, pp. 1178–1199, 2011.
[14] K. Martin, D. Hamai, M. Adamiak, S. Anderson, M. Begovic, G. Benmouyal, G. Brunello, J. Burger, J. Cai, B. Dickerson et al., “Exploring
the IEEE standard C37. 118–2005 synchrophasors for power systems,”
IEEE Transactions on Power Delivery, vol. 23, no. 4, pp. 1805–1811,
2008.
[15] G. N. Korres, “A robust algorithm for power system state estimation
with equality constraints,” IEEE Transactions on Power Systems, vol. 25,
no. 3, pp. 1531–1541, 2010.
[16] S. Bolognani and S. Zampieri, “On the existence and linear approximation of the power flow solution in power distribution networks,” IEEE
Transactions on Power Systems, vol. 31, no. 1, pp. 163–172, 2016.
[17] M. A. Woodbury, “Inverting modified matrices,” Statistical Research
Group Memorandum Reports. Princeton, NJ: Princeton University,
no. 42, 1950.
[18] W. Kersting, “Radial distribution test feeders,” IEEE Transactions on
Power Systems, vol. 6, no. 3, pp. 975–985, 1991.
[19] DistributionTestFeeders, “IEEE PES Distribution Systems Analysis
Subcommittee’s Radial Test Feeders,” 1991, accessed: 2016-10-20.
[Online]. Available: ewh.ieee.org/soc/pes/dsacom/testfeeders.html
| 3cs.SY
|
1
Data-Driven Inversion-Based Control:
closed-loop stability analysis for MIMO systems
arXiv:1801.08993v1 [cs.SY] 26 Jan 2018
Carlo Novara and Simone Formentin
Abstract—Data-Driven Inversion-Based Control (D2 -IBC) is
a recently introduced control design method for uncertain nonlinear systems, relying on a two degree-of-freedom architecture,
with a nonlinear controller and a linear controller running
in parallel. In this paper, extending to the MIMO case a
previous result holding for the SISO case, we derive a finite-gain
stability sufficient condition for a closed-loop system formed by
a nonlinear MIMO plant, connected in feedback with a D2 -IBC
controller.
I. I NTRODUCTION
Data-Driven Inversion-Based Control (D2 -IBC) is a recently introduced control design method for uncertain nonlinear systems, relying on a two degree-of-freedom architecture,
with a nonlinear controller and a linear controller running
in parallel, [1], [2]. Despite many different approaches for
joint design of identification and control have been already
proposed (see [3] and the reference therein for a comprehensive overview), as far as we are aware, D2 -IBC is the first
“identification for control" method for nonlinear dynamical
systems, where also stability guarantees are provided and
enforced directly in the identification algorithm. However,
the underlying stability sufficient condition is valid for Single
Input Single Output (SISO) systems only. In this paper, we
extend this sufficient condition to Multiple Input Multiple
Output (MIMO) systems.
Notation. A column vector x ∈ Rnx ×1 is denoted as
x = (x1 , . . . , xnx ). A row vector x ∈ R1×nx is denoted
⊤
as x = [x1 , . . . , xnx ] = (x1 , . . . , xnx ) , where ⊤ indicates
the transpose.
A discrete-time signal (i.e. a sequence of vectors) is denoted
with the bold style: x = (x1 , x2 , . . .), where xt ∈ Rnx ×1
and t = 1, 2, . . . indicates the discrete time; xi,t is the ith
component of the signal x at time t.
A regressor, i.e. a vector that, at time t, contains n present
and past values of a variable, is indicated with the bold style
and the time index: xt = (xt , . . . , xt−n+1 ).
The ℓp norms of a vector x = (x1 , . . . , xnx ) are defined as
(P
1
nx
|xi |p ) p , p < ∞,
. ( i=1
kxkp =
maxi |xi | ,
p = ∞.
C. Novara is with Politecnico di Torino (Italy), S. Formentin is
with Politecnico di Milano (Italy). Email: [email protected],
[email protected].
The ℓp norms of a signal x = (x1 , x2 , . . .) are defined as
(P
1
∞ Pnx
|xi,t |p ) p , p < ∞,
. ( t=1 i=1
kxkp =
maxi,t |xi,t | ,
p = ∞,
where xi,t is the ith component of the signal x at time t.
II. T HE D2 -IBC APPROACH
Consider a nonlinear discrete-time MIMO system in regression form:
yt+1 = g (y t , ut , ξt )
(1)
y t = (yt , . . . , yt−n+1 )
ut = (ut , . . . , ut−n+1 )
ξ t = (ξt , . . . , ξt−n+1 )
where ut ∈ U = [−ū, ū]nu ⊂ Rnu is the saturated input,
.
¯ ξ¯ nξ ⊂ Rnξ is a
yt ∈ Rny is the output, ξt ∈ Ξ = −ξ,
disturbance, and n is the system order. Both U and Ξ are
compact sets.
Suppose that the system (1) is unknown, but a set of
measurements is available:
.
D = {ũt , ỹt }0t=1−L
(2)
.
where ũt ∈ U , ỹt ∈ Y , Y = [−ȳ, ȳ]ny , ȳ = maxt,i |ỹt,i | <
∞, and ỹt,i is the ith component of ỹt . The tilde is used
to indicate the input and output samples of the data set (2),
which is supposed to be available at time t = 0 when the
controller needs to be designed. The input signals employed
to generate (2) are also assumed to be such that the system
output does not diverge.
.
Let R = [−r̄, r̄]ny , with 0 ≤ r̄ ≤ ȳ, be a domain
of interest for the trajectories of the system (1). The aim
is to control the system (1) in such a way that, starting
.
from any initial condition y 0 ∈ Y 0 = Rn×ny ⊂ Rn×ny ,
the system output sequence y = (y1 , y2 , . . .) tracks any
reference sequence r = (r1 , r2 , . . .) ∈ R ⊆ R∞ ⊂ ℓ∞ .
The set of all possible disturbance sequences is defined as
.
Ξ = {ξ = (ξ1 , ξ2 , . . .) : ξt ∈ Ξ, ∀t}.
To accomplish this task, we consider a feedback control
structure with two controllers, K nl and K lin , working in
parallel:K nl is a nonlinear controller used to guide the
system (1) along the trajectories of interest, while K lin is
a linear controller aimed to enhance the tracking precision.
2
A. Nonlinear controller design
The first step needed to design the nonlinear controller is
to identify from the data (2) a model for the system (1) of
the form
ŷt+1 = f (y t , ut ) ≡ f (q t , ut )
(3)
q t = (yt , . . . , yt−n+1 , ut−1 , . . . , ut−n+1 )
where ut and yt are the system input and output, and ŷt
is the model output. A parametric structure is taken for the
function f :
f (q t , ut ) =
N
X
αi φi (q t , ut )
(4)
i=1
where φi : Rn(ny +nu ) → R are polynomial basis functions
and αi ∈ Rny ×1 are parameter vectors that can be identified
by means of convex optimization.
Once a model of the form (3) has been identified, the
nl
current command action unl
t of the nonlinear controller K
is computed by the on-line inversion of the obtained models,
given the available regressor q t . This inversion is performed
by solving the following optimization problem:
unl
t = arg min Jt (ū)
ū∈U
(5)
where the current objective function Jt is
ny
2
X
ζi
µi 2
λi 2
Jt (ū) =
rt+1,i − fi (q t , ū) +
ū +
δū
ρ
ρui
ρui
i=1 yi
(6)
where rt+1,i is the ith component of rt+1 and fi is the
.
. ||y˜i 1−L ,...,y˜i 0 ||22
and ρui =
ith component of f ; ρyi =
L
||u˜i1−L ,...,u˜i0 ||22
L
are normalization constants computed form
the data set (2), δū = ū − ut−1 are the control action rate of
change with respect to the past input, µi ≥ 0 and λi ≥ 0 are
design parameters which allow us to set the trade off between
tracking precision and control action aggressiveness for every
input (control action magnitude and speed, respectively), and
1 ≥ ζi ≥ 0 are other design parameters which are needed to
set the priority of tracking for all the different output. The
controller K nl is fully defined by the law (5).
B. Linear controller design
The linear controller K lin is defined by the centralized
extended PID (Proportional Integrative Derivative) control
law
nθ
X
lin
Bi et−i
(7)
+
ulin
=
u
t
t−1
i=0
where et = rt − yt is the tracking error, nθ is the controller
order and Bi ∈ Rnu ×ny are the controller parameter matrices. All the entries of these matrices are contained in a
vector θ = (θ1 , . . . , θny ×nu ×nθ ). Note that, for nθ = 1 and
nθ = 2, the standard PI and PID controller are selected,
respectively. The goal of K lin in the proposed architecture
is to compensate for model-inversion errors and boost the
control performance by assigning a desired dynamics to the
resulting nonlinearly-compensated system.
In the considered setting, where most of the information
about the system is that inferrable from data, finding a good
control-oriented model of the error system, i.e. the system
describing the relationship between ulin
and yt , is not an
t
easy task. Therefore, in this paper, the Virtual Reference
Feedback Tuning (VRFT) method, originally developed in
[4] and extended to the MIMO case in [5], is employed and
adapted to the present setting. Its rationale is briefly recalled
next for self-consistency of the paper.
First, define the desired behavior for the closed loop
system by means of a discrete transfer function M . In the
present setting, this function is used to assign a specific
desired dynamic to the nonlinearly-compensated resulting
system. Typically, for ny = nu = m, M is a diagonal m×m
transfer function composed by m asymptotically stable lowpass filters, synthesizing the desired closed loop behavior
for each output. The virtual reference rationale permits the
design of K lin without identifying any model of the system,
based on the following observation: in a virtual operating
condition, where the closed-loop system behaves exactly as
M , the virtual reference signal rtv would be given by the
filtering of the output yt by the model M −1 . Since the
inverse model M −1 is non-causal, the filtering task must
be accomplished off-line using a set of available data. The
optimal controller is the one giving the measured ulin
as
t
output when fed by the virtual error evt = rtv − yt , as to
minimize the cost function
X
lin
JV R =
||ulin
(θ)evt ||22 .
(8)
t −K
t
Two versions of the VRFT method in the D2 -IBC setting are
available, corresponding to two cases, related to bijectivity
of the function f in (3). The details of these two versions
are not discussed here for the sake of brevity.
III. C LOSED - LOOP STABILITY ANALYSIS
The closed-loop system formed by the plant (1), controlled
in feedback by the parallel connection of K nl and K lin , is
described by the following equations:
yt+1 = g (y t , ut , ξ t )
lin
ut = unl
t + ut
nl
unl
rt+1 , y t , unl
t =K
t−1
lin
r t − y t , ulin
ulin
t−1
t = K
(9)
where K nl and K lin are defined in (5) and (7), respectively,
and ut ∈ U , ∀t. The reference initial condition is chosen as
r0 = y0 .
In the following, we study the stability properties of such a
close-loop system. The following stability notion is adopted.
3
Definition 1: A nonlinear system (possibly time-varying),
with inputs rtand ξt , and output yt , is finite-gain ℓ∞ stable
on Y 0 , R, Ξ if finite and non-negative constants Γr , Γξ
and Λ exist such that
kyk∞ ≤ Γr krk∞ + Γξ kξk∞ + Λ
for any (y 0 , r, ξ) ∈ Y 0 × R × Ξ.
Note that this finite-gain stability definition is more general
than the standard one, which is obtained for R = ℓ∞ and
Ξ = ℓ∞ , see e.g. [6].
In order to study how to guarantee finite-gain stability of
the feedback system (9), some additional assumptions are
introduced.
Assumption 1 (Lipschitzianity): The function g in (1) and
(9) is Lipschitz continuous on Y n × U n × Ξn . Without loss
of generality, it is also assumed that Y n × U n × Ξn contains
the origin.
This assumption is mild, since most real-world dynamic
systems are described by functions that are Lipschitz continuous on a compact set.
From Assumption 1, it follows that g can be written as
g (y t , ut , ξ t ) = g o (y t , ut ) + gtξ ξ t
.
where g o (y t , ut ) = g (y t , ut , 0) and gtξ ∈ Rny ×n is a timevarying parameter (dependent on y t , ut and ξ t ) bounded
on Y n × U n × Ξn as gtξ
≤ γξ , for some γξ < ∞.
∞
Assumption 1, together with (4), implies that the residue
function
.
∆ (y t , ut ) = g o (y t , ut ) − f (y t , ut )
is Lipschitz continuous on Y n × U n . Hence, a finite and
non-negative constant γy exists, such that
k∆ (y, u) − ∆ (y ′ , u)k∞ ≤ γy ky − y ′ k∞
for all y, y ′ ∈ Y n and all u ∈ U n .
Assumption 2 (Model accuracy): The following inequality
holds: γy < 1.
The meaning of this assumption is clear: it requires f to
accurately describe the variability of g with respect to y t .
Now, consider that
.
êt+1 = rt+1 − ŷt+1 = rt+1 − f (y t , ut )
lin
≡ rt+1 − f y t , ut rt+1 , rt , y t , unl
t−1 , ut−1
.
= F (y t , rt , v t ) ,
.
lin
where v t = (rt+1 , unl
t−1 , ut−1 ) ∈ V and V is a compact set.
Then, for any (y t , r t , v t ) ∈ Y n × Rn × V n ,
|êt+1 | ≤ Γy ky t k∞ + Γs kr t k∞ + Λe
(10)
where Γy , Γs , Λe < ∞. This inequality directly follows from
the fact that the model function f is Lipschitz continuous on
Y n × U n . Note that (10) does not imply that y ∈ Y ∞ .
Assumption 3 (Effective model inversion): The following
inequality holds: Γy ≤ 1 − γy .
This assumption is not restrictive: it is certainly satisfied if
µ = 0 and the reference r = (r1 , r2 , . . .) is a model solution
(i.e. rt+1 is in the range of f (y t , ·) for all t). Indeed, in
this case, ŷt+1 = rt+1 , ∀t, since K nl performs an exact
inversion of the model, see again (5) (K lin gives a null input
signal in this case). This implies that Γy = 0, Γs = 0 and
Λe = 0. Hence, if a sufficiently small µ is chosen and the
reference is sufficiently close to a system solution, supposing
that inequality (10) holds with a sufficiently small Γy is
reasonable.
To formulate our last assumption, define
1
(11)
λr r̄ + γξ ξ¯ + Λg
1 − λy
.
.
.
where λy = Γy + γy < 1, λr = λy + Γs and Λg = Λe +
maxu∈U n k∆ (0, u)k∞ . Note that ē is bounded, being the
sum of bounded quantities. In particular, for null disturbances
(ξ¯ = 0), exact modeling (f = g, ∆ = 0, γy = 0) and
reference signals properly chosen (Γy = 0, Γs = 0, Λe = 0),
we have ē = 0. In realistic situations, with reasonable
disturbances, sufficiently accurate modeling and reference
signals properly chosen, ē can be reasonably small (that is,
ē ≪ r̄).
.
ē =
Assumption 4 (Output domain exploration): The following
inequality holds: ȳ ≥ r̄ + ē.
This assumption requires that the set Y explored by the
output data is somewhat larger than the set R where the
trajectory of interest are defined. Note that it can always be
met just collecting data that sufficiently enlarge the set Y .
Closed-loop stability of the system (9) is stated by the
following result, which also provides a bound on the tracking
error.
Theorem 1: Consider the system (9) and let Assumptions
1-4 hold. Then:
(i) The feedback system (9), having inputs rt and ξt and
output yt , is finite-gain ℓ∞ stable on Y 0 , R, Ξ .
.
(ii) The tracking error signal e = r − y is bounded as
kek∞ ≤ ē.
(12)
Proof. The proof of the theorem is structured as follows.
Firstly, the tracking error is proven to be upper bounded by
a suitable combination of the norms of the output and the
reference. Secondly, it is shown that, under the assumption
of an effective model inversion, such a bound is equivalent to
a bound on the tracking error, whatever the output is (claim
(ii)). Claim (i) is derived as a straightforward consequence
of claim (ii).
To start with, consider that
.
et+1 = rt+1 − yt+1 = êt+1 − δyt
4
Then, yt ∈ Y , ∀t ≥ 0 and (12) holds (claim (ii)). Claim (i)
is a direct consequence of claim (ii) and the relation y =
r − e.
where
êt+1 = rt+1 − ŷt+1 = F (y t , r t , v t )
δyt = ∆ (y t , ut ) + gtξ ξ t .
The term êt+1 is bounded according to (10). Note that ky t k∞
could be unbounded. In order to derive a bound on δyt , we
can use Assumption 2 and observe that, for any y t ∈ Y n ,
k∆ (y t , ut )k∞ − k∆ (0, ut )k∞
≤ k∆ (y t , ut ) − ∆ (0, ut )k∞ ≤ γy ky t k∞ .
The following inequality thus holds for any y t ∈ Y n :
kδyt k∞ ≤ k∆ (y t , ut )k∞ + γξ kξ t k∞
¯
≤ γy ky t k + γξ kξt k + ∆,
∞
∞
.
¯ =
where ∆
maxu∈U n k∆ (0, u)k∞ < ∞. Hence,
ket+1 k∞ ≤ λy ky t k∞ + Γs kr t k∞ + Λe
¯
+γξ kξ k + ∆,
(13)
t ∞
which proves that the tracking error e is bounded by a
suitable combination of the norms of y t and r t , for any
y t ∈ Y n . Note that (13) in this form is of no use, since
ky t k∞ could be unbounded and thus the condition y t ∈ Y n
may not hold. However, (13) can be rewritten as
ket+1 k∞ ≤ λy ket k∞ + λy krt k∞ + Γs krt k∞
+γξ kξ t k∞ + Λg ,
(14)
that means,
ket+1 k∞ ≤ λy ket k∞ + w,
(15)
.
where w = λr r̄ + γξ ξ̄ + Λg . Inequality (15), again, holds
only if y t ∈ Y n .
Consider now that, by assumption, y 0 ∈ Rn ⊆ Y n . This
implies that, at time t = 0, inequality (15) holds. Being
e0 = r 0 − y 0 = 0 for the selected initialization of r0 , we
have
ke1 k∞ ≤ λy ke0 k∞ + w = w ≤ ē.
Since ke1 k∞ ≤ ē and |r1 | ≤ r̄, it follows from Assumption
4 that y1 ∈ Y . Consequently, the Lipschitzianity assumption
holds and (15) can be used also for t = 1, giving
ke2 k∞ ≤ λy ke1 k∞ + w ≤ λy w + w
P∞
w
≤ w k=0 λky ≤ 1−λ
= ē
y
where the geometric series sum has been obtained thanks
to the fact that, by Assumption 3, λy < 1. It follows that
y2 ∈ Y and, consequently, (15) can be used also for t = 2.
Iterating the above reasoning,
ke3 k∞ ≤ λy max {ke2 k∞ , ke1 k∞ } + w
≤ λy ke2 k∞ + w ≤ λ2y w + λy w + w ≤ ē
..
.
Pt
ket+1 k∞ ≤ w k=0 λky
P
w
k
≤w ∞
k=0 λy ≤ 1−λy = ē.
R EFERENCES
[1] S. Formentin, C. Novara, S. Savaresi, and M. Milanese, “Active braking
control system design: The D2-IBC approach,” IEEE/ASME Transactions on Mechatronics, vol. 20, no. 4, pp. 1573–1584, 2015.
[2] C. Novara, S. Formentin, S. Savaresi, and M. Milanese, “Data-driven
design of two degree-of-freedom nonlinear controllers: the D2-IBC
approach,” Automatica, vol. 72, pp. 19–27, 2016.
[3] H. Hjalmarsson, “From experiment design to closed-loop control,”
Automatica, vol. 41, no. 3, pp. 393–438, 2005.
[4] M. Campi, A. Lecchini, and S. Savaresi, “Virtual reference feedback
tuning: a direct method for the design of feedback controllers,” Automatica, vol. 38, no. 8, pp. 1337–1346, 2002.
[5] S. Formentin, S. Savaresi, and L. Del Re, “Non-iterative direct datadriven controller tuning for multivariable systems: theory and application,” IET Control Theory & Applications, vol. 6, no. 9, pp. 1250–1257,
2012.
[6] H. Khalil, Nonlinear Systems. Prentice Hall, 1996.
| 3cs.SY
|
A Parallel Simulator for Massive Reservoir Models Utilizing
Distributed-Memory Parallel Systems
Hui Liu 1∗, Lihua Shen1 , Yan Chen1,2∗ , Kun Wang1 , Bo Yang1 , Zhangxin Chen1∗
arXiv:1701.06254v1 [cs.CE] 23 Jan 2017
1
Department of Chemical and Petroleum Engineering
University of Calgary, 2500 University Drive NW, Calgary, AB, Canada
2 College of Mathematics and informatics
South China Agricultural University, Guangzhou 510642, China
Abstract
This paper presents our work on developing parallel computational methods for two-phase flow on
modern parallel computers, where techniques for linear solvers and nonlinear methods are studied and
the standard and inexact Newton methods are investigated. A multi-stage preconditioner for two-phase
flow is applied and advanced matrix processing strategies are studied. A local reordering method is
developed to speed the solution of linear systems. Numerical experiments show that these computational methods are effective and scalable, and are capable of computing large-scale reservoir simulation
problems using thousands of CPU cores on parallel computers. The nonlinear techniques, preconditioner and matrix processing strategies can also be applied to three-phase black oil, compositional and
thermal models.
Keywords: Multi-phase flow, oil-water, porous media, simulation, parallel computer
INTRODUCTION
Large-scale geological models in petroleum reservoirs and their field-scale simulation models with millions of grid blocks (cells, elements) are being applied to represent complex geological heterogeneity and
to capture their high-resolution phenomena, which lead to extremely long simulation run time because of
difficulties from solutions of linear and nonlinear systems with a huge number of grid blocks. Nowadays,
most of available commercial reservoir simulators have been developed for personal computers and workstations, which are limited by their memory size, memory speed and CPU performance. When simulating
large-scale reservoir models, it may take days or even longer to complete a simulation run. Parallel reservoir simulators and efficient numerical methods for parallel computers are essential to improve simulation
runs and simulators’ capability.
∗ Authors
to whom correspondence may be addressed.
[email protected]
1
Email addresses:
[email protected], [email protected],
Reservoir simulations have been studied for decades by researchers, and various techniques have been
developed to model new recovery processes and to accelerate computer simulations, including new reservoir models and numerical methods [27,28] , nonlinear methods [23,28] , linear solvers and preconditioner techniques [17,11,12,13,14,15] , and parallel computing [1,2] . Shiralkar and his collaborators [4] developed a parallel
simulator, FALCON, using FORTRAN 90 and high performance FORTRAN (HPF), which could run on
a variety of computing platforms, such as Cray T3D and T3E, SGI Origin 2000, Thinking Machines CM5
and IBM RS6000. Killough and Bhogeswara applied local refinement methods to reservoir simulations
and designed effective preconditioners for parallel linear solvers in their parallel simulator [7] . The parallel
simulator developed by Parashar et al. successfully handled multiple fault blocks with multiple physics [5] . Rutledge and his collaborators implemented a parallel compositional simulator for SIMD computers,
which solved pressure unknowns implicitly and saturation unknowns explicitly [3] . Kaarstad et al. presented a two-dimensional two-phase (oil and water) parallel reservoir simulator, which could solve large
problems with up to one million grid blocks [6] . Dogru et al. developed the parallel black oil simulator,
which could calculate models with a billion grid blocks [8,9] . Zhang et al. developed a parallel platform
for adaptive finite element methods and finite volume methods [35,25] , which was also used to develop a
parallel black oil simulator using discontinuous Galerkin methods. Linear systems from reservoir simulations are usually ill-conditioned, especially when heterogeneous geology is applied. The Krylov subspace
linear solvers and multi-grid solvers have been widely adopted in reservoir simulations. Advanced preconditioner techniques have also been proposed and applied to reservoir simulations, including point-wise
and block-wise incomplete factorization [11,12] , domain decomposition [26] , constrained pressure residual
(CPR) [13,14,22] , multi-stage [20] and FASP methods [21] .
This paper deals with developing fast numerical simulations for two-phase flow on distributed memory
parallel computers. A finite difference method is employed to compute a two-phase model due to its simplicity. Due to the high nonlinearity of this model, the standard Newton (or Newton-Raphson) method and
the inexact Newton method [23] are both investigated. The standard Newton method tends to solve linear
systems as accurately as possible, in which the solutions may be expensive to obtain when their condition
numbers are large. In contrast, the inexact Newton method just solves linear systems approximately and
the stop criterion for each linear system changes depending on its convergence history. For the linear
systems derived from the (inexact) Newton methods, they are usually ill-conditioned, especially when the
geological models are highly heterogeneous and the sizes of grid blocks of these models are large, which
are difficult to solve. An efficient multi-stage preconditioner [22] is applied to these linear systems, which
is based on the classical constrained pressure residual method (CPR). It solves the pressure equation using the AMG (algebraic multigrid) method to minimize errors of the solutions to linear systems and the
entire linear system using the restricted additive Schwarz (RAS) method. The AMG method and the RAS
method have excellent scalability so this preconditioner has excellent scalability. In addition, advanced
matrix processing techniques are also studied. Two decoupling strategies are applied to linear systems,
which are the Quasi-IMPES (implicit pressure and explicit saturation) strategy and the alternative block
factorization (ABF) strategy. A local potential reordering technique is also developed, which reorders
the unknowns by the potential of grid blocks. Although the techniques are applied for a two-phase flow
model, they are also efficient for three-phase black oil and compositional models. The nonlinear methods, the domain decomposition preconditioner, and matrix processing strategies can even be applied to a
thermal model. Numerical experiments show that these techniques are efficient and scalable for parallel
computing.
2
TWO-PHASE OIL-WATER FLOW MODEL
The two-phase model assumes that the simultaneous flow has two phases (oil and water) and two components (oil and water), and they are immiscible. It can also describe a case when solution gas stays in the
oil phase during the entire simulation period. The model also assumes temperature never changes.
Darcy’s law describes the flow of fluids through a porous media, which establishes a relationship
between the volumetric flow rate and the pressure gradient,
Q=−
KA∆p
,
µL
(1)
where K is the permeability of a given reservoir, A is the area in the flow direction, ∆p is the pressure
difference, µ is the viscosity of the fluid, and L is the length of the reservoir. In three-dimensional space,
its differential form law is
Q
K
q = = − ∇p.
(2)
A
µ
By combining Darcy’s law, mass conservation law and gravitational force, the two-phase model [27] is
described as the following equation,
∂
KKro
ρo ∇Φo ) + qo
(ϕso ρo ) =∇ · (
∂t
µo
(3)
KKrw
∂
(ϕsw ρw ) =∇ · (
ρw ∇Φw ) + qw .
∂t
µw
A no-flow condition is adopted as the boundary condition.
ϕ is the porosity of the reservoir, which is computed as,
ϕ = ϕr (1 + cr (po − pr )),
(4)
where cr is the rock compressibility factor and pr is the reference pressure of rock.
Let α (α = o, w) be a phase. Φα is phase potential (total absolute pressure), which is defined as,
Φα = pα + ρα gz,
(5)
where pα is phase pressure, ρα is phase density, g is the standard gravity, and z is the reservoir depth. The
density is a function of pressure, which is written as,
ρα = ρrα (1 + cα,r (pα − pα,r )).
(6)
Here, ρrα is the reference density at reference pressure pα,r , and cα,r is the reference compressibility factor.
µα is the viscosity of a phase, which is calculated as
µα = µα,r + cµ,α (pα − pα,r ),
(7)
where µα,r is the reference viscosity at reference pressure pα,r and cµ,α is the factor for pressure dependence of the phase viscosity.
In this model, oil and water occupy the reservoir. Their volumes are described by oil saturation, so ,
and water saturation, sw , which satisfy,
so + sw = 1.
(8)
3
The oil phase pressure, pw , and the water phase pressure, po , may be different due to capillary pressure
pc . They are related as,
pw = po − pc (sw ).
(9)
The capillary pressure pc is a function of water saturation and is user input. Krα is the relative permeability.
Krw and Kro are functions of water saturation, sw . K is the permeability, which has three components in x,
y and z directions, noted as Kx , Ky and Kz .
The Peaceman method is adopted for well modeling. A well may have many perforations. For each
perforation at a grid block m, its well rate qα,m is calculated by the following formula,
qα,m = Wi
ρα Krα
(pb − pα − ρα g(zb − z)),
µα
(10)
where pb is the bottom hole pressure defined at the reference depth zb , Wi is the well index, z is the depth
of the perforation in grid block m, and pα is phase pressure in grid block m. Wi , which is the well index,
defines the relationship among the well bottom hole pressure, flow rate, and grid block pressure [27] .
Various operation constraints may be applied to a well at different time stage, such as a fixed bottom
hole pressure, a fixed oil rate, a fixed water rate or a fixed liquid rate. When the fixed bottom hole pressure
condition is applied to some well, the constraint equation is,
pb = c,
(11)
where c is a constant and known. The fixed water rate condition is the following equation:
∑ qw,m = qw,
m
(12)
where qw is constant and known. For the fixed oil rate condition, its constraint equation is
∑ qo,m = qo,
m
(13)
where qo is fixed. For the fixed liquid rate condition, the constraint equation is
∑(qo,m + qw,m) = qo + qw.
m
(14)
NUMERICAL METHODS
To solve equation (3), the fully implicit method (FIM) is applied, and the oil phase pressure p, water saturation sw and the well bottom hole pressure pb are chosen as the unknowns. The time differentiation term
is discretized by the backward Euler difference scheme and the space differentiation terms are discretized
by the cell-centered finite difference method [27] .
Let un be the value of the function u at time step n, then its derivative at time step (n + 1) can be
approximated by backward difference,
(
∂u n+1 un+1 − un
)
=
,
∂t
∆t
4
(15)
where ∆t is time step. If we assume that d(d = x, y, z) is any space direction and A is the area of the faces
along d direction in a grid block, then the transmissibility term Tα,d of phase α is defined as
Tα,d =
KKrα
A
ρα ,
µα
∆d
where ∆d is the cell length along d direction. Then the nonlinear system is written as,
V
n+1
[(ϕρo so )n+1 − (ϕρo so )n ] = ∇ · (Ton+1 ∇Φn+1
o ) + qo
∆t
V [(ϕρw sw )n+1 − (ϕρw sw )n ] = ∇ · (T n+1 ∇Φn+1 ) + qn+1
w
w
w
∆t
(16)
(17)
where V is the volume of a specific grid block.
The transmissibility term is defined on each face of a cell. Therefore, a cell has six items defined. For
any face shared by two cells, the value of the transmissibility term is the same in each cell. In this case,
the mass of any component is conservative. In this paper, the upstream technique is employed to compute
the transmissibility term. For a grid block (i, j, k), the transmissibility term in x direction, (Tα,x )i± 1 , j,k , is
2
defined as
{
(Tα,x )i±1, j,k , if Φi±1, j,k ≥ Φi, j,k
(Tα,x )i± 1 , j,k =
,
(18)
2
(Tα,x )i, j,k ,
if Φi±1, j,k < Φi, j,k
whose value is from the cell with larger phase potential. Transmissibility terms in y and z directions are
defined the same way.
We assume a grid has n blocks (cells, elements) and τ wells exist in the model. The final nonlinear
system is represented by
−
F(→
x ) = 0,
(19)
where
and
→
−p
→
→
−
sw
x = −
→
−
pb
p1
→
−p = p2
···
pn
sw1
sw2
→
−
sw =
···
swn
,
(21)
(22)
,
,
pb1
pb2
→
−
pb =
··· .
pbτ
5
(20)
(23)
F is a nonlinear mapping from RN to RN (N = 2 × n + τ). The nonlinear system is solved by the
Newton method (or inexact Newton method). The linear systems are also an order of N. Here we mention
that the properties related to saturation are strongly nonlinear while the properties related to pressure are
weakly nonlinear.
Nonlinear Methods
In the reservoir simulations, a set of nonlinear equations are required to be solved at each time step, and
the objective is to find an x∗ ∈ RN such that ∥F(x∗ )∥ < ε, where ε is the stopping criterion (tolerance) for
nonlinear methods. Traditionally, the standard Newton method is applied to the nonlinear system (19),
and the stopping tolerance, ε, is small, such as 10−5 .
It is well-known that linear systems from reservoir simulations are hard to solve, and the solution time
may occupy more than 70% of the total simulation time [27] . To obtain enough accuracy, the stopping
tolerance for the standard Newton methods is set to be small, which can make the situation even worse. In
practice, we know that in the first few Newton iterations, the solution xl may be far from the real solution
x∗ . Therefore, it may not be necessary to solve the linear systems accurately [23] .
The inexact Newton method is an extension of the standard Newton method, which satisfies the following relationship during the solution procedure:
∥Ay − b∥ ≤ ηl ∥b∥,
(24)
where A is the Jacobian matrix and y is only an approximate solution depending on the parameter ηl . The
algorithm for the inexact Newton method is described in Algorithm 1 [23] .
Algorithm 1 The Inexact Newton Method
1: Given an initial guess x0 and stopping criterion ε, let l = 0 and assemble the right-hand side b.
2: while ∥b∥ ≥ ε do
3:
Assemble the Jacobian matrix A.
4:
Determine ηl .
5:
Find y such that
∥Ay − b∥ ≤ ηl ∥b∥ ,
(25)
6:
Let l = l + 1 and xl = xl−1 + y.
7: end while
8: x∗ = xl is the approximate solution of the nonlinear system, F(x) = 0.
ηl is a forcing term, which forces the residual of the linear system Ay = b to be small. The choice of the
forcing term is designed to achieve fast local convergence and to avoid over-solving the linear system [23] ,
three of which are listed as follows:
6
b(xl ) − rl−1
,
b(l−1 )
b(xl ) − rl−1
,
ηl =
l−1 )
b(x
(
)β
l)
b(x
,
γ
b(xl−1 )
(26)
where rl is the residual of the l-th iteration,
A typical value for β is
ηk ∈ [0.01, 0.1].
√
5+1
2 .
rl = b − Ay.
(27)
The second one is employed in this paper. Besides, ηk is forced to satisfy:
Solution of Linear Systems
−
−p is placed before the saturation unknown →
−
Since the pressure unknown →
sw and bottom hole pressure →
pb ,
the matrix A has the following structure:
A pp A ps A pσ
A = Asp Ass Asσ ,
(28)
Aσp Aσs Aσσ
where the sub-matrices A pp , Ass , and Aσσ are the matrices corresponding to the pressure unknowns, the
saturation unknowns, and the bottom hole pressure unknowns respectively. Other sub-matrices are coupled
items.
In this paper, the matrix A is decoupled first to weaken the relationship between the pressure and
saturation unknowns, which is described as,
D−1 Ax = D−1 b,
(29)
where D is the decoupling operator (matrix). Then a potential reordering technique is applied. In the end,
a multi-stage preconditioner are employed to work with Krylov solvers.
Matrix Decoupling Strategy
In order to weaken the strong coupling between the pressure unknowns and the saturation unknowns, a
decoupling operator is necessary to the Jacobian matrix A. The decoupling operator should be computationally cheap and effective. For the two-phase model, the Quasi-IMPES strategy [18] and the alternative
block factorization (ABF) strategy [15] are studied. A detailed study on these decoupling strategies can be
found in [20] .
The ABF method is written as
Diag(A pp ) Diag(A ps ) 0
DABF = Diag(Asp ) Diag(Ass ) 0 ,
0
0
Iσ
7
where Diag(B) stands for the diagonal of matrix B. The ABF method applies to the reservoir blocks only
and the well blocks keep unchanged. The Quasi-IMPES method is defined as follows:
In Diag(A ps )Diag(Ass )−1 0
DQI = 0
In
0 .
0
0
Iσ
The Quasi-IMPES decoupling strategy is simpler and less aggressive than the ABF strategy. It preserves
the saturation and well blocks. Asp contributes to the pressure blocks.
The new linear system is equivalent to the original linear system Ax = b. Therefore, it is also denoted
as Ax = b. For parallel computing, the matrix-matrix multiplication operation D−1 A and the matrix-vector
operation D−1 b are local, and there is no communication involved. We can see that the scalability of the
matrix decoupling strategies (the Quasi-IMPES strategy and the ABF strategy) is ideal.
Local Potential Reordering Technique
For the development of linear solvers, it is well-known that proper matrix reordering techniques improve
the efficiency of linear solvers and preconditioners. Some of the techniques reduce the bandwidth of a
given matrix, such as the RCM (reverse Cuthill-MaKee) reordering method. Some of the techniques reduce fill-ins of the ILU methods, such as the MD (minimum degree) reordering method. Proper reordering
techniques are important to the efficiency of linear solvers and preconditioners.
For reservoir simulation, upstream schemes are widely applied, which are determined by the potential
of a phase, such as the oil phase. Kwok and Tchelepi developed a potential-based reordering strategy
and a reduced Newton method was proposed based on this reordering technique [28] . In this paper, this
strategy is applied to reorder the pressure unknowns and the saturation unknowns. The oil phase potential,
Φo = po + ρo gz, is employed for each grid block. For the ith grid block, it has a potential Φi . Then Φ1 ,
Φ2 , · · · , and Φn form a set Φ = {Φ1 , Φ2 , · · · , Φn }.
Let us sort the set Φ in a descending order based on the values of Φi (1 ≤ i ≤ n). Then, for any Φi ,
its location in the sorted set Φ is determined, say Li , which is numbered from the left side. Now, we can
define a mapping Pm such that
Pm(i) = Li .
(30)
An example is given to demonstrate the definition of the mapping Pm. Assume that we have four
grid blocks, their potentials are Φ1 = 2.1, Φ2 = 1.3, Φ3 = 4.2 and Φ4 = 1, and the sorted set Φ is
{4.2, 2.1, 1.3, 1}. Then we have Pm(1) = 2 (Φ1 ), Pm(2) = 3 (Φ2 ), Pm(3) = 1 (Φ3 ), and Pm(4) = 4
(Φ4 ). We can see that the mapping Pm defines a permutation of the set {1, 2, · · · , n}.
We know that the unknown x has the following form:
→
−p
−
→
−
sw ,
x = →
(31)
→
−
pb
whose dimension is (2 × n + τ). Here, the pressure unknowns are numbered first, followed by the water
saturation unknowns and the bottom hole pressure unknowns. Now, let us define the mapping for the
whole linear system:
1 ≤ i ≤ n,
Li ,
(32)
Pt(i) = n + Li−n , n + 1 ≤ i ≤ 2 × n,
i,
2 × n + 1 ≤ i ≤ 2 × n + τ.
8
Then the mapping Pt defines a permutation of the set {1, 2, 3, · · · , 2 × n + τ}. If we apply this mapping to
the linear system Ax = b, the matrix à of the reordered linear system still has the same structure as A:
g
f
g
A
A
A
pp
ps
pσ
e=
f
f
A
(33)
A
A
Af
sp
ss
sσ .
g
f g
A
σp Aσs Aσσ
The reason that we use the potential reordering technique is that the upstream schemes are applied and the
fluid flows from grid blocks with high potential to grid blocks with low potential. After the reordering, the
saturation block Af
ss is close to a lower triangular matrix, which can be solved effectively by ILU methods.
The ILU methods are more efficient for the reordered linear systems.
A global reordering for distributed parallel systems is extremely difficult, which involves heavy communications. Therefore, the reordering strategy is only applied to the local grid on each processor. In
this case, there is no communication during the reordering process except information exchange required
to define the mapping Pt. The communication volume is small and the strategy is scalable to parallel
computing. The reordering technique here is called the local potential reordering technique.
Multi-stage Preconditioner
The pressure equation of the two-phase model is elliptic (or parabolic) and the pressure unknowns contribute most of an error to this model. It is well-known that the algebraic multigrid methods (AMG) are
the most effective methods to positive definite matrices. Wallis et al. introduced the classical constraint
pressure residual (CPR) method to oil-water model and black oil model.
To introduce our multi-stage preconditioner, some notation is required. We have introduced the x, the
−
−p , the saturation unknown →
−
pressure unknown →
sw and the bottom hole pressure unknown →
pb . A global
restriction operator for the pressure unknown is defined as
−p .
Πr x = →
(34)
→
−p
−
−p = →
Π p→
0 .
→
−
0
(35)
A prolongation operator Π p is defined as
Π p p has the same dimension as x.
We define the notation M (B)−1 b to represent the solution y for linear system By = b from an AMG
solver. If the system By = b is solved by the restricted additive Schwarz (RAS) method, we use the notation R(M)−1 b to represent the solution y. A CPR-like preconditioner, denoted by CPR-FPF preconditioner [22] , has been designed, which is described in Algorithm 2. The method is a three-stage preconditioner.
It is suitable for two-phase oil-water model, black oil model, compositional model and extended black
oil models. We should mention that the RAS method and AMG method are scalable, so the CPR-FPF
preconditioner is also scalable.
9
Algorithm 2 The CPR-FPF Preconditioner
1: x = R(A)−1 f .
2: r = f − Ax
3: x = x + Π p (M (A pp )−1 (Πr r)).
4: r = f − Ax
5: x = x + R(A)−1 r.
NUMERICAL EXPERIMENTS
An IBM Blue Gene/Q system is applied to benchmark our reservoir models. It uses 64-bit PowerPC A2
processor, whose performance is low compared with Intel processors. However, it has a strong network
relative to compute performance, which ensures its scalability. The IBM Blue Gene/Q is an excellent
system for parallel computing. In the following sections, our application runs only one MPI processor on
each CPU core. When we say using n p CPU cores, we mean that n p MPI processors are used.
Validation
Example 1 This case is from CMG IMEX (mxsmo031.dat). The permeability varies from cell to cell, and
the porosity varies in z direction. The mesh is 10× 10×3 with mesh size 1000ft. in x and y directions and
20, 30, 50 ft. in z direction from top to bottom. The depth of the top layer center is 926 ft. From top to
bottom, the permeability in x direction is 500, 50, 200 mD, same as the permeability in y direction. The
permeability in z direction is 60, 40, 20 mD. The porosity is 0.3.
Component properties: densities of gas, oil and water are 0.0647 lbm/ft3, 46.244 lbm/ft3 and 62.419
lbm/ft3.
The initial conditions are as follows: reference pressure 1000 psi at associated depth 100 ft, depth to
water-oil contact is 2000 ft.
There are one injection well and one production well. Both are vertical. Injection well has maximum
water injection rate 1000 bbl/day, maximum bottom hole pressure 2.0e+4 psi, well index 1.0e+5 with
perforation at cell [1 1 1]. Production well has maximum oil production rate 2000 stb/day, maximum
bottom hole pressure 100 psi. Well radius 0.25 ft with perforation at cell [10 10 3]. The well schedule is
listed in Table 1.
Table 1: The schedule of example two-phase model from mxsmo031.
time(days) injector(stw) producer(sto)
0
1000
shutin
1000
10000
10000
1200
shutin
unchanged
2000
unchanged
20000
8000
5000
shutin
9000
shutin
25000
17000
5000
shutin
20000
stop
stop
10
Oil-water Model : Oil Production Rate
10000
prsi well producer
CMG IMEX well producer
9000
Oil Production Rate (STB/day)
8000
7000
6000
5000
4000
3000
2000
1000
0
0
5000
10000
time (days)
15000
20000
Figure 1: Oil production rate for well producer of example two-phase model from mxsmo031 (unit:
stb/day).
Oil-water Model BHP
4500
prsi well injector
CMG IMEX well injector
prsi well producer
CMG IMEX well producer
4000
3500
BHP (psi)
3000
2500
2000
1500
1000
500
0
0
5000
10000
time (days)
15000
20000
Figure 2: Oil production rate for well injector and producer of example two-phase model from mxsmo031
(pressure unit: psi).
Figure 1 shows oil production rate from our methods and CMG IMEX, from which we can see they
match very well. Figure 2 compares bottom hole pressure from this paper and from CMG IMEX. Again,
the results match very well. The same happens in Figure 3, which compares water rate with results from
CMG IMEX. We can conclude that our methods and our implementation are correct.
Example 2 This case is also from CMG IMEX (mxspe1.dat). It is a 3D two-phase (oil-water) reservoir
model with constant porosity and permeability varying in the z-direction layers. The mesh is 10× 10×3
with mesh size 1000 ft.in x and y directions and 20.0, 30.0, 50.0 ft. in z direction from top to bottom. The
depth of the top layer center is 8335 ft. From top to bottom, the permeability in x direction is 500, 50, 200
mD, same as the permeability in y direction. The permeability in z direction is 60, 40, 20 mD. The porosity
is 0.3.
11
Oil-water Model : Water Rate
12000
prsi well injector
CMG IMEX well injector
prsi well producer
CMG IMEX well producer
Water Rate (STB/day)
10000
8000
6000
4000
2000
0
0
5000
10000
time (days)
15000
20000
Figure 3: Water rate for well injector and producer of example two phase model from mxsmo031 (pressure
unit: stb/day).
Component properties: densities of gas, oil and water are 0.0647 lbm/ft3, 46.244 lbm/ft3 and 62.238
lbm/ft3.
The initial conditions are as follows: reference pressure 4800 psi at associated depth 8400 ft, depth to
water-oil contact is 9500 ft.
There are one injection well and one production well. Both are vertical. Injection well has maximum
water injection rate 1.0e+5 bbl/day, maximum bottom hole pressure 2.0e+4 psi, well index 1.0e+5 with
perforation at cell [1 1 1]. Production well has maximum oil production rate 2.0e+4 stb/day, maximum
bottom hole pressure 1000 psi. Well radius 0.25 ft with perforation at cell [10 10 3]. The simulation time
is 3650 days.
Oil-water two-phase Model revised from SPE1 : Oil Production Rate
25000
prsi well producer
CMG IMEX well producer
Oil Production Rate (STB/day)
20000
15000
10000
5000
0
0
500
1000
1500
2000
time (days)
2500
3000
3500
4000
Figure 4: Oil production rate for well producer of example two-phase model from spe1 (unit: stb/day).
12
Oil-water two-phase model revised from SPE1 : BHP
14000
prsi well producer
CMG IMEX well producer
12000
BHP (psi)
10000
8000
6000
4000
2000
0
0
500
1000
1500
2000
time (days)
2500
3000
3500
4000
Figure 5: BHP for well injector and producer of example two-phase model from spe1 (pressure unit: psi).
Oil-water two-phase Model revised from SPE1 : Water Rate
0.5
prsi well producer
CMG IMEX well producer
0.45
0.4
Water Rate (STB/day)
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
500
1000
1500
2000
time (days)
2500
3000
3500
4000
Figure 6: Water rate for well injector and producer of example two-phase model from spe1 (pressure unit:
stb/day).
The results from this example are similar. We compare oil production rate, bottom hole pressure and
water rate with CMG IMEX, which are shown in Figure 4, 5, 6, respectively. Again, we can see that our
results match CMG IMEX, which prove our methods and implementation are correct.
Newton Methods
This section compares the standard Newton method and the inexact Newton method.
Example 3 The SPE10 project [33] is applied. The stopping criterion for both nonlinear methods is 1e-2.
The stopping criterion for the linear systems from the standard Newton method is 1e-4. The criterion
for the linear systems from inexact Newton method is determined automatically by the simulator. Other
settings are the same for these two methods. The maximal iterations for the nonlinear methods are 20.
13
The BICGSTAB method is the linear solver for all linear systems, and the maximal number of iterations
for BICGSTAB is 300. The preconditioner is the CPR-FPF method. The overlap for RAS method is 1. The
local potential reordering technique and the Quasi-IMPES strategy are applied. The simulation period
is 2,000 days and the maximal time step is 100 days. The numerical summaries for the standard Newton
method is shown in Table 2 and the summaries for the inexact Newton method is shown in Table 3. The
numerical results are compared with other simulators to validate our implementation, including average
pressure and oil rate, which are shown in Fig 7 and 8. Their scalability is shown by Figure 9.
Table 2: Numerical summaries of the Newton method
# procs
4
8
16
32
64
# Steps
51
50
53
51
52
# Ntn
240
236
247
240
244
# Slv
8817
8916
9258
8789
9393
# Avg. solver Time (s) Avg. time (s)
36.7
102546.6
427.2
37.7
46241.1
195.9
37.4
23753.3
96.1
36.6
11109.0
46.2
38.4
6236.8
25.5
Table 3: Numerical summaries of the inexact Newton method
# procs # Steps
4
88
8
91
16
84
32
96
64
92
# Ntn
522
542
499
582
551
# Slv # Avg. solver Time (s) Avg. time (s)
5231
10.0
87708.3
168.0
5712
10.5
44873.2
82.7
5010
10.0
19840.4
39.7
6177
10.6
12024.4
20.6
5851
10.6
5585.6
10.1
Fig 7 compares average pressure of all grid cells and the average pressure is compared with a commercial simulator, from which we can see our results (Avg Pres, UC ) match that simulator. Fig 8 compares
oil rate with results from several oil companies. We can also see that our results (UCalgary ) match other
simulators very well.
In Tables 2 and 3, the number of MPI processors, number of time steps, total nonlinear iterations, total
linear iterations, average linear iterations per Newton iteration (# Avg. solver), total simulation time and
average time per Newton iteration (Avg. time) are presented. From these two tables, we can see that the
standard Newton method requires much less time steps than the inexact Newton method, which means
that the average time step of the standard Newton method is larger. The total number of the nonlinear
iterations for the inexact Newton method is much larger than the standard Newton method. The results
show that the standard Newton method converges faster than the inexact Newton method for this project.
One reason is that the standard Newton method solves the linear systems much more accurately than the
inexact Newton method. However, for the linear iterations, we can see that the standard Newton method
requires three times more iterations than the inexact Newton method due to the accuracy required by the
standard Newton method. As a consequence, the standard Newton method uses more simulation time than
the inexact Newton method. Fig 9 shows that the parallel techniques applied in the manuscript are scalable
and effective.
14
6200
Avg Pres
Avg Pres, UC
6000
5800
5600
5400
5200
5000
4800
4600
4400
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Figure 7: Comparison of average pressure.
8000
UCalgary
Chevron
Geoquest
Landmark
Streamsim
7000
6000
5000
4000
3000
2000
1000
0
0
500
1000
1500
2000
Figure 8: Comparison of oil rate.
Matrix Processing
Example 4 The case is SPE10 project. The inexact Newton method is applied, its stopping criterion is
1e-2 and its maximal Newton iterations are 20. The restarted GMRES method is the linear solver for the
solution of Jacobian systems. The overlap for the RAS method is 1. The maximal number of iterations
for the solver is 100. The simulation period is 2,000 days and the maximal time step is 100 days. The
numerical summaries are given in Table 4. The performance of different stages of the solver is presented
in Table 5 and 64 MPI processors (CPU cores) are employed.
15
Standard Newton
Inexact Newton
Ideal
16
14
12
Speedup
10
8
6
4
2
0
0
10
20
30
40
# CPU cores
50
60
70
Figure 9: Scalability of the Newton and inexact Newton methods
Table 4: Numerical summaries of the matrix processing strategies
# procs
# Steps
NONE
161 (802.65)
PT
198
ABF
166
QI
77
PT-ABF
92
PT-QI
71
# Ntn
1808
1370
1075
466
550
403
# Slv # Avg. solver Time (s) Avg. time (s)
401342
221.98
97784.1
54.1
267182
195.0
74686.8
54.5
36043
33.5
14038.6
13.0
15965
34.2
5597.6
12.0
19930
36.2
8804.2
16.0
12976
32.1
5437.0
13.4
When the potential reordering and the matrix decoupling methods are both disabled (NONE ), the simulation fails, which indicates that the potential reordering and the matrix decoupling are necessary for
the preconditioner. When only the potential reordering method is applied (PT ), the simulation succeeds.
However, the number of its linear iterations are 267,182 and the number of its average linear iterations
per Newton iteration are 195. The result means that applying the potential reordering method alone is not
effective enough. When using the Quasi-IMPES strategy only (QI ), the numbers of the linear iterations
and the nonlinear iterations are 15,965 and 466, respectively, which is much more efficient than the potential reordering method. When using the ABF strategy only (ABF ), the number of the linear iterations
are 36,043. When using both the potential reordering method and the Quasi-IMPES strategy (PT-QI ),
the numbers of the linear iterations and the nonlinear iterations are 12,976 and 403, respectively, which is
better than using either the potential reordering method or the Quasi-IMPES strategy. When using both the
potential reordering method and the ABF strategy (PT-ABF ), its performance is also better than applying
the ABF strategy. We can conclude that the potential reordering method and the matrix decoupling strategies are important to the solution of the Jacobian systems. For the Quasi-IMPES and ABF decoupling
strategies, they have better performance when working with the potential reordering method.
16
Table 5: Performance of different stages of the linear solver and preconditioner
# Nprocs
32
64
128
PT (s)
0.696
0.364
0.193
ABF (s)
1.254
0.628
0.306
QI (s)
0.258
0.138
0.063
RAS (s)
1.048
0.622
0.354
AMG (s) Itr (s)
2.908
0.635
1.309
0.309
0.665
0.159
Table 5 shows the performance of the locally potential reordering (PT ), the ABF decoupling strategy
(ABF ), the Quasi-IMPES decoupling strategy (QI ), the assembling of the RAS method (RAS ), the assembling of the algebraic multigrid method (AMG ) and one solver iteration (Itr ). The results show that the
locally potential reordering takes around two solver iterations and the running time of the locally potential
reordering reduces if more MPI processors are applied. The ABF strategy requires much more time than
the Quasi-IMPES strategy and they require less time if more MPI processors are employed. The assembling of the RAS method and the AMG method take around six solver iterations. The total cost of the
assembling stage is around nine solver iterations, while Table 4 shows that the solver of each linear system
converges within around 35 linear iterations. We can see that the assembling stage takes around 20% of
the total solution time (including the assembling and solving of linear systems). The assembling stage and
the solving stage are also scalable.
Linear Solvers
This section studies the linear solvers, where the restarted GMRES(m) method and the BICGSTAB method
are tested.
Example 5 The case is SPE10 project. The inexact Newton method is applied and its stopping criterion is
1e-2. Its maximal Newton iterations are 20. The restarted GMRES(m) method is the GMRES(50) method.
The maximal number of (inner) iterations for the solvers is 100. The overlap for the RAS method is 1. The
potential reordering and the Quasi-IMPES strategies are applied. The simulation period is 2,000 days
and the maximal time step is 100 days. The numerical summaries for GMRES(50) and BICGSTAB are
presented in Tables 6 and 3, respectively.
Table 6: Numerical summaries of the restarted GMRES(50) method
# procs
4
8
16
32
64
# Steps
68
69
68
69
71
# Ntn
380
389
383
393
403
# Slv
11245
11758
11428
11415
12976
# Avg. solver Time (s) Avg. time (s)
29.5
82730.2
217.7
30.2
41324.1
106.2
29.8
20109.3
52.5
29.0
9875.1
25.1
32.1
5438.2
13.4
Tables 6 and 3 show that the GMRES(m) method uses smaller time steps and fewer Newton iterations
than the BICGSTAB method. However, the GMRES(m) method uses much more linear iterations than the
BICGSTAB method. The numerical results show that both the GMRES(m) method and the BICGSTAB
17
method are efficient in the simulations, whose average iterations per Newton iteration are around 30 and
10 for GMRES(50) and BICGSTAB, respectively. We should mention that the GMRES(m) solves one
preconditioning system in each inner linear iteration, while the BICGSTAB method solves two preconditioning systems in each linear iteration. However, the overall running time for both methods is comparable.
The average running time per Newton iteration also shows that they are scalable.
Scalability
This section tests the strong scalability of the parallel techniques developed.
Example 6 This example benchmarks a refined SPE10 case, where each grid block is refined into eight
grid blocks. This case has around nine millions of grid blocks and around 18 millions of unknowns.
The stopping criterion for the inexact Newton method is 1e-3 and the maximal Newton iterations are 20.
The BICGSTAB solver is applied and its maximal iterations are 100. The potential reordering and the
Quasi-IMPES decoupling strategy are applied. The overlap for RAS method is 1. The simulation period
is 20 days and the maximal time step is 10 days. The numerical summaries are shown in Table 7 and the
speedup (scalability) is shown in Figure 10.
Table 7: Numerical summaries of Example 6
# procs # Steps
16
37
32
35
64
38
128
38
256
37
512
40
1024
37
# Ntn
231
205
225
244
230
248
231
# Slv # Avg. solver Time (s) Avg. time (s)
2816
12.1
93900.4
406.4
2569
12.5
37854.6
184.6
3265
14.5
21699.5
96.4
3290
13.4
10514.2
43.0
3044
13.2
4889.9
21.2
3640
14.6
3071.8
12.3
3177
13.7
1555.7
6.7
This case is more difficult than the original SPE10 problem. The results from Table 7 show that the
simulations take around 40 time steps to complete 20 days, which means that the average time step is
small. The average linear iterations for each Newton iteration are between 12 and 15, which are slightly
more than that for the original SPE10 problem. The simulation times and Figure 10 show that the parallel
techniques here have good scalability.
Example 7 This example tests another refined SPE10 case, where each grid block is refined into 27 grid
blocks. This case has around 30 millions of grid blocks and around 60 millions of unknowns. The stopping
criterion for the inexact Newton method is 1e-2 and the maximal Newton iterations are 20. The BICGSTAB
solver is applied and its maximal iterations are 100. The potential reordering and the Quasi-IMPES
decoupling strategy are applied. The simulation period is 10 days and the maximal time step is 5 days.
The numerical summaries are shown in Tables 8 and 9, and the speedup (scalability) is shown in Figure
11.
18
70
Refined SPE10
Ideal
60
Speedup
50
40
30
20
10
0
0
200
400
600
# CPU cores
800
1000
1200
Figure 10: Scalability of Example 6
Table 8: Numerical summaries of Example 7, overlap 1
# procs # Steps
64
47
128
49
256
43
512
47
1024
47
2048
47
# Ntn
262
290
247
286
278
274
# Slv # Avg. solver Time (s) Avg. time (s)
2569
9.8
94344.7
360.0
2710
9.3
45517.1
156.9
2506
10.1
20313.8
82.2
2606
9.1
11222.0
39.2
2965
10.6
6177.3
22.2
2845
10.3
3217.7
11.7
Table 9: Numerical summaries of Example 7, overlap 2
# procs # Steps
64
47
128
45
256
46
512
46
1024
45
2048
46
# Ntn
263
276
272
260
253
257
# Slv # Avg. solver Time (s) Avg. time (s)
2153
8.1
89385.5
339.8
2312
8.3
42387.0
153.5
2506
9.2
22038.1
81.0
2412
9.2
10963.7
42.1
2457
9.7
5631.0
22.2
2486
9.6
3138.1
12.2
For this example, up to 2,048 MPI processors are employed. Tables 8 and 9 show numerical summaries
of the simulations with overlaps one and two, respectively. The total number of time steps is between 40
and 50 and the average time step is smaller than that in Example 6 due to the smaller size of grid blocks.
19
Refined SPE10, overlap 1
Refined SPE10, overlap 2
Ideal
30
25
Speedup
20
15
10
5
0
0
200
400
600
800
1000
1200
# CPU cores
1400
1600
1800
2000
2200
Figure 11: Scalability of Example 7
The tables show that the linear solver converges in around 10 iterations, which means that both the linear
solver and the preconditioner are efficient. When we increase the number of MPI processors, the average
linear iterations per Newton iteration increase slightly, which demonstrates the robustness of the linear
solver and the preconditioner. The running times and Figure 11 show good scalability; especially when
the number of MPI processors is not larger than 1,024, the scalability is almost ideal. For the case with
2,048 MPI processors, each node (computer card) runs 16 MPI processors, which share the memory and
network, and compete the computation and communication resources, the scalability is reduced slightly,
and the running time is slightly longer than that in the ideal condition. We emphasize that the results show
our parallel techniques and methods are efficient, robust and scalable.
CONCLUSIONS
This paper studies two-phase oil-water model and various numerical techniques are studied, including
nonlinear methods, linear solvers and advanced matrix processing strategies, which are mainly developed
for distributed memory parallel systems. Therefore, their effectiveness and scalability are important. The
matrix decoupling strategies are applied locally and no communication is involved. In this case, the
scalability is ideal. For the matrix reordering, the potential reordering strategy is applied, which is applied
locally and so is the matrix decoupling strategy. The communication volume for the matrix reordering
technique is small and this technique is also scalable for parallel systems. Numerical experiments for
the nonlinear methods, matrix processing techniques, linear solvers, and RAS method show that these
methods and techniques are efficient and scalable. The oil-water model is a simpler model for reservoir
simulation compared with the black oil, polymer flooding, compositional and thermal models. However,
those numerical techniques can be applied to these more complex models directly.
20
ACKNOWLEDGEMENTS
The support of Department of Chemical and Petroleum Engineering, University of Calgary and Reservoir
Simulation Research Group is gratefully acknowledged. The research is partly supported by NSERC,
AIEES, Foundation CMG, AITF iCore, IBM Thomas J. Watson Research Center, and the Frank and Sarah
Meyer FCMG Collaboration Centre for Visualization and Simulation. The research is also enabled in part
by support provided by WestGrid (www.westgrid.ca), SciNet (www.scinethpc.ca) and Compute Canada
Calcul Canada (www.computecanada.ca).
References
[1] K. Li, A. Dogru, A. McDonald, A. Merchant, A. Al-Mulhem, S. Al-Ruwaili, N. Sobh, and H. AlSunaidi, Improving the Performance of Mars Simulator on Cray-2 Supercomputer, Middle East Oil
Show, Society of Petroleum Engineers, Bahrain, 11-14 March 1995.
[2] J. Wheeler, and R. Smith, Reservoir Simulation on a Hypercube, SPE Reservoir Engineering, 1990,
5(04), 544.
[3] J. Rutledge, D. Jones, W. Chen, and E. Chung, The Use of Massively Parallel SIMD Computer for
Reservoir Simulation, SPE Computer Applications, 1992, 4(04), 16.
[4] G. Shiralkar, R. Stephenson, W. Joubert, O. Lubeck, and B. van Bloemen Waanders, Falcon: A
production quality distributed memory reservoir simulator, SPE Reservoir Simulation Symposium,
Society of Petroleum Engineers, Dallas, 8-11 Jun 1997.
[5] M. Parashar, J. Wheeler, G. Pope, K. Wang, and P. Wang, A New Generation EOS Compositional
Reservoir Simulator: Part II, Framework and Multiprocessing, SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, Dallas, 8-11 Jun 1997.
[6] T. Kaarstad, J. Froyen, P. Bjorstad, M. Espedal, A Massively Parallel Reservoir Simulator, SPE
Reservoir Simulation Symposium, Society of Petroleum Engineers, San Antonio, 12-15 Feb 1995.
[7] J. Killough, D. Camilleri, B. Darlow, J. Foster, Parallel Reservoir Simulator Based on Local Grid
Refinement, SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, Dallas, 8-11
Jun 1997.
[8] A. Dogru, H. Sunaidi, L. Fung, W. Habiballah, N. Al-Zamel, K. Li, A parallel reservoir simulator
for large-scale reservoir simulation. SPE Reservoir Evaluation & Engineering, 2002, 5(1), 11.
[9] A. Dogru, L. Fung, U. Middya, T. Al-Shaalan, J. Pita, A next-generation parallel reservoir simulator
for giant reservoirs, SPE/EAGE Reservoir Characterization & Simulation Conference, Society of
Petroleum Engineers, The Woodlands, 2-4 Feb 2009.
[10] Y. Saad, Iterative methods for sparse linear systems, 2nd edition, SIAM, Philadelphia, 2003.
[11] J. Wallis, Incomplete Gaussian elimination as a preconditioning for generalized conjugate gradient
acceleration, SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, San Francisco,
15-18 Nov 1983.
21
[12] G. Behie, and P. Forsyth, Incomplete factorization methods for fully implicit simulation of enhanced
oil recovery. SIAM Journal on Scientific and Statistical Computing, 1984, 5(3), 543.
[13] J. Wallis, R. Kendall and T. Little, Constrained residual acceleration of conjugate residual methods,
SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, Dallas, 10-13 Feb 1985.
[14] H. Cao, H. Tchelepi, J. Wallis, H. Yardumian, Parallel scalable unstructured CPR-type linear solver
for reservoir simulation. SPE Annual Technical Conference and Exhibition, Society of Petroleum
Engineers, Dallas, 9-12 Oct 2005.
[15] R. Bank, T. Chan, W. Coughran Jr., R. Smith, The Alternate-Block-Factorization procedure for systems of partial differential equations, BIT Numerical Mathematics, 1989, 29(4), 938.
[16] K. Stüben, T. Clees, H. Klie, B. Lou, M. Wheeler, Algebraic multigrid methods (AMG) for the
efficient solution of fully implicit formulations in reservoir simulation, SPE Reservoir Simulation
Symposium, Society of Petroleum Engineers, Houston, 26-28 Feb 2007.
[17] K. Stüben, A review of algebraic multigrid, Journal of Computational and Applied Mathematics,
2001, 128(1), 281.
[18] S. Lacroix, Y. Vassilevski, and M. Wheeler, Decoupling preconditioners in the implicit parallel accurate reservoir simulator (IPARS), Numerical linear algebra with applications, 2001, 8(8), 537.
[19] R. Scheichl, R. Masson, and R. Wendebourg, Decoupling and Block Preconditioning for Sedimentary
Basin Simulations, Computational Geosciences, 2003, 7, 295.
[20] T. Al-Shaalan, H. Klie, A. Dogru, M. Wheeler, Studies of Robust Two Stage Preconditioners for
the Solution of Fully Implicit Multiphase Flow Problems, SPE Reservoir Simulation Symposium,
Society of Petroleum Engineers, The Woodlands, 2-4 Feb 2009.
[21] X. Hu, W. Liu, G. Qin, J. Xu, C. Zhang, Development of a fast auxiliary subspace pre-conditioner
for numerical reservoir simulators, SPE Reservoir Characterisation and Simulation Conference and
Exhibition, Society of Petroleum Engineers, Abu Dhahi, 9-11 Oct 2011.
[22] H. Liu, K. Wang, Z. Chen, and K. Jordan, Efficient Multi-stage Preconditioners for Highly Heterogeneous Reservoir Simulations on Parallel Distributed Systems, SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, Houston, 23-25 Feb 2015.
[23] T. Chen, N. Gewecke, Z. Li, A. Rubiano, R. Shuttleworth, Y. Yang, and X. Zhong, Fast Computational Methods for Reservoir Flow Models, report-082009, University of Minnesota, 2009.
[24] A. Toselli, and O. Widlund, Domain decomposition methods: algorithms and theory, Vol 34,
Springer, New York, 2005.
[25] L. Zhang, T. Cui, and H. Liu, A set of symmetric quadrature rules on triangles and tetrahedra, J.
Comput. Math, 2009, 27(1), 89.
[26] X. Cai, and M. Sarkis, A restricted additive Schwarz preconditioner for general sparse linear systems,
SIAM Journal on Scientific Computing, 1999, 21(2), 792.
22
[27] Z. Chen, G. Huan, and Y. Ma, Computational methods for multiphase flows in porous media, Vol 2,
SIAM, Philadelphia, 2006.
[28] F. Kwok F, and H. Tchelepi, Potential-based reduced newton algorithm for nonlinear multiphase flow
in porous media, Journal of Computational Physics, 2007, 227(1), 706.
[29] R. Falgout, and U. Yang, Hypre: A library of high performance preconditioners, Lecture Notes in
Computer Science, 2002, 632.
[30] H. Liu, K. Wang, Z. Chen, K. Jordan, J. Luo, and H. Deng, A Parallel Framewrok for Reservoir
Simulators on Distributed-memory Supercomputers, SPE/IATMI Asia Pacific Oil & Gas Conference
and Exhibition, Society of Petroleum Engineers, Nusa Dua, 20-22 Oct 2015.
[31] H. Liu, Dynamic Load Balancing on Adaptive Unstructured Meshes, 10th IEEE International Conference on High Performance Computing and Communications, Dalian, 25-27 Sept 2008.
[32] A. Sedighi, Y. Deng, B. Zhang, Fairness of Task Scheduling in High Performance Computing Environments, Scalable Computing: Practice and Experience, 2014, 15(3), 273.
[33] M. Christie, and M. Blun, Tenth SPE comparative solution project: A comparison of upscaling
techniques. SPE Reservoir Evaluation & Engineering, 2001, 4(4), 308.
[34] G. Karypis, K. Schloegel, and V. Kumar, Parallel static and dynamic multi-constraint graph partitioning, Concurrency and Computation: Practice and Experience, 2002, 14(3), 219.
[35] L. Zhang, A Parallel Algorithm for Adaptive Local Refinement of Tetrahedral Meshes Using Bisection, Numer. Math.: Theory, Methods and Applications, 2009, 2, 65.
23
| 5cs.CE
|
A
Optimal Partitioning for Dual-Pivot Quicksort
arXiv:1303.5217v3 [cs.DS] 13 Oct 2015
Martin Aumüller and Martin Dietzfelbinger, Technische Universität Ilmenau
Dual-pivot quicksort refers to variants of classical quicksort where in the partitioning step two pivots are
used to split the input into three segments. This can be done in different ways, giving rise to different
algorithms. Recently, a dual-pivot algorithm proposed by Yaroslavskiy received much attention, because a
variant of it replaced the well-engineered quicksort algorithm in Sun’s Java 7 runtime library. Nebel and
Wild (ESA 2012) analyzed this algorithm and showed that on average it uses 1.9n ln n + O(n) comparisons
to sort an input of size n, beating standard quicksort, which uses 2n ln n+O(n) comparisons. We introduce a
model that captures all dual-pivot algorithms, give a unified analysis, and identify new dual-pivot algorithms
that minimize the average number of key comparisons among all possible algorithms up to a linear term.
This minimum is 1.8n ln n + O(n). For the case that the pivots are chosen from a small sample, we include
a comparison of dual-pivot quicksort and classical quicksort. Specifically, we show that dual-pivot quicksort
benefits from a skewed choice of pivots. We experimentally evaluate our algorithms and compare them to
Yaroslavskiy’s algorithm and the recently described 3-pivot quicksort algorithm of Kushagra et al. (ALENEX
2014).
Categories and Subject Descriptors: F.2.2 [Nonnumerical Algorithms and Problems]: Sorting and searching
General Terms: Algorithms
Additional Key Words and Phrases: Sorting, Quicksort, Dual-Pivot
ACM Reference Format:
Martin Aumüller and Martin Dietzfelbinger. 2014. Optimal Partitioning for Dual-Pivot Quicksort. ACM
Trans. Algor. V, N, Article A (January YYYY), 34 pages.
DOI:http://dx.doi.org/10.1145/0000000.0000000
1. INTRODUCTION
Quicksort [Hoare 1962] is a thoroughly analyzed classical sorting algorithm, described
in standard textbooks such as [Cormen et al. 2009; Knuth 1973; Sedgewick and Flajolet
1996] and with implementations in practically all algorithm libraries. Following the
divide-and-conquer paradigm, on an input consisting of n elements quicksort uses a
pivot element to partition its input elements into two parts, the elements in one part
being smaller than or equal to the pivot, the elements in the other part being larger
than or equal to the pivot, and then uses recursion to sort these parts. It is well known
that if the input consists of n elements with distinct keys in random order and the
pivot is picked by just choosing an element then on average quicksort uses 2n ln n +
A preliminary version of this work appeared in the Proceedings of the 40th International Colloquium on
Automata, Languages, and Programming (ICALP), Part 1, 33–44, 2013.
Author’s addresses: M. Aumüller; M. Dietzfelbinger, Fakultät für Informatik und Automatisierung, Technische Universität Ilmenau, 98683 Ilmenau, Germany; e-mail: {martin.aumueller,martin.dietzfelbinger}@tuilmenau.de
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights
for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted.
To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this
work in other works requires prior specific permission and/or a fee. Permissions may be requested from
Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or [email protected].
c YYYY ACM 1549-6325/YYYY/01-ARTA $15.00
DOI:http://dx.doi.org/10.1145/0000000.0000000
ACM Transactions on Algorithms, Vol. V, No. N, Article A, Publication date: January YYYY.
Fig. 1. Result of the partition step in dual-pivot quicksort schemes using two pivots p, q with p ≤ q. Elements
left of p are smaller than or equal to p, elements right of q are larger than or equal to q. The elements between
p and q are at least as large as p and at most as large as q.
O(n) comparisons.1 In 2009, Yaroslavskiy announced2 that he had found an improved
quicksort implementation, the claim being backed by experiments. After extensive
empirical studies, in 2009 a variant of Yaroslavskiy’s algorithm due to Yaroslavskiy,
Bentley, and Bloch became the new standard quicksort algorithm in Sun’s Java 7
runtime library. This algorithm employs two pivots to split the elements. If two pivots
p and q with p ≤ q are used, the partitioning step partitions the remaining n − 2
elements into 3 parts: elements smaller than or equal to p, elements between p and q,
and elements larger than or equal to q, see Fig. 1.3 Recursion is then applied to the
three parts. As remarked in [Wild and Nebel 2012], it came as a surprise that two pivots
should help, since in his thesis [Sedgewick 1975] Sedgewick had proposed and analyzed
a dual-pivot approach that was inferior to classical quicksort. Later, Hennequin in his
thesis [Hennequin 1991] studied the general approach of using k ≥ 1 pivot elements.
According to [Wild and Nebel 2012], he found only slight improvements that would not
compensate for the more involved partitioning procedure. (See [Wild and Nebel 2012]
for a short discussion.)
In [Wild and Nebel 2012] (see also the full version [Wild et al. 2015]), Nebel and
Wild formulated and analyzed a simplified version of Yaroslavskiy’s algorithm. (For
completeness, this algorithm is given as Algorithm 2 in Appendix B.2.) They showed that
it makes 1.9n ln n + O(n) key comparisons on average, in contrast to the 2n ln n + O(n)
of standard quicksort and the 32
15 n ln n + O(n) of Sedgewick’s dual-pivot algorithm. On
the other hand, they showed that the number of swap operations in Yaroslavskiy’s
algorithm is 0.6n ln n + O(n) on average, which is much higher than the 0.33n ln n + O(n)
swap operations in classical quicksort. In this paper we concentrate on the comparison
count as cost measure and on asymptotic results. We leave the study of other cost
measures to further investigations (which already have taken place, see [Martı́nez et al.
2015]). We consider other measures in experimental evaluations.
The authors of [Wild and Nebel 2012] state that the reason for Yaroslavskiy’s algorithm being superior were that his “partitioning method is able to take advantage
of certain asymmetries in the outcomes of key comparisons”. They also state that
“[Sedgewick’s dual-pivot method] fails to utilize them, even though being based on the
same abstract algorithmic idea”. So the abstract algorithmic idea of using two pivots
can lead to different algorithms with different behavior. In this paper we describe the
design space from which all these algorithms originate. We fully explain which simple
property makes some dual-pivot algorithms perform better and some perform worse
w.r.t. the average comparison count and identify optimal members (up to a linear term)
of this design space. The best ones use 1.8n ln n + O(n) comparisons on average—even
less than Yaroslavskiy’s method.
The first observation is that everything depends on the cost, i. e., the comparison
count, of the partitioning step. This is not new at all. Actually, in Hennequin’s thesis [Hennequin 1991] the connection between partitioning cost and overall cost for
1
In this paper ln denotes the natural logarithm and log denotes the logarithm to base 2.
archived version of the relevant discussion in a Java newsgroup can be found at
http://permalink.gmane.org/gmane.comp.java.openjdk.core-libs.devel/2628. Also see [Wild and Nebel
2012].
3 In accordance with tradition, we assume in this theoretical study that all elements have different keys. Of
course, in implementations equal keys are an important issue that requires a lot of care [Sedgewick 1977].
2 An
2
quicksort variants with more than one pivot is analyzed in detail. For dual-pivot
quicksort, Hennequin proved that if the (average) partitioning cost for n elements is
a · n + O(1), for a constant a, then the average cost for sorting n elements is
6
a · n ln n + O(n).
(1)
5
The partitioning cost of some algorithms presented in this paper will have a nonconstant lower order term. Utilizing the Continuous Master Theorem of [Roura 2001],
we prove that (1) describes the average cost even if the partitioning cost is a · n + O(n1−ε )
for some ε > 0. Throughout the present paper all that interests us is the constant factor
with the leading term. (The reader should be warned that for real-life n the linear
term, which can even be negative, can have a big influence on the average number of
comparisons.)
The second observation is that the partitioning cost depends on certain details of the
partitioning procedure. This is in contrast to standard quicksort with one pivot where
partitioning always takes n − 1 comparisons. In [Wild and Nebel 2012] it is shown that
Yaroslavskiy’s partitioning procedure uses 19
12 n + O(1) comparisons on average, while
16
Sedgewick’s uses 9 n + O(1) many. The analysis of these two algorithms is based on the
study of how certain pointers move through the array, at which positions elements are
compared to the pivots, which of the two pivots is used for the first comparison, and how
swap operations exchange two elements in the array. For understanding what is going
on, however, it is helpful to forget about concrete implementations with loops in which
pointers sweep across arrays and entries are swapped, and look at partitioning with
two pivots in a more abstract way. For simplicity we shall always assume that the input
is a permutation of {1, . . . , n}. Now pivots p and q with p < q are chosen. The task is to
classify the remaining n − 2 elements into classes “small” (s = p − 1 many), “medium”
(m = q − p − 1 many), and “large” (` = n − p many), by comparing these elements one
after the other with the smaller pivot or the larger pivot, or both of them if necessary.
Note that for symmetry reasons it is inessential in which order the elements are treated.
The only choice the algorithm can make is whether to compare the current element
with the smaller pivot or the larger pivot first. Let the random variable S2 denote the
number of small elements compared with the larger pivot first, and let L2 denote the
number of large elements compared with the smaller pivot first. Then the total number
of comparisons is n − 2 + m + S2 + L2 .
Averaging over all inputs and all possible choices of the pivots the term n − 2 + m
will lead to 43 n + O(1) key comparisons on average, independently of the algorithm.
Let W = S2 + L2 , the number of elements that are compared with the “wrong” pivot
first. Then E(W ) is the only quantity that is influenced by a particular partitioning
procedure.
In the paper, we will first devise an easy method to calculate E(W ). The result of
this analysis will lead to an asymptotically optimal strategy. The basic approach is
the following. Assume a partitioning procedure is given, and assume p, q and hence
s = p − 1 and ` = n − q are fixed, and let ws,` = E(W | s, `). Denote the average number
p
q
of elements compared to the smaller [larger] pivot first by fs,`
[fs,`
]. If the elements to be
classified were chosen to be small, medium, and large independently with probabilities
s/(n − 2), m/(n − 2), and `/(n − 2), resp., then the average number of small elements
q
compared with the large pivot first would be fs,`
· s/(n − 2), similarly for the large
elements. Of course, the actual input is a sequence with exactly s [m, `] small [medium,
large] elements, and there is no independence. Still, we will show that the randomness
in the order is sufficient to guarantee that, for some ε > 0,
q
p
ws,` = fs,`
· s/n + fs,`
· `/n + O(n1−ε ).
3
(2)
p
q
The details of the partitioning procedure will determine fs,`
and fs,`
, and hence ws,` up
1−ε
to O(n ). This seemingly simple insight has two consequences, one for the analysis
and one for the design of dual-pivot algorithms:
(i) In order to analyze the average comparison count of a dual-pivot algorithm (given
p
q
by its partitioning procedure) up to a linear term, determine fs,`
and fs,`
for this
1−ε
partitioning procedure. This will give ws,` up to O(n ), which must then be averaged
over all s, ` to find the average number of comparisons in partitioning. Then apply (1).
(ii) In order to design a good partitioning procedure w.r.t. the average comparison count,
q
p
try to make fs,`
· s/n + fs,`
· `/n small.
We shall demonstrate approach (i) in Section 4. An example: As explained in [Wild
q
and Nebel 2012], if s and ` are fixed, in Yaroslavskiy’s algorithm we have fs,`
≈ ` and
p
fs,`
≈ s + m. By (2) we get ws,` = (`s + (s + m)`)/n + O(n1−ε ). This must be averaged
over all possible values of s and `. The result is 14 n + O(n1−ε ), which together with
4
19
1−ε
), close to the result from [Wild and Nebel 2012].
3 n + O(1) gives 12 n + O(n
Principle (ii) will be used to identify an asymptotically optimal partitioning procedure
that makes 1.8n ln n + O(n) key comparisons on average. In brief, such a strategy should
achieve the following: If s > `, compare (almost) all entries with the smaller pivot first
p
q
(fs,`
≈ n and fs,`
≈ 0), otherwise compare (almost) all entries with the larger pivot first
p
q
(fs,`
≈ 0 and fs,`
≈ n). Of course, some details have to be worked out: How can the
algorithm decide which case applies? In which technical sense is this strategy optimal?
We shall see in Section 5 how a sampling technique resolves these issues.
In Section 6, we will consider the following simple and intuitive strategy: If more
elements have been classified as being large instead of being small so far, compare the
next element to the larger pivot first, otherwise compare it to the smaller pivot first. We
will show that this strategy is optimal w.r.t. minimizing the average comparison count.
In implementations of quicksort, the pivot is usually chosen as the median from a
small sample of 2k + 1 elements. Intuitively, this yields more balanced subproblems,
which are smaller on average. This idea already appeared in Hoare’s original publication
[Hoare 1962] without an analysis. This was later supplied by van Emden [van Emden
1970]. The complete analysis of this variant was given by Martinez and Roura√in
[Martı́nez and Roura 2001] in 2001. They showed that the optimal sample size is n.
For this sample size the average comparison count of quicksort matches the lowerorder bound of n log n + O(n) comparisons. In practice, one usually uses a sample
of size 3. Theoretically, this decreases the average comparison count from 2n ln n +
O(n) to 1.714..n ln n + O(n). This strategy has been generalized in the obvious way to
Yaroslavskiy’s algorithm as well. The implementation of Yaroslavskiy’s algorithm in
Sun’s Java 7 uses the two tertiles in a sample of size 5 as pivots, i.e., the elements
of rank 2 and 4. In Section 7 we will analyze the comparison count of dual-pivot
quicksort algorithms with this sampling strategy. Yaroslavskiy’s algorithm has an
average comparison count of 1.704..n ln n + O(n) in this case, while the optimal average
cost is 1.623..n ln n + O(n). In that section, we will also consider a question raised by
Wild, Nebel, and Martı́nez in [Wild et al. 2014, Section 8] for Yaroslavskiy’s algorithm,
namely of which rank the pivots should be to achieve minimum sorting cost. While
using the tertiles of the input seems the obvious choice for balancing reasons, in [Wild
et al. 2014] it is shown that for Yaroslavskiy’s algorithm this minimum is attained
for ranks ≈ 0.429n and ≈ 0.698n and is ≈ 1.4931n ln n. We will show that the simple
strategy “Always compare with the larger pivot first” achieves sorting cost ≈ 1.4427n ln n,
i. e., the lower bound for comparison-based sorting, when choosing the elements of rank
n/4 and n/2 as the two pivots.
4
As noted in [Wild et al. 2013], considering only key comparisons and swap operations
does not suffice for evaluating the practicability of sorting algorithms. In Section 8, we
will present experimental results that indicate the following: When sorting integers, the
comparison-optimal algorithms of Section 5 and Section 6 are slower than Yaroslavskiy’s
algorithm. However, an implementation of the simple strategy “Always compare with
the larger pivot first” performs very well both in C++ and in Java in our experimental
setup. We will also compare our algorithms to the fast three-pivot quicksort algorithm
described in [Kushagra et al. 2014]. While comparing these algorithms, we will provide
evidence that the theoretical cost measure “cache misses” described in [Kushagra et al.
2014] nicely predicts empirical cache behavior, and comes closest for correctly predicting
running time.
We emphasize that the purpose of this paper is not to arrive at better and better
quicksort algorithms by using all kinds of variations, but rather to thoroughly analyze
the situation with two pivots, showing the potential and the limitations of this approach.
2. BASIC APPROACH TO ANALYZING DUAL-PIVOT QUICKSORT
We assume the input sequence (a1 , . . . , an ) to be a random permutation of {1, . . . , n},
each permutation occurring with probability (1/n!). If n ≤ 1, there is nothing to do;
if n = 2, sort by one comparison. Otherwise, choose the first element a1 and the last
element an as the set of pivots, and set p = min(a1 , an ) and q = max(a1 , an ). Partition
the remaining elements into elements smaller than p (“small” elements), elements
between p and q (“medium” elements), and elements larger than q (“large” elements),
see Fig. 1. Then apply the procedure recursively to these three groups. Clearly,
each
pair p, q with 1 ≤ p < q ≤ n appears as set of pivots with probability 1/ n2 . Our cost
measure is the number of key comparisons needed to sort the given input. Let Cn be the
random variable counting this number. Let Pn denote the partitioning cost to partition
the n − 2 non-pivot elements into the three groups. As explained by Wild and Nebel
[Wild and Nebel 2012, Appendix A], the average number of key comparisons obeys the
following recurrence:
n−2
E(Cn ) = E(Pn ) +
X
2
(n − k − 1) · E(Ck ).
·3
n(n − 1)
(3)
k=0
If E(Pn ) = a · n + O(1), for a constant a, this can be solved (cf. [Hennequin 1991; Wild
and Nebel 2012]) to give
E(Cn ) =
6
a · n ln n + O(n).
5
(4)
In Section 3 we will show that (4) also holds if E(Pn ) = a · n + O(n1−ε ) for some ε > 0.
For the proof we utilize the Continuous Master Theorem from [Roura 2001].
In view of this simple relation it is sufficient to study the cost of partitioning. Abstracting from moving elements around in arrays, we arrive at the following “classification
problem”: Given a random permutation (a1 , . . . , an ) of {1, . . . , n} as the input sequence
and a1 and an as the two pivots p and q, with p < q, classify each of the remaining n − 2
elements as being small, medium, or large. Note that there are exactly s := p − 1 small
elements, m := q − p − 1 medium elements, and ` := n − q large elements. Although this
classification does not yield an actual partition of the input sequence, a classification
algorithm can be turned into a partitioning algorithm using only swap operations but no
additional key comparisons. Since elements are only compared with the two pivots, the
randomness of subarrays is preserved. Thus, in the recursion we may always assume
that the input is arranged randomly.
5
We make the following observations (and fix notation) for all classification algorithms.
One key comparison is needed to decide which of the elements a1 and an is the smaller
pivot p and which is the larger pivot q. For classification, each of the remaining n − 2
elements has to be compared against p or q or both. Each medium element has to be
compared to p and q. On average, there are (n − 2)/3 medium elements. Let S2 denote
the number of small elements that are compared to the larger pivot first, i. e., the
number of small elements that need 2 comparisons for classification. Analogously, let L2
denote the number of large elements compared to the smaller pivot first. Conditioning
on the pivot choices, and hence the values of s and `, we may calculate E(Pn ) as follows:4
X
1
E(Pn ) = (n − 1) + (n − 2)/3 + n
E(S2 + L2 | s, `).
(5)
2
s+`≤n−2
We call the third summand in (5) the additional cost term (ACT), as it is the only value
that depends on the actual classification algorithm.
3. ANALYZING THE ADDITIONAL COST TERM
We will use the following formalization of a partitioning procedure: A classification
strategy is given as a three-way decision tree T with a root and n − 2 levels numbered
0, 1, . . . , n − 3 of inner nodes as well as one leaf level. The root is on level 0. Each node
v is labeled with an index i(v) ∈ {2, . . . , n − 1} and an element l(v) ∈ {p, q}. If l(v) is p,
then at node v element ai(v) is compared with the smaller pivot first; otherwise, i. e.,
l(v) = q, it is compared with the larger pivot first. The three edges out of a node are
labeled σ, µ, λ, resp., representing the outcome of the classification as small, medium,
large, respectively. The label of edge e is called c(e). On each of the 3n−2 paths each
index occurs exactly once. Each input determines exactly one path w from the root to a
leaf in the obvious way; the classification of the elements can then be read off from the
node and edge labels along this path. We call such a tree a classification tree.
Identifying a path π from the root to a leaf lf by the sequence of nodes and edges
(v1 , e1 , v2 , e2 , . . . , vn−2 , en−2 , lf) on it, we define the cost cπ as
cπ = j ∈ {1, . . . , n − 2} | l(vj ) = q ∧ c(ej ) = σ or l(vj ) = p ∧ c(ej ) = λ .
For a given input, the cost of the path associated with this input exactly describes the
number of additional comparisons on this input. An example for such a classification
tree is given in Figure 2.
We let S2T [LT2 ] denote the random variable that for a random input counts the number
of small [large] elements classified in nodes with label q [p]. We now describe how we
can calculate the ACT of a classification tree T . First consider fixed s and ` and let
the input excepting the pivots be arranged randomly. For a node v in T , we let sv , mv ,
and `v , resp., denote the number of edges labeled σ, µ, and λ, resp., from the root to
v. By the randomness of the input, the probability that the element classified at v is
“small”, i. e., that the edge labeled σ is used, is exactly (s − sv )/(n − 2 − level(v)). The
probability that it is “medium” is (m − mv )/(n − 2 − level(v)), and that it is “large” is
(` − `v )/(n − 2 − level(v)). The probability pvs,` that node v in the tree is reached is then
just the product of all these edge probabilities on the unique path from the root to v.
The probability that the edge labeled σ out of a node v is used can then be calculated as
pvs,` · (s − sv )/(n − 2 − level(v)). Similarly, the probability that the edge labeled λ is used
is pvs,` · (` − `v )/(n − 2 − level(v)). Note that all this is independent of the actual ordering
4 We
use the following notation throughout this paper: To indicate that sums run over all
P
(s, `) with s, ` ≥ 0 and s + ` ≤ n − 2 we simply write s+`≤n−2 .
6
n
2
combinations
3:p
µ
σ
2:q
σ
4:p
µ
4:p
σ
λ
4:q
4:p
2:q
σ
a2 : σ
a3 : σ
a4 : σ
λ
µ
σ
λ
2:p
σ
...
µ
2:p
2:p
4:q
λ
λ
4:p
λ
a2 : σ
a3 : µ
a4 : µ
a2 : µ
a3 : µ
a4 : µ
a2 : λ
a3 : µ
a4 : µ
c2 : 0
c3 : 0
c4 : 1
c1 : 1
4:p
µ
...
a2 : λ
a3 : λ
a4 : λ
c5 : 3
Fig. 2. An example for a decision tree to classify three elements a2 , a3 , and a4 according to the pivots a1 and
a5 . Five out of the 27 leaves are explicitly drawn, showing the classification of the elements and the costs ci
of the specific paths.
in which the classification tree inspects the elements. We can thus always assume some
fixed ordering and forget about the label i(v) of node v.
By linearity of expectation, we can sum up the contribution to the additional comparison count for each node separately. Thus, we may calculate
E(S2T +LT2 | s, `) =
X
v∈T
l(v)=q
pvs,` ·
X
s − sv
` − `v
+
pvs,` ·
.
n − 2 − level(v)
n − 2 − level(v)
(6)
v∈T
l(v)=p
The setup developed so far makes it possible to describe the connection between a
classification tree T and its average comparison count in general. Let FpT and FqT be
two random variables that denote the number of elements that are compared with the
q
smaller and larger pivot first, respectively, when using T . Then let fs,`
= E FqT | s, `
p
resp. fs,`
= E FpT | s, ` denote the average number of comparisons with the larger resp.
smaller pivot first, given s and `. Now, if it was decided in each step by independent
random experiments with the correct expectations s/(n − 2), m/(n − 2), and `/(n − 2),
resp., whether an element is small, medium, or large, it would be clear that for example
q
fs,`
· s/(n − 2) is the average number of small elements that are compared with the
larger pivot first. The next lemma shows that one can indeed use this intuition in the
calculation of the average comparison count, except that one gets an additional O(n1−ε )
term due to the elements tested not being independent.
L EMMA 3.1. Let T be a classification tree. Let E(PnT ) be the average number of key
comparisons for classifying an input of n elements using T . Then there exists a constant
ε > 0 such that
X q
4
1
p
E(PnT ) = n + n
fs,` · s + fs,`
· ` + O n1−ε .
3
2 · (n − 2) s+`≤n−2
7
P ROOF. Fix p and q (and thus s, m, and `). We will show that
E(S2T + LT2 | s, `) =
q
p
fs,`
· s + fs,`
·`
n−2
+ O n1−ε .
(7)
(The lemma then follows by substituting this into (5).)
We call a node v in T on-track (to the expected values) if
s
1
s − sv
≤ 1/12
−
n − 2 n − level(v) − 2
n
`
1
` − `v
l(v) = p and
≤ 1/12 .
−
n − 2 n − level(v) − 2
n
l(v) = q and
or
(8)
Otherwise we call v off-track.
We first obtain an upper bound. Starting from (6), we calculate:
E(S2T + LT2 | s, `) =
X
pvs,` ·
v∈T,l(v)=q
s − sv
+
n − 2 − level(v)
X
pvs,` ·
v∈T,l(v)=p
` − `v
n − 2 − level(v)
X
s
`
=
pvs,` ·
+
pvs,` ·
+
n−2
n−2
v∈T,l(v)=q
v∈T,l(v)=p
X
s
s − sv
v
ps,`
−
+
n − 2 − level(v) n − 2
v∈T,l(v)=q
X
`
` − `v
v
ps,`
−
n − 2 − level(v) n − 2
X
v∈T,l(v)=p
≤
X
pvs,` ·
v∈T,l(v)=q
X
v∈T,l(v)=q
v on-track
s
+
n−2
pvs,`
n1/12
+
X
pvs,` ·
v∈T,l(v)=p
X
v∈T,l(v)=q
v off-track
pvs,` +
`
+
n−2
X
v∈T,l(v)=p
v on-track
pvs,`
n1/12
+
X
pvs,` ,
v∈T,l(v)=p
v off-track
(9)
where the last step follows by separating on-track and off-track nodes and using (8).
(For off-track nodes we use that the left-hand side of the inequalities in (8) is at most
1.) For the sums in the last line of (9), consider each level of the classification tree
separately. Since the probabilities pvs,` for nodes v on the same level sum up to 1, the
q
contribution of the 1/n1/12 terms is bounded by O(n11/12 ). Using the definition of fs,`
p
and fs,` , we continue as follows:
8
X
E(S2T +LT2 | s, `) ≤
pvs,` ·
v∈T,l(v)=q
=
=
s
+
n−2
q
p
fs,`
· s + fs,`
·`
n−2
q
fs,`
·s+
p
fs,`
n−2
n−3
X
X
X
`
+
n−2
pvs,` + O n11/12
v∈T
v off-track
v∈T,l(v)=p
pvs,` + O n11/12
X
+
pvs,` ·
v∈T,v off-track
·`
+
Pr(the node reached on level k is off-track) + O n11/12 , (10)
k=0
where in the last step we just rewrote the sum to consider each level in the classification
tree separately. So, to show (7) it remains to bound the sum in (10) by O(n1−ε ).
To do this, consider a random input that is classified using T . Using an appropriate
tail bound, viz. the method of average bounded differences, we will show that with very
high probability we do not reach an off-track node in the classification tree in the first
n − n3/4 levels. Intuitively, this means that it is highly improbable that underway the
observed fraction of small elements deviates very far from the average s/(n − 2).
Let Xj be the 0-1 random variable that is 1 if the j-th classified element is small;
let Yj be the 0-1 random variable that is 1 if the j-th classified element is large. Let
si = X1 + · · · + Xi and `i = Y1 + · · · + Yi for 1 ≤ i ≤ n − 2.
C LAIM 3.2. Let 1 ≤ i ≤ n − 2. Then
Pr(|si − E(si )| > n2/3 ) ≤ 2 exp(−n1/3 /2), and
Pr(|`i − E(`i )| > n2/3 ) ≤ 2 exp(−n1/3 /2).
P ROOF. We prove the first inequality. First, we bound the difference cj between
the expectation of si conditioned on X1 , . . . , Xj resp. X1 , . . . , Xj−1 for 1 ≤ j ≤ i. Using
linearity of expectation we may calculate
cj = E(si | X1 , . . . , Xj ) − E(si | X1 , . . . , Xj−1 )
=
j
X
i
X
Xk +
k=1
k=j+1
i
X
j−1
i
k=1
k=j
X
X s − sj−1
s − sj
−
Xk −
n−j−2
n−j−1
i
X
s − sj−1
n−j−1
k=j
k=j+1
i−j
i−j
i−j+1
= Xj − Xj ·
+ (s − sj−1 )
−
n−j−2
n−j−2 n−j−1
i−j
n−i−2
= Xj 1 −
− (s − sj−1 )
n−j−2
(n − j − 1)(n − j − 2)
i−j
s − sj−1
≤ max Xj 1 −
,
≤ 1.
n−j−2
n−j−1
= Xj +
s − sj−1 − Xj
−
n−j−2
9
In this situation we may apply the bound known as the method of averaged bounded
differences (see [Dubhashi and Panconesi 2009, Theorem 5.3]), which reads
!
t2
,
Pr(|si − E(si )| > t) ≤ 2 exp − P
2 j≤i c2j
and get
Pr(|si − E(si )| > n2/3 ) ≤ 2 exp
−n4/3
2i
,
which is not larger than 2 exp(−n1/3 /2).
Assume that |si − E(si )| = |si − i · s/(n − 2)| ≤ n2/3 . We have
s
n2/3
s
n2/3
s − si
s(1 − i/(n − 2))
≤
+
=
−
−
.
n−2 n−2−i
n−2
n−2−i
n−2−i
n−2−i
That means that for each of the first i ≤ n − n3/4 levels with very high probability we
are in an on-track node on level i, because the deviation from the ideal case that we
see a small element with probability s/(n − 2) is n2/3 /(n − 2 − i) ≤ n2/3 /n3/4 = 1/n1/12 .
Thus, for the first n − n3/4 levels the contribution of the sums of the probabilities of
off-track nodes in (10) is at most n11/12 . For the last n3/4 levels of the tree, we use that
the contribution of the probabilities that we reach an off-track node on level i is at most
1 for a fixed level.
11/12
This shows that the contribution of the sum
). This finishes the
in (10) is O(n
T
T
proof of the upper bound on E S2 + L2 | s, ` given in (10). The calculations for the
lower bound are similar and are omitted here.
There is the following technical complication when using this lemma in analyzing
a strategy that is turned into a dual-pivot quicksort algorithm: The cost bound is
a · n + O(n1−ε ), and Hennequin’s result (Equation (4)) cannot be applied directly to such
partitioning costs. However, the next theorem says that the leading term of (4) applies
to this situation as well, and the additional O(n1−ε ) term in the partitioning cost is
completely covered in the O(n) error term of (4).
T HEOREM 3.3. Let A be a dual-pivot quicksort algorithm that gives rise to a classification tree Tn for each subarray of length n. Assume E(PnTn ) = a · n + O(n1−ε ) for all n,
for some constants a and ε > 0. Then E CnA = 65 an ln n + O(n).
P ROOF. By linearity of expectation we may split the partitioning cost into two terms
t1 (n) = a · n and t2 (n) = K · n1−ε , solve recursion (3) independently for these two cost
terms, and add the results. Applying (4) for average partitioning cost t1 (n) yields an
average comparison count of 65 an ln n + O(n). Obtaining the bound of O(n) for the term
t2 (n) is a standard application of the Continuous Master Theorem of Roura [Roura
2001], and has been derived for the dual-pivot quicksort recurrence by Wild, Nebel, and
Martı́nez in a recent technical report [Wild et al. 2014, Appendix D]. For completeness,
the calculation is given in Appendix A.
Lemma 3.1 and Theorem 3.3 tell us that for the analysis of the average comparison
p
q
count of a dual-pivot quicksort algorithm we just have to find out what fs,`
and fs,`
are
for this algorithm. Moreover, to design a good algorithm (w.r.t. the average comparison
q
p
count), we should try to make fs,`
· s + fs,`
· ` small for each pair s, `.
10
4. ANALYSIS OF SOME KNOWN CLASSIFICATION STRATEGIES
In this section, we will study different classification strategies in the light of the
formulas from Section 3.
Oblivious Strategies. We will first consider strategies that do not use information of
previous classifications for future classifications. To this end, we call a classification
tree oblivious if for each level all nodes v on this level share the same label l(v) ∈ {p, q}.
This means that these algorithms do not react to the outcome of previous classifications,
but use a fixed sequence of pivot choices. Examples for such strategies are, e. g.,
— always compare to the smaller pivot first,
— always compare to the larger pivot first,
— alternate the pivots in each step.
Let fnq denote the average number of comparisons to the larger pivot first. By assumption
q
this value is independent of s and `. Hence these strategies make sure that fs,`
= fnq
p
and fs,` = n − 2 − fnq for all pairs of values s, `.
Applying Lemma 3.1 gives us
X
1
·
(f q · s + (n − 2 − fnq ) · `) + O(n1−ε )
· (n − 2) s+`≤n−2 n
q
X
X
n
−
2
−
f
4
fnq
n
s + n
` + O(n1−ε )
= n + n
·
·
3
·
(n
−
2)
2 · (n − 2)
2
s+`≤n−2
s+`≤n−2
X
4
1
5
= n + n ·
s + O(n1−ε ) = n + O(n1−ε ).
3
3
2
E(Pn ) =
4
n+
3
n
2
s+`≤n−2
Using Theorem 3.3 we get E(Cn ) = 2n ln n + O(n)—the leading term being the same
as in standard quicksort. So, for each strategy that does not adapt to the outcome
of previous classifications, there is no difference to the average comparison count of
classical quicksort. Note that this also holds for randomized strategies such as “flip a
coin to choose the pivot used in the first comparison”, since such a strategy can be seen
as a probability distribution on oblivious strategies.
Yaroslavskiy’s Algorithm. Following [Wild and Nebel 2012, Section 3.2], Yaroslavskiy’s
algorithm is an implementation of the following strategy Y: Compare ` elements to q
q
p
first, and compare the other elements to p first. We get that fs,`
= ` and fs,`
= s + m.
Applying Lemma 3.1, we get
X s`
4
1
(s + m)`
E(PnY ) = n + n
+
+ O(n1−ε ).
3
n
−
2
n
−
2
2
s+`≤n−2
Of course, it is possible to evaluate this sum by hand. We used Maple R to obtain
1−ε
E(PnY ) = 19
). Using Theorem 3.3 gives E(Cn ) = 1.9n ln n + O(n), as in [Wild
12 n + O(n
and Nebel 2012].
Sedgewick’s Algorithm. Following [Wild and Nebel 2012, Section 3.2], Sedgewick’s
algorithm amounts to an implementation of the following strategy S: Compare (on
average) a fraction of s/(s + `) of the keys with q first, and compare the other keys with p
q
p
first. We get fs,`
= (n − 2) · s/(s + `) and fs,`
= (n − 2) · `/(s + `).
11
Plugging these values into Lemma 3.1, we calculate
X s2
4
1
`2
16
S
E(Pn ) = n + n
+
+ O(n1−ε ) =
n + O(n1−ε ).
3
s
+
`
s
+
`
9
2
s+`≤n−2
Applying Theorem 3.3 gives E(Cn ) = 2.133... · n ln n + O(n), as known from [Wild and
Nebel 2012].
Obviously, this is worse than the oblivious strategies considered before.5 This is easily
explained intuitively: If the fraction of small elements is large, it will compare many
elements with q first. But this costs two comparisons for each small element. Conversely,
if the fraction of large elements is large, it will compare many elements to p first, which
is again the wrong decision.
Since Sedgewick’s strategy seems to do exactly the opposite of what one should do to
lower the comparison count, we consider the following modified strategy S 0 : For given p
and q, compare (on average) a fraction of s/(s + `) of the keys with p first, and compare
the other keys with q first. (S 0 simply uses p first when S would use q first and vice
versa.)
1−ε
Using the same analysis as above, we get E(Pn ) = 14
), which yields
9 n + O(n
E(Cn ) = 1.866... · n ln n + O(n)—improving on the standard algorithm and even on
Yaroslavskiy’s algorithm! Note that this has been observed by Wild in his Master’s
Thesis as well [Wild 2013].
Remark. Swapping the first comparison with p and q as in the strategy described
above is a general technique. In fact, if the leading coefficient of the average number of
comparisons for a fixed rule for choosing p or q first is α, e. g., α = 2.133... for strategy
S, then the leading coefficient of the strategy that does the opposite is 4 − α, e. g.,
4 − 2.133... = 1.866... as in strategy S 0 .
5. AN ASYMPTOTICALLY OPTIMAL CLASSIFICATION STRATEGY
Looking at the previous sections, all strategies used the idea that we should compare a
certain fraction of elements to p first, and the other elements to q first. In this section, we
will study the following strategy I: If s > ` then always compare with p first, otherwise
always compare with q first.
Of course, for an implementation of this strategy we have to deal with the problem
of finding out which case applies before the comparisons have been made. We shall
analyze a guessing strategy to resolve this.
5.1. Analysis of the Idealized Classification Strategy
Assume for a moment that for a given random input with pivots p, q the strategy
“magically” knows whether s > ` or not and correctly determines the pivot that should
be used for all comparisons. For fixed s and ` this means that for s > ` the classification
strategy makes exactly ` additional comparisons, and for s ≤ ` it makes s additional
comparisons.
When we start from (5), a standard calculation shows that for this strategy
X
4
1
3
E(Pn ) = n + n
min(s, `) = n + O(1).
(11)
3
2
2
s+`≤n−2
Applying (4), we get E(Cn ) = 1.8n ln n + O(n), which is by 0.1n ln n smaller than the
average number of key comparisons in Yaroslavskiy’s algorithm.
5 We
remark that in his thesis Sedgewick [Sedgewick 1975] focused on the average number of swaps, not on
the comparison count.
12
To see that this method is asymptotically optimal, recall that according to Lemma 3.1
q
the average comparison count is determined up to a linear term by the parameters fs,`
p
q
q
and fs,`
= n − 2 − fs,`
. Strategy I chooses these values such that fs,`
is either 0 or n − 2,
minimizing each term of the sum in Lemma 3.1—and thus minimizing the sum.
5.2. Guessing Whether s < ` or not
We explain how the idealized classification strategy just described can be approximated
by an implementation. The idea is to make a few comparisons and use the outcome as a
basis for a guess.
After p and q are chosen, classify the first ns = O(n1−ε ) many elements (the sample)
and calculate s0 and `0 , the number of small and large elements in the sample. If s0 < `0 ,
compare the remaining n − 2 − ns elements with q first, otherwise compare them with p
first. We say that the guess was correct if s0 < `0 and s < ` or s0 ≥ `0 and s ≥ `. In order
not to clutter up formulas, we will always assume that n1−ε is an integer. One would
otherwise work with dn1−ε e.
We incorporate guessing errors and sampling cost into (11) as follows:
1
4
E(Pn ) = n + n
3
2
X
Pr (guess correct | s, `) · min(s, `) +
Pr(guess wrong | s, `) · max(s, `) + O(n1−ε )
n/2 n−s
2 X X
4
= n + n
Pr(guess correct | s, `) · s +
3
2 s=0 `=s+1
Pr(guess wrong | s, `) · ` + O(n1−ε ),
s+`≤n−2
(12)
where the cost for comparing the elements in the sample is covered by the O(n1−ε ) term.
The following lemma says that for a wide range of values s and ` the probability of a
guessing error is exponentially small.
L EMMA 5.1. Let s and ` with s ≤ ` − n3/4 and ` ≥ n3/4 for n ∈ N be given. Let
ns = n2/3 . Then Pr(guess wrong | s, `) ≤ exp −n1/6 /18 .
P ROOF. Let Xi be a random variable that is −1 if the i-th P
classified element of the
ns
sample is large, 0 if it is medium, and 1 if it is small. Let d = i=1
Xi .
As in the proof of Lemma 3.1, we want to apply the method of averaged bounded
differences. Using the assumptions on the values of s and `, straightforward calculations
show that E(d) ≤ −ns /n1/4 = −n5/12 . Furthermore, we have that
ci = E(d | X1 , . . . , Xi ) − E(d | X1 , . . . , Xi−1 ) ≤ 3, for 1 ≤ i ≤ ns .
To see this, we let si and `i denote the number of small and large elements, respectively,
that are still present after the first i elements have been classified, i. e., X1 , . . . , Xi have
been determined. Let Yi be the 0-1 random variable that is 1 if and only if Xi is 1, and
13
let Zi be the 0-1 random variable that is 1 if and only if Xi is −1. We calculate:
|E(d | X1 , . . . , Xi ) − E(d | Xi , . . . , Xi−1 )|
ns
i
X
X
=
Xj +
Pr(Xj = 1 | X1 , . . . , Xi ) − Pr(Xj = −1 | X1 , . . . , Xi )
j=1
−
j=i+1
i−1
X
j=1
Xj −
ns
X
Pr(Xj = 1 | X1 , . . . , Xi−1 ) − Pr(Xj = −1 | X1 , . . . , Xi−1 )
j=i
X
ns
ns
X
si
`i
si−1
`i−1
= Xi +
−
−
−
n−i n−i
n−i+1 n−i+1
j=i+1
j=i
si−1 − Yi
`i−1 − Zi
si−1
`i−1
= Xi + (ns − i) ·
−
− (ns − i + 1) ·
−
n−i
n−i
n−i+1 n−i+1
(ns − i) · (Zi − Yi )
ns − i ns − i + 1
ns − i + 1 ns − i
= Xi +
+ si−1
−
+ `i−1
−
n−i
n−i
n−i+1
n−i+1 n−i
= Xi +
(ns − i) · (Zi − Yi )
n − ns
+ (`i−1 − si−1 ) ·
n−i
(n − i)(n − i + 1)
≤ |Xi | + |Zi − Yi | +
`i−1 − si−1
≤ 3.
n−i+1
The method of averaged bounded differences (see [Dubhashi and Panconesi 2009,
Theorem 5.3]) now implies that
!
t2
, for t > 0,
Pr(d > E(d) + t) ≤ exp − P
2 i≤ns c2i
which with t = n5/12 ≤ −E(d) yields
1/6
n
Pr(d > 0) ≤ exp −
.
18
Of course, we get an analogous result for s ≥ n3/4 and ` ≤ s − n3/4 .
Classification strategy SP works as follows: Classify the first ns = n2/3 elements. Let
0 0
s [` ] be the number of elements classified as being small [large]. If s0 > `0 then use p
for the first comparison for the remaining elements, otherwise use q.
We can now analyze the average number of key comparisons of this strategy turned
into a dual-pivot quicksort algorithm.
T HEOREM 5.2. The average comparison count of strategy SP turned into a dual-pivot
quicksort algorithm is 1.8n ln n + O(n).
P ROOF. We only have to analyze the expected classification cost. First, we classify
ns many elements. The number of key comparisons for these classifications is at most
2n2/3 = O(n1−ε ). By symmetry, we may focus on the case that s ≤ `. We distinguish the
following three cases:
14
(1) ` ≤ n3/4 : The contribution of terms in (12) satisfying this case is at most
3/4
n
`
2 XX
` = O(n1/4 ).
n
2
`=0 s=0
3/4
(2) ` − n
≤ s ≤ `: Let m1 (s, n) = min(s + n3/4 , n − s). The contribution of terms in (12)
satisfying this case is at most
n/2 m1 (s,n)
3/4
2 X
n
X
2
`=s
s=0
` = O(n3/4 ).
3/4
(3) ` ≥ n
and s ≤ ` − n . Let m2 (`, n) = min(n − `, ` − n3/4 ). Following Lemma 5.1,
the probability of guessing wrong is at most exp(−n1/6 /18). The contribution of this
case in (12) is hence at most
m2 (`,n)
m2 (`,n)
n
n
X
X
X
X
2
n
2
s + exp(−n1/6 /18)` = n
s + o(1) = + O(1).
n
6
2
2
3/4
3/4
s=0
s=0
`=n
`=n
Using these contributions in (12), we expect a partitioning step to make 32 n + O(n1−ε )
key comparisons. Applying Theorem 3.3, we get E(Cn ) = 1.8n ln n + O(n).
In this section we have seen an asymptotically optimal strategy. In the next section we
will present the optimal classification strategy. Unfortunately, it is even more unrealistic
than the idealized strategy I from above. However, we will give an implementation that
comes very close to the optimal strategy in terms of the number of comparisons.
6. THE OPTIMAL CLASSIFICATION STRATEGY AND ITS IMPLEMENTATION
We will consider two more strategies, an optimal (but not algorithmic) strategy, and an
algorithmic strategy that is optimal up to a very small error term.
We first study the (unrealistic!) setting where s and `, i. e., the number of small
resp. large elements, are known to the algorithm after the pivots are chosen, and the
classification tree can have different node labels for each such pair of values. Recall
that sv and `v , resp., denote the number of elements that have been classified as small
and large, resp., when at node v in the classification tree. We consider the following
strategy O: Given s and `, the comparison at node v is with the smaller pivot first if
s − sv > ` − `v , otherwise it is with the larger pivot first.6
T HEOREM 6.1. Strategy O is optimal, i. e., its ACT is at most as large as ACTT for
all classification trees T . When using O in a dual-pivot quicksort algorithm, we have
E(CnO ) = 1.8n ln n + O(n).
P ROOF. The proof of the first statement (optimality) is surprisingly simple. Fix the
two pivots, and consider equation (6). For each node v in the classification tree Strategy
O chooses the label that minimizes the contribution of this node to (6). So, it minimizes
each term of the sum, and thus minimizes the additional cost term in (5).
We now prove the second statement. For this, we first derive an upper bound of
1.8n ln n + O(n) for the average number of comparisons, and then show that this is tight.
For the first part, let an input with n entries and two pivots be given, so that there
are s small and ` large elements. Assume s ≥ `. Omit all medium elements to obtain a
reduced input (a1 , . . . , an0 ) with n0 = s + `. For 0 ≤ i ≤ n0 let si and `i denote the number
of small resp. large elements remaining in (ai+1 , . . . , an0 ). Let Di = si − `i . Of course we
6 This
strategy was suggested to us by Thomas Hotz (personal communication).
15
have D0 = s − ` and Dn0 = 0. Let i1 < i2 < · · · < ik be the list of indices i with Di = 0.
(In particular, ik = n0 .) Rounds i with Di = 0 are called zero-crossings. Consider some j
with Dij = Dij+1 = 0. The numbers Dij +1 , . . . , Dij+1 −1 are nonzero and have the same
positive [or negative] sign. The algorithm compares aij +2 , . . . , aij+1 with the smaller
[larger] pivot first, and aij +1 with the larger pivot first. Since {aij +1 , . . . , aij+1 } contains
the same number of small and large elements, the contribution
of this segment to the
additional comparison count is 12 (ij+1 − ij ) − 1 or 12 (ij+1 − ij ) .
If D0 > 0, i. e., s > `, all elements in {a1 , . . . , ai1 } are compared with the smaller
pivot first, and this set contains 12 (i1 − (s − `)) large elements (and 12 (i1 + (s − `)) small
elements), giving a contribution of 12 (i1 − (s − `)) to the additional comparison count.
Overall, the additional comparison count S2 + L2 (see the end of Section 2, in particular
(5)) of strategy O on the considered input is
k−1
n0 − (s − `)
i1 − (s − `) X ij+1 − ij
+
− k∗ =
− k∗ = ` − k∗ ,
2
2
2
j=1
(13)
for some correction term k ∗ ∈ {0, . . . , k − 1}.
Averaging the upper bound ` over all pivot choices, we obtain the following bound for
the additional cost term of strategy O:
X
X
1
E (S2 + L2 ) ≤ n · 2 ·
`+
` .
(14)
2
s+`≤n
`<s
`≤n/2
This gives an average number of at most 1.5n+O(1) comparisons. For such a partitioning
cost we can use (4) and obtain an average comparison count for sorting via strategy O
of at most 1.8n ln n + O(n).
It remains to show that this is tight. We shall see that the essential step in this
analysis is to show that the average (over all inputs) of the number of zero-crossings
(the number k from above, excepting the zero-crossing at position n) is O(log n). Again,
we temporarily omit medium elements to simplify calculations, i. e., we assume that the
number of small and large elements together is n0 . Fix a position n0 − 2i, for 1 ≤ i ≤ n0 /2.
If Dn0 −2i = 0, then there are as many small elements as there are large elements in
the last 2i positions of the input. Consequently, the input has to contain between i and
n0 − i small elements, otherwise no zero-crossing is possible. The probability that a
random input (excepting the two pivots) of n0 elements has exactly s small elements is
1/(n0 + 1), for 0 ≤ s ≤ n0 . Let Zn0 be the random variable that denotes the number of
zero-crossings for an input of n0 elements excepting the two pivots. We calculate:
E(Zn0 ) =
X
Pr(there is a zero-crossing at position n0 − 2i)
1≤i≤n0 /2
0
0
n /2 n −i
1 XX
Pr(Dn0 −2i = 0 | s small elements)
= 0
n + 1 i=1 s=i
n0 /2 n0 −i 2i
n0 /2 n0 /2 2i
n0 −2i
n0 −2i
1 X X i · s−i
2 X X i · s−i
= 0
≤ 0
,
n0
n0
n + 1 i=1 s=i
n + 1 i=1 s=i
s
s
where the last step follows by symmetry: replace s > n0 /2 by n0 − s.
16
√
By using the well-known estimate 2ii = Θ(22i / i) (which follows directly from
Stirling’s approximation), we continue by
E(Zn0 ) = Θ
=Θ
=Θ
1
n0
1
n0
1
n0
0
nX
/2
i=1
0
nX
/2
i=1
0
nX
/2
i=1
0
n /2
22i X
√
i s=i
n0 −2i
s−i
n0
s
0
n /2
22i X (n0 − s) · . . . · (n0 − s − i + 1) · s · . . . · (s − i + 1)
√
n0 · . . . · (n0 − 2i + 1)
i s=i
0
n /2−i i−1
X Y (n0 + 2j − 2k)(n0 − 2j − 2k)
n0 + 1
√
,
(n0 − 2k + 1)(n0 − 2k)
i(n0 −2i+1) j=0 k=0
(15)
where the last step follows by an index transformation using j = n0 /2−s and multiplying
22i into the terms of the rightmost fraction. We now obtain an upper bound for the
rightmost product:
i−1
Y
k=0
i−1
Y
(n0 + 2j − 2k)(n0 − 2j − 2k)
≤
(n0 − 2k + 1)(n0 − 2k)
1−
k=0
2j
n0 − 2k
2 !
≤
1−
2j
n0
2 !i
.
We substitute this bound into (15). Bounding the rightmost sum of (15) by an integral,
and using Maple R , we obtain
E(Zn0 ) = O
=O
1
n0
1
n0
0
nX
/2
0
√
i=1
n +1
i(n0 − 2i + 1)
n0 /2
E(Zn0 ) = O
X
i=1
n0 /2
1−
0
2t
n0
2 !i
dt + 1
Γ(i + 1)
n0 + 1
0
√
· n ·
+1 ,
Γ(i + 3/2)
i(n0 − 2i + 1)
1≤i≤n0 /2
X
involving the Gamma function Γ(x) =
we may continue by calculating
Z
R∞
0
√
tx−1 e−t dt. Since Γ(i+1)/Γ(i+3/2) = Θ(1/ i),
0
n /4
X1
n0 + 1
= O
+
i(n0 − 2i + 1)
i
i=1
n0 /2
X
i=n0 /4+1
1
= O(log n0 ).
n0 − 2i + 1
Now we generalize the analysis to the case that the input contains medium elements.
Let (a1 , . . . , an ) be a random input. Omit all medium elements to obtain a reduced
input (a1 , . . . , an0 ) with n0 = s + `. The additional cost term of strategy O on the reduced
input is the same as on the original input, since medium elements influence neither
the decisions of the strategy on elements of the reduced input nor the additional cost
term. Starting from (13), we may bound the difference between the additional cost term
E(S2 + L2 ) and the bound given in (14) by the average over all s, ` of the values k ∗ . This
17
O
σ
λ
σ
µ
C
σ
λ
σ
σ
σ
λ
σ
µ
σ
λ
−→
Next element
Next element
σ
σ
←−
Fig. 3. Visualization of the decision process when inspecting an element using strategy O (left) and C (right).
Applying strategy O from left to right uses that of the remaining elements three are small and one is large,
so it decides that the element should be compared with p first. Applying strategy C from right to left uses
that of the inspected elements two were small and only one was large, so it decides to compare the element
with p first, too. Note that the strategies would differ if, e. g., the rightmost element were a medium element.
is bounded by the average over all s, ` of the values Zn0 , hence by
X
X
X
1
1
`+
` − E(S2 + L2 ) ≤ n
E(Zs+` | s, `)
n 2
2
s+`≤n
`<s
2
`≤n/2
=
1
n
2
s+`≤n−2
X
O(log(s + `)) = O(log n).
s+`≤n−2
Since O(log n) is in O(n1−ε ), the influence of these O(log n) terms to the total average
sorting cost is bounded by O(n), see Theorem 3.3 and Appendix A.
While making an algorithm with minimum ACT possible, the assumption that the exact
number of small and large elements is known is of course not true for a real algorithm
or for a fixed tree. We can, however, identify a real, algorithmic partitioning strategy
whose ACT differs from the optimal one only by a logarithmic term. We study the
following strategy C: The comparison at node v is with the smaller pivot first if sv > lv ,
otherwise it is with the larger pivot first.
While O looks into the future (“Are there more small elements or more large elements
left?”), strategy C looks into the past (“Have I seen more small or more large elements
so far?”). It is not hard to see that for some inputs the number of additional comparisons
of strategy O and C can differ significantly. The next theorem shows that averaged over
all possible inputs, however, there is only a small difference.
T HEOREM 6.2. Let ACTO resp. ACTC be the ACT for classifying n elements using
strategy O resp. C. Then ACTC = ACTO +O(log n). When using C in a dual-pivot quicksort
algorithm, we get E(CnC ) = 1.8n ln n + O(n).
P ROOF. Assume that strategy O inspects the elements in the order an−1 , . . . , a2 ,
while C uses the order a2 , . . . , an−1 . If the strategies compare the element ai to different
pivots, then there are exactly as many small elements as there are large elements in
{a2 , . . . , ai−1 } or {a2 , . . . , ai }, depending on whether i is even or odd, see Figure 3.
The same calculation as the proof of the previous theorem shows that ACTC − ACTO
is O(log n), which—as mentioned in the proof of the previous theorem—sums up to
a total additive contribution of O(n) when using strategy C in a dual-pivot quicksort
algorithm.
Thus, dual-pivot quicksort with strategy C has average cost at most O(n) larger than
dual-pivot quicksort using the (unrealistic) optimal strategy O.
We have identified two asymptotically optimal strategies (SP and C) which can be
used in an actual algorithm. Note that both strategies have an additional summand of
18
O(n). Unfortunately, the solution of the dual-pivot quicksort recurrence, cf. Theorem 3.3,
does not give any information about the constant hidden in the O(n) term. However, the
average partitioning cost of strategy C differs by only O(log n) from the partitioning cost
of the optimal strategy O, while SP differs by O(n1−ε ). So one may be led to believe that
C has a (noticeably) lower average comparison count for real-world input lengths. In
Section 8 we will see that differences are clearly visible in the average comparison count
measured in experiments. However, we will also see that the necessity for bookkeeping
renders implementations of strategy C impractical.
7. CHOOSING PIVOTS FROM A SAMPLE
In this section, we consider the variation of quicksort where the pivots are chosen
from a small sample. Intuitively, this guarantees better pivots in the sense that the
partition sizes are more balanced. For classical quicksort, the Median-of-k strategy is
optimal w.r.t. minimizing the average comparison count [Martı́nez and Roura 2001],
which means that the median in a sample of k elements is chosen as the pivot. The
standard implementation of Yaroslavskiy’s algorithm in Sun’s Java 7 uses an intuitive
generalization of this strategy: it chooses the two tertiles in a sample of five elements
as pivots.7
We will compare dual-pivot quicksort algorithms which use the two tertiles of the
first five elements of the input as the two pivots with classical quicksort. Moreover, we
will see that the optimal pivot choices for dual-pivot quicksort are not the two tertiles of
a sample, but rather the elements of rank k/4 and k/2.
We remark that Wild, Nebel, and Martı́nez [Wild et al. 2014] provide a much more
detailed study of pivot sampling in Yaroslavskiy’s algorithm.
7.1. Choosing the Two Tertiles in a Sample of Size 5 as Pivots
We sort the first five elements and take the second and fourth elements aspivots. The
probability that p and q, p < q, are chosen as pivots is exactly (s · m · `)/ n5 . Following
Hennequin [Hennequin 1991, pp. 52–53], for average partitioning cost E(Pn ) = a · n +
O(1) we get
E(Cn ) =
20
1
· a · n ln n + O(n) =
· a · n ln n + O(n),
H6 − H2
19
(16)
where Hn denotes the n-th harmonic number.
When applying Lemma 3.1, we have average partitioning cost a · n + O(n1−ε ). Using
the same argument as in the proof of Theorem 3.3, the average comparison count
becomes 20/19 · a · n ln n + O(n) in this case. (This a special case of the much more
general pivot sampling strategy that is described and analyzed in [Wild et al. 2014,
Theorem 6.2].)
We will now investigate the effect on the average number of key comparisons in
Yaroslavskiy’s algorithm and the asymptotically optimal strategy SP from Section 5.
The average number of medium elements remains (n−2)/3. For strategy Y, we calculate
(again using Maple R )
X ` · (2s + m) · s · m · `
4
1
34
+ O(n1−ε ) =
n + O(n1−ε ).
E(PnY ) = n + n
3
n
−
5
21
5
s+`≤n−5
Applying (16), we get E(CnY ) = 1.704..n ln n + O(n) key comparisons on average. (Note
that [Wild et al. 2013] calculated this leading coefficient as well.) This is slightly better
than “clever quicksort”, which uses the median of a sample of three elements as a single
7
In an ordered set S = {x1 , . . . , xk }, the two tertiles are the elements of rank dk/3e and d2k/3e.
19
pivot element and achieves 1.714..n ln n + O(n) key comparisons on average [van Emden
1970]. For strategy SP, we get (similarly as in Section 5)
X
37
2
4
s · s · m · ` + O(n1−ε ) =
n + O(n1−ε ).
E(PnSP ) = n + n
3
24
5
s+`≤n−5
s≤`
Again using (16), we obtain E(CnSP ) = 1.623..n ln n + O(n), improving further on the
leading coefficient compared to clever quicksort and Yaroslavskiy’s algorithm.
7.2. Pivot Sampling in Classical Quicksort and Dual-Pivot Quicksort
In the previous subsection, we have shown that optimal dual-pivot quicksort using a
sample of size 5 clearly beats clever quicksort which uses the median of three elements.
We will now investigate how these two variants compare when the sample size grows.
The following proposition, which is a special case of [Hennequin 1991, Proposition
III.9 and Proposition III.10], will help in this discussion.
P ROPOSITION 7.1. Let a · n + O(1) be the average partitioning cost of a quicksort
algorithm A that chooses the pivot(s) from a sample of size k, for constants a and k. Then
the following holds:
(1) If k + 1 is even and A is a classical quicksort variant that chooses the median of
these k samples, then the average sorting cost is
1
· a · n ln n + O(n).
Hk+1 − H(k+1)/2
(2) If k + 1 is divisible by 3 and A is a dual-pivot quicksort variant that chooses the two
tertiles of these k samples as pivots, then the average sorting cost is
1
· a · n ln n + O(n).
Hk+1 − H(k+1)/3
Note that for classical quicksort we have partitioning cost n − 1. Thus, the average
1
sorting cost becomes Hk+1 −H
n ln n + O(n).
(k+1)/2
For dual-pivot partitioning algorithms, the probability that p and q, p < q, are chosen
as pivots in a sample of size k where k + 1 is divisible by 3 is exactly
q−p−1 n−q
p−1
(k−2)/3
(k−2)/3
n
k
(k−2)/3
.
SP
Thus, the average partitioning cost E(Pn,k
) of strategy SP using a sample of size k can
be calculated as follows:
4
2 X
s
m
`
SP
E(Pn,k ) = n + n
· s + O(n1−ε ).
(17)
3
(k
−
2)/3
(k
−
2)/3
(k
−
2)/3
k
s≤`
SP
Unfortunately, we could not find a closed form of E(Pn,k
). Some calculated values in
which classical and dual-pivot quicksort with strategy SP use the same sample size can
be found in Table I. These values clearly indicate that starting from a sample of size 5
classical quicksort has a smaller average comparison count than dual-pivot quicksort.8
This raises the question whether dual-pivot quicksort is inferior to classical quicksort
using the Median-of-k strategy.
8 Note
that this statement does not include that these two variants have different linear terms for the same
sample size.
20
Table I. Comparison of the leading term of the average cost of classical quicksort
and dual-pivot quicksort for specific sample sizes. Note that for real-world input sizes,
however, the linear term can make a big difference.
Sample Size
Median (QS)
Tertiles (DP QS)
5
1.622..n ln n
1.623..n ln n
11
1.531..n ln n
1.545..n ln n
17
1.501..n ln n
1.523..n ln n
41
1.468..n ln n
1.504..n ln n
7.3. Optimal Segment Sizes for Dual-Pivot Quicksort
It is known from, e.g., [Martı́nez and Roura 2001] that for classical quicksort in which
the pivot is chosen as the median of a fixed-sized sample, the leading term of the
average comparison count converges with increasing sample size to the lower bound
of (1/ ln 2) · n ln n = 1.4426..n ln n. Wild, Nebel, and Martı́nez observed in [Wild et al.
2014] that this is not the case for Yaroslavskiy’s algorithm, which makes at least
1.4931..n ln n − O(n) comparisons on average no matter how well the pivots are chosen.
In this section, we will show how to match the lower bound for comparison-based sorting
algorithms with a dual-pivot approach.
We study the following setting, which was considered in [Martı́nez and Roura 2001;
Wild et al. 2014] as well. We assume that for a random input of n elements9 we can
choose (for free) two pivots w.r.t. a vector ~τ = (τ1 , τ2 , τ3 ) such that the input contains
exactly τ1 n small elements, τ2 n medium elements, and τ3 n large elements. Furthermore,
we consider the (simple) classification strategy L: “Always compare with the larger pivot
first.”
The following lemma says that this strategy achieves the minimum possible average
comparison count for comparison-based sorting algorithms, 1.4426..n ln n, when setting
τ = ( 41 , 14 , 12 ).
P
L EMMA 7.2. Let ~τ = (τ1 , τ2 , τ3 ) with 0 < τi < 1 and i τi = 1, for i ∈ {1, 2, 3}. Assume
that for each input size n we can choose two pivots such that there are exactly τ1 · n small,
τ2 · n medium, and τ3 · n large elements. Then the comparison count of strategy L is10
p~τ (n) ∼
1 + τ + τ2
P 1
n ln n.
− 1≤i≤3 τi ln τi
This value is minimized for ~τ ∗ = (1/4, 1/4, 1/2) giving
∗
1
n ln n = 1.4426..n ln n.
p~τ (n) ∼
ln 2
P ROOF. On an input consisting of n elements, strategy L makes n + (τ1 + τ2 )n
comparisons. Thus, the comparison count of strategy L follows the recurrence
p~τ (n) = n + (τ1 + τ2 )n + p~τ (τ1 · n) + p~τ (τ2 · n) + p~τ (τ3 · n).
Using the Discrete Master Theorem from [Roura 2001, Theorem 2.3, Case (2.1)], we
obtain the following solution for this recurrence:
p~τ (n) ∼
1 + τ1 + τ2
n ln n.
P3
i=1 τi ln τi
−
∗
Using Maple R , one finds that p~τ is minimized for ~τ ∗ = ( 14 , 14 , 12 ), giving p~τ (n) ∼
1.4426..n ln n.
9 We
disregard the two pivots in the following discussion.
f (n) ∼ g(n) means that limn→∞ f (n)/g(n) = 1.
10 Here,
21
Table II. Overview of the algorithms considered in the experiments.
Abbreviation
QS
Y
L
S
SP
C
K
Full Name
Classical Quicksort
Yaroslavskiy’s Algorithm
Larger Pivot First
Sedgewick’s Algorithm (modified)
Sample Algorithm
Counting Algorithm
3-Pivot-Algorithm
Strategy
—
Section 4
Section 4
Section 4
Section 5
Section 6
—
Pseudocode
e.g., [Wild and Nebel 2012, Algo. 1]
Algorithm 2 (Page 29)
Algorithm 3 (Page 29)
Algorithm 5 (Page 31)
Algorithm 6 (Page 32)
Algorithm 7 (Page 32)
[Kushagra et al. 2014, Algo. A.1.1]
The reason why strategy L with this particular choice of pivots achieves the lower
bound is simple: it makes (almost) the same comparisons as does classical quicksort
using the median of the input as pivot. On an input of length n, strategy L makes 3/2n
key comparisons and then makes three recursive calls to inputs of length n/4, n/4, n/2.
On an input of length n, classical quicksort using the median of the input as the pivot
makes n comparisons to split the input into two subarrays of length n/2. Now consider
only the recursive call on the left subarray. After n/2 comparisons, the input is split into
two subarrays of size n/4 each. Now there remain two recursive calls on two subarrays
of size n/4, and one recursive call on a subarray of size n/2 (the right subarray of the
original input), like in strategy L. Since classical quicksort using the median of the
input clearly makes n log n key comparisons, this bound must also hold for strategy L.
8. EXPERIMENTS
We have implemented the methods presented in this paper in C++ and Java. Our
experiments were carried out on an Intel i7-2600 at 3.4 GHz with 16 GB RAM running
Ubuntu 13.10 with kernel version 3.11.0. For compiling C++ code, we used gcc in version
4.8. For compiling Java code, we used Oracle’s Java 8 SDK. All Java runtime tests were
preceeded by a warmup phase for the just-in-time compiler (JIT), in which we let each
algorithm sort 10 000 inputs of length 100 000.
For better orientation, the algorithms considered in this section are presented in
Table II. Pseudocode for the dual-pivot methods is provided in Appendix B. In the
following, we use a calligraphic letter both for the classification strategy and the actual
dual-pivot quicksort algorithm.
In Section 8.1, we experimentally evaluate the average comparison count of the
algorithms considered here. In Section 8.2, we focus on the actual running time needed
to sort a given input. The charts of our experiments can be found at the end of this
paper.
8.1. The Average Comparison Count
We first have a look at the number of comparisons needed to sort a random input of up
to 229 integers. We did not switch to a different sorting algorithm, e. g., insertion sort, to
sort short subarrays.
Figure 5 shows the results of our experiments for algorithms that choose the pivots
directly from the input. We see that for practical values of n the lower order terms in
the average comparison count have a big influence on the number of comparisons for
all algorithms. (E.g., for Yaroslavskiy’s algorithm this lower order term is dominated by
the linear term −2.46n, as known from [Wild and Nebel 2012].) Nevertheless, we may
conclude that the theoretical studies on the average comparison count correspond to
what can be observed in practice. Note that there is a difference between the optimal
strategies SP and C. We also see that the modified version of Sedgewick’s algorithm
beats Yaroslavskiy’s algorithm, as calculated in Section 4.
Figure 6 shows the same experiment for algorithms that choose the pivots from
a small sample. For dual-pivot quicksort variants, we used two different versions.
22
Algorithms Y, SP, C, L choose the tertiles of a sample of size 5 as pivots. QS is classical
quicksort using the median of three strategy. Y 0 is Yaroslavskiy’s algorithm with the
tertiles of a sample of size 11 as the pivots; L0 is algorithm L using the third- and sixthlargest element in a sample of size 11 as the pivots. This plot confirms the theoretical
results from Section 7. Especially, the simple strategy L beats Yaroslavskiy’s algorithm
for a sample of size 11 for the pivot choices introduced in Section 7.3.
8.2. Running Times
We now consider the running times of the algorithms for sorting a given input. We
restrict our experiments to sorting random permutations of the integers {1, . . . , n}. It
remains future work to compare the algorithms in more detail.
As a basis for comparison with other methods, we also include the three pivot quicksort algorithm described in [Kushagra et al. 2014] (about 8% faster than Yaroslavskiy’s
algorithm in their setup), which we call K. Similarly to classical quicksort, this algorithm uses two pointers to scan the input from left-to-right and right-to-left until these
pointers cross. Classifications are done in a symmetrical way: First compare to the
middle pivot; then (appropriately) compare either to the largest or smallest pivot. (In
this way, each element is compared to exactly two pivots.) Elements are moved to a
suitable position by the help of two auxiliary pointers. We note that the pseudocode
from [Kushagra et al. 2014, Algorithm A.1.1] uses multiple swaps when fewer assignments are sufficient to move elements. We show pseudocode of our implementation in
Appendix C.
In our experiments, subarrays of size at most 16, 20, and 23 were sorted by insertion
sort for classical quicksort, dual-pivot quicksort, and the three pivot quicksort algorithm,
respectively. Furthermore, strategy SP uses strategy L to sort inputs that contain no
more than 1024 elements.
C++ Experiments. We first discuss our results on C++ code compiled with gcc-4.8.
Since the used compiler flags might influence the observed running times, we considered
4 different compiler settings. In setting 1 we have compiled the source code with -O2, in
setting 2 with -O2 -funroll-loops, in setting 3 with -O3, and in setting 4 with -O3 -funrollloops. The option -funroll-loops tells the compiler to unroll (the first few iterations of)
a loop. In all settings, we used -march=native, which means that the compiler tries to
optimize for the specific CPU architecture we use during compilation.
The experiments showed that there is no significant difference between the settings
using -O2 and those using -O3. So, we just focus on the first two compiler settings
which use -O2. The running time results we obtained in these two settings are shown
in Figure 7 and Figure 8 at the end of this paper. We first note that loop unrolling
makes all algorithms behave slightly worse with respect to running times. Since a
single innocuous-looking compiler flag may have such an impact on running time, we
stress that our results do not allow final statements about the running time behavior of
quicksort variants. In the following, we restrict our evaluation to setting 1.
With respect to average running time, we get the following results, see Figure 7
(the discussion uses the measurements obtained for inputs of size 227 ): Yaroslavskiy’s
algorithm is the fastest algorithm, but the difference to the dual-pivot algorithm L
and the three pivot algorithm K is negligible. The asymptotically comparison-optimal
sampling algorithm SP cannot compete with these three algorithms w.r.t. running time.
On average it is about 7.2% slower than algorithm Y. Classical quicksort is about 8.3%
slower than Y. The slowest algorithm is the counting algorithm C, on average it is about
14.8% slower than Y.
Now we consider the significance of running time differences. For that, we let each
algorithm sort the same 1 000 inputs containing 227 items. In Table III we consider the
23
number of cases which support the hypothesis that an algorithm is a given percentage
faster than another algorithm. The table shows that the difference in running time is
about 1% smaller than the average suggested if we consider only significant running
time differences, i. e., differences that were observed in at least 95% of the inputs. We
conclude that there is no significant difference between the running times of the three
fastest quicksort variants.
Table III. Comparison of the actual running times of the algorithms on 1000 different inputs of size 227 . A table cell
in row labelled “A” and column labelled “B” contains a string “x%/y%/z%” and is read as follows: “In about 95%,
50%, and 5% of the cases algorithm A was more than x, y, and z percent faster than algorithm B, respectively.”
Y
L
K
SP
QS
Y
—
–/–/0.8%
—
—
—
L
–/–/0.9%
—
–/–/0.3%
—
—
K
–/1.0%/2.1%
–/0.9%/2.1%
—
—
—
SP
6.4%/7.2%/8.0%
5.8%/7.0%/8.5%
4.7%/6.0%/7.6%
—
—
QS
7.2%/8.3%/9.4%
7.0%/8.1%/9.5%
5.9%/7.2%/8.5%
–/1.0%/2.4%
—
C
13.9%/14.8%/15.8%
13.7%/14.6%/15.8%
12.3%/13.6%/14.9%
5.8%/7.1%/8.4%
4.7%/5.9%/7.2%
The good performance of the simple strategy “always compare to the larger pivot
first” is especially interesting, because it is bad from a theoretical point of view: it
makes 2n ln n comparisons and 0.6n ln n swaps on average. So, it does not improve in
both of these cost measures compared to classical quicksort, but is still faster. We may
try to explain these differences in measured running times by looking at the average
instruction count and the cache behavior of these algorithms. (The latter is motivated
by the cost measure “cache misses” considered in [Kushagra et al. 2014].) Table IV
shows measurements of the average total instruction count and the average number of
L1/L2 cache misses,11 both in total and in relation to algorithm K.
With respect to the average number of total instructions, we see that strategy K
needs the fewest instructions on average. It is about 1.2% better than algorithm L.
Furthermore, it is 7.1% better than the sampling algorithm SP, about 11% better than
Yaroslavskiy’s algorithm and about 12.7% better than classical quicksort. As expected,
the theoretically optimal strategy C needs by far the most instructions because of its
necessity for bookkeeping. However, focusing exclusively on the total instruction count
does not predict the observed running time behavior accurately. (In particular, Y should
be slower than classical quicksort, which has also been observed in [Wild et al. 2013]
with respect to Sun’s Java 7.)
The authors of [Kushagra et al. 2014] conjectured that another cost measure—the
average number of cache misses—explains observed running time behavior. From Table IV we see that strategy K shows the best performance with respect to cache misses,
confirming the theoretical results of [Kushagra et al. 2014]. In our experiments, we
observed that the relative difference of L1 cache misses between the 1-, 2-, and 3-pivot
algorithms match the theoretical results from [Kushagra et al. 2014]. (According to
their calculations, dual-pivot variants should incur about 16% more cache misses, while
classical quicksort should incur about 44% more cache misses.) This might explain why
classical quicksort is the slowest algorithm, but cannot explain why the dual-pivot
algorithms Y and L can compete with the three pivot algorithm.
With respect to L2 cache misses, we see that the cache misses are much higher
than predicted. Compared to the 3-pivot algorithm, classical quicksort incurs 178%
more cache misses, and the worst dual-pivot algorithm w.r.t. L2 misses has 49% more
cache misses. Since L2 misses have a much higher impact on the CPU cycles spent
waiting for memory, this might amplify differences between these algorithms. In general,
11 Such
statistics can be collected, e.g., by using the performance application programming interface from
http://icl.cs.utk.edu/papi/.
24
Table IV. Average number of total instructions and cache misses scaled by n ln n for the
algorithms considered in this section. In parentheses, these figures are set into relation to
the values measured for algorithm K. (The relative difference has been calculated directly
from the experimental data.) The figures were obtained by sorting 1 000 inputs of length 227
separately by each algorithm using the compiler flag -O2.
Algorithm
QS
Y
SP
C
L
K
total #instructions
10.58n ln n (+12.7%)
10.42n ln n (+11.0%)
10.05n ln n (+ 7.1%)
14.08n ln n (+50.1%)
9.50n ln n (+ 1.2%)
9.38n ln n ( — )
L1 misses
0.142n ln n (+47.9%)
0.111n ln n (+15.6%)
0.111n ln n (+15.6%)
0.111n ln n (+15.6%)
0.111n ln n (+15.6%)
0.096n ln n ( — )
L2 misses
0.030n ln n (+178.1%)
0.015n ln n (+ 39.9%)
0.014n ln n (+ 31.7%)
0.012n ln n (+ 11.8%)
0.016n ln n (+ 49.0%)
0.011n ln n (
— )
we note that only a fraction of about 1% of the instructions lead to cache misses.
Consequently, one has to accurately measure the cycles spent waiting for memory
in these situations. Investigating these issues is a possible direction for future work.
With respect to other cost measures, Martı́nez, Nebel, and Wild [Martı́nez et al. 2015]
analyzed branch misses in classical quicksort and Yaroslavskiy’s algorithm. Their result
is that the difference in the average number of branch mispredictions between these
two algorithms is too small to explain differences in empirical running time. Based
on our experimental measurements, we believe that this cost measure cannot explain
running time differences between dual-pivot quicksort algorithms, either.
We conclude that in the light of our experiments neither the average number of
instructions nor the average number of cache misses can fully explain empirical running
time.
Java Experiments. Figure 9 shows the running time measurements we got using
Oracle’s Java 8. Again, there is no significant difference between algorithms K, L, and
Y. The three pivot quicksort algorithm K is the fastest algorithm using Java 8 on our
setup. It is about 1.3% faster than Yaroslavskiy’s algorithm on average. The optimal
strategy C is about 37% slower on average than algorithm K. Classical quicksort and
the sampling strategy SP are about 7.7% slower on average.
9. CONCLUSION AND OPEN QUESTIONS
We have studied dual-pivot quicksort algorithms in a unified way and found optimal
partitioning methods that minimize the average number of key comparisons up to O(n).
This minimum is 1.8n ln n + O(n). We showed an optimal pivot choice for a simple dualpivot quicksort algorithm and conducted an experimental study of the practicability of
dual-pivot quicksort algorithms.
Several open questions remain. From a theoretical point of view, one should generalize
our algorithms to the case that three or more pivots are used. At the moment, the
optimal average comparison count for k-pivot quicksort is unknown for k ≥ 3. This is
the subject of ongoing work. From both a theoretical and an engineering point of view,
the most important question is to find a cost measure that accurately predicts empirical
running time behavior of variants of quicksort algorithms. Here we could only provide
evidence that neither standard measures like the average comparison count or the
average swap count, nor empirical measures like the average instruction count or the
cache behavior predict running time correctly when considered in isolation.
Acknowledgements. The authors thank Thomas Hotz for a useful suggestion regarding the optimal strategy of Section 6, and Pascal Klaue for implementing the algorithms
and carrying out initial experiments. The first author thanks Timo Bingmann, Michael
Rink, and Sebastian Wild for interesting and useful discussions. We thank the referees
of the conference and journal submissions for their insightful comments, which helped
a lot in improving the presentation. We especially thank the referee who pointed out a
25
way of showing that the lower order term of most algorithms considered in this paper is
O(n), not only o(n ln n).
REFERENCES
Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. 2009. Introduction to
Algorithms (3rd ed.). MIT Press.
Devdatt P. Dubhashi and Alessandro Panconesi. 2009. Concentration of Measure for the Analysis of Randomized Algorithms. Cambridge University Press.
Pascal Hennequin. 1991. Analyse en moyenne d’algorithmes: tri rapide et arbres de recherche. Ph.D. Dissertation. Ecole Politechnique, Palaiseau. http://www-lor.int-evry.fr/∼pascal/
C. A. R. Hoare. 1962. Quicksort. Comput. J. 5, 1 (1962), 10–15.
Donald E. Knuth. 1973. The Art of Computer Programming, Volume III: Sorting and Searching. AddisonWesley.
Shrinu Kushagra, Alejandro López-Ortiz, Aurick Qiao, and J. Ian Munro. 2014. Multi-Pivot Quicksort: Theory
and Experiments. In Proc. of the 16th Meeting on Algorithms Engineering and Experiments (ALENEX’14).
SIAM, 47–60.
Conrado Martı́nez, Markus E. Nebel, and Sebastian Wild. 2015. Analysis of Branch Misses in Quicksort. In
Proc. of the 12th Meeting on Analytic Algorithmics and Combinatorics (ANALCO’ 15). 114–128.
Conrado Martı́nez and Salvador Roura. 2001. Optimal sampling strategies in quicksort and quickselect.
SIAM J. Comput. 31, 3 (2001), 683–705.
Salvador Roura. 2001. Improved master theorems for divide-and-conquer recurrences. J. ACM 48, 2 (2001),
170–205.
Robert Sedgewick. 1975. Quicksort. Ph.D. Dissertation. Standford University.
Robert Sedgewick. 1977. Quicksort with equal keys. SIAM J. Comput. 6, 2 (1977), 240–268.
Robert Sedgewick and Philippe Flajolet. 1996. An Introduction to the Analysis of Algorithms. Addison-WesleyLongman.
M. H. van Emden. 1970. Increasing the efficiency of quicksort. Commun. ACM 13, 9 (1970), 563–567.
Sebastian Wild. 2013. Java 7’s Dual Pivot Quicksort. Master’s thesis. University of Kaiserslautern.
Sebastian Wild and Markus E. Nebel. 2012. Average case analysis of Java 7’s dual pivot quicksort. In Proc. of
the 20th European Symposium on Algorithms (ESA’12). 825–836.
Sebastian Wild, Markus E. Nebel, and Conrado Martı́nez. 2014. Analysis of Pivot Sampling in Dual-Pivot
Quicksort. CoRR abs/1412.0193 (2014).
Sebastian Wild, Markus E. Nebel, and Ralph Neininger. 2015. Average Case and Distributional Analysis of
Dual-Pivot Quicksort. ACM Transactions on Algorithms 11, 3 (2015), 22.
Sebastian Wild, Markus E. Nebel, Raphael Reitzig, and Ulrich Laube. 2013. Engineering Java 7’s Dual Pivot
Quicksort Using MaLiJAn. In Proc. of the 15th Meeting on Algorithms Engineering and Experiments
(ALENEX’13). 55–69.
A. MISSING DETAILS OF THE PROOF OF THEOREM ??
Here we solve recurrence (3) for partitioning cost E(Pn ) at most K · n1−ε . We use the
Continuous Master Theorem of Roura [Roura 2001], whose statement we will review
first.
T HEOREM A.1 ([R OURA 2001, T HEOREM 18]). Let Fn be recursively defined by
(
bn ,
for 0 ≤ n < N,
Pn−1
Fn =
tn + j=0 wn,j Fj , for n ≥ N,
where the toll function tn satisfies tn ∼ Knα logβ (n) as n → ∞ for constants K 6= 0, α ≥
0, β > −1. Assume there exists a function w : [0, 1] → R such that
Z (j+1)/n
n−1
X
wn,j −
w(z) dz = O(n−d ),
(18)
j/n
j=0
for a constant d > 0. Let H := 1 −
R1
0
z α w(z) dz. Then we have the following cases:
26
(1) If H > 0, then Fn ∼ tn /H.
(2) If H = 0, then Fn ∼ (tn ln n)/Ĥ, where
1
Z
z α ln(z)w(z) dz.
Ĥ := −(β + 1)
0
(3) If H < 0, then Fn ∼ Θ(nc ) for the unique c ∈ R with
Z 1
z c w(z) dz = 1.
0
We now solve recurrence (3) for tn = K · n1−ε , for some ε > 0. First, observe that
recurrence (3) has weights
wn,j =
6(n − j − 1)
.
n(n − 1)
We define the shape function w(z) as suggested in [Roura 2001] by
w(z) = lim n · wn,zn = 6(1 − z).
n→∞
Now we have to check (18) to see whether the shape function is suitable. We calculate:
Z (j+1)/n
n−1
X
wn,j −
w(z) dz
j/n
j=0
=6
n−1
X
j=0
=6
n−1
X
j=0
<6
n−1
X
j=0
n−j−1
−
n(n − 1)
Z
(j+1)/n
(1 − z) dz
j/n
n − j − 1 2j + 1
1
+
−
n(n − 1)
2n2
n
1
= O(1/n).
2n(n − 1)
Thus, w is a suitable shape function. By observing that
Z 1
H := 1 − 6
z 1−ε (1 − z) dz < 0,
0
we conclude that the third case of Theorem A.1 applies for our recurrence. Consequently,
we have to find the unique c ∈ R such that
Z 1
6
z c (1 − z) dz = 1,
0
which is true for c = 1. Thus, an average partitioning cost of at most K · n1−ε yields
average sorting cost of O(n).
B. DUAL-PIVOT QUICKSORT: ALGORITHMS IN DETAIL
B.1. General Setup
The general outline of a dual-pivot quicksort algorithm is presented as Algorithm 1.
To get an actual algorithm we have to implement a partition function that partitions
the input as depicted in Figure 1. A partition procedure in this paper has two output
27
Algorithm 1 Dual-Pivot-Quicksort (outline)
procedure Dual-Pivot-Quicksort(A, left, right)
1 if right − left ≤ THRESHOLD then
2
InsertionSort(A, left, right);
3
return ;
4 if A[right] > A[left] then
5
swap A[left] and A[right];
6 p ← A[left];
7 q ← A[right];
8 partition(A, p, q, left, right, posp , posq );
9 Dual-Pivot-Quicksort(A, left, posp - 1);
10 Dual-Pivot-Quicksort(A, posp + 1, posq - 1);
11 Dual-Pivot-Quicksort(A, posq + 1, right);
variables posp and posq that are used to return the positions of the two pivots in the
partitioned array.
For moving elements around, we make use of the following two operations.
procedure rotate3(a, b, c)
1 tmp ← a;
2 a ← b;
3 b ← c;
4 c ← tmp;
procedure rotate4(a, b, c, d)
1 tmp ← a;
2 a ← b;
3 b ← c;
4 c ← d;
5 d ← tmp;
B.2. Yaroslavskiy’s Partitioning Method
As mentioned in Section 4, Yaroslavskiy’s algorithm makes sure that for ` large elements
in the input ` or ` − 1 elements will be compared to the larger pivot first. How does it
accomplish this? By default, it compares to the smaller pivot first, but for each large
elements that it sees, it will compare the next element to the larger pivot first.
Algorithm 2 shows the partition step of (a slightly modified version of) Yaroslavskiy’s
algorithm. In contrast to the algorithm studied in [Wild and Nebel 2012], it saves
an index check at Line 8, and uses a rotate3 operation to save assignments. (In our
experiments this makes Yaroslavskiy’s algorithm about 4% faster.)
B.3. Algorithm Using “Always Compare to the Larger Pivot First”
Algorithm 3 presents an implementation of strategy L (“Always compare to the larger
pivot first”). Like Yaroslavskiy’s algorithm, it uses three pointers into the array. One
pointer is used to scan the array from left to right until a large element has been
found (moving small elements to a correct position using the second pointer on the
way). Subsequently, it scans the array from right to left using the third pointer until a
non-large element has been found. These two elements are then placed into a correct
position. This is repeated until the two pointers have crossed. The design goal is to check
as rarely as possible if these two pointers have met, since this event occurs infrequently.
(In contrast, Yaroslavskiy’s algorithm checks this for each element scanned by index k
in Algorithm 2.)
This strategy makes 2n ln n comparisons and 1.6n ln n assignments on average.
28
Algorithm 2 Yaroslavskiy’s Partitioning Method
procedure Y-Partition(A, p, q, left, right, posp , posq )
1 l ← left + 1; g ← right − 1; k ← l;
2 while k ≤ g do
3
if A[k] < p then
4
swap A[k] and A[l];
5
l ← l + 1;
6
else
7
if A[k] > q then
8
while A[g] > q do
9
g ← g − 1;
10
if k < g then
11
if A[g] < p then
12
rotate3(A[g], A[k], A[l]);
13
l ← l + 1;
14
else
15
swap A[k] and A[g];
16
g ← g − 1;
17
k ← k + 1;
18 swap A[left] and A[l − 1];
19 swap A[right] and A[g + 1];
20 posp ← l − 1; posq ← g + 1;
Algorithm 3 Always Compare To Larger Pivot First Partitioning
procedure L-Partition(A, p, q, left, right, posp , posq )
1 i ← left + 1; k ← right − 1; j ← i;
2 while j ≤ k do
3
while q < A[k] do
4
k ← k − 1;
5
while A[j] < q do
6
if A[j] < p then
7
swap A[i] and A[j];
8
i ← i + 1;
9
j ← j + 1;
10
if j < k then
11
if A[k] > p then
12
rotate3(A[k], A[j], A[i]);
13
i ← i + 1;
14
else
15
swap A[j] and A[k];
16
k ← k − 1;
17
j ← j + 1;
18 swap A[left] and A[i − 1];
19 swap A[right] and A[k + 1];
20 posp ← i − 1; posq ← k + 1;
29
Fig. 4. An intermediate partitioning step in Sedgewick’s algorithm.
B.4. Partitioning Methods Based on Sedgewick’s Algorithm
Algorithm 4 shows Sedgewick’s partitioning method as studied in [Sedgewick 1975].
Sedgewick’s partitioning method uses two pointers i and j to scan through the input.
It does not swap entries in the strict sense, but rather has two “holes” at positions i1
resp. j1 that can be filled with small resp. large elements. “Moving a hole” is not a
swap operation in the strict sense (three elements are involved), but requires the same
amount of work as a swap operation (in which we have to save the content of a variable
into a temporary variable [Sedgewick 1975]). An intermediate step in the partitioning
algorithm is depicted in Figure 4.
The algorithm works as follows: Using i it scans the input from left to right until it
has found a large element, always comparing to the larger pivot first. Small elements
found in this way are moved to a correct final position using the hole at array position
i1 . Subsequently, using j it scans the input from right to left until it has found a small
element, always comparing to the smaller pivot first. Large elements found in this
way are moved to a correct final position using the hole at array position j1 . Now it
exchanges the two elements at positions i resp. j and continues until i and j have met.
Algorithm 4 Sedgewick’s Partitioning Method
procedure S-Partition(A, p, q, left, right, posp , posq )
1 i ← i1 ← left; j ← j1 := right;
2 while true do
3
i ← i + 1;
4
while A[i] ≤ q do
5
if i ≥ j then
6
break outer while
7
if A[i] < p then
8
A[i1 ] ← A[i]; i1 ← i1 + 1; A[i] ← A[i1 ];
9
i ← i + 1;
10
j ← j − 1;
11
while A[j] ≥ p do
12
if A[j] > q then
13
A[j1 ] ← A[j]; j1 ← j1 − 1; A[j] ← A[j1 ];
14
if i ≥ j then
15
break outer while
16
j ← j − 1;
17
A[i1 ] ← A[j]; A[j1 ] ← A[i];
18
i1 ← i1 + 1; j1 ← j1 − 1;
19
A[i] ← A[i1 ]; A[j] ← A[j1 ];
20 A[i1 ] ← p; A[j1 ] ← q;
21 posp ← i1 ; posq ← j1 ;
Algorithm 5 shows an implementation of the modified partitioning strategy from
Section 4. In the same way as Algorithm 4 it scans the input from left to right until it
has found a large element. However, it uses the smaller pivot for the first comparison in
30
this part. Subsequently, it scans the input from right to left until it has found a small
element. Here, it uses the larger pivot for the first comparison.
Algorithm 5 Sedgewick’s Partitioning Method, modified
procedure S2-Partition(A, p, q, left, right, posp , posq )
1 i ← i1 ← left; j ← j1 := right;
2 while true do
3
i ← i + 1;
4
while true do
5
if i ≥ j then
6
break outer while
7
if A[i] < p then
8
A[i1 ] ← A[i]; i1 ← i1 + 1; A[i] ← A[i1 ];
9
else if A[i] > q then
10
break inner while
11
i ← i + 1;
12
j ← j − 1;
13
while true do
14
if A[j] > q then
15
A[j1 ] ← A[j]; j1 ← j1 − 1; A[j] ← A[j1 ];
16
else if A[j] < p then
17
break inner while
18
if i ≥ j then
19
break outer while
20
j ← j − 1;
21
A[i1 ] ← A[j]; A[j1 ] ← A[i];
22
i1 ← i1 + 1; j1 ← j1 − 1;
23
A[i] ← A[i1 ]; A[j] ← A[j1 ];
24 A[i1 ] ← p; A[j1 ] ← q;
25 posp ← i1 ; posq ← j1 ;
B.5. Algorithms for the Sampling Partitioning Method
The sampling method SP from Section 5 uses a mix of two classification algorithms:
“Always compare to the smaller pivot first”, and “Always compare to the larger pivot first”.
The actual partitioning method uses Algorithm 3 for the first ns = n/1024 classifications
and then decides which pivot should be used for the first comparison in the remaining
input. (This is done by comparing the two variables i and k in Algorithm 3.) If there are
more large elements than small elements in the sample it continues using Algorithm 3,
otherwise it uses Algorithm 6 below. If the input contains fewer than 1024 elements,
Algorithm 3 is used directly.
B.6. Algorithm for the Counting Strategy
Algorithm 7 is an implementation of the counting strategy from Section 6. It uses a
variable d which stores the difference of the number of small elements and the number
of large elements which have been classified so far. On average this algorithm makes
1.8n ln n + O(n) comparisons and 1.6n ln n assignments.
31
Algorithm 6 Simple Partitioning Method (smaller pivot first)
procedure SimplePartitionSmall(A, p, q, left, right, posp , posq )
1 l ← left + 1; g ← right − 1; k ← l;
2 while k ≤ g do
3
if A[k] < p then
4
swap A[k] and A[l];
5
l ← l + 1;
6
k ← k + 1;
7
else
8
if A[k] < q then
9
k ← k + 1;
10
else
11
swap A[k] and A[g]
12
g ← g − 1;
13 swap A[left] and A[l − 1];
14 swap A[right] and A[g + 1];
15 posp ← l − 1; posq ← g + 1;
Algorithm 7 Counting Strategy C
procedure C-Partition(A, p, q, left, right, posp , posq )
1 i ← left + 1; k ← right − 1; j ← i;
2 d ← 0;
. d holds the difference of the number of small and large elements.
3 while j ≤ k do
4
if d > 0 then
5
if A[j] < p then
6
swap A[i] and A[j];
7
i ← i + 1; j ← j + 1; d ← d + 1;
8
else
9
if A[j] < q then
10
j ← j + 1;
11
else
12
swap A[j] and A[k];
13
k ← k − 1; d ← d − 1;
14
else
15
while A[k] > q do
16
k ← k − 1; d ← d − 1;
17
if j ≤ k then
18
if A[k] < p then
19
rotate3(A[k], A[j], A[i]);
20
i ← i + 1; d ← d + 1;
21
else
22
swap A[j] and A[k];
23
j ← j + 1;
24 swap A[left] and A[i − 1];
25 swap A[right] and A[k + 1];
26 posp ← i − 1; posq ← k + 1;
32
C. A FAST THREE-PIVOT ALGORITHM
We give here the complete pseudocode for the three-pivot algorithm described in [Kushagra et al. 2014]. In contrast to the pseudocode given in [Kushagra et al. 2014, Algorithm
A.1.1], we removed two unnecessary bound checks (Line 5 and Line 10 in our code) and
we move misplaced elements in Lines 15–29 using less assignments. This is used in the
implementation of [Kushagra et al. 2014], as well.12 On average, this algorithm makes
1.846..n ln n + O(n) comparisons and 1.57..n ln n + O(n) assignments.
Algorithm 8 Symmetric Three-Pivot Sorting Algorithm
procedure 3-Pivot(A, left, right)
Require: right − left ≥ 2, A[left] ≤ A[left + 1] ≤ A[right]
1 p1 ← A[left]; p2 ← A[left + 1]; p3 ← A[right];
2 i ← left + 2; j ← i; k ← right − 1; l ← k;
3 while j ≤ k do
4
while A[j] < p2 do
5
if A[j] < p1 then
6
swap A[i] and A[j];
7
i ← i + 1;
8
j ← j + 1;
9
while A[k] > p2 do
10
if A[k] > p3 then
11
swap A[k] and A[l];
12
l ← l − 1;
13
k ← k − 1;
14
if j ≤ k then
15
if A[j] > p3 then
16
if A[k] < p1 then
17
rotate4(A[j], A[i], A[k], A[l]);
18
i ← i + 1;
19
else
20
rotate3(A[j], A[k], A[l]);
21
l ← l − 1;
22
else
23
if A[k] < p1 then
24
rotate3(A[j], A[i], A[k]);
25
i ← i + 1;
26
else
27
swap A[j] and A[k];
28
j ← j + 1; k ← k − 1;
29 rotate3(A[left + 1], A[i − 1], A[j − 1]);
30 swap A[left] and A[i − 2];
31 swap A[right] and A[l − 1];
32 3-Pivot(A, left, i − 3);
33 3-Pivot(A, i − 1, j − 2);
34 3-Pivot(A, j, l);
35 3-Pivot(A, l + 2, right);
12 Code
made available by Alejandro López-Ortiz.
33
Comparisons /n ln n
1.9
1.8
1.7
1.6
1.5
QS
S
SP
1.4
8
10
12
14
16
18
20
22
Y
C
L
24
26
28
30
Items [log2 (n)]
Fig. 5. Average comparison count (scaled by n ln n) needed to sort a random input of up to n = 229 integers.
We compare classical quicksort (QS), Yaroslavskiy’s algorithm (Y), the sampling algorithm (SP), the counting
algorithm (C), the modified version of Sedgewick’s algorithm (S), and algorithm L. Each data point is the
average over 400 trials.
Comparisons /n ln n
1.6
1.4
QS
S
C
L
1.2
Y
Y0
SP
L0
1
8
10
12
14
16
18
20
22
24
26
28
30
Items [log2 (n)]
Fig. 6. Average comparison count (scaled by n ln n) needed to sort a random input of up to n = 229 integers.
We compare classical quicksort (QS) with the Median-of-3 strategy, Yaroslavskiy’s algorithm (Y), the sampling
algorithm (SP), the counting algorithm (C), the modified version of Sedgewick’s algorithm from Section 4
(S), and algorithm L. Each of these dual-pivot algorithms uses the tertiles in a sample of size 5 as the two
pivots. Moreover, L0 is an implementation of strategy L which uses the third- and sixth-largest element from
a sample of size 11. Y 0 is Yaroslavskiy’s algorithm choosing the tertiles of a sample of size 11 as the two pivots.
Each data point is the average over 400 trials.
34
Time /n ln n [ns]
4.8
QS
C
L
Y
SP
K
24
26
4.6
4.4
4.2
4
3.8
12
14
16
18
20
22
28
Items [log2 (n)]
Fig. 7. Running time experiments in C++, setting 1 with compiler flags: -O2. Each data point is the average
over 1000 trials. Times are scaled by n ln n.
5.2
Time /n ln n [ns]
5
QS
C
L
Y
SP
K
24
26
4.8
4.6
4.4
4.2
4
3.8
12
14
16
18
20
22
28
Items [log2 (n)]
Fig. 8. Running time experiments in C++, setting 2 with compiler flags: -O2 -funroll-loops. Each data point
is the average over 1000 trials. Times are scaled by n ln n.
35
7.5
Time /n ln n [ns]
7
QS
Y
C
SP
L
K
6.5
6
5.5
14
16
18
20
22
24
26
Items [log2 (n)]
Fig. 9. Running time experiments in Java 8. For warming up the JIT, we let each algorithm sort 10 000
inputs consisting of 100 000 elements. Each data point is the average over 500 trials. Times are scaled by
n ln n.
36
| 8cs.DS
|
arXiv:1503.00966v2 [math.ST] 10 Apr 2017
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLING:
BOUNDS, ADAPTIVE RESAMPLING VIA ∞-ESS,
AND AN APPLICATION TO PARTICLE GIBBS
JONATHAN H. HUGGINS AND DANIEL M. ROY
Abstract. Sequential Monte Carlo (SMC) algorithms were originally designed
for estimating intractable conditional expectations within state-space models,
but are now routinely used to generate approximate samples in the context of
general-purpose Bayesian inference. In particular, SMC algorithms are often
used as subroutines within larger Monte Carlo schemes, and in this context,
the demands placed on SMC are different: control of mean-squared error is
insufficient—one needs to control the divergence from the target distribution
directly. Towards this goal, we introduce the conditional adaptive resampling
particle filter, building on the work of Gordon, Salmond, and Smith (1993),
Andrieu, Doucet, and Holenstein (2010), and Whiteley, Lee, and Heine (2016).
By controlling a novel notion of effective sample size, the ∞-ESS, we establish
the efficiency of the resulting SMC sampling algorithm, providing an adaptive
resampling extension of the work of Andrieu, Lee, and Vihola (2013). We apply
our results to arrive at new divergence bounds for SMC samplers with adaptive resampling as well as an adaptive resampling version of the Particle Gibbs
algorithm with the same geometric-ergodicity guarantees as its nonadaptive
counterpart.
1. Introduction
2. Preliminaries
3. Conditional αSMC
4. Main Results
5. Bounding the Divergence of SMC Samplers
6. The α-Particle Gibbs Sampler
Appendix A. Additional Proofs
Acknowledgments
References
1
8
10
12
18
20
23
33
33
1. Introduction
Sequential Monte Carlo (SMC) methods are a popular class of algorithms for approximate inference [Doucet et al., 2000, 2001, Del Moral et al., 2006, Kantas et al.,
2009, Doucet and Johansen, 2010, Künsch, 2013]. In the context of Bayesian inference, SMC produces a particle approximation to the posterior distribution as well
as an unbiased estimate of the marginal likelihood. Traditionally, particle approximations were built to estimate conditional expectations, and the analysis of SMC
methods focused on this operator perspective, by bounding the mean squared error
of the resulting estimates.
Date: Tuesday 11th April, 2017.
1
2
J. H. HUGGINS AND D. M. ROY
Increasingly, SMC methods are being used to produce approximate samples,
usually in the inner loop of other approximate inference algorithm. A key example is the class of particle Markov chain Monte Carlo (PMCMC) methods, which
aim to combine the best features of SMC and MCMC approaches by using SMC
as a proposal mechanism for a Metropolis–Hastings (“particle MH”) or approximate Gibbs (“particle Gibbs”) sampler [Holenstein, 2009, Andrieu et al., 2010].
Characterizing the efficiency of PMCMC methods is an active area of investigation [Andrieu and Roberts, 2009, Andrieu and Vihola, 2014, Andrieu et al., 2013,
Chopin and Singh, 2013, Lindsten et al., 2014, Lee and Latuszynski, 2014].
When SMC methods are employed for sampling, convergence guarantees from
the operator perspective are insufficient. In this work, we take up the sampling perspective, and study the distribution of a sample drawn from the SMC particle filter
approximation. Building off the work of Gordon et al. [1993], Andrieu et al. [2010,
2013], and Whiteley et al. [2016], we use conditional filters to derive a minorization
condition, lower bounding the density of the approximate sample’s distribution in
terms of that of the target distribution. The analysis extends to conditional SMC
as well as to adaptive resampling versions. One of our key contributions is a novel
notion of effective sample size, the ∞-ESS, which we use to establish the efficiency
of the adaptive SMC sampling algorithm. Thus, our results are both a sampling
analogue to the operator work of Whiteley et al. [2016] and an adaptive resampling
extension to the sampling work of Andrieu et al. [2013]. We apply our results to
arrive at new divergence bounds for SMC samplers with adaptive resampling as
well as an adaptive resampling version of the Particle Gibbs algorithm with the
same geometric-ergodicity guarantees as its nonadaptive counterpart.
In the remainder of this section we provide an overview of our contributions
for the special case of the the conditional adaptive resampling particle filter: we
introduce the conditional adaptive resampling particle filter, present our main theoretical results characterizing its performance, and describe an application to a novel
adaptive resampling Particle Gibbs algorithm.
1.1. (Conditional) adaptive resampling particle filters. We follow a similar
setup and notation to Del Moral [2004]. Let (ξt )t≥1 be an inhomogeneous Markov
chain on the measurable space (E, E) with transition kernels (Mt )t≥2 and initial
distribution M1 . Denote expectations with respect to the Markov chain by E[·].
Let gt : E → R+ , for t ≥ 1, be a sequence of E-measurable potential functions
Qt
on E, let g0 ≡ 1, and write gs:t (xs:t ) , τ =s gτ (xτ ). For t = 1, 2, . . . , define the
measure π1:t on E t given by
π1:t (dx1:t ) , γ1:t (dx1:t )/Zt ,
(1.1)
where
γ1:t (dx1:t ) ,
t
Y
gs (xs )Ms (xs−1 , dxs ) and Zt , γ1:t (1).
s=1
(We have written M1 (x0 , dx1 ) for M1 (dx1 ).) Equivalently,
π1:t (φ) ,
E [φ(ξ1:t )g1:t (ξ1:t )]
,
Zt
φ : E t → R measurable,
(1.2)
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
3
where Zt , E [g1:t (ξ1:t )] is the normalization constant.1
Towards the goal of efficiently approximating π1:t , we introduce a novel sequential Monte Carlo algorithm: the i-times conditional adaptive resampling particle
filter (ci ARPF), which is a generalization of the adaptive resampling particle filter [Gordon et al., 1993, Künsch, 2013] and the conditional SMC algorithm used
in particle Gibbs [Andrieu et al., 2010]. (In Section 3, we will introduce a further
generalization.) The integer parameter i ≥ 0 determines the number of fixed tra1
i
jectories y1:t
, . . . , y1:t
∈ E t required by the algorithm, which operates by generating
1:N
n N
a collection X̃ 1:t , {X̃1:t
}n=1 of N > i particles with corresponding nonnega1:N
tive weights Wt
, {Wtn }N
n=1 . When i = 0, we recover the standard (unconditional) adaptive resampling particle filter; when i = 1, we recover a generalization
of the conditional SMC algorithm that includes adaptive resampling. For time
s = 1, . . . , t, the measure π1:s is approximated by
i,N
π1:s
,
N
X
Wsn gs (Xsn )
δ n .
PN
k
k X̃1:s
k=1 Ws gs (Xs )
n=1
1:N
The ci ARPF algorithm iteratively constructs X̃ 1:N
as follows: The
1:t and Wt
first i particles are deterministically set to match the fixed trajectories:
Xsn = ysn ,
n
n
X̃1:s
= y1:s
,
s = 1, . . . , t and n = 1, . . . , i.
At time s = 1, the remaining N − i particles X1n , for n = i + 1, . . . , N , are sampled
independently and identically from M1 . The corresponding (length 1) trajectories
are
n
X̃1:1
= X1n ,
n = i + 1, . . . , N.
Furthermore, for all n = 1, . . . , N , W1n = 1.
The remaining particle trajectories are generated as follows: First, we introduce a cutoff parameter η ∈ [0, 1] and an effective sample size (ESS) function
ESS : RN
+ → [1, N ]. The ESS function measures how uniform the current weights
1:N
Ws are. Typically ESS(Ws1:N ) = 1 indicates that all but one weight is zero and
ESS(Ws1:N ) = N indicates all the weights are equal.
For each time s = 2, . . . , t:
1:N
• If ESS(Ws−1
) ≤ ηN , a resampling step is introduced. For n = 1, . . . , N ,
the weights are set to a common value
Wsn = Ws ,
N
1 X k
k
Ws−1 gs−1 (Xs−1
)
N
k=1
and, for n = i + 1, . . . , N , particle n’s “ancestor” at time s, denoted Ans , is
sampled independently, such that Ans = k, for k = 1, . . . , N , with probability
k
k
Ws−1
gs−1 (Xs−1
)
.
N Ws
1In the state-space setting, the potential g would be the conditional density (i.e., likelihood)
t
of the observation vt at time t as a function of unobserved state xt : i.e., gt (xt ) = pt (vt | xt ).
Then π1:t would be the posterior distribution of the unobserved state sequence given the observed
sequence.
4
J. H. HUGGINS AND D. M. ROY
1:N
• If ESS(Ws−1
) > ηN , then the algorithm does not resample the particles.
For n = 1, . . . , N , the weights are copied, i.e.,
n
n
Wsn = Ws−1
gs−1 (Xs−1
),
and, for n = i + 1, . . . , N , a record is made that particle n was its own
ancestor by setting Ans = n.
• Having sampled the ancestors, the algorithm propagates the particles forAn
s
ward. For n = i + 1, . . . , N , Xsn is sampled from Ms (Xs−1
, ·), and the
corresponding trajectories are set to
An
n
s
X̃1:s
= hX1:s−1
, Xsn i .
∗
In the final step of the algorithm, a single particle X̃1:t
is sampled from the full
i,N
approximation π1:t , and the algorithm yields an estimate of the normalization
constant Zt ,
Ẑt ,
N
1 X n
W gt (Xtn ).
N n=1 t
Let Ei,N
[·] denote the expectation operator with respect to the ci ARPF, and
y 1:i
1:t
write
i,N
i,N
P i,N (y 1:i
1:t , dx1:t ) , Py 1:i (dx1:t ) , Ey 1:i [δX̃ ∗ (dx1:t )]
1:t
∗
X̃1:t
1:t
1:t
y 1:i
1:t
for the law of
when the i fixed trajectories are
∈ (E t )i . We can now
i
describe in more precise terms how the c ARPF kernel P i,N generalizes several
0,N
well-known SMC kernels. When i = 0, π1:t
is the standard adaptive SMC particle
∗
approximation of π1:t and X̃1:t is a single sample from the particle approximation. When i = 1 and resampling is done at every step by taking η = 1, P 1,N
is exactly the conditional SMC kernel used in particle Gibbs samplers [Holenstein,
2009, Andrieu et al., 2010, 2013]. For general η ∈ (0, 1), we obtain a novel adaptive
resampling variant that we study in the sequel. In particular, under mild regularity
conditions, P 1,N defines a Markov kernel with invariant distribution π1:t .
1.2. Controlling ci ARPF efficiency with ∞-ESS. We can analyze the quality
of the ci ARPF kernel P i,N by quantifying the extent to which high-probability
sets under the target distribution also have high probability under the kernel. The
following theorem establishes a minorization condition for the i-times conditional
filter in terms of the (i + 1)-times conditional filter:
1
i
Theorem 1.1. For all t ≥ 1, i ≥ 0, N > i, and y1:t
, . . . , y1:t
∈ Et,
Z
Zt
i+1
t
π1:t (dy1:t
), S ⊆ E t measurable. (1.3)
P i,N (y 1:i
1:t , S) ≥ (1 − i/N )
i+1,N
S E 1:i+1 [Ẑt ]
y
1:t
The integral appearing in Eq. (1.3) has no simple form in general, but in many
settings we will be able to obtain a lower bound on the integrand that does not depend on the fixed trajectories y 1:i+1
1:t . In those cases, the integral is simply replaced
by this uniform lower bound. For i = 0, we are then immediately able to control
∗
numerous measures of divergence between π1:t and P 0,N (i.e., the law of X̃1:t
). For
example, in the case of total variation distance, we have the following corollary to
Theorem 1.1:
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
5
t
Corollary 1.2. If E1,N
y1:t [Ẑt /Zt ] ≤ Bt,N for all y1:t ∈ E , then
−1
dT V (π1:t , P 0,N ) ≤ 1 − Bt,N
.
(1.4)
For i = 1, a uniform lower bound assumption implies a minorization condition
on the kernel P 1,N (y1:t , dx1:t ), which in turn implies fast mixing of the Markov
chain with kernel P 1,N :
1
2
Corollary 1.3. If E2,N
[Ẑt /Zt ] ≤ Bt,N for all y1:t
, y1:t
∈ E t , then the Markov
y 1:2
1:t
chain with transition kernel P 1,N (y1:t , dx1:t ) is uniformly ergodic in total variation
distance and has invariant distribution π1:t . In particular, for all y1:t ∈ E t and
k ≥ 1,
k
−1
(1.5)
(1 − 1/N )t ,
dT V (π̃yk1:t , π1:t ) ≤ 1 − Bt,N
where π̃yk1:t , δy1:t [P 1,N ]k is the law of the Markov chain, with initial state y1:t ,
after k transitions.
In order to apply the corollaries, it remains to bound Ei,N
[Ẑt /Zt ]. Such a bound
y 1:i
1:t
was obtained for the nonadaptive conditional SMC kernel in Andrieu et al. [2013].
However, in our general adaptive resampling setting, one must make a careful choice
of effective sample size function. To this end, we introduce a generalized notion
of effective sample size, which includes several existing definitions as special cases.
p
For p ∈ (1, ∞], let p∗ , p−1
be the conjugate exponent of p (so 1/p + 1/p∗ = 1).
The p-effective sample size (p-ESS) of the weight vector w1:N ∈ RN
+ is
kw1:N kp1∗
.
kw1:N kpp∗
The following proposition highlights some elementary properties of p-ESS.
ESSp (w1:N ) ,
Proposition 1.4. The p-ESS has the following properties:
(1) For all p ∈ (1, ∞], 1 ≤ ESSp (w1:N ) ≤ N . The lower bound is achieved if and
only if all but one of the weights is zero. The upper bound is achieved if and
only if all the weights are equal.
(2) For 1 < p < q ≤ ∞, ESSp (w1:N ) ≥ ESSq (w1:N ) ≥ N −(1−q∗ /p∗ ) ESSp (w1:N ),
with equality if and only if K weights are equal and the rest are zero.
Part (1) demonstrates that the p-ESS satisfies basic properties one would expect
of a measure of effective sample size. Part (2) places the family of p-ESS measures
in a linear order: the larger the value of p, the more stringent the notion of effective
sample size.
The standard definition of effective sample size is precisely the 2-ESS. Whiteley et al.
[2016] provided a rigorous justification for the use of 2-ESS from the operator perspective: if adaptive resampling is used to guarantee that the 2-ESS does not fall
below ζN , for some fixed parameter ζ ∈ (0, 1], then the error bounds on the operator approximation match those of the nonadaptive sampler with ζN particles.
More formally, let ℑt (y1:t ) = yt be the projection onto the t-th component. Under
appropriate regularity conditions, for every bounded measurable φ : E → R and
real r ≥ 1,
i1/r
h
a(r)b(φ)
0,N
(φ ◦ ℑt ) − π(φ ◦ ℑt )|r
≤ √
sup ESS2 (Wt1:N ) ≥ ζN =⇒ sup E0,N |π1:t
,
ζN
t≥1
t≥1
6
J. H. HUGGINS AND D. M. ROY
where a(r) and b(φ) are explicit functions.
To upper bound Ei,N
[Ẑt /Zt ], however, we will require a lower bound on the
y 1:i
1:t
∞-ESS, which by Proposition 1.4(2) is a more stringent notion of effective sample
size than 2-ESS:
Assumption 1.A. There exists ζ ∈ (0, 1] such that ESS∞ (Ws1:N ) ≥ ζN for all
1 ≤ s ≤ t.
For the choice ESS = ESS∞ (i.e., under Assumption 1.A), we can bound the
estimate of the normalization constant. Let Gs,t (xs ) , E[gs:t−1 (ξs:t−1 ) | ξs = xs ]
for s = 1, . . . , t and G0,t , E[g1:t−1 (ξ1:t−1 )]. We now arrive at our second main
result:
1
i
Theorem 1.5. If Assumption 1.A holds, then for all i, t ≥ 1, N > i, y1:t
, . . . , y1:t
∈
t
E ,
Pt Pi
i
Zt−1 s=1 j=1 G0,s Gs,t+1 (y1:t
) − ζi
i,N
+ Θ(N −2 ).
(1.6)
Ey 1:i [Ẑt /Zt ] ≤ 1 +
1:t
ζN
Two possible further assumptions both lead to uniform bounds on Ei,N
[Ẑt ].
y 1:i
1:t
Assumption 1.B. The potentials satisfy gs , supx∈E gs (x) < ∞ for all 1 ≤ s ≤ t.
Assumption 1.C. There exists a constant β > 0 such that for any t, s ∈ N,
sup
x∈E
G0,t Gt,t+s (x)
≤ β.
G0,t+s
Corollary 1.6. Under the same conditions as Theorem 1.5, if Assumption 1.B
holds then
#
"
t
t
Y
i
i,N
−1
Ey 1:i [Ẑt /Zt ] ≤ 1 + Zt
−1
(1.7)
1+
gs
1:t
ζN
s=1
while if Assumption 1.C holds then
Ei,N
[Ẑt /Zt ]
y 1:i
1:t
t
β
≤ 1+
.
ζN
(1.8)
Combining Corollaries 1.2 and 1.6 yields the following guarantees for the ARPF
sampler:
Theorem 1.7. If Assumptions 1.A and 1.B hold then
Qt
tZt−1 s=1 g s
+ Θ(N −2 )
dT V (π1:t , P 0,N ) ≤
Q
ζN + tZt−1 ts=1 gs
(1.9)
while if Assumptions 1.A and 1.C hold then
dT V (π1:t , P 0,N ) ≤
βt
+ Θ(N −2 ).
ζN + βt
(1.10)
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
7
1.3. Applications to Particle Gibbs. In the language of state-space models,
the setting described so far involves approximating the posterior distribution of a
Markov chain given indirect stochastic observations of the chain’s values. However,
it is often the case that the chain and the potentials are controlled by a global
parameter θ ∈ Θ for which there is a prior distribution ̟(dθ). Replace Ms by
Msθ and gs by gsθ , then parameterize the other quantities defined previously in
terms of Ms and gs by θ. Let (Y, Y) , (E t , B(E t )). We will suppress much of the
time dependence when possible to make the notation less cluttered. The target
distribution on the product space (Θ × Y, B(Θ × Y )) is
π(dθ × dy) , γ(dθ × dy)/Z,
(1.11)
where
γ(dθ × dy) , ̟(dθ)
t
Y
gsθ (ys )Msθ (ys−1 , dys ) and Z , γ(1).
(1.12)
s=1
Let πθ (dy) and πy (dθ) denote the disintegrations of π along Θ and along Y , respectively.
The particle Gibbs sampler approximates the two-stage Gibbs kernel
Π(θ, y, dϑ × dz) , πy (dϑ)πϑ (dz).
(1.13)
In many settings, such as non-linear or non-Gaussian state-space models, it is possible to sample from πy (dϑ), but difficult to sample from πϑ (dz). The idea is to
replace πϑ (dz) with an SMC-based approximation Πϑ (y, dz) that leaves πϑ (dz)
invariant, leading to a kernel of the form πy (dϑ)Πϑ (y, dz).
We introduce the adaptive resampling particle Gibbs (ARPG) sampler, which
1,N
employs the cARPF kernel Pϑ,y
(dz) to approximate the conditional distribution
πϑ (dz) that would be used in a standard Gibbs sampler. The ARPG kernel is thus
given by
1,N
ΠN (θ, y, dϑ × dz) , πy (dϑ)Pϑ,y
(dz).
(1.14)
Theorem 1.1 and Corollaries 1.3 and 1.6 together with the results of Andrieu et al.
[2013] yield guarantees on the ergodicity properties of the cARPF kernel and the
ARPG sampler. Once instances of N are replaced by ζN , the guarantees essentially
match those provided by Andrieu et al. [2013] for the standard PG sampler.
Theorem 1.8. If Assumption 1.A holds, then the cARPF kernel and ARPG sampler have the following properties:
(1) If Assumption 1.B holds then there exists ǫt,N = 1 − Ct /N such that for any
θ ∈ Θ, y ∈ Y , and k ≥ 1,
dT V (δy [Pθ1,N ]k , πθ ) ≤ (1 − ǫt,N )k .
(1.15)
(2) If Assumption 1.C holds and N ≥ t/C + 1 for any fixed C > 0, then for any
t ≥ 1, Eq. (1.15) holds with
C
(1.16)
ǫt,N ≥ exp − (2β + ζ) .
ζ
(3) If either Assumption 1.B or Assumption 1.C holds, then whenever the Gibbs
sampler is geometrically ergodic the ARPG sampler is geometrically ergodic as
well.
8
J. H. HUGGINS AND D. M. ROY
At a high level, the results we have obtained highlight the role of the expected
value of Ẑt in the mixing properties of conditional SMC Markov chains and particle Gibbs (PG) samplers: In order to show geometric ergodicity for adaptive
resampling particle Gibbs samplers, it suffices to establish bounds on the expected
value of Ẑt under the twice-conditional filter, and the growth of the expectation as
t increases determining how well the particle Gibbs algorithm scales. Similarly, a
bound on the expected value of Ẑt under the once-conditional filter implies a bound
on dT V (π1:t , P 0,N ). Hence, as a slogan, good performance of (adaptive resampling)
particle Gibbs is equivalent to good performance of (adaptive) SMC for sampling.
2. Preliminaries
In this section, we fix some additional notation, introduce a few key additional
definitions, and then present αSMC [Whiteley et al., 2016], a generalization of the
adaptive resampling particle filter described in the introduction.
For a positive integer K, let [K] , {1, 2, . . . , K}. If xi , . . . , xj are elements of a
P
sequence,
write xi:j , hxi , xi+1 , . . . , xj i. We use the following conventions: ∅ = 0,
Q
∅ = 1, and 0/0 = 0.
Let (S, S), (S ′ , S ′ ) be measurable spaces. Then K : S × S ′ → R is a kernel if
K(·, B) is a (S, S)-measurable function for all B ∈ S ′ and K(x, ·) is measure on
(S ′ , S ′ ) for Rall x ∈ S. For a measure µ on (S, S)Rand kernels K, K ′ : S × S → R, let
µK(dy) , µ(dx)K(x, dy) and KK ′ (x, dz) , K(x, dy)K ′ (y, dz). We will often
use measures and kernels
function φ : S → R, let
R as operators. For a measurable
R
µ(φ) , Eξ∼µ [φ(ξ)] = φ(x)µ(dx) and K(φ)(x) , φ(y)K(x, dy). For measures
µ, ν on (S, S), we will write µ ≪ ν to denote that µ is absolutely continuous with
respect to ν, in which case we will write
R dµ/dν for the ν-almost everywhere (ν-a.e.)
unique function f satisfying µ(A) = A f dν, for all A ∈ S. When the choice is
clear from context, we may write B(S) for the σ-algebra of the space S.
For probability measures µ and ν on (S, S), the total variation distance between
µ and ν is
dT V (µ, ν) , sup |µ(A) − ν(A)|.
(2.1)
A∈S
If µ ≪ ν, then the Kullback–Liebler (KL) divergence is
dKL (µ||ν) , µ(log dµ/dν)
(2.2)
dχ2 (µ||ν) , ν([dµ/dν − 1]2 ) = µ(dµ/dν) − 1.
(2.3)
and the χ2 divergence is
Finally, we note that, when there is little risk of confusion, we will ignore
measure-theoretic niceties such as the distinction between equality and a.e.-equality.
Recall that (ξt )t≥1 is an inhomogeneous Markov chain on (E, E) with transition
kernels (Mt )t≥2 and initial distribution M1 , and that E[·] denotes expectation with
respect to the Markov chain. We will write M1 (x0 , ·) for M1 (·) when convenient
and, for all t ≥ 1 and xt−1 ∈ E, we will assume that Mt (xt−1 , ·) has a density with
respect to some common σ-finite dominating measure (which we denote by dx).
We will abuse notation and write Mt (xt−1 , xt ) for the density of Mt (xt−1 , ·) as xt .
Recall that, for each t ≥ 1, gt : E → R+ denotes a E-measurable potential function,
Qt
with g0 ≡ 1. Finally, recall that gs:t (xs:t ) , τ =s gτ (xτ ).
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
9
2.1. Target distributions. We now introduce some additional target distributions. (We will also repeat the definition of γ1:t and π1:t for completeness.)
Let φ1:t : E t → R and φt : E → R denote generic measurable functions. For each
t ≥ 1, the unnormalized predictive and updated measures are defined, respectively,
by
′
γ1:t
(φ1:t ) , E [φ1:t (ξ1:t )g1:t−1 (ξ1:t−1 )]
and
γ1:t (φ1:t ) , E [φ1:t (ξ1:t )g1:t (ξ1:t )]
and
γt (φt ) , E [φt (ξt )g1:t (ξ1:t )] .
with corresponding marginal versions
γt′ (φt ) , E [φt (ξt )g1:t−1 (ξ1:t−1 )]
Our ultimate goal is to approximate the normalized predictive and updated measures along with their marginal versions:
γ1:t (φ1:t )
,
Zt
γt (φt )
,
πt (φt ) ,
Zt
′
γ1:t
(φ1:t )
,
Zt′
γ ′ (φt )
ηt (φt ) , t ′ ,
Zt
η1:t (φ1:t ) ,
π1:t (φ1:t ) ,
and
(2.4)
(2.5)
where Zt , γt (1) and Zt′ , γt′ (1) are normalization constants.
2.2. The αSMC algorithm. In the introduction, adaptation in the particle filter was implemented via a simple multinomial resampling step, triggered when the
effective sample size fell below a fixed threshold. For the remainder of the article,
we will consider a more general mechanism for adaptation captured by the αSMC
algorithm introduced by Whiteley et al. [2016]. The αSMC algorithm can produce sequential importance sampling (SIS), sampling importance resampling (SIR,
also known as the bootstrap filter), and numerous other SMC variants as special
cases. Not only does the αSMC formulation aid in analyzing adaptive resampling
strategies, it provides a useful framework for devising novel adaptive schemes with
attractive computational properties, such as admitting parallelization even on resampling steps. In the remainder of this section, we outline the (unconditional)
αSMC algorithm. In the following section, we introduce a novel i-times conditional
version of αSMC, which will include the ci ARPF as a special case.
The αSMC algorithm, which is given as Algorithm 1, provides a flexible resampling mechanism: at each time t, a stochastic matrix αt−1 is chosen from a set AN
of N × N matrices. We denote the value in the n-th row and k-th column of αt−1
by αnk
t−1 . The αSMC estimators are
0,N
π1:t
,
N
X
Wtn gt (Xtn )
δ n ,
PN
k
k X̃1:t
n=1
k=1 Wt gt (Xt )
0,N
η1:t
,
N
X
n=1
Wtn
PN
k=1
Wtk
δX̃ n ,
1:t
πt0,N ,
N
X
Wtn gt (Xtn )
δXtn ,
PN
k g (X k )
W
t
t
t
n=1
k=1
and ηt0,N ,
N
X
Wtn
δXtn ,
PN
k
n=1
k=1 Wt
and the estimators of the normalization constants Zt and Zt′ are
Ẑt ,
N
1 X n
W gt (Xtn )
N n=1 t
and Ẑt′ ,
N
1 X n
W .
N n=1 t
Expectations with respect the law of the αSMC algorithm are written as E0,N [·].
10
J. H. HUGGINS AND D. M. ROY
Algorithm 1 αSMC
for n = 1, . . . , N do
Sample X1n from M1
n
Set X̃1:1
← X1n
n
Set W1 ← 1
end for
for t = 2, 3, . . . do
1:N
Select αt−1 from AN according to some function of X 1:t−1
and A1:N
1:t−2
for n = 1, . . . , N do
P
N
k
k
Set Wtn ← k=1 αnk
t−1 Wt−1
Dgt−1 (Xt−1 )
EN
k
k
αnk
t−1 Wt−1 gt−1 (Xt−1 )
n
Sample At−1 from Multi
Wn
t
k=1
n
At−1
Sample Xtn Dfrom Mt (Xt−1
E , ·)
n
At−1
n
n
Set X̃1:t ← X1:t−1 , Xt
end for
end for
SIS, SIR, and the standard adaptive algorithm are obtained as special cases of
αSMC as follows. SIS is recovered when αt−1 = I N , the N × N identity matrix,
while SIR is recovered when αt−1 = 11/N , the N × N matrix with all entries equal
to 1/N . The adaptive particle filter (APF) algorithm is obtained by setting αt−1
1:N
to 11/N if ESS2 (Wt−1
) ≤ ζN and to I N otherwise, where ζ ∈ (0, 1] is fixed.
3. Conditional αSMC
It is useful both algorithmically and analytically to generalize αSMC in such a
way that one or more particle trajectories is fixed ahead of time. The result, which
we will refer to as conditional αSMC, is a strict generalization of the conditional
adaptive particle filter given in the introduction. We will use conditional αSMC
to study the properties of (unconditional) αSMC, to design novel adaptive particle
Gibbs algorithms, and to analyze their mixing properties.
For this section, fix t ≥ 1, i ≥ 0, and N > i. The i-times conditional αSMC
(ci αSMC) process (or simply the cαSMC process when i = 1) is defined on the
space (E N × [N ]N × [N ]i )t−1 × E N × [N ] × [N ]i , and is essentially equivalent to
αSMC with the first i particle trajectories, but not their lineages,
Q fixed a priori.
′
If f 1:i ∈ [N ]i are indices of the first i particles, let D(f 1:i ) , j6=j ′ 1(f j 6= f j )
be the function that indicates whether the indices are distinct. As in αSMC, the
1:N
1:N
N t
matrix αt−1 ∈ AN is a function of X 1:N
1:t−1 and A1:t−2 . We have x1:t ∈ (E ) ,
1:i
i t
1:N
N t−1
∗
f 1:t ∈ ([N ] ) , a1:t−1 ∈ ([N ] ) , and at ∈ [N ], and use the notation
w1n , 1,
wtn ,
N
X
k
k
αnk
t−1 wt−1 gt−1 (xt−1 ),
(3.1)
k=1
and
1:N
rn (k|ws−1
, x1:N
1:s−1 ) ,
k
k
αnk
s−1 ws−1 gs−1 (xs−1 )
.
wsn
(3.2)
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
11
Algorithm 2 Conditional αSMC
Require: Fixed trajectory y1:t
Sample F11 uniformly from {1, . . . , N }
for n = 1, . . . , N do
if n = F11 then
Set X1n ← y1
else
Sample X1n from M1
end if
n
Set X̃1:1
← X1n
n
Set W1 ← 1
end for
for t = 2, 3, . . . do
1:N
1:N
Select αt−1 from AN according to some
! function of X 1:t−1 and A1:t−2
N
kF 1
Sample Ft1 from Multi
αt−1t−1
k=1
for n = 1, . . . , N do
PN
k
k
Set Wtn ← k=1 αnk
t−1 Wt−1 gt−1 (Xt−1 )
if n = Ft1 then
1
Set Ant−1 ← Ft−1
Set Xtn ← yt
else
D
EN
k
k
αnk
t−1 Wt−1 gt−1 (Xt−1 )
Sample Ant−1 from Multi
Wn
t
k=1
An
t−1
Mt (Xt−1
, ·)
Xtn
Sample
from
end if
D An
E
t−1
n
Set X̃1:t
← X1:t−1
, Xtn
end for
end for
D
EN
W k g (X k )
PN t tn t n
Sample A∗t from Multi
W g (X )
n=1
t
t
t
k=1
1
i
For fixed trajectories y1:t
, . . . , y1:t
∈ E t , the law of the ci αSMC process is given by
1:i
1:i
i
1:i
Pi,N
[X11:N ∈ dx1:N
1 , F1 = f1 ] , C1 D(f1 )
y1:i
1:t
N
i
Y
Y
1
fj
M1 (dxn1 ), (3.3)
δyj (dx11 )
1
N
1:i
j=1
n∈f
/ 1
for s = 2, . . . , t,
1:N
1:N
1:i
1:i
Pi,N
[Xs1:N ∈ dx1:N
s , As−1 = as−1 , Fs = fs |
y 1:i
1:t
1:N
1:N
1:N
1:i
1:i
X 1:N
1:s−1 = x1:s−1 , A1:s−2 = a1:s−2 , Fs−1 = fs−1 ]
, Csi D(fs1:i )
×
Y
n∈f
/ s1:i
i
Y
fjfj
fj
fj
j
s s−1
s
= fs−1
)
δysj (dxss )1(as−1
αs−1
j=1
an
s−1
1:N
n
rn (ans−1 |ws−1
, x1:N
1:s−1 )Ms (xs−1 , xs )
(3.4)
12
J. H. HUGGINS AND D. M. ROY
and
a∗
Pi,N
[A∗t
y1:i
1:t
=
a∗t
| X 1:N
1:t
=
1:N
x1:N
1:t , A1:t−1
=
a1:N
1:t−1 ]
a∗
w t gt (xt t )
.
, PN t
n
n
n=1 wt gt (xt )
(3.5)
The Csi terms are normalization constants that ensure the expressions are valid
A∗
∗
probabilities. Let X̃1:t
, X1:tt , let Ei,N
[·] denote the expectation operator with
y 1:i
respect to the ci αSMC, and write
1:t
i,N
i,N
P i,N (y 1:i
1:t , dx1:t ) , Py 1:i (dx1:t ) , Ey 1:i [δX̃ ∗ (dx1:t )]
1:t
1:t
1:t
∗
X̃1:t
.
for the law of
The normalization constants Csi arise because the lineages f 1:i
1:t of the fixed trai
jectories y 1:i
are
kept
distinct.
The
c
αSMC
kernel
enforces
distinct
lineages for
1:t
j
j′
′
the fixed trajectories since in general y1:t 6= y1:t for j 6= j and, from an algorithmic
standpoint, allowing overlapping lineages could lead to a substantial increase in
complexity, both in terms of implementation and computation. The distinct lineage requirement is enforced by the D(fs1:i ) terms. Since there is at most one fixed
trajectory when i = 0 or 1, Cs0 = Cs1 = 1 for all s ∈ [t].
Algorithm 2 provides pseudocode to sample from the law of the cαSMC process,
which is a necessary part of the particle Gibbs sampler described in Section 6.
Sampling from the law of the ci αSMC process for i > 1 is more delicate, but
unnecessary since these are only used for analytical purposes.
Remark 3.1. To recover the ci ARPF described in the introduction, let αt−1 = 11/N
1:N
if ESS(Wt−1
) ≤ ζN and let αt−1 = I N otherwise. Then note that by the symmetry
of 11/N and I N , instead of sampling them, we can set the lineage for the j-th fixed
trajectory to j: that is, set fsj = j for all j ∈ [i] and s ∈ [t].
▽
4. Main Results
We are now ready to undertake our study of the i-times conditional αSMC kernel
1:i
, dx1:t ). Specifically, our aim is to understand the conditions under which
P i,N (y 1:t
i
the c αSMC kernel is close to π1:t (dx1:t ). Formally, we will establish a minorization
condition for the i-times conditional filter in terms of the expected value of Ẑt under
the (i + 1)-times conditional filter. The remainder of the section presents conditions
under which the expected value of Ẑt can be bounded. One of the key assumptions
is that adaptation controls the ∞-ESS.
Of particular interest are the cases i = 0, which corresponds to the αSMC filter,
and i = 1, which corresponds to the conditional αSMC kernel. The former provides
approximate samples from π1:t . The latter can be used to define a Markov chain
with invariant distribution π1:t , producing an adaptive resampling particle Gibbs
sampler. We consider both these applications in, respectively, Sections 5 and 6.
4.1. A minorization condition for the ci αSMC kernel. For the remainder of
the article, we will work under the following assumption:
Assumption 4.D. For all N ≥ 1, all α ∈ AN are doubly stochastic.
Remark 4.1. Assumption 4.D is the same as Assumption (B++ ) in [Whiteley et al.,
2016], although there, the condition is stated as assuming each α admits the uniform
distribution as an invariant measure.
▽
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
13
Let
κN ,
max
n6=m,α∈AN
N
X
αkn αkm
k=1
and
κ′N , max{κN , 1/N }.
(4.1)
Our first main result provides control over how much the measure Pyi,N
1:i differs from
1:t
π1:t . The theorem gives a stronger result when i = 0 and gives a simpler result
when i = 1, by expressing the lower bound on P 1,N (y1:t , S) in terms the more
transparent quantity κ′N instead of the normalization terms Cs2 . For example, if
AN = {I N } then κN = 0, while if AN = {11/N }, then κN = 1/N , so in either case
κ′N = 1/N .
Theorem 4.2. If Assumption 4.D holds, then for all t ≥ 1, i ≥ 0, N > i, S ⊆ E t
1
i
measurable, and y1:t
, . . . , y1:t
, x1:t ∈ E t ,
Z
Zt
P i,N (y 1:i
,
S)
≥
π1:t (dx1:t ).
(4.2)
Qt
1:t
i+1,N
i+1
i
C
[
Ẑ
E
S y 1:i ,x
s /Cs ]
t
s=1
1:t
1:t
In particular, in the case of i = 0, we have
Zt
Zt
dP 0,N
≥ 1,N
(x1:t ) = E1,N
x1:t
dπ1:t
Ẑt
Ex1:t [Ẑt ]
(4.3)
and in the case of i = 1, we have
P
1,N
(y1:t , S) ≥
Z
S
Zt (1 − κ′N )t
E2,N
y1:t ,x1:t [Ẑt ]
π1:t (dx1:t ).
(4.4)
Remark 4.3. By identical arguments, Theorem 4.2 also holds in the marginal and
predictive cases. In the predictive cases, however, Ẑt′ and Zt′ replace, respectively, Ẑt
and Zt . In the predictive case, under Assumption 4.D, Ẑt′ = Ẑt−1 and Zt′ = Zt−1 , so
later results pertaining to Ei,N
[Ẑt ], such as Proposition 4.4 and Theorem 4.7, apply
y1:i
1:t
′
to Zt as well. The fact that Zt′ = Zt−1 follows immediately from the definitions.
To show that Ẑt′ = Ẑt−1 , apply Assumption 4.D:
X
XX
X
k
k
−1
k
k
Ẑt′ = N −1
Wtn = N −1
αnk
Wt−1
gt−1
= Ẑt−1 .
▽
t−1 Wt−1 gt−1 = N
n
n
k
k
We will prove Theorem 4.2 in two parts: first for the case of i = 0, then for
i ≥ 1. For the i = 0 case (corresponding to vanilla αSMC), we begin by writing
the joint density of the αSMC process as
1:N
ψ(x1:N
1:t , a1:t−1 )
,
N
Y
n=1
!
M1 (xn1 )
t Y
N
Y
s=2 n=1
!
an
s−1
1:N
n
, x1:N
rn (ans−1 |ws−1
1:s−1 )Ms (xs−1 , xs )
(4.5)
.
1:N
1:N
1
Under Assumption 4.D, the density of the cαSMC process with law P1,N
y1:t [X 1:t , A1:t−1 , F1:t ]
f1
t
can be written in the following “collapsed” form, by implicitly identifying x1:t
with
14
J. H. HUGGINS AND D. M. ROY
y1:t :
1:N
1
ψ̃(x1:N
1:t , a1:t−1 , f1:t )
1
Qt
fs1 fs−1
1:N
ψ(x1:N
1:t , a1:t−1 )
s=2 Is αs−1
=
(4.6)
1
1
f1 Q
1 |w1:N , x1:N )M (xfs−1 , xfs )
N M1 (x11 ) ts=2 rfs1 (fs−1
s
s
1:s−1
s−1
s−1
t
Y
Y
n
1 1
1 Y
a
f
f
s−1
s s−1
1:N
n
Is αs−1
rn (ans−1 |ws−1
, x1:N
M1 (xn1 )
=
1:s−1 )Ms (xs−1 , xs ) ,
N
1
1
s=2
n6=fs
n6=f1
fs1
1
where Is , 1(as−1 = fs−1
).
Proof of Theorem 4.2, i = 0 case. Consider the density
f1
1:N
1
1:N
1:N
1
t
π̃1:t (x1:N
1:t , a1:t−1 , f1:t ) , π1:t (x1:t )ψ̃(x1:t , a1:t−1 , f1:t ).
Then
f1
f1
1:N
t
t
ψ(x1:N
1:t , a1:t−1 )gt (xt )wt
P
N
1:N
1
n
n
π̃1:t (x1:N
1:t , a1:t−1 , f1:t )
n=1 gt (xt )wt
1
fs−1
f 1 Qt
fs1
ft1
ft1
1
1:N
M1 (x11 ) s=2 rfs1 (fs−1
|ws−1
, x1:N
1:s−1 )Ms (xs−1 , xs )gt (xt )wt
=
1
PN
fs1 fs−1
ft1 Qt
π1:t (x1:t
) s=2 Is αs−1
N −1 n=1 gt (xnt )wtn
1
1
1
1
fs1 fs−1
fs−1
fs−1
fs−1
f 1 Qt
f1
f1
f1
M1 (x11 ) s=2 αs−1
ws−1
gs−1 (xs−1
)Ms (xs−1
, xss )gt (xt t )wt t
=
1
PN
fs1 fs−1
ft1 Qt
f 1 Qt
π1:t (x1:t
) s=2 ws s s=2 Is αs−1
N −1 n=1 gt (xnt )wtn
1
Qt
fs−1
fs1
fs1
s=1 gs (xs )Ms (xs−1 , xs )
=
Qt
P
ft1
n
n
)N −1 N
π1:t (x1:t
s=2 Is
n=1 gt (xt )wt
1
Zt
,
=
Qt
Ẑt s=2 Is
Using the convention that 0/0 = 0, it follows that
P 0,N (dx1:t )
(
)
a∗
a∗
X Z
gt (xt t )wt t
1:N
1:N
ψ(x1:t , a1:t−1 ) PN
=
δ at (dx1:t ) dx1:N
1:t
n )wn x1:t
g
(x
t
∗
t
t
n=1
a1:N
,a
1:t−1 t
(
ft1
ft1 Qt
1:N
X Z
ψ(x1:N
1:t , a1:t−1 )gt (xt )wt
s=2 Is
=
1:i PN
1:N
1:N
π̃1:t (x1:t , a1:t−1 , f 1:t ) n=1 gt (xnt )wtn
1
a1:N
1:t−1 ,f1:t
)
1:N
1
1:N
× π̃1:t (x1:N
1:t , a1:t−1 , f1:t )δxft (dx1:t ) dx1:t
1:t
X
=
1
a1:N
1:t−1 ,f1:t
=
(
X
Z (
Zt
Z
Zt
1
a1:N
1:t−1 ,f1:t
Ẑt
)
1:N
1
π̃1:t (x1:N
t (dx1:t )
1:t , a1:t−1 , f1:t )δxf1:t
Ẑt
dx1:N
1:t
−ft1
ft
1:N
1
ψ̃(x1:N
1:t , a1:t−1 , f1:t )δx1:t (dx1:t )dx1:t
)
π1:t (dx1:t ).
(4.7)
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
15
The result follows from Lemma A.1.
We defer the proof of Theorem 4.2 in the i ≥ 1 case to Appendix A.2.
4.2. Bounding Ẑt under the ci αSMC kernel. In order to apply Theorem 4.2,
we must be able to control the quantity Ei,N
[Ẑt ]. As an initial step toward this
y 1:i
1:t
goal, we consider the SIR case:
Assumption SIR. For all s ∈ [t − 1], αs = 11/N .
For SIR, we can rewrite Ei,N
[Ẑt ] in an equivalent but more explicit form (Propoy 1:i
1:t
sition 4.4). Our goal will then be to provide general conditions under which Ei,N
[Ẑt ]
y 1:i
1:t
can be rewritten in a similar manner (Theorem 4.7).
Recall that G0,t , E[g1:t−1 (ξ1:t−1 )] and Gs,t (xs ) , E[gs:t−1 (ξs:t−1 ) | ξs = xs ] for
s ∈ [t] and xs ∈ E. For t ≥ 1, 1 ≤ ℓ ≤ s ≤ t + 1, let
Tt,ℓ,s , {hτ1 , . . . , τℓ i : t − s + 1 < τ1 < · · · < τℓ = t + 1}
and, for τ ∈ Tt,ℓ,s , define
Cℓ (τ , y1:t ) ,
ℓ−1
Y
Gτi ,τi+1 (yτi ).
(4.8)
(4.9)
i=1
We will sometimes write Cℓy (τ ) or Cℓτ (y1:t ) instead of Cℓ (τ , y1:t ). The following is
a straightforward generalization of [Andrieu et al., 2013, Proposition 9].
Proposition 4.4. If Assumption SIR holds, then for all t ≥ 1, i ≥ 1, N ≥ i,
i
1
∈ Et,
, . . . , y1:t
y1:t
Ei,N
[Ẑt ] =
y 1:i
1:t
t+1
1 X
(N − i)t+1−ℓ
Nt
ℓ=1
X
G0,τ1
τ ∈Tt,ℓ,t+1
In particular, in the case of i = 1, we have
E1,N
y1:t [Ẑt ] =
t+1
1 X
(N − 1)t+1−ℓ
Nt
ℓ=1
ℓ−1
Y
i
X
Gτm ,τm+1 (yτjm ).
(4.10)
m=1 j=1
X
G0,τ1 Cℓ (τ , y1:t ).
(4.11)
τ ∈Tt,ℓ,t+1
In order to obtain a version of Proposition 4.4 for the general ci αSMC case,
we will require that the algorithm enforce a lower bound on a carefully chosen
notion of effective sample size called ∞-ESS. The ∞-ESS is a member of a family
of effective sample size measures we call p-ESS, which also includes two commonly
used definitions as special cases.
p
be the conjugate exponent
Definition 4.5. For parameter p ∈ [1, ∞], let p∗ , p−1
of p (so 1/p+1/p∗ = 1). The p-effective sample size (p-ESS) of the weight vector
w1:N ∈ RN
+ is
1:N p∗
kw1:N k1
p>1
kw
kp
1:N
(4.12)
ESSp (w ) ,
1:N
kw
k
1
QN
p = 1.
n wn /kw1:N k1
n=1 (w
)
⊳
The following proposition highlights some elementary properties of p-ESS and
subsumes Proposition 1.4 (see Appendix A.3 for a proof).
16
J. H. HUGGINS AND D. M. ROY
Proposition 4.6. The p-ESS has the following properties:
(1) For all p ∈ [1, ∞], 1 ≤ ESSp (w1:N ) ≤ N . The lower bound is achieved if and
only if all but one of the weights is zero. The upper bound is achieved if and
only if all the weights are equal.
(2) For 1 < p < q ≤ ∞, ESSp (w1:N ) ≥ ESSq (w1:N ) ≥ N −(1−q∗ /p∗ ) ESSp (w1:N ),
with equality if and only if K weights are equal and the rest are zero.
(3) The 1-ESS satisfies
ESS1 (w1:N ) = lim ESSp (w1:N ) = ESSent (w1:N ) , eH(w
1:N
p↓1
where H(w1:N ) , −
wn
n kw 1:N k1
P
)
,
(4.13)
n
log kww
is the entropy.
1:N k
1
Parts (1) and (2) generalize their counterparts in Proposition 1.4 to all p ∈ [0, 1],
including the case p = 1. Part (3) shows that the 1-ESS corresponds to the entropic
ESS, which is a common choice of ESS in applications [Cornebise et al., 2008].
In order to obtain a bound on Ei,N
[Ẑt ], we will require a lower bound on the
y 1:i
1:t
∞-ESS of the weights, as formalized in Assumption 1.A. Our development follows
that of Whiteley et al. [2016], who used the 2-ESS lower bound guarantee to bound
the L2 norm of the weights in terms of their L1 norm. Similarly, we will use the
∞-ESS lower bound guarantee to bound the sup-norm of the weights in terms of
their L1 norm. Specifically, under Assumption 1.A, we have
ζN ≤ ESS∞ (Ws1:N ) =
kWs1:N k1
kWs1:N k1
=
,
kWs1:N k∞
supn Wsn
(4.14)
kW 1:N k
1
s
and so, for all n ∈ [N ] and s ∈ [t], Wsn ≤ ζN
. We can use this upper bound on
n
Ws to prove a result that is very similar to Proposition 4.4, but permits an arbitrary
adaptation scheme satisfying Assumptions 1.A and 4.D. (See Appendix A.4 for a
proof):
Theorem 4.7. If Assumptions 1.A and 4.D hold, then for all t ≥ 1, i ≥ 1, N ≥ i,
1
i
y1:t
, . . . , y1:t
∈ Et,
Ei,N
[Ẑt ]
y1:i
(4.15)
1:t
≤
t+1
X
1
N (ζN )t−1
X
(ζN )t+1−ℓ
ℓ=1 τ ∈Tt,ℓ,t+1
N −i
ζN
1(τ1 >1)
G0,τ1
ℓ−1
Y
i
X
Gτm ,τm+1 (yτjm ).
m=1 j=1
In particular, in the case of i = 1, we have
E1,N
y1:t [Ẑt ] ≤
t+1
X
1
N (ζN )t−1
X
ℓ=1 τ ∈Tt,ℓ,t+1
(ζN )t+1−ℓ
N −1
ζN
1(τ1 >1)
G0,τ1 Cℓy (τ ).
(4.16)
The gap between Proposition 4.4 and Theorem 4.7 is that most of the factors of
N − i in the former are replaced by factors of N in the latter. Luckily we are interested in the i = 1, 2 cases, so we expect the differences between the two quantities
to be fairly small. The following result, which is immediate upon expanding the
left-hand sides of Eqs. (4.10) and (4.15) and keeping only Ω(1/N ) terms, formalizes
this intuition:
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
17
1
i
Corollary 4.8. If Assumption SIR holds, then for all i, t ≥ 1, N > i, y1:t
, . . . , y1:t
∈
t
E ,
Pt Pi
i
Zt−1 s=1 j=1 G0,s Gs,t+1 (y1:t
) − ti
Ei,N
[
Ẑ
/Z
]
=
1
+
+ Θ(N −2 ).
(4.17)
t
t
y 1:i
1:t
N
1
i
If Assumptions 1.A and 4.D hold, then for all i, t ≥ 1, N > i, y1:t
, . . . , y1:t
∈ Et,
P
P
i
Zt−1 ts=1 ij=1 G0,s Gs,t+1 (y1:t
) − ζi
[
Ẑ
/Z
]
≤
1
+
+ Θ(N −2 ).
(4.18)
Ei,N
t
t
1:i
y 1:t
ζN
4.3. Quantitative bounds. Recall Assumptions 1.B and 1.C, either of which can
be used in conjunction with Theorem 4.7 to obtain uniform, quantitative bounds
on Ei,N
[Ẑt ] by following the approach of Andrieu et al. [2013]:
y 1:i
1:t
Assumption 1.B. The potentials satisfy gs , supx∈E gs (x) < ∞ for all 1 ≤ s ≤ t.
Assumption 1.C. There exists a constant β > 0 such that for any t, s ∈ N,
sup
x∈E
G0,t Gt,t+s (x)
≤ β.
G0,t+s
Assumption 1.C is implied by a standard “strong mixing” condition which is
often employed in SMC analyses (e.g., [Del Moral, 2004, Whiteley et al., 2016]).
See Andrieu et al. [2013] for details.
Proposition 4.9. If αs = 11/N for s ∈ [t − 1] and Assumption 1.B holds, then for
1
i
all t ≥ 1, i ≥ 1, N ≥ i, y1:t
, . . . , y1:t
∈ Et,
t #
"
t
Y
i
.
(4.19)
gs − 1 1 − 1 −
Ei,N
[Ẑt /Zt ] ≤ 1 + Zt−1
y 1:i
1:t
N
s=1
1
i
If Assumptions 1.A, 1.B and 4.D hold, then for all t ≥ 1, i ≥ 1, N ≥ i, y1:t
, . . . , y1:t
∈
t
E ,
"
#
t
t
Y
i
i,N
−1
Ey 1:i [Ẑt /Zt ] ≤ 1 + Zt
gs
1+
−1 .
(4.20)
1:t
ζN
s=1
Proof. The proof Eq. (4.19) is a straightforward generalization of that for [Andrieu et al.,
2013, Proposition 12] with some additional bookkeeping for i (instead of 2) fixed
trajectories. As for Eq. (4.20), we have
Ei,N
[Ẑt ] ≤
y 1:i
1:t
t+1
X
X
(ζN )−ℓ+1 G0,τ1
ℓ=1 τ ∈Tt,ℓ,t+1
t
Y
t+1
X
ℓ−1
Y
i
X
Gτm ,τm+1 (yτjm )
m=1 j=1
t
gs
iℓ−1 (ζN )−ℓ+1
≤ Zt +
ℓ
−
1
s=1
ℓ=2
t
t
Y X
t
gs
(ζN/i)−ℓ
= Zt +
ℓ
s=1
ℓ=1
#
"
t
t
Y
i
= Zt +
−1 .
gs
1+
ζN
s=1
(4.21)
18
J. H. HUGGINS AND D. M. ROY
Proposition 4.10. If αs = 11/N for s ∈ [t − 1] and Assumption 1.C holds, then
1
i
for all t ≥ 1, i ≥ 1, N ≥ i, y1:t
, . . . , y1:t
∈ Et,
t
i(β − 1)
.
(4.22)
Ei,N
[
Ẑ
/Z
]
≤
1
+
t
t
y 1:i
1:t
N
1
i
If Assumptions 1.A, 1.C and 4.D hold, then for all t ≥ 1, i ≥ 1, N ≥ i, y1:t
, . . . , y1:t
∈
t
E ,
t
iβ
Ei,N
[
Ẑ
/Z
]
≤
1
+
.
(4.23)
t
t
1:i
y 1:t
ζN
Proof. The proof of Eq. (4.22) is a simple generalization of that for [Andrieu et al.,
2013, Proposition 14]. As for Eq. (4.23), observe that for s ∈ [t + 1], G0,t+1 =
G
= G0,s ηs (Gs,t+1 ), so we can write for ℓ ∈ [t], τ ∈ Tt,ℓ,t+1 ,
G0,t+1 G0,t+1
0,s
Zt = G0,t+1 = G0,τk
ℓ−1
Y
ητi (Gτi ,τi+1 ).
(4.24)
i=1
Combined with Assumption 1.C and writing Ḡs,t , supx∈E Gs,t (x),
t+1
X
(ζN )−ℓ+1
X
G0,τ1
τ ∈Tt,ℓ,t+1
ℓ=1
≤ Zt + Zt
t+1
X
ℓ=2
t+1
X
(ζN )−ℓ+1
ℓ−1
i
YX
Gτm ,τm+1 (yτjm )
i=1 j=1
X
τ ∈Tt,ℓ,t+1
ℓ−1 i
G0,τ1 Y X Ḡτi ,τi+1
G0,τ1 i=1 j=1 ητi (Gτi ,τi+1 )
t
= Zt
(ζN )−ℓ+1 (iβ)ℓ−1
ℓ−1
ℓ=1
t
iβ
= Zt 1 +
.
ζN
(4.25)
To compare Eqs. (4.19) and (4.20), consider the Θ(1/N ) terms, which are, respectively,
Qt
Qt
tiZt−1 s=1 gs
ti[Zt−1 s=1 g s − 1]
and
.
N
ζN
Thus, up to a −ti term and a factor of 1/ζ, the two bounds are of the same leading
order in 1/N . The −it is likely an artifact of the analysis while the 1/ζ term
accounts for there being only ζN “effective particles.” The differences between
Eqs. (4.22) and (4.23) are identical.
5. Bounding the Divergence of SMC Samplers
0,N
∗
Recall that P 0,N (dx1:t ) is the distribution of X̃1:t
∼ π1:t
, a single sample from
0,N
0,N
the αSMC estimator of π1:t . Equivalently, P
(dx1:t ) = E0,N [π1:t
](dx1:t ) ,
0,N
0,N 0,N
E [π1:t (dx1:t )] is the expected value of the random measure π1:t . As a first
application of our results from Section 4, we consider bounding the distance between the measures π1:t and P 0,N . That is, for some divergence d(µ||ν) between
measures, can we bound d(π1:t ||P 0,N )? To the best of our knowledge, there has
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
19
been minimal investigation of this question, with [Del Moral, 2004, Chapter 8] a
notable exception. For example, under Assumption SIR, the bound
c
(5.1)
dKL (P 0,N ||πt ) ≤ ,
N
can be extracted as a special case of a more general propagation-of-chaos result
[Del Moral, 2004, Theorem 8.3.2].
Let F1 be the set of functions f : R+ → R that are monotonically increasing or
decreasing and satisfy f (1) = 0. We consider the class of monotonic divergences of
the form
dµ1
di,f (µ1 ||µ2 ) , µi f ◦
, i ∈ {1, 2}, f ∈ F1 .
(5.2)
dµ2
Table 1 lists some common divergences that can be written this way.
The following result characterizes the divergence between π1:t and P 0,N , only
assuming that f is concave.
−1 Pt
Proposition 5.1. Let Rt (y1:t ) , E1,N
y1:t [Ẑt /Zt ] and St , Zt
s=1 G0,s π1:t (Gs,t+1 )−
ζ. If Assumption 4.D holds, then for all concave f ∈ F1 ,
d1,f (π1:t ||P 0,N ) ≤ f (π1:t (Rt )).
In particular, if Assumption SIR holds, then
St − t
+ Θ(N −2 )
dKL (π1:t ||P 0,N ) ≤ log 1 +
N
S
−
t
t
dχ2 (π1:t ||P 0,N ) ≤
+ Θ(N −2 )
N
while if Assumptions 1.A and 4.D hold, then
St − ζ
−2
0,N
+ Θ(N )
dKL (π1:t ||P
) ≤ log 1 +
ζN
St − ζ
dχ2 (π1:t ||P 0,N ) ≤
+ Θ(N −2 ).
ζN
Proof. The general statement follows by applying Jensen’s inequality to Eq. (5.2),
then using Theorem 4.2. The special cases correspond to using the KL divergence
(version 1) and χ2 distance rows of Table 1 and applying Corollary 4.8.
Similar results for (sequential) importance sampling are included in Appendix A.5.
We can also consider the divergence between π1:t and P 0,N when E1,N
y1:t [Ẑt ] is
uniformly bounded:
Proposition 5.2. If Assumption 4.D holds and E1,N
y1:t [Ẑt /Zt ] ≤ Bt,N for all y1:t ∈
E t , then for all increasing f ∈ F1
di,f (π1:t ||P 0,N ) ≤ f (Bt,N )
and for all decreasing f ∈ F1 ,
−1
di,f (P 0,N ||π1:t ) ≤ f (Bt,N
).
20
J. H. HUGGINS AND D. M. ROY
Table 1. Divergences of the form Eq. (5.2). The operator (a)+
gives the positive part of a ∈ R.
Symbol
Name
KL divergence (version 1)
dKL
KL divergence (version 2)
dKL
dχ2
χ2 distance
total variation distance (version 1) dT V
total variation distance (version 2) dT V
i
1
2
1
2
2
f
a 7→ log a
a 7→ − log a
a 7→ a − 1
a 7→ (a − 1)+
a 7→ (1 − a)+
In particular,
dKL (π1:t ||P 0,N ) ≤ log Bt,N
dχ2 (π1:t ||P 0,N ) ≤ Bt,N − 1
Bt,N − 1
≤ Bt,N − 1.
dT V (π1:t , P 0,N ) ≤
Bt,N
Proof. The general statements follow immediately from Eq. (5.2) and Theorem 4.2.
The special cases correspond to using the KL divergence (version 1), χ2 distance,
and total variation distance (version 2) rows of Table 1. The second total variation
inequality holds since Bt,N ≥ 1.
The bounds in Proposition 5.2 for KL divergence, χ2 distance, and total variation distance are asymptotically equivalent if Bt,N → 1 as N → ∞. Combining
Proposition 5.2 with, for example, Proposition 4.10, yields quantitative bounds for
SIR and αSMC:
Corollary 5.3. If Assumptions 1.C and SIR hold, then
t(β − 1)
N
t(β
− 1)
+ O(N −2 )
dχ2 (π1:t ||P 0,N ) ≤
N
t(β − 1)
dT V (π1:t , P 0,N ) ≤
+ O(N −2 ).
N + t(β − 1)
dKL (π1:t ||P 0,N ) ≤
(5.3)
(5.4)
(5.5)
If Assumptions 1.A, 1.C and 4.D hold, then
tβ
ζN
tβ
+ O(N −2 )
dχ2 (π1:t ||P 0,N ) ≤
ζN
tβ
dT V (π1:t , P 0,N ) ≤
+ O(N −2 ).
ζN + tβ
dKL (π1:t ||P 0,N ) ≤
(5.6)
(5.7)
(5.8)
6. The α-Particle Gibbs Sampler
As a second application of our results from Section 4, we consider the mixing
properties of particle Gibbs with adaptive resampling. Recall from Section 1.3 that
we introduce a global parameter θ ∈ Θ with prior distribution ̟(dθ). Replace
Ms by Msθ and gs by gsθ , then parameterize the other quantities defined previously
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
21
in terms of Ms and gs by θ. The target distribution on the product space (Θ ×
Y, B(Θ × Y )), Y , E t , is
π(dθ × dy) , γ(dθ × dy)/Z,
(6.1)
where
γ(dθ × dy) ,
t
Y
gsθ (ys )Msθ (ys−1 , dys )̟(dθ)
and Z , γ(1).
(6.2)
s=1
Particle Gibbs samplers have kernels of the form πy (dθ)Πθ (y, dz), where Πθ (y, dz)
is an SMC-based kernel with invariant distribution πθ . The standard PG sampler
employs the iterated conditional SMC (i-cSMC) kernel [Andrieu et al., 2013]: that
1,N
is, Πθ = Pθ,y
and require Assumption SIR to hold.
We now introduce the novel α-particle Gibbs (αPG) sampler, which employs the
1,N
1,N
iterated conditional αSMC (i-cαSMC) kernel Pθ,y
, so Πθ = Pθ,y
. In Appendix A.6
we prove that the i-cαSMC kernel is reversible with respect to πθ and hence has
invariant distribution πθ .
The first step to proving mixing results for the i-cαSMC kernel and the αPG
sampler is to use Theorem 4.2 to obtain a sufficient condition for the i-cαSMC
transition kernel to satisfy a minorization condition.
Proposition 6.1. If Assumption 4.D holds and E2,N
θ,y 1:2 [Ẑt /Zt ] ≤ Bt,N for all θ ∈ Θ
and y 1 , y 2 ∈ Y , then
1,N
Pθ,y
(y, dx) ≥ εt,N πθ (dx),
where εt,N ,
(6.3)
(1−κ′N )t
Bt,N .
The constant ǫt,N , which determines mixing speed, can be found explicitly using
the quantitative bounds from Section 4.3. For example, using Assumption 1.C we
obtain the following:
Corollary 6.2. If Assumptions 1.A, 1.C and 4.D hold, then for all y ∈ Y ,
1,N
Pθ,y
(y, dx) ≥ εt,N πθ (dx),
(6.4)
where
εt,N ,
Furthermore, if ζN ≥
2βt
C(1−κ′N )−κ′N t
1 − κ′N
2β
1 + ζN
t
.
(6.5)
for some constant C > 0, then
ǫt,N ≥ e−C .
(6.6)
In particular, assuming κ′N ≤ B/N for some constant B ≥ 1, if N ≥ t/C + B, then
C
ǫt,N ≥ exp − (2β + ζB) .
ζ
(6.7)
22
J. H. HUGGINS AND D. M. ROY
Proof. The first part follows from Propositions 6.1 and 4.10. For the second part,
we then have
−t
2β −t
1 + ζN
1
2β
′
= 1+
+ κN
(6.8)
ǫt,N ≥
1 − κ′N
1 − κ′N ζN
−t
C
≥ e−C .
(6.9)
≥ 1+
t
The final part follows after noting that if κ′N ≤ B/N , then
1
1
2β
2β
2β
1
′
+ κN ≤
+ B/N =
+B .
1 − κ′N ζN
1 − B/N ζN
N −B ζ
(6.10)
Remark 6.3. In the case of the i-cSMC kernel, Corollary 6.2 is almost as tight as
[Andrieu et al., 2013, Corollary 14]: the former result replaces β − 1 with β.
▽
The minorization condition Eq. (6.3) implies uniform ergodicity and a number
of other types of convergence guarantees for the i-cSMC process. The following
generalizes [Andrieu et al., 2013, Theorem 1], which applies only to the i-cSMC
kernel and the PG sampler.
Theorem 6.4. Assume that Assumptions 1.A and 4.D hold.
I. Let N ≥ 2, and consider the i-cαSMC process with kernel P = Pθ1,N .
1. P is reversible with respect to π and defines a positive operator,
2. If the potentials are bounded then there exists ǫt,N = 1 − Ct /N such that
(i) for all y ∈ Y , P (y, dz) ≥ ǫt,N πθ (dz),
(ii) for every measure ν ≪ πθ and k ≥ 1,
dχ2 (νP k ||πθ ) ≤ dχ2 (ν||πθ )(1 − ǫt,N )k ,
(6.11)
(iii) for every y ∈ Y and k ≥ 1,
dT V (δy P k , πθ ) ≤ (1 − ǫt,N )k ,
(6.12)
ǫt,N ≥ εB,C,ζ > 0.
(6.13)
3. If Assumption 1.C also holds and there is a constant B > 0 such that
κ′N ≤ B/N , then for every C > 0, there exists εB,C,ζ > 0 such that for
N ≥ t/C + B and all t > 1,
II. If there exists β ≥ 1 such that, for all t, s ∈ N,
π- ess sup
θ,x
Gθ0,t Gθt,t+s (x)
≤ β,
Gθ0,t+s
(6.14)
or if
π- ess sup
θ
Qt
g θs
< ∞,
γθ (1)
s=1
(6.15)
then the αPG chain is geometrically ergodic whenever the Gibbs sampler is
geometrically ergodic.
Proof. Part I.1 follows from Lemma A.4. Parts I.2-3 follow from Proposition 6.1,
Corollary 6.2, and [Andrieu et al., 2013, Proposition 31]. Part II follows from
[Andrieu et al., 2013, Section 7].
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
23
Remark 6.5. Part I.3 means that if Assumption 1.C holds, then scaling N linearly
with t ensures a uniform convergence rate, as measured by χ2 -divergence or total
variation distance.
▽
Appendix A. Additional Proofs
A.1. Technical Lemma.
Lemma A.1. Let X and Y be random elements in Borel spaces (S, S) and (T, T ),
respectively, let ψ : S × T → R+ be a measurable, and let µ be the distribution of
X. If
then ν ≪ µ and
ν = E[ψ(X, Y )δX ],
(A.1)
dν
(X) = E[ψ(X, Y ) | X] a.s.
dµ
(A.2)
Proof. Because S is Borel, there exists an f satisfying f (X) = E[ψ(X, Y ) | X] a.s.
It follows from the chain rule of conditional expectation and then some elementary
manipulations that, for all A ∈ S,
Z
f (x)µ(dx),
ν(A) = E[f (X)δX (A)] = E[f (X)1A (X)] =
A
and so f is a version of the Radon–Nikodym derivative dν/dµ.
A.2. Proof of Theorem 4.2, i ≥ 1 case. First observe that we can write the
ci αSMC kernel as
h
i
i,N
δX̃ ∗ (dx1:t )
(A.3)
P i,N (y 1:i
1:t , dx1:t ) = Ey 1:i
1:t
1:t
hP
i
1:N
1:N
= Ei,N
(A.4)
k1:t ∈[N ]t Ik1:t (X 1:t , A1:t , dx1:t ) ,
y 1:i
1:t
where
∗
1:N
Ik1:t (x1:N
1:t , a1:t , dx1:t ) , δx̃kt (dx1:t )1(kt = at )
1:t
t−1
Y
1(ks = aks s+1 ).
(A.5)
s=1
Next note that
N
X
k1 =1
=
1:i
1:i
t
1(x̃k1:t
∈ S)Pi,N
[X11:N ∈ dx1:N
1 , F1 = f1 ]
y 1:i
N
X
k1 =1
≥
(A.6)
1:t
t
1(x̃k1:t
∈ S)C1i D(f11:i )
N Z
N C1i X
C1i+1 k =1 E
1
N
i
Y
Y
1
fj
M1 (dxn1 )
δyj (dx11 )
1
N
1:i
j=1
t
1(x̃k1:t
∈ S)C1i+1 D(f11:i , k1 )
×
N
Y
n∈f
/ 11:i ,k1
(A.7)
n∈f
/ 1
i
Y
j=1
M1 (dxn1 )M1 (dx1 )
1
fj 1
δyj (dx11 ) δx1 (dxk11 )
1
N
N
(A.8)
24
J. H. HUGGINS AND D. M. ROY
≥
N Z
N C1i X
i+1
1:i
1:i
t
1(x̃k1:t
[X11:N ∈ dx1:N
∈ S)Pi+1,N
= k1 ]M1 (dx1 )
1 , F1 = f1 , F1
y 1:i
1:t ,x1:t
C1i+1 k =1 E
1
(A.9)
For the remainder of the proof, to keep notation compact when writing laws, instead of writing, e.g., Xs1:N ∈ x1:N
or Fs1:i = fs1:i , whenever a random variable is
s
instantiated to be (the differential) of the lowercase version of itself, we will write
only the random variable: for example, Xs1:N or Fs1:i . Now, for s = 2, . . . , t,
N
X
ks =1
1:N
1:N
1:i
t
s
1(x̃k1:t
∈ S)1(ks−1 = aks−1
)Pi,N
[Xs1:N , As−1
, Fs1:i | X 1:N
]
1:s−1 , A1:s−2 , Fs−1(A.10)
y1:i
1:t
N
X
,
ks =1
t
s
1(x̃k1:t
∈ S)1(ks−1 = aks−1
)Csi D(fs1:i )
×
≥
Y
n∈f
/ s1:i
N
X
i
Y
fjfj
fj
fj
j
s s−1
s
= fs−1
)
αs−1
δysj (dxss )1(as−1
j=1
an
s−1
n
1:N
, x1:N
rn (ans−1 |ws−1
1:s−1 )Ms (xs−1 , xs )
Csi
k k
s s−1
i+1
ks =1 Cs αs−1
(A.11)
Z
E
t
∈ S)Csi+1 D(fs1:i , ks )
1(x̃k1:t
i
Y
fjfj
fj
fj
j
s s−1
s
= fs−1
)
αs−1
δysj (dxss )1(as−1
j=1
k k
s s−1
s
)
δxs (dxks s )1(ks−1 = aks−1
× αs−1
×
Y
n∈f
/ s1:i ,ks
(A.12)
an
s−1
1:N
1:N
rn (ans−1 |ws−1
, x1:s−1
)Ms (xs−1
, xns )
k
s−1
1:N
, x1:N
× rks (ks−1 |ws−1
1:s−1 )Ms (xs−1 , dxs )
=
N
X
Csi
i+1 ks ks−1
ks =1 Cs αs−1
Z
E
k
s−1
1:N
t
1(x̃k1:t
, x1:N
∈ S)rks (ks−1 |ws−1
1:s−1 )Ms (xs−1 , dxs ) (A.13)
1:N
i+1
1:i
i+1
1:i
[Xs1:N , A1:N
= ks | X 1:N
Pi+1,N
s−1 , Fs , Fs
1:s−1 , A1:s−2 , Fs−1 , Fs−1 = ks−1 ]
y1:i ,x1:t
1:t
Using Eqs. (A.9) and (A.14), we have (note that the terms such as those involving
a0 should be ignored)
X
k1:t ∈[N ]t
t
1(x̃k1:t
∈ S)1(kt = a∗t )
t
Y
s=2
1:N
1:i
s
1(ks−1 = aks−1
)Pi,N
[X 1:N
1:t , A1:t , F 1:t ]
y 1:i
1:t
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
X
=
k1:t ∈[N ]t
i,N
1:N
1:N
t
1(x̃k1:t
∈ S)1(kt = a∗t )Pi,N
[A∗t | X 1:N
, F11:i ]
1:t , A1:t−1 ]Py 1:i [X1
y1:i
1:t
t
Y
×
≥
X
k1:t ∈[N ]t
s=2
Z
×
×
Et
1:t
1:N
1:N
1:i
1:i
s
[Xs1:N , A1:N
1(ks−1 = aks−1
)Pi,N
s−1 , Fs | X 1:s−1 , A1:s−2 , Fs−1 ]
y1:i
1:t
(A.14)
t
N 1(x̃k1:t
∈ S)1(kt = a∗t )
Qt
Qt−1 ks ks−1
( s=1 Csi+1 /Csi )( s=1 αs−1
)
t
Y
s=2
t
Y
s=2
k
a∗
a∗
wt t gt (xt t )
PN
n
n
n=1 wt gt (xt )
s−1
1:N
, x1:N
rks (ks−1 |ws−1
1:s−1 )Ms (xs−1 , dxs )
1:t
1:N
1:i
i+1
[X 1:N
Pi+1,N
1:t , A1:t−1 , F 1:t , F1:t
y1:i
1:t ,x1:t
Z
X
Zt
S k ∈[N ]t
1:t
(A.15)
1:N
i+1
1:i
i+1
1:i
Pi+1,N
[Xs1:N , A1:N
= ks | X 1:N
s−1 , Fs , Fs
1:s−1 , A1:s−2 , Fs−1 , Fs−1 = ks−1 ]
y1:i ,x1:t
i+1
1:i
1:i
[X11:N ∈ dx1:N
× Pi+1,N
= k1 ]M1 (dx1 )
1 , F1 = f1 , F1
y1:i
1:t ,x1:t
X Z
N 1(x1,t ∈ S)
=
Qt
P
i+1
n
n
t
( s=1 Cs /Csi ) N
n=1 wt gt (xt )
k1:t ∈[N ]t E
=
25
Ẑt
Qt
i+1
i
s=1 Cs /Cs
(A.16)
= k1:t ]γ1:t (dx1:t )
1:N
1:i
i+1
Pi+1,N
[X 1:N
1:t , A1:t−1 , F 1:t , F1:t = k1:t ]π1:t (dx1:t ),
y1:i ,x1:t
1:t
(A.17)
from which Eq. (4.2) follows.
To prove Eq. (4.4), first note that under Assumption 4.D, the normalization
constants for the c2 αSMC process are given by
C12 ,
and, for s = 2, . . . , t,
Cs2
Thus, Cs2 ≤
1
1−κN
,
1−
N
X
N
N −1
(A.18)
1
2
kfs−1
kfs−1
αs−1
αs−1
k=1
for s = 2, . . . , t and hence Cs2 ≤
!−1
1
1−κ′N
.
(A.19)
for all s ∈ [t].
A.3. Proof of Proposition 4.6. For (1), the fact that ESS1 (w1:N ) = ESSent (w1:N )
is a straightforward algebraic manipulation. To prove the limit equality, observe
that, using the Taylor series for xp and log(1 + x), we have
!1/(1−p)
kw1:N kp1
(A.20)
lim PN
n p
p→1
n=1 (w )
!1/(1−p)
P∞
1:N
k1 (p − 1)k logk (kw1:N k1 )/k!
k=0 kw
= lim
(A.21)
PN P∞
k
n
k
n
p→1
n=1
k=0 w (p − 1) log (w )/k!
26
J. H. HUGGINS AND D. M. ROY
= lim
x→∞
P∞
1:N
k1 x−k logk (kw1:N k1 )/k!
k=0 kw
PN P∞
n −k logk (wn )/k!
n=1
k=0 w x
!x
(A.22)
!x
P
k
−k
1:N
exp(log(1 + ∞
x
log
(kw
k
)/k!))
1
k=1
(A.23)
= lim
P∞ PN
−k logk (wn )/k!))
x→∞
exp(log(1 + k=1 n=1 wn kw1:N k−1
1 x
P∞
P∞
exp(x m=1 (−1)m+1 [ k=1 x−k logk (kw1:N k1 )/k!]m )
= lim
PN
P
P∞
n
1:N k−1 x−k logk (wn )/k!]m )
m+1 [ ∞
x→∞ exp(x
1
n=1 w kw
k=1
m=1 (−1)
(A.24)
exp(log(kw1:N k1 ) + Θ(x−1 ))
Q
N
x→∞
n )w n /kw 1:N k1 + Θ(x−1 )
exp log
(w
n=1
= lim
= QN
kw1:N k1
n=1 (w
n )w n /kw 1:N k1
.
(A.25)
(A.26)
To prove the remaining parts, we make repeated use of the following:
1:N
Fact. For 1 ≤ r < s ≤ ∞, and any vector w1:N ∈ RN
ks ≤ kw1:N kr ≤
+ , kw
1/r−1/s
1:N
N
kw ks , with the lower (upper) bound achieved if and only if w1:N has
one non-zero entry (w1:N has all equal entries).
For (2), apply the Fact with r = 1, s = p > 1, and note that in this case
1/r − 1/s = 1 − 1/p = 1/p∗ . We then have 1 ≤ kw1:N k1 /kw1:N kp ≤ N 1/p∗ , proving
the result for p > 1. For p = 1, the result follows from part (1) and elementary
properties of the entropy.
For (3), in the case that p > 1, note that
kw1:N kq1∗ −p∗ ≥ N (q∗ −p∗ )/q∗ kw1:N kqq∗ −p∗ = N 1−p∗ /q∗ kw1:N kqq∗ −p∗
= N −p∗ (1/p−1/q) kw1:N kqq∗ −p∗ ,
(A.27)
where the final equality follows since
1 − p∗ /q∗ = 1 − p∗ (1 − 1/q) = 1 − p∗ + p∗ /q = −p∗ /p + p∗ /q.
(A.28)
We conclude that
1:N p∗
kw1:N kp1∗
kw k1
≥
kw1:N kp
N p∗ (1/p−1/q) kw1:N kpq ∗
(A.29)
≥
(A.30)
kw1:N kp1∗
kw1:N kq1∗ −p∗
p
∗
N p∗ (1/p−1/q) kw1:N kq N −p∗ (1/p−1/q) kvkqq∗ −p∗
1:N q∗
kw k1
=
kw1:N kq
1:N q∗
kw k1
≥
kw1:N kp
1:N p∗ −q∗ 1:N p∗
kw k1
kw kp
=
kw1:N k1
kw1:N kp
1:N p∗
kw k1
,
≥ N −(p∗ −q∗ )/p∗
kw1:N kp
(A.31)
(A.32)
(A.33)
(A.34)
where the first, third, and fourth inequalities follow from the Fact and the second
follows from Eq. (A.27).
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
27
The case of p = 1 follows from the p > 1 case and part (1).
A.4. Proof of Theorem 4.7. We prove the result for i = 1. The general case
follows from straightforward modifications.
For t ≥ 1, let Qt (xt−1 , dxt ) , gt−1 (xt−1 )Mt (xt−1 , dxt ), and for 0 ≤ s < t, let
Qs,t , Qs+1 Qs+2 · · · Qt ,
(A.35)
so Qt,t+1 = Qt . By convention Qt,t (xt , dyt ) = δxt (dyt ) and Q0,t (dxt ) is a measure,
not a probability kernel. Notice that for s ∈ [t], xs ∈ E, and φt : E → R,
Qs,t (xs )(φt ) = E[φt (ξt )gs:t−1 (ξs:t−1 ) | ξs = xs ]
(A.36)
and Q0,t (φt ) = M1 Q1,t (φt ). Generalizing these identities, we will abuse notation
and write, for s ∈ [t], xs ∈ E, and φs,t : E t−s → R,
Qs,t (xs )(φs,t ) , E[φs:t (ξs:t )gs:t−1 (ξs:t−1 ) | ξs = xs ]
(A.37)
and Q0,t (φ1:t ) , M1 Q1,t (φ1:t ). Note that Gs,t (y) = Qs,t (y)(1) for s ∈ [t − 1] and
G0,t = Q0,t (1).
We will use the abbreviated notation Qks,t (·) = Qs,t (·)(Xsk ) or Qs,t (·)(xks ), Gks,t =
Gs,t (Xsk ) or Gs,t (xks ), Gys,t = Gs,t (ys ), gsk = gs (Xsk ) or gs (xks ), and gsy = gs (ys ). The
variables are Xsk inside expectations and xks outside expectations. Throughout the
proof, when limits of a sum are not specified, the sum is from 1 to N .
1:N
1:N
1
, where by conven, and F1:s
, A1:s−1
Let Fs be the σ-algebra generated by X 1:s
tion we let F0 be the trivial σ-algebra. The proof relies on the following lemma.
Lemma A.2. If y1:t ∈ E t , then
(1) for s = 2, . . . , t and any functions φns : E → R, n ∈ [N ],
#
"
X
1,N
n
n
Ey1:t
φs (Xs ) | Fs−1
n
=
X
f 1f 1
f1
s s−1
αs−1
φss (ys ) +
X X X
fs1 n6=fs1
fs1
f 1f 1
s s−1
αs−1
k
(A.38)
k
αnk
s−1 ws−1 k
Qs−1,s (φns );
n
ws
(2) for τ ∈ [t − s],
"
#
X
X
1 X n n
1,N
n n
n
Ey1:t
Ws Gs,s+τ | Fs−1 ≤
ws−1 gs−1 Gys,s+τ +
ws−1
Gns−1,s+τ ;
ζN
n
n
n
(A.39)
and
(3) for s = 1, . . . , t − 1,
where,
At−s , (ζN )−s+1
X
n
Bt−s , (ζN )−s+1
X
n
i
h
N E1,N
y1:t Ẑt | Ft−s ≤ At−s + Bt−s ,
n
n
wt−s
gt−s
n
wt−s
s
X
X
ℓ=1 τ ∈Tt,ℓ,s
s
X
X
ℓ=1 τ ∈Tt,ℓ,s
(A.40)
(ζN )s−1−ℓ Gyt−s+1,τ1 Cℓy (τ ) (A.41)
(ζN )s−ℓ Gnt−s,τ1 Cℓy (τ ) .
(A.42)
28
J. H. HUGGINS AND D. M. ROY
Proof. For (1),
"
#
X
1,N
n
n
Ey1:t
φs (Xs ) | Fs−1
n
=
fs1
=
1
fs1 fs−1
αs−1
XX
k6=fs1
−f 1
as−1s
XX
fs1
1
fs1 fs−1
αs−1
−f 1
as−1s
=
X
=
X
Y
1,N
1:N
rk (aks−1 |ws−1
, x1:N
1:s−1 )Ey1:t
"
X
n
φns (Xsn ) | Fs−1 , A1:N
s−1
=
1
a1:N
s−1 , Fs
kak
ak
ak
s−1
s−1
s−1
i
h
Y αs−1
X
n
1
ws−1
gs−1 (xs−1
)
as−1
φfss (ys ) +
E φns (ξs ) | ξs−1 = xs−1
k
w
s
1
1
k6=fs
n6=fs
f 1f 1
f 1f 1
f1
X X X
s s−1
αs−1
f 1f 1
f1
X X X
s s−1
αs−1
s s−1
αs−1
φss (ys ) +
fs1 n6=fs1
fs1
s s−1
αs−1
φss (ys ) +
fs1 n6=fs1
fs1
φns (x)
k
f 1f 1
k
wsn Gs,s+τ (x),
For (2), choosing
=
"
#
X
E1,N
Wsn Gns,s+τ | Fs−1
y1:t
k
k
αnk
s−1 ws−1 gs−1 (xs−1 ) n
E φs (ξs ) | ξs−1 = xks−1
n
ws
k
αnk
s−1 ws−1 k
Qs−1,s (φns )
wsn
we have
(A.43)
n
=
X
f 1f 1
f1
s s−1
αs−1
ws s Gys,s+τ +
=
X
1
fs1 fs−1
f1
αs−1
ws s
+
≤ Gys,s+τ
X
fs1
f 1f 1
s s−1
αs−1
f 1f 1
s s−1
αs−1
k
X X X
fs1 n6=fs1
fs1
=
X X X
fs1 n6=fs1
fs1
Gys,s+τ
k
αnk
s−1 ws−1 k
Qs−1,s (wsn Gs,s+τ )
wsn
(A.44)
1
fs1 fs−1
k
k
αs−1
αnk
s−1 ws−1 Gs−1,s+τ
(A.45)
k
1
kws1:N k1 X X X fs1 fs−1
k
k
αs−1 αnk
+
s−1 ws−1 Gs−1,s+τ
ζN
1
n
fs
(A.46)
k
Gys,s+τ X n X X nk k
ws +
αs−1 ws−1 Gks−1,s+τ
ζN
n
n
(A.47)
k
X
Gys,s+τ X X nk k
k
k
αs−1 ws−1 gs−1
+
ws−1
Gks−1,s+τ
ζN
n
k
k
X
1 X k
y
k
k
ws−1 gs−1
Gs,s+τ +
ws−1
Gks−1,s+τ ,
=
ζN
=
k
=
fs1
(A.48)
(A.49)
k
where the inequality follows from Assumption 1.A, and we have repeatedly used
Assumption 4.D.
To show (3), we start by using (2) with s = t and τ = 1:
"
#
X
1 X k k y ζN X m m
1,N
n n
wt−1 gt−1 gt +
w G
(A.50)
Ey1:t
Wt gt | Ft−1 =
ζN
ζN m t−1 t−1,t+1
n
k
= At−1 + Bt−1 ,
(A.51)
Hence, (3) holds for s = 1. We now assume that the bound holds for some
s ∈ {1, . . . , t − 2} and establish that it also holds for s + 1. Using the inductive
#
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
29
hypothesis,
i
i
h
h
i
h
1,N
1,N
N E1,N
y1:t Ẑt | Ft−s−1 = Ey1:t N Ey1:t Ẑt | Ft−s | Ft−s−1
(A.52)
≤ E1,N
y1:t [At−s + Bt−s | Ft−s−1 ] .
(A.53)
Using (2), we have
(A.54)
A , E1,N
y1:t [At−s | Ft−s−1 ]
"
#
s
X
X
X
y
−s+1
s−1−ℓ y
1,N
n
n
= (ζN )
(ζN )
Gt−s+1,τ1 Cℓ (τ ) Ey1:t
Wt−s gt−s
n
ℓ=1 τ ∈Tt,ℓ,s
(A.55)
≤ (ζN )−s
×
X
s
X
X
ℓ=1 τ ∈Tt,ℓ,s
(ζN )s−1−ℓ Gyt−s+1,τ1 Cℓy (τ )
y
n
n
gt−s−1
gt−s
wt−s−1
+ ζN
n
X
n
wt−s−1
Gnt−s−1,t−s+1
n
(A.56)
!
and
(A.57)
B , E1,N
y1:t [Bt−s | Ft−s−1 ]
#
"
s
X
X
X
−s+1
n
= (ζN )
(ζN )s−ℓ E1,N
Wt−s
Gnt−s,τ1 | Ft−s−1 Cℓy (τ )
y1:t
n
ℓ=1 τ ∈Tt,ℓ,s
(A.58)
≤ (ζN )−s
s
X
X
(ζN )s−ℓ
X
n
n
wt−s−1
gt−s−1
Gyt−s,τ1
n
ℓ=1 τ ∈Tt,ℓ,s
+ ζN
X
n
wt−s−1
Gnt−s−1,τ1
n
(A.59)
Hence,
A + B ≤ (ζN )−s
X
n
+ (ζN )−s
s
X
X
n
n
wt−s−1
gt−s−1
(ζN )s−1−ℓ Gyt−s,t−s+1 Gyt−s+1,τ1 Cℓy (τ )
ℓ=1 τ ∈Tt,ℓ,s
X
n
+ (ζN )−s
X
n
+ (ζN )−s
X
n
s
X
X
n
n
(ζN )s−ℓ Gyt−s,τ1 Cℓy (τ )
wt−s−1
gt−s−1
ℓ=1 τ ∈Tt,ℓ,s
n
wt−s−1
n
wt−s−1
s
X
X
ℓ=1 τ ∈Tt,ℓ,s
s
X
X
ℓ=1 τ ∈Tt,ℓ,s
(ζN )s−ℓ Gnt−s−1,t−s+1 Gyt−s+1,τ1 Cℓy (τ )
(ζN )s−ℓ+1 Gnt−s−1,τ1 Cℓy (τ ) .
(A.60)
!
Cℓy (τ ) .
30
J. H. HUGGINS AND D. M. ROY
Summing the parenthesized double sums of the first two terms yields
s
s
X
X
X
X
(ζN )s−ℓ Gyt−s,τ1 Cℓy (τ )
(ζN )s−1−ℓ Gyt−s,t−s+1 Gyt−s+1,τ1 Cℓy (τ ) +
ℓ=1 τ ∈Tt,ℓ,s
ℓ=1 τ ∈Tt,ℓ,s
=
s+1
X
X
(ζN )s−ℓ Gyt−s,τ1 Cℓy (τ ) +
ℓ=1 τ ∈Tt,ℓ,s+1
τ1 =t−s+1
s+1
X
(ζN )s−ℓ Gyt−s,τ1 Cℓy (τ )
X
ℓ=1 τ ∈Tt,ℓ,s+1
τ1 >t−s+1
(A.61)
=
s+1
X
X
(ζN )s−ℓ Gyt−s,τ1 Cℓy (τ ),
(A.62)
ℓ=1 τ ∈Tt,ℓ,s+1
so the first two terms are equal to At−(s+1) . Summing the parenthesized double
sums of the last two terms yields
s
s
X
X
X
X
(ζN )s−ℓ+1 Gnt−s−1,τ1 Cℓy (τ )
(ζN )s−ℓ Gnt−s−1,t−s+1 Gyt−s+1,τ1 Cℓy (τ ) +
ℓ=1 τ ∈Tt,ℓ,s
ℓ=1 τ ∈Tt,ℓ,s
=
s+1
X
X
(ζN )s−ℓ+1 Gnt−s−1,τ1 Cℓy (τ ) +
ℓ=1 τ ∈Tt,ℓ,s+1
τ1 =t−s+1
s+1
X
(ζN )s−ℓ+1 Gnt−s−1,τ1 Cℓy (τ )
X
ℓ=1 τ ∈Tt,ℓ,s+1
τ1 >t−s+1
(A.63)
=
s+1
X
X
(ζN )s−ℓ+1 Gnt−s−1,τ1 Cℓy (τ ),
(A.64)
ℓ=1 τ ∈Tt,ℓ,s+1
so the last two terms are equal to Bt−(s+1) .
Using part (3) of Lemma A.2 with s = t − 1, we have
h i
1,N
N E1,N
y1:t Ẑt ≤ Ey1:t [A1 + B1 ] .
(A.65)
Therefore,
−t+2 1,N
Ey1:t
E1,N
y1:t [A1 ] = (ζN )
"
X
n
# t−1
X
g1n
X
ℓ=1 τ ∈Tt,ℓ,t−1
t−1
X
= (ζN )−t+2 (Gy1,2 + (N − 1)G0,2 )
(ζN )t−2−ℓ Gy2,τ1 Cℓy (τ )
X
′
ℓ=1 τ ∈Tℓ,t−1
(A.66)
(ζN )t−2−ℓ Gy2,τ1 Cℓy (τ )
(A.67)
and
t−1
X
−t+2
[B
]
=
(ζN
)
E1,N
1
y
1:t
#
Gn1,τ1 Cℓy (τ ) (A.68)
(ζN )t−1−ℓ E1,N
y1:t
X
(ζN )t−1−ℓ (Gy1,τ1 + (N − 1)G0,τ1 )Cℓy (τ ) .
ℓ=1 τ ∈Tt,ℓ,t−1
t−1
X
= (ζN )−t+2
"
X
X
ℓ=1 τ ∈Tt,ℓ,t−1
n
(A.69)
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
31
Hence, using arguments analogous to those from the proof of Lemma A.2 and the
fact that G0,1 = 1 yields
E1,N
y1:t [A1 + B1 ]
= (ζN )−t+2
t
X
X
(ζN )t−1−ℓ Gy1,τ1 Cℓy (τ )
(A.70)
ℓ=1 τ ∈Tt,ℓ,t
+
t
X
X
N −1
(ζN )t−ℓ G0,τ1 Cℓy (τ )
(ζN )−t+2
ζN
(A.71)
ℓ=1 τ ∈Tt,ℓ,t
= (ζN )−t+2
t+1
X
X
(ζN )t−ℓ G0,τ1 Cℓy (τ )
ℓ=1 τ ∈Tt,ℓ,t+1
τ1 =1
(A.72)
t+1
X
N −1
(ζN )−t+2
+
ζN
X
(ζN )
G0,τ1 Cℓy (τ )
1(τ1 >1)
t−ℓ
ℓ=1 τ ∈Tt,ℓ,t+1
τ1 >1
−t+2
= (ζN )
t+1
X
X
t−ℓ
(ζN )
ℓ=1 τ ∈Tt,ℓ,t+1
N −1
ζN
G0,τ1 Cℓy (τ ).
(A.73)
A.5. Divergence of importance samplers. The key quantity in this section is
the variance of the potentials:
h
2 i
Vt , V Zt−1 g1:t (ξ1:t ) = E Zt−1 g1:t (ξ1:t ) − 1
(A.74)
Theorem A.3. If αs = I N for all s ∈ [t − 1], then
Vt
dKL (π1:t ||P
) ≤ log 1 +
N
Vt
dχ2 (π1:t ||P 0,N ) ≤ .
N
0,N
(A.75)
(A.76)
Proof. By Theorem 4.2 and Jensen’s inequality
dP 0,N
Zt
(x1:t ) = E1,N
x1:t
dπ1:t
Ẑt
≥
PN
E1,N
x1:t [
k=1
(A.77)
N
k )]
Zt−1 g1:t (X̃1:t
N
.
=
N − 1 + Zt−1 g1:t (x1:t )
(A.78)
(A.79)
32
J. H. HUGGINS AND D. M. ROY
By definition of the χ2 divergence,
dπ1:t
−1
dP 0,N
dπ1:t dπ1:t
= M1:t
−1
dP 0,N dM1:t
N − 1 + Zt−1 g1:t −1
Zt g1:t − 1
≤ M1:t
N
S,N
dχ2 (π1:t ||π̄1,t
) = π1:t
M1:t (Zt−1 g1:t )2 − 1
N
V[Zt−1 g1:t (ξ1:t )]
.
=
N
=
(A.80)
(A.81)
(A.82)
(A.83)
(A.84)
The bound of the KL divergence follows from the elementary inequality dKL (µ||ν) ≤
log(1 + dχ2 (µ||ν)).
A.6. Invariant distribution of the i-cαSMC kernel.
1,N
Lemma A.4. Pθ,y
(dz) is reversible with respect to πθ (dz).
Proof. We mostly suppress dependence on θ since θ is fixed. We will show that the
cαSMC kernel is Gibbs sampler for the artificial joint density given in Eq. (4.7),
which we recall is
f1
1:N
1:N
1
1
1:N
t
π̃1:t (x1:N
1:t , a1:t−1 , f1:t ) , π1:t (x1:t )ψ̃(x1:t , a1:t−1 , f1:t ).
(A.85)
1:N
1
In particular, letting ω1:t , (x1:N
1:t , a1:t−1 , f1:t ), by definition,
f1
t
Pi,N
[dω 1:t ] = π̃1:t (dω 1:t | x̃1:t
= y1:t ).
y 1:i
(A.86)
1:t
Furthermore,
f1
a∗
t
π̃1:t (x̃1:t
= x̃1:tt | ω 1:t ) = π̃1:t (ft1 = a∗t | ω1:t )
∗
a∗
a∗
a∗
a∗
a∗ Q
a∗
s−1
s as−1
, xs s−1 )gs (xs s )αs−1
M1 (x1 1 )g1 (x1 1 ) ts=2 Is Ms (xs−1
∝
∗
∗
a∗ Qt
1:N , x1:N )M (xas−1 , xas−1 )
M1 (x1 1 ) s=2 ra∗s (a∗s−1 |ws−1
s
s s−1
1:s−1
a∗
a∗
= wt t gt (xt t )
1:N
1
∝ Pi,N
[a∗ | x1:N
1:t , a1:t−1 , f1:t ].
y1:i t
1:t
Reversibility now follows easily:
Z
1,N
Py (dz)π(dy) = π̃1:t (dz | ω 1:t )π̃1:t (dω 1:t | y)π(dy)
Z
= π̃1:t (dz | ω 1:t )π̃1:t (dy | ω1:t )π̃1:t (dω 1:t )
Z
= π̃1:t (dy | ω1:t )π̃1:t (dω 1:t | z)π(dz)
= Pz1,N (dy)π(dz).
SEQUENTIAL MONTE CARLO AS APPROXIMATE SAMPLINGS
33
Acknowledgments
The authors would like to thank Arnaud Doucet for critical feedback and numerous helpful suggestions; Cameron Freer for feedback on early versions of this
work; Josh Tenenbaum for discussions that helped to inspire this work; Vikash
Mansinghka for suggesting we investigate the expected value of SMC estimators;
and an anonymous referee for suggestions on improving the presentation of the results. JHH was supported by the U.S. Government under FA9550-11-C-0028 and
awarded by the DoD, Air Force Office of Scientific Research, National Defense Science and Engineering Graduate (NDSEG) Fellowship, 32 CFR 168a. This research
was carried out in part while DMR held a Research Fellowship at Emmanuel College, Cambridge, with funding also from a Newton International Fellowship through
the Royal Society, an NSERC Discovery Grant, Connaught Award, and U.S. Air
Force Office of Scientific Research grant #FA9550-15-1-0074.
References
Christophe Andrieu and G O Roberts. The pseudo-marginal approach for efficient
Monte Carlo computations. The Annals of Statistics, 37(2):697–725, April 2009.
Christophe Andrieu and Matti Vihola. Convergence properties of pseudo-marginal
Markov chain Monte Carlo algorithms. The Annals of Applied Probability, 2014.
Christophe Andrieu, Arnaud Doucet, and Roman Holenstein. Particle Markov
chain Monte Carlo methods. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 72(3):269–342, 2010.
Christophe Andrieu, Anthony Lee, and Matti Vihola. Uniform Ergodicity of the
Iterated Conditional SMC and Geometric Ergodicity of Particle Gibbs samplers.
arXiv.org, December 2013.
Nicolas Chopin and Sumeetpal S Singh. On the particle Gibbs sampler. arXiv.org,
April 2013.
Julien Cornebise, E Moulines, and Jimmy Olsson. Adaptive methods for sequential importance sampling with application to state space models. Statistics and
Computing, 18(4):461–480, 2008.
Pierre Del Moral. Feynman-Kac Formulae: Genealogical and Interacting Particles
Systems with Applications. Springer, New York, 2004.
Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential Monte Carlo samplers. Journal of the Royal Statistical Society: Series B (Statistical Methodology),
68(3):411–436, 2006.
Arnaud Doucet and Adam M Johansen. A tutorial on particle filtering and smoothing: fifteen years later. In D Crisan and B Rozovsky, editors, Handbook of Nonlinear Filtering. Cambridge University Press, Cambridge, 2010.
Arnaud Doucet, Simon J Godsill, and Christophe Andrieu. On sequential Monte
Carlo sampling methods for Bayesian filtering. Statistics and Computing, 10(3):
197–208, 2000.
Arnaud Doucet, Nando de Freitas, and Neil Gordon. Sequential Monte Carlo in
Practice. Springer, New York, 2001.
Neil J Gordon, David J Salmond, and Adrian FM Smith. Novel approach to
nonlinear/non-Gaussian Bayesian state estimation. IEE Proceedings F (Radar
and Signal Processing), 140(2):107–113, 1993.
Roman Holenstein. Particle Markov Chain Monte Carlo. PhD thesis, University
Of British Columbia, Vancouver, 2009.
34
J. H. HUGGINS AND D. M. ROY
Nicholas Kantas, Arnaud Doucet, Sumeetpal Sindhu Singh, and Jan Marian Maciejowski. An overview of sequential Monte Carlo methods for parameter estimation in general state-space models. In 15th IFAC Symposium on System
Identification, pages 774–785, 2009.
Hans R Künsch. Particle filters. Bernoulli, 19(4):1391–1403, September 2013.
Anthony Lee and Krzysztof Latuszynski. Variance bounding and geometric ergodicity of Markov chain Monte Carlo kernels for approximate Bayesian computation.
Biometrika, 101(3):655–671, 2014.
Fredrik Lindsten, Randal Douc, and E Moulines. Uniform ergodicity of the Particle
Gibbs sampler. arXiv.org, January 2014.
Nick Whiteley, Anthony Lee, and Kari Heine. On the role of interaction in sequential
Monte Carlo algorithms. Bernoulli, 22(1):494–529, 2016.
Massachusetts Institute of Technology
URL: http://jhhuggins.org/
E-mail address: [email protected]
University of Toronto
URL: http://danroy.org/
E-mail address: [email protected]
| 10math.ST
|
arXiv:1705.05295v2 [cs.DS] 19 May 2017
Constructing a Consensus Phylogeny from a
Leaf-Removal Distance
Cedric Chauve1 , Mark Jones2 , Manuel Lafond3 , Céline
Scornavacca4 , and Mathias Weller5
1
Department of Mathematics, Simon Fraser University, Burnaby,
BC, Canada,, [email protected]
2
Delft Institute of Applied Mathematics, Delft University of
Technology, P.O. Box 5, 2600 AA, Delft, the Netherlands,
[email protected]
3
Department of Mathematics and Statistics, University of Ottawa,
585 King Edward Ave., K1N 6N5 Ottawa, Canada,
[email protected]
4
Institut des Sciences de l’Evolution – Université Montpellier,
CNRS, IRD, EPHE, Place Eugène Bataillon, 34095 Montpellier,
France, [email protected]
5
Laboratoire d’Informatique, de Robotique et de Microélectronique
de Montpellier, 34392 Montpellier Cedex 5 - France,
[email protected]
January 3, 2018
Abstract
Understanding the evolution of a set of genes or species is a fundamental problem in evolutionary biology. The problem we study here takes as
input a set of trees describing possibly discordant evolutionary scenarios
for a given set of genes or species, and aims at finding a single tree that
minimizes the leaf-removal distance to the input trees. This problem is a
specific instance of the general consensus/supertree problem, widely used
to combine or summarize discordant evolutionary trees. The problem we
introduce is specifically tailored to address the case of discrepancies between the input trees due to the misplacement of individual taxa. Most
supertree or consensus tree problems are computationally intractable, and
we show that the problem we introduce is also NP-hard. We provide
tractability results in form of a 2-approximation algorithm and a parameterized algorithm with respect to the number of removed leaves. We also
introduce a variant that minimizes the maximum number d of leaves that
1
are removed from any input tree, and provide a parameterized algorithm
for this problem with parameter d.
1
Introduction
In the present paper, we consider a very generic computational biology problem: given a collection of trees representing, possibly discordant, evolutionary
scenarios for a set of biological entities (genes or species – also called taxa in
the following), we want to compute a single tree that agrees as much as possible
with the input trees. Several questions in computational biology can be phrased
in this generic framework. For example, for a given set of homologous gene sequences that have been aligned, one can sample evolutionary trees for this gene
family according to a well defined posterior distribution and then ask how this
collection of trees can be combined into a single gene tree, a problem known
as tree amalgamation [13]. In phylogenomics, one aims at inferring a species
tree from a collection of input trees obtained from whole-genome sequence data.
A first approach considers gene families and proceeds by computing individual
gene trees from a large set of gene families, and then combining this collection
of gene trees into a unique species tree for the given set of taxa; this requires
handling the discordant signal observed in the gene trees due to evolutionary
processes such as gene duplication and loss [10], lateral gene transfer [14], or
incomplete lineage sorting [12]. Another approach concatenates the sequence
data into a single large multiple sequence alignment, that is then partitioned
into overlapping subsets of taxa for which partial evolutionary trees are computed, and a unique species tree is then inferred by combining the resulting
collection of partial trees [11].
For example, the Maximum Agreement Subtree (MAST) problem considers
a collection of input trees1 , all having the same leaf labels and looks for a
tree of maximum size (number of leaves), which agrees with each of the input
trees. This problem is tractable for trees with bounded degree but NP-hard
generally [1]. The MAST problem is a consensus problem, because the input
trees share the same leaf labels set, and the output tree is called a consensus
tree. In the supertree framework, the input trees might not all have identical
label sets, but the output is a tree on the whole label set, called a supertree.
For example, in the Robinson-Foulds (RF) supertree problem, the goal is to
find a supertree that minimizes the sum of the RF-distances to the individual
input trees [15]. One way to compute consensus trees and supertrees that is
closely related to our work is to modify the collection of input trees minimally
in such a way that the resulting modified trees all agree. For example, in the
MAST problem, modifications of the input trees consist in removing a minimum
number of taxa from the whole label set, while in the Agreement Supertree
by Edge Contraction (AST-EC) problem, one is asked to contract a minimum
number of edges of the input trees such that the resulting (possibly non-binary)
1 All trees we consider here are uniquely leaf-labeled, rooted (i.e. are out-trees) and binary;
see next section for formal definitions.
2
trees all agree with at least one supertree [7]; in the case where the input trees
are all triplets (rooted trees on three leaves), this supertree problem is known as
the Minimum Rooted Triplets Inconsistency problem [4]. The SPR Supertree
problem considers a similar problem where the input trees can be modified with
the Subtree-Prune-and-Regraft (SPR) operator [16].
In the present work, we introduce a new consensus problem, called LR-Consensus.
Given a collection of input trees having the same leaf labels set, we want to remove a minimum number of leaves – an operation called a Leaf-Removal (LR) –
from the input trees such that the resulting pruned trees all agree. Alternatively,
this can be stated as finding a consensus tree that minimizes the cumulated leafremoval distance to the collection of input trees. This problem also applies to
tree amalgamation and to species tree inference from one-to-one orthologous
gene families, where the LR operation aims at correcting the misplacement of
a single taxon in an input tree.
In the next section, we formally define the problems we consider, and how
they relate to other supertree problems. Next we show that the LR-Consensus
problem is NP-hard and that in some instances, a large number of leaves need
to be removed to lead to a consensus tree. We then provide a 2-approximation
algorithm, and show that the problem is fixed-parameter tractable (FPT) when
parameterized by the total number of LR. However, these FPT algorithms have
impractical time complexity, and thus, to answer the need for practical algorithms, we introduce a variant of the LR-Consensus problem, where we ask if a
consensus tree can be obtained by removing at most d leaves from each input
tree, and describe an FPT algorithm with parameter d.
2
Preliminary notions and problems statement
Trees. All trees in the rest of the document are assumed to be rooted and
binary. If T is a tree, we denote its root by r(T ) and its leaf set by L(T ). Each
leaf is labeled by a distinct element from a label set X , and we denote by X (T )
the set of labels of the leaves of T . We may sometimes use L(T ) and X (T )
interchangeably. For some X ⊆ X , we denote by lcaT (X) the least common
ancestor of X in T . The subtree rooted at a node u ∈ V (T ) is denoted Tu and
we may write LT (u) for L(Tu ). If T1 and T2 are two trees and e is an edge of
T1 , grafting T2 on e consists in subdividing e and letting the resulting degree 2
node become the parent of r(T2 ). Grafting T2 above T1 consists in creating a
new node r, then letting r become the parent of r(T1 ) and r(T2 ). Grafting T2
on T1 means grafting T2 either on an edge of T1 or above T1 .
The leaf removal operation. For a subset L ⊆ X , we denote by T − L
the tree obtained from T by removing every leaf labeled by L, contracting the
resulting non-root vertices of degree two, and repeatedly deleting the resulting
root vertex while it has degree one. The restriction T |L of T to L is the tree
T − (X \ L), i.e. the tree obtained by removing every leaf not in L. A triplet
is a rooted tree on 3 leaves. We denote a triplet R with leaf set {a, b, c} by ab|c
3
if c is the leaf that is a direct child of the root (the parent of a and b being its
other child). We say R = ab|c is a triplet of a tree T if T |{a,b,c} = R. We denote
tr(T ) = {ab|c : ab|c is a triplet of T }.
We define a distance function dLR between two trees T1 and T2 on the same
label set X consisting in the minimum number of labels to remove from X so
that the two trees are equal. That is,
dLR (T1 , T2 ) = min{|X| : X ⊆ X and T1 − X = T2 − X}
Note that dLR is closely related to the Maximum Agreement Subtree (MAST)
between two trees on the same label set X , which consists in a subset X 0 ⊆ X of
maximum size such that T1 |X 0 = T2 |X 0 : dLR (T1 , T2 ) = |X | − |X 0 |. The MAST
of two binary trees on the same label set can be computed in time O(n log n),
where n = |X | [5], and so dLR can be found within the same time complexity.
Problem statements. In this paper, we are interested in finding a tree T on
X minimizing the sum of dLR distances to a given set of input trees.
LR-Consensus
Given: a set of trees T = {T1 , . . . , Tt } with X P
(T1 ) = . . . = X (Tt ) = X .
Find: a tree T on label set X that minimizes Ti ∈T dLR (T, Ti ).
We can reformulate the LR-Consensus problem as the problem of removing
a minimum number of leaves from the input trees so that they are compatible.
Although the equivalence between both formulations is obvious, the later formulation will often be more convenient. We need to introduce more definitions
in order to establish this equivalence.
A set of trees
S T = {T1 , . . . , Tt } is called compatible if there is a tree T such
that X (T ) = Ti ∈T X (Ti ) and T |X (Ti ) = Ti for every i ∈ [t]. In this case, we say
that T displays T . A list C = (X1 , . . . , Xt ) of subsets of X is a leaf-disagreement
P
for T if {T1 − X1 , . . . , Tt − Xt } is compatible. The size of C is i∈[t] |Xi |. We
denote by ASTLR (T ) the minimum size of a leaf-disagreement for T , and may
sometimes write ASTLR (T1 , . . . , Tt ) instead of ASTLR (T ). A subset X 0 ⊆ X
of labels is a label-disagreement for T if {T1 − X 0 , . . . , Tt − X 0 } is compatible.
Note that, if T = {T1 , T2 }, then the minimum size of a label-disagreement for
T is dLR (T1 , T2 ). We may now define the AST-LR problem (see Figure 1 for an
example).
Agreement Subtrees by Leaf-Removals (AST-LR)
Given: a set of trees T = {T1 , . . . , Tt } with X (T1 ) = . . . = X (Tt ) = X .
Find: a leaf-disagreement C for T of minimum size.
Lemma 1. Let T = {T1 , . . . , Tt } be a set ofP
trees on the same label set X , with
n = |X |. Given a supertree T such that v := Ti ∈T dLR (T, Ti ), one can compute
in time O(tn log(n)) a leaf-disagreement C of size at most v. Conversely, given
a leaf-disagreement C forPT of size v, one can compute in time O(tn log2 (tn))
a supertree T such that Ti ∈T dLR (T, Ti ) ≤ v.
4
T1
b
a
T2
c
d
a
T3
b
c
d
a
d
c
b
Figure 1: Example instance T = {T1 , T2 , T3 } of AST-LR with label set X =
{a, b, c, d}. The list (X1 = {d}, X2 = {b}, X3 = {a}) is a leaf-disagreement for T
of size 3.
From Lemma 12 both problems share the same optimality value, the NPhardness of one implies the hardness of the other and approximating one problem
within a factor c implies that the other problem can be approximated within a
factor c. We conclude this subsection with the introduction of a parameterized
variant of the AST-LR problem.
AST-LR-d
Input: a set of trees T = {T1 , . . . , Tt } with L(T1 ) = . . . = L(Tt ) = X , and an
integer d.
Question: Are there X1 , . . . , Xt ⊆ X such that |Xi | ≤ d for each i ∈ [t], and
{T1 − X1 , . . . , Tt − Xt } is compatible?
We call a tree T ∗ a solution to the AST-LR-d instance if dLR (Ti , T ∗ ) ≤ d for
each i ∈ [t].
Relation to other supertree/consensus tree problems. The most widely
studied supertree problem based on modifying the input trees is the SPR Supertree problem, where arbitrarily large subtrees can be moved in the input
trees to make them all agree (see [16] and references there). The interest of
this problem is that the SPR operation is very general, modelling lateral gene
transfer and introgression. The LR operation we introduce is a limited SPR,
where the displaced subtree is composed of a single leaf. An alternative to the
SPR operation to move subtrees within a tree is the Edge Contraction (EC)
operation, that contracts an edge of an input tree, thus increasing the degree
of the parent node. This operation allows correcting the local misplacement of
a full subtree. AST-EC is NP-complete but can be solved in O((2t)p tn2 ) time
where p is the number of required EC operations [7].
Compared to the two problems described above, an LR models a very specific type of error in evolutionary trees, that is the misplacement of a single
taxon (a single leaf) in one of the input trees. This error occurs frequently
in reconstructing evolutionary trees, and can be caused for example by some
evolutionary process specific to the corresponding input tree (recent incomplete
lineage sorting, or recent lateral transfer for example). Conversely, it is not well
2 All
missing proofs are provided in Appendix.
5
adapted to model errors, due for example to ancient evolutionary events that
impacts large subtrees. However, an attractive feature of the LR operation is
that computing the LR distance is equivalent to computing the MAST cost and
is thus tractable, unlike the SPR distance which is hard to compute. This suggests that the LR-Consensus problem might be easier to solve than the SPR
Supertree problem, and we provide indeed several tractability results. Compared to the AST-EC problem, the AST-LR problem is naturally more adapted
to correct single taxa misplacements as the EC operation is very local and the
number of EC required to correct a taxon misplacement is linear in the length
of the path to its correct location, while the LR cost of correcting this is unitary. Last, LR-Consensus is more flexible than the MAST problem as it relies on
modifications of the input trees, while with the way MAST corrects a misplaced
leaf requires to remove this leaf from all input trees. This shows that the problems AST-LR and AST-LR-d complement well the existing corpus of gene trees
correction models.
3
Hardness and approximability of AST-LR
In this section, we show that the AST-LR problem is NP-hard, from which the
LR-Consensus hardness follows. We then describe a simple factor 2 approximation algorithm. The algorithm turns out to be useful for analyzing the worst
case scenario for AST-LR in terms of the required number of leaves to remove,√as
we show that there are AST-LR instances that require removing about n − n
leaves in each input tree.
NP-hardness of AST-LR
We assume here that we are considering the decision version of AST-LR, i.e.
deciding whether there is a leaf-disagreement of size at most ` for a given `.
We use a reduction from the MinRTI problem: given a set R of rooted triplets,
find a subset R0 ⊂ R of minimum cardinality such that R \ R0 is compatible.
The MinRTI problem is NP-Hard [4] (even W [2]-hard and hard to approximate
within a O(log n) factor). Denote by M IN RT I(R) the minimum number of
triplets to remove from R to attain compatibility. We describe the reduction
here.
St Let R = {R1 , . . . , Rt } be an instance of MinRTI, with the label set L :=
i=1 X (Ri ). For a given integer m, we construct an AST-LR instance T =
{T1 , . . . , Tt } which is such that M IN RT I(R) ≤ m if and only if ASTLR (T ) ≤
t(|L| − 3) + m.
We first construct a tree Z with additional labels which will serve as our
main gadget. Let {Li }1≤i≤t be a collection of t new label sets, each of size
(|L|t)10 , all disjoint from each other and all disjoint from L. Each tree in our
AST-LR instance will be on label set X = L ∪ L1 ∪ . . . ∪ Lt . For each i ∈ [t],
let Xi be any tree with label set Li . Obtain Z by taking any tree on t leaves
6
a
X1
TL0
X2
b
c
X3
Figure 2: Construction of the tree T1 for an instance R = {R1 , R2 , R3 } of
MinRTI in which R1 = ab|c.
l1 , . . . , lt , then replacing each leaf li by the Xi tree (i.e. li is replaced by r(Xi )).
Denote by rZ (Xi ) the root of the Xi subtree in Z.
Then for each i ∈ [t], we construct Ti from Ri as follows. Let L0 = L \ X (Ri )
be the set of labels not appearing in Ri , noting that |L0 | = |L| − 3. Let TL0 be
any tree with label set L0 , and obtain the tree Zi by grafting TL0 on the edge
between rZ (Xi ) and its parent. Finally, Ti is obtained by grafting Ri above
Zi . See Figure 2 for an example. Note that each tree Ti has label set X as
desired. Also, it is not difficult to see that this reduction can be carried out in
polynomial time. This construction can now be used to show the following.
Theorem 1. The AST-LR and LR-Consensus problems are NP-hard.
The idea of the proof is to show that in the constructed AST-LR instance,
we are ”forced” to solve the corresponding MinRTI instance. In more detail, we
show that M IN RT I(R) ≤ m if and only if ASTLR (T ) ≤ t(|L| − 3) + m. In one
direction, given a set R0 of size m such that R \ R0 is compatible, one can show
that the following leaf removals from T make it compatible: remove, from each
Ti , the leaves L0 = L\X (Ri ) that were inserted into the Z subtree, then for each
Ri ∈ R0 , remove a single leaf in X (Ri ) from Ti . This sums up to t(|L| − 3) + m
leaf removals. Conversely, it can be shown that there always exists an optimal
solution for T that removes, for each Ti , all the leaves L0 = L \ X (Ri ) inserted
in the Z subtree, plus an additional single leaf l from m trees Ti1 , . . . , Tim such
that l ∈ L. The corresponding triplets Ri1 , . . . , Rim can be removed from R so
that it becomes compatible.
Approximating AST-LR and bounding worst-case scenarios
Given the above result, it is natural to turn to approximation algorithms in order
to solve AST-LR or LR-Consensus instances. It turns out that there is a simple
factor 2 approximation for LR-Consensus which is achieved by interpreting the
problem as finding a median in a metric space. Indeed, it is not hard to see
that dLR is a metric (over the space of trees on the same label set X ). A direct
consequence, using an argument akin to the one in [9, p.351], is the following.
Theorem 2. The following is a factor P
2 approximation algorithm for LR-Consensus:
return the tree T ∈ T that minimizes Ti ∈T dLR (T, Ti ).
7
Theorem 2 can be used to lower-bound the ‘worst’ possiblepinstance of
AST-LR. We show that in some cases, we can only keep about |X | leaves
√
per tree. That is, there are instances for which ASTLR (T ) = Ω(t(n − n)),
where t is the number of trees and n = |X |. The argument is based on a probabilistic argument, for which we will make use of the following result [3, Theorem
4.3.iv].
√
Theorem 3 ([3]). For any constant c > e/ 2, there is some n0 such that for
all n ≥ n0 , the following holds: if T1 and T2 are two binary trees on n leaves
√
chosen randomly, uniformly and independently, then E[dLR (T1 , T2 )] ≥ n − c n.
√
Corollary 1. There are instances of AST-LR in which Ω(t(n − n)) leaves need
to be deleted.
The above is shown by demonstrating that, by picking
P a set T of t random trees, the expected optimal sum of distances minT Ti ∈T dLR (T, Ti ) is
√
Ω(t(n − n). This is not direct though, since the tree T ∗ that minimizes this
sum is not itself random, and so we cannot apply Theorem 3 directly on T ∗ . We
can however, show that the tree T 0 ∈ T obtained using the
√ 2-approximation,
which is random, has expected sum of distances Ω(t(n − n)). Since T ∗ requires, at best, half the leaf deletions of T 0 , the result follows. Note that finding
a non-trivial upper bound on ASTLR (T ) is open.
4
Fixed-parameter tractability of AST-LR and AST-LR-d.
An alternative way to deal with computational hardness is parameterized complexity. In this section, we first show that AST-LR is fixed-parameter-tractable
with respect to q := ASTLR (T ). More precisely, we show that AST-LR can be
solved in O(12q n3 ) time, where n := |X |. We then consider an alternative parameter d, and show that finding a tree T ∗ , if it exists, such that dLR (Ti , T ∗ ) ≤ d
for every input tree Ti , can be done in O(cd d3d (n3 + tn log n)) time for some
constant c.
4.1
Parameterization by q
The principle of the algorithm is the following. It is known that a set of trees T =
{T1 , . . . , TSt } is compatible if and only if the union of their triplet decomposition
tr(T ) = Ti ∈T tr(Ti ) is compatible [2]. In a step-by-step fashion, we identify
a conflicting set of triplets in tr(T ), each time branching into the (bounded)
possible leaf-removals that can resolve the conflict. We stop when either tr(T )
is compatible after the performed leaf-removals, or when more than q leaves
were deleted.
We employ a two phase strategy. In the first phase, we eliminate direct
conflicts in tr(T ), i.e. if at least two of ab|c, ac|b and bc|a appear in tr(T ),
then we recursively branch into the three ways of choosing one of the 3 triplets,
and remove one leaf in each Ti disagreeing with the chosen triplet (we branch
8
into the three possible choices, either removing a, b or c). The chosen triplet is
locked in tr(T ) and cannot be changed later.
When the first phase is completed, there are no direct conflicts and tr(T )
consists in a full set of triplets on X . That is, for each distinct a, b, c ∈ X , tr(T )
contains exactly one triplet on label set {a, b, c}. Now, a full set of triplets is
not necessarily compatible, and so in the second phase we modify tr(T ), again
deleting leaves, in order to make it compatible. Only the triplets that have
not been locked previously can be modified. This second phase is analogous to
the FPT algorithm for dense MinRTI presented in [8]. The dense MinRTI is a
variant of the MinRTI problem, introduced in Section 3, in which the input is
a full set of triplets and one has to decide whether p triplets can be deleted to
attain compatibility.
Theorem 4 ([8]). A full set of triplets R is compatible if and only if for any
set of four labels {a, b, c, d}, R does not contain the subset {ab|c, cd|b, bd|a} nor
the subset {ab|c, cd|b, ad|b}.
One can check, through an exhaustive enumeration of the possibilities, that
there are only four ways to correct a conflicting set of triplets R1 , R2 , R3 where
R1 = ab|c, R2 = cd|b, R3 ∈ {bd|a, ad|b}. We can: (1) transform R1 to bc|a; (2)
transform R1 to ac|b; (3) transform R2 to bd|c; (4) transform R3 to ab|d. This
leads to a O(4p n3 ) algorithm for solving dense MinRTI: find a conflicting set of
four labels, and branch on the four possibilities, locking the transformed triplet
each time.
For the second phase of AST-LR, we propose a slight variation of this algorithm. Each time a triplet R is chosen and locked, say R = ab|c, the trees
containing ac|b or bc|a must loose a, b or c. We branch into these three possibilities. Thus for each conflicting 4-set, there are four ways of choosing a triplet,
then for each such choice, three possible leaves to delete from a tree. This gives
12 choices to branch into recursively. Algorithm 1 is described in detail in the
Appendix and its analysis yields the following.
Theorem 5. AST-LR can be solved in time O(12q tn3 ).
Although Theorem 5 is theoretically interesting as it shows that AST-LR is in
FPT with respect to q, the 12q factor might be too high for practical purposes,
motivating the alternative approach below.
4.2
Parameterization by maximum distance d
We now describe an algorithm for the AST-LR-d problem, running in time
O(cd d3d (n3 + tn log n)) that, if it exists, finds a solution (where here c is a
constant not depending on d nor n).
We employ the following branch-and-bound strategy, keeping a candidate
solution at each step. Initially, the candidate solution is the input tree T1 and, if
T1 is indeed a solution, we return it. Otherwise (in particular if dLR (T1 , Ti ) > d
for some input tree Ti ), we branch on a set of “leaf-prune-and-regraft” operations
9
on T1 . In such an operation, we prune one leaf from T1 and regraft it somewhere
else. If we have not produced a solution after d such operations, then we halt this
branch of the algorithm (as any solution must be reachable from T1 by at most
d operations). The resulting search tree has depth at most d. In order to bound
the running time of the algorithm, we need to bound the number of “leaf-pruneand-regraft” operations to try at each branching step. There are two steps to
this: first, we bound the set of candidate leaves to prune, second, given a leaf,
we bound the number of places where to regraft it. To bound the candidate set
of leaves to prune, let us call a leaf x interesting if there is a solution T ∗ , and
minimal sets X1 , Xi ⊆ X of size at most d, such that (a) T1 − X1 = T ∗ − X1 ,
(b) Ti − Xi = T ∗ − Xi , and (c) x ∈ X1 \ Xi , where Ti is an arbitrary input tree
for which dLR (T1 , Ti ) > d. It can be shown that an interesting leaf x must exist
if there is a solution. Moreover, though we cannot identify x before we know
T ∗ , we can nevertheless construct a set S of size O(d2 ) containing all interesting
leaves. Thus, in our branching step, it suffices to consider leaves in S.
Assuming we have chosen the correct x, we then bound the number of places
to try regrafting x. Because of the way we chose x, we may assume there
is a solution T ∗ and Xi ⊆ X such that |Xi | ≤ d, Ti − Xi = T ∗ − Xi and
x ∈
/ Xi . Thus we may treat Ti as a “guide” on where to regraft x. Due to
the differences between T1 , Ti and T ∗ , this guide does not give us an exact
location in T1 to regraft x. Nevertheless, we can show that the number of
candidate locations to regraft x can be bounded by O(d). Thus, in total we
have O(d3 ) branches at each step in our search tree of depth d, and therefore
have to consider O((O(3d ))d ) = O(cd d3d ) subproblems.
Theorem 6. AST-LR-d can be solved in time O(cd d3d (n3 + tn log n)), where c
is a constant not depending on d or n.
5
Conclusion
To conclude, we introduced a new supertree/consensus problem, based on a simple combinatorial operator acting on trees, the Leaf-Removal. We showed that,
although this supertree problem is NP-hard, it admits interesting tractability
results, that compare well with existing algorithms. Future research should
explore if various simple combinatorial operators, that individually define relatively tractable supertree problems (for example LR and EC) can be combined
into a unified supertree problem while maintaining approximability and fixedparameter tractability.
Acknowledgement: MJ was partially supported by Labex NUMEV (ANR-10LABX-20) and Vidi grant 639.072.602 from The Netherlands Organization for Scientific Research (NWO). CC was supported by NSERC Discovery Grant 249834. CS was
partially supported by the French Agence Nationale de la Recherche Investissements
dAvenir/Bioinformatique (ANR-10-BINF-01-01, ANR-10- BINF-01-02, Ancestrome).
ML was supported by NSERC PDF Grant.
10
References
[1] A. Amir and D. Keselman. Maximum agreement subtree in a set of evolutionary trees: Metrics and efficient algorithms. SIAM J. Comput., 26:1656–
1669, 1997.
[2] D. Bryant. Building trees, hunting for trees, and comparing trees. PhD
thesis, Bryant University, 1997.
[3] D. Bryant, A. McKenzie, and M. Steel. The size of a maximum agreement
subtree for random binary trees. Dimacs Series in Discrete Mathematics
and Theoretical Computer Science, 61:55–66, 2003.
[4] J. Byrka, S. Guillemot, and J. Jansson. New results on optimizing rooted
triplets consistency. Discrete Appl. Math., 158:1136–1147, 2010.
[5] R. Cole, M. Farach-Colton, R. Hariharan, T. M. Przytycka, and M. Thorup.
An O(nlog n) algorithm for the maximum agreement subtree problem for
binary trees. SIAM J. Comput., 30:1385–1404, 2000.
[6] Y. Deng and D. Fernández-Baca. Fast Compatibility Testing for Rooted
Phylogenetic Trees. In Combinatorial Pattern Matching 2016, volume 54
of LIPIcs. Leibniz Int. Proc. Inform., pages 12:1–12:12, 2016.
[7] D. Fernández-Baca, S. Guillemot, B. Shutters, and S. Vakati. Fixedparameter algorithms for finding agreement supertrees. SIAM J. Comput.,
44:384–410, 2015.
[8] S. Guillemot and M. Mnich. Kernel and fast algorithm for dense triplet
inconsistency. Theoret. Comput. Sci., 494:134–143, 2013.
[9] D. Gusfield. Algorithms on strings, trees and sequences: computer science
and computational biology. Cambridge university press, 1997.
[10] M. Hellmuth, N. Wieseke, M. Lechner, H.-P. Lenhof, M. Middendorf, and
P. F. Stadler. Phylogenomics with paralogs. Proc. Natl. Acad. Sci. USA,
112:2058–2063, 2015.
[11] E. D. Jarvis et al. Whole-genome analyses resolve early branches in the
tree of life of modern birds. Science, 346:1320–1331, 2014.
[12] C. Scornavacca and N. Galtier. Incomplete lineage sorting in mammalian
phylogenomics. Sys. Biol., 66:112–120, 2017.
[13] C. Scornavacca, E. Jacox, and G. J. Szollösi. Joint amalgamation of most
parsimonious reconciled gene trees. Bioinformatics, 31:841–848, 2015.
[14] G. J. Szollösi, B. Boussau, S. S. Abby, E. Tannier, and V. Daubin. Phylogenetic modeling of lateral gene transfer reconstructs the pattern and relative
timing of speciations. Proc. Natl. Acad. Sci. USA, 109:17513–17518, 2012.
11
[15] P. Vachaspati and T. Warnow. FastRFS: fast and accurate robinson-foulds
supertrees using constrained exact optimization. Bioinformatics, 33:631–
639, 2017.
[16] C. Whidden, N. Zeh, and R. G. Beiko. Supertrees based on the subtree
prune-and-regraft distance. Sys. Biol., 63:566–581, 2014.
12
A
Omitted proofs
Here we give proofs for several results whose proofs were omitted in the main
paper. Note that the proof of Theorem 6 is deferred to its own section.
Lemma 1 (restated). Let T = {T1 , . . . , Tt } be a setPof trees on the same
label set X . Then, given a supertree T such that v := Ti ∈T dLR (T, Ti ), one
can compute in time O(tn log n) a leaf-disagreement C of size at most v, where
n = |X |. Conversely, given a leaf-disagreement C forPT of size v, one can
compute in time O(tn log2 (tn)) a supertree T such that Ti ∈T dLR (T, Ti ) ≤ v.
Proof. In the first direction, for each Ti ∈ T , there is a set Xi ⊆ X of size
dLR (T, Ti ) such that Ti − Xi = T − Xi . Moreover, Xi can be found in time
O(n log n). Thus (X1 , . . . , Xt ) is a leaf-disagreement of the desired size and
can be found in time O(tn log n). Conversely, let C = (X1 , . . . , Xt ) be a leafdisagreement of size v. As T 0 = {T1 − X1 , . . . , Tt − Xt } is compatible, there is a
tree T that displays T 0 , and it is easy to see that the sum of distances between
T and T 0 is at most the size of C. As for the complexity, it is shown in [6] how
to compute in time O(tn log2 (tn)), given a set of trees T 0 , a tree T displaying
T 0 if one exists.
We next consider the case where T consists only of two trees.
Lemma 2. Let T1 , T2 be two trees on the same label set X . Then ASTLR (T1 , T2 ) =
dLR (T1 , T2 ). Moreover, every optimal leaf-disagreement C = (X10 , X20 ) for T1 and
T2 can be obtained in the following manner: for every label-disagreement X 0 of
size dLR (T1 , T2 ), partition X 0 into X10 , X20 .
Proof. Let X 0 ⊂ X such that |X 0 | = dLR (T1 , T2 ) and T1 − X 0 = T2 − X 0 . Then
clearly, for any bipartition (X10 , X20 ) of X 0 , T10 := T1 − X10 and T20 := T2 − X20
are compatible, since the leaves that T10 and T20 have in common yield the same
subtree, and leaves that appear in only one tree cannot create incompatibility.
In particular, ASTLR (T1 , T2 ) ≤ dLR (T1 , T2 ).
Conversely, let C = (X10 , X20 ) be a minimum leaf-disagreement. We have
0
X1 ∩ X20 = ∅, for if there is some ` ∈ X10 ∩ X20 , then ` could be reinserted into
one of the two trees without creating incompatibility. Thus C is a bipartition
of X 0 = X10 ∪ X20 . Moreover, we must have T1 − X 0 = T2 − X 0 , implying
|X 0 | ≥ dLR (T1 , T2 ). Combined with the above inequality, |X 0 | = dLR (T1 , T2 ),
and the Lemma follows.
It follows from Lemma 2 that any optimal label-disagreement X 0 can be
turned into an optimal leaf-disagreement, which is convenient as X 0 can be
found in polynomial time. We will make heavy use of this property later on.
Note that the same type of equivalence does not hold when 3 or more trees
are given, i.e. computing a MAST of three trees does not necessarily yield a
leaf-disagreement of minimum size. Consider for example the instance T =
{T1 , T2 , T3 } in Figure 1. An optimal leaf-disagreement for T has size 2 and
13
consists of any pair of distinct leaves. On the other hand, an optimal leafdisagreement for T has size 3, and moreover each leaf corresponds to a different
label.
Theorem 1 (restated). The AST-LR and LR-Consensus problems are NPhard.
Proof. We begin by restating the reduction from MinRTI to AST-LR.
St Let R = {R1 , . . . , Rt } be an instance of MinRTI, with the label set L :=
i=1 X (Ri ). For a given integer m, we construct an AST-LR instance T =
{T1 , . . . , Tt } which is such that M IN RT I(R) ≤ m if and only if ASTLR (T ) ≤
t(|L| − 3) + m.
We first construct a tree Z with additional labels which will serve as our
main gadget. Let {Li }1≤i≤t be a collection of t new label sets, each of size
(|L|t)10 , all disjoint from each other and all disjoint from L. Each tree in our
AST-LR instance will be on label set X = L ∪ L1 ∪ . . . ∪ Lt . For each i ∈ [t],
let Xi be any tree with label set Li . Obtain Z by taking any tree on t leaves
l1 , . . . , lt , then replacing each leaf li by the Xi tree (i.e. li is replaced by r(Xi )).
Denote by rZ (Xi ) the root of the Xi subtree in Z.
Then for each i ∈ [t], we construct Ti from Ri as follows. Let L0 = L \ X (Ri )
be the set of labels not appearing in Ri , noting that |L0 | = |L| − 3. Let TL0 be
any tree with label set L0 , and obtain the tree Zi by grafting TL0 on the edge
between rZ (Xi ) and its parent. Finally, Ti is obtained by grafting Ri above
Zi . See Figure 2 for an example. Note that each tree Ti has label set X as
desired. Also, it is not difficult to see that this reduction can be carried out in
polynomial time.
We now show that M IN RT I(R) ≤ m if and only if ASTLR (T ) ≤ t(|L| −
3) + m.
(⇒) Let R0 ⊂ R such that |R0 | ≤ m and R∗ := R \ R0 is compatible, and let
T (R∗ ) be a tree displaying R∗ . Note that |R∗ | ≥ t − m. We obtain a AST-LR
solution by first deleting, in each Ti ∈ T , all the leaves labeled by L \ X (Ri )
(thus Ti becomes the tree obtained by grafting Ri above Z). Then for each
deleted triplet Ri ∈ R0 , we remove any single leaf of Ti labeled by some element
in X (Ri ). In this manner, no more than t(|L| − 3) + m leaves get deleted.
Moreover, grafting T (R∗ ) above Z yields a tree displaying the modified set of
trees, showing that they are compatible.
(⇐) We first argue that if T admits a leaf-disagreement C = (X1 , . . . , Xt ) of
size at most t(|L| − 3) + m, then there is a better or equal solution that removes,
in each Ti , all the leaves labeled by L \ X (Ri ) (i.e. those grafted in the Zi tree).
For each i ∈ [t], let Ti0 = Ti − Xi , and denote T 0 = {T10 , . . . , Tt0 }. Suppose that
there is some i ∈ [t] and some ` ∈ L \ X (Ri ) such that ` ∈ X (Ti0 ).
We claim that ` ∈
/ X (Tj0 ) for every i 6= j ∈ [t]. Suppose otherwise that
0
` ∈ X (Tj ) for some j 6= i. Consider first the case where ` ∈
/ X (Rj ). Note that
by the construction of Zi and Zj , for every xi ∈ X (Xi ) ∩ X (Ti0 ) ∩ X (Tj0 ) and
every xj ∈ X (Xj ) ∩ X (Ti0 ) ∩ X (Tj0 ), Ti0 contains the `xi |xj triplet whereas Tj0
contains the `xj |xi triplet. Since these triplets are conflicting, no supertree can
14
contain both and so no such xi , xj pair can exist, as we are assuming that a
supertree for Ti0 and Tj0 exists. This implies that one of X (Xi ) ∩ X (Ti0 ) ∩ X (Tj0 )
or X (Xj ) ∩ X (Ti0 ) ∩ X (Tj0 ) must be empty. Suppose without loss of generality
that the former is empty. Then each xi ∈ Xi must have been deleted in at least
one of Ti or Tj . As |X (Xi )| = (|L|t)10 > t(|L| − 3) + m, this contradicts the
size of the solution C. In the second case, we have ` ∈ X (Rj ). But this time, if
there are xi ∈ X (Xi ) ∩ X (Ti0 ) ∩ X (Tj0 ) and xj ∈ X (Xj ) ∩ X (Ti0 ) ∩ X (Tj0 ), then
Tj0 contains the xi xj |` triplet, again conflicting with the `xi |xj triplet found in
Ti . As before, we run into a contradiction since too many Xi or Xj leaves need
to be deleted. This proves our claim.
We thus assume that ` only appears in Ti0 . Let Rj ∈ R such that ` ∈ X (Rj ),
noting that ` does not appear in Tj0 . Consider the solution T 00 obtained from T 0
by removing ` from Ti0 , and placing it back in Tj0 where it originally was in Tj .
Formally this is achieved by replacing, in the leaf-disagreement C, Xi by Xi ∪{`}
and Xj by Xj \ {`}. Since ` still appears only in one tree, no conflict is created
and we obtain another solution of equal size. By repeating this process for
every such leaf `, we obtain a solution in which every leaf labeled by L \ X (Ri )
is removed from Ti0 . We now assume that the solution T 0 has this form.
Consider the subset R0 = {Ri ∈ R : |X (Ti0 ) ∩ X (Ri )| < 3}, that is those
triplets Ri for which the corresponding tree Ti had a leaf removed outside of the
Zi tree. By the form of the T 0 solution, at least t(|L| − 3) removals are done in
the Zi trees, and as only m removals remain, R0 has size at most m. We show
that R \ R0 is a compatible set of triplets. Since T 0 is compatible, there is a
tree T that displays each Ti0 ∈ T 0 , and since each triplet of R \ R0 belongs to
some Ti0 , T also displays R \ R0 . This concludes the proof.
Theorem 2 (restated). The following is a factor 2 approximation
algorithm
P
for LR-Consensus: return the tree T ∈ T that minimizes Ti ∈T dLR (T, Ti ).
∗
Proof. Let
solution for LR-Consensus, i.e. T ∗ is a tree minP T be an optimal
∗
imizing Ti ∈T dLR (Ti , T ), and let T be chosen as described in the theorem
statement. Moreover let T 0 be the tree of T minimizing dLR (T 0 , T ∗ ). By the
triangle inequality,
X
X
X
dLR (T 0 , Ti ) ≤
(dLR (T 0 , T ∗ ) + dLR (T ∗ , Ti )) ≤ 2
dLR (T ∗ , Ti )
Ti ∈T
Ti ∈T
Ti ∈T
where the last inequality is due to the fact that dLR (T 0P
, T ∗ ) ≤ dLR (T ∗ , Ti )
0
for all i, by our choicePof T . Our choice of T implies
Ti ∈T dLR (T, Ti ) ≤
P
0
∗
d
(T
,
T
)
≤
2
d
(T
,
T
).
i
Ti ∈T LR
Ti ∈T LR i
Corollary 1 (restated). There are instances of AST-LR in which Ω(t(n −
leaves need to be deleted.
√
n))
Proof. Let T = {T1 , . . . , Tt } be a random set of t trees chosen uniformly and
independently. For large enough n, the expected sum of distances between each
15
pair of trees is
X
E
X
dLR (Ti , Tj ) =
E[dLR (Ti , Tj )] ≥
1≤i<j≤t
1≤i<j≤t
√
t
(n − c n)
2
Pt
for some constant c, by Theorem 3. Let S := minT i=1 dLR (T, Ti ) be the random variable corresponding to the
Ptminimum sum of distances. By Theorem 2,
there is a tree T 0 ∈ T such that i=1 dLR (T 0 , Ti ) ≤ 2S. We have
X
dLR (Ti , Tj ) ≤
1≤i<j≤t
X
dLR (Ti , T 0 ) + dLR (T 0 , Tj )
1≤i<j≤t
= (t − 1)
t
X
dLR (Ti , T 0 )
i=1
≤ (t − 1)2S
Since, in general for two random variables X and Y , always having X ≤ Y
implies E[X] ≤ E[Y ], we get
X
√
t
(n − c n) ≤ E
dLR (Ti , Tj ) ≤ E[(t − 1)2S] = 2(t − 1)E[S]
2
1≤i<j≤t
√
√
yielding E[S] ≥ t/4(n − c n) = Ω(t(n − n)), and so there must exist an
instance T satisfying the statement.
Theorem 5 (restated). AST-LR can be solved in time O(12q tn3 ).
Proof. We provide an algorithm, Algorithm 1, for AST-LR, and prove its correctness and complexity.
We first argue that the algorithm is correct. First observe that the algorithm only returns TRUE when a conflict-free set of triplets is attained without
deleting more than q leaves, and so there are no false-positives. Moreover, it
is not hard to see that the first phase of the algorithm tries every possible way
of obtaining a full set of triplets from tr(T ) using at most q leaf removals. Indeed, for every set of 3 labels a, b, c that are present in a direct conflict, the
algorithm branches into the 3 ways of locking ab|c, ac|b or bc|a and for each
tree Ti disagreeing with the chosen triplet, all three ways of removing a leaf to
agree with the chosen triplet are tested. In a similar manner, for each dense set
D of triplets that is attained, each way of freeing D from conflicting 4-sets is
evaluated (not only in the ways of choosing a triplet to resolve the conflict, but
also in the ways of removing leaves from the trees of T so that they agree with
the locked triplets). It follows that if tr(T ) can be made compatible by deleting
at most q leaves, then some leaf in the branch tree created by the algorithm will
return TRUE, and so there are no false-negatives.
16
Algorithm 1 Recursive AST-LR FPT algorithm.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
procedure mastrl(T , q, phase, F )
T is the set of input trees, q is the maximum number of leaves to delete,
phase is the current phase number (initially 1), F is the set of locked triplets
so far
if q < 0 or F contains conflicting triplets then
Return FALSE
else if there is ab|c ∈ F and a tree Ti ∈ T such that ac|b ∈ tr(Ti ) or
bc|a ∈ Ti then
Branching: If one of the following calls returns True
mastrl((T \ {Ti }) ∪ {Ti − {a}}, q − 1, phase, F ) //remove a from
Ti
mastrl((T \ {Ti }) ∪ {Ti − {b}}, q − 1, phase, F ) //remove b from
Ti
mastrl((T \ {Ti }) ∪ {Ti − {c}}, q − 1, phase, F ) //remove c from
Ti
then Return True, otherwise Return False
else if phase = 1 then
if there are a, b, c ∈ X such that at least 2 of ab|c, ac|b or bc|a appear
in T then
Branching: If one of the following calls returns True
mastrl(T , q, phase, F ∪ {ab|c})
mastrl(T , q, phase, F ∪ {ac|b})
mastrl(T , q, phase, F ∪ {bc|a})
then Return True, otherwise Return False
else
Return mastrl(T , q, 2, F ) //enter phase 2
else if phase = 2 then
if there is a conflicting set {a, b, c, d} in tr(T ) ∪ F then
Branching: If one of the following calls returns True
mastrl(T , q, phase, F ∪ {ac|b})
mastrl(T , q, phase, F ∪ {bc|a})
mastrl(T , q, phase, F ∪ {bd|c})
mastrl(T , q, phase, F ∪ {ab|d})
then Return True, otherwise Return False
else
Return True //There are no conflicts ⇒ tr(T ) ∪ F is compatible
17
As for the complexity, when the algorithm enters the ‘else if’ block of line 4, it
branches into 3 cases that decrement q. When it enters the ‘if’ block of line 11,
it branches into 3 cases but q is not decremented. However each of these 3
recursive calls immediately leads to the ‘else if’ block on line 4, and so this case
can be seen as branching into 9 cases. Similarly, when the algorithm enters the
‘if’ block of line 19, it branches into 4 cases, each of which leads to the 3 subcases
following line 4. Thus 12 cases are considered. Therefore, the branching tree
created by the algorithm has degree at most 12 and depth at most q, and so
at most 12q cases are considered. Finally, each call to the algorithm requires
time O(tn3 ) since this is the time required to identify conflicting sets of triplets
within the t trees.
B
Leaf Prune-and-Regraft Moves
Here we introduce the notion of leaf prune-and-regraft (LPR) moves, which
will be used in the proof of Theorem 6, and which may be of independent
interest. In an LPR move, we prune a leaf from a tree and then regraft it
another location (formal definitions below). LPR moves provide an alternate
way of characterizing the distance function dLR - indeed, we will show that
dLR (T1 , T2 ) ≤ k if and only if there is a sequence of at most k LPR moves
transforming T1 into T2 .
Definition 1. Let T be a tree on label set X . A LPR move on T is a pair (`, e)
where ` ∈ X and e ∈ {E(T − {`}), ⊥}. Applying (`, e) consists in grafting ` on
the e edge of T − {`} if e 6=⊥, and above the root of T − {`} if e =⊥.
An LPR sequence L = ((`1 , e1 ), . . . , (`k , ek )) is an ordered tuple of LPR
moves, where for each i ∈ [k], (`i , ei ) is an LPR move on the tree obtained after
applying the first i − 1 LPR moves of L. We may write L = (`1 , . . . , `k ) if the
location at which the grafting takes place does not need to be specified. We say
that L turns T1 into T2 if, by applying each LPR move of L in order on T1 , we
obtain T2 .
See Figure 4 for an example of an LPR sequence.
In the following statements, we assume that T1 and T2 are two trees on label
set X . We exhibit an equivalence between leaf removals and LPR sequences,
then show that the order of LPR moves in a sequence do not matter in terms
of turning one tree into another - in particular any leaf can be displaced first.
Lemma 3. There is a subset X ⊆ X such that T1 − X = T2 − X if and
only if there is an LPR sequence (x1 , x2 , . . . , xk ) turning T1 into T2 such that
X = {x1 , . . . , xk }.
Proof. If T1 = T2 then the proof is trivial, so we will assume this is not the case.
We prove the lemma by induction on |X|.
For the base case, suppose that X = {x}. If T1 − X = T2 − X, then let
Tm = T1 − X = T2 − X. We find an LPR move (x, e) with e ∈ E(Tm ) ∪ {⊥}
turning T1 into T2 . Observe that T2 can be obtained by grafting x on Tm , either
18
a
a
b
c
d
e
a
c
b
c
e
b
d
e
d
Figure 3: Sequence of trees showing the LPR sequence L = ((b, f ), (d, ⊥)),
where f is the edge between the root and the least common ancestor of a and
c in the first tree.
on an edge uv, in which case we set e = uv, or above the root, in which case we
set e =⊥. Since Tm = T1 − {x}, it follows that (x, e) is an LPR move turning
T1 into T2 . In the other direction, assume there is an LPR move (x, e) turning
T1 into T2 . Observe that for any tree T 0 derived from T1 by an LPR move using
x, T 0 − {x} = T1 − {x}. In particular, T2 − {x} = T1 − {x} and we are done.
For the induction step, assume that |X| > 1 and that the claim holds for
any X 0 such that |X 0 | < |X|. If T1 − X = T2 − X, then define Tm = T1 − X,
and let x be an arbitrary element of X. We will first construct a tree T10 such
that T1 − {x} = T10 − {x} and T10 − (X \ {x}) = T2 − (X \ {x}).
Observe that T2 − (X \ {x}) can be obtained by grafting x in Tm . Let e = uv
if this grafting takes place on an edge of Tm with v being the child of u, or e =⊥
if x is grafted above Tm , and in this case let v = r(Tm ). Let v 0 = vT1 −{x} be
the node in T1 − {x} corresponding to v.
Let T10 be derived from T1 − {x} by grafting x onto the edge between v 0
and its parent if v 0 is non-root, and grafting above v 0 otherwise. It is clear
that T1 − {x} = T10 − {x}. Furthermore, by our choice of v 0 we have that
T10 − (X \ {x}) = T2 − (X \ {x}).
Now that we have T1 −{x} = T10 −{x} and T10 −(X \{x}) = T2 −(X \{x}), by
the inductive hypothesis there is an LPR sequence turning T1 into T10 consisting
of a single move (x, e), and an LPR sequence (x1 , x2 , . . . , xk−1 ) turning T10 into
T2 such that {x1 , . . . , xk0 } = (X \ {x}). Then by concatenating these two
sequences, we have an LPR sequence (x1 , x2 , . . . , xk ) turning T1 into T2 such
that X = {x1 , . . . , xk }.
19
For the converse, suppose that there is an LPR sequence (x1 , x2 , . . . , xk )
turning T1 into T2 such that X = {x1 , . . . , xk }. Let T10 be the tree derived from
T1 by applying the first move in this sequence. That is, there is an LPR move
(x1 , e) turning T1 into T10 , and there is an LPR sequence (x2 , . . . , xk ) turning
T10 into T2 . Then by the inductive hypothesis T1 − {x1 } = T10 − {x1 } and
T10 − {x2 , . . . , xk } = T2 − {x2 , . . . , xk }. Thus, T1 − X = T10 − X = T2 − X, as
required.
Lemma 4. If there is an LPR sequence L = (x1 , . . . , xk ) turning T1 into T2 ,
then for any i ∈ [k], there is an LPR sequence L0 = (x01 , . . . , x0k ) turning T1 into
T2 such that x01 = xi and {x1 , . . . , xk } = {x01 , . . . , x0k }.
Proof. Consider again the proof that if T1 − X = T2 − X then there is an LPR
sequence (x1 , . . . xk ) turning T1 into T2 such that X = {x1 , . . . , xk } (given in the
proof of Lemma 3). When |X| > 1, we construct this sequence by concatenating
the LPR move (x, e) with an LPR sequence of length |X| − 1, where x is an
arbitrary element of X. As we could have chosen any element of X to be x,
we have the following: If T1 − X = T2 − X then for each x ∈ X, there is an
LPR sequence (x1 , . . . , xk ) turning T1 into T2 such that X = {x1 , . . . , xk } and
x1 = x.
Thus our proof is as follows: Given an LPR sequence L = (x1 , . . . , xk )
turning T1 into T2 and some i ∈ [k], Lemma 3 implies that T1 − {x1 , . . . , xk } =
T2 − {x1 , . . . , xk }. By the observation above, this implies that there is an LPR
sequence (x01 , . . . , x0k ) turning T1 into T2 such that {x1 , . . . , xk } = {x01 , . . . , x0k }
and x01 = x.
C
Proof of Theorem 6
This section makes use of the concept of LPR moves, which are introduced in
the previous section. As discussed in the main paper, we employ a branch-andbound style algorithm, in which at each step we alter a candidate solution by
pruning and regrafting a leaf. That is, we apply an LPR move.
The technically challenging part is bound the number of possible LPR moves
to try. To do this, we will prove Lemma 7, which provides a bound on the number
of leaves to consider, and Lemma 10, which bounds the number of places a leaf
may be regrafted to.
Denote by tr(T ) the set of rooted triplets of a tree T . Two triplets R1 ∈
tr(T1 ) and R2 ∈ tr(T2 ) are conflicting if R1 = ab|c and R2 ∈ {ac|b, bc|a}. We denote by conf (T1 , T2 ) the set of triplets of T1 for which there is a conflicting triplet
in T2 . That is, conf (T1 , T2 ) = {ab|c ∈ tr(T1 ) : ac|b ∈ tr(T2 ) or bc|a ∈ tr(T2 )}.
Finally we denote by conf set(T1 , T2 ) = {{a, b, c} : ab|c ∈ conf (T1 , T2 )}, i.e.
the collection of 3-label sets formed by conflicting triplets. Given a collection
C = {S1 , . . . , S|C| } of sets, a hitting set of C is a set S such that S ∩ Si 6= ∅ for
each Si ∈ C.
Lemma 5. Let X ⊆ X . Then T1 − X = T2 − X if and only if X is a hitting
set of conf set(T1 , T2 ).
20
Proof. It is known that for two rooted trees T1 , T2 that are leaf-labelled and
binary, T1 = T2 if and only if tr(T1 ) = tr(T2 ) [2]. Note also that tr(T − X) =
{ab|c ∈ tr(T1 ) : X ∩ {a, b, c} = ∅} for any tree T and X ⊆ X .
Therefore we have that T1 − X = T2 − X if and only if tr(T1 − X) =
tr(T2 − X), which holds if and only if for every a, b, c ∈ X \ X, if ab|c ∈ tr(T1 )
then ab|c ∈ tr(T2 ). This in turn occurs if and only if X is a hitting set for
conf set(T1 , T2 ).
In what follows, we call X ⊆ X a minimal disagreement between T1 and T2
if T1 − X = T2 − X and for any X 0 ⊂ X, T1 − X 0 6= T2 − X 0 .
Lemma 6. Suppose that d < dLR (T1 , T2 ) ≤ d0 +d with d0 ≤ d, and that there is a
tree T ∗ and subsets X1 , X2 ⊆ X such that T1 −X1 = T ∗ −X1 , T2 −X2 = T ∗ −X2
and |X1 | ≤ d0 , |X2 | ≤ d. Then, there is a minimal disagreement X between T1
and T2 of size at most d + d0 and x ∈ X such that x ∈ X1 \ X2 .
Proof. Let X 0 = X1 ∪ X2 . Observe that T1 − X 0 = T ∗ − X 0 = T2 − X 0 and
|X 0 | ≤ d+d0 . Letting X be the minimal subset of X 0 such that T1 −X = T2 −X,
we have that X is a minimal disagreement between T1 and T2 and |X| ≤ d + d0 .
Furthermore as |X| ≥ dLR (T1 , T2 ) > d, |X \X2 | > 0, and so there is some x ∈ X
with x ∈ X \ X2 = X1 \ X2 .
We are now ready to state and prove Lemma 7.
Lemma 7. Suppose that dLR (T1 , T2 ) ≤ d for some integer d. Then, there is
some S ⊆ X such that |S| ≤ 8d2 , and for any minimal disagreement X between
T1 and T2 with |X| ≤ d, X ⊆ S. Moreover S can be found in time O(n2 ).
We will call S as described in Lemma 7 a d-disagreement kernel between
T1 and T2 . Thus Lemma 6 essentially states that if T1 isn’t a solution and
dLR (T1 , T2 ) > d, then for T1 to get closer to a solution, there is a leaf x in the
dLR (T1 , T2 )-disagreement kernel that needs to be removed and regrafted in a
location that T2 ‘agrees with’. Lemma 7 in turn gives us a set S of size at most
8d2 such that the desired x must be contained in S.
Proof. By Lemma 5, it is enough to find a set S such that S contains every
minimal hitting set of conf set(T1 , T2 ) of size at most d.
We construct S as follows.
Let X be a subset of X of size at most d such that T1 − X = T2 − X. As
previously noted, this can found in time O(n log n) [5].
For notational convenience, for each x ∈ X we let x1 , x2 be two new labels,
and set X1 = {x1 : x ∈ X}, X2 = {x2 : x ∈ X}. Thus, X1 , X2 are disjoint
“copies” of X. Let T10 be derived from T1 by replacing every label from X
with the corresponding label in X1 , and similarly let T20 be derived from T2 by
replacing every label from X with the corresponding label in X2 .
Let TJ be a tree with label set (X \ X) ∪ X1 ∪ X2 such that TJ − X2 = T10
and TJ − X1 = T20 . The tree TJ always exists and can be found in polynomial
time. Intuitively, we can start from T10 , and graft the leaves of X2 where T2
21
T1
x
a
b
c
dc
T2
e
f
g
y
h
a
b
y
c
dc
e
x
f
g
h
TJ
c
x1
a
b
e
d
f
y2 x2
g
y1
h
Figure 4: Construction of the tree TJ , given two trees T1 , T2 , with X = {x, y}
such that T1 − X = T2 − X.
“wants” them to be. See Figure 4 for an example. Algorithm 2 gives a method
for constructing TJ , and takes O(n2 ) time.
In addition, let L be the set of all labels in X \ X that are descended in TJ
from lcaTJ (X1 ∪ X2 ), and let R = X \ (L ∪ X). Thus, L, X, R form a partition
of X , and L, X1 , X2 , R form a partition of the labels of TJ .
For the rest of the proof, we call {x, y, z} a conflict triple if {x, y, z} ∈
conf set(T1 , T2 ).
We first observe that no triple in conf set(T1 , T2 ) contains a label in R.
Indeed, consider a triple {x, y, z}. Any conflict triple must contain a label from
X, so assume without loss of generality that x ∈ X, z ∈ R. If x ∈ X, y ∈ L, z ∈
R, then we have that TJ contains the triplets x1 y|z, x2 y|z, and so T1 and T2 both
contain xy|z, and {x, y, z} is not a conflict triple. Similarly if x, y ∈ X, z ∈ R,
then TJ contains the triplets x1 y1 |z, x2 y2 |z, and again {x, y, z} is not a conflict
triple. If x ∈ X and y, z ∈ R, then the triplet on {x1 , y, z} in TJ depends only
on the relative positions in TJ of y, z and lcaTJ (X1 ∪ X2 ). Thus we get the
same triplet if we replace x1 with x2 , and so {x, y, z} is not a conflict triple.
This concludes the proof that no triple in conf set(T1 , T2 ) contains a label
in R. Having shown this, we may conclude that any minimal disagreement
between T1 and T2 is disjoint from R, and so our returned set S only needs to
contain labels in L ∪ X.
Now consider the tree T ∗ = TJ |X1 ∪X2 , i.e. the subtree of TJ restricted to
the labels in X1 ∪ X2 . Thus in the example of Figure 4, T ∗ is the subtree of TJ
22
Algorithm 2 Algorithm to construct “Join tree” of T10 , T20
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
procedure join-trees(T10 , T20 , L0 , X10 , X20 )
T10 is a tree on L0 ∪ X10 , T20 is a tree on L0 ∪ X20 , T10 |L0 = T20 |L0 . Output: A
tree TJ on L0 ∪ X10 ∪ X20 such that TJ |L0 ∪X10 = T10 and TJ |L0 ∪X20 = T20
if L0 ∪ X10 = ∅ then
Return T20
else if L0 ∪ X20 = ∅ then
Return T10
else if X10 ∪ X20 = ∅ then
Return T10
Set r1 = root of T10 , u, v the children of r1
Set r2 = root of T20 , w, z the children of r2
Set X1u = descendants of u in X10 , L1u = descendants of u in L0
Set X1v = descendants of v in X10 , L1v = descendants of v in L0
Set X2w = descendants of w in X20 , L2w = descendants of w in L0
Set X2z = descendants of z in X20 , L2z = descendants of z in L0
if L1u = L2w and L1v = L2z then
Set Tlef t = join-trees(T10 |L1u ∪X1u , T20 |L1u ∪X2w , L1u , X1u , X2w )
Set Tright = join-trees(T10 |L1v ∪X1v , T20 |L1v ∪X2z , L1v , X1v , X2z )
else if L1u = L2z and L1v = L2w then
Set Tlef t = join-trees(T10 |L1u ∪X1u , T20 |L1u ∪X2z , L1u , X1u , X2z )
Set Tright = join-trees(T10 |L1v ∪X1v , T20 |L1v ∪X2w , L1v , X1v , X2w )
else if L1u =∅ then
Set Tlef t = join-trees(T10 |X1u , T20 |∅ , ∅, X1u , ∅)
Set Tright = join-trees(T10 |L0 ∪X1v , T20 , L0 , X1v , X20 )
else if L1v = ∅ then
Set Tlef t = join-trees(T10 |L0 ∪X1u , T20 , L0 , X1u , X20 )
Set Tright = join-trees(T10 |X1v , T20 |∅ , ∅, X1v , ∅)
else if L2w = ∅ then
Set Tlef t = join-trees(T10 |∅ , T20 |X2w , ∅, ∅, X2w )
Set Tright = join-trees(T10 , T20 |L0 ∪X2z , L0 , X10 , X2z )
else if L2z = ∅ then
Set Tlef t = join-trees(T10 , T20 |L0 ∪X2w , L0 , X10 , X2w )
Set Tright = join-trees(T10 |∅ , T20 |X2z , ∅, ∅, X2z )
. If none of the above cases holds, then T10 |L0 6= T20 |L0 , contradicting the
requirements on the input
Set TJ = the tree on L0 ∪ X10 ∪ X20 whose root has Tlef t and Tright as
children.
Return TJ .
23
spanned by {x1 , x2 , y1 , y2 }. We will now use the edges of T ∗ to form a partition
of L, as follows. For any edge uv in T ∗ with u the parent of v, let s(uv) denote
the set of labels y ∈ X such that y has an ancestor which is an internal node on
the path from u to v in TJ , but y is not a descendant of v itself. For example in
Figure 4, if u is the least common ancestor of x1 , y1 and v is the least common
ancestor of x1 , y2 , then uv is an edge in T ∗ and s(uv) = {c, d, e}.
Observe that {s(uv) : uv ∈ E(T ∗ )} forms a partition of L. (Indeed, for any
l ∈ L, let u be the minimal element in T ∗ on the path in TJ between l and
lcaTJ (X1 ∪ X2 ) (note that u exists as lcaTJ (X1 ∪ X2 ) itself is in T ∗ ). As u is
in T ∗ , both of its children are on paths in TJ between u and a child of u in T ∗ .
In particular, the child of u which is an ancestor of l is an internal node on the
path between u and v in TJ , for some child v of u in T ∗ , and l is not descended
from v by construction. It is clear by construction that all s(uv) are disjoint.)
The main idea behind the construction of S is that we will add X to S,
together with O(d) labels from s(uv) for each edge uv in T ∗ . As the number of
edges in T ∗ is 2(|X1 ∪ X2 | − 1) = O(d), we have the required bound of O(d2 )
on |S|.
So now consider s(uv) for some edge uv in T ∗ . In order to decide which
labels to add to S, we need to further partition s(uv). Let u = u0 u1 . . . ut = v
be the path in TJ from u to v. For each i ∈ [t − 1] (note that this does not
include i = 0), we call the set of labels descended from ui but not ui+1 a dangling
clade. Observe that the dangling clades form a partition of s(uv). Thus in the
example of Figure 4, if u is the least common ancestor of x1 , y1 and v is the
least common ancestor of x1 , y2 , then for the edge uv the dangling clades are
{c} and {d, e}.
We now make the following observations about the relation between s(uv)
and triples in conf set(T1 , T2 ).
Observation 1: if {x, y, z} is a conflict triple and x ∈ s(uv), y, z ∈
/ s(uv),
then {x0 , y, z} is also a conflict triple for any x0 ∈ s(uv). (The intuition behind
this is that there are no labels appearing ’between’ x and x0 that are not in
s(uv).)
Observation 2: for any triple {x, y, z} with x, y ∈ s(uv), {x, y, z} is a
conflict triple if and only if x, y are in different dangling clades and z ∈ X
with zi descended from v, z3−i not descended from u1 for some i ∈ [2] (recall
that z1 ∈ X1 and z2 ∈ X2 ). To prove one direction, it is easy to see that if
the conditions hold, then Ti displays either xz|y or yz|x (depending on which
dangling clade appears ’higher’), and T3−i displays xy|z. For the converse,
observe first that z ∈ X as X is a hitting set for conf set(T1 , T2 ) and x, y ∈
/ X.
Then if xy are in the same dangling clade, we have that both T1 and T2 display
xy|z. So x, y must be in different dangling clades. Next observe that each of
z1 , z2 must either be descended from v or not descended from u1 , as otherwise v
would not be the child of u in T ∗ . If z1 , z2 are both descended from v or neither
are descended from u1 , then T1 and T2 display the same triplet on {x, y, z}.
So instead one must be descended from v and one not descended from u1 , as
required.
Using Observations 1 and 2, we now prove the following:
24
Observation 3: for any minimal disagreement X 0 between T1 and T2 , one
of the following holds:
• X 0 ∩ s(uv) = ∅;
• s(uv) ⊆ X 0 ;
• s(uv) \ X 0 forms a single dangling clade.
To see this, let X 0 be any minimal hitting set of conf set(T1 , T2 ) with s(uv)∩
X 6= ∅ and s(uv) \ X 0 6= ∅. As X 0 is minimal, any x ∈ s(uv) ∩ X 0 must be in a
conflict triple {x, y, z} with y, z ∈
/ X 0 . As X is a hitting set for conf set(T1 , T2 ),
at least one of y, z must be in X. If y, z ∈
/ s(uv), then by Observation 1 {x0 , y, z}
0
is also a conflict triple for any x ∈ s(uv) \ X 0 . But this is a contradiction as
{x0 , y, z} has no elements in X 0 . Then one of y, z must also be in s(uv). Suppose
without loss of generality that y ∈ s(uv). We must also have that z ∈ X, as X
is a hitting set for conf set(T1 , T2 ) and x, y ∈
/ X. By Observation 2, we must
have that one of z1 , z2 is descended from v, and the other is not descended from
u1 . This in turn implies (again by Observation 2) that for any x0 ∈ s(uv) \ X 0 , if
x0 and y are in different dangling clades then {x0 , y, z} is a conflict triple. Again
this is a contradiction as {x0 , y, z} has no elements of X 0 , and so we may assume
that all elements of s(uv) \ X 0 are in the same dangling clade.
It remains to show that every element of this dangling clade is in s(uv) \ X 0 .
To see this, suppose there exists some x ∈ X 0 in the same dangling clade as
the elements of s(uv) \ X 0 . Once again we have that x is in some conflict triple
{x, y, z} with y, z ∈
/ X 0 , and if y, z ∈
/ s(uv) then {x0 , y, z} is also a conflict triple
0
0
for any x ∈ s(uv) \ X , a contradiction. So we may assume that one of y, z is
in s(uv) \ X 0 . But all elements of s(uv) \ X 0 are in the same dangling clade as
x, and so by Observation 2 {x, y, z} cannot be a conflict triple, a contradiction.
So finally we have that all elements of s(uv) \ X 0 are in the same dangling clade
and all elements of this clade are in s(uv) \ X 0 , as required.
With the proof of Observation 3 complete, we are now in a position to
construct S. For any minimal hitting set X 0 of conf set(T1 , T2 ) with size at
most d, by Observation 3 either X 0 ∩ s(uv) = ∅, or s(uv) ⊆ X 0 (in which
case |s(uv)| ≤ d), or s(uv) \ X 0 forms a single dangling clade C (in which case
|s(uv) \ C| ≤ d).
So add all elements of X to S. For all uv ∈ E(TJ ) and any dangling clade
C of labels in s(uv), add s(uv) \ C to S if |s(uv) \ C| ≤ d. Observe that this
construction adds at most 2d labels from s(uv) to S.
Thus, in total, we have that the size of S is at most |X| + 2d|E(TJ )| ≤
d + 2d(2(|X1 ∪ X2 | − 1)) ≤ d + 2d(4d − 2) = 8d2 − 3d ≤ 8d2 .
Algorithm 3 describes the full procedure formally. The construction of TJ
occurs once and as noted above takes O(n2 ) time. As each other line in the
algorithm is called at most n times and takes O(n) time, the overall running
time of the algorithm O(n2 ).
0
The last ingredient needed for Theorem 6 is Lemma 10, which shows that if
a leaf x of T1 as described in Lemma 6 has to be moved, then there are not too
25
Algorithm 3 Algorithm to construct a d-disagreement kernel between T1 and
T2
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
procedure disagreement-kernel(d, T1 , T2 )
T1 and T2 are trees on X ,d an integer.
Output: A set S ⊆ X such that for every minimal disagreement X
between T1 and T2 with |X| ≤ d, X ⊆ S.
Find X such that |X| ≤ d and T1 − X = T2 − X
Set S = X
Let X1 , X2 be copies of X and replace T1 , T2 with corresponding trees
T10 , T20 on (X \ X) ∪ X1 , (X \ X) ∪ X2 .
Let TJ = join-trees(T10 , T20 , (X \ X), X1 , X2 )
Let T ∗ = TJ |X1 ∪X2
for uv ∈ E(T ∗ ) do
Let u = u0 u1 . . . ut = v be the path in TJ from u to v
Let s(uv) = {l ∈ X \ X : l is descended from u1 but not from v}
Set p = |s(uv)| − d
. Any clade C has |C| ≥ p iff |s(uv) \ C| ≤ d
for i ∈ [t] do
Set C = {l ∈ s(uv) : l is descended from ui but not from ui+1 } .
C is a single ’dangling clade’
if |C| ≥ p then
Set S = S ∪ (s(uv) \ C)
Return S.
many ways to regraft it in order to get closer to T ∗ .
In the course of the following proofs, we will want to take observations about
one tree and use them to make statements about another. For this reason it’s
useful to have a concept of one node ”corresponding” to another node in a
different tree. In the case of leaf nodes this concept is clear - two leaf nodes
are equivalent if they are assigned the same label- but for internal nodes there
is not necessarily any such correspondence. However, in the case that one tree
is the restriction of another to some label set, we can introduce a well-defined
notion of correspondence:
Given two trees T, T 0 such that T 0 = T |X for some X ⊆ X (T ), and a node
u ∈ V (T 0 ), define the node uT of T by uT = lcaT (LT 0 (u)). That is, uT is the
least common ancestor, in T , of the set of labels belonging to descendants of u
in T 0 . We call uT the node corresponding to u in T .
We note two useful properties of uT here:
Lemma 8. For any T, T 0 , X ⊆ X (T ) such that T 0 = T |X and any u, v ∈ V (T 0 ),
uT is an ancestor of vT if and only if u is an ancestor of v.
Proof. If u is an ancestor of v then LT 0 (v) ⊆ LT 0 (u), which implies that uT
is an ancestor of vT . For the converse, observe that for any Z ⊆ X, any
label in X descending from lcaT (Z) in T is also descending from lcaT 0 (Z) in
T 0 . In particular letting Z = LT 0 (u), we have LT (uT ) ∩ X = LT (lcaT (Z)) ∩
26
X ⊆ LT 0 (lcaT 0 (Z)) = LT 0 (lcaT 0 (LT 0 (u))) = LT 0 (u) ⊆ LT (uT ) ∩ X. Thus
LT 0 (u) = LT (uT ) ∩ X and similarly LT 0 (v) = LT (vT ) ∩ X. Then we have
that uT being an ancestor of vT implies LT (vT ) ⊆ LT (uT ), which implies that
LT 0 (v) = LT (vT ) ∩ X ⊆ LT (uT ) ∩ X = LT 0 (u), which implies that u is an
ancestor of v.
Lemma 9. For any T 00 , T 0 , T and Y ⊆ X ⊆ X (T ) such that T 0 = T |X and
T 00 = T 0 |Y , (uT 0 )T = uT .
Proof. It is sufficient to show that any node in V (T ) is a common ancestor of
LT 0 (lcaT 0 (Z)) if and only if it is a common ancestor of Z, where Z = LT 00 (u) (as
this implies that the least common ancestors of these two sets are the same).
It is clear that if v ∈ V (T ) is a common ancestor of LT 0 (lcaT 0 (Z)) then it
is also a common ancestor of Z, as Z ⊆ LT 0 (lcaT 0 (Z)). For the converse,
observe that as T 0 = T |X and Z ⊆ X, any label in X descended from lcaT 0 (Z)
in T 0 is also descended from lcaT (Z) in T . This implies LT 0 (lcaT 0 (Z)) ⊆
LT (lcaT (Z)), and so any common ancestor of Z in T is also a common ancestor
of LT 0 (lcaT 0 (Z)).
We are now ready to state and prove Lemma 10
Lemma 10. Suppose that d < dLR (T1 , T2 ) ≤ d0 + d with d0 ≤ d, and that
there are X1 , X2 ⊆ X , and a tree T ∗ such that T1 − X1 = T ∗ − X1 , T2 − X2 =
T ∗ − X2 , |X1 | ≤ d0 , |X2 | ≤ d, and let x ∈ X1 \ X2 . Then, there is a set P of
trees on label set X that satisfies the following conditions:
• for any tree T 0 such that dLR (T 0 , T ∗ ) < dLR (T1 , T ∗ ) and T 0 can be obtained
from T1 by pruning a leaf x and regrafting it, T 0 ∈ P ;
• |P | ≤ 18(d + d0 ) + 8;
• P can be found in time O(n(log n + 18(d + d0 ) + 8)).
The idea behind the proof is as follows: by looking at a subtree common
to T1 and T2 , we can identify the location that T2 “wants” x to be positioned.
This may not be the correct position for x, but we can show that if x is moved
too far from this position, we will create a large number of conflicting triplets
between T2 and the solution T ∗ . As a result, we can create all trees in P by
removing x from T1 and grafting it on one of a limited number of edges.
Proof. For the purposes of this proof, we will treat each tree T as “planted”, i.e.
as having an additional root of degree 1, denoted r(T ), as the parent of what
would normally be considered the “root” of the tree. (That is, r(T ) is the parent
of lcaT (X (T )). Note that trees are otherwise binary. We introduce r(T ) as a
notational convenience to avoid tedious repetition of proofs - grafting a label
above a tree T can instead be represented as grafting it on the edge between
r(T ) and its child. For the purposes of corresponding nodes, if T 0 = T − X then
(r(T 0 ))T = r(T ). This allows us to assume that every node in T is a descendant
of uT for some node u in T 0 .
27
A naive method for constructing a tree in P is the following: Apply an LPR
move (x, e) on T1 , such that x is moved to a position that T2 “wants” x to be in.
There are at least two problems with this method. The first is that, since T1 and
T2 have different structures, it is not clear where in T1 it is that T2 “wants” x
to be. We can partially overcome this obstacle by initially considering a subtree
common to both T1 and T2 . However, because T2 will want to move leafs that
will not be moved in T1 , it can still be the case that even though T2 “agrees”
with T ∗ on x, T2 may want to put x in the “wrong” place, when viewed from
the perspective of T1 . For this reason we have to give a counting argument to
show that if x is moved “too far” from the position suggested by T2 , it will
create too many conflicting triplets, which cannot be covered except by moving
x. We make these ideas precise below.
Let P ∗ be the set of all trees T 0 such that dLR (T 0 , T ∗ ) < dLR (T1 , T ∗ ) and
T can be obtained from T1 by an LPR move on x. Thus, it is sufficient to
construct a set P such that |P | ≤ 18(d + d0 ) + 8 and P ∗ ⊆ P .
We first construct a set Xm ⊆ X such that |Xm | ≤ d + d0 , x ∈ Xm , and
T1 − Xm = T2 − Xm . Note that the unknown set (X1 ∪ X2 ) satisfies these
properties, as T1 − (X1 ∪ X2 ) = T ∗ − (X1 ∪ X2 ) = T2 − (X1 ∪ X2 ), and so such
a set Xm must exist. We can find Xm in time O(n log n) by applying MAST on
(T1 − {x}, T2 − {x}) [5].
Now let Tm be the tree with labelset X \ Xm such that Tm = T1 − Xm =
T2 − Xm . Note that for any T 0 in P ∗ , we have that T 0 − {x} = T1 − {x} and
therefore T 0 − Xm = T1 − Xm = Tm .
Informally, we now have a clear notion of where T2 “wants” x to go, relative
to Tm . There is a unique edge e in Tm such that grafting x on e will give the
tree T2 − (Xm \ {x}). If we assume that this is the “correct” position to add
x, then it only remains to add the remaining labels of Xm back in a way that
agrees with T1 (we will describe how this can be done at the end of the proof).
Unfortunately, grafting x onto the obvious choice e does not necessarily lead to
a graph in P ∗ . This is due to the fact that T2 can be “mistaken” about labels
outside of Xm .
To address this, we have try grafting x on other edges of Tm . There are too
many edges to try them all. We therefore need the following claim, which allows
us to limit the number of edges to try.
0
Claim: In O(n) time, we can find y ∈ V (Tm ) and Z ⊆ V (Tm ), |Z| ≤ 4,
such that:
S
• For any T 0 in P ∗ , x ∈ LT 0 (yT 0 ) \ z0 ∈Z LT 0 (zT0 0 )
S
• |LTm (y) \ z0 ∈Z LTm (z 0 )| ≤ 8(d + d0 )
Informally, the claim identifies a node y and set of nodes Z in Tm , such that x
should be added as a descendant of y but not of any node in Z, and the number
of such positions is bounded. Algorithm 4 describes the formal procedure to
produce y and Z. The proof of the claim takes up most of the remainder of our
proof; the reader may wish to skip it on their first readthrough.
28
Algorithm 4 FPT algorithm to restrict possible locations of x given
(Tm , T1 , T2 , x, d, d0 )
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
procedure location-restriction(Tm , T2 , Xm , x, d, d0 )
T1 , T2 are two trees, Tm is a common subtree of T1 and T2 such that Tm =
T2 − Xm , x is a label that cannot be moved in T2 (but must be moved in
T1 ), d is the maximum number of leaves we can remove in a tree, d0 is the
maximum number of leaves we can move in T1 . Output is a pair (y, Z) with
y ∈ V (Tm ), Z ⊆ V (Tm ), such that we may assume x is a descendant of y
but not a descendant of any z 0 ∈ Z, and the number of labels like this in
Tm is O(d). For this pseudocode, every tree T has a degree-1 root r(T ).
0
= T2 − (Xm \ {x})
Set Tm
0
0
0 (z) \ x| ≥ d + d , or
such that |LTm
Set z = lowest ancestor of x in Tm
0
return (r(Tm ), ∅) if no such z exists.
0
0
0 (y) \ LT 0 (z)| ≥ d + d ,
Set y = lowest ancestor of z in Tm
such that |LTm
m
0
or r(Tm ) if no such ancestor exists.
. Find sets Z = Z1 ∪ Z2 of nodes that cover all but a
bounded number of the descendants of y, and such that we can rule out x
being descended from any z 0 in Z.
0
Let z1 , z2 be the children of z such that x is descended from z1 in Tm
0
0
0
0
0
Set Z1 = {z descended from z2 : |LTm (z )| ≥ d + d and |LTm (z2 ) \
0
0
0
0 (z )| ≥ d + d , and this does not hold for any ancestor of z }
LTm
0
Let y1 , y2 be the children of y such that x is descended from y1 in Tm
0
0
0
0
0
Set Z2 = {y descended from y2 : |LTm (y )| ≥ d + d and |LTm (y2 ) \
0
0
0
0 (y )| ≥ d + d , and this does not hold for any ancestor of y }
LTm
. Note that |Z1 | ≤ 2, |Z2 | ≤ 2.
0
Set y ∗ = node of Tm for which y is the corresponding node in Tm
Set Z = {z ∗ in Tm : z 0 ∈ Z1 ∪ Z2 is the node corresponding to z ∗ in Tm }
Return (y ∗ , Z)
29
0
0
Proof. Let Tm
= T2 − (Xm \ {x}). Note that Tm
− {x} = Tm . We will use the
0
presence of x in Tm to identify the node y and set Z. (Technically, this means
0
the nodes we find are nodes in Tm
rather than Tm . However, we note that apart
the parent of x and x itself, neither of which will be added to {y} ∪ Z, every
0
0 corresponding to some node v in Tm . For the sake
node in Tm
is the node vTm
0
of clarity, we ignore the distinction and write v to mean vTm
throughout this
proof. The nodes in {y} ∪ Z should ultimately be replaced with the nodes in
Tm to which they correspond.)
We first identify two nodes z, y of Tm as follows:
0
0
0 (z)\{x}| ≥ d+d . If no
• Let z be the least ancestor of x in Tm
such that |LTm
0
0
0
such x exists, then X (Tm ) ≤ d + d and we may return y = r(Tm ), Z = ∅.
0
0
0 (y)\LT 0 (z)| ≥ d+d .
such that |LTm
• Let y be the least ancestor of z in Tm
m
0
If no such y exists, set y = r(Tm ).
Using this definition, we will show that x must be a descendant of yT 0 for
any T 0 ∈ P ∗ . We first describe a general tactic for restricting the position of x
in T 0 , as this tactic will be used a number of times.
Suppose that for some T 0 ∈ P ∗ there is a set of d+d0 triplets in conf set(T 0 , T2 )
whose only common element is x. Then let X 0 ⊆ X be a set of labels such that
T 0 − X 0 = T ∗ − X 0 and |X 0 | = dLR (T 0 , T ∗ ) ≤ dLR (T1 , T ∗ ) − 1 ≤ d0 − 1.
Note that T2 − (X 0 ∪ X2 ) = T ∗ − (X 0 ∪ X2 ) = T 0 − (X 0 ∪ X2 ), and therefore (X 0 ∪ X2 ) is a hitting set for conf set(T 0 , T2 ). As |X 0 ∪ X2 | ≤ d + d − 1
and there are d + d0 triplets in conf set(T 0 , T2 ) whose only common element
is x, it must be the case that x ∈ X 0 ∪ X2 . As x ∈
/ X2 , we must have
x ∈ X 0 . But this implies that T1 − X 0 = T 0 − X 0 = T ∗ − X 0 and therefore
dLR (T1 , T ∗ ) ≤ |X 0 | = dLR (T 0 , T ∗ ) ≤ dLR (T1 , T ∗ ) − 1, a contradiction. Thus we
may assume that such a set of triplets does not exist.
We now use this idea to show that x ∈ LT 0 (yT 0 ), for any T 0 ∈ P ∗ . Indeed,
0
suppose x ∈
/ LT 0 (yT 0 ). We may assume y 6= r(Tm
) as otherwise yT 0 = r(T 0 ) by
0
definition and so LT 0 (yT 0 ) = X (T ). Then let z1 , . . . , zd+d0 be d + d0 labels in
0
0 (z) \ {x}. Let y1 , . . . , yd+d0 be d + d labels in LT 0 (y) \ LT 0 (z). Observe
LTm
m
m
0
0
that for each i ∈ [d + d ], Tm (and therefore T2 ) contains the triplet (zi x|yi ),
but T 0 contains the triplet (zi yi |x). Therefore conf set(T 0 , T2 ) contains d + d0
sets whose only common element is x. As this implies a contradiction, we must
have x ∈ LT 0 (yT 0 ).
0 (y)| maybe be very large.
Note however that |LTm
In order to provide a
bounded range of possible
positions
for
x,
we
still
need
to
find a set Z of nodes
S
0
0 (y) \
0 (z ))| is bounded, and such that we can show
such that |LTm
L
0
T
z ∈Z
m
x∈
/ LT 0 (zT0 0 ) for any z 0 ∈ Z.
We now construct a set Z1 of descendants of z as follows:
0
• Let z1 , z2 be the children of z in Tm
such that x is descended from z1 .
0
0 (z2 )| ≤ 3(d + d ) then set Z1 = ∅.
• If |LTm
30
• Otherwise, let Z1 be the set of highest descendants z 0 of z2 , such that
0
0
0
0
0 (z )| ≥ d + d
0 (z2 ) \ LT 0 (z )| ≥ d + d
|LTm
and |LTm
(i.e. by highest
m
0
00
descendant we mean such that z has no ancestor z with the same properties).
0
0 (z1 )| ≤ d+d by our choice of z. It follows that if |LT 0 (z2 )| ≤
Note that |LTm
m
0
0
0 (z)| ≤ 4(d + d ). If on the other hand |LT 0 (z2 )| > 3(d +
3(d + d ) then |LTm
m
d0 ) then Z1 is non-empty. Indeed, let z 0 be a lowest descendant of z2 with
0
0
0
0
0
0 (z )| ≥ d + d , and observe that |LT 0 (z )| ≤ 2(d + d ). Then either z ∈ Z1 ,
|LTm
m
0
0
0
0
0
0
0
or |LTm (z2 ) \ LTm (z )| ≤ d + d , in which case |LTm (z2 )| ≤ |LTm (z2 ) \ LTm (z 0 )| +
0
0
0
0
0 (z )| ≤ d + d + 2(d + d ) = 3(d + d ).
|LTm
We also have that |Z1 | ≤ 2. Indeed, let z10 , z20 , z30 be three distinct nodes in
0
Z1 , and suppose without loss of generality that (z10 z20 |z30 ) ∈ tr(Tm
). Then setting
0
0
0
0
0
0
0
0 (z )| ≥
z = lcaTm (z1 , z2 ), we have that z is an ancestor of z1 such that |LTm
0
0
0
0
0
0
0
d+d and |LTm (z2 )\LTm (z )| ≥ |LTm (z3 )| ≥ d+d , a contradiction by minimality
of z1 .
S
0
0
0 (z) \
0
We have that |LTm
z 0 ∈Z1 LTm (z ))| ≤ 4(d + d ). Indeed, if Z1 = ∅ then
0
0
|LTm (z)| ≤ 4(d + d ) as described above. Otherwise, let z 0 be an element of Z1
0
0
0
0 (zp )| ≥ |LT 0 (z )| ≥ d + d ,
. Clearly |LTm
and zp its parent, zs its sibling in Tm
m
0
0
0 (z2 ) \ LT 0 (zp )| < d + d . If |LT 0 (zs )| ≥ d + d
and so as zp ∈
/ Z1 we have |LTm
m
m
0
0
0
0
0
0
then
z
∈
Z
(since
|L
(z
)
\
L
(z
)|
≥
|L
(z
)|
≥
d
+
d
),
and
so
|L
(z)
\
s
1
Tm 2
Tm s
Tm
Tm
S
0
0
0 (z ))| ≤ |LT 0 (z1 )| + |LT 0 (z2 ) \ LT 0 (zp )| ≤ 2(d + d ). Otherwise,
z 0 ∈Z1 LTm
m
m
m
S
0
0 (z) \
0
0
0
0
0
|LTm
z 0 ∈Z1 LTm (z ))| ≤ |LTm (z1 )| + |LTm (z2 ) \ LTm (zp )| + |LTm (zs )| ≤
3(d + d0 ).
S
0
0 (z) \
0
We have now shown that |Z1 | ≤ 2 and that |LTm
z 0 ∈Z1 LTm (z ))| ≤
4(d + d0 ). The final property of Z1 we wish to show is that for any z 0 ∈ Z1 and
any T 0 ∈ P , x ∈
/ LT 0 (zT0 0 ).
0 (z2 ) \
So suppose x ∈ LT 0 (zT0 0 ). Let ẑ1 , . . . , ẑd+d0 be d + d0 labels in LTm
0
0
0 (z ). Also, z1 and z2 were already taken. Let w1 , . . . , wd+d0 be d + d labels
LTm
0
0
0
0 (z ). Then for each i ∈ [d + d ], T
in LTm
m (and therefore T2 ) contains the triplet
(ẑi wi |x), but T 0 contains the triplet (xwi |ẑi ). Therefore conf set(T 0 , T2 ) contains
d + d0 sets whose only common element is x. As this implies a contradiction,
we must have x ∈
/ LT 0 (zT0 0 ).
We now define a set Z2 of descendants of y:
0
• If y = r(Tm
), set Z2 = ∅.
0
• Otherwise, let y1 , y2 be the children of y in Tm
such that z is descended
from y1 .
0
0 (y2 )| ≤ 3(d + d ) then set Z2 = ∅.
• If |LTm
• Otherwise, let Z2 be the set of highest descendants y 0 of y2 , such that
0
0
0
0
0
0 (y )| ≥ d + d and |LT 0 (y2 ) \ LT 0 (y )| ≥ d + d (i.e. such that y has
|LTm
m
m
00
no ancestor y with the same properties).
31
In a similar way S
to the proofs for Z1 , we can show that |Z2 | ≤ 2, that
0
0
0 (y) \ LT 0 (z)) \
0
|(LTm
/ LT 0 (yT0 0 ) for
y 0 ∈Z2 LTm (y ))| ≤ 4(d + d ), and that x ∈
m
0
0
∗
any y ∈ Z2 and any T ∈ P .
0
0 (y1 ) \ LT 0 (z)| ≤ d + d by our choice of y. It follows that
Note that |LTm
m
0
0
0 (y2 )| ≤ 3(d + d ) then |LT 0 (y) \ LT 0 (z)| ≤ 4(d + d ). If on the other
if |LTm
m
m
0
0
0 (y2 )| > 3(d + d ), then Z2 is non-empty. Indeed, let y be a lowest
hand |LTm
0
0
0
0
0 (y )| ≥ d + d , and observe that |LT 0 (y )| ≤ 2(d + d ).
descendant of y2 with |LTm
m
0
0
0
0 (y2 ) \ LT 0 (y )| ≤ d + d , in which case |LT 0 (y2 )| ≤
Then either y ∈ Z2 , or |LTm
m
m
0
0
0
0
0
0 (y2 ) \ LT 0 (y )| + |LT 0 (y )| ≤ d + d + 2(d + d ) = 3(d + d ).
|LTm
m
m
We also have that |Z2 | ≤ 2. Indeed, let y10 , y20 , y30 be three distinct nodes
0
in Z2 , and suppose without loss of generality that (y10 y20 |y30 ) ∈ tr(Tm
). Then
0
0
0
0
0
0 (y , y ), we have that y
setting y = lcaTm
is
an
ancestor
of
y
such
that
1 2
1
0
0
0
0
0
0 (y )| ≥ d + d and |LT 0 (y2 ) \ LT 0 (y )| ≥ |LT 0 (y )| ≥ d + d , a contradiction
|LTm
3
m
m
m
by minimality of y1 .
S
0
0
0 (y) \ LT 0 (z)) \
0
We have that |(LTm
y 0 ∈Z2 LTm (y ))| ≤ 4(d + d ). Indeed, if
m
0
0
0 (ŷ) \ LT 0 (z)| < d + d for any ancestor ŷ of
y = r(Tm ) then by construction |LTm
m
0
) rather than
z (noting that otherwise there would be no reason to set y as r(Tm
0
0
0
0 (y)\LT 0 (z)| < d+d . If y 6= r(T
the child of r(Tm )), and so in particular |LTm
m)
m
0
0 (y)\LT 0 (z)| ≤ 4(d+d ) as described above. Otherwise, let
and Z2 = ∅ then |LTm
m
0
0 (yp )| ≥
. Clearly |LTm
y 0 be an element of Z2 and yp its parent, ys its sibling in Tm
0
0
0
0 (y )| ≥ d + d , and so as yp ∈
0 (y2 ) \ LT 0 (yp )| < d + d . If
|LTm
/ Z2 we have |LTm
m
0
0
0
0 (ys )| ≥ d+d then ys ∈ Z2 (since |LT 0 (y2 )\LT 0 (ys )| ≥ |LT 0 (y )| ≥ d+d ),
|LTm
m
m
m
S
0
0 (y) \ LT 0 (z)) \
0
0
0 (z)| + |LT 0 (y2 ) \
and so |(LTm
Tm
y 0 ∈Z2 LTm (y ))| ≤ |LTm (y1 ) \ LS
m
m
0
0
0 (yp )| ≤ 2(d + d ).
0 (y) \ LT 0 (z)) \
0
LTm
Otherwise, |(LTm
y 0 ∈Z2 LTm (y ))| ≤
m
0
0 (y1 ) \ LT 0 (z)| + |LT 0 (y2 ) \ LT 0 (yp )| + |LT 0 (ys )| ≤ 3(d + d ).
|LTm
m
m
m
m
S
0
0 (y)\LT 0 (z))\
0
We have now shown that |Z2 | ≤ 2 and |(LTm
y 0 ∈Z2 LTm (y ))| ≤
m
0
0
4(d + d ). The final property of Z2 we wish to show is that for any y ∈ Z2 and
any T 0 ∈ P ∗ , we have that x ∈
/ LT 0 (yT0 0 ).
0
0 (y2 ) \
So suppose x ∈ LT 0 (yT 0 ). Let ŷ1 , . . . , ŷd+d0 be d + d0 labels in LTm
0
0
0
0
0 (y ). Let w1 , . . . , wd+d0 be d+d labels in LT 0 (y ). Then for each i ∈ [d+d ],
LTm
m
0
0
Tm (and therefore T2 ) contains the triplet (ŷi wi |x), but T contains the triplet
(xwi |ŷi ). Therefore conf set(T 0 , T2 ) contains d + d0 sets whose only common
element is x. As this implies a contradiction, we must have x ∈
/ LT 0 (yT0 0 ).
Now that Z1 and Z2 have been constructed, let Z = Z1 ∪ Z2 . Note that
|Z| ≤ 4. Algorithm 4 describes the construction of y and Z formally (see
Figure 5).
We have shown above that for any T 0 ∈ P ∗ , x isSdescended from yT 0 in T 0 and
0
0
0
0 (z) \
not
any z 0 ∈ Z, and so x ∈ LT 0 (yT 0 ) \ z0 ∈Z
|LTm
S from zT 0 for
S LT (zT 0 ). As
0
0
0
0
0
0
0
L
(z
))|
≤
4(d
+
d
)
and
|(L
(y)
\
L
(z))
\
L
(y
))|
≤
4(d +
Tm
Tm
z 0 ∈Z1 Tm
y 0 ∈Z2 Tm
S
0
0
0 (y) \
0
d0 ), we have |LTm
L
(z
))|
≤
8(d
+
d
).
0
z ∈Z Tm
0 (u)|
To analyze the complexity, note that we can calculate the value of |LTm
for all u in O(n) time using a depth-first search approach, together with the
0 (u)| = |LT 0 (u1 )| + |LT 0 (u2 )| for any node u with children u1 , u2 .
fact that |LTm
m
m
Then we can find z in O(n) time, and once we have found z we can find y, and
thence z1 , z2 , y1 , y2 , in O(n) time. Similarly, once these nodes are found we can
32
y
y1
z
z1
y2
Z2
z2
Z1
x
Figure 5: Construction of y and Z = Z1 ∪ Z2 , see Algorithm 4. Dashed edges
represent parts of the tree that were omitted. Triangles represent parts of the
tree that may contain more than d + d0 leaves.
find the members of Z in O(n) time.
Using the claim, we may now construct a set P 0 of O ≤ 16(d+d0 )+8 trees on
X \(Xm \{x}), such for any T 0 ∈ P ∗ , P 0 contains the tree T 00 = T 0 −(Xm \{x}).
Indeed, letSF be the set of arcs uv in Tm thatSexist on a path from y to a node in
(LTm (y) \ z∈Z LTm (z 0 )) ∪ Z. As |LTm (y) \ z0 ∈Z LTm (z 0 )| ≤ 8(d + d0 ), |Z| ≤ 4
and Tm is a binary tree, we have |F | ≤ 16(d + d0 ) + 8. For each e ∈ F , let Te be
the tree obtained from Tm by grafting x onto the arc e . Let P 0 = {Te : e ∈ F }.
Let T 0 be a tree in P ∗ and consider T 00 = T 0 − (Xm \ {x}). Note that
00
T − {x} = T 0 − Xm = Tm . Therefore uT 00 is well-defined for every node
u ∈ V (Tm ), and every node in T 00 is equal to uT 00 for some u ∈ V (Tm ), except
for x and its parent in T 00 . So let w be the parent of x in T 00 , uT 00 the parent
of w in T 00 , and vT 00 the child of w in T 00 that is not x. Observe that T 00 can
be obtained from Tm by grafting x onto the arc uv. Then it is enough to show
that uv ∈ F .
To see that uv ∈ F , first note that for each z 0 ∈ {y} ∪ Z, zT0 00 is well-defined
and (zT0 00 )T 0 = zT0 0 (see Lemma 9). Then as x is descended from (yT 00 )T 0 = yT 0 in
T 0 , x is descended from yT 00 in T 00 (Lemma 8). Similarly, as x is not descended
from (zT0 00 )T 0 = zT0 0 inST 0 for any z 0 ∈ Z, x is not descended from zT 00 in T 00 .
Thus x ∈ LT 00 (yT 00 ) \ z0 ∈Z LT 00 (zT0 00 ). It follows that uT 00 is a descendant of
yT 00 in T 00 (note that yT 00 6= w, as w is not the least common ancestor of any set
0
of labels
S in X (Tm )). Also, vT 00 is not a descendant of zT 00 for any z ∈ Z, unless
00
00
00
vT ∈ z∈Z zT , as otherwise x would be a descendant
of such a zT . SThus, vT 00
S
is either a member or an ancestor of (LT 00 (yT 00 ) \ z0 ∈Z LT 00 (zT0 00 ))) ∪ z0 ∈Z zT0 00 .
It follows using
is a descendant in Tm of y, and v is an ancestor
S Lemma 8 that uS
of (LTm (y) \ z0 ∈Z LTm (z 0 ))) ∪ z0 ∈Z z 0 . Then uv ∈ F , as required.
Now that we have constructed our set P 0 , it remains to find, for each Te ∈ P 0 ,
every tree T 0 on X such that T 0 − (Xm \ {x}) = Te and T 0 − {x} = T1 − {x}.
33
This will give us our set P , as for every T 0 ∈ P ∗ , T 0 − (Xm \ {x}) is a tree Te
in P 0 , and T 0 − {x} = T1 − {x}.
Let e = uv, where u, v ∈ V (Tm ), and let T1e be the subtree of T1 − {x}
whose root is v, and has as its label set v together with all labels in Xm \ {x}
descended from u but not v. Then we have to try every way of adding x into this
tree. If T1e contains t labels from Xm , then there are 2t − 1 places to try adding
x. Therefore P will have at most 2|Xm | ≤ 2(d + d0 ) additional trees compared
to P 0 , and so |P | ≤ 18(d + d0 ) + 8. Algorithm 5 gives the full procedure to
construct P .
Algorithm 5 FPT algorithm to find candidate trees for (T1 , Ti , x)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
procedure candidate-trees(T1 , T2 , x, d, d0 )
T1 , T2 are two trees, x is a label that cannot be moved in T2 (but must be
moved in T1 ), d is the maximum number of leaves we can remove in a tree, d0
is the maximum number of leaves we can move in T1 . For this pseudocode,
every tree T has a degree-1 root r(T ).
0
0
0
= (T2 −
| ≤ d0 + d − 1 and (T1 − {x}) − Xm
such that |Xm
Find Xm
0
{x}) − Xm
0
∪ {x}
Set Xm = Xm
Set Tm = T1 − Xm
Set (y, Z) = location-restriction(Tm , T2 , Xm , x, d, d0 )
. y, Z are
nodes in Tm such that roughly speaking, we may assume x must become a
descendant of y but not of any z 0 ∈ Z.
Set U = {u ∈ V (Tm ) : u ∈ Z or u is a leaf descended from y but not
from any z 0 ∈ Z}
Set F = {uv ∈ E(Tm ) : uv is on a path from y to U } . F is the set of
edges we could graft x onto.
Set P = ∅
. Given F we now begin constructing P .
Set T10 = T1 − {x}
for e = uv ∈ F with u the parent of v do
. Try grafting x on e
Set uT10 = the node in T10 corresponding to u
Set vT10 = the node in T10 corresponding to v
Set Xe = set of labels l in Xm \ {x} for which l has an ancestor v 0 in
T10 with v 0 descended from uT10 , vT10 descended from v 0
. Xe is the set of leaves of T1 for which we have to subdivide e.
Set U = vT10 ∪ Xe
Set Ee = {u0 v 0 ∈ E(T10 ) : u0 v 0 is on a path from uT10 to U }
for u0 v 0 ∈ Ee do
Constuct T 0 from T10 by grafting x on u0 v 0
Set P = P ∪ {T 0 }
Return P
To analyze the complexity, recall that we find Xm , and therefore construct
0
Tm and Tm
, in O(n log n) time. As shown above, we can find the node y and
set Z in O(n) time. Given y and Z, the set of arcs F can be found in O(n)
34
time using a depth-first search approach. For each e ∈ F it takes O(n) time to
construct Te , and so the construction of P 0 takes O(|F |n) = O((16(d+d0 )+8)n)
time. Finally, the construction of of P from P 0 takes O(|P |n) = O((18(d + d0 ) +
8)n) time. Putting it all together, we have that the construction of P takes
O(n(log n + 18(d + d0 ) + 8)) time.
We will call the set of trees P described in Lemma 10 the set of candidate
trees for (T1 , T2 , x).
We are finally ready to give the proof of Theorem 6
Theorem 6 (restated). AST-LR-d can be solved in time O(cd d3d (n3 +tn log n)),
where c is a constant not depending on d or n.
Proof. The outline for our algorithm is as follows. We employ a branch-andbound algorithm, in which at each step we attempt to modify the input tree
T1 to become close to a solution. We keep track of an integer d0 , representing
the maximum length of an LPR sequence between T1 and a solution. Initially
set d0 = d. At each step, if dLR (T1 , Ti ) ≤ d for each Ti ∈ T then T1 is a
solutioon, and we are done. Otherwise, there must exist sime Ti for which
dLR (T1 , Ti ) ≥ d + d0 . In this case, we calculate the (d + d0 ) disagreement kernel
S between T1 and Ti (using the procedure of Lemma 7), and for each x ∈ S,
attempt to construct a set P of trees as in Lemma 10. For each T 0 ∈ P , we try
replacing T1 with T 0 , reducing d0 by 1, and repeating the procedure. Algorithm 6
describes the full procedure formally.
We claim that Algorithm 6 is a correct algorithm for AST-LR-d, and runs in
time O(cd d3d (n2 + tn log n)), for some constant c not depending on n or d.
First notice that if, in a leaf node of the branch tree created by Algorithm 6,
a tree T ∗ is returned, this occurs at line 3 in which case it has been verified that
T ∗ is indeed a solution. As an internal node of the branch tree returns a tree if
and only if a child recursive call also returns a tree (the for loop on line 9), this
shows that when the algorithm outputs a tree T ∗ , it is indeed a solution.
We next show that if a solution exists, then Algorithm 6 will return one.
Suppose that T admits a solution, and let T ∗ be a solution that minimizes
d1 := dLR (T1 , T ∗ ), with d1 ≤ d0 . We show that one leaf of the branch tree
created by the algorithm returns T ∗ (and thus the root of the branch tree also
returns a solution, albeit not necessarily T ∗ ). This is done by proving that in
one of the recursive calls made to mastrl-distance on line 13, the tree T 0
obtained from T1 satisfies dLR (T 0 , T ∗ ) = d1 − 1. By applying this argument
inductively, this shows that the algorithm will find T ∗ at some node of depth
d1 in the branch tree of the algorithm.
First notice that since dLR is a metric, for each Ti ∈ T , dLR (T1 , Ti ) ≤
dLR (T1 , T ∗ )+dLR (T ∗ , Ti ) ≤ d0 +d, and so the algorithm will not return F ALSE
on line 5.
If T1 isn’t a solution, then there is a tree of T , say T2 w.l.o.g., such that
dLR (T1 , T2 ) > d. Notice that in this case, all the conditions of Lemma 6 are
satisfied, i.e. dLR (T1 , T2 ) > d, and there are sets X1 , X2 ⊆ X both of size at
most d such that T1 − X1 = T ∗ − X1 , T2 − X2 = T ∗ − X2 . Thus there is a
35
minimal disagreement X between T1 and T2 , |X| ≤ d0 + d, and x ∈ X such that
x ∈ X1 \ X2 . By Lemma 3, there is an LPR sequence L = (x1 , . . . , xk ) turning
T1 into T ∗ , where {x1 , . . . , xk } = X1 . As x ∈ X1 , by Lemma 4, the leaves
appearing in L can be reordered, and we may assume that x = x1 . Finally by
Lemma 10, if T 0 satisfies dLR (T 0 , T ∗ ) ≤ d1 − 1 and T 0 can be obtained from T1
by an LPR move on x, then T 0 ∈ P . As we are making one recursive call to
mastrl-distancefor each tree in P , this proves that one such call replaces T1
by T 0 such that dLR (T 0 , T ∗ ) = d1 − 1.
As for the complexity, recall from Lemma 7 that the (d + d0 )-disagreement
kernel S computed in line 8 contains at most 8d2 labels.Therefore when Algorithm 6 enters the ’for’ loop of line 9, it branches into at most 8d2 cases, one
for each x ∈ S. Within each of these cases, the algorithm enters at most |P |
recursive calls, each of which decrements d0 . As |P | ≤ 18(d+d0 )+8 ≤ 36d+8 by
Lemma 10, a single call of the algorithm splits into at most 8d2 (36d+8) = O(d3 ),
each of which decrements d0 . Therefore, the branching tree created by the algorithm has degree at most cd3 (for some constant c) and depth at most d, and
so O(cd d3d ) cases are considered.
As dLR (T1 , Ti ) can be calculated in O(n log n) time for each Ti , a single call
of lines 2-5 of the algorithm takes O(tn log n) time. A single call of lines 6-8
takes O(n2 ) time by Lemma 10. Thus the total time for all calls of lines 2-8 is
O(cd d3d n(n2 + t log n). Each call of line 10 occurs just before a recursive call to
the algorithm, as so line 10 is called at most O(cd d3d ) times. A single call of line
10 takes O(n(log n+18(d+d0 )+8)) = O(n(log n+36d)) time by Lemma 10, and
so the total time for all calls of line 10 is O(cd d3d n(log n + 36d)). Thus in total,
we have that the running time of the algorithm is O(cd d3d (n2 + n(t log n + 36d)).
As we may assume d ≤ n, this simplifies to O(cd d3d (n2 + tn log n)).
36
Algorithm 6 FPT algorithm for parameter d.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
procedure mastrl−distance(T = (T1 , T2 , . . . , Tt ), d, d0 )
T is the set of input trees (represented as a sequence to distinguish T1 from
the other trees), d is the maximum number of leaves we can remove in a
tree, d0 is the maximum number of leaves we can move in T1 , which should
be initially set to d.
if dLR (T1 , Ti ) ≤ d for each Ti ∈ T then
Return T1
else if there is some Ti ∈ T such that dLR (T1 , Ti ) > d0 + d then
Return FALSE #handles the d0 ≤ 0 case
else
. here we ‘guess’ a leaf prune-and-regraft move on T1
Choose Ti ∈ T such that dLR (T1 , Ti ) > d
Set S = disagreement-kernel(d + d0 , T1 , Ti )
for x ∈ S do . we are ‘guessing’ that x should go where Ti wants
it.
Set P = candidate-trees(T1 , Ti , x, d, d0 )
T ∗ = F ALSE
for T ∈ P do
T 0 = mastrl−distance((T, T2 , . . . , Tt ), d, d0 − 1)
if T 0 is not F ALSE, let T ∗ := T 0
Return T ∗
37
| 8cs.DS
|
Submitted to the Annals of Statistics
arXiv: arXiv:1701.04889
EFFICIENT AND ADAPTIVE LINEAR REGRESSION IN
SEMI-SUPERVISED SETTINGS
arXiv:1701.04889v2 [stat.ME] 19 Aug 2017
By Abhishek Chakrabortty ∗,† and Tianxi Cai†
University of Pennsylvania and Harvard University
We consider the linear regression problem under semi-supervised
settings wherein the available data typically consists of: (i) a small or
moderate sized ‘labeled’ data, and (ii) a much larger sized ‘unlabeled’
data. Such data arises naturally from settings where the outcome, unlike the covariates, is expensive to obtain, a frequent scenario in modern studies involving large databases like electronic medical records
(EMR). Supervised estimators like the ordinary least squares (OLS)
estimator utilize only the labeled data. It is often of interest to investigate if and when the unlabeled data can be exploited to improve
estimation of the regression parameter in the adopted linear model.
In this paper, we propose a class of ‘Efficient and Adaptive SemiSupervised Estimators’ (EASE) to improve estimation efficiency. The
EASE are two-step estimators adaptive to model mis-specification,
leading to improved (optimal in some cases) efficiency under model
mis-specification, and equal (optimal) efficiency under a linear model.
This adaptive property, often unaddressed in the existing literature,
is crucial for advocating ‘safe’ use of the unlabeled data. The construction of EASE primarily involves a flexible ‘semi-non-parametric’
imputation, including a smoothing step that works well even when
the number of covariates is not small; and a follow up ‘refitting’ step
along with a cross-validation (CV) strategy both of which have useful practical as well as theoretical implications towards addressing
two important issues: under-smoothing and over-fitting. We establish
asymptotic results including consistency, asymptotic normality and
the adaptive properties of EASE. We also provide influence function
expansions and a ‘double’ CV strategy for inference. The results are
further validated through extensive simulations, followed by application to an EMR study on auto-immunity.
1. Introduction. In recent years, semi-supervised learning (SSL) has
emerged as an exciting new area of research in statistics and machine learning. A detailed discussion on SSL including its practical relevance, the pri∗
Corresponding author; previously at Harvard University during the time of this work.
This research was partially supported by National Institutes of Health grants R01
HL089778, U54 HG007963, and U01 HL121518.
MSC 2010 subject classifications: 62F35, 62J05, 62F12, 62G08.
Keywords and phrases: Semi-supervised Linear Regression, Semi-parametric Inference,
Model Mis-specification, Adaptive Estimation, Semi-non-parametric Imputation.
†
1
2
A. CHAKRABORTTY AND T. CAI
mary question of interest in SSL, and the existing relevant literature can
be found in Chapelle, Schölkopf and Zien (2006) and Zhu (2008). A typical semi-supervised (SS) setting is characterized by two types of available
data: (i) a small or moderate sized ‘labeled’ data, L, containing observations for both an outcome Y and a set of covariates X of interest, and (ii)
an ‘unlabeled’ data, U, of much larger size but having observations only
for the covariates X. By virtue of its large size, U essentially gives us the
distribution of X, denoted henceforth by PX . Such a setting arises naturally
whenever the covariates are easily available so that unlabeled data is plentiful, but the outcome is costly or difficult to obtain, thereby limiting the size
of L. This scenario is directly relevant to a variety of practical problems,
especially in the modern ‘big data’ era, with massive unlabeled datasets (often electronically recorded) becoming increasingly available and tractable.
A few familiar examples include machine learning problems like text mining,
web page classification, speech recognition, natural language processing etc.
Among biomedical applications, a particularly interesting problem where
SSL can be of great use is the statistical analysis of electronic medical records
(EMR) data. Endowed with a wealth of de-identified clinical and phenotype
data for large patient cohorts, EMR linked with bio-repositories are increasingly gaining popularity as rich resources of data for discovery research (Kohane, 2011). Such large scale datasets obtained in a cost-effective and timely
manner are of great importance in modern medical research for addressing
important questions such as the biological role of genetic variants in disease
susceptibility and progression (Kohane, 2011). However, one major bottleneck impeding EMR driven research is the difficulty in obtaining validated
phenotype information (Liao et al., 2010) since they are labor intensive or
expensive to obtain. Thus, gold standard labels and genomic measurements
are typically available only for a small subset nested within a large cohort. In
contrast, digitally recorded data on the clinical variables are often available
on all subjects, highlighting the necessity and utility of developing robust
SSL methods that can leverage such rich source of auxiliary information to
improve phenotype definition and estimation precision.
SSL primarily distinguishes from standard supervised methods by making
use of U, an information that is ignored by the latter. The ultimate question
of interest in SSL is to investigate if and when the information on PX in U
can be exploited to improve the efficiency over a given supervised approach.
In recent years, several graph based non-parametric SSL approaches have
been proposed (Zhu, 2005; Belkin, Niyogi and Sindhwani, 2006) for regression or classification. These approaches essentially target non-parametric SS
estimation of E(Y |X) and therefore, for provable improvement guarantees,
SEMI-SUPERVISED LINEAR REGRESSION
3
must rely implicitly or explicitly on assumptions relating PX to PY |X (the
conditional distribution of Y given X), as duly noted and characterized more
formally in Lafferty and Wasserman (2007). For non-parametric classification problems, the theoretical underpinnings of SSL including its scope and
the consequences of using U have been also studied earlier by Castelli and
Cover (1995, 1996). More parametric SS approaches, still aimed mostly at
prediction, have also been studied for classification, including the ‘generative
model’ approach (Nigam et al., 2000; Nigam, 2001) which is based on modeling the joint distribution of (Y, X) as an identifiable mixture of parametric
models, thereby implicitly relating PY |X and PX . However, these approaches
depend strongly on the validity of the assumed mixture model, violation of
which can actually degrade their performance compared to the supervised
approach (Cozman and Cohen, 2001; Cozman, Cohen and Cirelo, 2003).
However SS estimation problems, especially from a semi-parametric point
of view, has been somewhat less studied in SSL. Such problems are generally
aimed at estimating some (finite-dimensional) parameter θ0 ≡ θ0 (P), where
P = (PY |X , PX ), and the key to the potential usefulness of U in improving
estimation of θ0 lies in understanding when θ0 (P) relates to PX . For simple
parameters like θ0 (P) = E(Y ), unless E(Y |X) is a constant, θ0 clearly depends on PX and hence, improved SS estimation is possible compared to the
supervised estimator Y L , the sample mean of Y based on L. The situation is
however more subtle for other choices of θ0 , especially those where θ0 is the
target parameter corresponding to an underlying parametric working model
for PY |X . This includes the least squares parameter, as studied in this paper, targeted by a working linear model for E(Y |X). Such models are often
adopted due to their appealing simplicity and interpretability.
In general, for such cases, if the adopted working model for PY |X is correct and θ0 is not related to PX , then one cannot possibly gain through SSL
by using the knowledge of PX (Zhang and Oles, 2000; Seeger, 2002). On
the other hand, under model mis-specification, θ0 may inherently depend on
PX , and thus imply the potential utility of U in improving the estimation.
However, inappropriate use of U may lead to degradation of the estimation
precision. This therefore signifies the need for robust and efficient SS estimators that are adaptive to model mis-specification, so that they are as efficient
as the supervised estimator under the correct model and more efficient under model mis-specification. To the best of our knowledge, work done along
these lines is relatively scarce in the SSL literature, one notable exception
being the recent work of Kawakita and Kanamori (2013), where they use a
very different approach based on density ratio estimation, building on the
more restrictive approach of Sokolovska, Cappé and Yvon (2008). However,
4
A. CHAKRABORTTY AND T. CAI
as we observe in our simulation studies, the extent of the efficiency gain actually achieved by these approaches can be quite incremental, at least in finite
samples. Further, the seemingly unclear choice of the ideal (nuisance) model
to be used for density ratio estimation can also have a significant impact on
the performance, both finite sample and asymptotic, of these estimators.
We propose here a class of Efficient and Adaptive Semi-Supervised Estimators (EASE) in the context of linear regression problems. We essentially
adopt a semi-parametric perspective wherein the adopted linear ‘working’
model can be potentially mis-specified, and the goal is to obtain efficient and
adaptive SS estimators of the regression parameter through robust usage of
U. The EASE are two-step estimators with a simple and scalable construction based on a first step of ‘semi-non-parametric’ (SNP) imputation which
includes a smoothing step and a follow-up ‘refitting’ step. In the second step,
we regress the imputed outcomes against the covariates using the unlabeled
data to obtain our SNP imputation based SS estimator, and then further
combine it optimally with the supervised estimator to obtain the final EASE.
Dimension reduction methods are also employed in the smoothing step to
accommodate higher dimensional X, if necessary. Further, we extensively
adopt cross-validation (CV) techniques in the imputation, leading to some
useful theoretical properties (apart from practical benefits) typically not observed for smoothing based two-step estimators. We demonstrate that EASE
is guaranteed to be efficient and adaptive in the sense discussed above, and
also achieves semi-parametric optimality whenever the SNP imputation is
‘sufficient’ or the linear model holds. We also provide data adaptive methods
to optimally select the directions for smoothing when dimension reduction
is desired, and tools for inference with EASE.
The rest of this paper is organized as follows. In Section 2, we formulate
the SS linear regression problem. In Section 3, we construct a family of SS
estimators based on SNP imputation and establish all their properties, and
further propose the EASE as a refinement of these estimators. For all our
proposed estimators, we also address their associated inference procedures
based on ‘double’ CV methods. In Section 4, we discuss a kernel smoothing
based implementation of the SNP imputation and establish all its properties. In Section 5, we discuss SS dimension reduction techniques, useful for
implementing the SNP imputation. Simulation results and an application to
an EMR study are shown in Section 6, followed by concluding discussions
in Section 7. Proofs of all theoretical results and associated technical materials, and further numerical results and discussions are distributed in the
Appendix and the Supplementary Material [Chakrabortty and Cai (2017)].
SEMI-SUPERVISED LINEAR REGRESSION
5
2. Problem Set-up.
Data Representation. Let Y ∈ R denote the outcome random variable and
X ∈ Rp denote the covariate vector, where p is fixed, and let Z = (Y, X0 )0 .
Then the entire data available for analysis can be represented as S = (L∪U),
where L = {Zi ≡ (Yi , X0i )0 : i = 1, . . . , n} consists of n independent and
identically distributed (i.i.d.) observations from the joint distribution PZ of
Z, U = {Xi : i = n + 1, . . . , n + N } consists of N i.i.d. observations from PX ,
and L ⊥
⊥ U. Throughout, for notational convenience, we use the subscript ‘j’
to denote the unlabeled observations, and re-index without loss of generality
(w.l.o.g.) the N observations in U as: U = {Xj : j = n + 1, . . . , n + N }.
Assumption 2.1 (Basic Assumptions). (a) We assume that Z has finite
2nd moments and Σ ≡ Var(X) is positive definite, denoted as Σ 0. We
also assume, for simplicity, that X has a compact support X ⊆ Rp .
(b) We assume N n i.e. n/N → 0 as n, N → ∞, and L and U arise
from the same underlying distribution, i.e. Z ∼ PZ for all subjects in S.
→
−→
−
−
Notations. Let Γ = E( X X 0 ) 0, where ∀ v ∈ Rp , →
v = (1, v0 )0 ∈ R(p+1) .
Let L2 (PX ) denote the space of all R-valued measurable functions of X
having finite L2 norm with respect to (w.r.t.) PX , and for any g(.) ∈ L2 (PX ),
→
−→
−
let Σ(g) 0 denote the (p+1)×(p+1) matrix Γ−1 E[ X X 0 {Y −g(X)}2 ]Γ−1 .
Lastly, let k · k denote the L2 vector norm, and for any integer a ≥ 1, let
Ia denote the identity matrix of order a, and Na [µ, Ω] denote the a-variate
Gaussian distribution with mean µ ∈ Ra and covariance matrix Ωa×a 0.
Remark 2.1. Assumption 2.1 (b) enlists some fundamental characteristics of SS settings. Indeed, the condition of L and U being equally distributed has usually been an integral part of the definition of SS settings
(Chapelle, Schölkopf and Zien, 2006; Kawakita and Kanamori, 2013). Interpreted in missing data terminology, it entails that Y in U are ‘missing
completely at random’ (MCAR), with the missingness/labeling being typically by design. Interestingly, the crucial assumption of MCAR, although
commonly required, has often stayed implicit in the SSL literature (Lafferty
and Wasserman, 2007). It is important to note that while the SS set-up can
be viewed as a missing data problem, it is quite different from standard
ones, since with n/N → 0 i.e. |U| |L|, the proportion of Y observed in S
tends to 0 in SSL. Hence, the ‘positivity assumption’ typical in missing data
theory, requiring this proportion to be bounded away from 0, is violated
here. It is also worth noting that owing to such violations, the analysis of
SS settings under more general missingness mechanisms such as ‘missing at
6
A. CHAKRABORTTY AND T. CAI
random’ (MAR) is considerably more complicated and to our knowledge, the
literature for SS estimation problems under such settings is virtually nonexistent. Furthermore, for such problems, the traditional goal in SSL, that
of improving upon a ‘supervised’ estimator, can become unclear without
MCAR, unless an appropriately weighted version of the supervised estimator is considered. Given these subtleties and the traditional assumptions
(often implicit) in SSL, the MCAR condition is assumed for most of this
paper, although a brief discussion on possible extensions of our proposed SS
estimators to MAR settings is provided in the Supplementary Material.
2.1. The Target Parameter and Its Supervised Estimator. We consider
the linear regression working model given by:
→
−
(2.1)
Y = X 0 θ + , with E( | X) = 0,
where, θ ∈ R(p+1) is an unknown regression parameter. Accounting for the
potential mis-specification of the working model (2.1), we define the target
parameter of interest as a model free parameter, as follows:
Definition 2.1. The target parameter θ 0 for linear regression may be
→
−
→
−
defined as the solution to the normal equations: E{ X(Y − X 0 θ)} = 0 in
→
−
θ ∈ R(p+1) , or equivalently, θ 0 = argmin E(Y − X 0 θ)2 .
θ∈R(p+1)
→
−
Existence and uniqueness of θ 0 in 2.1 is clear. Further, X 0 θ 0 is the L2
projection of E(Y |X) ∈ L2 (PX ) onto the subspace of all linear functions of X
and hence, is the best linear predictor of Y given X. The linear model (2.1) is
correct (else, mis-specified ) if and only if E(Y |X) lies in this space (in which
→
−
case, E(Y |X) = X 0 θ 0 ). When the model is correct, θ 0 depends only on PY |X ,
not on PX . Hence, improved estimation of θ 0 through SSL is impossible in
this case unless further assumptions relating θ 0 to PX are made. On the
other hand, under model mis-specification, the normal equations defining
θ 0 inherently depend on PX , thereby implying the potential utility of SSL
in improving the estimation of θ 0 in this case.
b the solution
The usual supervised estimator of θ 0 is the OLS estimator θ,
Pn →
−
→
−0
−1
in θ to the equation: n
i=1 X i (Yi − X i θ) = 0, the normal equations
based on L. Under Assumption 2.1 (a), it is well known that as n → ∞,
(2.2)
1
1
b − θ 0 ) = n− 2
n 2 (θ
n
X
1
d
ψ 0 (Zi ) + Op n− 2
→ N(p+1) [0, Σ(gθ0 )],
i=1
→
−
→
−
→
−
where ψ 0 (Z) = Γ−1 { X(Y − X 0 θ 0 )} and gθ (X) = X 0 θ ∀ θ ∈ R(p+1) .
SEMI-SUPERVISED LINEAR REGRESSION
7
Our primary goal is to obtain efficient SS estimators of θ 0 using the entire
b It is worth noting that the
data S and compare their efficiencies to that of θ.
estimation efficiency of θ 0 also relates to the predictive performance of the
fitted linear model since its out-of-sample prediction error is directly related
to the mean squared error (w.r.t. the Σ metric) of the parameter estimate.
3. A Family of Imputation Based Semi-Supervised Estimators.
If Y in U were actually observed, then one would simply fit the working
model to the entire data in S for estimating θ 0 . Our general approach is
precisely motivated by this intuition. We first attempt to impute the missing
Y in U based on suitable training of L in step (I). Then in step (II), we fit the
linear model (2.1) to U with the imputed outcomes. Clearly, the imputation
is critical. Inaccurate imputation would lead to biased estimate of θ 0 , while
inadequate imputation would result in loss of efficiency. We next consider SS
estimators constructed under two imputation strategies for step (I) including
a fully non-parametric imputation based on kernel smoothing (KS), and a
semi-non-parametric (SNP) imputation that involves a smoothing step and
a follow up ‘refitting’ step. Although the construction of the final EASE is
based on the SNP imputation strategy, it is helpful to begin with a discussion
of the first strategy in order to appropriately motivate and elucidate the
discussion on EASE and the SNP imputation strategy.
3.1. A Simple SS Estimator via Fully Non-Parametric Imputation. We
present here an estimator based on a fully non-parametric imputation involving KS when p is small. For simplicity, we shall assume here that X is continuous with a density f (·). Let m(x) = E(Y | X = x) and l(x) = m(x)f (x).
Consider the local constant KS estimator of m(x),
1 Pn
b
l(x)
i=1 {Kh (Xi , x)}Yi
nhp
(3.1)
m(x)
b
=
=
,
P
n
1
fb(x)
i=1 Kh (Xi , x)
nhp
where Kh (u, v) = K{(u−v)/h} with K : Rp → R being some suitable kernel
function and h = h(n) > 0 being the bandwidth. With m(·)
b
as defined in
(3.1), we now fit (2.1) to the imputed unlabeled data: [{m(X
b j ), X0j }0 : j =
bnp of θ 0 as the solution in θ to:
n + 1, ..., n + N ] and obtain a SS estimator θ
(3.2)
n+N
−
→
−
1 X →
X j {m(X
b j ) − X 0j θ} = 0.
N
j=n+1
Here and throughout in our constructions of SS estimators, L with either
the true or the imputed Y is not included in the final fitting step mostly
8
A. CHAKRABORTTY AND T. CAI
due to technical convenience in the asymptotic analysis of our estimators,
and also due to the fact that the contribution of L, included in any form, in
the final fitting step is asymptotically negligible since n/N → 0.
bnp , we require uniform (in L∞ norm)
In order to study the properties of θ
convergence of m(·)
b
to m(·), a problem that has been extensively studied
in the non-parametric statistics literature (Newey, 1994; Andrews, 1995;
Masry, 1996; Hansen, 2008) under fairly general settings and assumptions.
In particular, we would assume the following regularity conditions to hold:
Assumption 3.1. (i) K(·) is a symmetric q th order kernel for some integer q ≥ 2. (ii) K(·) is bounded, Lipschitz continuous and has a bounded
support K ⊆ Rp . (iii) E(|Y |s ) < ∞ for some s > 2. E(|Y |s | X = x)f (x) and
f (x) are bounded on X . (iv) f (x) is bounded away from 0 on X . (v) m(·)
and f (·) are q times continuously differentiable with bounded q th derivatives
on some open set X0 ⊇ X . (vi) For any δ > 0, let Aδ ⊆ Rp denote the set
{(x − X)/δ : x ∈ X }. Then, for small enough δ, Aδ ⊇ K almost surely (a.s.).
Conditions (i)-(v) are fairly standard in the literature. In (v), the set X0 is
needed mostly to make the notion of differentiability well-defined, with both
m(·) and f (·) understood to have been analytically extended over (X0 \X ).
Condition (vi) implicitly controls the tail behaviour of X, requiring that
perturbations of X in the form of (X + δφ) with φ ∈ K (bounded) and δ
bnp .
small enough, belong to X a.s. [PX ]. We now present our result on θ
1
1
Theorem 3.1. Suppose n 2 hq → 0 and (log n)/(n 2 hp ) → 0 as n → ∞,
1
1
1
and let rn = n 2 hq + (log n)/(n 2 hp ) + (n/N ) 2 . Then, under Assumption 3.1,
(3.3)
n
X
1
d
bnp − θ 0 = n− 12
n2 θ
ψ eff (Zi ) + Op (rn ) → N(p+1) [0, Σ(m)],
i=1
→
−
where ψ eff (Z) = Γ−1 [ X{Y − m(X)}].
Remark 3.1. Theorem 3.1 establishes the efficient and adaptive nabnp . The asymptotic variance Σ(m) of θ
bnp satisfies Σ(g) − Σ(m)
ture of θ
2
0 ∀ g(·) ∈ L (X) and the inequality is strict unless g(·) = m(·) a.s. [PX ].
bnp is asymptotically optimal among the class of all regular and
Hence, θ
asymptotically linear (RAL) estimators of θ 0 with influence function (IF) of
→
−
bnp is more
the form: Γ−1 [ X{Y − g(X)}] with g(·) ∈ L2 (PX ). In particular, θ
b
efficient than θ whenever (2.1) is mis-specified, and equally efficient when
(2.1) is correct i.e. m(·) = gθ0 (·). Further, it can also be shown that ψ eff (Z)
SEMI-SUPERVISED LINEAR REGRESSION
9
is the ‘efficient’ IF for estimating θ 0 under the semi-parametric model MX ≡
{(PY |X , PX ) : PX is known, PY |X is unrestricted upto Assumption 2.1 (a)}.
bnp also globally achieves the semi-parametric efficiency bound under
Thus, θ
MX . Lastly, note that at any parametric sub-model in MX that corresponds
b also achieves optimality, thus showing that under
to (2.1) being correct, θ
b if the linear model is correct.
MX , it is not possible to improve upon θ
Remark 3.2. The asymptotic results in Theorem 3.1 require a kernel of
order q > p and h smaller in order than the ‘optimal’ bandwidth order hopt =
O(n−1/(2q+p) ). This under-smoothing requirement, often encountered in twostep estimators involving a first-step smoothing (Newey, Hsieh and Robins,
1998), generally results in sub-optimal performance of m(.).
b
The optimal
under-smoothed bandwidth order for Theorem 3.1 is given by: O(n−1/(q+p) ).
3.2. SS Estimators Based on Semi-Non-Parametric (SNP) Imputation.
The simple and intuitive imputation strategy in Section 3.1 based on a fully
non-parametric p-dimensional KS is however often undesirable in practice
owing to the curse of dimensionality. In order to accommodate larger p, we
now propose a more flexible SNP imputation method involving a dimension
reduction, if needed, followed by a non-parametric calibration. An additional
‘refitting’ step is proposed to reduce the impact of bias from non-parametric
estimation and possibly inadequate imputation due to dimension reduction.
We also introduce some flexibility in terms of the smoothing methods, apart
from KS, that can be used for the non-parametric calibration.
Let r ≤ p be a fixed positive integer and let Pr = [p1 , .., pr ]p×r be any
rank r transformation matrix. Let XPr = P0r X. Given (r, Pr ), we may now
consider approximating the regression function E(Y |X) by smoothing Y
over the r dimensional XPr instead of the original X ∈ Rp . In general, Pr
can be user-defined and data dependent. A few reasonable choices of Pr
are discussed in Section 5. If Pr depends only on the distribution of X, it
may be assumed to be known given the SS setting considered. If Pr also
depends on the distribution of Y , then it needs to be estimated from L and
the smoothing needs to be performed using the estimated Pr .
For approximating E(Y |X), we may consider any reasonable smoothing
technique T . Some examples of T include KS, kernel machine regression and
smoothing splines. Let m(x; Pr ) denote the ‘target function’ for smoothing
Y over XPr using T . For notational simplicity, the dependence of m(x; Pr )
and other quantities on T is suppressed throughout. For T := KS, the appropriate target function is given by: m(x; Pr ) = mPr (P0r x), where mPr (z) ≡
E(Y | XPr = z). For basis function expansion based methods, m(x; Pr ) will
typically correspond to the L2 projection of m(x) ≡ E(Y | X = x) ∈ L2 (PX )
10
A. CHAKRABORTTY AND T. CAI
onto the functional space spanned by the basis functions associated with
T . The results in this section apply to any choice of T that satisfies the
required conditions. In Section 4, we provide more specific results for the
implementation of our methods using T := KS.
Note that we do not assume m(x; Pr ) = m(x) anywhere, and hence the
name ‘semi-non-parametric’. Clearly, with Pr = Ip and T := KS, it reduces
to a fully non-parametric approach. We next describe the two sub-steps
involved in step (I) of the SNP imputation: (Ia) smoothing, and (Ib) refitting.
b r and
(Ia) Smoothing Step. With Pr and m(x; Pr ) as defined above, let P
b
m(x;
b
Pr ) respectively denote their estimators based on L. In order to address
potential overfitting issues in the subsequent steps, we further consider generalized versions of these estimators based on K-fold CV for a given fixed integer K ≥ 1. For any K ≥ 2, let {Lk }K
k=1 denote a random partition of L into
−
K disjoint subsets of equal sizes, nK = n/K, with index sets {Ik }K
k=1 . Let Lk
−
denote the set excluding Lk with size nK = n − nK and respective index set
b r,k and m
b r,k ) denote the corresponding estimators based on
Ik− . Let P
b k (x; P
−
Lk . Further, for notational consistency, we define for K = 1, Lk = L−
k = L;
−
−
b
b
b
b r ).
b k (x; Pr,k ) = m(x;
b
P
Ik = Ik = {1, ..., n}; nK = nK = n; Pr,k = Pr and m
(Ib) Refitting Step. In this step, we fit the linear model to L using X as predictors and the estimated m(X; Pr ) as an offset. To motivate this, we recall
that the fully non-parametric imputation given in Section 3.1 consistently
estimates E(Y |X), the L2 projection onto a space that always contains the
→
−
working model space, i.e. the linear span of X. This need not be true for
the SNP imputation, since we do not assume m(X; Pr ) = m(X) necessarily.
The refitting step essentially ‘adjusts’ for this so that the final imputation,
combining the predictions from these two steps, targets a space that contains the working model space. In particular, for T := KS with r < p, this
step is critical to remove potential bias due to inadequate imputation.
Interestingly, it turns out that the refitting step should always be performed, even when m(X; Pr ) = m(X). It plays a crucial role in reducing
the bias of the resulting SS estimator due to the inherent bias from nonparametric curve estimation. In particular, for T := KS with any r ≤ p, it
ensures that a bandwidth of the optimal order can be used, thereby eliminating the under-smoothing issue as encountered in Section 3.1. The target
parameter for the refitting step is simply the regression coefficient obtained
from regressing the residual Y −m(X; Pr ) on X and may be defined as: η Pr ,
→
−
→
−
the solution in η ∈ R(p+1) to the equation: E[ X{Y − m(X; Pr ) − X 0 η}] = 0.
SEMI-SUPERVISED LINEAR REGRESSION
11
b (Pr ,K) , the solution in η to the equation:
For any K ≥ 1, we estimate η Pr as η
(3.4)
n
−1
K X
X
→
−
−0
b r,k ) − →
X i {Yi − m
b k (Xi ; P
X i η} = 0.
k=1 i∈Ik
For Xi ∈ Lk , the estimate of m(Xi ; Pr ) to be used as an offset is obtained
b r,k ) that is based on data in L− . For K ≥ 2, with L− ⊥
from m
b k (· ; P
k
k ⊥ Lk , the
residuals are thus estimated in a cross-validated manner. For K = 1 however,
b r ) is estimated using the entire L which can lead to considerable
m(·
b ;P
underestimation of the true residuals owing to over-fitting and consequently,
substantial finite sample bias in the resulting SS estimator of θ 0 . This bias
can be effectively reduced by using the CV approach with K ≥ 2. We next
estimate the target function for the SNP imputation given by:
−
µ(x; P ) = m(x; P ) + →
x 0η
(3.5)
as:
r
(3.6)
br,K ) = K−1
µ
b(x; P
r
K
X
Pr
−
b r,k ) + →
b (Pr ,K) ,
m
b k (x; P
x 0η
k=1
b r,k }K . For notational simplicity, we suppress throughout
br,K = {P
where P
k=1
K
the inherent dependence of µ
b(· ; ·) itself on K and {L−
k }k=1 . Note that
similar to m(X; Pr ), we also do not assume µ(X; Pr ) = m(X). Apart from
the geometric motivation for the refitting step and its technical role in bias
→
−
reduction, it also generally ensures the condition: E[ X{Y − µ(X; Pr }] = 0,
regardless of the true underlying m(X). This condition is a key requirement
for the asymptotic expansions, in Theorem 3.2, of our resulting SS estimabr,K ), we now construct our final SS estimator as follows.
tors. Using µ
b(· ; P
SS Estimator from SNP Imputation. In step (II), we fit the linear model to
br,K ), X0 }0 : j = n + 1, ..., n + N ]
the SNP imputed unlabeled data: [{b
µ(Xj ; P
j
b
and obtain a SS estimator θ (Pr ,K) of θ 0 given by:
n+N
−
−0
1 X →
b
br,K ) − →
(3.7) θ (Pr ,K) is the solution in θ to
X j {b
µ(Xj ; P
X j θ} = 0.
N
j=n+1
For convenience of further discussion, let us define: ∀ k ∈ {1, . . . , K},
(3.8)
(3.9)
b r,k ) = m
b r,k ) − m(x; Pr ) ∀ x ∈ X , and
b k (x; Pr , P
∆
b k (x; P
−b
−
b k (x) = →
b r,k ) − EX {→
b r,k )} ∀ x ∈ X ,
b k (x; Pr , P
G
x∆
X ∆k (X; Pr , P
b k (·)
where EX (·) denotes expectation w.r.t. X ∈ U. The dependence of G
b
on (Pr , Pr,k ) and PX is suppressed here for notational simplicity. We now
b(P ,K) .
present our main result summarizing the properties of θ
r
12
A. CHAKRABORTTY AND T. CAI
Theorem 3.2. Suppose that T satisfies: (i) supx∈X |m(x; Pr )| < ∞ and
b r ) − m(x; Pr )| = Op (cn ) for some cn = o(1). With G
b k (.)
(ii) supx∈X |m(x;
b
P
− 21 PK P
b
as in (3.9), define Gn,K = n
k=1
i∈Ik Gk (Xi ). Then, for any K ≥ 1,
(3.10)
n
1
2
n
X
− 12
b
θ (Pr ,K) − θ 0 = n
ψ(Zi ; Pr ) − Γ−1 Gn,K + Op (c∗n,K ),
i=1
1
1
→
−
where ψ(Z; Pr ) = Γ [ X{Y − µ(X; Pr )}] and c∗n,K = cn− + n− 2 + (n/N ) 2
K
= o(1). Further, for any fixed K ≥ 2, Gn,K = Op (cn− ), so that
−1
K
(3.11)
1
b(P ,K) − θ 0 = n− 12
n2 θ
r
n
X
ψ(Zi ; Pr ) + Op (cn− + c∗n,K ),
i=1
K
which converges in distribution to N(p+1) [0, Σ{µ(· ; Pr )}].
Remark 3.3. If the imputation is ‘sufficient’ so that µ(x; Pr ) = m(x),
b(P ,K) , for any K ≥ 2, enjoys the same set of optimality properties as
then θ
r
bnp (while requiring less stringent assumptions
those noted in Remark 3.1 for θ
about K(·) and h, if KS is used). If µ(x; Pr ) 6= m(x), then it is however unb(P ,K) is always more efficient than θ.
b This will be addressed
clear whether θ
r
in Section 3.3 where we develop the final EASE.
Remark 3.4. Apart from the fairly mild condition (i), Theorem 3.2 only
b r ) w.r.t. m(· ; Pr ) for establishing the
requires uniform consistency of m(·
b ;P
1
b(P ,K) for any K ≥ 2.
n 2 -consistency and asymptotic normality (CAN) of θ
r
The uniform consistency typically holds for a wide range of smoothing methods T under fairly general conditions. For T := KS in particular, we provide
explicit results in Section 4 under mild regularity conditions that allow the
use of any kernel order and the associated optimal bandwidth order. This is
a notable relaxation from the stringent requirements for Theorem 3.1 that
necessitate under-smoothing and the use of higher order kernels.
b(P ,1) has not yet been established.
Remark 3.5. The CAN property of θ
r
The term Gn,K in (3.10) behaves quite differently when K = 1, compared to
K ≥ 2 when it has a nice structure due to the inherent ‘cross-fitting’ involved,
and can be controlled easily, and quite generally, under mild conditions as
noted in Remark 3.4. For K = 1 however, Gn,K is simply a centered empirical
process devoid of any such structure and in general, controlling it requires
stronger conditions and the use of empirical process theory (see for instance
b(P ,1)
Van der Vaart (2000) for relevant results). We derive the properties of θ
r
for the case of T := KS in Theorem 4.2 using a different approach however,
specialized for KS estimators, in order to control Gn,1 .
13
SEMI-SUPERVISED LINEAR REGRESSION
3.3. Efficient and Adaptive Semi-Supervised Estimators (EASE). To ensure adaptivity even when µ(x; Pr ) 6= m(x), we now define the final EASE
b and θ
b(P ,K) . Specifically, for any fixed
as an optimal linear combination of θ
r
b(P ,K) (∆) = θ
b + ∆(θ
b(P ,K) − θ)
b is a CAN
(p + 1) × (p + 1) matrix ∆, θ
r
r
b and θ
b(P ,K) are, and an optimal ∆ can be
estimator of θ 0 whenever θ
r
selected easily to minimize the asymptotic variance of the combined estimator. For simplicity, we focus here on ∆ being a diagonal matrix with
bE
b
b
∆ = diag(δ1 , ..., δp+1 ). Then the EASE is defined as θ
(Pr ,K) ≡ θ (Pr ,K) (∆)
b being any consistent estimator (see Section 3.4 for details) of the
with ∆
minimizer ∆ = diag(δ 1 , ..., δ p+1 ), where ∀ 1 ≤ l ≤ (p + 1),
n
o
Cov ψ 0[l] (Z), ψ [l] (Z; Pr ) − ψ 0[l] (Z)
n
o
(3.12)
δ l = − lim
,
↓0
Var ψ [l] (Z; Pr ) − ψ 0[l] (Z) +
and for any vector a, a[l] denotes its lth component. Note that in (3.12),
the and the limit outside are included to formally account for the case:
ψ 0[l] (Z) = ψ [l] (Z, Pr ) a.s. [PZ ], when we define δ l = 0 for identifiability.
bE
b(P ,K) (∆) are asymptotiIt is straightforward to show that θ
and θ
(Pr ,K)
r
bE
cally equivalent, so that θ
(Pr ,K) is a RAL estimator of θ 0 satisfying:
n
E
X
1
d
− 12
b
−
θ
=
n
ψ(Zi ; Pr , ∆) + op (1) → N(p+1) [0, ΣPr (∆)],
n2 θ
0
(Pr ,K)
i=1
as n → ∞, where ψ(Z; Pr , ∆) = ψ 0 (Z) + ∆{ψ(Z; Pr ) − ψ 0 (Z)} and
ΣPr (∆) = Var{ψ(Z; Pr , ∆)}. Note that when either the linear model holds
or the SNP imputation is sufficient, then ψ(Z; Pr , ∆) = ψ eff (Z), so that
bE
θ
is asymptotically optimal in the sense of Remark 3.1. Further, when
(Pr ,K)
bE
neither cases hold, θ
(Pr ,K) is no longer optimal, but is still efficient and
b Lastly, if the imputation is certain to be sufficient
adaptive compared to θ.
bE
b(P ,K) .
(for example, if r = p and T := KS), we may simply define θ
=θ
(Pr ,K)
r
Remark 3.6. It can be shown that under MX , defined in Remark 3.1,
the class of all possible IFs achievable by RAL estimators of θ 0 is given by:
IF θ0 ,MX = {ψ g (Z) ≡ ψ eff (Z)+g(X) : E{g(X)} = 0, g[j] (·) ∈ L2 (PX ) ∀ j}.
b θ
b(P ,K) and θ
bE
The IFs achieved by θ,
are clearly members of this class.
r
(Pr ,K)
The SNP imputation, for various choices of the imputation function µ(· ; Pr ),
b(P ,K) , θ
bE
therefore equips us with a family of RAL estimator pairs {θ
}
r
(Pr ,K)
14
A. CHAKRABORTTY AND T. CAI
bE
for estimating θ 0 . The IF of θ
(Pr ,K) is further guaranteed to dominate that
b and when µ(· ; Pr ) = m(·), it also dominates all other IFs ∈ IFθ ,M .
of θ,
0
X
3.4. Inference for EASE and the SNP Imputation Based SS Estimators.
b(P ,K)
We now provide procedures for making inference about θ 0 based on θ
r
E
b
and θ
obtained using K ≥ 2. We also employ a ‘double’ CV to over(Pr ,K)
come bias in variance estimation due to over-fitting. A key step involved in
the variance estimation is to obtain reasonable estimates of {µ(Xi ; Pr )}ni=1 .
b (Pr ,K) in (3.4) was constructed via CV, the corresponding estiAlthough η
br,K ) in (3.6), of µ(x; Pr ) is likely to be over-fitted for Xi ∈ L.
mate, µ
b(x; P
To construct bias corrected estimates of µ(Xi ; Pr ), we first obtain K separate doubly cross-validated estimates of η Pr , {b
η k(Pr ,K) : k = 1, ..., K}, with
P
b k(Pr ,K) , for each k, being the solution in η to k0 6=k Sk0 (η) = 0, where
η
Sk0 (η) =
P
→
−
i∈Ik0 X i {Yi
−0
b r,k0 ) − →
−m
b k0 (Xi ; P
X i η}
∀ k 0 ∈ {1, . . . , K}.
For each k and k 0 6= k, Sk0 (η) is constructed such that {Zi : i ∈ Ik0 } used for
b r,k0 ) that is based on L−0 ⊥
0
b k(Pr ,K) is independent of m
obtaining η
b k0 (· ; P
k ⊥ Lk .
Then, for each Xi ∈ Lk and k ∈ {1, . . . , K}, we may estimate µ(Xi ; Pr ) as:
−0 k
b r,k ) + →
br,K ) = m
b (Pr ,K) .
µ
bk (Xi ; P
b k (Xi ; P
X iη
b k(Pr ,K) to reduce over-fitting bias
We exclude Sk (η) in the construction of η
br,K )} which we now use for estimating the IFs.
in the residuals {Yi − µ
bk (Xi ; P
For each Zi ∈ Lk and k ∈ {1, .., K}, we estimate ψ 0 (Zi ) and ψ(Zi ; Pr ),
b and θ
b(P ,K) , respectively as:
the corresponding IFs of θ
r
−
→
− b
−
b (Zi ) = Γ
b (Zi ; Pr ) = Γ
b −1 {→
b −1 [→
br,K )}],
ψ
X i (Yi − X 0i θ)}
and ψ
X i {Yi −b
µk (Xi ; P
0
k
b denotes any consistent estimator of Γ from L and/or U (for example,
where Γ
Pn+N →
− →
−0
− →
−0
−1
b = Γn ≡ n−1 Pn →
b
Γ
i=1 X i X i based on L, or Γ = ΓN ≡ N
j=n+1 X j X j
based on U). Then, Σ{µ(· ; Pr )} in (3.11) may be consistently estimated as:
b
Σ{µ(·
; Pr )} = n−1
K X
X
b (Zi ; Pr )ψ
b 0 (Zi ; Pr ).
ψ
k
k
k=1 i∈Ik
To estimate the combination matrix ∆ in (3.12) and the asymptotic variance, ΣPr (∆), of EASE consistently, let us define, ∀ 1 ≤ l ≤ (p + 1),
P
P
b
b
b
σ
bl,12 = − n−1 K
k=1
i∈Ik ψ 0[l] (Zi ){ψ k[l] (Zi ; Pr ) − ψ 0[l] (Zi )},
P
P
2
b
b
σ
bl,22 = n−1 K
k=1
i∈Ik {ψ k[l] (Zi ; Pr ) − ψ 0[l] (Zi )} ,
15
SEMI-SUPERVISED LINEAR REGRESSION
1
and δbl = σ
bl,12 /(b
σl,22 + n ) for some sequence n → 0 with n 2 n → ∞. Then,
b = diag(δb1 , ..., δbp+1 ) and
we estimate ∆ and ΣPr (∆) respectively as: ∆
b P (∆)
b = n−1
Σ
r
K X
X
b (Zi ; Pr , ∆)
b 0 (Zi ; Pr , ∆),
b ψ
b
ψ
k
k
k=1 i∈Ik
b (Z; Pr , ∆)
b (Z) + ∆{
b (Z; Pr ) − ψ
b (Z)} ∀ k ∈ {1, . . . , K}.
b =ψ
b ψ
where ψ
k
0
k
0
Normal confidence intervals (CIs) for the parameters of interest can also be
constructed accordingly based on these variance estimates.
4. Implementation Based on KS. We next detail the specific implementation of the SNP imputation based on KS estimators. With T := KS,
the target function for the smoothing is given by: m(x; Pr ) = mPr (P0r x) ≡
E(Y | XPr = P0r x). For simplicity, we assume that XPr is continuous with a
density fPr (·) and support XPr ≡ {P0r x : x ∈ X } ⊆ Rr . Let us now consider
the following class of local constant KS estimators for m(x; Pr ):
1 P
b0
b0
− r
i∈Ik− {Kh (Pr,k Xi , Pr,k x)}Yi
nK h
b
(4.1) m
b k (x; Pr,k ) =
∀ 1 ≤ k ≤ K,
1 P
b0
b0
i∈I − Kh (Pr,k Xi , Pr,k x)
n− hr
K
k
where Kh (·) and h are as in Section 3.1 with K(·) now being a suitable
kernel on Rr . In the light of Theorem 3.2, we focus primarily on establishing
b r) ≡ m
b r,1 ) in (4.1) with K = 1,
the uniform consistency of m(x;
b
P
b 1 (x; P
b r . For establishing the
accounting for the additional estimation error from P
desired result, we shall assume the following regularity conditions to hold:
Assumption 4.1. (i) K(·) is a symmetric kernel of order q ≥ 2 with
finite q th moments. (ii) K(·) is bounded, integrable and is either Lipschitz
continuous with a compact support or, has a bounded derivative ∇K(·)
which satisfies: k∇K(z)k ≤ Λkzk−ρ ∀ z ∈ Rr with kzk > L, where Λ > 0,
L > 0 and ρ > 1 are some fixed constants, and k.k denotes the standard
L2 vector norm. (iii) XPr ⊆ Rr is compact. E(|Y |s ) < ∞ for some s > 2.
E(|Y |s | XPr = z)fPr (z) and fPr (z) are bounded on XPr . (iv) fPr (z) is
bounded away from 0 on XPr . (v) mPr (z) and fPr (z) are both q times
continuously differentiable with bounded q th derivatives on some open set
X0,Pr ⊇ XPr . Additional Conditions (required only when Pr needs to be
estimated): (vi) K(·) has a bounded and integrable derivative ∇K(·). (vii)
∇K(·) satisfies: k∇K(z1 ) − ∇K(z2 )k ≤ kz1 − z2 k φ(z1 ) ∀ z1 , z2 ∈ Rr such
that kz1 − z2 k ≤ L∗ , for some fixed constant L∗ > 0, and some bounded
and integrable function φ : Rr → R+ . (viii) ∇K(·) is Lipschitz continuous
16
A. CHAKRABORTTY AND T. CAI
on Rr . (ix) E(X | XPr = z) and E(XY | XPr = z) are both continuously
differentiable with bounded first derivatives on X0,Pr ⊇ XPr .
Assumption 4.1, mostly adopted from Hansen (2008), imposes some mild
smoothness and moment conditions most of which are fairly standard, except
perhaps the conditions on K(·) in (vi)-(viii) all of which are however satisfied
by the Gaussian kernel among others. We now propose the following result.
b r − Pr ) = Op (αn ) for some αn = o(1) with
Theorem 4.1. Suppose (P
αn = 0 identically if Pr is known. Let q be the order of the kernel K(.) in
(4.1) for some integer q ≥ 2. Define:
1
1
log n 2
log n 2
2 −(r+2)
an,1 = αn
+ αn h
+ αn , an,2 =
+ hq
nhr+2
nhr
and assume that each of the terms involved in an,1 = o(1) and an,2 = o(1).
b r ), based on (4.1), satisfies:
Then, under Assumption 4.1, m(x;
b
P
(4.2)
b r ) − m(x; Pr )| = Op (an,1 + an,2 ).
supx∈X |m(x;
b
P
b r)
Remark 4.1. Theorem 4.1 establishes the L∞ error rate of m(x;
b
P
under mild regularity conditions and restrictions on h. Among its various
b r ) at the
implications, the rate also ensures uniform consistency of m(x;
b
P
−1/(2q+r)
optimal bandwidth order: hopt = O(n
) for any kernel order q ≥ 2
and any r ≤ p, as long as αn = o(n−(r+2)/(4q+2r) ) which always includes:
1
αn = O(n− 2 ) and αn = 0. These two cases are particularly relevant in prac1
tice as Pr being finite dimensional, n 2 -consistent estimators of Pr should
typically exist. For both cases, using hopt results in an,1 to be of lower order
(for q > 2) or the same order (for q = 2) compared to that of the main term
an,2 , so that the usual optimal rate prevails as the overall error rate.
b(P ,K) for K = 1. We now address the CAN property of
Properties of θ
r
b(P ,K) for K = 1 under the KS framework. Based on (3.10) and Remark
θ
r
3.5, the only step required for this is to effectively control the term Gn,K in
(3.10). The following result is in this regard. It involves Lemmas A.2-A.3 as
the main technical tools which may themselves be of independent interest.
b r)
Theorem 4.2. Let K = 1, T := KS, Gn,K be as in (3.10), and m(x;
b
P
be the KS estimator based on (4.1). Let αn , an,1 and an,2 be as in Theorem
b r − Pr ) = Op (αn ). Assume that a∗ and a∗ are o(1), where
4.1 with (P
n,1
n,2
a∗n,1 = αn +
αn
1
2
n h(r+1)
1
1
1
+n 2 αn2 h−2 +n 2 a2n,1 +n 2 an,1 an,2
and
1
a∗n,2 = n 2 a2n,2 .
SEMI-SUPERVISED LINEAR REGRESSION
17
Then, under Assumption 4.1, Gn,K = Op (a∗n,1 + a∗n,2 ) = op (1). Further, let
c∗n,K be as in Theorem 3.2 with cn = (an,1 + an,2 ). Then, using (3.10),
(4.3)
n
X
1
b(P ,K) − θ 0 = n− 12
n2 θ
ψ(Zi , Pr ) + Op (c∗n,K + dn ),
r
i=1
1
d
b(P ,K) − θ 0 ) →
N(p+1) [0, Σ{µ(· ; Pr )}].
where dn = a∗n,1 + a∗n,2 . Hence, n 2 (θ
r
Remark 4.2. Note that the term a∗n,2 always requires q > r/2 in order
to converge to 0, thus showing the contrasting behavior of the case K = 1
compared to K ≥ 2 where no such higher order kernel restriction is required.
1
Nevertheless, when αn = O(n− 2 ) or αn = 0, the optimal bandwidth order:
hopt = O(n−1/(2q+r) ) can indeed be still used as long as q > r/2 is satisfied.
Despite these facts and all the theoretical guarantee in Theorem 4.2, emb(P ,1) can be substantially
pirical evidence however seems to suggest that θ
r
biased in finite samples, in part due to over-fitting.
Remark 4.3. Technical benefits of refitting and CV: Suppose that
Pr = Ip , so that the SNP imputation with T := KS is indeed sufficient.
Further, assume that all of Theorems 3.1-4.2 hold, so that the estimators
bnp , θ
b(P ,1) , and θ
b(P ,K) (K ≥ 2) are comparable and all asymptotically
θ
r
r
optimal. However, their constructions are quite different which can signifibnp is based on KS only, and
cantly affect their finite sample performances. θ
requires stringent under-smoothing and a kernel of order q > p (Remark
b(P ,1) is based on KS and refitting (although the KS itself is certain to
3.2); θ
r
be sufficient), and requires no under-smoothing but needs a (weaker) kernel
b(P ,K) (K ≥ 2) additionorder condition (q > p/2) (Remark 4.2); while θ
r
ally involves CV, and requires no under-smoothing or higher order kernel
conditions (Remark 3.4). This highlights the critical role played by refitting
and CV, apart from their primary roles in the SNP imputation, in removing any under-smoothing and/or higher order kernel restrictions when T :=
KS, and this continues to hold for any other (r, Pr ) as well. In particular,
it shows, rather surprisingly, that refitting should be performed in order to
avoid under-smoothing even if the smoothing is known to be sufficient.
Remark 4.4. As mentioned in Section 3.2, T := KS along with possible
dimension reduction is just one reasonable choice of T for implementing the
SNP imputation, all technical requirements for which have been thoroughly
established in Section 4. In general, other smoothing methods, as long as
the requirements are satisfied, can also be equally used as choices of T . One
such choice could be kernel machine (KM) regression (with possibly no use
18
A. CHAKRABORTTY AND T. CAI
of dimension reduction, as KM uses penalization to effectively regularize the
target even with Pr = Ip ). We leave its implementation details to the reader
as they are readily available in a multitude of references, and also skip any
theoretical treatment, considering the primary goal and scope of this paper.
However, detailed numerical results are presented in Section 6 for this choice
of T as well to illustrate the wider applicability of our proposed methods.
5. Choices of Pr : Dimension Reduction Techniques. We next discuss choosing and estimating the matrix Pr (r < p) to be used for dimension
reduction, if required, in the SNP imputation, and which can play an important role in the sufficiency of the imputation. Simple choices of Pr include r
leading principal component directions of X or any r canonical directions of
X. Note that under the SS setting, Pr is effectively known if it only involves
the distribution of X, as is true for these choices. We now focus primarily on
the case where Pr also depends on the distribution of Y and hence, is unknown. Such a choice of Pr is often desirable to ensure that the imputation is
as ‘sufficient’ as possible for predicting Y . Several reasonable choices of such
Pr and their estimation are possible based on sufficient dimension reduction
(s.d.r.) methods like Sliced Inverse Regression (SIR) (Li, 1991), Principal
Hessian Directions (PHD) (Li, 1992; Cook, 1998), Sliced Average Variance
Estimation (SAVE) (Cook and Weisberg, 1991; Cook and Lee, 1999) etc.
In particular, we focus here on SIR where the choice of Pr is given by: P0r
1
= Σ− 2 Pr , with Pr being the r leading eigenvectors of M = Var{E(X | Y )},
1
where X = Σ− 2 (X − µ), with µ = E(X), denotes the standardized version
of X. It is well known (Li, 1991) that these directions lead to an optimal
(in some appropriate sense) r-dimensional linear transformation of X that
can be predicted by Y . Apart from these general optimality, they also have
deeper implications in the context of s.d.r. We refer the reader to Li (1991)
and other relevant references in the s.d.r. literature for further details.
For estimating P0r , we consider the SIR algorithm of Li (1991) and further
− b
K
propose a SS modification to it. With K and {L−
k , Ik , Pr,k }k=1 as before,
−
b k ) denote the estimates of (µ, Σ) based on L and define X(k) =
let (b
µk , Σ
k
− 12
b (X − µ
b k ). Then, the original SIR algorithm estimates P0r based on L−
Σ
k
k
as follows: (i) Divide the range of {Yi }i∈I − into H slices {I1 , .., IH }, where
k
H may depend on n−
bh,k denote the proportion of
K . For 1 ≤ h ≤ H, let p
b
{Yi }i∈I − in slice Ih ; (ii) For each Ih , let Mh,k denote the sample average of
k
PH
(k)
b
b h,k M
b0
the set: {Xi ∈ L−
bh,k M
h=1 p
h,k
k : Yi ∈ Ih }; (iii) Estimate M as: Mk =
1
b0 = Σ
b−2 P
b r,k , where P
b r,k denotes the r leading eigenvectors
and P0r as: P
k
r,k
19
SEMI-SUPERVISED LINEAR REGRESSION
b k . However, the SIR algorithm often tends to give unstable estimates
of M
of P0r , especially for the directions corresponding to the smaller eigenvalues
of M. To improve the efficiency in estimating P0r , we now propose a semisupervised SIR (SS-SIR) algorithm as follows.
∗
− b
K
b ) denote the esSS-SIR Algorithm. Given {L−
µ∗k , Σ
k
k , Ik , Pr,k }k=1 , let (b
1
∗−
−
2
(k∗)
b
b k ). Then
timates of (µ, Σ) based on Lk ∪ U and define X
= Σk (X − µ
the SS-SIR proceeds as follows. Step (i) stays the same as in SIR. In step
∗ =Y
(ii), for each k, and each j ∈ {n + 1, ..., n + N }, we impute Yj as Yj,k
bij,k ,
(k∗)
(k∗)
b ∗ be the samwhere bij,k = argmin − kX
− X k2 . For each Ih , let M
i
(k∗)
{Xi
i∈Ik
ple average of the set:
in step (iii), we estimate M
∗− 12
j
h,k
(k∗)
−
∗
∈ Lk : Yi ∈ Ih } ∪ {Xj ∈ U : Yj,k
P
0
H
b∗
b∗ M
b∗ =
bh,k M
as: M
h=1 p
h,k h,k and
k
∈ Ih }. Then
then, P0r as:
b ∗.
b
b 0∗ = Σ
b ∗ denotes the r leading eigenvectors of M
b ∗ , where P
P
P
k
r,k
k
r,k
r,k
The SS-SIR algorithm aims to improve the estimation of P0r by making
use of U in step (ii) through a nearest neighbour approximation for the un−
observed Y in U using L−
k . With nK large enough and m(·) smooth enough,
the imputed and the true underlying Y should belong to the same slice with
a high probability. Thus, the set of X’s belonging to a particular slice is
now ‘enriched’ and consequently, improved estimation of M and P0r is expected. The proposed method based on a nearest neighbor approximation is
also highly scalable and while other smoothing based approximations may
be used, they can be computationally intensive. The SS-SIR algorithm is
1
fairly robust to the choice of H, and H = O(n 2 log n) seems to give fairly
satisfactory performance. The slices may be chosen to have equal width or
1
equal number of observations. For SIR, n 2 -consistency of the estimates are
well established (Li, 1991; Duan and Li, 1991; Zhu and Ng, 1995) for various
formulations under fairly general settings (without any model based assumptions). The theoretical properties of SS-SIR, although not derived here, are
expected to follow similarly. Our simulation results (not shown here) further
suggest that SS-SIR significantly outperforms SIR, leading to substantially
improved estimation of θ 0 from the proposed methods.
6. Numerical Studies.
6.1. Simulation Studies. We conducted extensive simulation studies to
examine the finite sample performance of our proposed point and interval estimation procedures as well as to compare with existing methods. Throughout we let n = 500, N = 10000, and considered p = 2, 10 and 20. For our
CV based methods, we let K = 5. The true values of the target parameter
20
A. CHAKRABORTTY AND T. CAI
θ 0 were estimated via monte carlo with a large sample size of 50, 000. For
each configuration, the results were summarized based on 500 replications.
Results for p = 2 are summarized in the Supplementary Material, and the
discussions below focus primarily on p = 10 and 20.
We generated X ∼ Np [0, Ip ] and restricted X to [−5, 5]p to ensure its
boundedness. Given X = x, we generated Y ∼ N1 [m(x), 1], where we considered four different choices of m(x) :
(i) Linear : m(x) = x0 bp ;
(ii) Non-linear one component (NL1C): m(x) = (x0 bp ) + (x0 bp )2 ;
(iii) Non-linear two component (NL2C): m(x) = (x0 bp )(1 + x0 δ p ); and
(iv) Non-linear three component (NL3C): m(x) = (x0 bp )(1 + x0 δ p ) + (x0 ω p )2 ;
(1)
where, for each setting, we considered bp = bp
(2)
≡ (10p/2 , 00p/2 )0 and bp =
bp ≡ 1p , and set δ p = (00p/2 , 10p/2 )0 and ω p = (1, 0, 1, 0, . . . , 1, 0)0p×1 , where
for any a, 1a = (1, . . . , 1)0a×1 and 0a = (0, . . . , 0)0a×1 . Through appropriate choices of bp , δ p and ω p , as applicable, these models can incorporate
commonly encountered linear, quadratic and interaction effects.
For each setting, we used two choices of the smoothing method: (a)
T := KS2,P2 denoting KS with 2-dimensional smoothing over P0r X ≡ P02 X,
where P2 was estimated via SIR with H = 100 slices of equal width, followb r,k )}K were obtained via KS using a Gaussian kernel;
ing which {m
b k (x; P
k=1
(b) T := KM where we let Pr = Ip and then estimated {m
b k (x; Ip )}K
k=1 using
kernel machine (KM) regression based on a radial basis function (RBF) kernel. Throughout, h for KS, and all tuning parameters for KM were selected
via least squares CV. For (a), with X ∼ Np [0, Ip ], results from Li (1991) imply that the SNP imputation with r = 2 is sufficient for models (i)-(iii), and
insufficient for model (iv). For comparison, we also implemented two other
SS estimators: the density ratio based “DRESS” estimator of Kawakita and
Kanamori (2013) and the estimator of Sokolovska, Cappé and Yvon (2008)
called “MSSL” by Kawakita and Kanamori (2013). The density ratio estimation for the DRESS estimator was implemented using either (i) linear bases
3
{1, (X[j] )pj=1 } (DRESS1 ); or (ii) cubic bases {1, (Xd[j] )p,
j=1,d=1 } (DRESS3 ).
First, we compare the various estimators with respect to their efficiencies
based on empirical mean squared error. In Table 1, we present the efficiencies
of the proposed SNP and EASE estimators as well as other SS estimators
relative to the OLS. As expected, under model mis-specification, our estimators are substantially more efficient than the OLS with the relative efficiency
(RE) as high as near 5 fold when p = 10 and 3 fold when p = 20, for the
non-linear models. The efficiency gain is generally lower for p = 20 than for
p = 10, likely a consequence of overfitting of the non-parametric estimators
SEMI-SUPERVISED LINEAR REGRESSION
21
involved in the SNP imputation for larger p. Comparing EASE to SNP, the
EASE generally perform better for both linear and non-linear settings, as
expected. Comparing the two smoothers, it appears that T := KM generally
attains higher efficiency compared to that of T := KS2,P2 . This is in part
due to the high variability in the SIR direction estimation which impacts
the performance of the resulting SS estimator in finite samples. Interestingly,
none of the existing SS estimators perform well with REs ranging only from
about 0.9 to 1.1 across all settings.
(a) p = 10
Setting
(I)
(II)
Models
Linear
NL1C
NL2C
NL3C
Linear
NL1C
NL2C
NL3C
OLS
(Ref.)
1
1
1
1
1
1
1
1
SNP EASE
(T := KS2,P2 )
0.895
0.983
4.481
4.424
2.683
2.700
2.772
2.795
0.841
0.989
4.511
4.585
3.596
3.634
3.280
3.301
SNP EASE
(T := KM)
0.772
0.985
4.501
5.543
4.268
5.055
4.481
5.560
0.657
0.993
4.416
5.471
4.405
5.497
4.636
5.566
Other SS Estimators
DRESS1 DRESS3 MSSL
0.982
0.927
0.982
1.136
1.110
1.135
1.120
1.016
1.119
1.102
1.025
1.103
0.981
0.924
0.981
1.132
1.030
1.130
1.127
1.042
1.128
1.110
1.079
1.109
(b) p = 20
Setting
(I)
(II)
Models
Linear
NL1C
NL2C
NL3C
Linear
NL1C
NL2C
NL3C
OLS
(Ref.)
1
1
1
1
1
1
1
1
SNP EASE
(T := KS2,P2 )
0.673
0.986
2.256
2.288
1.414
1.388
1.539
1.531
0.519
0.991
2.290
2.346
1.899
1.917
1.937
1.949
SNP EASE
(T := KM)
0.740
0.981
2.680
3.630
2.661
3.544
2.605
3.510
0.609
0.989
2.669
3.660
2.766
3.963
2.682
3.702
Other SS Estimators
DRESS1 DRESS3 MSSL
0.956
0.866
0.956
1.035
0.920
1.035
1.032
0.922
1.033
1.049
0.931
1.051
0.958
0.872
0.958
1.032
0.908
1.031
1.036
0.917
1.036
1.046
0.958
1.046
b(P ,K) (SNP) and θ
bE
Table 1: Efficiencies of θ
(Pr ,K) (EASE) using T := KS2,P2
r
b (OLS)
or T := KM, as well as DRESS1 , DRESS3 and MSSL, relative to θ
with respect to the empirical mean squared error (MSE) under models (i),
(1)
(2)
(ii), (iii) and (iv) each with: (I) bp = bp or, (II) bp = bp .
We next examine the performance of the proposed inference procedures. In
Table 2(a) and (b), we present the bias, empirical standard error (ESE), the
average of the estimated standard error (ASE) and the coverage probability
(CovP) of the 95% CIs for each component of θ 0 when p = 10 under the
linear and NL2C models. In general, the EASE with both the KS and the KM
smoothers have negligible biases although the KM based estimator appears
to have slightly lower biases. The ASEs are close to the ESEs and the CovPs
are close to the nominal level, suggesting that the variance estimators work
22
A. CHAKRABORTTY AND T. CAI
well in practice with K = 5.
(a) OLS and EASE for the linear model.
Parameter
α0 = 0
β01 = 1
β02 = 1
β03 = 1
β04 = 1
β05 = 1
β06 = 0
β07 = 0
β08 = 0
β09 = 0
β010 = 0
OLS
Bias
-0.001
0.002
-0.001
-0.001
-0.002
-0.004
-0.000
0.003
-0.001
-0.002
0.003
b
(θ)
ESE
0.043
0.044
0.044
0.046
0.045
0.048
0.045
0.046
0.045
0.047
0.045
EASE
Bias
-0.001
-0.003
-0.005
-0.005
-0.006
-0.008
-0.001
0.003
-0.001
-0.002
0.003
E
b(P ,K) ; T := KS2,P )
(θ
2
r
ESE
ASE CovP
0.043 0.044
0.95
0.045 0.044
0.94
0.044 0.044
0.94
0.046 0.044
0.95
0.045 0.044
0.94
0.049 0.044
0.92
0.045 0.044
0.94
0.046 0.044
0.93
0.045 0.044
0.95
0.048 0.044
0.94
0.045 0.044
0.94
bE
EASE (θ
(Pr ,K) ; T := KM)
Bias
ESE
ASE CovP
0.000 0.043 0.044
0.96
0.004 0.047 0.044
0.93
0.000 0.045 0.044
0.95
-0.004 0.045 0.044
0.94
0.001 0.047 0.044
0.94
-0.001 0.046 0.044
0.95
0.001 0.045 0.044
0.95
0.001 0.043 0.044
0.96
-0.000 0.048 0.044
0.94
0.000 0.045 0.045
0.95
-0.002 0.045 0.045
0.94
(b) OLS and EASE for the NL2C model.
Parameter
α0 = 0
β01 = 1
β02 = 1
β03 = 1
β04 = 1
β05 = 1
β06 = 0
β07 = 0
β08 = 0
β09 = 0
β010 = 0
OLS
Bias
-0.015
0.000
-0.004
-0.015
-0.001
0.013
-0.010
0.006
-0.008
0.002
-0.008
b
(θ)
ESE
0.239
0.260
0.269
0.249
0.267
0.260
0.281
0.277
0.277
0.279
0.272
EASE
Bias
-0.016
0.015
0.017
0.018
0.016
0.019
0.008
0.002
-0.004
0.003
0.002
E
b(P ,K) ; T := KS2,P )
(θ
2
r
ESE
ASE CovP
0.146 0.136
0.93
0.159 0.160
0.96
0.173 0.158
0.93
0.156 0.158
0.95
0.164 0.159
0.94
0.164 0.158
0.94
0.164 0.155
0.94
0.166 0.155
0.93
0.167 0.156
0.94
0.160 0.157
0.95
0.160 0.155
0.95
bE
EASE (θ
(Pr ,K) ; T := KM)
Bias
ESE
ASE CovP
0.013 0.105 0.096
0.93
-0.004 0.124 0.112
0.93
0.010 0.127 0.113
0.93
-0.000 0.118 0.113
0.95
0.007 0.124 0.113
0.93
0.002 0.120 0.113
0.94
0.005 0.119 0.112
0.94
0.011 0.116 0.111
0.95
-0.001 0.120 0.112
0.95
0.007 0.118 0.113
0.95
0.004 0.130 0.111
0.91
(c) All other SS estimators for the models in (a) and (b) above.
DRESS1
Bias
ESE
-0.001 0.043
-0.002 0.044
0.000 0.045
0.006 0.045
0.003 0.045
-0.004 0.047
-0.001 0.045
-0.000 0.048
-0.004 0.043
-0.001 0.048
-0.003 0.047
Linear Model
DRESS3
Bias
ESE
-0.001 0.044
-0.001 0.046
0.001 0.046
0.006 0.047
0.003 0.046
-0.004 0.049
-0.001 0.046
-0.001 0.050
-0.003 0.044
-0.001 0.049
-0.003 0.049
MSSL
Bias
ESE
-0.001 0.043
-0.002 0.044
0.000 0.045
0.006 0.045
0.003 0.045
-0.004 0.047
-0.001 0.045
-0.000 0.048
-0.004 0.043
-0.001 0.048
-0.003 0.047
DRESS1
Bias
ESE
-0.004 0.223
-0.014 0.266
0.005 0.257
-0.013 0.256
-0.005 0.262
0.002 0.250
-0.017 0.239
-0.022 0.260
-0.011 0.241
-0.020 0.256
-0.020 0.252
NL2C Model
DRESS3
Bias
ESE
-0.003 0.226
-0.009 0.279
0.006 0.266
-0.019 0.281
-0.007 0.274
-0.007 0.266
-0.009 0.247
-0.019 0.270
-0.013 0.261
-0.019 0.259
-0.022 0.269
MSSL
Bias
ESE
-0.004 0.223
-0.014 0.266
0.006 0.257
-0.011 0.256
-0.005 0.262
0.002 0.252
-0.017 0.239
-0.022 0.260
-0.010 0.241
-0.020 0.256
-0.020 0.252
Table 2: Coordinate-wise bias, ESE, ASE and CovP of EASE, obtained using
T := KS2,P2 or T := KM, for estimating θ 0 under the linear and NL2C
(1)
models with p = 10 and bp = bp . Shown also are the corresponding bias
and ESE of the OLS, as well as the DRESS1 , DRESS3 and MSSL estimators.
SEMI-SUPERVISED LINEAR REGRESSION
23
As shown in Table 2(c), the other SS estimators tend to have slightly
larger biases and substantially larger standard errors (SEs) compared to our
estimators under the NL2C model.
6.2. Application to EMR Data. We applied our proposed SS procedures
to an EMR study of rheumatoid arthritis (RA), a systemic auto-immune
disease (AD), conducted at the Partners HealthCare (Liao et al., 2010).
The study cohort consists of 3854 RA patients with blood samples stored.
The outcome of interest is the (logarithm of) anti-CCP (antibodies to cyclic
citrullinated polypeptide), a biomarker that is often used to determine subtypes of RA. Due to cost constraints, anti-CCP was measured only for a
random subset of n = 355 patients, thereby leading to a SS set-up. To investigate the validity of the MCAR assumption, we report in Table II in
the Supplementary Material summary measures of the distributions in the
labeled and unlabeled data for each of the predictors, as well as p-values
from various tests for assessing equality of those distributions. The results
suggest that the MCAR assumption is appropriate in this study.
We relate the log anti-CCP level to a set of p = 24 clinical variables X
related to ADs, including age, gender, race; total counts of codified and/or
narrative mentions extracted from physicians’ notes via natural language
processing (NLP) for various RA related conditions including RA, Lupus,
Polymyalgiarheumatica (PmR), Spondyloarthritis (SpA), as well as various
RA medications; indicators of seropositivity and radiological evidence of
erosion; mentions of rheumatoid factor (RF), as well as anti-CCP positivity
from prior medical history. Since the tests for RF and anti-CCP were not
always ordered, missing indicators for these variables were also included. All
count variables were transformed as: x → log(1 + x) to increase stability of
the model fitting. All predictors were normalized to have unit variance.
We obtained the OLS, the EASE using both T := KS2,P2 and T := KM
in the smoothing step, as well as the DRESS1 estimator for comparison.
For EASE, we again used K = 5 and for the KS2,P2 smoother, P2 was
obtained using SIR with H = 80 slices of equal width. In Table 3, we present
the coordinate-wise estimates of the regression parameters along with their
estimated SEs and the corresponding p-values based on these estimates.
Overall, the point estimators from all methods are quite close to each other.
Our proposed EASE, with both KS and KM smoothers, is substantially
more efficient than the OLS across all coordinates with efficiency ranging
from about 1.4 to 2.4. The DRESS1 estimator improved estimation for a
few coordinates but the efficiency remains comparable to the OLS for most
coordinates. This again suggests the advantage of our proposed estimators
24
A. CHAKRABORTTY AND T. CAI
compared to both OLS and other SS estimators.
Predictors
Age
Gender
Race
Lupus
PmR
RA
SpA
Other ADs
Erosion
Seropositivity
Anti-CCPprior
Anti-CCPmiss
RF
RFmiss
Azathioprine
Enbrel
Gold salts
Humira
Infliximab
Leflunomide
Methotrexate
Plaquenil
Sulfasalazine
Other meds.
Est
.105
-.032
-.041
.038
-.075
.015
-.137
-.022
.076
.056
.572
.527
.128
.085
-.080
.138
.138
-.051
.003
-.027
-.021
-.043
-.114
-.042
b
OLS (θ)
SE
.076
.059
.065
.066
.044
.089
.102
.078
.070
.062
.136
.123
.081
.085
.071
.070
.050
.068
.069
.069
.073
.069
.074
.074
Pval
.168
.589
.534
.563
.088
.862
.177
.775
.278
.370
.000
.000
.113
.316
.263
.048
.006
.453
.968
.697
.775
.540
.125
.570
EASE (KS2,P2 )
Est
SE
Pval
.106 .064 .099
-.028 .050 .570
-.042 .055 .452
.048 .052 .359
-.076 .031 .013
.012 .076 .879
-.133 .072 .063
-.024 .058 .679
.078 .059 .184
.054 .053 .310
.557 .110 .000
.520 .097 .000
.125 .066 .059
.084 .070 .233
-.074 .056 .185
.133 .058 .021
.136 .043 .002
-.049 .057 .391
.008 .057 .887
-.023 .058 .693
-.024 .061 .699
-.038 .057 .503
-.116 .063 .064
-.052 .060 .385
RE
1.40
1.41
1.40
1.59
2.07
1.37
2.02
1.79
1.44
1.37
1.54
1.61
1.49
1.46
1.62
1.48
1.37
1.43
1.50
1.40
1.42
1.47
1.39
1.52
Est
.094
-.027
-.044
.021
-.074
.008
-.128
-.018
.085
.041
.567
.508
.149
.137
-.075
.136
.147
-.057
.000
-.031
-.025
-.044
-.105
-.052
DRESS1
SE
.073
.058
.067
.063
.031
.080
.075
.067
.069
.061
.123
.115
.079
.080
.062
.073
.050
.067
.067
.071
.073
.070
.072
.071
Pval
.199
.638
.511
.731
.016
.923
.089
.792
.221
.496
.000
.000
.059
.088
.225
.064
.003
.389
.994
.660
.728
.532
.145
.466
RE
1.09
1.04
.95
1.11
2.10
1.23
1.82
1.35
1.03
1.05
1.23
1.15
1.05
1.12
1.33
.91
1.01
1.03
1.07
.93
1.01
.98
1.06
1.10
EASE (KM)
Est
SE
Pval
.104 .064 .103
-.031 .049 .524
-.040 .055 .462
.037 .051 .464
-.075 .030 .014
.016 .075 .832
-.136 .066 .038
-.022 .056 .692
.076 .058 .189
.055 .052 .296
.568 .107 .000
.523 .096 .000
.127 .066 .054
.084 .070 .231
-.079 .053 .132
.137 .057 .017
.137 .042 .001
-.051 .056 .360
.003 .055 .959
-.026 .057 .644
-.022 .060 .720
-.042 .057 .460
-.113 .061 .065
-.042 .059 .473
RE
1.42
1.44
1.41
1.70
2.04
1.30
2.37
1.93
1.47
1.41
1.60
1.64
1.49
1.48
1.83
1.49
1.40
1.49
1.57
1.45
1.46
1.48
1.45
1.59
Table 3: Estimates (Est) of the regression coefficients based on OLS, EASE
obtained using either T := KS2,P2 or T := KM, as well as DRESS1 , along
with their estimated standard errors (SE) and the corresponding p-values
(Pval.) for testing the null effect of each predictor. Shown also are the relative
efficiencies (RE) of all the estimators compared to the OLS.
We also estimated the prediction errors (PEs) for each of the fitted linear
models based on the aforementioned estimation methods via CV. To remove
potential randomness in the CV partitions, we averaged over 10 replications
of leave-5-out CV estimates. The PE was about 1.28 for EASE with both
smoothers, 1.29 for OLS and 1.30 for DRESS1 . For prediction purposes, we
may also directly employ non-parametric estimates of the conditional mean
rather than the fitted linear models. The PE in fact is slightly larger when we
b r ) or µ
br,K ). The PE was 1.34 for KS and 1.33 for KM based
use m(x;
b
P
b(x; P
b
br,K ). This
on m(x;
b
Pr ), and 1.30 for KS and 1.28 for KM based on µ
b(x; P
confirms that while the linear model may be mis-specified, it may often be
preferable to non-parametric models in practice as it may achieve simplicity
without substantial loss in prediction performance.
7. Discussion. We have developed in this paper an efficient and adaptive estimation strategy for the SS linear regression problem. The adaptive
property possessed by the proposed EASE is crucial for advocating ‘safe’ use
of the unlabeled data and is often unaddressed in the existing literature. In
SEMI-SUPERVISED LINEAR REGRESSION
25
general, the magnitude of the efficiency gain with EASE depends on the inherent degree of non-linearity in E(Y |X) and the extent of sufficiency of the
underlying SNP imputation. In particular, if the imputation is sufficient or
bE
the working linear model is correct, θ
(Pr ,K) is further optimal among a wide
class of estimators. We obtained theoretical results along with IF expansions
b(P ,K) and θ
bE
for θ
(Pr ,K) substantiating all our claims and also validated them
r
based on numerical studies. The double CV method further facilitates accurate inference, overcoming potential over-fitting issues in finite samples due
to smoothing. An R code for implementing EASE is available upon request.
The proposed SNP imputation, the key component of EASE, apart from
being flexible and scalable, enjoys several useful properties. The refitting
b(P ,K) , and for
step and CV play a crucial role in reducing the bias of θ
r
T := KS in particular, eradicate any under-smoothing or higher order kernel requirements: two undesirable, yet often inevitable, conditions required
1
for n 2 -consistency of two-step estimators based on a first step of smoothing.
b(P ,1) compared
Theorem 4.2, apart from showing the distinct behaviour of θ
r
b(P ,K) for K ≥ 2, also highlights the key role of CV in completely removto θ
r
ing kernel order restrictions, apart from addressing over-fitting issues. The
error rates in the results of Theorems 4.1-4.2 are quite sharp and account
b r . The regularity conditions required are also
for any estimation error from P
fairly mild and standard in the literature. The continuity assumption on X
in Sections 3.1 and 4 is mostly for the convenience of proofs, and the results
continue to hold for more general X. Lastly, while we have focussed here on
linear regression for simplicity, our methods can indeed be easily adapted to
other regression problems such as logistic regression for binary outcomes.
When the goal is solely that of prediction, one obviously does not have to
employ linear regression models, and models that incorporate non-linear efb r ) or µ
br,K ),
fects can be helpful. For such settings, the estimators m(x;
b
P
b(x; P
obtained as by-products of our SNP imputation, can themselves serve as potentially useful non-linear predictors. These SNP estimators may substantially outperform naive non-parametric estimators such as a p-dimensional
KS estimator, as demonstrated in Table III of the Supplementary Material
for the models considered in our simulation studies. In practice, when the
covariates are substantially correlated and the dimension of p is not small
as in the EMR example, it is unclear whether non-linear models necessarily
provide better prediction performance than the linear models. Under such
settings, the linear model also has a clear advantage due to its simplicity. Furthermore, while prediction is a vitally important goal of predictive
modeling, association analysis under interpretable models is key to clinical
studies for discovery research and efficient estimation of the corresponding
26
A. CHAKRABORTTY AND T. CAI
model parameters remains an important task.
b(P ,K) . While (3.11)
We end with a comment on the choice of K ≥ 2 in θ
r
holds for any K ≥ 2, the error term in (3.11) depends on K through cn−
1
K
cn− are
and more precisely, through e
cn− = K 2 cn− . Since K is fixed, cn− and e
K
K
K
K
asymptotically equivalent. But for a given n, cn− is expected to decrease with
K
K, while e
cn− is likely to increase. It is however desirable that both are small
K
since cn− inherently controls the efficiency of the SNP imputation, while e
cn−
K
K
b
directly controls the bias of θ (P ,K) . Hence, a reasonable choice of K ≥ 2 may
r
c2n− ) for some λ ≥ 0. Since the (first order)
be based on minimizing: (c2n− +λe
K
K
b(P ,K) is independent of K, this is equivalent to a
asymptotic variance of θ
r
b(P ,K) with λ denoting
penalized minimization of the asymptotic MSE of θ
r
the weightage of the (lower order) bias relative to the (first order) variance.
In general, the optimal K should be inversely related to λ. Conversely, choice
of any K may be viewed to have an associated regularization effect (through
λ) resulting in a ‘variance-bias trade-off’ with smaller K leading to lower bias
at the cost of some efficiency, and higher K leading to improved efficiency
in lieu of some bias. In practice, we find that K = 5 works well, and K = 10
tends to give slightly smaller MSE at the cost of increased bias.
APPENDIX A
A.1. Preliminaries. The following Lemmas A.1-A.3 would be useful
in the proofs of the main theorems. The proofs of these lemmas, as well as
Theorems 3.1, 4.1 and 4.2, can be found in the Supplementary Material.
Lemma A.1. Let Z ∈ Rl be any random vector and g(Z) ∈ Rd be any
measurable function of Z, where l and d are fixed. Let Sn = {Zi }ni=1 ⊥
⊥ Sm =
{Zj }m
be
two
random
samples
of
n
and
m
i.i.d.
observations
of
Z
respecj=1
bn (·) be any estimator of g(·) based on Sn such that the random
tively. Let g
sequence: Tbn ≡ supz∈χ kb
gn (z)k is Op (1), where χ ⊆ Rl denotes the support
b n,m denote the (double) random sequence: m−1 P
bn (Zj ),
of Z. Let G
Zj ∈Sm g
b n,m ) = EZ {b
and let Gn denote the random sequence: ESm (G
gn (Z)}, where
EZ (·) denotes expectation w.r.t. Z ∈ Sm ⊥
⊥ Sn , and all expectations involved
are assumed to be finite almost surely (a.s.) [Sn ] ∀ n.
1
Then: (a) Gn,m − Gn = Op (m− 2 ), and (b) as long as g(.) has finite 2nd
P
1
moments, m−1 Zj ∈Sm g(Zj ) − EZ {g(Z)} = Op (m− 2 ).
Controlling Empirical Processes Indexed by KS Estimators. The next two
lemmas would be useful in the proof of Theorem 4.2. They may also be
27
SEMI-SUPERVISED LINEAR REGRESSION
of more general use in other applications that involve controlling empirical
processes indexed by kernel estimators - both linear and ratio-type estimators, where the smoothing is further allowed to be performed over a possibly
lower dimensional and estimated transformation of the original covariate X.
These allowances make the technical analyses of such processes considerably
more involved and nuanced. The results of these lemmas and the techniques
used in their proofs may therefore be of independent general interest.
Suppose Assumption 2.1 (a) holds, and consider the KS framework introP
(%)
duced in Section 4. Let ϕ
ePr (w) = (nhr )−1 ni=1 Kh (w, P0r Xi )Yi% , for % =
(0)
(1)
0, 1. Let feP (·) = ϕ
e (·), e
lP (·) = ϕ
e (·) and m
e P (·) = e
lP (·)/feP (·). Next,
r
let
(0)
ϕPr (·)
Pr
= fPr (·) and
r
(1)
ϕPr (·)
Pr
r
r
r
= lPr (·), where lPr (·) = mPr (·)fPr (·). For
(%)
(%)
0
each % ∈ {0, 1}, let
e(%) (x; Pr ) = ϕ
ePr (P0r x).
r ) = ϕPr (Pr x) and ϕ
Further, let fe(·) = ϕ
e(0) (·), e
l(·) = ϕ
e(1) (·) and m(·)
e =e
l(·)/fe(·).
Lastly, let Pn denote the empirical probability measure on Rp based on
{Xi }ni=1 , and for any measurable (and possibly vector-valued) function γ(·)
1 R
of X, where γ(·) can be random itself, let G∗n (γ) = n 2 γ(x)(Pn − PX )(dx),
1
the (centered) n 2 -scaled empirical process indexed by γ(·). Lemmas A.2-A.3
together, among other more general implications, establish explicit rates of
g (·)−g(·)}, for any linear or ratio-type kernel
convergence of the quantity G∗n {b
estimator gb(·) of the type discussed above and its corresponding target g(·).
ϕ(%) (x; P
Lemma A.2. Consider the set-up introduced above. For any fixed integer
d ≥ 1, let λ(·) be any Rd -valued measurable function of X that is bounded
1
1
1
(1)
a.s. [PX ]. Define: bn = n− 2 h−r + hq and an,2 = (log n) 2 (nhr )− 2 + hq .
1
(1)
Assume bn = o(1) for (A.1) and n 2 a2n,2 = o(1) for (A.2) below. Then,
under Assumption 4.1 (i)-(v), and ∀ % ∈ {0, 1},
(A.1)
G∗n [λ(·){ϕ
e(%) (· ; Pr ) − ϕ(%) (· ; Pr )}] = Op (b(1)
n ) = op (1), and
(A.2)
G∗n [λ(·){m(·
e ; Pr ) − m(· ; Pr )}] = Op (n 2 a2n,2 ) = op (1).
1
b r ) = (nhr )−1 Pn Kh (P
b 0 x, P
b 0 Xi )Y % ∀ % ∈ {0, 1}, where P
br
Let ϕ
b(%) (x; P
r
r
i=1
i
is as in Section 3.2 and all other notations are the same as in the set-up of
b r) = ϕ
b r ) and b
b r) = ϕ
b r ). Then:
Lemma A.2. Let fb(x; P
b(0) (x; P
l(x; P
b(1) (x; P
b r ) be as
Lemma A.3. Consider the set-up of Lemma A.2. Let ϕ
b(%) (x; P
b r − Pr ) = Op (αn ) for
above, and let λ(·) be as in Lemma A.2. Suppose (P
1
(2)
(2)
some αn = o(1). Assume bn = o(1), where bn = αn + n− 2 αn h−(r+1) +
1
n 2 αn2 (h−2 + n−1 h−(r+2) ). Then, under Assumption 4.1,
b r) − ϕ
(A.3) G∗n [λ(·){ϕ
b(%) (· ; P
e(%) (· ; Pr )}] = Op (b(2)
n ) = op (1) ∀ % ∈ {0, 1}.
28
A. CHAKRABORTTY AND T. CAI
A.2. Proof of Theorem 3.2. Let Γn =
n
T(1)
n =
1
n
Pn →
− →
−0
i=1 X i X i , and
K
−
− b
1 X→
1XX→
(2)
b r,k ).
X i {Yi − µ(Xi ; Pr )} , Tn,K =
X i ∆k (Xi ; Pr , P
n
n
i=1
k=1 i∈Ik
Then, using (3.4)-(3.8), it is straightforward to see that:
(A.4)
(A.5)
→
−
→
−
→
−
E[ X{Y − µ(X; Pr )}] ≡ E[ X{Y − m(X; Pr ) − X 0 η Pr }] = 0, and
(2)
b (Pr ,K) − η Pr = T(1)
Γn η
n − Tn,K .
Under (A.4), Assumptions 2.1 (a) and (i), it follows from Lemma A.1 (b)
1
(1)
that Tn = Op (n− 2 ). Next, due to assumption (ii) and boundedness of X,
(2)
kTn,K k ≤ n−1
K X
X
−
b r,k )|} = Op (c − ).
b k (x; Pr , P
supx∈X {k→
x k |∆
n
K
k=1 i∈Ik
1
Finally, under Assumption 2.1 (a), we have: Γn = Γ+Op (n− 2 ) using Lemma
1
−1
A.1 (b). Further, since Γn is invertible a.s., Γ−1
+ Op (n− 2 ). Using all
n =Γ
(1)
(2)
(1)
−1
these facts, we then have: (b
η (Pr ,K) − η Pr ) = Γ−1
n (Tn − Tn,K ) = Γ (Tn −
1
(2)
1
Tn,K ) + Op {n− 2 (n− 2 + cn− )}. Thus,
K
(A.6)
(2)
1
−1
(b
η (Pr ,K) − η Pr ) = Γ−1 (T(1)
+ n− 2 cn− ).
n − Tn,K ) + Op (n
K
Next, let us define:
ΓN = N
−1
n+N
X
n+N
X →
→
− →
−0
−
→
−
(1)
−1
X j X j , RN = N
X j {µ(Xj ; Pr ) − X 0j θ 0 },
j=n+1
b (K) = N −1
and R
N,n
j=n+1
n+N
X
→
−
br,K ) − µ(Xj ; Pr )}.
X j {b
µ(Xj ; P
j=n+1
Then, using (3.7), we have:
b(P ,K) − θ 0 ) = N −1
ΓN ( θ
r
n+N
X
→
−
−0
(1)
b (K) .
br,K ) − →
X j [b
µ(Xj ; P
X j θ 0 ] = RN + R
N,n
j=n+1
b (K) = ΓN (b
b (K) , where
Next, using (3.4)-(3.8), we have: R
η (Pr ,K) − η Pr ) + S
N,n
N,n
− b
b (K) = K−1 PK {N −1 Pn+N →
b
S
k=1
j=n+1 X j ∆k (Xj ; Pr , Pr,k )}.
N,n
29
SEMI-SUPERVISED LINEAR REGRESSION
(1)
b(P ,K) − θ 0 ) = ΓN (b
b (K) .
Hence, we have: ΓN (θ
η (Pr ,K) − η Pr ) + RN + S
r
N,n
Now, under assumptions (i)-(ii) and Assumption 2.1 (a), we have:
PK
→
−b
b
(I)
k=1 supx∈X k x ∆k (x; Pr , Pr,k )k = Op (1),
− 21
b (K) = K−1 PK S
b∗
so that using Lemma A.1 (a), S
), where
k=1 n,k + Op (N
N,n
→
−
∗
b = EX { X ∆
b r,k )} ∀ 1 ≤ k ≤ K;
b k (X; Pr , P
S
n,k
1
1
→
−
→
−
(1)
(II) RN = E[ X{µ(X; Pr ) − X 0 θ 0 }] + Op (N − 2 ) = Op (N − 2 )
→
−
→
−
from Lemma A.1 (b) and E[ X{µ(X; Pr ) − X 0 θ 0 }] = 0 due to (A.4) and 2.1;
1
−1
and lastly, (III) Γ−1
+ Op (N − 2 ). It then follows from (I)-(III) that
N =Γ
b(P ,K) − θ 0 = (b
θ
η (Pr ,K) − η Pr ) + K−1 Γ−1
r
(A.7)
K
X
b ∗ + Op (N − 21 ).
S
n,k
k=1
Using (A.6) and (3.9) in (A.7), we then have:
n
K n
o
X
X
1 X b
−1 1
b(P ,K) − θ 0 = 1
θ
G
(X
)
+ Op (bn,K ),
ψ(Z
;
P
)
−
Γ
i
i
r
k
r
n
K
nK
i=1
where bn,K = n−1 + n
i∈Ik
k=1
− 21
cn− + N
− 12
. It follows, as claimed in (3.10), that
K
1
1
b(P ,K) − θ 0 ) = n− 2
n 2 (θ
r
(A.8)
n
X
ψ(Zi ; Pr ) − Γ−1 Gn,K + Op (c∗n,K )
i=1
We next show that Gn,K = Op (cn− ) for any fixed K ≥ 2. To this end,
K
1 P
(n)
b k (Xi ), D
b r,k )| and C =
b k = supx∈X |∆
b k (x; Pr , P
let Tk = (nK )− 2 i∈Ik G
→
−
supx∈X k x k < ∞. For any subset A ⊆ L, let PA denote the joint distribution
of the observations in A, and let EA (·) denote expectation w.r.t. PA . By
1 P
(n)
definition, Gn,K = K− 2 K
T = Op (cn− ) if and only if given any > 0,
k=1 k
K
∃ M > 0 such that P kGn,K k > M cn− ≤ ∀ n. Note that for any M > 0,
K
− 12
P kGn,K k > M cn− ≤ P K
K
≤
K
X
− 21
P K
k=1
p+1
K X
X
(A.9) =
k
k=1 l=1
M cn−
(n)
EL− PLk
!
(n)
kTk k
> M cn−
K
k=1
kTk k >
"
K
X
p+1
K X
X
!
K
≤
K
(n)
P |Tk[l] | >
k=1 l=1
(
(n)
|Tk[l] |
(
>
)#
M cn−
K
1
1
K 2 (p + 1) 2
L−
k
,
)
M cn−
K
1
2
1
K (p + 1) 2
30
A. CHAKRABORTTY AND T. CAI
where the steps follow from repeated use of Bonferroni’s inequality and other
1
(n)
standard arguments. Now, conditional on L−
⊥ Lk , with K ≥ 2), nK2 Tk is
k (⊥
→
− b
b r,k )}i∈I which,
a centered sum of the i.i.d. random vectors { X i ∆k (Xi ; Pr , P
k
bk < ∞
due to assumption (ii) and the compactness of X , are bounded by: C D
(n)
a.s. [PL− ] ∀ k, n. Hence, applying Hoeffding’s inequality to Tk[l] ∀ l, we have:
k
(
)
2 c2
M
M
c
−
−
nK
nK
(n)
−
L
≤
2
exp
−
(A.10) PLk |Tk[l] | > 1
1
k
b2
2(p + 1)KC 2 D
K 2 (p + 1) 2
k
a.s. [PL− ] ∀ n; for each k ∈ {1, ..., K} and ∀ 1 ≤ l ≤ (p + 1).
k
b k = Op (c − ), (c − /D
b k ) ≥ 0 is stochastically bounded away
Now, since D
nK
nK
from 0. Thus, ∀ k, and for any given > 0, ∃ δ(k, ) > 0 (independent of n)
b k ) ≤ δ(k, )} ≤ ∗ ∀ n, where ∗ = /{4K(p+1)} > 0.
such that: PL− {(cn− /D
K
k
e
Let δ(K,
) = min{δ(k, ) : k = 1, ..., K} > 0 (as K is fixed). Let A(k, )
e
b k ) ≤ δ(K,
denote the event: {(c − /D
)}, and let Ac (k, ) be its complement.
nK
e
b k ) > δ(K,
). Thus, the
Then, PL− {A(k, )} ≤ ∗ , while on Ac (k, ), (cn− /D
K
k
bound in (A.10) is dominated by: 2 exp[−M 2 δe2 (K, )/{2(p + 1)KC 2 }] on
Ac (k, ), and trivially by 2 on A(k, ) ∀ k. Plugging the bound of (A.10) into
(A.9) and using all these facts, we then have:
p+1
K X
M 2 c2n−
X
K
P kGn,K k > M cn− ≤
EL− 2 exp −
K
k
b2
2(p + 1)KC 2 D
k
k=1 l=1
2
2
p+1
K X
M cn−
X
K
1 c
+ 1A(k,)
=
EL− 2 exp −
k
b 2 A (k,)
2(p + 1)KC 2 D
k
k=1 l=1
"
)
(
#
p+1
K
XX
M 2 δe2 (K, )
≤
2 exp −
PL− {Ac (k, )} + 2 PL− {A(k, )}
k
k
2(p + 1)KC 2
k=1 l=1
)
#
"
(
M 2 δe2 (K, )
+ ∗
≤ 2K(p + 1) exp −
2(p + 1)KC 2
(A.11) ≤ + = (with some suitable choice M for M ),
2 2
where the last step follows from noting the definition of ∗ and choosing M
to be any M large enough such that 4 exp[− M 2 δe2 (K, )/{2(p + 1)KC 2 }] ≤
/{K(p + 1)}. Thus, (A.11) shows Gn,K = Op (cn− ) for any fixed K ≥ 2. This
K
further establishes (3.11) and all its associated implications. The proof of
Theorem 3.2 is now complete.
SEMI-SUPERVISED LINEAR REGRESSION
31
ACKNOWLEDGEMENTS
The authors would like to thank Dr. James Robins and Dr. Eric Tchetgen
Tchetgen for many helpful discussions throughout the progress of this paper,
as well as the editor Dr. Edward George, the anonymous associate editor
and the two referees for their useful comments and suggestions that helped
significantly in improving and revising the original version of this article.
SUPPLEMENTARY MATERIAL
Supplement to “Efficient and Adaptive Linear Regression in
Semi-Supervised Settings”
(doi: COMPLETED BY THE TYPESETTER; .pdf). The supplement includes: (i) Supplementary results for the simulation studies and the real data
analysis; (ii) Discussions on generalization of the proposed SS estimators to
MAR settings; (iii) Proof of Lemma A.1; (iv) Proof of Theorem 3.1; (v)
Proof of Theorem 4.1; and (vi) Proofs of Lemmas A.2-A.3 and Theorem 4.2.
SEMI-SUPERVISED LINEAR REGRESSION
i
SUPPLEMENT TO “EFFICIENT AND ADAPTIVE LINEAR
REGRESSION IN SEMI-SUPERVISED SETTINGS”
BY ABHISHEK CHAKRABORTTY AND TIANXI CAI
University of Pennsylvania and Harvard University
This supplementary document contains additional numerical results as well
as technical materials, including proofs etc., that could not be accommodated in the main article.
I. Numerical Studies: Supplementary Results.
I.1. Simulation Results for p = 2. For p = 2, we investigated three choices
of m(x) as follows:
(Linear): m(x) = x1 + x2 ;
(NL-Iλ(k) ): m(x) = x1 + x2 + λ(k) x1 x2 for λ(1) = 0.5 and λ(2) = 1; and
(NL-Qγ (k) ): m(x) = x1 + x2 + γ (k) (x21 + x22 ) for γ (1) = 0.3 and γ (2) = 1.
Since the dimension is low, we implemented EASE using the KS and KM
smoothers with P2 = I2 for both, i.e. without any dimension reduction.
For comparison, the other SS estimators were also obtained. In Table I,
we summarize the efficiencies of all the estimators relative to OLS, based
e the
on the empirical mean squared error (MSE), where for any estimator θ,
2
e 0 k averaged over the 500 replications.
empirical MSE is summarized as kθ−θ
Models
Linear
NL-Iλ(1)
NL-Iλ(2)
NL-Qγ (1)
NL-Qγ (2)
OLS
(Ref.)
1
1
1
1
1
SNP EASE
(T := KS)
0.897
0.995
1.229
1.243
2.261
2.261
2.241
2.215
4.096
4.144
SNP EASE
(T := KM)
0.920
0.988
1.338
1.355
2.301
2.267
2.500
2.550
4.612
4.641
Other SS Estimators
DRESS1 DRESS3 MSSL
0.993
0.963
0.993
1.072
1.039
1.072
1.217
1.181
1.216
1.187
2.063
1.187
1.352
3.217
1.352
Table I: Efficiencies of SNP and EASE, obtained using T := KS or KM, as
well as DRESS1 , DRESS3 and MSSL, relative to OLS with respect to the
empirical MSE under the various models considered with p = 2.
For this setting, all estimators have comparable efficiency under the linear
model, as expected. Under the non-linear models, the EASE estimators are
substantially more efficient than the OLS and also more efficient than the
other SS estimators. For the non-linear models with quadratic effects, the
DRESS3 is also substantially more efficient than the OLS while our EASE
estimator performs even better. For the non-linear models with interaction
ii
A. CHAKRABORTTY AND T. CAI
effects, the efficiency gain was very modest when employing existing SS
estimation procedures while it was quite substantial for EASE.
I.2. Supplementary Results for the Data Example. We present in Table
II some summary measures of the distributions in the labeled and unlabeled
data for each of the predictors in the data example, and also report p-values
for diagnostic tests aimed at detecting any possible differences in the labeled
and unlabeled data distributions for each of the predictors.
Predictors
Age
Gender
Race
Lupus
PmR
RA
SpA
Other ADs
Erosion
Seropositivity
Anti-CCPprior
Anti-CCPmiss
RF
RFmiss
Azathioprine
Enbrel
Gold salts
Humira
Infliximab
Leflunomide
Methotrexate
Plaquenil
Sulfasalazine
Other meds.
(Intercept)
Labeled Data
Mean
Sd
4.090 0.241
0.786 0.411
0.696 0.461
0.230 0.520
0.057 0.336
4.171 1.071
0.073 0.343
0.251 0.642
0.577 0.495
0.369 0.483
0.386 0.629
0.645 0.479
0.949 0.772
0.307 0.462
0.121 0.397
0.738 0.858
0.346 0.568
0.856 0.833
0.386 0.673
0.549 0.740
1.417 0.638
0.248 0.464
0.535 0.752
0.163 0.380
–
–
Unlabeled Data
Mean
Sd
4.070
0.272
0.799
0.401
0.673
0.469
0.251
0.600
0.078
0.382
4.084
1.079
0.066
0.313
0.271
0.690
0.567
0.494
0.395
0.489
0.405
0.718
0.610
0.488
0.897
0.846
0.324
0.468
0.137
0.419
0.722
0.822
0.336
0.555
0.917
0.835
0.400
0.672
0.555
0.743
1.389
0.669
0.273
0.496
0.554
0.734
0.189
0.403
–
–
P-values from Diagnostic Tests
T-test Wilcoxon Test PS Model
0.151
0.373
0.475
0.556
0.547
0.574
0.371
0.378
0.456
0.461
0.877
0.689
0.269
0.326
0.255
0.144
0.055
0.255
0.716
0.932
0.780
0.582
0.780
0.540
0.709
0.701
0.743
0.331
0.335
0.231
0.592
0.471
0.564
0.192
0.198
0.089
0.227
0.184
0.199
0.516
0.520
0.410
0.463
0.449
0.734
0.740
0.921
0.574
0.729
0.791
0.962
0.193
0.190
0.177
0.711
0.607
0.789
0.895
0.912
0.973
0.435
0.584
0.549
0.331
0.463
0.398
0.661
0.458
0.724
0.223
0.192
0.322
–
–
0.000
Table II: Comparison of the means and standard deviations (sd) from the
labeled and unlabeled data for each predictor in the data example. Shown
also are the p-values obtained from various diagnostic tests, testing for possible differences in the distributions of each of the predictors in the labeled
and unlabeled data, including a two-sample T-test (with possibly unequal
variances in the two populations), a Wilcoxon rank sum test, and a test
obtained by fitting a parametric logistic regression model for the propensity
score (PS) of missingness, with all the predictors included as covariates, and
then testing for the null effect of each of the predictors in the fitted model.
iii
SEMI-SUPERVISED LINEAR REGRESSION
I.3. Simulation Results on the Prediction Error. In Table III, we present
the out-of-sample mean squared prediction error for various SNP imputation
estimators under all the models considered in the simulation studies for
p = 2, 10 and 20. The results suggest that the SNP imputation estimators
for both KS and KM based smoothers perform substantially better than the
naive non-parametric estimator based on a p-dimensional kernel smoothing.
(a) p = 2
Models
Linear
NL-Iλ(1)
NL-Iλ(2)
NL-Qγ (1)
NL-Qγ (2)
T := KS
m
bT
µ
bT
0.36 0.35
0.34 0.33
0.29 0.28
0.29 0.29
0.20 0.20
T := KM
m
bT
µ
bT
0.35 0.35
0.32 0.32
0.27 0.27
0.27 0.27
0.17 0.18
(b) p = 10 and 20
Models
(I)
(II)
Linear
NL1C
NL2C
NL3C
Linear
NL1C
NL2C
NL3C
T := KS2,P2
m
bT
µ
bT
0.186 0.174
0.143 0.131
0.241 0.228
0.275 0.254
0.106 0.096
0.147 0.131
0.179 0.163
0.218 0.199
p = 10
T := KM
m
bT
µ
bT
0.214 0.204
0.147 0.154
0.150 0.156
0.138 0.146
0.133 0.126
0.135 0.143
0.135 0.142
0.137 0.144
KSp
0.260
0.290
0.554
0.458
0.313
0.658
0.487
0.527
T := KS2,P2
m
bT
µ
bT
0.143 0.099
0.337 0.332
0.557 0.566
0.543 0.552
0.095 0.054
0.341 0.336
0.423 0.422
0.424 0.424
p = 20
T := KM
m
bT
µ
bT
0.130 0.119
0.280 0.317
0.271 0.299
0.282 0.327
0.083 0.073
0.283 0.331
0.272 0.302
0.282 0.318
KSp
0.698
0.921
1.000
0.974
0.755
1.095
0.980
0.996
Table III: Mean squared prediction errors (PEs), relative to Var(Y ), of the
b r ) and µ
br,K ), with T :=
SNP imputation estimators m
b T = m(x;
b
P
bT = µ
b(x; P
KS2,P2 or T := KM, under the various models discussed in the simulation
studies. Shown also are the corresponding PEs for the fully non-parametric
KS estimator, KSp , for comparison, in the case of p = 10 and 20.
II. Generalization to the Missing at Random (MAR) Case. Our
SS estimation methods proposed so far assume that the underlying Y for
subjects in U are MCAR, a standard (and often implicit) assumption made
in SSL. In this section, we provide some discussions on possible generalizations of our SS methods to the MAR case. Such generalizations might be
desirable for settings where the availability of Y is not determined by design. To this end, let N̄ = N + n denote the sample size of the entire data
S = L ∪ U. Then, S = {Zi ≡ (Ri , Ri Yi , Xi ) : i = 1, . . . , N̄ } consists of N̄
i.i.d. realizations of Z = (R, RY, X), where R ∈ {0, 1} denotes the indicator
iv
A. CHAKRABORTTY AND T. CAI
of Y being observed. As opposed to the stronger MCAR setting with the
assumption R ⊥
⊥ (Y, X) and the probability law of S being determined by
the law of (Y, X), we now have: under the MAR setting, R ⊥
⊥ Y | X and the
probability law of S is determined by PZ , the law of Z. For notational ease,
we also let PN̄ denote the empirical measure for S, and for any function e(·)
P
of Z, possibly random andR vector-valued, we let PN̄ (e) = N̄ −1 N̄
i=1 e(Zi ),
and PZ (e) = EZ {e(Z)} = e(z)dPZ (z).
P
Under a SS set-up as above, we have: n = N̄
i=1 Ri is a random quantity
and n/N̄ → 0 in probability. It is important to note that πN̄ ≡ P(R = 1)
must depend on N̄ . Let πN̄ (X) = P(R = 1 | X) be the “propensity score”,
assumed to be strictly greater than 0 almost surely (a.s.) for any given N̄
1
−2
and let bN̄ = [E{πN̄
(X)}]− 2 . Then, under the above set-up, we assume that
E{πN̄ (X)} = πN̄ → 0, bN̄ → 0 and N̄ bN̄ → ∞ as N̄ → ∞. This decaying
sampling probability is the main factor that distinguishes SSL from standard
missing data problems and contributes to the complexity of devising and
analyzing SS estimators which, even for the MCAR setting, were seen to only
1
1
have a convergence rate of (N̄ bN̄ )− 2 , rather than N̄ − 2 , with bN̄ = (n/N̄ ).
For simplicity, we shall first assume that πN̄ (X) is known and next detail
how we may extend our proposed procedures to obtain SS estimators of θ 0 ,
→
−
→
−
the solution to: φ(θ) ≡ E{ X(Y − X 0 θ)} = 0, under the MAR setting.
To derive efficient SS estimators of θ 0 based on S under MAR, we first
b(P ,K) in Section 3.2 in fact remains
note that our proposed SNP estimator θ
r
valid even under the MAR setting, whenever the imputation is sufficient
i.e. µ(X; Pr ) equals the true conditional mean m(X) ≡ E(Y |X). For the
general case allowing for insufficient imputation, we need to modify our
SNP imputation to account for the MAR setting. To this end, we note that
→
−
→
−0
−
→
−0
R →
φ(θ) ≡ E{ X(Y − X θ)} = E
(1)
X(Y − X θ) ,
πN̄ (X)
→
−
→
−
(2)
= E[ X{m(X) − X 0 θ}],
→
−
→
−0
−
R →
(3)
= E[ X{m(X) − X θ}] + E
X{Y − m(X)} .
πN̄ (X)
More generally, for any µ(·) ∈ L2 (PX ) satisfying:
−
→
−
R →
X{Y − µ(X)} ≡ E[ X{m(X) − µ(X)}] = 0,
(4)
E
πN̄ (X)
it is easy to see that the following representation of φ(·) holds under MAR:
→
−
→
−0
−
R →
(5)
φ(θ) = E[ X{µ(X) − X θ}] + E
X{Y − µ(X)} .
πN̄ (X)
v
SEMI-SUPERVISED LINEAR REGRESSION
Let µ
b(·) be any estimator of µ(·) based on S. Motivated by (5), we may
bMAR ≡ θ
bMAR, µ(·) , as the solution to:
then devise a SS estimator of θ 0 , θ
(6)
b (θ) ≡ N̄ −1
φ
N̄
N̄
X
→
−
→
−
X i {b
µ(Xi ) − X 0i θ} +
i=1
−
Ri →
X i {Yi − µ
b(Xi )} .
πN̄ (Xi )
P →
− →
−0
Then, letting ΓN̄ = N̄ −1 N̄
i=1 X i X i , it is straightforward to show that:
bMAR − θ 0 = P (T) + P (S) − P (b
(7)
ΓN̄ θ
N̄
N̄
N̄ e), where
→
−
→
−
→
−
where T(Z) = X{µ(X) − X 0 θ 0 }, S(Z) = {R/πN̄ (X)} X{Y − µ(X)}, and
→
−
R
b
− 1 X{b
µ(X) − µ(X)}.
e(Z) ≡
πN̄ (X)
Convergence rates of the terms in (7) need careful analysis as the asymp1
totics here is non-standard, with the dominating rate being slower than N − 2 .
To this end, note that PN̄ (T) is a simple centered i.i.d. average of vari1
ables with bounded variance. Hence, TN̄ = O(N − 2 ) indeed. On the other
hand, PN̄ (S) has a slower convergence rate since the variance of S, VN̄ ,
diverges due to the πN̄ (·) ↓ 0 appearing in the denominator. Under mild
moment conditions, it can be shown that bN̄ VN̄ converges to a positive definite matrix V with kVk < ∞. Hence, using concentration inequalities, and
assuming N̄ bN̄ → ∞, it can be shown that the convergence rate of PN̄ (S) is
1
O{(N bN̄ )− 2 }. Further, using CLT for triangular arrays, it can be shown un1
d
der suitable conditions that (N bN̄ ) 2 PN̄ (S) → N(p+1) [0, V]. Lastly, to control
the term PN̄ (b
e), note that PZ (b
e) = 0. Therefore, GN̄ is a centered empirical
P
process indexed by µ
b(·) − µ(·). Hence, as long as EX [{b
µ(X) − µ(X)}2 ] → 0,
and µ
b(·) − µ(·) lies in a P−Donsker class with probability → 1, it can be
shown using results from empirical process theory (see Van der Vaart (2000)
1
for instance) that (N bN̄ ) 2 PN̄ (b
e) = op (1). Finally, note that ΓN̄ 0 a.s., and
−1
−1
− 21
ΓN̄ = Γ +Op (N ). Hence, under suitable regularity conditions, we have:
1
bMAR − θ 0 )
(N̄ bN̄ ) 2 (θ
=
1
(N̄ bN̄ ) 2 Γ−1
N̄
1 X
S(Zi ) + op (1)
N̄ i=1
d
→ N(p+1) [0, Γ−1 VΓ−1 ], with V as defined above.
Having now provided an abstract sketch of the construction of the estimators and their properties, we next briefly discuss the choice of the
vi
A. CHAKRABORTTY AND T. CAI
‘imputation’ function µ(·), and its estimator µ
b(·) inherent in the construcbMAR ≡ θ
bMAR, µ(·) . With (r, Pr , P
b r ) as defined in Section 3.2 and
tion of θ
{K(·), h, Kh (·, ·)} as in Section 4, we may modify the SNP estimator in Section 3.2 under the MAR setting as follows: consider µ(X) ≡ µ(X; Pr ) =
→
−
m(X; Pr ) + X 0 η Pr , where m(X; Pr ) = E(Y |P0r X), and η Pr satisfies
−
→
−0
R →
E
X{Y − m(X; Pr ) − X η Pr } = 0.
πN̄ (X)
b r)
This will ensure that (4) holds. Then, we may estimate µ(X; Pr ) as µ
b(X; P
→
−
b r ) + X 0η
b Pr , where
= m(X;
b
P
PN̄
b r) =
m(x;
b
P
Ri
b0
b0
i=1 πN̄ (Xi ) Yi Kh (Pr Xi , Pr x)
,
PN̄
Ri
b0
b0
i=1 πN̄ (Xi ) Kh (Pr Xi , Pr x)
b Pr satisfies: N̄ −1
η
N̄
X
i=1
and
−
−0
Ri →
b r) − →
b Pr } = 0.
X i {Yi − m(X
b i; P
X iη
πN̄ (Xi )
Thus, to accommodate the MAR setting, one essentially needs to implement
appropriately weighted versions of both the smoothing and the refitting steps
in our original SNP imputation. Of course, while we have chosen the smoothing method T to be the weighted KS here for illustration, other reasonable
choices of T such as an appropriately weighted KM may also be used. Under
bMAR
MCAR, with πN̄ (X) ≡ πN̄ ≡ n/N̄ and bN̄ = (n/N̄ ), the estimator θ
indeed becomes (asymptotically) equivalent to the SNP estimators obtained
earlier in Section 3.2. Further, with various choices of µ(·), the SNP imputation strategy again equips us with a family of SS estimators of θ 0 under
the MAR setting, with µ(·) = m(·) leading to the optimal estimator.
bMAR is derived with a known π (·), for simplicity.
The above estimator θ
N̄
In practice, πN̄ (·) is typically unknown and a consistent estimator π
bN̄ (·)
b
may be constructed. Then, one may modify θ MAR by replacing πN̄ (·) with
π
bN̄ (·) in all the steps. The resulting estimator will have an expansion similar
to (7) but with extra error terms accounting for the variability in π
bN̄ (·),
which need to be properly controlled. The theoretical analysis will be more
involved since establishing the convergence rates and asymptotic expansion
for π
bN̄ (·) − πN̄ (·) is also non-standard due to bN̄ → 0. Lastly, the score
bMAR has the additional benefit of ‘double
equation (5) used to construct θ
robustness’, in the sense that even if π
bN̄ (·) is inconsistent for πN̄ (·), as long
bMAR, µ(·) is consistent for θ 0 . On the other
as µ
b(·) estimates the true m(·), θ
hand, as long as π
bN̄ (·) targets the true πN̄ (·), then for any choice of µ(·),
SEMI-SUPERVISED LINEAR REGRESSION
vii
bMAR, µ(·) is consistent for θ 0 . For the MCAR case, π
θ
bN̄ (·) ≡ π
bN̄ = n/N̄
is always consistent for πN̄ (·) and in fact this is exactly what allowed us
to achieve a family of SNP estimators, all consistent for θ 0 , for various
choices of the SNP imputation function µ(·). For a specific choice of µ(·), this
b Note however that under MAR,
also included the supervised estimator θ.
b
θ simply denotes the naive ‘complete case’ estimator and in general, is not
consistent for θ 0 unless πN̄ (·) is trivially a constant or m(·) is exactly linear.
III. Proof of Lemma A.1. Firstly, since d is fixed, it suffices to prove
(j)
b (j)
b
the result for any arbitrary scalar coordinate G
n,m ≡ Gn,m (say) and Gn ≡
b n,m and Gn respectively, for any j ∈ {1, . . . , d}. For any data
G n (say) of G
∗
S and S , we let PS and PS,S∗ denote the joint probability distributions of
the observations in S and (S, S∗ ) respectively, ES (·) denote the expectation
w.r.t PS , and PS|S∗ denote the conditional probability distribution of the
observations in S given S∗ .
1
To show that Gbn,m − G n = Op (m− 2 ), we first note that since Sn ⊥
⊥ Sm ,
n
o
1
1
PSn ,Sm |Gbn,m − G n | > m− 2 t = ESn PSm |Gbn,m − G n | > m− 2 t Sn ,
b n,m − Gn is a centered average of
for any t > 0. Now, conditional on Sn , G
m
{b
gn (Zj )}j=1 which are i.i.d. and bounded by Tbn < ∞ a.s. [PSn ] ∀ n. Hence,
applying Hoeffding’s inequality, we have for any n and m,
!
2 t2
1
2m
a.s. [PSn ].
(8) PSm |Gbn,m − G n | > m− 2 t Sn ≤ 2 exp −
4m2 Tbn2
Now, since Tbn ≥ 0 is Op (1), we have: for any given > 0, ∃ δ() > 0 such
that: PSn {Tbn > δ()} ≤ /4 ∀ n. Let A() denote the event: {Tbn > δ()} and
let Ac () denote its complement. Then, using (8), we have: ∀ n and m,
(
!)
2 t2
1
2m
−
PSn ,Sm |Gbn,m − G n | > m 2 t ≤ ESn 2 exp −
4m2 Tbn2
(
!)
"
!
#
t2
t2
= ESn 2 exp −
= ESn 2 exp −
1Ac () + 1A()
2Tbn2
2Tbn2
t2
c
≤ 2 exp − 2
PSn {A ()} + 2 PSn {A()}
2δ ()
t2
≤ 2 exp − 2
+
≤ + = (for some suitable choice of t),
2δ ()
2
2 2
viii
A. CHAKRABORTTY AND T. CAI
where the last step follows by choosing t ≡ t to be any large enough t
such that exp{−t2 /2δ 2 ()} ≤ /4. Such a choice of t clearly exists. This
establishes the first claim (a) in Lemma A.1. The second claim (b) in Lemma
A.1 is a trivial consequence of the Central Limit Theorem (CLT).
IV. Proof of Theorem 3.1. To show Theorem 3.1, we first note that
1
1
under Assumption 3.1 (i)-(v), and letting an = (log n) 2 (nhp )− 2 + hq , the
following holds:
(9)
supx∈X |m(x)
b
− m(x)| = Op (an ) = supx∈X |fb(x) − f (x)|.
(9) is a fairly standard result and we only provide a sketch of its proof as
follows. Under Assumption 3.1 (ii)-(iii), using Theorem 2 of Hansen (2008),
supx∈X |b
l(x) − EL {b
l(x)}| = Op (a∗n ) = supx∈X |fb(x) − EL {fb(x)}|, where a∗n =
1
1
(log n) 2 (nhp )− 2 . Next, using standard arguments based on Taylor series
expansions of l(·) and m(·) under their assumed smoothness, and noting
that K(·) is a q th order kernel having finite q th moments, we obtain:
supx∈X |EL {b
l(x)} − l(x)| = O(hq ) = supx∈X |EL {fb(x)} − f (x)|.
Combining these two results, and the definitions of m(.) and m(.)
b along with
Assumption 3.1 (iv), we have (9). Next, note that using (3.2), we have:
bnp − θ 0 ) = EU [N −1
ΓN (θ
n+N
X
1
→
−
→
−
X j {m(X
b j ) − X 0j θ 0 }] + Op (N − 2 )
j=n+1
1
→
−
= EX [ X{m(X)
b
− m(X)}] + Op (N − 2 ),
−
→
−
where the first step is due to Lemma A.1 (a) with supx∈X k→
x {m(x)−
b
x 0 θ 0 }k
→
−
→
−
0
b
− m(x)| + |m(x) − x θ 0 |}] = Op (1) due to (9) and
≤ supx∈X [k x k{|m(x)
→
−
the boundedness of X and m(·), while the last step uses: EX [ X{m(X) −
→
−0
X θ 0 }] = 0 which follows from the definitions of θ 0 and m(·). It then follows
1
−1
further, using Γ−1
+ Op (N − 2 ), that
N =Γ
n 1
1
−
2
bnp − θ 0 ) = n 12 Γ−1 EX [→
n 2 (θ
X{m(X)
b
− m(X)}] + Op
.
N
P
Letting φn (X) = (nhp )−1 ni=1 K{(X − Xi )/h}{Yi − m(X)}, and expanding
the first term in the above equation, we now obtain:
(10)
1
1
bnp − θ 0 = Γ−1 T(1) + T(2) + Op n 2 ,
n2 θ
n,1
n,1
N
SEMI-SUPERVISED LINEAR REGRESSION
ix
1
→
−
(1)
where Tn,1 = n 2 EX { Xφn (X)/f (X)} and
h→
i
1
−
(2)
Tn,1 = n 2 EX Xφn (X){fb(X)−1 − f (X)−1 }
1
→
−
= n 2 EX [ X {m(X)
b
− m(X)} {f (X) − fb(X)}/f (X)]
n
o
1
1
−
(11) ≤ n 2 supx∈X k→
x k |m(x)
b
− m(x)| fb(x)/f (x) − 1
= Op n 2 a2n ,
where the last step in (11) follows from (9), Assumption 3.1 (iv) and the
(1)
boundedness of X. For Tn,1 , we have:
(1)
1
Z
1
n
X
Tn,1 = n 2
(12)
1
→
−
x φn (x)dx = n− 2
X
= n2
n Z
X
i=1
n−1
Z
Ai,n
i=1
→
−
x h−p Kh (x − Xi ) {Yi − m(x)} dx
X
−−−−−−−→
(Xi + hψ i ) K (ψ i ) {Yi − m(Xi + hψ i )} dψ i ,
where ψ i = (x − Xi )/h and Ai,n = {ψ i ∈ Rp : (Xi + hψ i ) ∈ X }. Now, since
K(·) is zero outside the bounded set K, the ith integral in (12) only runs over
(Ai,n ∩ K). Further, since h = o(1), using Assumption 3.1 (vi), Ai,n ⊇ K a.s.
[PL ] or, (Ai,n ∩ K) = K a.s. [PL ] ∀ 1 ≤ i ≤ n with n large enough. Thus, for
large enough n, (12) can be written as:
(1)
Tn,1
(13)
(14)
=n
− 21
1
= n2
n Z
X
−−−−−−−→
(Xi + hψ i ) K (ψ i ) {Yi − m(Xi + hψ i )} dψ i a.s. [PL ]
i=1
n
X
K
h→
i
−
n−1 X i {Yi − m(Xi )} + Op (hq )
i=1
n
X
1
= n− 2
1
→
−
X i {Yi − m(Xi )} + Op n 2 hq ,
i=1
where (13), and hence (14), follows from standard arguments based on Taylor
series expansions of m(Xi + hψ i ) around m(Xi ) under the assumed smoothness of m(·), and using the fact that K(·) is a q th order kernel. Combining
1
1
(10), (11) and (14), and noting that under our assumptions, (n 2 a2n + n 2 hq )
1
1
= O{n 2 hq + (log n)(n 2 hp )−1 }, the result of Theorem 3.1 now follows.
x
A. CHAKRABORTTY AND T. CAI
1
1
V. Proof of Theorem 4.1. Let an,2 = (log n) 2 (nhr )− 2 + hq . Then, we
first note that
(%)
(%)
supw∈XPr |ϕ
ePr (w) − ϕPr (w)| = Op (an,2 ),
(15)
∀ % ∈ {0, 1}.
To see this, note that under Assumption 4.1 (ii)-(iii), Theorem 2 of Hansen
1
1
(2008) applies, and we have for dn = (log n) 2 (nhr )− 2 ,
(%)
(%)
supw∈XPr |ϕ
ePr (w) − EL {ϕ
ePr (w)}| = Op (dn )
∀ % ∈ {0, 1}.
Next, using standard arguments based on a q th order Taylor series expansion
(%)
of ϕPr (·) and noting that K(·) is a q th order kernel, we obtain:
(%)
(%)
ePr (w)} − ϕPr (w)| = O(hq )
supw∈XPr |EL {ϕ
∀ % ∈ {0, 1}.
Combining these two results gives (15). Further,
supx∈X |m(x;
e
Pr ) − m(x; Pr )| = supw∈XPr |m
e Pr (w) − mPr (w)|
(
)
e
lPr (w) − lPr (w)
|lPr (w)| |lPr (w)|
≤ supw∈XPr
+ supw∈XPr
−
fPr (w)
feP (w)
feP (w)
r
r
(16) = Op (an,2 ),
where the last step follows from repeated use of (15) and Assumption 4.1
b r )− ϕ
(iii)-(iv). Next, we aim to bound supx∈X |ϕ
b(%) (x; P
e(%) (x; Pr )| to account
b r . Using a first order Taylor series
for the potential estimation error of P
expansion of K(.) under Assumption 4.1 (vi), we have: ∀ % ∈ {0, 1},
n
X
x
−
X
1
i
0
0
b r − Pr )
b r) − ϕ
∇K (wi,x )(P
Yi%
ϕ
b (x; P
e (x; Pr ) =
nhr
h
i=1
o
o
n
n
0 c (1)
b r − Pr )0 M
c (2)
b
(17)
= trace (Pr − Pr ) Mn,%,x + trace (P
n,%,x ,
(%)
(%)
where
0
n
0
1 X
0 Pr x − Pr X i
(x − Xi ) ∇K
Yi% and
nhr+1
h
i=1
0
n
0
X
1
0
0 Pr x − Pr Xi
=
(x − Xi ) ∇K (wi,x ) − ∇K
Yi% ,
r+1
nh
h
c (1)
M
n,%,x =
c (2)
M
n,%,x
i=1
with wi,x ∈ Rr being ‘intermediate’ points satisfying: kwi,x −P0r (x−Xi )h−1 k
b 0 (x−Xi )h−1 −P0 (x−Xi )h−1 k ≤ Op (αn h−1 ). The last bound, based on
≤ kP
r
r
xi
SEMI-SUPERVISED LINEAR REGRESSION
b r − Pr ) = Op (αn ) and the compactness of X , is uniform in (i, x). For any
(P
matrix A = [aij ], let kAkmax denote the max-norm of A, and |A| denote the
matrix [|aij |]. Now, Assumption 4.1 (viii) implies: k∇K(w1 ) − ∇K(w2 )k ≤
Bkw1 −w2 k ∀ w1 , w2 ∈ Rr , for some constant B < ∞. Then using the above
c (2)
arguments, we note that ∀ % ∈ {0, 1}, k supx∈X |M
n,%,x |kmax is bounded by:
)
(
n
B X
P0r x − P0r Xi
supx∈X
|Yi% |
kx − Xi k wi,x −
nhr+1
h
i=1
(
)
n
b r − Pr )0 (x − Xi )
B X
(P
%
≤ supx∈X
|Yi |
kx − Xi k
nhr+1
h
i=1
≤
sup
n
o
b r − Pr )0 (x − X)k
kx − Xkk(P
x∈X ,X∈X
n
α
B X %
n
|Y
|
≤
O
.
p
i
nhr+2
hr+2
i=1
The first two steps above use the triangle inequality, the Lipschitz continuity
of ∇K(·) and the definition of wi,x , while the next two use the compactness
of X , the uniform bound obtained in the last paragraph, the Law of Large
b r − Pr ) = Op (αn ). Thus, we have:
Numbers (LLN), and that (P
2
o
n
(2)
0
b r − Pr ) M
c n,%,x = Op αn
(18) supx∈X trace (P
∀ % ∈ {0, 1}.
hr+2
c (1)
c (1)
c (1,1) c (1,2)
Now for bounding M
n,%,x , let us first write it as: Mn,%,x = Mn,%,x − Mn,%,x ,
P
(1,2)
c n,%,x
c (1,1)
=
where M
= (nhr+1 )−1 ni=1 x∇K 0 {P0r (x − Xi )/h}Yi% and M
n,%,x
P
%
n
0
0
r+1
−1
(nh )
i=1 Xi ∇K {Pr (x − Xi )/h}Yi ∀ % ∈ {0, 1}. Then, under Assumption 4.1 (iii), (vi) and (vii), using Theorem 2 of Hansen (2008) along
with the compactness of X , we have: for each s ∈ {1, 2} and % ∈ {0, 1},
(19)
supx∈X
c (1,s)
c (1,s)
M
n,%,x − EL Mn,%,x
max
≤ Op
log n
nhr+2
1
2
.
Now, ∀ % ∈ {0, 1}, let ν (%) (w) = E{Y % | XPr = w}fPr (w) and ξ (%) (w) =
E{XY % | XPr = w}fPr (w). Further, let {∇ν (%) (w)}r×1 and {∇ξ (%) (w)}p×r
denote their respective first order derivatives. Then, ∀ % ∈ {0, 1}, we have:
c (1,1)
supx∈X EL M
n,%,x
max
0
Z
x
(%)
0 Pr x − w
dw
= supx∈X r+1 ν (w) ∇K
h
h
max
Z
(%)0
0
(20)
= O(1),
Pr x + hψ K(ψ)dψ
= supx∈X x ∇ν
max
xii
A. CHAKRABORTTY AND T. CAI
=
(21)
=
(1,2)
c n,%,x
supx∈X EL M
max
0
Z
(%)
0 Pr x − w
−(r+1)
ξ (w) ∇K
dw
supx∈X h
h
Z
supx∈X D(x) ∇ξ (%) P0r x + hψ K(ψ)dψ
max
= O(1),
max
where, ∀ x ∈ X , D(x) denotes the p × p diagonal matrix: diag(x[1] , . . . , x[p] ).
In both (20) and (21), the first step follows from definition, the second from
standard arguments based on integration by parts (applied coordinate-wise)
and change of variable, while the last one is due to compactness of X and
a medley of the conditions in Assumption 4.1 namely, boundedness and
integrability of K(·) and ∇K(·), (iii) and (v) for (20) so that ∇ν (%) (·) is
bounded on XPr , and (ix) for (21). It now follows that for each % ∈ {0, 1},
c (1)
supx∈X EL M
(22)
= O(1).
n,%,x
max
1
1
Letting d∗n = (log n) 2 (nhr+2 )− 2 , we now have from (19) and (22):
o
n
b 0 − P0 )M
c (1)
(23) supx∈X trace (P
= Op (αn d∗n + αn ) ∀ % ∈ {0, 1}.
n,%,x
r
r
Applying (23) and (18) to (17) using the triangle inequality, we have ∀ %,
(
)
1
2
2
(log
n)
α
n
(%)
(%)
b r )− ϕ
+ αn
(24) supx∈X |ϕ
b (x; P
e (x; Pr )| = Op
.
1 + αn
hr+2
(nhr+2 ) 2
b r) = b
b r )/fb(x; P
b r) = ϕ
b r )/ϕ
b r ).
Finally, note that m(x;
b
P
l(x; P
b(1) (x; P
b(0) (x; P
Repeated use of (24), along with (16) and Assumption 4.1 (iii)-(iv), leads
to:
b r ) − m(x; Pr )
supx∈X m(x;
b
P
b r ) − m(x;
≤ supx∈X m(x;
b
P
e
Pr ) + supx∈X |m(x;
e
Pr ) − m(x; Pr )|
(
)
b
e
e
b r) − e
l(x; P
l(x; Pr )
l(x; Pr )
l(x; Pr )
≤ supx∈X
+
−
+ Op (an,2 )
b r)
b r)
fb(x; P
fe(x; Pr ) fb(x; P
(
)
1
αn2
(log n) 2
(25) ≤ Op
+ αn
+ Op (an,2 ) = Op (an,1 + an,2 ).
1 + αn
hr+2
(nhr+2 ) 2
The proof of Theorem 4.1 is now complete.
SEMI-SUPERVISED LINEAR REGRESSION
xiii
VI. Proofs of Lemmas A.2-A.3 and Theorem 4.2.
VI.1. Proof of Lemma A.2. First note that for each % ∈ {0, 1},
Z
n X
n
X
(n,%)
ϕ
e(%) (x; Pr )Pn (dx) = n−2
Hi1 ,i2
i1 =1 i2 =1
(n,%)
is a V-statistic, where Hi1 ,i2 = h−r λ(Xi1 )Yi%2 K{P0r (Xi1 − Xi2 )/h}. Using
the V-statistic projection result given in Lemma 8.4 of Newey and McFadden
(1994), it then follows that for each % ∈ {0, 1},
n
o
G∗n λ(·)[ϕ
e(%) (· ; Pr ) − EL {ϕ
e(%) (· ; Pr )}]
h
i
1
1
1
(n,%)
(n,%)
= n− 2 Op E(kHi1 ,i1 k) + {E(kHi1 ,i2 k2 )} 2 = Op n− 2 h−r ,
(26)
% having finite
The last step follows from K(·) and λ(·)nbeing bounded and Y o
1
(%)
e? (· ; Pr )}] is a centered
2nd moments. Now, observe that n 2 G∗n λ(·)[EL {ϕ
sum of i.i.d. random vectors bounded by:
n
o
(%)
Dn,% = supx∈X kλ(x)k |EL {ϕ
e? (x; Pr )}| = O(hq ) ∀ % ∈ {0, 1},
˜ with population limit ξ(·), we use
where throughout, for any estimator ξ(·)
˜ − ξ(·).
the notation ξ˜? (·) to denote its centered version given by: ξ˜? (·) = ξ(·)
q
e? (x; Pr )}| =
Here, Dn,% = O(h ) since λ(·) is bounded and supx∈X |EL {ϕ
(%)
(%)
q
ePr (w)} − ϕPr (w)| = O(h ), as argued while proving (15).
supw∈XPr |EL {ϕ
Hence, ∃ a constant κ% > 0 such that hq /Dn,% ≥ κ% ∀ n. Then, using
Hoeffding’s Inequality, we have: ∀ n, given any > 0 and any M = M ()
large enough,
d
n
o
X
M 2 h2q
M hq
(%)
∗
≤ 2d exp −
⇒
P Gn λ[l] (·)[EL {ϕ
e? (· ; Pr )}] > 1
2
2dDn,%
d2
l=1
!
h
n
o
i
M 2 κ2%
(%)
∗
q
e? (· ; Pr )}] > M h ≤ 2d exp −
≤ ⇒
P Gn λ(·)[EL {ϕ
2d
n
o
(%)
(27)
G∗n λ(·)[EL {ϕ
e? (· ; Pr )}] = Op (hq ) ∀ % ∈ {0, 1}.
Combining (26) and (27) using the linearity of G∗n (·), we then have (A.1).
Next, to show (A.2), let f (x; Pr ) = ϕ(0) (x; Pr ) and l(x; Pr ) = ϕ(1) (x; Pr ).
Then, we write
e (1) (·) − T
e (2) (·) − T
e (3) (·) + T
e (4) (·)}],
G∗n [λ(·){m
e ? (· ; Pr )}] = G∗n [λ(·){T
n,Pr
n,Pr
n,Pr
n,Pr
xiv
A. CHAKRABORTTY AND T. CAI
where
e
e (1) (x) = l? (x; Pr ) ,
T
n,Pr
f (x; Pr )
e
e
e (3) (x) = l? (x; Pr )f? (x; Pr ) , and
(28) T
n,Pr
fe(x; Pr )f (x; Pr )
(1)
e
e (2) (x) = f? (x; Pr )l(x; Pr ) ,
T
n,Pr
f (x; Pr )2
2
e
e (4) (x) = l(x; Pr )f? (x; Pr ) .
T
n,Pr
fe(x; Pr )f (x; Pr )2
(2)
Since λPr (x) ≡ λ(x)f (x; Pr )−1 and λPr (x) ≡ λ(x)l(x; Pr )f (x; Pr )−2 are
bounded a.s. [PX ] due to Assumption 4.1 (iii)-(iv) and the boundedness of
λ(·), using these as choices of ‘λ(·)’ in (A.1), we have:
(1)
e (1) (·)} = Op (b(1) ),
G∗n {λPr (·)e
l? (· ; Pr )} = G∗n {λ(·)T
n
n,Pr
(2)
e (2) (·)} = Op (b(1) ).
G∗n {λPr (·)fe? (· ; Pr )} = G∗n {λ(·)T
n
n,Pr
e (s) (x)k ≤ Op (a2 ) which follows
Further, for each s ∈ {3, 4}, supx∈X kT
n,2
n,Pr
from repeated use of (15) along with Assumption 4.1 (iii)-(iv). Consequently,
e (s) (·)} is bounded
with λ(·) bounded a.s. [PX ], for each s ∈ {3, 4}, G∗n {λ(·)T
n,Pr
1
by: Op (n 2 a2n,2 ). Combining all these results using the linearity of G∗n (·), we
1
(1)
1
e ? (· ; Pr )} = Op (bn + n 2 a2n,2 ) = Op (n 2 a2n,2 ), thus
finally obtain: G∗n {λ(·)m
leading to (A.2). The proof of the lemma is now complete.
VI.2. Proof of Lemma A.3. Throughout this proof, all additional notations introduced, if not explicitly defined, are understood to have been
adopted from the proof of Theorem 4.1 in Section V. Now, using (17),
0 c (2)
b r )−ϕ
b 0 −P0 )M
c (1)
b0
ϕ
b(%) (x; P
e(%) (x; Pr ) = trace{(P
n,%,x }+trace{(Pr −Pr )Mn,%,x },
r
r
c (1)
c (1,1) c (1,2)
and M
n,%,x = Mn,%,x − Mn,%,x , as defined in Section V. Thus,
n (1,1)
o
b
b(1,2) (·) + ζ
b(2) (·) ,
b r) − ϕ
G∗n [λ(·){ϕ
b(%) (· ; P
e(%) (· ; Pr )}] = G∗n ζ
(·)
−
ζ
n,%,λ
n,%,λ
n,%,λ
where ∀ (ω) ∈ {(1, 1), (1, 2), (2)}, % ∈ {0, 1}, and x ∈ X ,
n
o
b(ω) (x) = λ(x) trace (P
b 0 − P0 )M
c (ω)
(29)
ζ
n,%,x .
n,%,λ
r
r
Then, ∀ s ∈ {1, 2} and l ∈ {1, . . . , d}, each element of
Z
n X
n
X
(n,s)
−2
c (1,s)
Hl,% (i1 , i2 )
λ[l] (x)M
n,%,x Pn (dx) = n
i1 =1 i2 =1
is a V-statistic, where
(n,s)
Hl,% (i1 , i2 ) = h−(r+1) λ[l] (Xi1 )Yi%2 U(s) (i1 , i2 )∇K 0 {P0r (Xi1 − Xi2 )/h}
xv
SEMI-SUPERVISED LINEAR REGRESSION
with U(1) (i1 , i2 ) = Xi1 and U(2) (i1 , i2 ) = Xi2 . Hence, similar to the proof
of (26), using Lemma 8.4 of Newey and McFadden (1994) with X compact,
∇K(·) and λ(·) bounded, and Y % having finite 2nd moments, we have: for
each l ∈ {1, . . . , d}, s ∈ {1, 2} and % ∈ {0, 1},
1
h
n
oi
− 2 −(r+1)
c (1,s) − EL λ[l] (·)M
c (1,s)
=
O
n
G∗n λ[l] (·)M
h
.
p
n,%,(·)
n,%,(·)
max
b r − Pr ) = Op (αn ) that for each s and %,
It then follows from (P
h (1,s)
n (1,s) oi
1
b
b
(30)
G∗ ζ
(·) − EL ζ
(·) = Op αn n− 2 h−(r+1) .
n
n,%,λ
n,%,λ
1
c (1,s) }] is
Next, for any given l, s and %, each element of n 2 G∗n [EL {λ[l] (·)M
n,%,(·)
a centered sum of i.i.d. random variables which are bounded by:
n
o
c (1,s)
supx∈X kλ(x)k |EL (M
= O(1),
n,%,x )|
max
where the order follows from (20), (21) and the boundedness of λ(·). Hence,
b r −Pr ) =
similar to the proof of (27), using Hoeffding’s inequality and that (P
Op (αn ), we have: ∀ l ∈ {1, . . . , d}, s ∈ {1, 2} and % ∈ {0, 1},
(31)
G∗n
h
n
oi
c (1,s)
EL λ[l] (·)M
n,%,(·)
max
= Op (1) ⇒
G∗n
(1,s)
b
EL ζ n,%,λ (·)
= Op (αn ).
For any matrix A, let us denote by A[a,b] the (a, b)th element of A. Now,
b(2) (.)} in (29), note that kG∗ {ζ
b(2) (·)}k is bounded by:
to control G∗ {ζ
n
n,%,λ
n
1
n 2 supx∈X kλ(x)k
XZ
n,%,λ
c (2)
b 0 − P0 )[b,a] M
(P
n,%,x
r
r
a,b
1
2
≤ n rp
b r − Pr
{kλ(x)k kx − Xk} P
sup
x∈X ,X∈X
(32)
[a,b]
max
(Pn + PX )(dx)
b %∗
Z
n
1
b %∗ ,
≤ Op n 2 αn Z
n
b r − Pr ) = Op (αn ) and the boundedness
where the last step follows from (P
R (%)
%∗
bn = Z
b n (x) (Pn + PX )(dx) with
of X and λ(·), and Z
b (%) (x) = n−1
Z
n
0
n
X
|Yi% |
Pr (x − Xi )
∇K(wi,x ) − ∇K
.
hr+1
h
i=1
b r − Pr )0 (x − Xi )h−1 k ≤ Op (αn h−1 )
Now, kwi,x − P0r (x − Xi )h−1 k ≤ k(P
uniformly in (i, x), as noted while proving (18). Further, with L∗ , as defined
xvi
A. CHAKRABORTTY AND T. CAI
b r −Pr )0 (x−Xi )h−1 k ≤
in Assumption 4.1 (vii), let An denote the event: {k(P
∗
b
L ∀ x ∈ X , i = 1, .., n}. Then, with (Pr − Pr ) = Op (αn ), X compact and
1
αn h−1 = o(1) since n 2 αn2 h−2 = o(1) as assumed, it follows that P(An ) → 1.
Using these along with Assumption 4.1 (vii) and the function φ(.) defined
therein, we have: on An with P(An ) → 1,
0
n
X
b r − Pr )0 (x − Xi )
|Yi% | (P
Pr (x − Xi )
φ
nhr+1
h
h
i=1
0
n
%
X
|Y
|
P
(x
−
X
)
√
i
r
i
b r − Pr
≤ rp sup kx − Xk P
.
φ
nhr+2
h
max
x∈X ,X∈X
b (%) (x) ≤
Z
n
i=1
b %∗
e %∗
e %∗
Thus, Z
n ≤ Op αn Zn , where Zn =
e (%) (x)
Z
n
=n
−1
n
X
e (%) (x; Zi ),
Z
n
and
R
e (%)
Z
n (x)(Pn + PX )(dx),
e (%) (x; Z)
Z
n
i=1
|Y % |
= r+2 φ
h
P0r (x − X)
h
.
0
Let Z0 ≡ (Y 0 , X0 )0 ∼ PZ be generated independent of L, and define:
e (1) = n−1
U
n,%
n
X
e (%) (X0 ; Zi )},
EX0 {Z
n
e (2) = n−1
U
n,%
e (%) (Xi ; Z0 )},
EZ0 {Z
n
i=1
i=1
e (1,1) = E{Z
e (%) (X0 ; Z0 )},
U
n,%
n
Then, first note that:
Z
n
X
R
and
e (k) = E{Z
e (%) (X0 ; Z)k } for k = 1, 2.
V
n,%
n
e (%)
e (1)
Z
n (x)PX (dx) = Un,% . Further, since
e (%) (x)Pn (dx) = n−2
Z
n
n X
n
X
e (%) (Xi ; Zi )
Z
n
1
2
i1 =1 i2 =1
is a V-statistic, we have:
Z
e (%) (x)Pn (dx) = U
e (1) + U
e (2) − V
e (1) + Op {n−1 U
e (1,1) + n−1 (V
e (2) ) 21 }
Z
n
n,%
n,%
n,%
n,%
n,%
using Lemma 8.4 of Newey and McFadden (1994). Then, with all notations
as above, we have:
e (1,1) + n−1 (V
e (2) ) 21 ≤ Op n−1 h−(r+2) ,
(33)
n−1 U
and
n,%
n,%
SEMI-SUPERVISED LINEAR REGRESSION
xvii
Z
n
1 X %
w − P0r Xi
fPr (w)dw
φ
|Yi |
nhr+2
h
XPr
i=1
(
)
Z
n
BPr X
%
≤
|Yi |
φ(ψ i )dψ i ,
nh2
An
Xi
i=1
)
Z
(
n
X
BPr
%
−1
φ(ψ)dψ
n
|Yi | ≤ Op h−2 ,
≤ 2
h
Rr
e (1) =
U
n,%
(34)
i=1
where ψ i = h−1 (w − P0r Xi ) ∀ i, Anx = {ψ : (P0r x + hψ) ∈ XPr } ∀ x ∈ X ,
and BPr = supw∈XPr fPr (w) < ∞. The error rate in (33) follows since φ(·)
e (1)
is bounded and Y % has finite 2nd moments, while that of U
n,% follows from
Assumption 4.1 (iii), integrability of φ(·), and LLN applied to the sequence
e (2)
e (1)
{Yi% }ni=1 having finite 2nd moments. Now, note that U
n,% − Vn,% is a centered
0
n
e (%)
average of [EZ0 {Z
n (Xi ; Z )}]i=1 which are i.i.d. and bounded by:
0
Z
Pr x − w
1
(%)
(%)
e
φ
mPr (w)fPr (w)dw,
sup EZ {Zn (x; Z)} = sup r+2
h
x∈X
x∈X h
XPr
(%)
where mPr (w) = E(|Y |% | XPr = w) ∀ % ∈ {0, 1} and w ∈ XPr . Using the
integrability of φ(·), we then have:
0
(%) Z
CP r
Pr x − w
φ
dw
r+2
h
x∈X h
XPr
Z
(%) Z
CPr
φ(−ψ) dψ ≤ 2
φ(ψ)dψ = O h−2 ,
h
An
Rr
x
e (%) (x; Z)} ≤ sup
sup EZ {Z
n
x∈X
(%)
C
≤ sup P2r
x∈X h
(%)
(%)
where CPr = supw∈XPr mPr (w)fPr (w) < ∞ due to Assumption 4.1 (iii),
and Anx = {ψ : (P0r x + hψ) ∈ XPr }, as before. It then follows, similar to the
proof of (27), from a simple application of Hoeffding’s inequality that
e (2) − V
e (1) = Op n− 21 h−2 .
(35)
U
n,%
n,%
−2 + n−1 h−(r+2) ). Hence,
e %∗
Using (33)-(35), we finally have: Z
n = Op (h
Z
%∗
b
b (%) (x) (Pn + PX )(dx) ≤ Op αn Z
e %∗ = Op αn + αn
(36) Zn = Z
,
n
n
h2
nhr+2
!
1
1
n
o
n 2 αn2
n 2 αn2
∗ b(2)
(37) and
Gn ζ n,%,λ (·) ≤ Op
+
∀ % ∈ {0, 1},
h2
nhr+2
xviii
A. CHAKRABORTTY AND T. CAI
where the final bound in (37) follows from (32). The desired result in (A.3)
now follows by applying (30), (31) and (37) to (29) using the linearity of
G∗n (·). The proof of the lemma is now complete. (Note that conditions (i),
(iv) and (viii) in Assumption 4.1 were actually not used in this proof).
VI.3. Proof of Theorem 4.2. Finally, to establish the result of Theorem
−
4.2, let λ0 (x) = →
x which is measurable and bounded on X . Further, with
G∗n (·) as defined in Appendix A.1, note that Gn,K for K = 1 is given by:
b r ) − m(·
Gn,K = G∗n {λ0 (·)m
e ? (· ; Pr )} + G∗n [λ0 (·){m(·
b ;P
e ; Pr )}],
(38)
due to linearity of G∗n (·). Now, using Lemma A.2, we have:
1
G∗n {λ0 (·)m
e ? (· ; Pr )} = Op (n 2 a2n,2 ) = Op (a∗n,2 ).
(39)
b r ) − m(·
The second term G∗n [λ0 (·){m(·
b ;P
e ; Pr )}] in (38) can be written as:
(1)
(2)
(3)
(4)
b
b
b
b
G∗n [λ0 (·){T
n,Pr (·) − Tn,Pr (·) − Tn,Pr (·) + Tn,Pr (·)}]
1
1
2
∗
2
2
= Op b(2)
n + n an,1 + n an,1 an,2 = Op an,1 ,
(40)
where with slight abuse of notation,
b e
a−e
a
b (2) (x) = a(b − b) ,
b (1) (x) = b
,
T
T
n,Pr
n,Pr
b
b2
a−e
a)(eb − b) (b
a−e
a)(bb − eb)
b (3) (x) = (b
T
+
, and
n,Pr
eb bb
b eb
a(bb − eb)2
(e
a − a)(bb − eb) a(bb − eb)(eb − b)(b + eb)
b (4) (x) = e
T
+
,
−
n,Pr
bb b2
b2
(b eb)(b bb)
with (a, b) = {l(x; Pr ), f (x; Pr )}, (e
a, eb) = {e
l(x; Pr ), fe(x; Pr )} and (b
a, bb) =
b r ), fb(x; P
b r )}}.
{b
l(x; P
For (40), the starting expansion is due to a linearization similar to (28),
(1)
while the final rate is due to the following: note that λ0,Pr (·) ≡ b−1 λ0 (·) and
(2)
λ0,Pr (·) ≡ ab−2 λ0 (·) are both bounded a.s. [PX ] due to Assumption 4.1 (iii)(iv) and the boundedness of λ0 (·). Hence, using these as choices of ‘λ(·)’ in
(1)
b (1) (·)}] = Op (b(2)
Lemma A.3, we have: G∗n {(b
a−e
a)λ0,Pr (·)} = G∗n [λ0 (·){T
n )
n,Pr
(2)
(2)
(2)
∗
∗
b
e
b
and G {(b − b)λ
(·)} = G [λ0 (·){T
(·)}] = Op (bn ) respectively. Furn
0,Pr
n
n,Pr
b (s) (x)k ≤ Op (a2 + an,1 an,2 )
ther, note that for each s ∈ {3, 4}, supx∈X kT
n,1
n,Pr
which follows from repeated use of (15), (24) along with Assumption 4.1
SEMI-SUPERVISED LINEAR REGRESSION
xix
(iii)-(iv). Consequently, with λ0 (x) bounded a.s. [PX ], for each s ∈ {3, 4},
b (s) (·)}] is bounded by: Op (n 21 a2 + n 12 an,1 an,2 ). Combining all
G∗n [λ0 (·){T
n,1
n,Pr
these results using the linearity of G∗n (·) and noting that with a∗n,2 = o(1),
(2)
1
1
(bn + n 2 a2n,1 + n 2 an,1 an,2 ) = O(a∗n,1 ), (40) now follows and, along with (39)
and (38), implies: Gn,K = Op (a∗n,1 + a∗n,2 ) as claimed in Theorem 4.2. Lastly,
using this in (3.10), the expansion in (4.3) and its associated implications
follow. The proof of Theorem 4.2 is now complete.
References.
Andrews, D. W. K. (1995). Nonparametric Kernel Estimation for Semiparametric Models. Econometric Theory 11 560-586.
Belkin, M., Niyogi, P. and Sindhwani, V. (2006). Manifold Regularization : A Geometric Framework for Learning from Labeled and Unlabeled Examples. The Journal of
Machine Learning Research 7 2399-2434.
Castelli, V. and Cover, T. M. (1995). The Exponential Value of Labeled Samples.
Pattern Recognition Letters 16 105-111.
Castelli, V. and Cover, T. M. (1996). The Relative Value of Labeled and Unlabeled
Samples in Pattern Recognition with an Unknown Mixing Parameter. IEEE Transactions on Information Theory 42 2102-2117.
Chakrabortty, A. and Cai, T. (2017). Supplement to “Efficient and Adaptive Linear
Regression in Semi-Supervised Settings”. DOI: Completed by the typesetter.
Chapelle, O., Schölkopf, B. and Zien, A. (2006). Semi-Supervised Learning. MIT
Press, Cambridge, MA, USA.
Cook, R. D. (1998). Principal Hessian Directions Revisited (with Discussion). Journal of
the American Statistical Association 93 84-100.
Cook, R. D. and Lee, H. (1999). Dimension Reduction in Binary Response Regression.
Journal of the American Statistical Association 94 1187-1200.
Cook, R. D. and Weisberg, S. (1991). Discussion of “Sliced Inverse Regression” by
K.-C. Li. Journal of the American Statistical Association 86 328-332.
Cozman, F. G. and Cohen, I. (2001). Unlabeled Data Can Degrade Classification Performance of Generative Classifiers. Technical Report No. HPL-2001-234, HP Laboratories,
Palo Alto, CA, USA.
Cozman, F. G., Cohen, I. and Cirelo, M. C. (2003). Semi-Supervised Learning of
Mixture Models. In Proceedings of the Twentieth ICML 99-106.
Duan, N. and Li, K.-C. (1991). Sliced Regression: A Link-Free Regression Method. The
Annals of Statistics 19 505-530.
Hansen, B. E. (2008). Uniform Convergence Rates for Kernel Estimation with Dependent
Data. Econometric Theory 24 726-748.
Kawakita, M. and Kanamori, T. (2013). Semi-Supervised Learning with Density-Ratio
Estimation. Machine Learning 91 189-209.
Kohane, I. S. (2011). Using Electronic Health Records to Drive Discovery in Disease
Genomics. Nature Reviews Genetics 12 417-428.
Lafferty, J. D. and Wasserman, L. (2007). Statistical Analysis of Semi-Supervised
Regression. Advances in Neural Information Processing Systems 20 801-808.
Li, K.-C. (1991). Sliced Inverse Regression for Dimension Reduction. Journal of the American Statistical Association 86 316-327.
xx
A. CHAKRABORTTY AND T. CAI
Li, K.-C. (1992). On Principal Hessian Directions for Data Visualization and Dimension
Reduction: Another Application of Stein’s Lemma. Journal of the American Statistical
Association 87 1025-1039.
Liao, K. P., Cai, T., Gainer, V. et al. (2010). Electronic Medical Records for Discovery
Research in Rheumatoid Arthritis. Arthritis Care and Research 62 1120-1127.
Masry, E. (1996). Multivariate Local Polynomial Regression for Time Series: Uniform
Strong Consistency and Rates. Journal of Time Series Analysis 17 571-600.
Newey, W. K. (1994). Kernel Estimator of Partial Means and a Generalized Variance
Estimator. Econometric Theory 10 1-21.
Newey, W. K., Hsieh, F. and Robins, J. (1998). Undersmoothing and Bias Corrected
Functional Estimation. Technical Report No. 98-17, Dept. of Economics, MIT, USA.
Newey, W. K. and McFadden, D. (1994). Large Sample Estimation and Hypothesis
Testing. Handbook of Econometrics 4 2111–2245.
Nigam, K. P. (2001). Using Unlabeled Data to Improve Text Classification. PhD thesis,
Carnegie Mellon University, USA. CMU-CS-01-126.
Nigam, K., McCallum, A. K., Thrun, S. and Mitchell, T. (2000). Text Classification
from Labeled and Unlabeled Documents Using EM. Machine Learning 39 103-134.
Seeger, M. (2002). Learning with Labeled and Unlabeled Data. Technical Report No.
EPFL-REPORT-161327, University of Edinburgh, UK.
Sokolovska, N., Cappé, O. and Yvon, F. (2008). The Asymptotics of Semi-Supervised
Learning in Discriminative Probabilistic Models. In Proceedings of the Twenty Fifth
ICML 984-991.
Van der Vaart, A. W. (2000). Asymptotic Statistics 3. Cambridge University Press.
Zhang, T. and Oles, F. J. (2000). The Value of Unlabeled Data for Classification Problems. In Proceedings of the Seventeenth ICML 1191-1198.
Zhu, X. (2005). Semi-Supervised Learning through Graphs. PhD thesis, Carnegie Mellon
University, USA. CMU-LTI-05-192.
Zhu, X. (2008). Semi-Supervised Learning Literature Survey. Technical Report No. 1530,
Computer Sciences, University of Wisconsin-Madison, USA.
Zhu, L.-X. and Ng, K. W. (1995). Asymptotics of Sliced Inverse Regression. Statistica
Sinica 5 727-736.
Abhishek Chakrabortty
Department of Statistics
University of Pennsylvania
3730 Walnut Street
Jon M. Huntsman Hall, 4th Floor
Philadelphia, PA 19104, USA.
E-mail: [email protected]
Tianxi Cai
Department of Biostatistics
Harvard University
655 Huntington Avenue
Building 2, 4th Floor
Boston, MA 02115, USA.
E-mail: [email protected]
| 10math.ST
|
arXiv:1702.02926v2 [cs.FL] 1 Sep 2017
THE WORD PROBLEM OF Zn IS A MULTIPLE
CONTEXT-FREE LANGUAGE
MENG-CHE “TURBO” HO
Abstract. The word problem of a group G = hΣi can be defined as the set of
formal words in Σ∗ that represent the identity in G. When viewed as formal
languages, this gives a strong connection between classes of groups and classes
of formal languages. For example, Anisimov showed that a group is finite if and
only if its word problem is a regular language, and Muller and Schupp showed
that a group is virtually-free if and only if its word problem is a context-free
language. Above this, not much was known, until Salvati showed recently that
the word problem of Z2 is a multiple context-free language, giving first such
example. We generalize Salvati’s result to show that the word problem of Zn
is a multiple context-free language for any n.
1. Introduction
Fixing an alphabet set Σ, the set of finite strings in Σ is denoted by Σ∗ =
{a1 a2 · · · an | n ∈ N, ai ∈ Σ}. Any subset of Σ∗ is called a formal language.
As an attempt to analyze natural languages, Chomsky [Cho56] introduced the
Chomsky-Schützenberger hierarchy, which includes regular, context-free, contextsensitive, and recursively enumerable languages, each one a larger class than the
former. In this paper, we will be focusing on multiple context-free languages, another
generalization of context-free languages introduced by Seki et al [SMFK91].
There are many ways to associate formal languages to a group. Indeed, for a
group G generated by a finite set Σ = Σ−1 , we define its word problem W (G) to
be the subset of Σ∗ consisting of all strings that represent the identity element of
G. Note that the language (and even the set of alphabets) depends on the choice
of generating set. However, we will see that this dependence is negligible for the
languages we are interested in here.
There had been various work associating classes of languages and classes of
groups via word problems. Anı̄sı̄mov [Anı̄71] showed that G is finite if and only if
W (G) is regular. Muller and Schupp showed that G is virtually free if and only if
W (G) is context-free. Indeed, they showed the following stronger theorem:
Theorem 1 (Muller, Schupp [MS83]). For a finitely generated group G, the following are equivalent:
(1) G is virtually free.
(2) W (G) is context-free.
Date: September 5, 2017.
The material is based upon work supported by the National Science Foundation under Grant
No. DMS-1440140 while the author was in residence at the Mathematical Sciences Research
Institute in Berkeley, California, during the Fall 2016 semester.
1
2
MENG-CHE “TURBO” HO
(3) There exists a constant K such that every closed path in the Cayley graph
of G can be triangulated by chords with length ≤ K.
Recently, Salvati [Sal15] showed that the word problem of Z2 is a 2-multiple
context-free language, giving first such example. In the proof, he used some geometric methods that seems to be specific to the two-dimensional case. Naderhof
[Ned16] gave a shorter proof of Salvati’s result, avoiding some of the pieces that are
two-dimensional specific. In this paper, by using a topological lemma from Burago
[Bur92], we also avoid the geometric pieces and are able to generalize Silvati’s theorem and show that the word problem of Zn is multiple context-free:
Theorem 2. The word problem W (Zn ) is an (8 n+1
− 2)-multiple context-free
2
language for every n.
However, this yields that W (Z2 ) is a 6-multiple context-free language, which is
slightly weaker than Salvati’s result which says that W (Z2 ) is a 2-multiple contextfree language. Thus, one natural question will be improving our result to get the
following conjecture proposed by Nederhof:
Conjecture 3 (Nederhof [Ned17]). For any n, the language W (Zn ) is n-multiple
context-free.
In the same paper, Nederhof also showed that W (Zn ) is not a K-multiple contextfree language for any K < n, so this conjecture is indeed sharp.
To end the introduction, we propose the following open problem that may be
interesting to group theorists. For more recent progress on this topic, the reader
may also want to see [GKS] and [KS].
Question 4. For which finitely-generated groups G is W (G) multiple context-free?
This paper is organized as follows: We give the necessary definition and background in section 2. We then set up our notation and introduce the main lemma
in section 3. Finally we show the main Theorem in section 4.
2. Definition and background
We follow [SMFK91] and [Ned17] to give the definition of multiple context-free
grammars:
Definition 5. A K-multiple context-free grammar is a tuple (Σ, N, S, P ), where Σ
is the set of terminals, N is the set of non-terminals, S ∈ N is the start symbol,
and P is the set of production rules, such that:
• Each non-terminal I(x1 , x2 , . . . , xm ) has an associated natural number m ≤
K, its arity.
• The arity of S is 1.
• Each of the production rule is of the following form for some p ≥ 0:
p
2
1
)
), · · · , Ip (y1p , . . . , ym
), I2 (y12 , . . . , ym
I0 (x1 , x2 , . . . , xm0 ) ← I1 (y11 , y21 , . . . , ym
p
2
1
where Ii has arity mi , each xi is a word in the variables yjl and the terminals
Σ, and each of yjl shows up at most once in all of x1 , x2 , . . . , xn .
An instance of a production rule or non-terminal is obtained by replacing the
occurance(s) of each variable by a string in Σ∗ . A sequence of rule instances is valid
if every non-terminal instances on the right hand side of some rule instances in the
THE WORD PROBLEM OF Zn IS A MULTIPLE CONTEXT-FREE LANGUAGE
3
sequence already appears on the left hand side of some rule instances earlier in the
sequence. We say a sequence ends in a non-terminal instance if it is the left hand
side of the last rule instance in the sequence.
We say a grammar admits a string s in Σ∗ if there is a valid sequence of rule
instances, called a derivation, ending in S(s). We say a grammar produces the
language which is the collection of all strings in Σ∗ that it admits.
Definition 6. A language is called K-multiple context-free if it can be produced by
some K-multiple context-free grammar. A language is called multiple context-free
if it is K-multiple context-free for some K.
Example 7. Consider the 2-multiple context-free grammar with terminals a, b, c, d,
a binary non-terminal I = I(x, y), the start symbol S = S(x), and production rules
I(ǫ, ǫ) ←
I(axb, cyd) ← I(x, y)
S(xy) ← I(x, y).
Then the following is a derivation for an bn cn dn :
I(ǫ, ǫ) ←
I(ab, cd) ← I(ǫ, ǫ)
I(a2 b2 , c2 d2 ) ← I(ab, cd)
..
.
I(an bn , cn dn ) ← I(an−1 bn−1 , cn−1 dn−1 )
S(an bn cn dn ) ← I(an bn , cn dn )
Indeed, this is the only kind of strings this grammar admits, so the language
produced by this grammar is {an bn cn dn | n ≥ 0}.
The class of multiple context-free languages is a cone [SMFK91] in the sense of
[GKS]. A well-known consequence of being a cone is the following theorem from
[GKS], which we only state in the case of multiple context-free languages:
Theorem 8. Let W (G) be the word problem of G with respect to a fixed generating
set Σ. Suppose W (G) is multiple context-free. Then
(1) The word problem of G is multiple context-free with respect to any generating set of G.
(2) The word problem of any finitely-generated subgroup of G is multiple contextfree.
(3) The word problem of any supergroup of G of finite index is multiple contextfree.
4
MENG-CHE “TURBO” HO
3. Notations and main lemma
For any curve α in Rn , we will write α to be the displacement vector of α, and
|α| to be the length of α. Note that when α is the curve on the Cayley graph of
Zn corresponding to a word w, α is just the image of the word w in Zn and |α| is
just the length of the word. We also write αβ to mean the concatenation of the
two curves α and β.
We will need the following lemma by Burago to prove the main theorem. We
include the proof here for completeness.
Lemma 9 (Burago [Bur92]). Let α(t) : [0, 1] → Rn be a continuous curve. Then
there exist [t1 , s1 ], [t2 , s2 ], . . . , [tk , sk ] such that 0 ≤ t1 ≤ s1 ≤ t2 ≤ s2 ≤ · · · ≤ tk ≤
sk ≤ 1 with k = [ n+1
2 ] and
k
X
(α(si ) − α(ti )) =
i=1
1
(α(1) − α(0)).
2
Proof. We will construct an odd map f : S n → Rn . For x = (x1 , x2 , . . . , xn+1 ) ∈
S n ⊂ Rn , we write y0 = 0, y1 = x21 , y2 = x21 + x22 , . . . , yn+1 = x21 + · · · + x2n+1 = 1.
We define
n+1
X
f (x) =
sign(xi )(α(yi ) − α(yi−1 )).
i=1
One checks that f is indeed an odd map, and since every odd map from S n → Rn
has a zero, we can pick some x ∈ S n such that f (x) = 0. Then one of {[yi−1 , yi ] |
xi > 0} and {[yi−1 , yi ] | xi < 0} satisfies the condition, by possibly attaching some
empty intervals in the end.
Remark 10. One sees that k = [ n+1
2 ] is sharp by considering the curve along the
edges of an n-cube.
4. Proof of the main Theorem
We restate and prove the main Theorem. To make the proof clearer, we shall
suppress the embedding when talking about curves.
Theorem 2. The word problem W (Zn ) is an (8 n+1
− 2)-multiple context-free
2
language for every n.
Proof. We will denote the standard generators of Zn by a1 , a2 , . . . , an , and their
inverses by A
1 , A2 , .. . , An .
n+1
and m = 8k − 2. We consider the m-multiple context-free
Let k =
2
grammar with terminals Σ = {a1 , A1 , a2 , A2 , . . . , An }, an m-ary non-terminals I =
I(x1 , x2 , . . . , xm ), start symbol S = S(x), and production rules
S(x1 x2 · · · xm ) ← I(x1 , x2 , . . . , xm )
I(ǫ, ǫ, . . . , ǫ) ←
I(a1 , A1 , ǫ, ǫ, . . . , ǫ) ←
I(a2 , A2 , ǫ, ǫ, . . . , ǫ) ←
..
.
I(an , An , ǫ, ǫ, . . . , ǫ) ←
THE WORD PROBLEM OF Zn IS A MULTIPLE CONTEXT-FREE LANGUAGE
5
I(z1 , z2 , . . . , zm ) ← I(x1 , x2 , . . . , xm ), I(xm+1 , xm+2 , . . . , x2m )
for every σ ∈ Sym(2m) and z1 z2 · · · zm = xσ(1) xσ(2) · · · xσ(2m) .
By the very first production rule, it suffices to show that x1 x2 · · · xm ∈ W (Zn )
if and only if there is a valid sequence ending in I(x1 , x2 , . . . , xm ). The if direction follows immediately from the definition of the production rules and commutativity of Zn . To show the only if direction, we will induct on the length
of x1 x2 · · · xm ∈ W (Zn ). The base cases where |x1 · · · xm | ≤ m are straight forward. Now suppose |x1 · · · xm | > m. We may assume that x1 x2 · · · xm/2 6= 0
and xm/2+1 xm/2+2 · · · xm 6= 0, otherwise we may apply induction hypothesis on
x1 x2 · · · xm/2 = 0 and xm/2+1 xm/2+2 · · · xm = 0.
Now for every i, we write vi to be the curve in Rn that represents xi , which is
contained in the Cayley graph of Zn . Applying Lemma 9 to the curve v1 · · · vm/2 ,
we get 2k points on the curve, breaking the curve into 2k + 1 pieces u0 , u1 , . . . , u2k ,
with
(1)
i=k
X
u2i−1
i=1
i=2k
i=k
X
1 X
=
ui =
u2i .
2 i=0
i=0
Applying Lemma 9 to the curve vm/2+1 · · · vm in the same way, we get
(2)
i=k
X
u′2i−1 =
i=1
i=k
i=2k
1 X ′ X ′
ui =
u2i .
2 i=0
i=0
Claim. We may choose the endpoints of all the ui ’s and u′i ’s to be either on lattice
points, or mid-lattice points, i.e. the midpoints of two adjacent lattice points.
Proof. Fix ui ’s and u′i ’s that has the fewest number of endpoints not being a lattice
point or a mid-lattice point. Suppose the claim is false, then there is some endpoint
x that is not a lattice point or a mid-lattice point. Without loss, say this is an
endpoint of ui , and its first coordinate is not a multiple of 21 . This forces all its
other coordinates to be integral, as u1 u2 · · · um is a curve contained in the lattice
Zn .
But each coordinate of the middle term of (1) is a multiple of 12 , thus there must
be another endpoint y of some uj (possibly with i = j) whose first coordinate is
also not a multiple of 12 . Now depending on how x and y appear in (1), either we
may replace x by x + δe1 and y by y + δe1 and still keep (1), or we may replace x
by x + δe1 and y by y − δe1 and still keep (1). Say we are in the first case, then
choose the smallest positive δ that makes x + δe1 or y + δe1 a multiple of 21 , and
replace x and y correspondingly. This gives another choice of ui ’s and u′i ’s which
has at least one fewer endpoint not being a lattice point or a mid-lattice point, a
contradiction.
Now, the 2k points together with the endpoints of vi ’s, which are lattices points
′
by definition, give a refinement w1 , w2 , . . . , wm/2+2k and w1′ , w2′ , . . . , wm/2+2k
of the
′
ui ’s and ui ’s. Notice that some of these may be empty curves if some of these points
coincide. Now rewriting (1) and (2) in terms of wi ’s and wi′ ’s, we get
X
i∈S
wi =
i=m/2
X
1 X
xi =
wi
2 i=1
i∈S
6
MENG-CHE “TURBO” HO
X
wi′ =
i∈T
1
2
i=m
X
xi =
i=m/2+1
X
wi′
i∈T
for some S, T ⊂ {1, 2, . . . , m/2 + 2k}, and S = {1, 2, . . . , m/2 + 2k} \ S, T =
{1, 2, . . . , m/2 + 2k} \ T . Notice that since wi ’s and wi′ ’s are refinements of ui ’s and
u′i ’s, so we have |S|, |T |, |S|, |T | ≥ k. Without loss, we will assume that |S| ≥ |S|,
and |T | ≤ |T |. Thus |S|, |T | ≤ m/2 + 2k − k = 5k − 1, and |S|, |T | ≤ (m/2 +
2k)/2 = 3k − 1/2, so |S|, |T | ≤ 3k − 1. Hence |S| + |T | ≤ 8k − 2 = m, and
|S| + |T | ≤ 8k − 2 = m.
i=m
P
Now observe that
xi = 0 by the assumption that x1 x2 · · · xm ∈ W (Zn ). Thus
we have
(3)
i=1
X
i∈S
wi +
X
i∈T
wi′ =
X
i∈S
wi +
X
wi′ = 0.
i∈T
Claim. There is a partition {wi | 1 ≤ i ≤ m/2 + 2k} of v1 v2 · · · vm/2 and {wi′ | 1 ≤
i ≤ m/2 + 2k} of vm/2+1 · · · vm which is a refinement of xi ’s such that:
(a) This partition satisfies (3).
(b) Not all curves in the first part of (3) are empty. Also, not all curves in the
second part of (3) are empty.
(c) All endpoints are lattice points.
Proof. The previous construction gives a partition satisfying (a) with every endpoint being either a lattice point or a mid-lattice point. It also satisfies (b) by (1)
and (2) and the assumption that x1 x2 · · · xm/2 6= 0.Thus, we may choose a partition
satisfying (a) and (b) whose endpoints are either a lattice point or a mid-lattice
point, with fewest number of endpoints being a mid-lattice point. Supposing the
claim is false, let x be an endpoint that is a mid-lattice point. If both of the curves
adjacent to it are in the same sum, we may simply replace x to one of the closest
lattice point, which reduces the number of endpoints being a mid-lattice point in
our partition while still satisfying (a) and (b), a contradiction.
Now without loss, we may assume x is adjacent to w1 and w2 , with 1 ∈ S and
2∈
/ S, and assume x is not integral in the first coordinate, with all other coordinates
being integral. Note that in this case x contributes positively to the first part of
(3), and negatively to the second part of (3). As the sums equal to 0, there must
be another endpoint y which is also not integral in the first coordinate. By the
argument in the previous paragraph, the two curves adjacent to y cannot be in the
same sum. Thus, similar to the case of x, y must contribute to either the first or
second part of (3) positively, and the other part negatively.
We now replace x with x + 21 e1 and y with y ± 12 e1 , depending on whether y
contributes positively or negatively to the second part of (3). This new partition
still satisfies (a). If this replacement breaks (b), say all curves in the first part of
(3) are empty, then this replacement must have reduced the total length of curves
in the first part of (3) by 1. So, we may instead replace x with x − 21 e1 and y with
y ∓ 21 e1 , and increase the the total length of curves in the first part of (3) by 1, while
reducing the total length of the second part by 1. However, as the total length of
all curves are > m, this second replacement will satisfy (b).
Either case, this will replace the mid-lattice points x and y by lattice points,
thus reducing the number of mid-lattice points of the partition. So we have a new
THE WORD PROBLEM OF Zn IS A MULTIPLE CONTEXT-FREE LANGUAGE
7
partition satisfying (a) and (b) but has fewer endpoints being a mid-lattice point,
a contradiction.
Now we finally get a partition on which (3) holds and all its endpoints are lattice
points. Thus, each part of the partition actually represent a word in Zn . We shall
write y1 , y2 , . . . , ym to be the words given by {wi | i ∈ S} and {wi′ | i ∈ T }, and
z1 , z2 , . . . , zm to be the words given by {wi | i ∈
/ S} and {wi′ | i ∈
/ T }.PWe let
m
yP
j = ǫ for j > |T | + |S|, and zk = ǫ for k > |T | + |S|. Then, (3) says that
0 yi =
m
0 zi = 0 and condition (b) ensures that |y1 · · · ym |, |z1 · · · zm | < |x1 · · · xm |. Thus
by induction hypothesis we have two valid sequences ending in I(y1 , y2 , . . . , ym )
and I(z1 , z2 , . . . , zm ), respectively. Since yi ’s and zi ’s is a refinement of xi ’s, there
is also an instance of one of the last production rules that combines the yi ’s and
zi ’s back to xi ’s. Concatenating the two sequences with this production rule will
give a valid sequence ending in I(x1 , x2 , . . . , xm ). This completes the proof.
We finish with the following corollary by combining our main Theorem with
Theorem 8 to get a more general statement.
Corollary 11. The word problem of any finitely-generated virtually-abelian group
is multiple context-free.
References
[Anı̄71]
A. V. Anı̄sı̄mov, The group languages, Kibernetika (Kiev) (1971), no. 4, 18–24.
MR 0301981
[Bur92]
D. Yu. Burago, Periodic metrics, Representation theory and dynamical systems,
Adv. Soviet Math., vol. 9, Amer. Math. Soc., Providence, RI, 1992, pp. 205–210.
MR 1166203
[Cho56]
Noam Chomsky, Three models for the description of language, IRE Transactions on
information theory 2 (1956), no. 3, 113–124.
[GKS]
Robert H. Gilman, Robert P. Kropholler, and Saul Schleimer, Groups whose word
problems are not semilinear, in preperation.
[KS]
Robert P. Kropholler and Davide Spiriano, The class of groups with mcf word problem
is closed under free products, in preparation.
[MS83]
David E. Muller and Paul E. Schupp, Groups, the theory of ends, and context-free
languages, J. Comput. System Sci. 26 (1983), no. 3, 295–310. MR 710250
[Ned16]
M.-J. Nederhof, A short proof that o2 is a mcfl, Proceedings of the 54th Annual Meeting of the Annual Meeting of the Association for Computational Linguistics (2016),
no. 1117-1126.
, Free word order and mcfls, From Semantics to Dialectometry: Festschrift for
[Ned17]
John Nerbonne (2017), 273–282.
[Sal15]
Sylvain Salvati, MIX is a 2-MCFL and the word problem in Z2 is captured by the
IO and the OI hierarchies, J. Comput. System Sci. 81 (2015), no. 7, 1252–1277.
MR 3354791
[SMFK91] Hiroyuki Seki, Takashi Matsumura, Mamoru Fujii, and Tadao Kasami, On multiple
context-free grammars, Theoretical Computer Science 88 (1991), no. 2, 191–229.
| 4math.GR
|
Worst-case Optimal Submodular Extensions for Marginal Estimation
arXiv:1801.06490v1 [cs.LG] 10 Jan 2018
Pankaj Pansari1,2
1: University of Oxford
[email protected]
Chris Russell 2,3
2: The Alan Turing Institute
[email protected]
Abstract
Submodular extensions of an energy function
can be used to efficiently compute approximate marginals via variational inference. The
accuracy of the marginals depends crucially on
the quality of the submodular extension. To
identify the best possible extension, we show
an equivalence between the submodular extensions of the energy and the objective functions
of linear programming (LP) relaxations for
the corresponding MAP estimation problem.
This allows us to (i) establish the worst-case
optimality of the submodular extension for
Potts model used in the literature; (ii) identify
the worst-case optimal submodular extension
for the more general class of metric labeling;
and (iii) efficiently compute the marginals for
the widely used dense CRF model with the
help of a recently proposed Gaussian filtering
method. Using synthetic and real data, we
show that our approach provides comparable
upper bounds on the log-partition function to
those obtained using tree-reweighted message
passing (TRW) in cases where the latter is
computationally feasible. Importantly, unlike
TRW, our approach provides the first practical algorithm to compute an upper bound on
the dense CRF model.
1
Introduction
The desirable optimization properties of submodular
set functions have been widely exploited in the design
of approximate MAP estimation algorithms for discrete
conditional random fields (CRFs) [Boykov et al., 2001;
Kumar et al., 2011]. Submodularity has also been
recently used to design an elegant variational inference
algorithm to compute the marginals of a discrete CRF
Accepted at AISTATS 2018.
M. Pawan Kumar1,2
3: University of Surrey
[email protected]
by minimising an upper-bound on the log-partition
function. In the initial work of [Djolonga and Krause,
2014], the energy of the CRF was restricted to be
submodular. In a later work [Zhang et al., 2015], the
algorithm was extended to handle more general Potts
energy functions. The key idea here is to define a
large ground set such that its subsets represent valid
labelings, sublabelings or even incorrect labelings (these
may assign two separate labels to a random variable
and hence be invalid). Given the large ground set, it
is possible to define a submodular set function whose
value is equal to the energy of the CRF for subsets that
specify a valid labeling of the model. We refer to such
a set function as a submodular extension of the energy.
For a given energy function, there exists a large number
of possible submodular extensions. The accuracy of the
variational inference algorithm depends crucially on
the choice of the submodular extension. Yet, previous
work has largely ignored the question of identifying
the best extension. Indeed, the difficulty of identifying submodular extensions of general energy functions
could be a major reason why the experiments of [Zhang
et al., 2015] were restricted to the special case of models
specified by the Potts energy functions.
In this work, we establish a hitherto unknown connection between the submodular extension of the Potts
model proposed by Zhang et al. [2015], and the objective function of an accurate linear programming (LP)
relaxation of the corresponding MAP estimation problem [Kleinberg and Tardos, 2002]. This connection
has three important practical consequences. First, it
establishes the accuracy of the submodular extension
of the Potts model, via the UGC-hardness worst-case
optimality of the LP relaxation. Second, it provides
an accurate submodular extension of the hierarchical
Potts model, via the LP relaxation of the corresponding MAP estimation problem proposed by Kleinberg
and Tardos [2002]. Since any metric can be accurately
approximated as a mixture of hierarchical Potts models [Bartal, 1996, 1998], this result also provides a
computationally feasible algorithm for estimating the
marginals for metric labeling. Third, it establishes the
Worst-case Optimal Submodular Extensions for Marginal Estimation
equivalence between the subgradient of the LP relaxation and the conditional gradient of the problem of
minimising the upper bound of the log-partition. This
allows us to employ the widely used dense CRF, since
the subgradient of its LP relaxation can be efficiently
computed using a recently proposed modified Gaussian filtering algorithm [Ajanthan et al., 2017]. As a
consequence, we provide the first efficient algorithm to
compute an upper bound of the log-partition function
of dense CRFs. This provides complementary information to the popular mean-field inference algorithm
for dense CRFs, which computes a lower bound on
the log-partition [Koltun and Krahenbuhl, 2011]. We
show that the quality of our solution is comparable to
tree reweighted message passing (TRW) [Wainwright
et al., 2005] for the case of sparse CRFs. Unlike our
approach, TRW is computationally infeasible for dense
CRFs, thereby limiting its use in practice. Using dense
CRF models, we perform stereo matching on standard
data sets and obtain better results than [Koltun and
Krahenbuhl, 2011]. The complete code is available at
https://github.com/pankajpansari/denseCRF.
2
Preliminaries
We now introduce the notation and definitions that we
will make use of in the remainder of the paper.
Submodular Functions Given a ground set U =
{1, . . . , N }, denote by 2U its power set. A set function
F : 2U → R is submodular if, for all subsets A, B ⊆ U ,
we have
F (A ∪ B) + F (A ∩ B) ≤ F (A) + F (B).
(1)
The set function FPis modular if there exists s ∈ RN
such that F (A) = k∈A sk ∀ A ⊆ 2U . P
Henceforth, we
will use the shorthand s(A) to denote k∈A sk .
Extended Polymatroid Associated with any submodular function F is a special polytope known as the
extended polymatroid defined as
EP (F ) = {s ∈ RN | ∀A ⊆ U : s(A) ≤ F (A)},
(2)
where s denotes the modular function s(.) considered
as a vector.
Lovasz Extension For a given set function F with
F (∅) = 0, the value of its Lovasz extension f (w) :
RN → R is defined as follows: order the components of
w in decreasing order such that wj1 ≥ wj2 ≥ · · · ≥ wjN ,
where (j1 , j2 , . . . , jN ) is the corresponding permutation
of the indices. Then,
f (w) =
N
X
The function f is an extension because it equals F on
the vertices of the unit cube. That is, for any A ⊆ V ,
f (1A ) = F (A) where 1A is the 0-1 indicator vector
corresponding to the elements of A.
Property 1. By Edmond’s greedy algorithm [Edmonds, 1970], if w ≥ 0 (non-negative elements),
f (w) =
wjk (F (j1 , . . . , jk ) − F (j1 , . . . , jk−1 )) .
(3)
(4)
Property 1 implies that an LP over EP (F ) can be
solved by computing the value of the Lovasz extension
using equation (3).
Property 2. The Lovasz extension f of a submodular
function F is a convex piecewise linear function.
Property 2 holds since f (w) is the pointwise maximum
of linear functions according to equation (4).
CRF and Energy Functions A CRF is defined as a
graph on a set of random variables X = {X1 , . . . , XN }
related by a set of edges N . We wish to assign
every variable Xa one of the labels from the set
L = {1, 2, . . . , L}. The quality of a labeling x is given
by an energy function defined as
X
X
E(x) =
φa (xa ) +
φab (xa , xb ),
(5)
a∈X
(a,b)∈N
where φa and φab are the unary and pairwise potentials
respectively. In computer vision, we often think of X as
arranged on a grid. A sparse CRF has N defined by 4connected or 8-connected neighbourhood relationships.
In a dense CRF, on the other hand, every variable is
connected to every other variable.
The energy function also defines a probability distribution P (x) as follows:
(
1
exp(−E(x))
if x ∈ LN ,
P (x) = Z
(6)
0
otherwise.
P
The normalization factor Z = x∈LN exp(−E(x)) is
known as the partition function.
Inference There are two types of inference problems
in CRFs:
(i) Marginal inference: We want to compute the
marginal probabilities P (Xa = i) for every a =
1, 2, . . . , N and i = 1, 2, . . . , L.
(ii) MAP inference: We want to find a labeling with the
minimum energy, that is, minx∈LN E(x). Equivalently,
MAP inference finds the mode of P (x).
3
k=1
max hw, si.
s∈EP (F )
Review: Variational Inference Using
Submodular Extensions
We now summarise the marginal inference method of
Zhang et al. [2015]. To do this, we need to first define
Pankaj Pansari1,2 , Chris Russell
2,3
, M. Pawan Kumar1,2
problem as
s∈EP (F )
Figure 1: Illustration of 1-of-L encoding used in [Zhang
et al., 2015] with 4 variables and 3 labels. The blue labeling,
corresponding to X1 = 1, X2 = 3, X3 = 2, X4 = 1, is
valid. The yellow labeling, corresponding to X2 = 2, X3 =
1, 3, X4 = 3, is invalid since X3 has been assigned multiple
labels and X1 has been assigned none.
submodular extensions.
Submodular Extensions A submodular extension
is defined using a ground set such that some of its
subsets correspond to valid CRF labelings. To such an
extension, we need an encoding scheme which gives the
sets corresponding to valid CRF labelings.
One example of an encoding scheme is the 1-of-L encoding, illustrated in figure 1. Let each variable Xa
take one of L possible labels. In this scheme, we represent the set of possible assignments for Xa by the
set Va = {va1 , va2 , . . . , vaL }. If Xa is assigned label i,
then we select the element vai . Extending to all variables, our ground set becomes V = ∪N
a=1 Va . A valid
assignment A ⊆ V assigns each variable exactly one
label, that is, |A ∩ Va | = 1 for all Va . We denote the
set of valid assignments by M where M = ∩N
a=1 Ma
and Ma = {A : |A ∩ Va | = 1}.
Using our ground set V , we can define a submodular
function F which equals E(x) for all sets corresponding
to valid labelings, that is, F (Ax ) = E(x), x ∈ LN
where Ax is the set encoding of x. We call such a
function F a submodular extension of E(x).
Upper-Bound on Log-Partition Using a submodular extension F and given any s ∈ EP (F ), we can
obtain an upper-bound on the partition function as
Z=
X
A∈M
exp(−F (A)) ≤
X
exp(−s(A)),
(7)
A∈M
where M is the set of valid labelings. The upperbound is the partition function of the distribution
Q(A) ∝ exp(−s(A)), which factorises fully because
s(.) is modular. Since s ∈ EP (F ) is a free parameter,
we can obtain good approximate marginals of the distribution P (·) by minimising the upper-bound. After
taking logs, we can equivalently write our optimisation
X
g(s), where g(s) = log
min
exp(−s(A)). (8)
A∈M
Conditional Gradient Algorithm The conditional gradient algorithm (algorithm 1) [Frank and
Wolfe, 1956] is a good candidate for solving problem
(8) due to two reasons. First, problem (8) is convex.
Second, as solving an LP over EP(F) is computationally
tractable (property 1), the conditional gradient can be
found efficiently. The algorithm starts with an initial
solution s0 (line 1). At each iteration, we compute the
conditional gradient s∗ (line 3), which minimises the
linear approximation g(sk ) + ∇g(sk )T (s − sk ) of the objective function. Finally, s is updated by either (i) fixed
step size schedule, as in line 7 of algorithm 1, or (ii) by
doing line search sk+1 = min0≤γ≤1 g(γs∗ + (1 − γ)sk ).
Algorithm 1 Upper Bound Minimisation using Conditional Gradient Descent
1: Initialize s = s0 ∈ EP (F )
2: for k = 1 to MAX_ITER do
3:
s∗ = argmins∈EP (F ) h∇g(sk ), si
4:
if hs∗ − sk , ∇g(sk )i ≤ then
5:
break
6:
end if
7:
sk+1 = sk + γ(s∗ − sk ) with γ = 2/(k + 2)
8: end for
9: return s
4
Worst-case Optimal Submodular
Extensions via LP Relaxations
Worst-case Optimal Submodular Extensions
Different choices of extensions F change the domain in
problem (8), leading to different upper bounds on the
log-partition function. How does one come up with an
extension which yields the tightest bound?
In this paper, we focus on submodular extension families F(.) which for each instance of the energy function
E(.) belonging to a given class E gives a corresponding
submodular extension F(E). We find the extension
family Fopt that is worst-case optimal. This implies
that there does not exist another submodular extension
family F that gives a tighter upper bound for problem
(8) than Fopt for all instances of the energy function in
E. Formally,
@F :
min
s∈EP (F (E))
g(s) ≤
min
g(s) ∀ E(.) ∈ E.
s∈EP (Fopt (E))
(9)
Note that our problem is different from taking a given
energy model and obtaining a submodular extension
which is optimal for that model. Also, we seek a closedform analytical expression for F. For the sake of clarity,
Worst-case Optimal Submodular Extensions for Marginal Estimation
Tightest LP Relaxation Before describing our set
encoding and submodular extension, we briefly outline
the LP relaxation of the corresponding MAP estimation
problem. To this end, we define indicator variables yai
which equal 1 if Xa = i, and 0 otherwise. The following
LP relaxation is the tightest known for Potts model in
the worst-case, assuming the Unique Games Conjecture
to be true [Manokaran et al., 2008]
XX
(P-LP) min E(y) =
φa (i)yai +
y
a
i
X X wab
· |yai − ybi |
2
i
(a,b)∈N
s.t y ∈ ∆.
Figure 2: An illustration of the worst-case optimal submodular extension for Potts model for a chain graph of
4 variables, each of which can take 3 labels. The figure
shows the way to compute the extension values of the set
A = {v1,1 , v4,1 , v3,2 }.
in the analysis that follows we use F to represent F(E)
where the meaning is clear from context. The two
classes of energy functions we consider in this paper
are Potts and hierarchical Potts families.
Using LP Relaxations If we introduce a temperature parameter in P (x) (equation (6)) by using E(x)/T
and decrease T , the resulting distribution starts to
peak more sharply around its mode. As T → 0,
marginal estimation becomes the same as MAP inference since the resulting distribution P 0 (x) has mass
1 at its mode x∗ and is 0 everywhere else. Given the
MAP solution x∗ , one can compute the marginals as
P (Xi = j) = [x∗i = j], where [.] is the Iverson bracket.
(11)
The set ∆ is the union of N probability simplices:
∆ = {ya ∈ RL |ya ≥ 0 and h1, ya i = 1},
(12)
where y is the vector of all variables and ya is the
component of y corresponding to Xa .
Set Encoding We choose to use the 1-of-L encoding
for Potts model as described in section 3. With the
encoding scheme for Potts model above, g(s) can be
factorised and problem (8) can be rewritten as:
min
s∈EP (F )
N
X
log
a=1
L
X
exp(−sai ).
(13)
i=1
(See Remark 1 in appendix)
Marginal Estimation with Temperature We
now introduce a temperature parameter T > 0 to
problem (13) which divides E(x), or equivalently divides s belonging to EP (F ). Also, since T > 0, we
can multiply the objective by T leaving the problem
unchanged. Without changing the solution, we can
transform problem (13) as follows
Motivated by this connection, we ask if one can introduce a temperature parameter to our problem (8) and
transform it to an LP relaxation in the limit T → 0?
We can then hope to use the tightest LP relaxations
of MAP problems known in literature to find worstcase optimal submodular extensions. We answer this
question in affirmative. Specifically, in the following
two sections we show how one can select the set encoding and submodular extension to convert problem
(8) to the tightest known LP relaxations for Potts and
hierarchical Potts models. Importantly, we prove the
worst-case optimality of the extensions thus obtained.
Worst-case Optimal Submodular Extension
We now connect our marginal estimation problem (8)
with LP relaxations using the following proposition.
5
Proposition 1. Using the 1-of-L encoding scheme, in
the limit T → 0, problem (34) for Potts model becomes:
Potts Model
The Potts model, also known as the uniform metric,
specifies the pairwise potentials φab (xa , xb ) in equation
(5) as follows:
φab (xa , xb ) = wab · [xa 6= xb ],
min
s∈EP (F )
gT (s) =
N
X
T · log
a=1
L
X
i=1
exp(−
sai
).
T
− min f (y)
y∈∆
(14)
(15)
where f (.) is the Lovasz extension of F (.).
(Proof in appendix)
(10)
where wab is the weight associated with edge (a, b).
The above problem is equivalent to an LP relaxation
of the corresponding MAP estimtation problem (see
Pankaj Pansari1,2 , Chris Russell
2,3
, M. Pawan Kumar1,2
s
1-1
1-2
1-3
2-1
φ (3)
φ1(1) φ1(2) 1
2-2
2-3
φ2(1) φ2(2)
φ2(3)
t
Figure 3: An st-graph specifying the worst-case optimal
submodular extension for Potts model for 2 variables with 3
labels each and connected to each other. There is a node for
each variable and each label, that is, for all elements of the
ground set. The nodes have been labeled as ‘variable-label’,
hence node 1-1 represents the element v11 and so on. The
solid blue arcs model the unary potentials, and the dotted
red arcs represent the pairwise potentials. Each dotted red
arc has weight w12 /2.
Remark 2 in appendix). We note that gT (s) in problem
(34) becomes the objective function of an LP relaxation
in the limit T → 0. We seek to obtain the worst-case
optimal submodular extension by making gT (s) same
as the objective of (P-LP) as T → 0. Since at T = 1,
problems (34) and (13) are equivalent, this gives us the
worst-case optimal extension for our problem (13) as
well.
The question now becomes how to recover the worstcase optimal submodular extension using E(y). The
following propositions answers this question.
Proposition 2. The worst-case optimal submodular
extension for Potts model is given by FP otts (A) =
PL
i=1 Fi (A), where
Fi (A) =
X
φa (i)[|A ∩ {vai }| = 1]+
a
X
(a,b)∈N
wab
· [|A ∩ {vai , vbi }| = 1]
2
(16)
Also, E(y) in (P-LP) is the Lovasz extension of FP otts .
(Proof in appendix)
Proposition 2 paves the way for us to identify the worstcase optimal extension for hierarchical Potts model,
which we discuss in the following section.
6
Hierarchical Potts
Potts model imposes the same penalty for unequal assignment of labels to neighbouring variables, regardless
of the label dissimilarity. A more natural approach is
to vary the penalty based on how different the labels
are. A hierarchical Potts model permits this by specifying the distance between labels using a tree with the
following properties:
Figure 4: An hierarchical Potts model instance illustrating
notations with 2 meta-labels (blue) and 4 labels (yellow).
All labels are at the same level. r = 2, that is, edge-length
decreases by 2 at each level. Also, distance between labels 1
and 3 is d(1, 3) = 0.5 + 1 + 1 + 0.5 = 3.
1. The vertices are of two types: (i) the leaf nodes
representing labels, and (ii) the non-leaf nodes,
except the root, representing meta-labels.
2. The lengths of all the edges from a parent to its
children are the same.
3. The lengths of the edges along any path from the
root to a leaf decreases by a factor of at least r ≥ 2
at each step.
4. The metric distance between nodes of the tree is
the sum of the edge lengths on the unique path
between them.
A subtree T of an hierarchical Potts model is a tree
comprising all the descendants of some node v (not
root). Given a subtree T , lT denotes the length of the
tree-edge leading upward from the root of T and L(T )
denotes the set of leaves of T . We call the leaves of the
tree as labels and all other nodes of the tree expect the
root as meta-labels. Figure 4 illustrates the notations
in the context of a hierarchical Potts model.
Tightest LP Relaxation We use the same indicator
variables yai that were employed
P in the LP relaxation of
Potts model. Let ya (T ) = i∈L(T ) yai . The following
LP relaxation is the tightest known for hierarchical
Potts model in the worst-case, assuming the Unique
Games Conjecture to be true [Manokaran et al., 2008]
XX
e
(T-LP) min E(y)
=
φa (i)yai +
y
X
a
i
wab
X
(a,b)∈N
such that
lT · |ya (T ) − yb (T )|
T
y ∈ ∆.
(17)
The set ∆ is the same domain as defined in equation
(57). We rewrite this LP relaxation using indicator
variables zai for all labels and meta-labels as
(T-LP-FULL)
e
min E(z)
such that
z ∈ ∆0
(18)
Worst-case Optimal Submodular Extensions for Marginal Estimation
where ∆0 is the convex hull of the vectors satisfying
X
zai = 1, zai ∈ {0, 1} ∀a ∈ X , i ∈ L (19)
i∈L
and zai =
X
zaj . ∀a ∈ X , i ∈ R − L
(20)
j∈L(Ti )
The details of the new relaxation (T-LP-FULL) can be
found in the appendix.
Set Encoding For any variable Xa , let the set of
possible assignment of labels and meta-labels be the
set Va = {va1 , . . . , vaM }, where M is the total number
of nodes in the tree except the root. Our ground set is
V = ∪N
a=1 Va of size N · M .
A consistent labeling of a variable assigns it one label,
and all meta-labels on the path from root to the label.
Let us represent the set of consistent assignments for
Xa by the set Pa = {pa1 , . . . , paL }, where pai is the
collection of elements from Va for label i and all metalabels on the path from root to label i.
The set of valid labelings A ⊆ V assigns each variable
exactly one consistent label. This constraint can be
formally written as M = ∩N
a=1 Ma where Ma has
exactly one element from Pa .
Let s0ai be the sum of the components of s corresponding
to the elements of pai , that is,
X
s0ai =
st .
(21)
t∈pai
Using our encoding scheme, we rewrite problem (8) as:
min
s∈EP (F )
N
X
log
a=1
L
X
exp(−s0ai ).
(22)
i=1
s∈EP (F )
gT (s) =
N
X
T · log
a=1
L
X
i=1
exp(−
s0ai
).
T
Proposition 4. The worst-case optimal submodular extension
hierarchical Potts model is given by
Pfor
M
Fhier (A) = i=1 Fi (A), where
Fi (A) =
X
φ0a (i)[|A ∩ {vai }| = 1]+
a
X
wab · lTi · [|A ∩ {vai , vbi }| = 1]
e
Also, E(z)
in (T-LP-FULL) is the Lovasz extension of
Fhier . (Proof in appendix)
Since any finite metric space can be probabilistically
approximated by mixture of tree metric [Bartal, 1996],
the worst-case optimal submodular extension for metric
energies can be obtained using Fhier . Note that Fhier
reduces to FP otts for Potts model. One can see this by
considering the Potts model as a star-shaped tree with
edge weights as 0.5.
7
Fast Conditional Gradient
Computation for Dense CRFs
Dense CRF Energy Function A dense CRF is
specified by the following energy function
X
X X
E(x) =
φa (xa ) +
φab (xa , xb ). (26)
− min0 f (z)
(24)
a∈X b∈X ,b6=a
Note that every random variable is a neighbour of
every other random variable in a dense CRF. Similar
to previous work [Koltun and Krahenbuhl, 2011], we
consider the pairwise potentials to be to be Gaussian,
that is,
X
(m)
φ(i, j) = µ(i, j)
w(m) k(fa(m) , fb ),
(27)
m
(m)
k(fa(m) , fb )
Proposition 3. In the limit T → 0, problem (52) for
hierarchical Potts energies becomes:
(25)
(a,b)∈N
(23)
Worst-case Optimal Submodular Extension
The following proposition connects the marginal estimation problem (8) with LP relaxations:
z∈∆
The above problem is equivalent to an LP relaxation
of the corresponding MAP estimtation problem (see
Remark 3 in appendix). Hence, gT (s) becomes the
objective function of an LP relaxation in the limit
e
T → 0. We seek to make this objective same as E(z)
of (T-LP-FULL) in the limit T → 0. The question
now becomes how to recover the worst-case optimal
e
submodular extension from E(z).
a∈X
Marginal Estimation with Temperature Similar to Potts model, we now introduce a temperature
parameter T > 0 to problem (22). The transformed
problem becomes
min
(Proof in appendix).
= exp
−||fa − fb ||2
2
.
(28)
The term µ(i, j) is known as label compatibility function
between labels i and j. Potts model and hierarchical
Potts models are examples of µ(i, j). The other term
is a mixture of Gaussian kernels k(., .) and is called
(m)
the pixel compatibility function. The terms fa are
features that describe the random variable Xa . In
practice, similar to [Koltun and Krahenbuhl, 2011], we
Pankaj Pansari1,2 , Chris Russell
2,3
, M. Pawan Kumar1,2
use x, y coordinates and RGB values associated to a
pixel as its features.
However, using the equivalence of relaxed LP objectives
and the Lovasz extension of submodular extensions in
proposition 6, we are able to compute s∗ in O(N L) time.
Specifically, we use the algorithm of Ajanthan et al.
[2017], which provides an efficient filtering procedure to
compute the subgradient of the LP relaxation objective
E(y) of (P-LP).
Proposition 5. Computing the subgradient of E(y)
in (P-LP) is equivalent to computing the conditional
gradient for the submodular function FP otts .
(Proof in appendix).
A similar observation can be made in case of hierarchical Potts model. Hence we have the first practical
algorithm to compute upper bound of log-partition
function of a dense CRF for Potts and metric energies.
8
Experiments
Using synthetic data, we show that our upper-bound
compares favorably with TRW for both Potts and hierarchical Potts models. For comparison, we restrict
ourselves to sparse CRFs, as the code available for TRW
does not scale well to dense CRFs. We also perform
stereo matching using dense CRF models and compare our results with the mean-field-based approach of
[Koltun and Krahenbuhl, 2011]. All experiments were
run on a x86-64, 3.8GHz machine with 16GB RAM. In
this section, we refer to our algorithm as Submod and
mean field as MF.
8.1
Upper-bound Comparison using
Synthetic Data
Data We generate lattices of size 100 × 100, where
each lattice point represents a variable taking one of
20 labels. The pairwise relations of the sparse CRFs
are defined by 4-connected neighbourhoods. The unary
potentials are uniformly sampled in the range [0, 10].
We consider (a) Potts model and (b) hierarchical Potts
models with pairwise distance between labels given by
the trees of figure 5. The pairwise weights are varied
in the range {1, 2, 5, 10}. We compare the results of
our worst-case optimal submodular extension with an
(a) Tree 1
(b) Tree 2
Figure 5: The hierarchical Potts models defining pairwise
distance among 20 labels used for upper-bound comparison
with TRW. Blue nodes are the meta-labels and yellow nodes
are labels. All the edges at a particular level have the same
edge weights. The sequence of weights from root level to leaf
level is 1, 0.5 for tree 1 and 1, 1, 0.5 for tree 2. The yellow
node is shown to clump together 4 and 5 leaf nodes for tree
1 and 2 respectively.
s
φ1(1)
1-1
φ1(3)
1-2
φ2(1)
2-1
φ1(2)
1-3
φ2(3)
2-3
2-2
φ2(2)
t
Figure 6: An st-graph specifying the alternate submodular
extension for Potts model for 2 variables with 3 labels each
and connected to each other. The convention used is same
as in figure 3. Each dotted red arc has weight w12 /2. This
alternate extension was also used to derive the extension
for hierarchical Potts model.
4.0
(Submodular bound - TRW bound)/|TRW bound|
Algorithm 1 assumes that the conditional gradient s∗ in
step 3 can be computed efficiently. This is certainly not
the case for dense CRFs, since computing s∗ involves
N L function evaluations of the submodular extension
F , where N is the number of variables, and L is the
number of labels. Each F evaluation has complexity
O(N ) using the efficient Gaussian filtering algorithm
of [Koltun and Krahenbuhl, 2011]. However, computation of s∗ would still be O(N 2 ) this way, which is
clearly impractical for computer-vision applications
where N ∼ 105 − 106 .
Potts Optimal Extension
Potts Alternate Extension
Hierarchical Tree 1 Optimal Extension
Hierarchical Tree 1 Alternate Extension
Hierarchical Tree 2 Optimal Extension
Hierarchical Tree 2 Alternate Extension
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.5
1
2
3
4
5
6
Weight
7
8
9
10
Figure 7: Upper-bound comparison using synthetic
data. The plot shows the ratio (Submod bound - TRW
bound)/|TRW bound| averaged over 100 unary instances as
a function of pairwise weights using the worst-case optimal
and alternate extension for Potts and hierarchical Potts
models. We observe that the worst-case optimal extension
(solid) results in tighter bounds as compared to the respective
alternate extensions (dotted). Also, the worst-case optimal
extension bounds are in similar range as the TRW bounds.
Figure best viewed in color.
Worst-case Optimal Submodular Extensions for Marginal Estimation
alternate submodular extension as given in figure 2.
Method For our algorithm, we use the standard
schedule γ = 2/(k + 2) to obtain step size γ at iteration k. We run our algorithm till convergence - 100
iterations suffice for this. The experiments are repeated
for 100 randomly generated unaries for each model and
each weight. For TRW, we used the MATLAB toolbox
of [Domke, 2013]. The baseline code does not optimise
over tree distributions. We varied the edge-appearance
probability in trees over the range [0.1 - 0.5] and found
0.5 to give tightest upper bound.
Results We plot the ratio of the normalised difference of the upper bound values of our method with
TRW as a function of pairwise weights. The ratios are
averaged over 100 instances of unaries. Figure 7 shows
the plots for Potts and hierarchical Potts models for
the worst-case optimal and alternate extension. We
find that the optimal extension (solid) results in tighter
upper-bounds than the alternate extension (dotted) for
both models. To see the reason for this, we observe
that the representation of the submodular function
using figure 6 necessitates that φa (i) be non-negative.
This implies that F (A) values are larger for the worstcase optimal extension of figure 3 as compared to the
alternate extension. Hence the minimisation problem
8 has the same objective function g(s) for both cases
but the domain EP (F ) of equation (2) is larger for the
optimal extension, thereby resulting in better minima.
Figure 7 also indicates that our algorithm with optimal
extension provides similar range of upper bound as
TRW, thereby providing empirical justification of our
method. Note that the TRW upper bound has to be
tighter than our method. This is because the TRW
makes use of the standard LP relaxation [Chekuri et al.,
2004] which involves marginal variables for nodes as
well as edges. On the other hand, our method makes
use of the LP relaxation proposed by Kleinberg and
Tardos [2002] which involves marginal variables only
for nodes. The standard LP relaxation is tighter than
Kleinberg-Tardos relaxation, and hence TRW results in
better approximation. However, TRW does not scale
well with neighborhood size, thereby prohibiting its use
in dense CRFs.
8.2
Stereo Matching using Dense CRFs
Data We demonstrate the benefit our algorithm for
stereo matching on images extracted from the Middlebury stereo matching dataset [Scharstein et al., 2001].
We use dense CRF models with Potts compatibility
term and Gaussian pairwise potentials. The unary
terms are obtained using the absolute difference matching function of [Scharstein et al., 2001].
Method We use the implementation of mean-field
algorithm for dense CRFs of [Koltun and Krahenbuhl,
2011] as our baseline. For our algorithm, we make use
of the modified Gaussian filtering implementation for
dense CRFs by [Ajanthan et al., 2017] to compute the
conditional gradient at each step. The step size γ at
each iteration is selected by doing line search. We run
our algorithm till 100 iterations, since the visual quality
of the solution does not show much improvement beyond this point. We run mean-field up to convergence,
with a threshold of 0.001 for change in KL-divergence.
Results Figure 8 shows some example solutions obtained by picking the label with maximum marginal
probability for each variable for mean-field and for our
algorithm. We also report the time and energy values
of the solution for both methods. Though we are not
performing MAP estimation, energy values give us a
quantitative indication of the quality of solutions. For
the full set of 21 image pairs (2006 dataset), the average
ratio of the energies of the solutions from our method
compared to mean-field is 0.943. The avearge time
ratio is 10.66. We observe that our algorithm results
in more natural looking stereo matching results with
lower energy values for all images. However, mean-field
runs faster than our method for each instance.
9
Discussion
We have established the relation between submodular
extension for the Potts model and the LP relaxation
for MAP estimation using Lovasz extension. This allowed us to identify the worst-case optimal submodular
extension for Potts as well as the general metric labeling problems. It is worth noting that it might still be
possible to obtain an improved submodular extension
for a given problem instance. The design of a computationally feasible algorithm for this task is an interesting
direction of future research. While our current work has
focused on pairwise graphical models, it can be readily applied to high-order potentials by considering the
corresponding LP relaxation objective as the Lovasz extension of a submodular extension. The identification
of such extensions for popular high-order potentials
such as the P n Potts model or its robust version could
further improve the accuracy of important computer
vision applications such as semantic segmentation.
Pankaj Pansari1,2 , Chris Russell
(a) Venus GT
(a) Cones GT
(b) MF solution
(c) Submod solution
60.32s, 1.83e+07
469.75s, 1.55e+07
(b) MF solution
(c) Submod solution
239.14s, 2.68e+07
1082.72s, 1.27e+07
2,3
, M. Pawan Kumar1,2
(a) Tsukuba GT
(a) Teddy GT
(b) MF solution
(c) Submod solution
14.93s, 8.21e+06
215.22s, 4.12e+06
(b) MF solution
(c) Submod solution
555.30s, 2.36e+07
1257.86s, 1.58e+07
Figure 8: Stereo matching using dense CRFs with Potts compatibility and Gaussian pairwise potentials. We compare our
solution with the mean-field algorithm of Koltun and Krahenbuhl [2011]. We observe that our method gives better-looking
solutions with lower energy value at the cost of higher computational time.
References
Ajanthan, T., Desmaison, A., Bunel, R., Salzmann, M.,
Torr, P., and Kumar, M. (2017). Efficient continuous
relaxations for dense crf. In CVPR.
Bartal, Y. (1996). Probabilistic approximation of metric spaces and its algorithmic applications. In Foundations of Computer Science.
Bartal, Y. (1998). On approximating arbitrary metrices
by tree metrics. In ACM Symposium on Theory of
Computing.
Boykov, Y., Veksler, O., and Zabih, R. (2001). Fast
approximate energy minimization via graph cuts.
PAMI.
Chekuri, C., Khanna, S., Naor, J., and Zosin, L. (2004).
A linear programming formulation and approximation algorithms for the metric labeling problem.
SIAM Journal on Discrete Mathematics.
Djolonga, J. and Krause, A. (2014). From map to
marginals: Variational inference in bayesian submodular models. In NIPS.
Domke, J. (2013). Learning graphical model parameters
with approximate marginal inference. PAMI.
Edmonds, J. (1970). Submodular functions, matroids,
and certain polyhedra. Combinatorial Optimization
— Eureka, You Shrink!
Frank, M. and Wolfe, P. (1956). An algorithm for
quadratic programming. Naval research logistics
quarterly.
Kleinberg, J. and Tardos, E. (2002). Approximation
algorithms for classification problems with pairwise
relationships: Metric labeling and markov random
fields. IEEE Symposium on the Foundations of Computer Science.
Koltun, V. and Krahenbuhl, P. (2011). Efficient inference in fully connected crfs with gaussian edge
potentials. NIPS.
Kumar, M., Veksler, O., and Torr, P. (2011). Improved
moves for truncated convex models. JMLR.
Manokaran, R., Naor, J., Raghavendra, P., and
Schwartz, R. (2008). Sdp gaps and ugc hardness
for multiway cut, 0-extension, and metric labeling.
In ACM Symposium on Theory of Computing.
Scharstein, D., Szeliski, R., and Zabih, R. (2001). A
taxonomy and evaluation of dense two-frame stereo
correspondence algorithms. In Stereo and MultiBaseline Vision.
Wainwright, M., Jaakkola, T., and Willsky, A. (2005).
A new class of upper bounds on the log partition
function. IEEE Transactions on Information Theory.
Zhang, J., Djolonga, J., and Krause, A. (2015). Higherorder inference for multi-class log-supermodular models. In ICCV.
Worst-case Optimal Submodular Extensions for Marginal Estimation
Appendix
1
Proofs for Potts Model Extension
Remark 1
We show using induction over the number of variables that with 1-of-L encoding for Potts,
X
exp(−s(A)) =
N X
L
Y
exp(−sai ).
(29)
a=1 i=1
A∈M
Proof. Let t be the number of variables, V t be the corresponding ground set and Mt be the sets corresponding
to valid labelings. Equation (29) clearly holds for t = 1.
Let us assume that the relation holds for t = N , that is,
X
N
exp(−s(A )) =
N X
L
Y
exp(−sai )
(30)
a=1 i=1
AN ∈MN
For t = N + 1,
X
exp(−s(AN +1 )) =
L
X
X
exp(−s(AN ) − sN +1,i )
i=1 AN ∈MN
AN +1 ∈MN +1
=
L
X
i=1
=
L
X
exp(−s(AN ))
AN ∈MN
exp(−sN +1,i )
N
+1 X
L
Y
N X
L
Y
exp(−sai )
a=1 i=1
i=1
=
X
exp(−sN +1,i )
exp(−sai )
(31)
a=1 i=1
Remark 2 Given any submodular extension F (.) of a Potts energy function E(.), its Lovasz extension f (.)
defines an LP relaxation of the MAP problem for E(.) as
min f (y).
y∈∆
(32)
Proof. By definition of a submodular extension and the Lovasz extension, E(x) = F (Ax ) = f (1Ax ) for all valid
labelings x. Also, from property 1, f (y) is maximum of linear functions. Hence, f (y) is a piecewise linear
relaxation of E(x).
The domain ∆ is a polytope formed by union of N probability simplices
∆ = {ya ∈ RL |ya 0 and h1, ya i = 1}
(33)
With objective as maximum of linear functions and domain as a polytope, we have an LP relaxation of the
corresponding MAP problem.
Pankaj Pansari1,2 , Chris Russell
2,3
, M. Pawan Kumar1,2
Proposition 6. In the limit T → 0, the following problem for Potts energies
min
s∈EP (F )
gT (s) =
N
X
T · log
a=1
L
X
exp(−
i=1
sai
).
T
(34)
becomes
− min f (y).
y∈∆
(35)
Proof. In the limit of T → 0, we can rewrite the above problem as
N
X
min
s∈EP (F )
a=1
max(−sai ).
i
(36)
In vector form, the problem becomes
min
max −hy, si
s∈EP (F ) y∈∆
(37)
= − max minhy, si
(38)
∆ = {ya ∈ RL |ya 0 and h1, ya i = 1}
(39)
s∈EP (F ) y∈∆
∆ is the union of N probability simplices:
where ya is the component of y corresponding to the a-th variable. By the minimax theorem for LP, we can
reorder the terms:
− min max hy, si
(40)
y∈∆ s∈EP (F )
Recall that maxs∈EP (F ) hy, si is the value of the Lovasz extension of F at y, that is, f (y). Hence, as T → 0, the
marginal inference problem converts to minimising the Lovasz extension under the simplices constraint:
− min f (y)
y∈∆
(41)
Proposition 7. The objective function E(y) of the LP relaxation (P-LP) is the Lovasz extension of FP otts (A) =
PL
i=1 Fi (A), where
X
Fi (A) =
φa (i)[|A ∩ {vai }| = 1]+
a
X
(a,b)∈N
wab
· [|A ∩ {vai , vbi }| = 1].
2
(42)
Proof. Since FP otts is sum of Ising models Fi , we first focus on a particular label i and then generalize. Consider
a graph with only two variables Xa and Xb with an edge between them. The ground set in this case is {vai , vbi }.
Let the corresponding relaxed indicator variables be y = {yaj , ybj }, such that yai , ybi ∈ [0, 1] and assume yai > ybi .
The Lovasz extension is:
f (y) = yai · [Fi ({vai }) − Fi ({})] + ybi · [Fi ({vai , vbi }) − Fi ({vai })]
wab
wab
= yai · [ φa (j) +
− 0] + ybi · [(φa (j) + φb (j)) − φa (j) +
]
2
2
wab
= φa (j) · yai + φb (j) · ybi +
· (yai − ybi )
2
(43)
Worst-case Optimal Submodular Extensions for Marginal Estimation
In general for both orderings of yab and ybi , we can write
wab
· |yai − ybi |
2
Extending Lovasz extension (equation (44)) to all variables and labels gives E(y) in (P-LP).
f (y) = φa (j) · yai + φb (j) · ybi +
2
(44)
Proofs for Hierarchical Potts Model Extension
Transformed Tightest LP Relaxation We take (T-LP) and rewrite it using indicator variables for all labels
and meta-labels. Let R denote the set of all labels and meta-labels, that is, all nodes in the tree apart from the
root. Also, let L denote the set of labels, that is, the leaves of the tree. Let Ti denote the subtree which is rooted
at the i-th node. We introduce an indicator variable zai ∈ {0, 1}, where
(
yai
if i ∈ L
zai =
(45)
ya (Ti )
if i ∈ R − L
We need to extend the definition of unary potentials to the expanded label space as follows:
(
φa (i)
if i ∈ L
where φ0a (i) =
0
if i ∈ R − L
(46)
We can now rewrite problem (T-LP) in terms of new indicator variables zai :
XX
e
(T-LP-FULL) min E(z)
=
φ0a (i) · zai +
i∈R a∈X
X X
wab · lTi · |zai − zbi |
i∈R (a,b)∈N
such that
z ∈ ∆0
(47)
where ∆0 is the convex hull of the vectors satisfying
X
zai = 1, zai ∈ {0, 1} ∀a ∈ X , i ∈ L
(48)
i∈L
and zai =
X
zaj . ∀a ∈ X , i ∈ R − L
(49)
j∈L(Ti )
Constraint (49) ensures consistency among labels and meta-labels, that is, if a label is assigned then all the
meta-labels which lie on the path from the root to the label should be assigned as well. We are now going to
identify a suitable set encoding and the worst-case optimal submodular extension using (T-LP-FULL).
Remark 3 Given any submodular extension F (.) of a hierachical Potts energy function E(.), its Lovasz extension
defines an LP relaxation of the corresponding MAP estimation problem as
min f (z).
z∈∆0
(50)
Proof. By definition of a submodular extension and the Lovasz extension, E(x) = F (Ax ) = f (1Ax ) for all valid
labelings x. Also, from property 1, f (y) is maximum of linear functions. Hence, f (y) is a piecewise linear
relaxation of E(x).
We can write the domain ∆0 as
∆0 = {ya ∈ RM |ya 0, h1, yalabel i = 1, ya (pai ) = 1 or ya (pai ) = 0∀i ∈ [1, L]}
(51)
where ya is the component of y corresponding to the a-th variable, yalabel is the component of ya corresponding
to the L labels, and ya (pai ) is the component of ya corresponding to the elements of pai .
Since ∆0 is defined by linear equalities and inequalities, it is a polytope. With objective as maximum of linear
functions and domain as a polytope, we have an LP relaxation of the corresponding MAP problem.
Pankaj Pansari1,2 , Chris Russell
2,3
, M. Pawan Kumar1,2
Proposition 8. In the limit T → 0, the following problem for hierarchical Potts energies
min
s∈EP (F )
gT (s) =
N
X
T · log
a=1
L
X
exp(−
i=1
s0ai
).
T
(52)
becomes:
− min0 f (z).
(53)
z∈∆
Proof. In the limit of T → 0, we can rewrite the above problem as
N
X
min
s∈EP (F )
a=1
max(−s0ai ).
(54)
i
In vector form, the problem becomes
min
max −hz, s0 i
(55)
s∈EP (F ) z∈∆
= − max minhz, s0 i
(56)
s∈EP (F ) z∈∆
where ∆ = {za ∈ RL |za 0 and h1, za i = 1}
(57)
0
where za is the component of z corresponding to the a-th variable. We can unpack s using
X
s0ai =
st .
(58)
t∈pai
and rewrite problem (56) as
− max
min hy, si
s∈EP (F ) y∈∆0
(59)
The new constraint set ∆0 ensures that the binary entries of labels and meta-labels is consistent:
where ∆0 = {ya ∈ RM |ya 0, h1, yalabel i = 1,
ya (pai ) = 1 or ya (pai ) = 0∀i ∈ [1, L]}
(60)
where ya is the component of y corresponding to the a-th variable, yalabel is the component of ya corresponding
to the L labels, and ya (pai ) is the component of ya corresponding to the elements of pai .
By the minimax theorem for LP, we can reorder the terms:
− min0 max hy, si
y∈∆ s∈EP (F )
(61)
Recall that maxs∈EP (F ) hy, si is the value of the Lovasz extension of F at y, that is, f (y). Hence, as T → 0, the
marginal inference problem converts to minimising the Lovasz extension under the constraints ∆0 :
− min0 f (y).
y∈∆
(62)
e
Proposition 9. The objective function E(z)
of (T-LP-FULL) is the Lovasz extension of Fr−HST (A) =
PM
F
(A),
where
i
i=1
X
φ0a (i)[|A ∩ {vai }| = 1]+
Fi (A) =
a
X
(a,b)∈N
wab · lTi · [|A ∩ {vai , vbi }| = 1].
(63)
Worst-case Optimal Submodular Extensions for Marginal Estimation
Proof. We observe that Fr−HST is of exactly the same form as FP otts , except that the Ising models Fi are defined
over not just labels, but meta-labels as well. Using the same logic as in the proof of proposition 7, each Fi is the
Lovasz extension of
X
X
ei (z) =
E
φ0a (i) · zai +
wab · lTi · |zai − zbi |
(64)
a∈X
and the results follows.
(a,b)∈N
| 1cs.CV
|
1
Compressive Sensing for Millimeter Wave
Antenna Array Diagnosis
arXiv:1612.06345v4 [cs.IT] 9 Oct 2017
Mohammed E. Eltayeb, Tareq Y. Al-Naffouri, and Robert W. Heath, Jr.
Abstract—The radiation pattern of an antenna array depends
on the excitation weights and the geometry of the array. Due
to wind and atmospheric conditions, outdoor millimeter wave
antenna elements are subject to full or partial blockages from a
plethora of particles like dirt, salt, ice, and water droplets. Handheld devices are also subject to blockages from random finger
placement and/or finger prints. These blockages cause absorption
and scattering to the signal incident on the array, modify the
array geometry, and distort the far-field radiation pattern of the
array. This paper studies the effects of blockages on the far-field
radiation pattern of linear arrays and proposes several array
diagnosis techniques for millimeter wave antenna arrays. The
proposed techniques jointly estimate the locations of the blocked
antennas and the induced attenuation and phase-shifts given
knowledge of the angles of arrival/departure. Numerical results
show that the proposed techniques provide satisfactory results in
terms of fault detection with reduced number of measurements
(diagnosis time) provided that the number of blockages is small
compared to the array size.
Index Terms—Antenna arrays, fault diagnosis, compressed
sensing, millimeter wave communication.
I. I NTRODUCTION
T
HE abundance of bandwidth in the millimeter wave
(mmWave) spectrum enables gigabit-per-second data
rates for cellular systems and local area networks [1], [2].
MmWave systems make use of large antenna arrays at both
the transmitter and the receiver to provide sufficient receive
signal power. The use of large antenna arrays is justified by
the small carrier wavelength at mmWave frequencies which
permits large number of antennas to be packed in small form
factors.
Due to weather and atmospheric effects, outdoor mmWave
antenna elements are subject to blockages from flying debris or
particles found in the air as shown in Fig. 1. The term “blockage” here refers to a physical object partially or completely
blocking a subset of antenna elements and should not be
confused with mmWave channel blockage. MmWave antennas
on handheld devices are also subject to blockage from random
finger placement and/or fingerprints on the antenna array.
Part of this work was presented in the 2016 IEEE Global Communications
Conference [3]. Mohammed E. Eltayeb is with the California State University, Sacramento, California, USA (e-mail: [email protected]),
Robert W. Heath, Jr. is with the University of Texas at Austin, USA
(email: [email protected]), and Tareq Y. Al-Naffouri is with King Abdullah University of Science and Technology, Saudi Arabia (e-mail:
[email protected]). This research was partially supported by the
U.S. Department of Transportation through the Data-Supported Transportation
Operations and Planning (D-STOP) Tier 1 University Transportation Center
and by the Texas Department of Transportation under Project 0-6877 entitled
Communications and Radar-Supported Transportation Operations and Planning (CAR-STOP).
Partial or complete blockage of some of the antenna elements
reduces the amount of energy incident on the antenna [4],
[5]. For instance, it is reported in [6] that 90% of a 76.5
GHz signal energy will be absorbed by a water droplet of
thickness 0.23 mm. A thin water film caused by, for example
a finger print, is also reported to cause attenuation and a
phase shift on mmWave signals [6]. Moreover, snowflakes,
ice stones, and dry and damp sand particles are reported
to cause attenuation and/or scattering [4]-[8]. Because the
size of these suspended particles is comparable to the signal
wavelength and antenna size, random blockages caused by
these particles will change the antenna geometry and result
in a distorted radiation pattern [9], [10]. Random changes
in the array’s radiation pattern causes uncertainties in the
mmWave channel. It is therefore important to continuously
monitor the mmWave system, reveal any abnormalities, and
take corrective measures to maintain efficient operation of
the system. This necessitates the design of reliable and low
latency array diagnosis techniques that are capable of detecting
the blocked antennas and the corresponding signal power loss
and/or phase shifts caused by the blocking particles. Once a
fault has been detected, pattern correction techniques proposed
in, for example, [9]-[13] can be employed to calculate new
excitation weights for the array.
Several array diagnostic techniques, which are based on
genetic algorithms [14], [15], matrix inversion [16], exhaustive
search [17], and MUSIC [18], have been proposed in the
literature to identify the locations of faulty antenna elements.
These techniques compare the radiation pattern of the array
under test (AUT) with the radiation pattern of an “error free”
reference array. For large antenna arrays, the techniques in
[14]-[18] require a large number of samples (measurements) to
obtain reliable results. To reduce the number of measurements,
compressed sensing (CS) based techniques have recently been
proposed in [19]-[21]. Despite their good performance, the
techniques in [14]-[21] are primarily designed to detect the
sparsity pattern of a failed array, i.e. the locations of the
failed antennas and not necessarily the complex blockage
coefficients. Moreover, the CS diagnosis techniques proposed
in [19] and [20] have the following limitations: (i) They require
measurements to be made at multiple receive locations and
are not suitable when both the transmitter and the receiver
are fixed. (ii) They assume fault-free receive antennas, i.e.
faults at the AUT only, however, faults can occur at both the
transmitter and the receiver. (iii) They can not exploit correlation between faulty antennas to further reduce the diagnosis
time. (iv) They do not estimate the effective antenna element
gain, i.e. the induced attenuation and phase shifts caused
2
by blockages. These estimates can be used to re-calibrate
the array. v) They do not optimize the CS measurement
matrices, i.e. the restricted isometry property might not be
satisfied. While the CS technique proposed in [21] performs
joint fault detection/estimation and angle-of-arrival/departure
(AoA/D) estimation, it is not suitable for mmWave systems
as it requires a separate RF chain for each antenna element.
The high diagnosis time required by the techniques proposed
in [14]-[18] and the limitations of the CS based techniques
proposed in [19]-[21] motivate the development of new array
diagnosis techniques suitable for mmWave systems.
In this paper, we develop low-complexity array diagnosis
techniques for mmWave systems with large antenna arrays.
These techniques account for practical assumptions on the
mmWave hardware in which the analog phase shifters have
constant modulus and quantized phases, and the number of
RF chains is limited (assumed to be one in this paper). The
main contributions of the paper can be summarized as follows:
• We investigate the effects of random blockages on the
far-field radiation pattern of linear uniform arrays. We
consider both partial and complete blockages.
• We derive closed-form expressions for the mean and variance of the far-field radiation pattern as a function of the
antenna element blockage probability. These expressions
provide an efficient means to evaluate the impact of the
number of antenna elements and the antenna element
blockage probability on the far-field radiation pattern.
• We propose a new formulation for mmWave antenna
diagnosis which relaxes the need for multi-location measurements, captures the sparse nature of blockages, and
enables efficient compressed sensing recovery.
• We consider blockages at the transmit and/or receive
antennas and propose two CS based array diagnosis
techniques. These techniques identify the locations and
the induced attenuation and phase shifts caused by unstructured blockages.
• We exploit the two dimensional structure of mmWave
antenna arrays and the correlation between the blocked
antennas to further reduce the array diagnosis time when
structured blockages exist at the receiver.
• We evaluate the performance of the proposed array diagnosis techniques by simulations in a mmWave system
setting, assuming that both the transmit and receive
antennas are equipped with a single RF chain and 2-bit
phase shifters.
The remainder of this paper is organized as follows. In
Section II, we formulate the array diagnosis problem and
study the effects of random blockages on the far-field radiation
pattern of linear arrays. In Section III, we introduce the
proposed array diagnosis technique assuming a fault free
transmit array and in Section IV, we introduce the proposed
array diagnosis technique when faults are present at both the
transmit and receive arrays. In Section V we provide some
numerical results and conclude our work in Section VI.
II. P ROBLEM F ORMULATION
We consider a two-dimensional (2D) planar antenna array
with Nx equally spaced elements along the x-axis and Ny
Fig. 1: An example of an outdoor millimeter wave antenna array
with different suspended particles partially blocking the array. The
suspended particles, with different absorption and scattering properties, modify the array geometry.
equally spaced elements along the y-axis; nonetheless, the
model and the corresponding algorithms can be adapted to
other antenna structures as well. Each antenna element is
described by its position along the x and y axis, for example,
the (Nx , Ny )th antenna refers to an antenna located at the Nx th
position along the x-axis and the Ny th position along the yaxis. The ideal far-field radiation pattern of this planar array
in the direction (θ, φ) is given by [22]
Ny −1 Nx −1
f (θ, φ) =
X X
wn,m ejm
2πdx
λ
sin θ cos φ jn
e
2πdy
λ
sin θ sin φ
,(1)
n=0 m=0
where dx and dy are the antenna spacing along the x and y
axis, λ is the wavelength, and wn,m is the (n, m)th complex
antenna weight.
Let ax (θ, φ) ∈ C Nx ×1 and ay (θ, φ) ∈ C Ny ×1 be two
vectors where the mth entry of ax (θ, φ) is [ax (θ, φ)]m =
2πdx
ejm λ sin θ cos φ and the nth entry of ay (θ, φ) is [ay (θ, φ)]n =
2πdy
ejn λ sin θ sin φ . Also, let the matrix W ∈ C Ny ×Nx be a
matrix of antenna weights, where the (n, m)th entry of W
is [W]n,m = wn,m . Then (1) can be reduced to
T
f (θ, φ) = vec(W) a(θ, φ),
(2)
where vec(W) is the Nx Ny × 1 column vector obtained by
stacking the columns of the matrix W on top of one another,
the vector a(θ, φ) = ax (θ, φ) ⊗ ay (θ, φ) is the 1D array
response vector, and the operator ⊗ represents the Kronecker
product. The formulation in (2) allows us to represent the 2D
array as a 1D array, and as a result, simplify the problem
formulation.
In the presence blockages, the far-field radiation pattern of
the array in (2) becomes
T
g(θ, φ) = vec(W) (b ◦ a(θ, φ)),
{z
}
| {z } |
x
(3)
z
where operator ◦ represents the Hadamard product, the vector
x ∈ C Nx Ny ×1 is a vector of antenna weights and the vector
z ∈ C Nx Ny ×1 is the equivalent array response vector. The nth
entry of the vector b ∈ C Nx Ny ×1 is defined by
3
0
Original pattern
Pattern after phase correction
Damaged pattern
-5
Normalized amplitude (dB)
-10
-15
-20
-25
-30
-35
-40
0
20
40
60
80
100
120
140
160
180
Directional angle φ (degrees)
Fig. 2: Original and damaged beam patterns of a 16 element (4 × 4)
d
planar array with θ = 90 degrees and dλx = λy = 0.5. The third,
fifth and thirteenth array elements of the equivalent 1D array (see (2))
are blocked with b3 = 0.37 + j0.22, b5 = −0.1 + j0.34, and b1 3 =
−0.64−j0.1. Blockages result in an increase in the sidelobe level and
a decrease in gain. Phase correction improves beam pattern, however,
more complex precoder design is required for pattern correction.
bn =
αn ,
1,
if the nth element is blocked
otherwise,
(4)
where n = 1, ..., Nx Ny , αn = κn ejΦn , 0 ≤ κn ≤ 1 and
0 ≤ Φn ≤ 2π are the resulting absorption and scattering
coefficients at the nth element. A value of κn = 0 represents
maximum absorption (or blockage) at the nth element, and
the scattering coefficient Φn measures the phase-shift caused
by the particle suspended on the nth element. This makes bn
a random variable, i.e. bn = αn with probability Pb if the
nth antenna is blocked and bn = 1 with probability 1 − Pb
otherwise. It is clear from (3) that blockages will change the
array manifold and result in a distorted radiation pattern as
shown in Fig. 2. The resulting pattern is a function of the
number of the particles suspended on the array and their
corresponding dielectric constants.
In Tables I and II, we summarize the effects of blockages on
a linear array. Specifically, we tabulate the mean and variance
of a distorted far-field radiation pattern of a linear array subject
to blockages. The total number of blockages is assumed to be
fixed, however, block locations and intensities (for the case
of partial blockage) are assumed to be random. For ease of
exposition, we study the effects on the azimuth direction only.
A similar analysis can be performed for the elevation pattern.
Derivations of the results in Tables I and I can be found in [23]
and are omitted for space limitation. From Table I, we observe
that both complete and partial blockages reduce the amplitude
of the main lobe. This reduces the beamforming gain of the
array. We also observe that complete blockages have no effect
on the variance of the of the main lobe, and hence do not
cause randomness in the main lobe. Random partial blockages,
however, randomize the main lobe and lead to uncertainties
in the mmWave channel. From Table II, we observe that
complete and partial blockages distort the sidelobes of the farfield radiation pattern. Table II also shows that the variance of
this distortion is a function of the blockage intensity and the
antenna element blockage probability Pb .
In the following sections, we propose several array diagnosis
(or blockage detection/estimation) techniques for mmWave
antenna arrays. Once array diagnosis is complete, the estimated attenuation and phase shifts caused by blockages can
be used to calibrate the array. Fig. 2, for example, shows
the resulting pattern when phase correction is applied to
the affected antenna elements. As shown, the resulting beam
pattern is slightly improved, however, more complex precoder
design is required to modify the excitation weights of the
antenna elements and calibrate the array. The calibration process could for example focus on maximizing the beamforming
gain and/or minimize the sidelobe level, and as a result,
reduce the uncertainty in the mmWave channel. While the
excitation weights of failed arrays can be easily modified
in digital antenna architectures, additional hardware, e.g. RF
chains, antenna switches, subarrays, etc., might be required
to generate more degrees of freedom for the precoder design.
Hybrid architectures (see e.g. [24] and [25]) can be used to
modify the excitation weights of the failed array and also
reduce the diagnosis time since each RF chain can now
obtain independent measurements. Beam pattern correction
and precoder design for failed arrays is beyond the scope
of this work and is left for future work. Before we proceed
with the proposed techniques, we lay down the following
assumptions: (i) The number of blockages is assumed to be
small compared to the array size. (ii) The channel between
the transmitter and the receiver is line-of-sight (LoS) with a
single dominant path. In the case of multi-path, the transmitter
waits for a period τ , where τ is proportional to the channel’s
delay spread, before it transmits the following training symbol.
(iii) The transmit and receive array manifolds as well as the
transmitter’s angle-of-departure and the receiver’s angle-ofarrival are known at the receiver. The AoA/D can be known
a priori, obtained by, for example, using prior sub-6 GHz
channel information [28], or provided by an infrastructure
via a lower frequency control channel. (iv) Blockages remain
constant for a time interval which is larger than the diagnosis
time.
III. FAULT D ETECTION AT THE R ECEIVER
In the previous section, we showed that blockages distort
the beam pattern of the array. To mitigate the effects of
blockages, it is imperative to design reliable array diagnosis
techniques that detect the fault locations and estimate the
values of the blockage coefficients with minimum diagnosis
time. Array diagnosis can be initiated after channel estimation.
For example, the optional training subfield of the SC PHY
IEEE 802.11ad frame could be utilized for periodic array
diagnosis. Since the system performance is greatly affected
by the antenna architecture and precoder design, which we
do not undertake in this work, we do not simulate the effect
of CSI training loss in this paper and simply focus on the
array diagnosis problem. Note, however, that beamforming at
4
TABLE I: Mean and variance of the far-field beam pattern of a linear array steered at φ = φT as a function of the antenna element blockage
probability Pb and the blockage coefficient αn . The blockage coefficient is constant if the array is subject to a single type of blockage, and
random if the array is subject to multiple types of blockages. Derivation of results can be found in [23] and are omitted for space limitations
Fault Type
Complete blockage (αn∈Ik = 0)
Partial blockage (αn∈Ik = β)
Partial blockages (αn random)
Mean
1 − Pb
1 − Pb + Pb β
1 − Pb + Pb E[αn ]
Variance
0
0
Pb var[αn ]
TABLE II: Mean and variance of the far-field beam pattern of a linear array steered at φ 6= φT as a function of the antenna element blockage
probability Pb and the blockage coefficient αn ; γ = πdλx (cos(φ) − cos(φT )). Derivation of results can be found in [23] and are omitted for
space limitation.
Fault Type
Complete blockage (αn∈Ik = 0)
Mean
sin(Nx γ) j(Nx −1)γ
e
(1 − Pb ) N sin(γ)
Partial blockage (αn∈Ik = β)
(1 − Pb (1 − β)) N
Partial blockage (αn random)
x
(1 − Pb (1 −
A. Generic Fault Detection
To start the array diagnosis process, the receiver with the
AUT requests a transmitter, with known location, i.e. θ and φ,
to transmit K training symbols (known to the receiver). The
receiver with the AUT generates a random beam to receive
each training symbol as shown in Fig. 3(a), i.e., random
antenna weights are used at the AUT to combine each training
symbol. Mathematically, the kth output of the AUT can be
written as
√
ρsxT
k z + ek ,
(5)
where k = 1, ..., K, ρ is the effective signal-to-noise ratio
(SNR) which includes the path loss, s = 1 is the training
symbol, and ek ∼ CN (0, 1) is the additive noise. The entries
of the weighing vector xk at the kth instant are chosen
uniformly and independently at random. Equipped with the
path-loss and angular location of the transmitter, the receiver
generates the ideal pattern f (θ, φ) using (1). Subtracting the
ideal beam pattern from the received signal in (5) we obtain
T
yk = hk (θ, φ) − fk (θ, φ) = xT
k (b ◦ a(θ, φ)) − xk a(θ, φ) + ẽk
= xT
k (c ◦ a(θ, φ)) + ẽk ,
x
sin(Nx γ) j(Nx −1)γ
e
x sin(γ)
sin(Nx γ) j(Nx −1)γ
E[αn ])) N sin(γ) e
x
both the transmitter and receiver leads to larger coherence time
[26], [27] and the angular variation is typically an order of
magnitude slower than the conventional coherence time [26].
Since all AoDs/AoAs are assumed to be known a priori in this
paper, the loss in CSI training time becomes unsubstantial.
In this section we propose three array diagnosis techniques.
The first technique is generic in the sense that it does not
exploit the block structure of blockages while the second and
third techniques exploit the dependencies between blocked
antenna elements to further reduce the array diagnosis time.
hk (θ, φ) =
Variance
Pb
(1 − Pb )
N
(6)
ek
where ẽk = √
ρ , and the nth entry of the vector c is [c]n = 0
when there is no blockage at the nth element, and [c]n = bn −1
otherwise. Let q = c ◦ a(θ, φ), after K measurements we
obtain
x1,1
x2,1
..
.
x1,2
x2,2
..
.
xK,1
yK
| {z } |
xK,2
y1
y2
..
.
1
(1 − Pb
Nx
Pb
(1 − Pb
Nx
=
y
1
(1 − Pb
Nx
Pb E[αn ]2 )
+ Pb |β|2 ) −
+
E[α2n ]
−
···
···
..
.
x1,NR
x2,NR
..
.
···
xK,NR
{z
X
}|
+ Pb β)2
q1
q2
..
.
+
ẽ1
ẽ2
..
.
ẽK
qNR
{z } | {z }
q
e
or equivalently
y = Xq + e,
(7)
where NR = Nx Ny is the total number of antennas at
the receiver. Assuming that the number of blocked antenna
elements is small, i.e. S NR , the vector q in (7) becomes
sparse with S non-zero elements that represent the locations
of the blocked antennas.
To mitigate the effects of blockages, it is desired to first
detect the locations of the blocked antennas and then the
complex random variable bn with few measurements (or
diagnostic time). Note that the system in (7) requires K ≥ NR
measurements to estimate the vector q. While this might be
acceptable for small antenna arrays, mmWave systems are usually equipped with large antenna arrays to provide sufficient
link budget [29]. Scaling the number of measurements with
the number of antennas would require more measurements
and would increase the array diagnosis time. In the following,
we show how we can (i) detect the locations of the blocked
antennas, and (ii) estimate the complex blockage coefficients
bn with K NR measurements by exploiting the sparsity
structure of the vector q under the assumption that blockages
remain constant for a time interval which is larger than
the diagnosis time. If the time interval is small, a hybrid
architecture (see e.g. [24] and [25]) can be employed with
multiple RF chains and each RF chain can obtain independent
measurements. This reduces the diagnosis time by a factor of
NRF , where NRF is the number of RF chains.
1) Sparsity Pattern Detection: Compressive sensing theory
permits efficient reconstruction of a sensed signal with only
a few sensing measurements. While there are many different
methods used to solve sparse approximation problems (see,
e.g., [31], [32]), we employ the least absolute shrinkage and
selection operator (LASSO) as a recovery method. We adopt
5
Transmitter
Receiver
Transmitter
Receiver
(a)
(b)
Fig. 3: An example of real time array diagnosis. (a) Diagnosis of the receiver array. Transmitter sends K training symbols and the receiver
receives each training symbol using a random receive beam. (b) Joint diagnosis of both the transmit and the receive arrays. Transmitter sends
training symbols using kt random transmit beams and the receiver receives each training symbol using kr random receive beams.
the LASSO since it does not require the support of the vector
q to be known a priori. This makes it a suitable detection
technique as blockages are random in general. The LASSO
estimate of (7) is given by [31]
arg
min
ν∈CNR ×1
1
ky − Xνk22 + Ωσkνk1 ,
2
(8)
where σ is the standard derivation of the noise ẽ, and Ω is a
regularization parameter. The antennas weights, i.e. the entries
of the matrix X, are randomly and uniformly selected from the
set {1 + j, 1 − j, −1 + j, −1 − j} in this paper, i.e. weights can
be applied to a mmWave antenna with 2-bit phase shifters. In
the special case of 1-bit phase shifters, X becomes a Bernoulli
matrix which is known to satisfy the coherence property with
high probability [30], [32], [33].
2) Attenuation and Induced Phase Shift Estimation: Once
the support S, where S = {n : qn 6= 0}, one can apply
estimation techniques such as least squares (LS) estimation to
estimate and refine the complex coefficient bn∈S . To achieve
this, the columns of X which are associated with the non-zero
entries of q are removed to obtain XS ∈ CK×S . Hence, the
vector y in equation (7) can now be written as
y = XS qS + e,
(9)
where qS is obtained by pruning the zero entries of q. Since
K > S, the entries of qS can be estimated via LS estimation.
In particular, one can write the LS estimate after successful
sparsity pattern recovery as [34]
q̂S = (X∗S XS )−1 X∗S y = qS + ě,
number of measurements without scarifying robustness. In
this section, we reformulate the CS problem to exploit this
structure and reduce the array diagnosis time. To formulate
the problem, we first rewrite f (θ, φ) in (2) and g(θ, φ) in (3)
as
(10)
where q̂S is a noisy estimate of qS , and the entries of the
output noise vector ě are Gaussian random variables as linear
operations preserve the Gaussian noise distribution. Note that
the nth entry of the vector q is [q]n = (bn −1)an , where an is
the nth entry of of the vector a(θ, φ) (see (6)-(7)). Therefore,
+ 1| and
the estimated attenuation coefficient κ̂n∈S = | aq̂n∈S
n∈S
q̂n∈S
the estimated induced phase Φ̂n∈S = ∠ an∈S + 1 .
B. Exploiting the Block-Structure of Blockages
Due to the small antenna element size, it is likely that
a suspended particle will block several neighboring antenna
elements as shown in Fig. 1. This results in a block sparse
structure which could be exploited to substantially reduce the
T
f (θ, φ) = vec(W) vec(A),
(11)
and
T
g(θ, φ) = vec(W) vec(A ◦ B),
(12)
where A = ay (θ, φ)aT
x (θ, φ) is the array response matrix, W
is the weighting matrix, and B is the sparse blockage matrix,
i.e the entries of B are “1” in the case of no blockage and a
random variable (see (4)) in the case of a blockage. Substituting (11) and (12) in (5)-(6), the kth received measurement
(after subtracting it form the ideal pattern) becomes
T
T
yk = vec(Wk ) vec(A) − vec(Wk ) vec(A ◦ B) + ẽk (13)
T
= vec(Wk ) vec(As ) + ẽk ,
(14)
where Wk is the kth random weighting matrix, the innovation
matrix As = A − As is sparse, and ek is the additive noise.
Observe that the columns of the matrix A are either all 0’s,
in the case of no blockage, or contains a block of non-zero
entries. We exploit this structure to reduce the number of
measurements and as a result, reduce the array diagnosis time.
Let X = [vec(W1 ), vec(W2 ), · · · , vec(WK )]T be the
measurement matrix which consists of K random antenna
weights and q = vec(As ), then after K measurements the
innovation vector becomes
y = Xq + e.
(15)
Observe that (15) is similar to (7) with the exception that
the new formulation allows the vector q to be block sparse.
While the structure of the all 0 columns of As is known,
the structure of the non-zero elements is unknown. This
makes the block structure of the vector q in (15) random
and a function of the number of blockages and their size.
To complete the array diagnosis process, we employ the expanded block sparse Bayesian learning algorithm with bound
optimization (EBSBL-BO) proposed in [35] to recover the
block sparse matrix q. The EBSBL-BO algorithm exploits the
intra-block correlation of the sparse vector to improve recovery
performance without requiring prior knowledge of the block
6
structure. From q, the amplitude and induced phase shifts can
be estimated as shown in Section III-A2.
C. Extension to Complete Group-Blockages
When blockages are complete, i.e. αn = 0, and span
multiple neighboring antennas, the innovation matrix As in
(14) becomes sparse with J groups (or clusters) of non-zero
entries at the locations of the faults as shown in Fig. 4. In this
section, we exploit the structure of these faults and propose a
technique that identifies the locations of these faults with just
Nx + Ny measurements provided that the number of groups is
small, independent of the group size. Recall Nx is the number
of antennas along the x-axis and Ny is the number of antennas
along the y-axis. The idea is to decompose the matrix As into
two dense vectors as shown in Fig. 4. The first vector is a
weighted sum of the rows of As while the second vector is
a weighted sum of the columns of As . Since the number of
unknown is now Nx + Ny , only Nx + Ny measurements are
required to recover the vectors. Once the vectors are recovered,
the intersection of the non-zero elements of both vectors
provides the location of potential faults. Equipped with the
ideal matrix A, the locations of the faults can be refined via an
exhaustive search over all possible locations. In what follows,
we formulate the problem and show how this is performed.
To formulate the problem, we rewrite ideal far-field radiation pattern in (1) as
f (θ, φ) = wT Ap,
(16)
and the damaged pattern in (3) as
g(θ, φ) = wT (A ◦ B)p,
(17)
where w ∈ C Ny ×1 and p ∈ C Nx ×1 represent the receive
antenna weights, and the matrix A ∈ C Ny ×Nx is the antenna
response matrix. Substituting (16) and (17) in (5)-(6), the
kth received measurement (after subtracting it form the ideal
pattern) becomes
yk = wkT Apk − wkT (A ◦ B)pk + ẽk = wkT As pk + ẽk , (18)
To start the diagnosis process, the receiver obtains Ny measurements using the weighting vector w while fixing the
weighting vector p. After Ny measurements, the receiver
obtains Nx measurements using the weighting vector p while
fixing the weighting vector w. Mathematically, the innovation
vector can be written as
y1 = w1T As p0 + ẽ1
..
.. ..
.
. .
yNy
T
= wN
As p0 + ẽNy
y
yNy +1 = w0T As p1 + ẽNy +1
..
.. ..
.
. .
yNy +Nx
= w0T As pNx + ẽNy +Nx ,
where the weighting vectors w0 ∈ C Ny ×1 and p0 ∈ C Nx ×1
consist of random weighting entries and are fixed throughout
the diagnosis stage. Note that the term As p0 represents a
weighted sum of all the columns of matrix As , and the term
w0T As represents a weighted sum of the rows of As (see Fig.
(4)). To simplify above system of equations, let x1 = As p0 ,
x2 = (w0T As )T , and the matrix Φ be
W 01
Φ=
,
(19)
02 P
where the weighting matrices W ∈ C Ny ×Ny and P ∈ C Nx ×Nx
are both orthonormal matrices, the matrix 01 is an all zero
matrix of size Ny × Nx , the matrix 02 is an all zero matrix of
size Nx × Ny . Then, the innovation vector can be simplified
to
ẽ1
y1
x1
W 01
..
..
+
, (20)
= 0
.
.
x2
P
2
{z
} | {z }
|
ẽNy +Nx
yNy +Nx
Φ
x
or equivalently
y = Φx + e.
(21)
To recover the locations of the blockages, we estimate the
vectors x1 and x2 in (20) as follows
x̂ = Φ∗ y = Φ∗ Φx + Φ∗ e = x + é,
(22)
where x̂ is a noisy estimate of x, the estimates x̂1 = [x̂]1:Ny
and x̂2 = [x̂](Ny +1):(Ny +Nx ) . The intersection of the indices of
the non-zero elements in x̂1 and x̂1 correspond to potential
blocked/fault locations.
Using x̂1 and x̂2 , we form the approximate binary matrix
B̃ as follows
1, if [x1 ]m and [x2 ]n are non-zero
[B̃]m,n =
(23)
0, otherwise.
The binary matrix B̃ can be refined as follows
(A − (A ◦ D))p0
B̂ = arg min x̂ −
(w0T (A − (A ◦ D)))T
D∈{0,1}
, (24)
2
where the matrix D is a submatrix of the matrix [B̃]. The zero
entries of B̂ correspond to the locations of the blocked/faulty
antennas.
Remarks:
1) For the special case of a single group blockage, i.e. J =
1, the zero entries of the matrix B̃ in (23) correspond to
the locations of the faulty antennas and the search step
in (24) is not required.
2) When the number of groups increases, the computational
complexity in (24) increases and as a result, conventional
compressed sensing recovery techniques might be more
favorable in this case.
3) For large group sizes, the matrix As becomes dense and
compressed sensing recovery techniques may fail in this
case. The technique proposed in this section is able to
identify the locations of the faulty antenna elements with
just Ny +Nx measurements. Even if compressed sensing
recovery techniques succeed, the required number of
measurements would be much higher than Ny + Nx .
4) When blockages are partial, the entries of the matrix D
in (24) would be complex instead of binary numbers,
7
weighting matrix at the transmitter, and E is the additive
noise matrix. Equipped with the weighting matrices F and
W, the receiver generates the ideal vector YI = W∗ AI F,
where AI = a(θ, φ)a∗T (θ, φ), and subtracts it from Yr in (26)
to obtain
Y = Yr − YI = W∗ As F + E,
(27)
where As ∈ CNR ×NT is a sparse matrix. The non-zero entries
of the columns of As represent the indices of faulty antennas
at the transmitter and the non-zero entries of the rows of As
represent the entries of the faulty antennas at the receiver.
To formulate the CS recovery problem, we vectorize the
measurement matrix Y in (27) to obtain
Fig. 4: An example of an innovation matrix (i.e. As = A − A ◦ B)
of a 128 element antenna with 19 blocked/faulty elements. The
equivalent dense vectors are obtained by summing the rows and
columns of the matrix As . Intersection of the indices of the nonzero elements (of the dense vectors) correspond to the location of
potentially blocked/faulty elements.
and the complexity of the exhaustive search step would
be high if not prohibitive. A two-stage setting where
the entries of B̃ are independently examined might be
favorable in this case.
IV. J OINT FAULT D ETECTION AT THE T RANSMITTER AND
THE R ECEIVER
In the previous section, we assumed that the transmit
antenna is free from blockages. When blockages exist at
the transmit antenna, the receiver receives distorted training
symbols and as a result, the technique proposed in the previous
section will fail. In this section, we propose a detection
technique that jointly detects blockages at both the transmitter and the receiver. For this technique, we assume that
the receiver is equipped with the path-loss, angular location
and array manifold of the transmitter. To start the diagnosis
process, the transmitter sends a training symbol s = 1 using
kt random antenna weights (assumed to be known by the
receiver). The receiver generates kr random antenna weights to
receives each training symbol, thus making the total number of
measurements K = kt kr . Let aT (θ, φ) = atx (θ, φ) ⊗ aty (θ, φ)
be the NT × 1 1D transmit array response vector. The kth
received measurement at the receiver can be written as
yk = wi∗ (b ◦ a(θ, φ))(bT ◦ aT (θ, φ))∗ fo + ẽk ,
(25)
where wi is a random weighting vector at the receiver, fo
is a random weighting vector at the transmitter, and bT is a
vector of complex coefficients that result from the absorption
and scattering caused by the particles blocking the transmit
antenna array. After K measurements the receiver obtains the
following measurement matrix
Yr = W∗ AF + E,
(26)
where Yr is the received measurement matrix, W ∈
CNR ×kr is a random weighting matrix at the receiver, A =
(b ◦ a(θ, φ))(bT ◦ aT (θ, φ))∗ is the NR × NT equivalent array response matrix, the matrix F ∈ CNT ×kt is a random
vec(Y) = vec(W∗ As F) + vec(E)
= (FT ⊗ W∗ ) vec(As ) + vec(E),
{z
} | {z } | {z }
|
U
g
(28)
(29)
e
where the vector y = vec(Y), the matrix U ∈ CK×NT NR is
the effective CS sensing matrix, and the vector g ∈ CNT NR ×1
is the effective sparse vector. Note that if the matrices W
and F in (28) both satisfy the coherence property, examples
include Gaussian, Bernoulli and Fourier matrices [30], then the
matrix U = W ⊗ F also satisfies the coherence property and
it can be applied in standard CS techniques. For simplicity, the
entries of W and F are chosen uniformly and independently at
random from the set {1+j, 1−j, −1+j, −1−j} in this paper.
This corresponds to 2-bit phase shifters at both the transmitter
and the receiver.
A. Sparsity Pattern Detection and Least Squares Estimation
The LASSO estimate of (28) is given by
arg
min
ν∈CNT NR ×1
1
ky − Uνk22 + Ωσe kνk1 ,
2
(30)
where σe is the standard derivation of the noise e. Once the
support S, where S = {i : gi 6= 0}, of the vector g is
estimated, the columns of U which are associated with the
non-zero entries of g are removed to obtain US . Hence, the
vector y in (28) becomes
y = US gS + e,
(31)
where gS is obtained by pruning the zero entries of g. The LS
estimate after successful sparsity pattern recovery becomes
ĝS = (U∗S US )−1 U∗S y = gS + e,
(32)
where ĝS is a noisy estimate of gS .
B. Attenuation and Induced Phase Shift Estimation
Let r be a vector of size NT NR × 1 and its ith entry is
[r]i = [gS ]i in (32), if i ∈ S and ri = 0 otherwise. Reshaping
r into NR rows and NT columns we obtain an estimate Âs
of the sparse matrix As in (27). The non-zero columns of
Âs represent the IDs of the transmit array faulty antennas,
and the non-zero rows of Âs represent the IDs of the receive
array faulty antennas. Let the set Ir contain the indices of the
zero rows of Âs and the set It contain the indices of the zero
8
columns of Âs . Removing the rows associated with the IDs
of the faulty receive antennas from Âs we obtain the matrix
Ar = [Âs ]Ir ,: , and the sparse vector that represents the faulty
transmit antennas becomes
(ar (θ, φ)∗ Ar )∗
,
(33)
q̂t =
kar (θ, φ)k22
where the vector ar (θ, φ) results from selecting the Ir entries
from the vector a(θ, φ) in (25). Based on (33), the estimated
attenuation coefficient and induced phase of the ith antenna
t ]i
element at the transmit array becomes κ̂t,i = | [aT[q̂
(θ,φ)]i + 1|
t ]i
and Φ̂t,i = ∠ [aT[q̂
(θ,φ)]i + 1 .
Similarly, removing the columns associated with the IDs
of the faulty antennas from Âs we obtain the matrix At =
[Âs ]:,It , and the sparse vector that represents the faulty receive
antennas becomes
At at (θ, φ)
,
(34)
q̂r =
kat (θ, φ)k22
where the vector at (θ, φ) results from selecting the It entries
from the vector aT (θ, φ) in (25). From (34), the estimated
attenuation coefficient and induced phase of the ith antenna
[q̂r ]i
element at the receive array becomes κ̂r,i = | [a(θ,φ)]
+ 1| and
i
[q̂r ]i
Φ̂r,i = ∠ [a(θ,φ)]i + 1 .
V. N UMERICAL VALIDATION
In this section, we conduct numerical simulations to evaluate the performance of the proposed techniques. We consider
d
a 2D planar array, with dλx = λy = 0.5, that experiences
random and independent blockages with probability Pb . To
generate the random blockages, the values of κ and Φ in (4) are
chosen uniformly and independently at random from the set
{i ∈ R : 0 ≤ i ≤ 1} and {0, .., 2π} respectively. We adopt the
success probability, i.e. the probability that all faulty antennas
are detected, and the normalized mean square error (NMSE) as
a performance measure to quantify the error in detecting the
blocked antenna locations and estimating the corresponding
blockage coefficients (κ and Φ). The NMSE is defined by
NMSE =
kv − v̂k22
.
kvk22
(35)
When blockages only exist at the receiver, v = c ◦ a(θ, φ)
(see (6)), and the ith entry of the estimated vector v̂ is
[v̂]i∈S = [q̂]i∈S in (10) and zero otherwise. When blockages
exist at both the receiver and the transmitter, v = (bT ◦aT )−aT
(see (25) and (27)) and v̂ = q̂t in (33) when detecting
blockages at the transmitter array, and v = (b ◦ a) − a and
v̂ = qˆr in (34) when detecting blockages at the receiver array.
To implement the LASSO, we use the function SolveLasso
included in the SparseLab toolbox [36]. As a benchmark,
we compare the NMSE of the proposed techniques with the
NMSE of the Genie aided LS estimate which indicates the
optimal estimation performance when the exact locations of
the faulty antennas is known, i.e., the support S in (10) and
(32) is assumed to be provided by a Genie.
In Figs. 5-8 we consider faults at the receiver antenna array
and assume that the transmitter antenna array is fault free. To
study the effect of the number of measurements (or diagnosis
time) on the performance of the proposed fault detection technique in Section III-A, we plot the NMSE when the receive
array is subject to blockages with different probabilities in Fig.
5a. For all cases, we observe that the NMSE decreases with
increasing number of measurements K. The figure also shows
that for sufficient number of measurements (on the order of
K ∼ O(Pb NR log NR )), the NMSE of the proposed technique
matches the NMSE obtained by the Genie-aided LS technique.
This indicates for sufficient number of measurements, the
proposed technique successfully detects the locations of the
blocked antennas and the corresponding blockage coefficients
with K NR measurements. The figure also shows that as
the blockage probability increases, more measurements are
required to reduce the NMSE. The reason for this is that as the
blockage probability increases, the average number of blocked
antennas increases as well. Therefore, more measurements are
required to estimate the locations of the blocked antennas
and the corresponding blockage coefficients. In the event of
large number of blockages or fast varying blockages, a hybrid
antenna architecture with a few RF chains can be adopted to
reduce the array diagnosis time.
For comparison, we plot the performance of the CS-based
diagnosis technique proposed in [19] which requires measurements to be randomly taken at NR locations, in Figs. 5a and 5b.
This technique is chosen as it is based on analog beamforming,
and therefore, it can be applied to mmWave systems. For
both Pb = 0.01 and Pb = 0.1, the proposed technique
provides a lower NMSE while requiring lower number of
measurements. For instance, to obtain a target NMSE of 30dB with Pb = 0.01, the proposed technique requires 45
measurements while the algorithm proposed in [19] requires
110 measurements, taken at 110 independent locations. This
makes the proposed algorithm superior as it is able to obtain
lower NMSE with lower number of measurements without the
need for taking measurements at multiple locations.
To examine the effect of the array size on the number of
measurements, we plot the NMSE of the proposed technique
with NR = 512 in Fig. 5b. The figure shows that the
NMSE decreases with increasing number of measurements
K. Nonetheless, this decrease occurs at a lower rate when
compared to the case when NR = 256 in Fig. 5a. This is
particularly observed for higher blockage probabilities. The
reason for this is that as the array size increases, the average
number of blocked antennas increases as well. Therefore, more
measurements are required to estimate the locations of the
blocked antennas and the corresponding blockage coefficients.
Similar to the case when NR = 256, the proposed technique
results in lower NMSE with lower diagnosis time when
compared to [19]. This is mainly due to the CS sensing matrix
which in this paper is optimized to satisfy the coherence
property, thereby requiring lower number of measurements.
The effect of the SNR on the performance of the proposed
technique is shown in Fig. 6. The figure shows that for
sufficient number of measurements, the NMSE obtained by
the proposed technique approaches the NMSE obtained by the
Genie-aided techniques. The figure also shows that required
number of measurements is a function of the receive SNR. For
9
10
5
CS [19] + LS; Pb = 0.1
CS [19] + LS; Pb = 0.1
0
Genie-aided LS, P b = 0.1
Genie-aided LS, P b = 0.1
CS [19] + LS; Pb = 0.01
0
CS [19] + LS; Pb = 0.01
-5
Proposed, P b = 0.1
5
Proposed, P b = 0.1
Proposed, P b = 0.01
Proposed, P b = 0.01
-10
NMSE (dB)
NMSE (dB)
-10
Genie-aided LS, P b = 0.01
-5
Genie-aided LS, P b = 0.01
-15
-20
-15
-20
-25
-25
-30
-30
-35
-35
-40
20
40
60
80
100
120
140
160
180
-40
20
200
Number of measurements
40
60
80
100
120
140
160
180
200
Number of measurements
(b) NR = 16 × 32 = 512.
(a) NR = 16 × 16 = 256.
Fig. 5: Detection and estimation of faults in a 2D receive planar array subject to random partial blockages with different blockage probability
Pb ; ρ = 10 dB, and NR = 16 × 16 = 256. Blockages do not occur in groups and the technique proposed in Section III-A is used for
diagnosis.
5
5
∆θ~ U(-1,1) , ∆φ ~ U(-1,1)
∆θ~ U(-0.5,0.5), ∆φ ~ U(-0.5,0.5)
∆θ~ U(-0.25,0.25), ∆φ ~ U(-0.25,0.25)
∆θ = 0, ∆φ =0
Proposed
0
Genie-aided LS
0
-5
-5
-10
-15
NMSE (dB)
NMSE (dB)
-10
SNR = 5, 10, 15 dB
-20
-15
-20
-25
-25
-30
-30
-35
-35
-40
-45
20
40
60
80
100
120
140
160
180
200
Number of measurements
-40
20
40
60
80
100
120
140
160
180
200
Number of measurements
Fig. 6: Detection and estimation of faults in a 2D receive planar
array subjected to random partial blockages with different receiver
SNR; Pb = 0.1, and NR = 16 × 16 = 256. Blockages do not occur
in groups and the technique proposed in Section III-A is used for
diagnosis.
Fig. 7: Detection and estimation of faults in a 2D receive planar
array subject to random partial blockages with AoD estimation errors;
Pb = 0.1; ρ = 10 dB, and NR = 16 × 16 = 256. Blockages do not
occur in groups and the technique proposed in Section III-A is used
for diagnosis. Dashed lines represent Gene aided LS estimation.
instance, for a receive SNR of 15 dB and 120 measurements,
the NMSE of the proposed technique is similar to the NMSE
obtained by the Genie-aided technique. As the SNR decreases
to 5 dB, more than 200 measurements are required to match
the NMSE of the Genie aided technique. To reduce the number
of measurements, one can increase the receive SNR by either
placing more antennas at the transmitter to increase the array
gain or reduce the transmitter-receiver distance to minimize
the path-loss.
To investigate the effect of imperfect AoD/AoA estimation
on the NMSE performance, we plot the NMSE of the proposed
technique in Fig. 7 under the assumption of imperfect azimuth
and elevation angles of departure, i.e. φ = φ + ∆φ, and
θ = θ + ∆θ, where ∆φ and ∆θ are uniformly distributed
random variables. Fig. 7 shows that angular deviations (as
small as ±0.25◦ ) result in a 10 dB loss in NMSE performance
and this loss increases with larger angular deviations. The
reason for this NMSE increase is due to the AoD/AoA
mismatch which increases the system noise. Fig. 7 also shows
that for sufficient number of measurements, the NMSE of
the proposed technique matches the NMSE of the Gene-aided
LS estimation technique (which assumes perfect knowledge
of the location of blockages). This suggests that majority of
the NMSE results from blockage coefficient estimation errors
rather than blockage location detection errors.
In Fig. 8 we study the impact of random multipath on the
NMSE performance. Similar to the imperfect AoD/AoA case,
multipath introduces interference which increases the noise
10
10
(a) Location and intensity of blockage coefficients; |1-B|
Proposed: multi-path
Gene aided LS: multi-path
Proposed: no multi-path
Genie-aided LS: no multi-path
5
0.8
5
0
y
0.6
10
-5
0.2
-10
NMSE (dB)
0.4
15
-15
0
2
4
6
8
10
12
14
16
x
(b) Reconstruction error in dB using conventional CS
-20
-10
-25
5
-15
y
-30
-35
10
-40
20
15
-20
-25
40
60
80
100
120
140
160
180
Number of measurements
4
6
8
10
12
14
x
(c) Reconstruction error in dB using Block CS
16
-6
-8
5
-10
-12
y
Fig. 8: Detection and estimation of faults in a 2D receive planar array
subject to random partial blockages in the presence of random indirect
(or scattered) path; Pb = 0.1; ρ = 10 dB, NR = 16 × 16 = 256.
Total number of path = 3 (one direct path plus two random path with
random delays uniformly distributed across the diagnosis period), and
90% of the energy is located in the direct path.
-30
2
-14
10
-16
-18
15
2
4
6
8
10
12
14
16
x
floor of the system and, as a result, deteriorate the NMSE
performance. We consider a direct path and 2 multipath with
random delays and gains. For this setup, 90% of the received
signal energy is contained in the direct path and the scattered
path delays are uniformly distributed across the diagnosis time.
As expected, Fig. 8 shows that the NMSE is lower in the
presence of multipath and the NMSE does not decrease with
increasing number of measurements. This is mainly attributed
to multipath interference which is treated as noise in this paper.
In Figs. 9-12 we consider blockages that span a group of
neighboring antennas. More specifically, in Fig. 9 we consider
blockages that affect 16 neighboring antennas. To model the
random shapes of the blocking particles, we assume that each
antenna element within a block experiences a random blockage
intensity, i.e. the values of κ and Φ in (4) are chosen uniformly
and independently at random from the set {i ∈ R : 0 ≤ i ≤ 1}
and {0, .., 2π}. In Fig. 9(a), we plot an example of a 256
element array subject to 2 blockages and each blockage affects
a group of 16 antennas with random intensities. In Fig. 9(b),
we plot the reconstruction error when the proposed technique
is used to detect and estimate the blockage coefficients without
exploiting the block structure of blockages. As shown, the
proposed technique, using conventional CS, identifies locations of the blocked antennas, however, it also results in some
false positives which can be reduced by increasing the number
of measurements. In Fig. 9(c), we plot the reconstruction
error when we exploit the block structure of blockages to
detected and estimate the blockage coefficients. As shown,
this technique leverages the dependencies between the values
and locations of the blockages in its recovery process and as
a result it minimizes false positives and reduces the number
of required measurements.
To highlight the benefit of exploiting correlation between the
Fig. 9: Comparison of the proposed array diagnosis techniques in
detecting antenna faults in an NR = 16 × 16 = 256 element antenna;
(a) Location and intensity of blockages incident on the array (|1−B|,
where 1 is an all-ones matrix). (b) Reconstruction error (|B − B̂|)
in dB when ignoring the block structure. The reconstructed matrix
is denoted by B̂. (c) Reconstruction error when exploiting the block
structure of blockages using EBSBL-BO+LS. For (b) and (c), ρ =
10 dB, number of blocks is 2, block size J = 16, and the number of
measurements are fixed to 130.
blocked antenna elements, we compare the NMSE achieved
when (i) ignoring the block structure of the blockages and
using conventional CS recovery, and (ii) exploiting the block
structure of blockages by implementing the technique proposed in Section III-B in Fig. 10 and Fig. 11. Specifically,
we consider random blockages and each blockage spans Γ
antennas with random blockage intensity. For array diagnosis,
we plot the performance of the proposed techniques with
and without exploiting the block structure. For Γ = 32 and
J = 1, Fig. 10 shows that lower NMSE can be achieved
when exploiting the block structure. For fixed number of
blockages, Fig. 10 shows that the performance gap decreases
with decreasing block size Γ. As the block size decreases, the
correlation between the blocked antenna elements diminishes
and hence the performance of this technique comes closer to
that achieved by techniques that ignore the block structure
of blockages. In Fig. 11 we compare the performance of the
proposed techniques when exploiting and ignoring the block
structure of blockages for different receive SNRs. For all cases,
the plots show a clear performance benefit when exploiting the
block structure in the array diagnosis process.
11
0
NMSE (dB)
-5
-10
-15
-20
-25
40
LASSO, S=32x1
EBSBL-B0, S=4x8
EBSBL-B0, S=16x2
EBSBL-B0, S=32x1
50
60
70
80
90
100
Number of measurements
Fig. 10: Comparison of the proposed techniques in detecting faults in
a 2D receive planar array subjected to random group blockages. The
proposed technique that ignores the block structure uses conventional
LASSO as a recovery method, while the technique that exploits
the block structure uses the EBSBL-B0 [35] algorithm for recovery.
Blockages affect a group of S = Γ × J antenna elements, where Γ is
the group size and J is the number of groups. Each antenna element
within a group experiences a random blockage intensity; ρ = 10 dB,
NR = 16 × 16 = 256.
0
NMSE (dB)
-5
-10
-15
-20
-25
40
LASSO; S=16x2, SNR = 0dB
LASSO; S=16x2, SNR = 10dB
EBSBL-B0; S=16x2, SNR = 0dB
EBSBL-B0; S=16x2, SNR = 10dB
50
60
70
80
90
100
Number of measurements
Fig. 11: Comparison of the proposed techniques in detecting faults
in a 2D receive planar array subjected to random group blockages
and different noise levels. Each antenna element within a group
experiences a random blockage intensity; NR = 16 × 16 = 256.
In Figs. 12a and 12b, we consider complete blockages and
compare the performance of the technique proposed in Section
III-C with conventional CS recovery techniques (LASSO) for
fixed number of measurements K = 16. Fig. 12a shows that
the proposed technique achieves higher success probability
when compared to conventional CS recovery techniques. The
success probability of the proposed technique is highest when
the number of blocks is J = 1. When the number of blocks increases, however, the search space increases (see (24)), and as
a result, we observe a performance hit especially at low SNR.
In Fig. 12b we fix the total number of blockages and study
the impact of the block size on the success probability. Fig.
12b shows that for fixed number of blockages, higher success
probability is achieved for larger block sizes and the success
probability decreases with increasing number of groups. As the
number of groups increases, the search space in (24) increases,
and in the presence of noise, false alarms could occur. This
reduces the success probability. Note that in Figs. 12a and 12b,
conventional CS techniques fails since the required number of
required measurements is K > S log Ny Nx which is much
more than Ny + Nx .
In Figs. 13-14, we assume that both the transmit and
receive arrays are subject to random blockages. To analyze
the effect of the number of measurements (or diagnosis time)
on the NMSE performance when jointly detecting faults on
both the transmit and receive arrays, we plot the NMSE of
the technique proposed in Section IV and the Genie-aided
technique in Fig. 13. For both cases, we observe that the
NMSE decreases with increasing number of measurements K,
and for sufficient K, the NMSE of the proposed technique
matches the NMSE obtained by the Genie-aided technique.
The figure also shows that the NMSE of the receive array
is lower than the NMSE of the transmit array. This is due to
the transmit-receive array size difference. For a fixed blockage
probability Pb , larger array sizes encounter more blockages on
average, and therefore, require more measurements to detect
faults.
The success probability when detecting faulty antennas in
both the transmit and the receive arrays is plotted in Fig. 14.
Fig. 14 (a) shows that for both the transmit and receive array,
the success probability increases with increasing number of
measurements. Fig. 14 (b) shows that lower number of measurements are required to detect faulty antennas at the receiver
when the transmit array size is larger than the receive array
size. As the transmit array size increases, the average number
of blocked antennas increases as well, thereby requiring more
measurements to successfully detect the fault locations.
VI. C ONCLUSIONS
In this paper, we investigated the effects of blockages on
mmWave linear antenna arrays. We showed that both complete
and partial blockages distort the far-field beam pattern of a
linear array and partial blockages result in a higher beam
pattern variance when compared to complete blockages. To
detect blockages, we proposed several compressed sensing
based array diagnosis techniques. The proposed techniques
do not require the AUT to be physically removed and do
not require any hardware modification. When faults exist at
the receiver only, we showed that the proposed techniques
reliably detect the locations of the blocked antennas, if any,
and estimate the corresponding attenuation and phase-shift
coefficients caused by the blocking particles. Moreover, we
showed that the dependencies between the blocked antennas
can be exploited to further reduce estimation errors and lower
the diagnosis time. When faults exist at both the receiver and
the transmiter, we showed that reliable detection and estimatation of blockages can be achievded. Nonetheless, high number
12
1
1
S=4x1
S=4x2
S=4x4
S=4x1; LASSO
S=4x2; LASSO
S=4x4; LASSO
Success Probability
0.8
0.7
0.8
Exploiting
2D/block structure
0.6
S=16x2
S=8x4
S=4x8
S=16x2; LASSO
0.9
Success Probability
0.9
0.5
0.4
0.3
0.2
0.7
Exploiting 2D/block
structure
0.6
0.5
0.4
0.3
0.2
Conventional CS
0.1
0
-20
0.1
-15
-10
-5
0
SNR (dB)
5
10
Conventional CS
0
-20
15
-15
-10
-5
0
5
10
15
SNR (dB)
(a) Fixed block-size.
(b) Variable block size.
Fig. 12: Detection of faults in a 2D receive planar array subjected to complete (non-partial) blockages that occur in groups as a function of
the receive SNR; K = 16 measurements, and NR = 8 × 8 = 64.
1
5
Proposed (transmit array)
Genie-aided LS (transmit array)
Proposed (receive array)
Genie-aided LS (receive array)
0
(a) N T /N R = 1
1
0.9
0.9
-5
(b) NT /N R = 4
0.8
0.8
-10
0.7
NMSE (dB)
-15
-20
-25
Success probability
Success probability
0.7
0.6
0.5
0.6
0.5
0.4
0.4
-30
0.3
0.3
-35
0.2
Receive array
0.2
-40
Receive array
0.1
Transmit array
-45
100
150
200
250
300
350
400
450
Number of measurements
Fig. 13: Joint detection of faults in 2D transmit and receive planar
arrays both subject to random partial blockages; Pb = 0.1 for both
arrays, ρ = 0 dB, NR = 16, NT = 32, and NT NR = 512.
of measurements are required in this case, even if compressed
sensing is used. For all cases, the estimated coefficients can be
used to calculate new antenna excitations weights to recalibrate
the transmit/receive antennas. Due to their reliability and
low diagnosis time, the proposed techniques can be used to
perform real-time mmWave antenna array diagnosis at the
receiver and enhance the mmWave communication link.
R EFERENCES
[1] F. Boccardi, R. Heath Jr., A. Lozano, T. Marzetta, and P. Popovski, “Five
disruptive technology directions for 5G,” IEEE Commun. Mag., vol. 52,
no. 2, pp. 74-80, Feb. 2014.
[2] T. Rappaport, R. Heath Jr., R. Daniels, and J. Murdock, Millimeter wave
wireless communications. Pearson Education, 2014.
[3] M. Eltayeb, T. Al-Naffouri, and R. Heath Jr., “Compressive sensing
for vehicular millimeter wave antenna array diagnosis,” in IEEE Global
Commun. Conference, Washington, DC USA, Dec. 2016 pp. 1-6.
0.1
200
300
400
Number of measurements
Transmit array
500
0
200
300
400
500
Number of measurements
Fig. 14: Joint detection of faults in 2D transmit and receive planar
arrays all subject to constant blockages; Pb = 0.1 for both arrays,
[bT ]i ∈ {0, 1}, [b]i , ∈ {0, 1}, and NT NR = 1024. In (a) NR = 32,
and NT = 32, and in (b) NR = 16, and NT = 64.
[4] S. Dudzinsky, Jr., Atmospheric effects on terrestrial millimeter-wave
communications, The Rand Corporation, R-1335-ARPA, March 1974.
[5] Y. Kuga, F. Ulaby, T. Haddock, and R. DeRoo, “Millimeter-wave radar
scattering from snow: 1. Radiative transfer model”, Radio Sci., vol. 26,
no. 2, pp. 329-341, 1991.
[6] A. Arage, W. Steffens, G. Kuehnle, and R. Jakoby, “Effects of water and
ice layer on automotive radar” in Proc. of the German Microwave Conf.,
2006.
[7] E. Vinyaikin, M. Zinicheva, and A. Naumov, “Attenuation and phase
variation of millimeter and centimeter radio waves in a medium consisting
of dry and wet dust particles,” Radiophysics and Quantum Electronics,
vol. 37, no. 11, 1994.
[8] F. Ulaby, T. Haddock, R. Austin, and Y. Kuga, “Millimeter-wave radar
scattering from snow: 2. Comparison of theory with experimental observations”, Radio Sci., vol. 26, no. 2, pp. 343-351, 1991.
[9] M. Li and Y. Lu, “Source bearing and steering-vector estimation using
partially calibrated arrays,” IEEE Trans. Aerosp. Electron. Syst., vol. 45,
pp. 1361-1372, Oct. 2009.
13
[10] R. Mailloux “Array failure correction with a digitally beamformed
array,” IEEE Trans. Antennas Propag., vol. 44, no. 12, pp. 1543-1550,
Dec. 1996.
[11] M. Li, M. McGuire, K. S. Ho and G. Hayward, “Array element failure
correction for robust ultrasound beamforming and imaging,” in IEEE
Ultrasonics Symposium, San Diego, CA, 2010, pp. 29-32.
[12] B.-K. Yeo and Y. Lu, “Array failure correction with a genetic algorithm,”
IEEE Trans. Antennas Propag., vol. 47, no. 5, pp. 823-828, 1999.
[13] J. Rodrguez, F. Ares, E. Moreno, and G. Franceschetti, “Genetic algorithm procedure for linear array failure correction,” Electronics Letters,
vol. 36, no. 3, pp. 196-198, 2000.
[14] R. Iglesias, F. Ares, M. Delgado, J. Rodriguez, J. Bregains, and S.
Barro, “Element failure detection in linear antenna arrays using casebased reasoning,” IEEE Antennas Propag. Mag., vol. 50, no. 4, pp. 198204, Aug. 2008.
[15] J. Rodriguez, F. Ares, E. Moreno, and G. Franceschetti, “Genetic
algorithm procedure for linear array failure correction,” Electron. Lett.,
vol. 36, no. 3, pp. 196-198, Feb. 2000.
[16] O. Bucci, M. Migliore, G. Panariello, and G. Sgambato, “Accurate
diagnosis of conformal arrays from near-field data using the matrix
method,” IEEE Trans. Antennas Propag., vol. 53, no. 3, pp. 1114-1120,
Mar. 2005.
[17] J. R.-Gonzalez, F. A.-Pena, M. F.-Delgado, R. Iglesias, and S. Barro,
“Rapid method for finding faulty elements in antenna arrays using far
field pattern samples,” IEEE Trans. Antennas Propag., vol. 57, no. 6, pp.
1679-1683, June 2009.
[18] A. Buonanno and M. D’Urso, “On the diagnosis of arbitrary geometry
fully active arrays,” in the Eur. Conf. Antennas Propag. (EuCAP),
Barcelona, Apr. 2010.
[19] M. Migliore, “A compressed sensing approach for array diagnosis from
a small set of near-field measurements,“ IEEE Trans. Antennas Propag.,
vol. 59, no. 6, pp. 2127-2133, June 2011.
[20] G. Oliveri, P. Rocca, and A. Massa, “Reliable diagnosis of large linear
arrays-Bayesian compressive sensing approach,” IEEE Trans. Antennas
Propag., vol. 60, no. 10, pp. 4627-4636, Oct. 2012.
[21] B. Friedlander and T. Strohmer, “Bilinear compressed sensing for array
self-calibration,” in 48th Asilomar Conference on Signals, Systems and
Computers, Pacific Grove, CA, 2014, pp. 363-367.
[22] C. Balanis, Antenna Theory, Analysis, and Design, 1997, Wiley.
[23] M. Eltayeb, T. Al-Naffouri, and R. Heath, “Compressive
sensing for millimeter wave antenna array diagnosis,” archived,
https://arxiv.org/abs/1612.06345, arXiv:1612.06345v3.
[24] A. Alkhateeb, O. Ayach, and R. Heath Jr., “Channel estimation and
hybrid precoding for millimeter wave cellular systems,” IEEE J. on Select.
Areas in Commun., vol. 8, no. 5, pp. 831-846, Oct. 2014.
[25] M. Eltayeb, A. Alkhateeb, R. Heath Jr., and T. Y. Al-Naffouri, “Opportunistic beam training with hybrid analog/digital codebooks for mmwave
systems,” in Proc. of IEEE Global Conf. on Signal and Info. Proc., Dec.
2015.
[26] V. Va, J. Choi, and R. Heath, “The impact of beamwidth on temporal
channel variation in vehicular channels and its implications,” IEEE Trans.
Veh. Technol., vol. 66, no. 6, pp. 5014-5029, June 2017.
[27] D. Chizhik, “Slowing the time-fluctuating MIMO channel by beamforming,” IEEE TWC, vol. 3, no. 5, pp. 1554-1565, Sep. 2004.
[28] A. Ali, N. Gonzlez-Prelcic and R. Heath Jr., “Estimating Millimeter
Wave Channels using Out-of-Band Measurements”, in Proceedings of
Information Theory and Applications (ITA) Workshop, 2016.
[29] M. Akdeniz, Y. Liu, M. Samimi, S. Sun, S. Rangan, T. Rappaport,
and E. Erkip, “Millimeter wave channel modeling and cellular capacity
evaluation,” IEEE J. on Selected Areas in Commun., vol. 32, no. 6, pp.
1164-1179, June 2014.
[30] M. Wainwright, “Sharp thresholds for high-dimensional and noisy
sparsity recovery using `1 -constrained quadratic programming (Lasso),”
IEEE Trans. Inf. Theory, vol. 55, no. 5, pp. 2183-2202, May 2009.
[31] E. Candes and Y. Plan, “Near-ideal model selection by l1 minimization,”
Ann. Statist., vol. 37, pp. 2145-2177, 2009.
[32] A. Fletcher, S. Rangan, and V. Goyal, “Necessary and sufficient conditions for sparsity pattern recovery,” IEEE Trans. Inform. Theory, vol. 55,
no. 12, pp. 5758-5772, Dec. 2009.
[33] M. Rudelson and R. Vershynin, “On sparse reconstruction from fourier
and gaussian measurements,” Communications on Pure and Applied
Mathematics, vol. 61, pp. 1025-1045, 2008.
[34] A. Sayed, Fundamentals of adaptive filtering, John Wiley & Sons, 2003.
[35] Z. Zhang and B. D. Rao, “Extension of SBL Algorithms for the
Recovery of Block Sparse Signals With Intra-Block Correlation,” IEEE
Trans. Sig. Process., vol. 61, no. 8, pp. 2009-2015, April, 2013.
[36] http://sparselab.stanford.edu.
| 7cs.IT
|
arXiv:1506.05607v5 [cs.SY] 22 Aug 2017
1
Unbounded-Time Analysis of Guarded LTI
Systems with Inputs by Abstract
Acceleration
Dario Cattaruzza Alessandro Abate Peter Schrammel Daniel Kroening
Abstract—Reachability analysis of continuous and
discrete time systems is a hard problem that has
seen much progress in the last decades. In many
cases the problem has been reduced to bisimulations
with a number of limitations in the nature of the
dynamics, soundness, or time horizon. In this article
we focus on sound safety verification of UnboundedTime Linear Time-Invariant (LTI) systems with inputs
using reachability analysis. We achieve this by using
Abstract Acceleration, which over-approximates the
reach tube of a system over unbounded time by using
abstraction . The technique is applied to a number of
models and the results show good performance when
compared to state-of-the-art tools.
Reachability in linear programs, however, is a
formidable challenge for automatic analysers: the
problem is undecidable despite the restriction to
linear transformations (i.e., linear dynamics) and
linear guards.
The goal of this article is to push the frontiers
of unbounded-time reachability analysis: we aim at
devising a method that is able to reason soundly
about unbounded trajectories. We present a new
approach for performing abstract acceleration. Abstract acceleration [26], [27], [34] approximates the
effect of an arbitrary number of loop iterations (up to
I. Introduction
Linear loops are an ubiquitous programming pat-
infinity) with a single, non-iterative transfer function
that is applied to the entry state of the loop (i.e., to
tern. Linear loops iterate over continuous variables, the set of initial conditions of the linear dynamics).
which are updated with a linear transformation. Lin- This article extends the work in [34] to systems
ear loops may be guarded, i.e., terminate if a given with non-deterministic inputs elaborating the details
linear condition holds. Inputs from the environment
can be modelled by means of non-deterministic
choices within the loop. These features make linear
loops expressive enough to capture the dynamics of
omitted in [39].
The key contributions of this article are:
many hybrid dynamical models. The usage of such
models in safety-critical embedded systems makes
linear loops a fundamental target for formal methods.
Many high-level requirements for embedded con-
1) We present a new technique to include inputs
(non-determinism) in the abstract acceleration
trol systems can be modelled as safety properties, i.e.
deciding reachability of certain bad states, in which
of general linear loops.
2) We introduce the use of support functions in
the system exhibits unsafe behaviour. Bad states may,
in linear loops, be encompassed by guard assertions.
complex spaces, in order to increase the precision of previous abstract acceleration methods.
2
Since the transformations A and B are linear, and
vector sums preserve convexity, the sets Xn =
II. Preliminaries
A. Linear Loops with Inputs
Simple linear loops are programs expressed in the
form:
while(Gx ≤ h) x := Ax + Bu,
ψ := Gx ≤ h is a linear constraint on the states (with
G ∈ Rr×p and h ∈ Rr ), u ∈ Rq is a non-deterministic
input, and A ∈ R p×p and B ∈ R p×q are linear transformations characterising the dynamics of the system.
In particular, the special instance where ψ = ⊤ (i.e.,
“while true”) represents a time-unbounded loop with
no guards, for which the discovery of a suitable
invariant (when existing) is paramount. As evident
at a semantical level, this syntax can be interpreted
as the dynamics of a discrete-time LTI model with
inputs, under the presence of a guard set which, for
ease of notation, we denote as G = {x | Gx ≤ h}.
In the remaining of this work we will also use the
notation Mi,∗ to represent the rows of a matrix and
M∗, j its columns.
B. Model Semantics
The traces of the model starting from an initial
set X0 ⊆ R p , with inputs restricted to U ⊆ Rq , are
u0
u1
u2
sequences x0 −→ x1 −→ x2 −→ . . ., where x0 ∈ X0
and ∀k ≥ 0, xk+1 = τ(xk , uk ), where
as the union of the reachable sets over n iterations.
S
Moreover, X̂ = n≥0 τn (X0 , U) extends the previous
notion over an unbounded time horizon.
C. Support Functions
1) Support Function Definition: A support function is a convex function on R p which describes
the distance of a supporting hyperplane for a given
geometrical set in R p .
Support functions may be used to describe a set by
defining the distance of its convex hull with respect
to the origin, given a number of directions. More
specifically, the distance from the origin to the hyperplane that is orthogonal to the given direction and
that touches its convex hull at its farthest. Finitely
sampled support functions are template polyhedra in
which the directions are not fixed, which helps avoiding wrapping effects [25]. The larger the number of
directions provided, the more precisely represented
the set will be.
In more detail, given a direction v ∈ R p , the
support function of a non-empty set X ⊆ R p in the
direction of v is defined as
ρX : R p → R,
where x · v =
vectors.
states reached from X by τ in one step:
(2)
We furthermore denote the set of states reached from
X0 via τ in n steps (n-reach set), for n ≥ 0:
0
τ (X0 , U) = X0
n
(4)
(1)
We extend the notation above to convex sets of
states and inputs (X and U), and denote the set of
τ(X, U) = {τ(x, u) | x ∈ X, u ∈ U}
We define the n-reach tube
[
X̂n = τ̂n (X0 , U) =
τk (X0 , U)
k∈[0,n]
where x ∈ R p is a valuation on the state variables,
τ(xk , uk ) = { Axk + Buk | Gxk ≤ h ∧ uk ∈ U}
τn (X0 , U) are also convex.
Pp
i=0
ρX (v) = sup{x · v : x ∈ X} .
xi vi is the dot product of the two
Support functions do not exclusively apply to
convex polyhedra, but in fact to any set X ⊆ R p
represented by a general assertion θ(X). We will
restrict ourselves to the use of convex polyhedra, in
which case the support function definition translates
to solving the linear program
n−1
τ (X0 , U) = τ(τ
(X0 , U) ∩ G, U)
(3)
ρX (v) = max{x · v | Cx ≤ d} .
(5)
3
which is consistent with the real case when θ = 0.
The reason why eiθ cannot be extracted out is be2) Support Functions Properties: Several properties of support functions allow us to reduce operational complexity. The most significant are [23]:
ρkX (v) = ρX (kv) = kρX (v) : k ≥ 0
ρAX (v) = ρX (AT v) : A ∈ R p×p
ρX1 ⊕X2 (v) = ρX1 (v) + ρX2 (v)
ρX (v1 + v2 ) ≤ ρX (v1 ) + ρX (v2 )
ρconv(X1 ∪X2 ) (v) = max{ρX1 (v), ρX2 (v)}
ρX1 ∩X2 (v) ≤ min{ρX1 (v), ρX2 (v)}
As can be seen by their structure, some of these properties reduce complexity to lower-order polynomial
or even to constant time, by turning matrix-matrix
multiplications (O(p3 )) into matrix-vector (O(p2 )), or
into scalar multiplications.
3) Support Functions in Complex Spaces:
The literature does not state, as far as we found
any description of the use of support functions
in complex spaces. Since this is relevant to using
our technique, we extend the definition of support
functions to encompass their operation on complex
spaces.
A support function in a complex vector field
is a transformation:
cause it is a rotation, and therefore follows the same
rules as a matrix multiplication,
cos θ sin θ
iθ
v .
ρX (e v) , ρX
− sin θ cos θ
Since matrices using pseudo-eigenvalues are real, all
other properties remain the same. An important note
is that when using pseudo-eigenvalues, conjugate
eigenvector pairs must be also converted into two
separate real eigenvectors, corresponding to the real
and the imaginary parts of the pair.
III. The Polyhedral Abstract Domain
A. Convex Polyhedra
A polyhedron is a topological element in R p
with flat polygonal (2-dimensional) faces. Each face
corresponds to a hyperplane that creates a halfspace,
and the intersections of these hyperplanes are the
edges of the polyhedron. A polyhedron is said to be
convex if its surface does not intersect itself and a
line segment joining any two points of its surface is
contained in the interior of the polyhedron. Convex
polyhedra are better suited than general polyhedra
as an abstract domain, mainly because they have a
simpler representation and operations over convex
ρX (v) : C p → R = sup{|x · v| | x ∈ X ⊆ C p , v ∈ C p }.
polyhedra are in general easier than for general
The dot product used here is the Euclidean Internal
polyhedra. There are a number of properties of
convex polyhedra that make them ideal for abstract
Product of the vectors, which is commonly defined
in the complex space as:
a·b=
p
X
i=0
ai bi ,
a, b ∈ C p
We are interested in the norm of the complex value,
which is a 1-norm given our definition of dot product:
|a · b| = |re(a · b)| + |im(a · b)|
Returning to our support function properties, we
now have:
ρX (reiθ v) = rρX (eiθ v),
interpretation of continuous spaces, including their
ability to reduce an uncountable set of real points
into a countable set of faces, edges and vertices.
Convex polyhedra retain their convexity across linear
transformations, and are functional across a number
of operations because they have a dual representation [21]. The mechanism to switch between these
two representations is given in section III-B5
1) Vertex Representation: Since every edge in the
polyhedron corresponds to a line between two vertices and every face corresponds to the area enclosed
4
by a set of co-planar edges, a full description of the
polyhedron is obtain by simply listing its vertices.
Since linear operations retain the topological properties of the polyhedron, performing these operations
on the vertices is sufficient to obtain a complete
description of the transformed polyhedron (defined
dron is
V ′ = {v′ = v + t | v ∈ V}
Given an inequality representation X and a translation vector t, the transformed polyhedron corresponds to
X ′ = {x | Cx ≤ d + Ct}
by the transformed vertices). Formally, a polyhedron
is a set V ∈ R p such that v ∈ V is a vertex of the
polyhedron.
2) Linear Transformation: Given a vertex rep-
2) Inequality Representation: The dual of the
Vertex representation is the face representation. Each
resentation V and a linear transformation L, the
transformed polyhedron is
face corresponds to a bounding hyperplane of the
V ′ = LV
polyhedron (with the edges being the intersection
of two hyperplanes and the vertices the intersection
Given an inequality representation X and a linear
of 3 or more), and described mathematically as a
function of the vector normal to the hyperplane. If
transformation L, the transformed polyhedron corresponds to
we examine this description closely, we can see that
it corresponds to the support function of the vector
normal to the hyperplane. Given this description we
formalise the following: A convex polyhedron is a
p
o
n
X ′ = x | C(L+ )T x ≤ ρ′X (L+ CT )
where L+ represents the pseudo-inverse of L. In the
case when the inverse L−1 exists, then
topological region in R described by the set
X = {x ∈ R p | Cx ≤ d, C ∈ Rm×p , d ∈ Rm }
where the rows Ci,∗ for i ∈ [1, m] correspond to the
transposed vectors normal to the faces of the polyhedron and di for i ∈ [1, m] their support functions.
For simplicity of expression, we will extend the use
of the support function operator as follows:
ρ′X : Rm×p → Rm
ρX (MT )
1,∗
ρX (MT )
2,∗
ρ′X (MT ) =
..
.
ρX (MTm,∗ )
B. Operations on Convex Polyhedra
Several operations of interest can be performed on
convex polyhedra
1) Translation: Given a vertex representation V
and a translation vector t, the transformed polyhe-
n
o
X ′ = x | C(L−1 )T x ≤ d)
From this we can conclude that linear transformations are better handled by a vertex representation,
except when the inverse of the transformation exists
and is know a-priori. This works makes use of this
last case to avoid continuous swapping in representations.
3) Set Sums: The addition of two polyhedra is a
slightly more complex matter. The resulting set is
one such that for all possible combinations of points
inside both original polyhedra, the sum is contained
in the result. This operation is commonly known as
the Minkowski sum, namely
A ⊕ B = {a + b | a ∈ A, b ∈ B}
Given two vertex representations V1 and V2 the
resulting polyhedron
V = conv(V1 ⊕ V2 )
5
where conv(·) is the convex hull of the set of vertices
contained in the Minkowski sum.
5) Vertex Enumeration: The vertex enumeration
problem corresponds to the algorithm required to
Let
obtain a list of all vertices of a polyhedron given an
inequality description of its bounding hyperplanes.
X1 = {x | C1 x ≤ d1 }
Given the duality of the problem, it is also possible
to find the bounding hyperplanes given a vertex de-
X2 = {x | C2 x ≤ d2 }
scription if the chosen algorithm exploits this duality.
be two sets, then
In this case the description of V is given in the forms
of a matrix inequality Vx ≤ [ 1 1 · · · 1 ]T with
X = X1 ⊕ X2 = {x | Cx ≤ d},
V = [ v1 · · · vm ]T , vi ∈ V. Similarly, A can be
described as a set containing each of its rows. At
where
C1
d + ρ′X (CT1 )
2
, d = 1
C =
C2
d2 + ρ′X1 (CT2 )
.
Because these sets correspond to systems of inequalities, they may be reduced removing redundant
constraints. Note that if C1 = C2 then
X = X1 ⊕ X2 = {x | C1 x ≤ d1 + d2 },
4) Set Hadamard Product:
Lemma 1. Given two vertex representations V1 and
V2 the resulting polyhedron
V = V1 ◦ V2 = conv({v = v1 ◦ v2 | v1 ∈ V1 , v2 ∈ V2 )
where ◦ represents the Hadamard (coefficient-wise)
product of the vectors, contains all possible combinations of products between elements of each set.
Proof: Given a convex set X, we have:
the time of writing, there are two algorithms that
efficiently solve the vertex enumeration problem. lrs
is a reverse search algorithm, while cdd follows the
double description method. In this work we use the
cdd algorithm for convenience in implementation
(the original cdd was developed for floats, whereas
lrs uses rationals). The techniques presented here can
be applied to either.
Let
C = {x | Ax ≥ 0, A ∈ Rn×p , x ∈ R p }
be the polyhedral cone represented by A. The pair
( A, V) is said to be a double description pair if
C = {λT V | V ∈ R p , λ ∈ R|V|
≥0 }
V is called the generator of X. Each element in V lies
in the cone of X, and its minimal form (smallest m)
has a one-to-one correspondence with the extreme
X ′ = {xi j | xi , x j ∈ X, xi j = txi +(1−t)x j , t ∈ [0, 1]} ⊆ X
rays of X if the cone is pointed (i.e., it has a vertex
at the origin). This last can be ensured by translating
Given xi ∈ X, y j ∈ Y, zi, j = xi ◦ y j ∈ Z
a polyhedral description so that it includes the origin,
and then translating the vertices back once they have
xi j ∈ X ′ , yk ∈ Y, zi,k , z j,k ∈ Z ⇒ zi j,k ∈ Z
xi j ∈ X, ym , yn ∈ Y, zi j,m , zi j,n ∈ Z ⇒ zi j,mn ∈ Z
This equation proves that given v11 , v12 ∈ V1 ,
v21 , v22 ∈ V2 and u, t ∈ [0, 1],
(tv11 + (1 − t)v12 ) (uv21 + (1 − u)v22 ) ∈ V
been discovered (see section III-B).
We will also point out that
n
{x | Ax ≤ b} = x′ | [ −A
x
p
′
where x ∈ R and x =
1
b ] x′ ≥ 0
∈ R p+1 .
o
The vertex enumeration algorithm starts by finding
a base CK which contains a number of vertices of
6
the polyhedron. This can be done by pivoting over a
number of different rows in A and selecting the feasi-
computing the reachset (3) using a function f such
that f (·) = τn (·). In the case of systems without
ble visited points, which are known to be vertices of
the polyhedron (pivoting p times will ensure at least
inputs, this equation is xn = An x0 . We will use this
property and others derived from it to calculate our
one vertex is visited if the polyhedron is non-empty). abstract matrices.
CK is represented by AK which contains the rows
used for the pivots. The base CK is then iteratively
expanded to CK+i by exploring the ith row of A until
CK = C. The corresponding pairs ( AK+i , VK+i ) are
constructed using the information from ( AK , VK ) as
follows:
(6)
Hi− = {x | Ai,∗ x < 0},
(7)
= {x | Ai,∗ x = 0}
(8)
Hi0
be the spaces outside inside and on the ith hyperplane
and
VK+
VK−
= {v j ∈
= {v j ∈
Hi+ },
Hi− },
VK0 = {v j ∈ Hi0 }
The basic steps required to evaluate a reach tube
using abstract acceleration can be seen in figure 1.
1) The process starts by doing eigendecomposition
Let AK ∈ RnK ×p , Ai,∗ ∈ R1×p , VK ∈ R p ,
Hi+ = {x | Ai,∗ x > 0},
B. Overview of the Algorithm
(9)
(10)
(11)
the existing vertices lying on each of these spaces.
Then
of the dynamics ( A) in order to transform the
problem into a simpler one.
2) A variety of off-the-shelf tools may be used,
but since larger problems require numerical
algorithms for scalability, a second step involves
upper-bounding the error in order to obtain
sound results. In such cases, all subsequent
steps must be performed using interval arithmetic.
3) The inverse of the generalised eigenvectors
must be calculated soundly.
4) The problem gets transformed into canonical
form by multiplying both sides of the equation
VK+i = VK+ ∪ VK− ∪ VK0 ∪ VKi
(12)
VKi = ( Ai,∗ v+ )v− − ( Ai,∗ v− )v+ | v− ∈ V − , v+ ∈ V +
(13)
For the proof see [21].
IV. Abstract Matrices in abstract acceleration
A. Acceleration Techniques
Acceleration of a transition system is a method
that seeks to precisely describe the transition
by S−1 :
′
∩ G′ + U ′ where
Xk′ = J Xk−1
X ′ = S−1 X, U ′ = S−1 BU and G′ = {x | GSx ≤ h}
5) We calculate the number of iterations as explained in section VI. If there are no guards, we
use n = ∞. It is worth noting that this number
need not be exact: if we overapproximate the
number of iterations, the resulting reachtube
will overapproximate the desired one.
relations over a number of steps using a concise
description of the overall transition between the first
6) we overapproximate the dynamics of the variable inputs (for parametric or no inputs this step
and final step. Namely, it looks for a direct formula
to calculate the postimage of a loop from the initial
will be ignored) using the techniques described
in section V-D
states of the loop. Formally, given the dynamics
in equation (1) an acceleration formula aims at
7) we calculate the abstract dynamics using the
techniques described in section V-A
7
A
1. Calculate Ŝ, Ĵ
Eigenspace
2. Restore
Soundness
3. Find
Inverse
S
S−1
J
5. Find Number
of Iterations
G′
X0′
U′
4. Transform
problem into
Eigenspace
J, n
G
X0
BU
X0′ , U ′
7. Get Abstract
6. Semispherical
Dynamics
Jb Approximation
S
X♯
10. Get
Reach Tube
X0′ , Ub′
J
X ′♯
9. Find Eigen
Reach Tube
V0′
8. Find
Vertices
Fig. 1. Block diagram describing the different steps used to
calculate the abstract reach tube of a system.
8) we evaluate the vertices of the combined inputinitial eigenspace to be used as source for the
reachtube calculation
9) we use a sound simplex algorithm to evaluate
made up by the corresponding eigenvectors. We can
then easily compute An = SJ n S−1 , where
n
J 1
..
J n =
(15)
.
n
Jr
n n−p +1
n n n−1
s
λ s 1 λ s
...
p s −1 λ s
..
n n−1
n
.
λs
λs
n
1
J s∈[1,r] = .
..
..
.
.
.
.
λns
(16)
The abstract matrix An is computed as an abstraction
over a set of vectors mk ∈ R p , k ∈ [1 n] of entries of
Jk.
Let I s = [ 1 0 · · · 0 ] ∈ R ps . The vector mk
is obtained by the transformation ϕ−1 :
mk = [ I1 J k1
···
Ir J kr ] ∈ R p
(17)
the convex set product of the abstract dynamics
(used as the tableau) and the initial set (whose
such that J k = ϕ(mk ).
vertices are used as the obejctive functions
alongside a set of template directions for the
If J is diagonal [34], then mk equals the vector
of powers of eigenvalues [λk1 , . . . , λkr ]. An interval
result).
abstraction can thus be simply obtained by computing the intervals [min{λ0s , λns }, max{λ0s , λns }], s ∈
10) since we have calculated our result in the
eigenspace, we transform the reachtube back
into the normal space by multiplying by S .
C. Computation of Abstract Matrices
We define the abstract matrix An as an over-
approximation of the union of the powers of the maS
trix Ak such that An ⊇ k∈[0,n] Ak and its application
to the initial set X0
X̂n♯ = An X0 ⊇ X̂n
(14)
Next we explain how to compute such an abstract
matrix. For simplicity, we first describe this computation for matrices A with real eigenvalues, whereas
[1, r]. We observe that the spectrum of the interval
matrix σ(An ) (defined as intuitively) is an overS
approximation of k∈[0,n] σ( Ak ).
In the case of the sth Jordan block J s with geometric non-trivial multiplicity p s (λi = λi−1 = . . .),
observe that the first row of J ns contains all (possibly)
distinct entries of J ns . Hence, in general, the vector
section ms is the concatenation of the (transposed)
n n−p +1 T
s
first row vectors λns , n1 λn−1
of
s , . . . , p s −1 λ s
J ns .
Since the transformation ϕ transforms the vector
m into the shape of (16) of J n , it is called a matrix
shape [34]. We then define the abstract matrix as
the extension to the complex case will be addressed
in Section IV-D. Similar to [34], we first have to com-
An = {S ϕ(m) S−1 | Φm ≤ f } ,
pute the Jordan normal form of A. Let A = SJS−1
where J is the normal Jordan form of A, and S is
where the constraint Φm ≤ f is synthesised from
intervals associated to the individual eigenvalues and
(18)
8
to their combinations. More precisely, we compute
polyhedral relations: for any pair of eigenvalues (or
To deal with complex numbers in eigenvalues and
eigenvectors, [34] employs the real Jordan form for
conjugate eigenvalues λ = reiθ and λ∗ = re−iθ (θ ∈
cos θ
r
− sin θ
sin θ
.
cos θ
Although this equivalence will be of use once we
evaluate the progression of the system, calculating
powers under this notations is often more difficult
than handling directly the original matrices with
complex values.
In Section IV-C, in the case of real
eigenvalues we
in the power
150
100
50
D. Abstract Matrices in Complex Spaces
[0, π]), so that
λ 0
is replaced by
0 λ∗
n=5
200
λn2
binomials) within J, we find an over-approximation
of the convex hull containing the points
n
o
mk | k ∈ [1, n] ⊆ {m | Φm≤ f }
250
n=4
n=1
n=3
n=2
0
0
5
10
20
15
25
30
35
λn1
Fig. 2. Polyhedral faces from an R2 subspace, where (λn1 , λn2 )
so that λ1 =2, λ2 =3, 1≤n≤5. Bold purple lines represent supports
found by this article. The dotted grey and dashed red polytopes
show logahedral approximations (box and octagon) used in [34].
Note the scales (sloped dashed lines are parallel to the x=y
line, and dashed red polytope hides two small faces yielding an
octagon).
1) Positive Real Eigenvalues
The exponential curve is cut along the diagonal
have abstracted the entries
matrix J ns by ranges of
between the maximum and minimum eigenvalues to create a support function for the
eigenvalues [min{λ0s · · · λns }
max{λ0s · · · λns }].
In the complex case we can do something
corresponding hyperplane. A third point taken
from the curve is used to test the direction of
similar by
form λ s
the corresponding template vector. An arbitrary
number of additional hyperplanes are selected
[min{r0s · · · rns }
rewriting eigenvalues into polar
=
r s eiθs and abstracting by
max{r0s · · · rns }]ei[0
, min(θ s ,2π)]
.
V. General Abstract Acceleration with Inputs
A. Using Support Functions on Abstract Acceleration
by picking pairs of adjacent points in the curve
and creating the corresponding support function
as shown in Figure 2.
2) Complex Conjugate Eigenvalue pairs
As an improvement over [34], the rows in Φ and
In the case of complex conjugate pairs, the
eigenvalue map corresponds to a logarithmic
f (see (18)) are synthesised by discovering support
functions in these sets. The freedom of directions
spiral. In this case, we must first extract the
number of iterations required for a full cycle.
provided by these support functions results in an
improvement over the logahedral abstractions used in
For convergent eigenvalues, only the first n iter-
previous works (see Figures 2 - 5). The mechanism
by which this works follows the convex properties
of the exponential progression. There are four cases
cases to consider 1 :
1 these
explain in detail the procedure alluded in [8]
ations have an effect on the support functions,
while in the divergent case only the last n
iterations are considered. Support functions are
found for adjacent pairs, checking for the location of the origin point (first point for convergent eigenvalues, last for divergent eigenvalues).
9
0.8
2
0.6
n=3
apex
n=3
n=2
0.4
1.5
n=2
λn2
λn2
n=1
0.2
0
n=1
1
−0.2
−0.4
−0.6 −0.4 −0.2
0.5
0
0.2
0.4
0.6
0.8
0
0.2
Fig. 3. Polyhedral faces from an R2 complex conjugate subspace,
where (λn1 , λn2 ) so that λ1 =0.8 + 0.4i, λ2 =0.8 − 0.4i, 1≤n≤14. Bold
purple lines represent supports found by this article. The blue
dotted line shows the support function that excludes the origin
(n=1), which is replaced by the support function projecting from
said origin.
0.4
0.6
0.8
λn1
λn1
Fig. 4. Polyhedral faces from an R2 Jordan block subspace,
where (λn1 , λn2 ) so that λ1 =0.8, λ2 =0.8, 1≤n≤15. Bold purple lines
represent supports found by this article. The blue dotted line
shows the support function that excludes the origin (n=1), which
is replaced by the support function projecting from said origin.
account for both sides of the axis on the latIf the origin falls outside the support function,
we look for an interpolant point that closes the
spiral tangent to the origin. This last check is
performed as a binary search over the remaining
points in the circle (whose supporting planes
would exclude the origin) to achieve maximum
tightness (see Figure 3).
3) Equal Eigenvalues
When two eigenvalues are the same, the resulting support functions are those orthogonal to
the x = y plane, intersecting the square created
by the maximum and minimum values.
4) Jordan Blocks of size > 1
In the case of eigenvalues with geometric mul-
ter. These form mirror images that are merged
during the abstraction. To make matters simple,
we use the magnitude of a complex eigenvalue
and evaluate whether the dynamics are concave
or convex with respect to the mirroring plane.
See Figure 5.
Note that if both eigenvalues are negative and/or
conjugate pairs from a different pair, the mirror
image would be taken on both axes, resulting
in a hyperrectangle. For a tighter bound in the
purely convergent case, we find the convex hull
of a point cloud for a small time horizon and
merge it with the hyperrectangle for the infinite
time horizon thereon.
tiplicities > 1, the shape of the function is
similar to its corresponding unit size block. In
An additional drawback of [34] is that calculating
the exact Jordan form of any matrix is computa-
the convergent case, since the convexity can be
tionally expensive and hard to achieve for large-
sharp, it is important to find the apex of the
upper diagonals in order to minimise the over-
dimensional matrices. We will instead use numerical
algorithms in order to get an approximation of the
approximation. See Figure 4.
5) Negative Eigenvalues and mixed types
When mapping a positive real eigenvalue to a
complex conjugate or negative one, we must
Jordan normal form and account for numerical errors.
In particular, if we examine the nature of (14),
we find out that the numerical operations are not
iterative, therefore the errors do not accumulate with
10
vides a 100-fold plus improvement, which can make
a difference towards rendering verification practically
15
λ2 > |λ1 | > 1
10
λn2
λ2 = |λ1 |
feasible. For a full description on the numerical
processes described here see [9]
|λ1 | > λ2 > 1
5
B. Abstract Matrices in Support Functions
|λ1 | > 1 > λ2
Since we are describing operations using abstract
matrices and support functions, we briefly review the
0
−10
0
10
λn1
Fig. 5. Polyhedral faces from an R2 subspace, with different
convexities (note that the blue and orange plots are convex w.r.t.
the λn2 -axis, whereas the green and brown are concave). Dotted
purple lines represent supports for some of these layouts.
nature of these operations and the properties that the
support functions retain within this domain. Let X ∈
R p be a space and A ∈ R p×p an abstract matrix for
the same space. From the definition we have
[
A=
Sϕ(m)S−1 : Φm ≤ f
which leads to
time. We use properties of eigenvalues to relax f by
finding the maximum error in the calculations that
can be determined by computing the norm δmax =
|A − ŜĴSˆ−1 |, where Ĵ and Ŝ are the numerically
ρAX (v) = ρSϕ(m)S−1 X (v) = ρϕ(m)S−1 X ST v ,
(19)
where
n
o
ρϕX (v) = sup ρϕ (x ◦ v) | x ∈ X
(20)
interval matrices and all operations are performed
using interval arithmetic with outward rounding in
ρϕ (v) = sup{m · ϕ−1 (v) | Φm ≤ f }
(21)
order to ensure soundness. In the following we will
presume exact results and use the regular notation to
Here, x ◦ y is the Hadamard product, where
(x ◦ y)i = xi yi , and ϕ−1 (·) is the reverse operation
calculated eigenvalues and eigenvectors of A. The
notation above is used to represent the matrices as
describe the algorithms. The constraints Φm < f are
then computed by considering the ranges of eigenvalues λ s ±δmax (represented in Fig. 2 as the diameter of
the blue dots). The outward relaxation of the support
functions ( f ), which follows a principle similar to
that introduced in [22], reduces the tightness of the
over-approximation, but ensures the soundness of the
abstract matrix An obtained. It is also worth noting
that the transformation matrices into and from the
eigenspace will also introduce over-approximations
due to the intervals. One can still use exact arithmetic
with a noticeable improvement over previous work;
however, for larger-scale systems the option of using
floating-point arithmetic, while taking into account
errors and meticulously setting rounding modes, pro-
and
of ϕ(·) in order to align the elements on v with the
elements in m. In the case of conjugate pairs this
is equivalent to multiplying the vector section by
h 1 1 i
, and in the case of a Jordan block by an
1
−1
upper triangular matrix of all ones.
We may also define
ρAX (v) = sup {ρ aX (v), ∀a ∈ A}
n
o
= sup Sϕ(m)S−1 x · v, ∀x ∈ X
n
o
= sup ϕ(m)S−1 x · ST v, ∀x ∈ X
n
o
= sup ρϕ (S−1 x ◦ ST v), ∀x ∈ X .
(22)
In order to simplify the nomenclature we write
ρAX (v) = ρX (AT v).
(23)
11
following representation:
C. Acceleration of Parametric Inputs
Let us now consider the following overapproximation for τ on sets:
n−1
X
k=0
♯
τ (X0 , U) = AX0 ⊕ BU
(24)
Unfolding (3) (ignoring the presence of the guard set
G for the time being), we obtain
Xn = An X0 ⊕
X
Ak BU
k∈[0,n−1]
What is left to do is to further simplify the sum
P
k
k∈[0,n−1] A BU. We can exploit the following simple results from linear algebra.
Lemma 2. If I − A is invertible, then
n−1
X
k=0
λ =
k
n
1−λn
1−λ
λ=1
λ,1
⇒ (I − An )(I − A)−1 = SDn S−1
−1k 1 − λni
d(λi , n, k) =
k + 1 (1 − λi )k+1
! n− j−1
k
X
λi
−1k− j n
+
k
−
j
j
−
1
(1
− λi )k− j
j=1
0
j<i
n
i = j ∧ λi = 1
1−λni
i = j ∧ λi , 1
1−λi
Dni, j =
0
gm(λi ) = 1
n+1
λi = 1
k+1
d(λi , n, j − i)
λi , 1
(25)
where gm(·) is the geometric multiplicity of the given
Ak = (I − An )(I − A)−1
eigenvalue.
. If furthermore lim An = 0, then
n→∞
lim
n→∞
n
X
k=0
D. Acceleration of Variable Inputs
k
A = (I − A)
−1
The result in the previous section can be only
.
This lemma presents a difficulty in the nature of A.
The inverse (I − A)−1 , does not exist for eigenvalues
directly applied under restricted conditions in the
case of variable inputs. For instance whenever ∀k >
0, uk = uk−1 . In order to generalise it (in particular to
non-constant inputs), we will over-approximate BU
of 1, i.e. we need 1 < σ( A), where σ(A) is the over the eigenspace by a semi-spherical enclosure
spectrum (the set of all the eigenvalues) of matrix A. with centre u′c and radius U ′ . To this end, we first
b
In order to overcome this problem, we introduce the rewrite
eigen-decomposition of A = SJS−1 , (and trivially
U J′ = S−1 BU = {u′c } ⊕ Ud′ ,
I = SIS−1 ), and by the distributive and transitive
properties we obtain
where u′c is the centre of the interval hull of U J′ :
n
(I − A )(I − A)
−1
n
−1 −1
= S(I − J )(I − J) S
.
1
1
= (ρU′J (vi ) + ρU′J (−vi )) | vi j =
0
2
j=i
.
This allows us to accelerate the eigenvalues individj,i
Pn−1 k
ually, using the property k=0 1 = n for eigenvalues
of λ = 1. Using the properties above, and trans- We then over-approximate Ud′ via Ub′ , by the maxiu′ci
lating the problem into the generalised eigenspace
to accounting for unit eigenvalues, we obtain the
mum radius in the directions of the complex eigenvalues as (cf. illustration in Figure 6). Let
12
40
Λ = {λi | i ∈ [1, p], λ∗i , λi−1 }
λi < Λ
λi , λ∗i+1
of a vector by removing the elements where λi < Λ.
Extending this to matrices we have
Fb : R
o×pb
→R
Fb (C) = Cb where (Cb )i,∗ = fb (Ci,∗ )
0
0, 0
λi = λ∗i+1
and red(·) is a function that reduces the dimension
o×p
u′c
λ2
0
|v |
fb (v) = red(vb ) where (vb )i =
q i
2
vi + v2i+1
Ud′
20
fb : R p → R pb such that
U′
−20
−40
Ub′
−40
−20
0
20
40
λ1
(26)
Finally
Fig. 6. Relaxation of an input set within a complex subspace, in
order to make it invariant to matrix rotations. Dashed lines and
curves denote translated quantities onto the origin.
Ud′ = {u | C′u u ≤ d′u }
Ud′ ⊆ Ub′ = {u | Fb (C′u ) fb (u) ≤ fb (d′u )}
BU ⊆ Ub ⊕ Uc | Ub = SUb′ and Uc = {Su′c }
(27)
Since the description of Ub′ is no longer polyhedral
in R p , we will also create a semi-spherical overapproximation J b of J in the directions of the complex eigenvectors, in a similar way as we generated
Ub′ for Ud′ . More precisely,
J b1
..
where
J b =
.
n
J br
λ j ∈ J b s = |λi | ∈ J s ∩ Λ
∀s ∈ [1 r]
gm(J b s ) = gm(J s )
(31)
Shifting our attention from reach sets to tubes,
we can now over-approximate the reach tube by
Theorem 1. The abstract acceleration
(28)
Jordan block.
Definition 1. Given a matrix A = SJS−1 and a
vector x, we define the following operations:
Fb∗ ( A, x) = S fb−1 Fb (J) fb (S−1 x)
(29)
′
′
−1
Fb ( A, x) = fb Fb (J) fb (x )
(30)
Finally, we refer to the accelerated sets
o
n
Ubn = Fb∗ ((I − An ), Fb∗ ((I − A)−1 , u)) | u ∈ Ub
n
Ucb
= Ucn ⊕ Ubn
n
Xn ⊆ An X0 ⊕ Ucb
abstract acceleration of the three summands in (31),
as follows.
where gm(·) is the geometric multiplicity of the
Ucn = (I − An )(I − A)−1 Uc
Returning to our original equation for the n-reach
set, we obtain2
τ♯n (X0 , U) =def An X0 ⊕ Bn Uc ⊕ Bnb Ub
(32)
is an over-approximation of the n-reach tube, namely
X̂n ⊆ τ♯n (X0 , U).
Proof: The proof is derived from that in [34]
for An X0 , and extends it as in the developments
presented above.
E. Combining Abstract Matrices
One important property of the abstract matrices
An , Bn and Bnb is that they are correlated. In the
that ∀U b′ , U c′ , U d′ ; ∃U b , U c , U d : U b′ = S−1 BU b so
that
= S−1 BU c and U d′ = S−1 BU d . Hence, this inclusion is
also valid in the original state space.
2 Note
U c′
13
case of parametric inputs, this correlation is linear
and described by the acceleration defined in Lemma
with J b defined by (28). This model provides a
tighter over-approximation than (32) since the accel-
(2). In the case of Bnb this relationship is not linear
(see Eq. 27). However, we can still find a linear over-
erated dynamics of the inputs are now constrained
by the accelerated dynamics of the system.
approximation of the correlation between Bnb and An
based on the time steps k. Given two orthonormal
p
q
spaces X ∈ R ∧ U ∈ R and a transition equation
The most important task remaining is how to
Xk+1 = AXk + BU,
calculate the number of iterations dealing with the
presence of the guard set G.
which is related to
Given a convex polyhedral guard expressed as the
assertion {x | Gx ≤ h}, we define Gi,∗ as the ith row
ρXk+1 (v) = ρ AXk (v) + ρ BU (v),
we define a space
x
′
X =
|
x
∈
X,
u
∈
U
Bu
of G and hi as the corresponding element of h. We
denote the normal vector to the ith face of the guard
as gi = GTi . The distance of the guard to the origin
is thus γi = | hgi | .
so that
with
i
Given a convex set X, we may now describe
its position with respect to each face of the guard
T
A v
= ρX ′ DT v′ ,
ρXk+1 (v) = ρXk′
k
v
A
D =
0
0
,
I
Accelerating Xk+1 , we obtain
through the use of its support function alongside
the normal vector of the hyperplane (for clarity, we
v
v = .
v
assume the origin to be inside set X):
′
ρX ( gi ) ≤ γi ,
ρXn (v) = ρ An X0 (v) + ρ(I− An )(I− A)−1 BU (v) = ρX0′ DnT v′ ,
with
n
A
n
D =
0
0
(I − An )(I − A)−1
in the case of parametric inputs. More generally, the
diagonal elements of Dn correspond to the diagonal
P
k
elements of An and n−1
k=0 A B, which means we can
construct
An 0
n
(33)
D =
| ρXn (v) = ρX0′ (DnT v′ ).
0 Bn
We can then apply this abstraction to (27) and obtain:
ρXn (v) =
′
ρX0′ (DnT
b v )
An
n
Db =
0
VI. Abstract Acceleration with Guards:
Estimation of the number of Iterations
where
v
0
, v′ =
Bnb
fb (v)
Bnb = SFb−1 (I − Jbn )(I − J b )−1 Fb (S−1 )
(34)
inside the hyperplane,
−ρX (−gi ) ≥ γi ,
outside the hyperplane.
Applying this to equation (31) we obtain:
ρXn ( gi ) = ρX0 ( Ani T gi ) + ρUcbn ( gi ) ≤ γi
(35)
ρXn (−gi ) = ρX0 (−A
(36)
ni T
gi ) + ρUcbn (−gi ) ≤ −γi
From the inequalities above we can determine
up to which number of iterations ni the reach tube
remains inside the corresponding hyperplane, and
starting from which iteration ni the corresponding
reach set goes beyond the guard:
In order for a reach set to be inside the guard
it must therefore be inside all of its faces, and we
can ensure it is fully outside of the guard set when
it is fully beyond any of them. Thus, we have n =
min{ ni }, and n = min{ ni }.
We have not however discussed why these two
cases are important. Looking at the transition in
equation (1), we can easily derive that if Gxk h
14
the postimage of all subsequent iterations is empty.
Therefore, any overapproximation henceforth will
diagonal. In such a case, the progression of the
support function in each direction is monotonically
only add imprecision. We will use the bounds n and
n to create a tighter overapproximation. Let
increasing (or decreasing) and it is therefore very
easy to find a bound for its progression. We note
X̂n♯ = An X0 ⊕ Bn U
that the envelope of rotating dynamics will always
contain the true dynamics and is therefore a sound
(n-reachtube)
Xn♯ = An X0 ⊕ Bn U
(n-reachset)
X̂n♯ | n = τ An−n−1 Xn♯ ⊕ Bn U ∩ G, U
X̂n♯
X̂n♯ | n
∪
overapproximation. We will initially assume that γi
is positive and then extend to the general case.
X̂n♯
(37)
o
n
This double step prevents the set x | x ∈ X̂n♯ , x < Xn♯
to be included in further projections, thus reducing
=
Let ρX0 ( AnT gi ) = ρX0′ (J nT g′i ) such that
g′i = S−1 gi
X0 = {x | CX0 x ≤ dX0 }
the size of the overapproximation.
X0′ = S−1 X0 = {x | SC X0 x ≤ dX0 }
Computing the maximum ni such that (35) is
satisfied is not easy because the unknown ni occurs
in the exponent of the equation. However, since an
intersection with the guard set will always return a
sound over-approximation, we do not need a precise
value. We can over-approximate it by decomposing
gi into the generalised eigenspace of A. Let gi =
Pp
−1
or
j=1 ki j v j +res( gi ), where v j are row vectors of S
−1
−S such that ki j ≥ 0, and res( gi ) is the component
Let
Λσ = {λi : i ∈ [1, p],
p
fσ (v) : R → R
i−1
^
(λ∗i , λ j ∧ λi , λ j )}
j=1
pb
of gi that lies outside the range of S. Notice that
since S has an inverse, it is full rank and therefore
fσ (v) = red(vσ )
0
r
P
where (vσ )i =
v2j
j∈[1,p]∧(λ j =λi ∨λ j =λ∗i )
res( gi ) = 0 and subsequently not relevant. It is also
Fσ (C) = Cσ where (Cσ )i,∗ = fσ (Ci,∗ ).
important to note that S is the matrix of generalised
eigenvectors of A and therefore we are expressing
our guard in the generalised eigenspace of A.
p
p
X
X
nT
nT
ρX0 ( A gi ) = ρX0 ki j A v j ≤
ki j ρX0 AnT v j
j=1
j=1
(38)
A. Overestimating the Iterations of a loop without
inputs
We start by looking into the approximation of
the inside bound (i.e. the iterations for which the
reachtube remains fully inside the guard). Since
rotating dynamics and Jordan shapes will have a
complex effect on the behaviour of the equation, we
seek to transform the Jordan form into a real positive
λi < Λσ
λi ∈ Λσ
Fσ : Ro×p → Ro×r
and red(·) is a function that reduces the dimension of a vector by removing the elements where
λi < Λσ . This reduction is not extrictly necessary, but it enables a faster implementation by reducing dimensionality. Correspondingly, given J =
diag ({J s | s ∈ [1, r]})
σ1
0
J σ =
0
0
0
σ2
0
0
···
···
..
.
···
0
σr
0
0
(39)
where σ s = ||J s ||2 is the maximum singular value
(hence the induced norm [36]) of the Jordan block
J s.
15
Finally, let
x′c =
1
vi j =
0
1
(ρX ′ (vi ) + ρX0′ (−vi )),
2 0
j=i
j,i
Xσ′ = {x | Fσ (SC X0 ) fσ (x) ≤ fσ (dX0 − SCX0 x′c )}
′
′
) | Xcσ
= { fσ (x′c )} ⊕ Xσ′
X0′ ⊆ fσ−1 (Xcσ
(40)
Proof: This follows from the developments unfolded above. Notice that the sequence ni is monotonically increasing, before it breaks the inequality. As
such any local minimum represents a sound underapproximation of the number of loop iterations. Note
that in the case where γi ≤ 0 we must first translate
the system coordinates such that γi > 0. This is
and vσ = fσ (v).
Using eigenvalue and singular value properties, we
obtain ρX0 ( AnT v j ) ≤ σ j n ρXcσ ((vσ ) j ) | j ∈ [1, r], and
therefore:
p
X
ρX0 ( AnT gi ) ≤
ki j σnj ρXcσ ((vσ ) j ))
(41)
j=1
Since we have no inputs, ρUcn ( gi ) + ρUbn ( gi ) = 0,
hence we may solve for ni :
ρX0 ( Ani T gi ) ≤
p
X
j=1
n
ki j σ j i ρXcσ ((vσ ) j ) ≤ γi
(42)
simply done by replacing x′ = x + c and operating
over the resulting system where γi′ = ρ c ( gi ) + γi .
Mathematically this is achieved as follows: first
we get c by finding the center of the interval hull
of G (if G is open in a given direction we may pick
any number in that direction for the corresponding
row of c). Next we
xk A Ac
=
1
0
1
where
To separate the divergent element of the dynamics
from the convergent one, let us define
ki j = max ki j ρXcσ ((vσ ) j ) , 0
transform the dynamics into
xk−1 B
x
+
uk | k−1
1
0
1
x G
G =
|
1 0
′
Gc xk−1
1
1
∈ G′
h
≤
1
σ = max ({σ s | s ∈ [1, p]}) .
This step will allow us to track effectively which
B. Underestimating the Iterations of a loop without
trajectories are likely to hit the guard and when, since
it is only the divergent element of the dynamics that
inputs
can increase the reach tube in a given direction.
Replacing (42), we obtain
p
X
σj
σ
!n
In order to apply a similar techniques to (36) we
must find an equivalent under-approximation. In the
case of equation (42), the σ j esure that the equation
(43)
diverges faster than the real dynamics, hence the
iteration found is an upper bound to the desired
which allows to finally formulate an iteration scheme
iteration. In this case we want the opposite, hence
we look for a model where the dynamics diverge
σ
n
j=1
ki j
≤ γi ,
for approximating n.
Proposition 1. An iterative under-approximation of
the number of iterations n can be computed by
starting with ni = 0 and iterating over
p
!ni
X
σ
j
, (44)
ni ≥ n = logσ (γi ) − logσ ki j
σ
j=1
substituting ni = n on the right-hand side until we
meet the inequality.
slower. In this case it is easy to demonstrate that
λb j = |λ j | represents these slower dynamics.
ni T
ρX0 (−A
gi ) ≤
p
X
j=1
ki j λb j ni ρXcσ (−(vσ ) j ) ≤ −γi (45)
which reduces to
σ
n
p
X
j=1
k−i j
λb j
σ
!n
+σ
n
p
X
j=1
k+i j ≤ −γi ,
(46)
16
where
k−i j = min ki j ρXcσ (−(vσ ) j ) , 0
k+i j = max ki j ρXcσ (−(vσ ) j ) , 0
cosines and removing the constants:
!
p
ni X
((n − ni )θ j − αi j )2
ci j 1 −
max σ
2
j=1
p
X
⇒ min ci j ((n − ni )θ j − αi j )2
j=1
p
X
2
2
⇒ min ci j θ j (n − ni ) + ci j αi j θ j (n − ni )
j=1
An additional consideration must also be made
regarding the rotational nature of the dynamics. In
the previous case we did not care about the rotational
alignment of the set Xn with respect to the vector
gi , because any rotation would move the set inside
the guard. In this case, although the magnitude of
the resulting vector is greater than the required one,
the rotation may cause it to be at an angle that
keeps the set inside the guard. We must therefore
account for the rotating dynamics in order to find
the point where the angles align with the guard. In
order to do this, let us first fix the magnitudes of
the powered eigenvalues, in the case of convergent
dynamics we will assume they have converged a full
The solution to this equation is
Pp
j=1 ci j αi j θ j
| n ∈ [ni , ni + nθ ]
n = ni − P p
2 j=1 ci j θ2j
(47)
The second part of the equation is expected to be
a positive value. When this is not the case, the
dominating dynamics will have a rotation θ j ≥ pi2 .
In such cases we must explicitly evaluate the set of
up to 4 iterations after ni . If the resulting
bound does
T
not satisfy the original inequality: ρX0 Ani gi ≥ γi ,
we replace ni = n until it does 3 .
Proposition 2. An iterative under-approximation of
the number of iterations n can be computed by
rotation in to make our equation strictly divergent.
starting with ni ′ = 0 and iterating over
Let θ = min{θ j | j ∈ [1, p]}, where θ j are the angles
p
!ni ′ X
p
X
λ
2π
b
j
′
of the complex conjugate eigenvalues. Let nθ = θ
+
ni ≤ n = logσ (γi ) − logσ k−i j
k+i j
σ
j=1
j=1
be the maximum number of iterations needed for
′
T
any of the dynamics to complete a full turn. Then
(48)
ni = ni ′ + k | ρX0 A(ni +k) gi ≥ γi
ni +nθ
ni +n
≤ |λ j |
| |λi | ≤ 1, n ∈ 0nθ .
at any given turn |λ j |
where k is the result of equation (47). we substitute
This means that any bound we find on the iterations
for ni = n on the right-hand side until we break
will be necessarily smaller than the true value. Our
the inequality, and then find k such that the second
problem becomes the solution to:
inequality holds.
max σ
−1
ni
p
X
j=1
ci j cos((n − ni )θ j − αi j )
αi j = cos ( gi · v j )
n
− λb j i
ki j σ
ci j =
k− λb j ni +nθ
ij σ
Since we are explicitly verifying the inequality,
there is no further proof required.
C. Estimating the Iterations of a loop with inputs
For the case with inputs, we will use the same
|λ j | ≥ 1
|λ j | < 1
The problem is simplified by underapproximating the
paradigm explained in the previous section after performing a mutation that transforms the system with
3 this
is a tighter value than work shown on previous versions
(2π)m
, where
of this paper where we overapproximated using nθ = Q
j θj
m is the number of conjugate pairs.
17
inputs into an over-approximating system without
inputs.
′
′
Let Xcσ
, Ucσ
be the corresponding sets of initial
states and inputs obtained by applying equation (40)
′
′
to X0′ and U J′ , and let U Jσ
= (I − J σ )−1 Ucσ
. The
accelerated resulting system may be represented by
the equations
′
′
⊕ (I − J nσ )U Jσ′
(Xcσ
)n = J nσ Xcσ
nT
′ ) (v) = ρX ′
J nT
ρ(Xcσ
σ v + ρU ′Jσ (v) − ρU ′Jσ J σ v
n
cσ
(49)
′
Xcσ
,u
Proof: Since the elements of the sums are
convergent, we have
ni ≥ nk ⇒ kni ≥ knk i.e. |kni | ≤ |knk |
⇒ logσ kni + kn ≥ logσ kni + kn
k
i
which means that nk in equation (51) is smaller than
or n in equation (44) (nk ≤ n ≤ ni | ni ≥ nk ).
In the case of equation (48), the explicit evaluation
of the guard at each cycle executes the behaviour
described here.
′
U Jσ
}
Let us now define (XU)σ = {x−u | x ∈
∈
which allows us to translate the system into
(50)
ρ((XU)′σ )n (v) = ρ(XU)′σ J nT
σ v
which has the same shape as the equations in the
E. Maintaining Geometric Multiplicity
A second step in optimising the number of itera-
previous section. We may now apply the techniques
tions comes from adding granularity to the bounding
semi-spherical abstraction by retaining the geometric
described above to find the bounds on the iterations.
multiplicity using the matrix J b .
D. Narrowing the estimation of the iterations
Lemma 3. Given a matrix A with eigenvalues
{λ s | s ∈ [1, r]}, where each eigenvalue λ s has a
The estimations above are very conservative, but
we may use further techniques to obtain tighter
bounds on the number of iterations. In the first
instance we note that we have eliminated all negative
terms in the sums in equation (44). Reinstating these
geometric multiplicity p s and corresponding generalised eigenvectors {v s,i | i ∈ [1, p s]},
∀n ≥ 0, An vis = λns v s,i +
terms can cause us to lose monotonicity, but we
may still create an iterative approach by fixing the
negative value at intermediate stages. Let ni be our
existing bound for the time horizon before reaching
σ ni
σ ni
P
P
a guard, and kn = pj=1 ki j σj , kni = pj=1 ki j σj
i
the corresponding negative and positive terms of the
equation. We may now find upper and lower bounds
for ni by replacing the equation
ni ≥ nk = logσ (γi ) − logσ kni + kn
k
(51)
where nk is the bound found in the previous stage.
Some stages of this process will provide an unsound
result, but they will also provide an upper bound to
our number of iterations. In fact, every second stage
will provide a monotonically increasing sound bound
which will be tighter than the one in equation (44).
i−1
X
j=1
j
λn−
s
j−1
Y
(n − k)v s,i− j
k=0
j−1
Q
(n
−
k)
i−1
X k=0
n
v
= λ s v s,i +
(52)
s,i− j
j
λs
j=1
Proof: By definition, given an eigenvector v s
of A, then Avs = λ s v s [32]. Similarly a generalised eigenvector v s,i of A satisfies the equation
( A − λ s I) v s,i = v s,i−1 and vs,1 = vs hence
Av s,i = λ s vs,i + v s,i−1
An v s,1 = λns v s,1
An v s,i = An−1 (λ s v s,i + v s,i−1 ) = λ s An−1 vs,i + An−1 v s,i−1
= λ2s An−2 v s,i + λ s An−2 v s,i−1 + An−1 vs,i−1
= · · · = λns v s,i +
n−1
X
j=0
λ sj An− j−1 v s,i−1
18
10,000
From here we recursively expand the formula for
An− j−1 v s,i−1 and obtain:
An v s,i = λns vs,i +
n−1
X
j−1
λ sj λn−
v s,i−1
s
+
j=0
n−1 X
n−2
X
λks An−k−2 v s,i−2
ρ( g)
1,000
n( fσ , fσ )
n( fb , fb )
n( fσ , fb )
100
n( fb , fb )
j=0 k=0
= λns vs,i + nλn−1
s v s,i−1 + n
n−2
X
λ sj An− j−2 vs,i−2
10
j=0
= · · · = λns v s,i +
i−1
X
j
λn−
s
j=1
j−1
Y
(n − k)v s,i− j
1
Let i′ denote the position of fb (λ j ) within the
block J bs it belongs to, such that its corresponding
generalised eigenvector is identified as vbs,i′ = fb (v j ).
Then
ρX0′ (J nT g′i )
pb
X
≤
ki j ρX0 J nb T fb (v j )
j=1
k−1
Q
′
(n − m)
pb
i −1
X
X
m=0
′ −k
v
≤
ki j λb nj ρX0 vbs,i′ +
bs,i
k
λ
b
j
j=1
k=1
k−1
Q
′
(n
−
m)
p
i
−1
b
X
X m=0
′ −k
v
ρ
≤
ki j λb nj ρX0 vbs,i′ +
bs,i
X
0
k
λ
b
j
j=1
k=1
≤
pb
X
ki′ j0 λb nj
j=1
+
i′
X
ki′ jm λb nj
p sY
−i′ −1
m=0
m=1
(n − m)
(53)
I order to manage the product on the right hand
side we use slightly different techniques for over- and
under-approximations. For ni we first find an upper
bound n′i using equation (44) and ki j = ki′ j 0 + ki′ j m
and then do a second iteration using ki j = ki′ j 0 +
′
p s −i
Q−1 ′
ki′ j m
(ni − m) which ensures the true value is
m=0
under the approximation. In the case of ni , we also
ki′ j 0
start with ki j =
iterative process.
+
ki′ j m
0
2
4
and update it during the
8
6
iteration
k=0
10
Fig. 7. Progression of the support function of a system for a
given guard. Blue dots are real values. The dashed green line overapproximates the progression using singular values (sec VI-A),
the dashed yellow line underapproximates them using eigenvalue
norms (sec VI-B), whereas the continuous purple lines represent
the tighter overapproximation maintaining the gemoetric multiplicity (sec VI-E). We can see how the purple line finds a better bound
for ni , while the ni bound is conservative for both approaches.
Mind the logarithmic scale.
Let us look at the following example:
3 0 0 0 0 0
0 2 1 0 0 0
0 0 2 0 0 0
J = 0 0 0 −2 0 0
0 0 0 0 −1 1
S =
0
0
1
0
0
0
0
0
0
3
−4
0
0
0
J σ =
x′0 =
0
0
0
0
1
0
0
0
3
0
0
0
0
2
0
3
0
0
0
1
Gx ≤ 300 | G=
G= 1
−1
0
0
0
1
0
0
0
−1
0
0
0
0
1
1
0
0
0
0
0
1
0
0
0
√
2
1
1
1
1
3
−3
1
1
2
1
2
4
1
4
−3
1
T
S
The progression of the system along the support
function and corresponding bounds as described in
the previous section are shown in figure 7
19
we assume that part of the system is coded, and
further assume that it is possible to discretise the
1,000,000
100,000
physical environment for simulation. Algorithm 1
shows a pseudo-code fragment for the temperature
ρ( g)
10,000
1,000
n( fb , fb )
control problem. We use the read function to
100
Algorithm 1 Temperature Control Loop
States: temp=temperature, heat=heat output.
n( fb , fb )
10
1
Inputs: set=set-point, amb=ambient temperature.
0
5
10
15
20
iteration
Fig. 8. Progression of the support function of a rotational system
for a given guard. Blue dots are real values (negative values are
missing due to the log scale). Continuous purple lines represent
the overapproximation. The steep vertical line at 19 is due to the
alignment of the rotations with the guard at this point. The point
at iteration 14 appears below the line because of the higher point
at iteration 9. The model will either find that this boundary was
met at iteration 9 or push it forward to 19.
1:
temp=5+read(35);
2:
heat=read(1);
while(temp< 400 && heat< 300)
3:
4:
5:
0
0
0
0
0
0
0
0
1.1e0.5i
0
0
0
0
0
0
1.1e−0.5i
we get the results in figure 8. In this case we can
see that the rotational dynamics force an increase of
the initially calculated iteration to account for the
effects of the rotation.
F. Case Study
We have selected a known benchmark to illustrate
the discussed procedure: the room temperature control problem [17]. The temperature (variable temp)
of a room is controlled to a user-defined set point
(set), which can be changed at any time through a
heating (heat) element, and is affected by ambient
temperature (amb) that is out of the control of the
system.
We formalise the description of such a system both
via a linear loop and via hybrid dynamics. Observe
that since such a system may be software controlled,
amb=5+read(35);
7:
set=read(300);
temp=.97 temp + .02 amb + .1 heat;
8:
heat=heat + .05 set;
6:
9:
Changing the eigenvalues to:
2e−0.2i
0
0
0
0
2e0.2i
0
0
√ −0.3i
0
2e
0
0
√ 0.3i
J = 0
2e
0
0
0
0
0
0
{
}
represent non-deterministic values between 0 and
the maximum given as argument. Alternatively, this
loop corresponds to the following hybrid dynamical
model:
0.97 0.1 temp
temp
=
heat k+1
−0.05 1
heat k
0.02
0 amb
,
+
0
0.05
set k
with initial condition
temp
heat
[5 40]
∈
,
[0
1]
0
non-deterministic inputs
amb [5 40]
∈
set k
[0 300]
and guard set
temp
G=
heat
1
|
0
0
1
,
temp 400
<
.
300
heat
In this model the variables are continuous and
take values over the real line, whereas within the
code they are represented as long double precision floating-point values, with precision of ±10−19 ,
20
moreover the error of the approximate Jordan form
computation results in δmax < 10−17 . Henceforth we
300
focus on the latter description, as in the main text of
this work. The eigen-decomposition of the dynamics
200
A = S JS −1 ⊆ SJS−1 where
0.798 ± 10−14 0.173 ± 10−15
S =
0 ± 10−19
0.577 ± 10−14
0.985 ± 10−16 0.069 ± 10−17
J =
−0.069 ± 10−17 0.985 ± 10−16
1.253 ± 10−12 −0.376 ± 10−13
−1
.
S =
0 ± 10−18
1.732 ± 10−12
The discussed over-approximations of the reach-sets
indicate that the temperature variable intersects the
guard at iteration n = 32. Considering the pseudoeigenvalue matrix (described in the extended version
heat
is (the values are rounded to three decimal places):
heat = 300
100
0
0
100
200
temp
300
400
Fig. 9. The abstractly accelerated tube (yellow, dashed boundary),
representing an over-approximation of the thermostat reach tube
(dark blue). The set of initial conditions is shown in black,
whereas successive reach sets are shown in white. The guards
and the reach set that crosses them are close to the boundary in
red.
for the case of complex eigenvalues) along these
iterations, we use Equation (18) to find that the corre-
in Figure 9, where for the sake of clarity we display
sponding complex pair remains within the following
boundaries:
only 8 directions of the 16 constraints. This results
in a rather tight over-approximation that is not much
A32
r
=
−i
r
B32 =
−i
i
r
i
r
0.4144
0.0691
0.1082
0.9159
1
0
1
6.145
<
<
<
<
<
<
<
<
r
i
r+i
i−r
r
i
r+i
i−r
<
<
<
<
<
<
<
<
0.985
0.7651
1.247
0.9389
13.41
17.98
29.44
6.514
looser than the convex hull of all reach sets obtained
by [20] using the given directions. In Figure 9, we
can see the initial set in black colour, the collection
of reach sets in white, the convex hull of all reach
sets in dark blue (as computed by [20]), and finally
the abstractly accelerated set in light yellow (dashed
lines). The outer lines represent the guards.
The reach tube is calculated by multiplying these
abstract matrices with the initial sets of states and inputs, as described in Equation (32), by the following
inequalities:
[5 40]
[5 40]
32
#
32
X̂32 =A
+ B
[0 1]
[0 300]
−24.76 <
temp
< 394.5
"
#
heat
< 253
temp −30.21 <
=
−40.85
<
temp
+
heat
< 616.6
heat
−86.31 < temp − heat < 843.8
G. Calculating the Number of Iterations for Continuous Dynamics
Following the same steps as in Lemma 3 we
develop an equivalent for continuous dynamics.
The negative values represent the lack of restriction
Lemma 4. Given a matrix A with eigenvalues
in the code on the lower side and correspond to system cooling (negative heating). The set is displayed
{λ s | s ∈ [1, r]}, where each eigenvalue λ s has a
geometric multiplicity p s and corresponding gener-
21
alised eigenvectors {v s,i | i ∈ [1, p s]},
∀t ≥ 0, At vis = eλs v s,i +
i−1
X
Table II gives the comparison of our implementation using different levels of precision (long double,
t j eλs vs,i− j
256 bit, and 1024 bit floating-point precision) with
the original abstract acceleration for linear loops
j=1
i−1
X
t j v s,i− j
= eλs v s,i +
(54)
without inputs (J) [34] (where inputs are fixed to
constants). This shows that our implementation gives
Proof: The proof derives again from the taylor
expansion.
tighter over-approximations on most benchmarks
j=1
∞
X
∞
X tk
tk
λks v s,i + kλk−1
v s,i =
s v s,i−1
k!
k!
k=0
k=0
∞ k
∞ k
X
X
t k
t
=
λ s v s,i +
kλk−1
s v s,i−1
k!
k!
k=0
k=0
e At v s,i =
Ak
(column ‘improved’). While on a limited number of
instances the current implementation is less precise
(Fig. 2 gives a hint why this is happening), the
overall increased precision is owed to lifting the lim-
(55)
itation on directions caused by the use of logahedral
abstractions.
At the same time, our implementation is faster –
The rest of the proof follows the same expansion as
in 3
even when used with 1024 bit floating-point precision – than the original abstract acceleration (using
Given the similarity of equation (54) with (52) we
may apply exactly the same techniques described in
rationals). The fact that many bounds have improved
with the new approach, while speed has increased by
section VI-E to the continuous case.
several orders of magnitude, provides evidence of the
advantages of the new approach.
The speed-up is due to the faster Jordan form
= eλi t (vs,i + tv s,i−1 )
VII. Experimental Results
The algorithm has been implemented in C++ using the eigen-algebra package (v3.2), with double
precision floating-point arithmetic, and has been
tested on a 1.6 GHz core 2 duo computer.
computation, which takes between 2 and 65 seconds
for [34] (using the ATLAS package), whereas our
implementation requires at most one second. For the
last two benchmarks, the polyhedral computations
blow up in [34], whereas our support function approach shows only moderately increasing runtimes.
A. Comparison with other unbounded-time ap- The increase of speed is owed to multiple factors,
proaches.
as detailed in Table III. The difference of using long
In a first experiment we have benchmarked our double precision floating-point vs. arbitrary precision
implementation against the tools InterProc [33] and
Sting [12]. We have tested these tools on different
arithmetic is negligible, as all results in the given
examples match exactly to 9 decimal places. Note
scenarios, including guarded/unguarded, stable/un-
that, as explained above, soundness can be ensured
by appropriate rounding in the floating-point compu-
stable and complex/real loops with inputs (details
in Table I).4 It is important to note that in many
tations.
instances, InterProc (due to the limitations of widening) and Sting (due to the inexistence of tight
B. Comparison with bounded-time approaches.
polyhedral, inductive invariants) are unable to infer
finite bounds at all.
In a third experiment, we compare our method
with the LGG algorithm [28] used by SpaceEx [20].
4 The tool and
the benchmarks
http://www.cprover.org/LTI/.
In order to set up a fair comparison we have provided the implementation of the native algorithm
are
available
from
22
characteristics
improved
analysis time [sec]
name
type dim inputs bounds IProc Sti IProc
Sti
J+I
parabola i1
¬s,¬c,g 2
1
80
+25 +28 0.007 237
0.049
parabola i2
¬s,¬c,g 2
1
80
+24 +35 0.008 289
0.072
cubic i1
¬s,¬c,g 3
1
120
+44 +50 0.015 704
0.097
cubic i2
¬s,¬c,g 3
1
120
+35 +55 0.018 699
0.124
oscillator i0
s,c,¬g 2
0
56
+24 +24 0.004
0.990 0.021
oscillator i1
s,c,¬g 2
0
56
+24 +24 0.004
1.060 0.024
inv pendulum
s,c,¬g 4
0
16
+8 +8 0.009
0.920 0.012
convoyCar2 i0 s,c,¬g 3
2
12
+9 +9 0.007
0.160 0.043
convoyCar3 i0 s,c,¬g 6
2
24
+15 +15 0.010
0.235 0.513
convoyCar3 i1 s,c,¬g 6
2
24
+15 +15 0.024
0.237 0.901
convoyCar3 i2 s,c,¬g 6
2
24
+15 +15 0.663
0.271 1.416
convoyCar3 i3 s,c,¬g 6
2
24
+15 +15 0.122
0.283 2.103
type: s – stable loop, c – complex eigenvalues, g – loops with guard; dim: system dimension (variables); bounds: nb. of half-planes
defining the polyhedral set;
IProc is [33]; Sti is [12]; J+I is this work;
improved: number of bounds newly detected by J+I over the existing tools (IProc, Sti)
TABLE I
Experimental comparison of unbounded-time analysis tools with inputs
in [28]. We have run both methods on the con- ∼ 4 ms for each iteration on a 6-dimensional problem
voyCar example [34] with inputs, which presents with octagonal abstraction). This can be improved
an unguarded, scalable, stable loop with complex
dynamics, and focused on octahedral abstractions.
by the use of zonotopes, or by careful selection
of the directions along the eigenvectors, but this
For convex reach sets, the approximations computed
by abstract acceleration are quite tight in comparison
comes at a cost on precision. Even when finding
combinations that outperform our approach, this will
to those computed by the LGG algorithm. However,
storing finite disjunctions of convex polyhedra, the
only allow the time horizon of the LGG approach
to be slightly extended before matching the analysis
LGG algorithm is able to generate non-convex reach
tubes, which are arguably more proper in case of
time from abstract acceleration, and the reachable
states will still remain unknown beyond the extended
oscillating or spiralling dynamics. Still, in many
applications abstract acceleration can provide a tight
time horizon.
over-approximation of the convex hull of those nonconvex reach sets.
Table IV gives the results of this comparison. For
simplicity, we present only the projection of the
The evident advantage of abstract acceleration is
its speed over finite horizons without much precision
loss, and of course the ability to prove properties for
unbounded-time horizons.
bounds along the variables of interest. As expected,
C. Scalability
the LGG algorithm performs better in terms of
tightness, but its runtime increases with the number
Finally, in terms of scalability, we have an expected O(n3 ) complexity worst-case bound (from
of iterations. Our implementation of LGG using
Convex Polyhedra with octagonal templates is slower
than the abstractly accelerated version even for small
time horizons (our implementation of LGG requires
the matrix multiplications in equation (32)). We
have parameterised the number of cars in the convoyCar example [34] (also seen in Table II), and
experimented with up to 33 cars (each car after
23
characteristics
improved
analysis time (sec)
name
type
dim
bounds
tighter
looser
J (jcf)
mpfr+(jcf)
mpfr
ld
parabola i1
¬s,¬c,g
3
80
+4(5%)
0(0%)
2.51
( 2.49) 0.16 (0.06) 0.097 0.007
parabola i2
¬s,¬c,g
3
80
+4(5%)
0(0%)
2.51
( 2.49) 0.26 (0.06) 0.101 0.008
cubic i1
¬s,¬c,g
4
120
0(0%)
0(0%)
2.47
( 2.39) 0.27 (0.20) 0.110 0.013
cubic i2
¬s,¬c,g
4
120
0(0%)
0(0%)
2.49
( 2.39) 0.32 (0.20) 0.124 0.014
oscillator i0
s,c,¬g
2
56
0(0%)
-1(2%)
2.53
( 2.52) 0.12 (0.06) 0.063 0.007
oscillator i1
s,c,¬g
2
56
0(0%)
-1(2%)
2.53
( 2.52) 0.12 (0.06) 0.078 0.008
inv pendulum
s,c,¬g
4
12
+8(50%)
0(0%)
65.78
(65.24) 0.24 (0.13) 0.103 0.012
convoyCar2 i0
s,c,¬g
5
12
+9(45%)
0(0%)
5.46
( 4.69) 3.58 (0.22) 0.258 0.005
convoyCar3 i0
s,c,¬g
8
24
+10(31%)
-2(6%)
24.62
(11.98) 3.11 (1.01) 0.552 0.051
convoyCar3 i1
s,c,¬g
8
24
+10(31%)
-2(6%)
23.92
(11.98) 4.94 (1.01) 0.890 0.121
convoyCar3 i2
s,c,¬g
8
24
+10(31%)
-2(6%)
1717.00
(11.98) 6.81 (1.01) 1.190 0.234
convoyCar3 i3
s,c,¬g
8
24
+10(31%)
-2(6%)
1569.00
(11.98) 8.67 (1.01) 1.520 0.377
type: s – stable loop, c – complex eigenvalues, g – loops with guard; dim: system dimension (including fixed inputs); bounds: nb.
of half-planes defining the polyhedral set; improved: number of bounds (and percentage) that were tighter (better) or looser (worse)
than [34];
J is [34]; mpfr+ is this article using 1024bit mantissas (e < 10−152 ); mpfr uses a 256bit mantissa (e < 10−44 ); ld uses a 64bit
mantissa (e < 10−11 ); here e is the accumulated error of the dynamical system; jcf: time taken to compute Jordan form
TABLE II
Experimental comparison with previous work
the first requires 3 variables, so that for example
(33 − 1) × 3 = 96 variables), and have adjusted
the time bounded analysis is in most cases unsound
since it cannot reason about the unbounded time case
the initial states/inputs sets. We report an average
of 10 runs for each configuration. These results
(we not that a proof of the existence of a fix-point for
the given horizon would restore such soundness by
demonstrate that our method scales to industrial-size
many tools do not attempt to find such proof which
problems.
is left to the user). Unbounded-time solutions are
therefore preferred when such soundness is required,
# of variables
runtime (s)
3
6
12
24
48 96
0.004 0.031 0.062 0.477 5.4 56
although they are often either less precise or slower
than their bounded counterparts.
VIII. Related Work
There are several approaches that solve the safety
problem for the linear and other cases such as hybrid
systems. They are broadly divided into two categories due to the inherent nature of these. Namely
Optimisation
Eigen vs. ATLAS 5
Support functions vs. generators
long double vs. multiple precision arithmetic
interval vs. regular arithmetic
Total
Speed-up
2–10
2–40
5–200
.2–.5
4–80000
TABLE III
Performance improvements by feature
A. Time-Bounded Reachability Analysis
The first approach is to surrender exhaustive analysis over the infinite time horizon, and to restrict
the exploration to system dynamics up to some
given finite time bound. Bounded-time reachability
is decidable, and decision procedures for the resulting satisfiability problem have made much progress
in the past decade. The precision related to the
bounded analysis is offset by the price of uncertainty:
behaviours beyond the given time bound are not
considered, and may thus violate a safety requirement. Representatives are STRONG [15], HySon [7],
24
name
run time
car acceleration
car speed
car position
this article
100 iterations
unbounded
166 ms
166 ms
[-0.820 1.31]
[-1.262 1.31]
[-1.013 5.11]
[-4.515 6.15]
[43.7 83.4]
[40.86 91.9]
LGG
100 iterations 200 iterations 300 iterations
50 ms
140 ms
195 ms
[-0.815 1.31] [-0.968 1.31] [-0.968 1.31]
[-1.013 4.97] [-3.651 4.97] [-3.677 4.97]
[44.5 83.4]
[44.5 88.87] [44.5 88.87]
TABLE IV
Comparison on convoyCar2 benchmark, between this work and the LGG algorithm [28]
CORA [1], HYLAA [3] and SpaceEx [20].
B. Unbounded Reachability Analysis
The second approach, epitomised in static anal-
Set-based simulation methods generalise guaran- ysis methods [30], explores unbounded-time horiteed integration [6], [37]from enclosing intervals to zons. It employs conservative over-approximations to
relational domains. They use precise abstractions
with low computational cost to over-approximate
sets of reachable states up to a given time horizon.
achieve completeness and decidability over infinite
time horizons.
Unbounded techniques attempt to infer a loop
Early tools used polyhedral sets (HyTech [31] and
PHAVer [19]), polyhedral flow-pipes [10], ellip-
invariant, i.e., an inductive set of states that includes
all reachable states. If the computed invariant is
soids [5] and zonotopes [24]. A breakthrough was
disjoint from the set of bad states, this proves that the
been achieved by [25], [28], with the representation of convex sets using template polyhedra and
latter are unreachable and hence that the loop is safe.
However, analysers frequently struggle to obtain an
support functions. This method is implemented in
the tool SpaceEx [20], which can handle dynamical
invariant that is precise enough with acceptable computational cost. The problem is evidently exacerbated
systems with hundreds of variables. Although it
may use exact arithmetic to maintain soundness, it
by non-determinism in the loop, which corresponds
to the case of open systems. Prominent representa-
performs computations using floating-point numbers:
this is a deliberate choice to boost performance,
tives of this analysis approach include Passel [35],
Sting [12], and abstract interpreters such as Astrée
which, although quite reasonable, its implementation
is numerically unsound and therefore does not pro-
[4] and InterProc [33]. Early work in this area has
used implementations of abstract interpretation and
vide genuine formal guarantees. In fact, most tools
using eigendecomposition over a large number of
widening [13], which are still the foundations of
most modern tools. The work in [30] uses abstract
variables (more than 10) are numerically unsound
due to the use of unchecked floating-point arithmetic.
interpretation with convex polyhedra over piecewiseconstant differential inclusions. Dang and Gawl-
Another breakthrough in performance was done by
itza [14] employ optimisation-based (max-strategy
HYLAA [3] which was the first tool to solve all
high order problems of hundreds and thousands
iteration) with linear templates for hybrid systems
with linear dynamics. Relational abstractions [38]
dimensions. Other approaches use specialised con- use ad-hoc “loop summarisation” of flow relations,
straint solvers (HySAT [18], iSAT [16]), or SMT while abstract acceleration focuses on linear relations
encodings [11], [29] for bounded model checking
of hybrid automata.
analysis [26], [27], which is common in program
analysis.
25
C. Abstract Acceleration
References
Abstract acceleration [26], [27], [34] captures the
effect of an arbitrary number of loop iterations with a
single, non-iterative transfer function that is applied
to the entry state of the loop (i.e., to the set of initial
conditions of the linear dynamics). Abstract acceleration has been extended from its original version
to encompass inputs over reactive systems [40] but
restricted to subclasses of linear loops, and later to
general linear loops but without inputs [34].
The work presented in this article lifts these limitations by presenting abstract acceleration for general
linear loops with inputs [8], developing numeric
[1] M. Althoff. An introduction to cora 2015. In ARCH@
CPSWeek, pages 120–151, 2015.
[2] E. Asarin, T. Dang, and A. Girard. Hybridization methods
for the analysis of nonlinear systems. Acta Informatica,
43(7):451–476, 2007.
[3] S. Bak and P. S. Duggirala. Hylaa: A tool for computing
simulation-equivalent reachability for linear systems. In
Proceedings of the 20th International Conference on Hybrid
Systems: Computation and Control, HSCC 2017, Pittsburgh,
PA, USA, April 18-20, 2017, pages 173–178, 2017.
[4] B. Blanchet, P. Cousot, R. Cousot, J. Feret, L. Mauborgne,
A. Miné, D. Monniaux, and X. Rival. A static analyzer
for large safety-critical software. In PLDI, pages 196–207.
ACM, 2003.
techniques for scalability and extending the domain
to continuous time systems.
[5] O. Botchkarev and S. Tripakis. Verification of hybrid
systems with linear differential inclusions using ellipsoidal
approximations. In HSCC, LNCS, pages 73–88. Springer,
2000.
IX. Conclusions and Future Work
[6] O. Bouissou. Analyse statique par interprétation abstraite
de systèmes hybrides. PhD thesis, École Polytechnique,
2008.
We have presented an extension of the Abstract
Acceleration paradigm to guarded LTI systems (linear loops) with inputs, overcoming the limitations of
existing work dealing with closed systems. We have
decisively shown the new approach to over-compete
state-of-the-art tools for unbounded-time reachability
analysis in both precision and scalability. The new
approach is capable of handling general unboundedtime safety analysis for large scale open systems
with reasonable precision and fast computation times.
Conditionals inside loops and nested loops are out
of the scope of this paper.
Work to be done is extending the approach to nonlinear dynamics, which we believe can be explored
via hybridisation techniques [2], and to formalise the
framework for general hybrid models with multiple
guards and location-dependent dynamics, with the
aim to accelerate transitions across guards rather than
integrate individual accelerations on either side of the
guards.
a) Acknowledgments.: We would like to thank
Colas Le Guernic for his constructive suggestions
and comments on the paper.
[7] O. Bouissou, S. Mimram, and A. Chapoutot. Hyson: Setbased simulation of hybrid systems. In Rapid System Prototyping (RSP), 2012 23rd IEEE International Symposium on,
pages 79–85. IEEE, 2012.
[8] D. Cattaruzza, A. Abate, P. Schrammel, and D. Kroening.
Unbounded-time analysis of guarded LTI systems with inputs by abstract acceleration. In SAS, volume 9291 of LNCS,
pages 312–331. Springer, 2015.
[9] D. Cattaruzza, A. Abate, P. Schrammel, and D. Kroening.
Sound numerical computations in abstract acceleration. In
International Workshop on Numerical Software Verification,
pages 38–60. Springer, 2017.
[10] A. Chutinan and B. H. Krogh. Computing polyhedral
approximations to flow pipes for dynamic systems. In CDC,
pages 2089–2094. IEEE Computer Society, 1998.
[11] A. Cimatti, S. Mover, and S. Tonetta. SMT-based verification of hybrid systems. In AAAI Conference on Artificial
Intelligence. AAAI Press, 2012.
[12] M. A. Colón, S. Sankaranarayanan, and H. B. Sipma. Linear
invariant generation using non-linear constraint solving. In
CAV, pages 420–432. Springer, 2003.
[13] P. Cousot and R. Cousot. Abstract interpretation: A unified
lattice model for static analysis of programs by construction
or approximation of fixpoints. In POPL, pages 238–252,
1977.
[14] T. Dang and T. M. Gawlitza. Template-based unbounded
time verification of affine hybrid automata. In APLAS,
LNCS, pages 34–49. Springer, 2011.
[15] Y. Deng, A. Rajhans, and A. A. Julius. STRONG: A
trajectory-based verification toolbox for hybrid systems. In
26
Quantitative Evaluation of Systems, volume 8054 of LNCS,
pages 165–168. Springer, 2013.
[16] A. Eggers, M. Fränzle, and C. Herde. SAT Modulo ODE:
A direct SAT approach to hybrid systems. In ATVA, volume
5311 of LNCS, pages 171–185. Springer, 2008.
[17] A. Fehnker and F. Ivancic. Benchmarks for hybrid systems
verification. In HSCC, pages 326–341. Springer, 2004.
[18] M. Fränzle and C. Herde. HySAT: An efficient proof engine
for bounded model checking of hybrid systems. Formal
Methods in System Design, 30(3):179–198, 2007.
[19] G. Frehse. PHAVer: Algorithmic verification of hybrid
systems past HyTech. In HSCC, volume 3414 of LNCS,
pages 258–273. Springer, 2005.
[20] G. Frehse, C. L. Guernic, A. Donzé, R.
R. Ripado, A. Girard, T. Dang, and O.
Scalable verification of hybrid systems.
6806 of LNCS, pages 379–395. Springer,
Ray, O. Lebeltel,
Maler. SpaceEx:
In CAV, volume
2011.
[21] K. Fukuda and A. Prodon. Double description method
revisited. In Combinatorics and computer science, pages
91–111. Springer, 1996.
[22] S. Gao, J. Avigad, and E. M. Clarke. δ-complete decision
procedures for satisfiability over the reals. In Automated
Reasoning, pages 286–300. Springer, 2012.
[23] P. K. Ghosh and K. V. Kumar. Support function representation of convex bodies, its application in geometric
computing, and some related representations. Computer
Vision and Image Understanding, 72:379–403, 1998.
[24] A. Girard. Reachability of uncertain linear systems using
zonotopes. In HSCC, volume 3414 of LNCS, pages 291–
305. Springer, 2005.
[25] A. Girard, C. L. Guernic, and O. Maler. Efficient computation of reachable sets of linear time-invariant systems with
inputs. In HSCC, volume 3927 of LNCS, pages 257–271.
Springer, 2006.
[26] L. Gonnord and N. Halbwachs. Combining widening and
acceleration in linear relation analysis. In SAS, LNCS, pages
144–160. Springer, 2006.
[27] L. Gonnord and P. Schrammel. Abstract acceleration in
linear relation analysis. Science of Computer Programming,
93(Part B):125–153, 2014.
[28] C. L. Guernic and A. Girard. Reachability analysis of hybrid
systems using support functions. In CAV, volume 5643 of
LNCS, pages 540–554. Springer, 2009.
[29] S. Gulwani and A. Tiwari. Constraint-based approach for
analysis of hybrid systems. In CAV, volume 5123 of LNCS,
pages 190–203. Springer, 2008.
[30] N. Halbwachs, P. Raymond, and Y.-E. Proy. Verification of
linear hybrid systems by means of convex approximations.
In SAS, volume 864 of LNCS, pages 223–237. Springer,
1994.
[31] T. A. Henzinger, P.-H. Ho, and H. Wong-Toi. HyTech: A
model checker for hybrid systems. Journal on Software
Tools for Technology Transfer, 1(1-2):110–122, 1997.
[32] R. A. Horn and C. R. Johnson. Matrix analysis. Cambridge
university press, 2012.
[33] B. Jeannet.
Interproc analyzer for recursive
programs
with
numerical
variables,
2010.
http://pop-art.inrialpes.fr/interproc/interprocweb.cgi.
[34] B. Jeannet, P. Schrammel, and S. Sankaranarayanan. Abstract acceleration of general linear loops. In POPL, pages
529–540. ACM, 2014.
[35] T. T. Johnson and S. Mitra.
Passel: A verification
tool for parameterized networks of hybrid automata, 2012.
https://publish.illinois.edu/passel-tool/.
[36] P. Lancaster and M. Tismenetsky. The Theory of Matrices.
Academic Press, 2nd edition, 1984.
[37] R. Löhner. Einschließung der Lösung gewöhnlicher Anfangsund Randwertaufgaben und Anwendungen. PhD thesis,
Universität Karlsruhe, 1988.
[38] S. Sankaranarayanan and A. Tiwari. Relational abstractions
for continuous and hybrid systems. In CAV, volume 6806
of LNCS, pages 686–702. Springer, 2011.
[39] P. Schrammel. Unbounded-time reachability analysis of
hybrid systems by abstract acceleration. In Embedded
Software, pages 51–54. IEEE, 2015.
[40] P. Schrammel and B. Jeannet. Applying abstract acceleration
to (co-)Reachability analysis of reactive programs. Journal
of Symbolic Computation, 47(12):1512–1532, 2012.
| 3cs.SY
|
arXiv:1610.07253v1 [math.GR] 24 Oct 2016
DUAL ORE’S THEOREM FOR DISTRIBUTIVE
INTERVALS OF SMALL INDEX
SEBASTIEN PALCOUX
Abstract. This paper proves a dual version of a theorem of Oystein Ore for every distributive interval of finite groups [H, G] of
index |G : H| < 9720, and for every boolean interval of rank < 7.
It has applications to representation theory for every finite group.
1. Introduction
Oystein Ore has proved that a finite group is cyclic if and only if its
subgroup lattice is distributive [3]. He has extended one side as follows:
Theorem 1.1 ([3]). Let [H, G] be a distributive interval of finite
groups. Then ∃g ∈ G such that hHgi = G.
We have conjectured the following dual version of this theorem:
Conjecture 1.2. Let [H, G] be a distributive interval of finite groups.
Then ∃V irreducible complex representation of G, with G(V H ) = H
(Definition 3.1); this property will be called linearly primitive.
The interval [1, G] is linearly primitive if and only if G is linearly
primitive (i.e. admits a faithful irreducible complex representation).
We will see that Conjecture 1.2 reduces to the boolean case, because a
distributive interval is bottom boolean (i.e. the interval generated by its
atoms is boolean). As application, Conjecture 1.2 leads to a new bridge
between combinatorics and representation theory of finite groups:
Definition 1.3. Let [H, G] be any interval. We define the combinatorial invariant bbℓ(H, G) as the minimal length ℓ for a chain of subgroups
H = H0 < H1 < · · · < Hℓ = G
with [Hi , Hi+1 ] bottom boolean. Then, let bbℓ(G) := bbℓ(1, G).
Application 1.4. Assuming Conjecture 1.2, bbℓ(G) is a non-trivial
upper bound for the minimal number of irreducible complex representations of G generating (for ⊕ and ⊗) the left regular representation.
2010 Mathematics Subject Classification. 20D60, 05E15, 20C15, 06C15.
Key words and phrases. group; representation; lattice; distributive; boolean.
1
2
SEBASTIEN PALCOUX
Remark 1.5. If the normal subgroups of G are also known, note that
cf ℓ(G) := min{bbℓ(H, G) | H core-free}
is a better upper bound. For more details on the applications, see [1, 4].
This paper is dedicated to prove Conjecture 1.2 for [H, G] boolean
of rank < 7, or distributive of index |G : H| < 9720. For so, we will
use the following new result together with two former results:
Theorem 1.6. Let [H, G] be a boolean interval and L a coatom with
|G : L| = 2. If [H, L] is linearly primitive, then so is [H, G].
Theorem 1.7 ([4]). A distributive interval [H, G] with
n
X
1
≤2
|K
:
H|
i
i=1
for K1 , . . . , Kn the minimal overgroups of H, is linearly primitive.
Theorem 1.8 ([1]). A boolean interval [H, G] with a (below) nonzero
dual Euler totient, is linearly primitive.
X
ϕ̂(H, G) :=
(−1)ℓ(H,K) |G : K|
K∈[H,G]
P
Remark 1.9 ([1]). The Euler totient ϕ(H, G) = (−1)ℓ(K,G) |K : H|
is the number of cosets Hg with hHgi = G, so ϕ > 0 by Theorem 1.1;
but in general ϕ̂ 6= ϕ. We extend ϕ to any distributive interval as
ϕ(H, G) = |T : H| · ϕ(T, G)
with [T, G] the top interval of [H, G], so that for n =
Y
Y
pini −1 · (pi − 1)
ϕ(1, Z/n) =
i
Q
i
pni i ,
i
which is the usual Euler totient ϕ(n). Idem for ϕ̂ and bottom interval.
We will also translate our planar algebraic proof of Theorem 1.7 in
the group theoretic framework (one claim excepted).
Contents
1. Introduction
2. Preliminaries on lattice theory
3. A dual version of Ore’s theorem
4. The proof for small index
5. Acknowledgments
References
1
3
4
7
15
15
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 3
2. Preliminaries on lattice theory
Definition 2.1. A lattice (L, ∧, ∨) is a partially ordered set (or poset)
L in which every two elements a, b have a unique supremum (or join)
a ∨ b and a unique infimum (or meet) a ∧ b.
Example 2.2. Let G be a finite group. The set of subgroups K ⊆ G
is a lattice, denoted by L(G), ordered by ⊆, with K1 ∨ K2 = hK1 , K2 i
and K1 ∧ K2 = K1 ∩ K2 .
Definition 2.3. A sublattice of (L, ∧, ∨) is a subset L′ ⊆ L such that
(L′ , ∧, ∨) is also a lattice. Let a, b ∈ L with a ≤ b, then the interval
[a, b] is the sublattice {c ∈ L | a ≤ c ≤ b}.
Definition 2.4. A finite lattice L admits a minimum and a maximum,
called 0̂ and 1̂.
Definition 2.5. An atom is an element a ∈ L such that
∀b ∈ L, 0̂ < b ≤ a ⇒ a = b.
A coatom is an element c ∈ L such that
∀b ∈ L, c ≤ b < 1̂ ⇒ b = c.
Definition 2.6. The top interval of a finite lattice L is the interval
[t, 1̂] with t the meet of all the coatoms. The bottom interval is the
interval [0̂, b] with b the join of all the atoms.
Definition 2.7. The length of a finite lattice L is the greatest length ℓ
of a chain 0 < a1 < a2 < · · · < aℓ = 1 with ai ∈ L.
Definition 2.8. A lattice (L, ∧, ∨) is distributive if ∀a, b, c ∈ L:
a ∨ (b ∧ c) = (a ∨ b) ∧ (a ∨ c)
(or equivalently, ∀a, b, c ∈ L, a ∧ (b ∨ c) = (a ∧ b) ∨ (a ∧ c)).
Lemma 2.9. The reverse lattice and the sublattices of a distributive
lattice are also distributive. Idem for concatenation and direct product.
Definition 2.10. A distributive lattice is called boolean if any element
b admits a unique complement b∁ (i.e. b ∧ b∁ = 0̂ and b ∨ b∁ = 1̂).
Example 2.11. The subset lattice of {1, 2, . . . , n}, for union and intersection, is called the boolean lattice Bn of rank n (see B3 below).
4
SEBASTIEN PALCOUX
{1, 2, 3}
{1, 2} {1, 3} {2, 3}
{1}
{2}
{3}
∅
Remark 2.12. Any finite boolean lattice is isomorphic to some Bn .
Theorem 2.13 (Birkhoff’s representation theorem or FTFDL [5]).
Any finite distributive lattice embeds into a finite boolean lattice.
Corollary 2.14. The top and bottom intervals of a distributive lattice
are boolean.
Proof. See [5, items a-i p254-255], together with Lemma 2.9.
3. A dual version of Ore’s theorem
In this section, we will state the dual version of Ore’s theorem, and
prove it for any boolean interval of rank ≤ 4, after Theorem 1.7 proof.
Definition 3.1. Let W be a representation of a group G, K a subgroup
of G, and X a subspace of W . We define the fixed-point subspace
W K := {w ∈ W | kw = w , ∀k ∈ K}
and the pointwise stabilizer subgroup
G(X) := {g ∈ G | gx = x , ∀x ∈ X}
Lemma 3.2. [1, Section 3.2] Let G be a finite group, H, K two subgroups, V a complex representation of G and X, Y two subspaces. Then
(1) H ⊆ K ⇒ V K ⊆ V H
(2) X ⊆ Y ⇒ G(Y ) ⊆ G(X)
(3) V H∨K = V H ∩ V K
(4) H ⊆ G(V H )
(5) V G(V H ) = V H
(6) [H ⊆ K and V K ( V H ] ⇒ K 6⊆ G(V H )
Lemma 3.3. [1] Let V1 , . . . , Vr be the irreducible complex representations of a finite group G (up to equivalence), and H a subgroup. Then
|G : H| =
r
X
i=1
dim(Vi ) dim(ViH ).
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 5
Definition 3.4. An interval of finite groups [H, G] is called linearly
primitive if there is an irreducible complex representation V of G such
that G(V H ) = H.
Remark 3.5. The interval [1, G] is linearly primitive iff G is linearly
primitive (i.e. it admits an irreducible faithful complex representation).
The dual version of Ore’s Theorem 1.1 is the following:
Conjecture 3.6. A distributive interval [H, G] is linearly primitive.
Lemma 3.7. A boolean interval [H, G] of rank 1 is linearly primitive.
Proof. Note that [H, G] is of rank 1 iff H is a maximal subgroup of G.
Let V be a non-trivial irreducible complex representation of G with
V H 6= ∅, by Lemma 3.2 (4), H ⊆ G(V H ) . If G(V H ) = G then V must be
trivial (by irreducibility), so by maximality G(V H ) = H.
Lemma 3.8. [1, Lemma 3.37] An interval [H, G] is linearly primitive
if its bottom interval [H, B] is so (see Definition 2.6).
Proposition 3.9. An interval [H, G] satisfying
n
X
1
≤1
|Ki : H|
i=1
with K1 , . . . , Kn the minimal overgroups of H, is linearly primitive.
Proof. First, by Lemmas 3.7, 3.8, we can assume n > 1. By assumption
Pn |G:H|
Pn
i=1 |Ki :H| ≤ |G : H|, so
i=1 |G : Ki | ≤ |G : H|. Let V1 , . . . , Vr be
the irreducible complex representations of G. By Lemma 3.3
n
n X
r
r
n
X
X
X
X
|G : Ki | =
dim(Vα ) dim(VαKi ) =
dim(Vα )[
dim(VαKi )].
i=1 α=1
i=1
If ∀α,
P
Ki
i Vα
=
α=1
i=1
VαH , then
n
X
dim(VαKi ) ≥ dim(VαH ),
i=1
Pn
Pn
and
Pn so i=1 |G : Ki | ≥ |G : H|, but i=1 |G : Ki | ≤ |G : H|, then
i=1 |G : Ki | = |G : H|. So ∀α,
n
X
dim(VαKi ) = dim(VαH ),
i=1
P
but for V1 trivial, we get that n = ni=1 dim(V1Ki ) = dim(V1H ) = 1,
contradiction with n > 1.
P
Else there is α such that i VαKi ( VαH , then by Lemma 3.2 (6),
Ki 6⊆ G(VαH ) ∀i, which means that G(VαH ) = H by minimality.
6
SEBASTIEN PALCOUX
Corollary 3.10. If a subgroup H of G admits at most two minimal
overgroups, then [H, G] is linearly primitive. In particular, a boolean
interval of rank n ≤ 2 is linearly primitive.
P
Proof. i |Ki1:H| ≤ 12 + 12 = 1; the result follows by Proposition 3.9.
We can upgrade Proposition 3.9 in the distributive case as follows:
Theorem 3.11. A distributive interval [H, G] satisfying
n
X
1
≤2
|Ki : H|
i=1
with K1 , . . . , Kn the minimal overgroups of H, is linearly primitive.
Proof. By Lemma 3.8, Corollaries 2.14 and 3.10, we can assume the
interval to be boolean of rank n > 2.
If ∃α such that
X
(⋆)
VαKi∨Kj ( VαH
i,j,i6=j
then by Lemma 3.2 (6), ∀i, j with i 6= j, Ki ∨Kj 6⊆ G(VαH ) . If G(VαH ) = H
then ok, else by the boolean structure and minimality ∃i such that
G(VαH ) = Ki . Now Li := Ki∁ (see Definition 2.10) is a maximal subgroup
of G, so by Lemma 3.7, there is β such that G(V Li ) = Li .
β
Claim: ∃Vγ ≤ Vα ⊗ Vβ such that Ki ∩ G(VγH ) , G(VγH ) ∩ Li ⊆ Ki ∩ Li .
Proof: See the first part of [4, Theorem 6.8] proof; it exploits (⋆) in a
tricky way (we put this reference because we didn’t find an argument
which avoids the use of planar algebras).
By H ⊆ G(VγH ) , distributivity and Claim, we conclude as follows:
G(VγH ) = G(VγH ) ∨ H = G(VγH ) ∨ (Ki ∧ Li ) = (G(VγH ) ∧ Ki ) ∨ (G(VγH ) ∧ Li )
⊆ (Ki ∧ Li ) ∨ (Ki ∧ Li ) = H ∨ H = H
Else, ∀α,
X
VαKi ∨Kj = VαH .
i,j,i6=j
Ki ∨Kj
∀k, ∀(i, j) with i 6= j, ∃s ∈ {i, j} with s 6= k, but Vα
X
VαKs = VαH .
s6=k
It follows that ∀i, ∀α,
X
j6=i
dim(VαKj ) ≥ dim(VαH ).
⊆ VαKs , so
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 7
Now if ∃α∀i, VαKi ( VαH then (by Lemma 3.2 (6) and minimality)
P
K
G(VαH ) = H. Else ∀α∃i, VαKi = VαH , but j6=i dim(Vα j ) ≥ dim(VαH ), so
X
dim(VαKj ) ≥ 2 dim(VαH )
j
By using Lemma 3.3 and taking V1 trivial, we get
X
XX
X
X
|G : Ki | =
[
dim(Vα ) dim(VαKi )] =
dim(Vα )[
dim(VαKi )]
i
i
≥n+2
α
X
α
i
dim(Vα ) dim(VαH ) = 2|G : H| + (n − 2).
α6=1
It follows that
n
X
i=1
n−2
1
≥2+
|Ki : H|
|G : H|
which contradicts the assumption because n > 2.
Corollary 3.12. A rank n boolean interval [H, G] with |Ki : H| ≥ n/2
for any minimal overgroup Ki of H, is linearly primitive. In particular,
a boolean interval of rank n ≤ 4 is linearly primitive.
P
Proof. i |Ki1:H| ≤ n × n2 = 2; the result follows by Theorem 3.11.
In the next section, we get a proof at any rank n < 7.
4. The proof for small index
This section will prove dual Ore’s theorem, for any boolean interval of
rank < 7, and then for any distributive interval of index |G : H| < 9720.
Lemma 4.1. Let [H, G] be a boolean interval of rank 2 and let K, L
the atoms. Then (|G : K|, |G : L|) and (|K : H|, |L : H|) 6= (2, 2).
Proof. If |G : K| = |G : L| = 2, then K and L are normal subgroups
of G, and so H = K ∧ L is also normal. So G/H is a group and
[1, G/H] = [H, G] as lattices, but a boolean lattice is distributive, so
by Ore’s theorem, G/H is cyclic; but it has two subgroups of index 2,
contradiction. If |K : H| = |L : H| = 2, then H is a normal subgroup
of K and L, so of G = H ∨ K, contradiction as above.
Note the following immediate generalization:
Remark 4.2. Let [H, G] be boolean of rank 2, with K and L the atoms.
• If H is a normal subgroup of K and L, then |K : H| =
6 |L : H|.
• If K and L are normal subgroups of G then |G : K| =
6 |G : L|.
8
SEBASTIEN PALCOUX
Remark 4.3. Let G be a finite group and H, K two subgroups, then
|H| · |K| = |HK| · |H ∩ K| (Product Formula). It follows that
|H| · |K| ≤ |H ∨ K| · |H ∧ K|
Corollary 4.4. Let [H, G] be a boolean interval of finite groups and A
an atom. Any K1 , K2 ∈ [H, A∁ ] with K1 ⊂ K2 satisfy
|K1 ∨ A : K1 | ≤ |K2 ∨ A : K2 |
Moreover if |G : A∁ | = 2 then |K ∨ A : K| = 2, ∀K ∈ [H, A∁ ].
Proof. Suppose that K1 ⊂ K2 . By Remark 4.3,
|K1 ∨ A| · |K2 | ≤ |(K1 ∨ A) ∨ K2 | · |(K1 ∨ A) ∧ K2 |
but K1 ∩ K2 = K1 , K1 ∪ K2 = K2 and A ∧ K2 = H, so by distributivity
|K1 ∨ A| · |K2 | ≤ |K2 ∨ A| · |K1 |
Finally, A∁ ∨ A = G and ∀K ∈ [H, A∁ ], K ⊂ A∁ , so if |G : A∁ | = 2,
then
2 ≤ |K ∨ A : K| ≤ |A∁ ∨ A : A∁ | = 2,
It follows that |K ∨ A : K| = 2.
Lemma 4.5. Let [H, G] rank 2 boolean with K, L the atoms. Then
|K : H| = 2 ⇔ |G : L| = 2.
Proof. If |G : L| = 2 then |K : H| = 2 by Corollary 4.4.
If |K : H| = 2 then H ⊳ K and K = H ⊔ Hτ with τ H = Hτ and
(Hτ )2 = H, so Hτ 2 = H and τ 2 ∈ H. Now L ∈ (H, G) open, then
τ Lτ −1 ∈ (τ Hτ −1 , τ Gτ −1 ) = (H, G), so by assumption τ Lτ −1 ∈ {K, L}.
If τ Lτ −1 = K, then L = τ −1 Kτ = K, contradiction. So τ Lτ −1 = L.
Now H = Hτ 2 ⊂ Lτ 2 , and τ 2 ∈ H ⊂ L, so Lτ 2 = L. It follows that
hL, τ i = L ⊔ Lτ . But by assumption, G = hL, τ i, so |G : L| = 2.
Corollary 4.6. If a boolean interval [H, G] admits a subinterval [K, L]
of index 2, then there is an atom A with L = K ∨ A and |G : A∁ | = 2.
Proof. Let [K, L] be the edge of index |L : K| = 2. By the boolean
structure, there is an atom A ∈ [H, G] such that L = K ∨ A. Let
K = K1 < K2 < · · · < Kr = A∁
be a maximal chain from K to A∁ . Let Li = Ki ∨ A, then the interval
[Ki , Li+1 ] is boolean of rank 2, now |L1 : K1 | = 2, so by Lemma 4.5
2 = |L1 : K1 | = |L2 : K2 | = · · · = |Lr : Kr | = |G : A∁ |.
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 9
Remark 4.7. Let [H, G] of index |G : H| = 2. Then G = H ⋊ Z/2 if
|H| is odd, but it’s not true in general if |H| even1.
The following theorem was pointed out by Derek Holt2.
Theorem 4.8. Let G be a finite group, N a normal subgroup of prime
index p and π an irreducible complex represenation of N. Exactly one
of the following occurs:
(1) π extends to an irreducible representation of G,
(2) IndG
N (π) is irreducible.
Proof. It is a corollary of Clifford theory, see [2] Corollary 6.19.
Theorem 4.9. Let [H, G] be a boolean interval and L a coatom with
|G : L| = 2. If [H, L] is linearly primitive, then so is [H, G].
Proof. Let the atom A := L∁ . As an immediate corollary of the proofs
of Lemma 4.5 and Corollary 4.6, there is τ ∈ A such that ∀K ∈ [H, L],
Kτ = τ K and τ 2 ∈ H ⊂ K, so K ∨ A = K ⊔ Kτ and G = L ⊔ Lτ . By
assumption, [H, L] is linearly primitive, which means the existence of
an irreducible complex representation V of L such that L(V H ) = H.
Assume that πV extends to an irreducible representation πV+ of G.
Note that G(V+H ) = H ⊔ Sτ with
S = {l ∈ L | πV+ (lτ ) · v = v, ∀v ∈ V H }
If S = ∅ then G(V+H ) = H, ok. Else S 6= ∅ and note that
πV+ (lτ ) · v = v ⇔ πV+ (τ ) · v = πV (l−1 ) · v
but πV+ (τ )(V H ) ⊂ V H and τ 2 ∈ H, so ∀l1 , l2 ∈ S and ∀v ∈ V H ,
πV (l1 l2 )−1 · v = πV+ (τ 2 ) · v = v
It follows that S 2 ⊂ H. Now, HS = S, so HS 2 = (HS)S = S 2 ,
which means that S 2 is a disjoint union of H-coset, then |H| divides
|S 2 |, but S 2 ⊂ H and S 6= ∅, so S 2 = H. Let s0 ∈ S, then the maps
S ∋ s 7→ s0 s ∈ H and H ∋ h 7→ hs0 ∈ S are injective, so |S| = |H|. If
S 6= H, then A = H ⊔Hτ and G(V+H ) = H ⊔Sτ are two different groups
containing H with index 2, contradiction with the boolean structure
by Lemma 4.1. So we can assume that H = S. Now the extension
V+ is completely characterized by πV+ (τ ), and we can make an other
irreducible extension V− characterized by πV− (τ ) = −πV+ (τ ). As above,
G(V−H ) = H ⊔ S ′ τ with
S ′ = {l ∈ L | πV− (lτ ) · v = v, ∀v ∈ V H }.
1http://math.stackexchange.com/a/1609599/84284
2http://math.stackexchange.com/a/1966655/84284
10
SEBASTIEN PALCOUX
But πV− (lτ ) = −πV+ (lτ ), so
S ′ = {l ∈ L | πV+ (lτ ) · v = −v, ∀v ∈ V H }.
Then S ∩ S ′ = ∅, but S = H, so S ′ 6= H, contradiction as above.
Next, we can assume that πV does not extend to an irreducible representation of G. So πW := IndG
L (πV ) is irreducible by Theorem 4.8.
We need to check that G(W H ) = H. We can see W as V ⊕ τ V , with
πW (l) · (v1 + τ v2 ) = πV (l) · v1 + τ [πV (τ −1 lτ ) · v2 ],
with l ∈ L, and
πW (τ ) · (v1 + τ v2 ) = πV (τ 2 ) · v2 + τ v2
Then
W H = {v1 +τ v2 ∈ W | πV (h)·v1 = v1 and πV (τ −1 hτ )·v2 = v2 , ∀h ∈ H}
But τ −1 Hτ = H, so W H = V H ⊕ τ V H . Finally, according to πW (l)
and πW (τ ) above, we see that G(W H ) ⊂ L, and then G(W H ) = H.
Remark 4.10. It seems that we can extend Theorem 4.9, replacing
|G : L| = 2 by L⊳G (and so |G : L| = p prime), using Theorem 4.8 and
Remark 4.2. In the proof, we should have K ∨A = K ⊔Kτ ⊔· · ·⊔Kτ p−1 ,
τ p ∈ H, S p = H and πV− (τ ) = e2πi/p πV+ (τ ). We didn’t check the details
because we don’t need this extension.
Corollary 4.11. Let [H, G] be a boolean interval with an atom A satisfying |A : H| = 2. If [H, A∁ ] is linearly primitive, then so is [H, G].
Proof. Immediate by Corollary 4.6 and Theorem 4.9.
One of the main result of the paper is the following:
Theorem 4.12. A boolean interval [H, G] of rank n < 7, is linearly
primitive.
Proof. Let K1 , . . . , Kn be the atoms of [H, G]. By Corollary 4.11, we
can assume that |Ki : H| =
6 2, ∀i. Now n ≤ 6 and |Ki : H| ≥ 3, then
n
X
i=1
1
1
≤ 6 × = 2.
|Ki : H|
3
The result follows by Theorem 3.11.
For the upper bound on the index of distributive interval we will
need a former result (proved group theoretically in [1]):
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 11
Theorem 4.13. [1, Theorem 3.24] A boolean interval [H, G] with a
(below) nonzero dual Euler totient is linearly primitive.
X
ϕ̂(H, G) :=
(−1)ℓ(H,K) |G : K|
K∈[H,G]
with ℓ(H, K) the rank of [H, K].
Conjecture 4.14. A rank n boolean interval has ϕ̂ ≥ 2n−1 .
Remark 4.15. If Conjecture 4.14 is correct, then its lower bound is
optimal, because realized by the interval [1 × S2n , S2 × S3n ].
Lemma
4.16. Let [H, G]Pbe a boolean interval of rank n and index
Q ri
pi with pi prime and i ri = n. Then for any atom A and any
K ∈ [H, A∁ ], |K ∨ A : K| = pi for some i.
W
Proof. Let A1 , · · · , Ar be the atoms of [H, G] such that K = ri=1 Ai ,
let Ar+1 = A and Ar+2 , . . . , An all the other atoms. By considering the
corresponding maximal chain we have that
|G : H| = |A1 : H| · |A1 ∨ A2 : A1 | · · · |K ∨ A : K| · · · |G : A∁n−1 |
It’s a product of n numbers > 1 and the result is composed by n prime
numbers, so by the fundamental theorem of arithmetic, any component
above is prime, then |K ∨ A : K| = pi for some i.
Lemma 4.17. Let [H, G] be a boolean interval of rank n and index pn
with p prime. Then ϕ̂(H, G) = (p − 1)n > 0.
P
Proof. By Lemma 4.16, ϕ̂(H, G) = k (−1)k nk pk = (p − 1)n
Remark 4.18. Lemma 4.17 is coherent with Conjecture 4.14 because
if p = 2 then n = 1 by Lemma 4.1.
Proposition 4.19. Let [H, G] be a boolean interval of rank n and index
pn−1 q, with p, q prime and p ≤ q. Then
q−p
1
ϕ̂(H, G) = (p − 1)n [1 +
(1 −
)] ≥ (p − 1)n > 0.
p
(1 − p)m
with m be the number of coatoms L ∈ [H, G] with |G : L| = q.
Proof. If m = 0, then by Lemma 4.16, Corollary 4.4 and p ≤ q, for any
atom A ∈ [H, G] and ∀K ∈ [H, A∁ ], |K ∨ A : K| = p, so |G : H| = pn
and ϕ̂(H, G) = (p − 1)n by Lemma 4.17, ok.
Else m ≥ 1. We will prove the formula by induction. If n = 1, then
m = 1 and ϕ̂(H, G) = q − 1, ok. Next, assume it is true at rank < n.
Let L be a coatom with |G : L| = q, then for A = L∁ ,
ϕ̂(H, G) = q ϕ̂(H, L) − ϕ̂(A, G)
12
SEBASTIEN PALCOUX
Now |L : H| = pn−1 so by Lemma 4.17, ϕ̂(H, L) = (p − 1)n−1 . But
|A : H| = p or q. If |A : H| = p then |G : A| = pn−2 q and by induction
1
q−p
(1 −
)].
ϕ̂(A, G) = (p − 1)n−1 [1 +
p
(1 − p)m−1
Else |A : H| = q, |G : A| = pn−1 , m = 1 and the same formula works.
Then
q−p
1
ϕ̂(H, G) = (p − 1)n−1 [q − 1 −
(1 −
)]
p
(1 − p)m−1
1
q−1 q−p 1
+
(
−
)]
= (p − 1)n [
p−1
p 1 − p (1 − p)m
1
q−p
1
q−1 q−p
−
(1 +
)+
(1 −
)]
= (p − 1)n [
p−1
p
p−1
p
(1 − p)m
1
q−p
(1 −
)]
= (p − 1)n [1 +
p
(1 − p)m
The result follows.
Definition 4.20. A chain H1 ⊂ · · · ⊂ Hr+1 is of type (k1 , . . . , kr ) if
∃σ ∈ Sr with kσ(i) = |Hi+1 : Hi | (so that we can choose (ki )i increasing).
Remark 4.21. The proof of Proposition 4.19 is working without assuming p, q prime, but assuming type (p, . . . , p, q) for every maximal
chain of [H, G]. For p prime and q = p2 we deduce that at rank n and
index pn+1 , there is 1 ≤ m ≤ n such that
ϕ̂(H, G) = (p − 1)n+1 + (p − 1)n − (−1)m (p − 1)n+1−m ≥ (p − 1)n+1
If there is no edge of index 2, we can also take q = 2p or (p, q) = (3, 4).
Lemma 4.22. A boolean interval [H, G] of index |G : H| = an bc and
rank n+2 with 3 ≤ a ≤ b ≤ c ≤ 12, 1 ≤ n ≤ 6 and every maximal chain
of type (a, . . . , a, b, c), has a dual Euler totient ϕ̂(H, G) ≥ (a − 1)n+2 .
Proof. This is checked by computer calculation using the following iterative method. Let L be a coatom just that |G : L| = c and A = L∁ .
Then ϕ̂(H, G) = cϕ̂(H, L) − ϕ̂(A, G). Now |L : H| = an b so we can use
Propoposition 4.19 formula for ϕ̂(H, L). Next there are three cases:
|A : H| = a, b or c. If |A : H| = c then, by Corollary 4.4, ∀K ∈ [H, L],
|K ∨ A : K| = c, so ϕ̂(H, G) = (c − 1)ϕ̂(H, L). If |A : H| = b, then
|G : A| = an c so we can use Propoposition 4.19 formula for ϕ̂(A, G).
Else |A : H| = a and |G : A| = an−1 bc, so we iterate the method.
Remark 4.23. Let [H, G] be a boolean interval and A an atom such
that ∀K ∈ [H, A∁ ], |K ∨ A : K| = |A : H|. So ϕ̂(H, A∁ ) = ϕ̂(A, G) and
ϕ̂(H, G) = |A : H|ϕ̂(H, A∁ ) − ϕ̂(A, G) = (|A : H| − 1)ϕ̂(A, G).
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 13
Corollary 4.24. Let [H, G] be a boolean interval such that for any
atom A and ∀K ∈ [H, A∁ ], |K ∨ A : K| = |A : H|. Then
n
Y
ϕ̂(H, G) =
(|Ai : H| − 1) > 0.
i=1
with A1 , . . . , An all the atoms of [H, G].
Proof. By Remark 4.23 and induction.
Lemma 4.25. Let [H, G] boolean of rank 2 and index < 32. Let K, L
be the atoms, a = |G : K|, b = |G : L|, c = |L : H| and d = |K : H|.
a
G
b
K
L
d
H
c
If a 6= 7, then (a, b) = (c, d).
If a = 7 and a 6= c then a = b = 7 and c = d ∈ {3, 4}.
Proof. We can check by GAP3 that there are exactly 241 boolean intervals [H, G] of rank 2 and index |G : H| < 32 (up to equivalence).
They all satisfy (a, b) = (c, d), except [D8 , P SL2 (7)] and [S3 , P SL2 (7)],
for which (a, b) = (7, 7) and (c, d) = (3, 3) or (4, 4).
Corollary 4.26. Let [H, G] be a boolean interval having a maximal
chain such that the product of the index of two different edges is < 32,
and no edge has index 7. Then [H, G] satisfies Corollary 4.24.
Proof. Consider such a maximal chain
H = K0 ⊂ K1 ⊂ · · · ⊂ Kn = G
and A1 , . . . , An the atoms of [H, G] such that Ki = Ki−1 ∨ Ai . Now, ∀i
and ∀j < i, [Kj−1, Kj ∨ Ai ] is boolean of rank 2, so by Lemma 4.25,
|Ki : Ki−1 | = |Ki−2 ∨ Ai : Ki−2 | = |Ki−3 ∨ Ai : Ki−2 | = · · · = |Ai : H|
Next, ∀i and ∀j ≥ i, let Lj−1 = Kj ∧ A∁i , then [Lj , Kj+2 ] is boolean of
rank 2 and by Lemma 4.25,
|Ki : Ki−1 | = |Ki+1 : Li | = |Ki+2 : Li+1 | = · · · = |G : A∁i |
Finally, by Corollary 4.4, ∀K ∈ [H, A∁i ],
|Ai : H| ≤ |K ∨ Ai : K| ≤ |G : A∁i |
but |Ai : H| = |Ki : Ki−1 | = |G : A∁i |; the result follows.
3The
GAP Group, http://www.gap-system.org, version 4.8.3, 2016.
14
SEBASTIEN PALCOUX
Remark 4.27. A combinatorial argument could replace the use of
Corollary 4.4 in the proof of Corollary 4.26.
Remark 4.28. Here is the list of all the numbers < 10125 which are
product of at least seven integers ≥ 3; first with exactly seven integers:
2187 = 37
2916 = 36 4
3645 = 36 5
3888 = 35 42
4374 = 36 6
4860 = 35 41 5
5103 = 36 7
5184 = 34 43
5832 = 36 8
6075 = 35 52
6480 = 34 42 5
6561 = 36 9
6804 = 35 41 7
6912 = 33 44
7290 = 36 10
7776 = 35 41 8
8019 = 36 11
8100 = 34 41 52
8505 = 35 51 7
8640 = 33 43 5
8748 = 36 12
9072 = 34 42 7
9216 = 32 45
9477 = 36 13
9720 = 35 41 10;
next with exactly eight integers: 6561 = 38 , 8748 = 374; nothing else.
We can now prove the main theorem of the paper:
Theorem 4.29. A distributive interval [H, G] of index |G : H| < 9720,
is linearly primitive.
Proof. By Lemma 3.8, Corollary 2.14 and Theorem 4.12, we can assume
the interval to be boolean of rank n ≥ 7, and without edge of index
2 by Corollary 4.6 and Theorem 4.9. So by Theorem 4.13, it suffices
to check that for every index (except 9720) in the list of Remark 4.28,
any boolean interval as above with this index has a nonzero dual Euler
totient. We can assume the rank to be 7, because at rank 8, the indices
38 and 37 4 are checked by Lemma 4.17 and Remark 4.21, and there is
nothing else at rank > 8. Now, any maximal chain for such a boolean
interval of index 35 41 5 has type (3, . . . , 3, 4, 5), so it is checked by Corollary 4.26. Idem for index 36 10 with (3, . . . , 3, 10) or (3, . . . , 3, 5, 6). The
index 36 7 is checked by Proposition 4.19. For the index 36 12, if there is
a maximal chain of type (3, . . . , 3, 6, 6), 62 > 32 but using Lemma 4.25
with a, b, c, d ∈ {3, 6} we can deduce that (a, b) = (c, d), so the proof
of Corollary 4.26 is working; else 12 must appears in every maximal
chain, so that the proof of Proposition 4.19 works with q = 12. We can
do the same for every index, except 35 41 7, 35 41 8, 35 51 7, 34 42 7, 35 41 10.
For index 35 41 8, if there is a maximal chain of type (3, . . . , 3, 4, 4, 6),
then ok by Corollary 4.26, else (because there is no edge of index 2)
every maximal chain is of type (3, . . . , 3, 4, 8), so ok by Lemma 4.22.
We can do the same for every remaining index except 35 41 10 = 9720,
the expected upper bound.
Remark 4.30. The tools above don’t check 35 41 10 because the possible maximal chain types are (3, . . . , 3, 4, 5, 6), (3, . . . , 3, 4, 10) and
(3, . . . , 3, 5, 8). The first is ok by Corollary 4.26, but not the two last
because 4 · 10 = 5 · 8 = 40 > 32. So there is not necessarily a unique
maximal chain type, and Lemma 4.22 can’t be applied. Nevertheless,
more intensive computer investigation can probably leads beyond 9720.
DUAL ORE’S THEOREM FOR DISTRIBUTIVE INTERVAL, SMALL INDEX 15
5. Acknowledgments
I would like to thank Derek Holt for showing me a theorem on representation theory used in this paper. This work is supported by the
Institute of Mathematical Sciences, Chennai.
References
[1] Mamta Balodi and Sebastien Palcoux, On boolean intervals of finite groups,
arXiv:1604.06765v5, 25pp. submitted to Trans. Amer. Math. Soc.
[2] I. Martin Isaacs, Character theory of finite groups, Dover Publications, Inc.,
New York, 1994. Corrected reprint of the 1976 original [Academic Press, New
York; MR0460423 (57 #417)]. MR1280461
[3] Oystein Ore, Structures and group theory. II, Duke Math. J. 4 (1938), no. 2,
247–269, DOI 10.1215/S0012-7094-38-00419-3. MR1546048
[4] Sebastien Palcoux, Ore’s theorem for cyclic subfactor planar algebras and applications, arXiv:1505.06649v10, 50pp. submitted to Pacific J. Math.
[5] Richard P. Stanley, Enumerative combinatorics. Volume 1, 2nd ed., Cambridge
Studies in Advanced Mathematics, vol. 49, Cambridge University Press, Cambridge, 2012. MR2868112
Institute of Mathematical Sciences, Chennai, India
E-mail address: [email protected]
| 4math.GR
|
Journal of Scheduling manuscript No.
(will be inserted by the editor)
A parameterized complexity view on non-preemptively scheduling
interval-constrained jobs: few machines, small looseness, and small slack
arXiv:1508.01657v2 [cs.DS] 24 Mar 2016
René van Bevern · Rolf Niedermeier · Ondřej Suchý
Submitted: August 7, 2015
Accepted: March 23, 2016
Abstract We study the problem of non-preemptively scheduling n jobs, each job j with a release time t j , a deadline d j ,
and a processing time p j , on m parallel identical machines.
Cieliebak et al (2004) considered the two constraints |d j −
t j | ≤ λ p j and |d j − t j | ≤ p j + σ and showed the problem to
be NP-hard for any λ > 1 and for any σ ≥ 2. We complement their results by parameterized complexity studies: we
show that, for any λ > 1, the problem remains weakly NPhard even for m = 2 and strongly W[1]-hard parameterized
by m. We present a pseudo-polynomial-time algorithm for
constant m and λ and a fixed-parameter tractability result for
the parameter m combined with σ .
Keywords release times and deadlines · machine minimization · sequencing within intervals · shiftable intervals ·
fixed-parameter tractability · NP-hard problem
1 Introduction
Non-preemptively scheduling jobs with release times and
deadlines on a minimum number of machines is a well-studied problem both in offline and online variants (Chen et al
René van Bevern is supported by grant 16-31-60007 mol a dk of the
Russian Foundation for Basic Research (RFBR).
Ondřej Suchý is supported by grant 14-13017P of the Czech Science
Foundation.
René van Bevern
Novosibirsk State University, Novosibirsk, Russian Federation, E-mail:
[email protected]
Rolf Niedermeier
Institut für Softwaretechnik und Theoretische Informatik, TU Berlin,
Germany, E-mail: [email protected]
Ondřej Suchý
Faculty of Information Technology, Czech Technical University in
Prague, Prague, Czech Republic, E-mail: [email protected]
2016; Chuzhoy et al 2004; Cieliebak et al 2004; Malucelli
and Nicoloso 2007; Saha 2013). In its decision version, the
problem is formally defined as follows:
I NTERVAL -C ONSTRAINED S CHEDULING
Input: A set J := {1, . . . , n} of jobs, a number m ∈ N of
machines, each job j with a release time t j ∈ N, a deadline d j ∈ N, and a processing time p j ∈ N.
Question: Is there a schedule that schedules all jobs onto
m parallel identical machines such that
1. each job j is executed non-preemptively for p j time units,
2. each machine executes at most one job at a time, and
3. each job j starts no earlier than t j and is finished by d j .
For a job j ∈ J, we call the half-open interval [t j , d j ) its time
window. A job may only be executed during its time window.
The length of the time window is d j − t j .
We study I NTERVAL -C ONSTRAINED S CHEDULING with
two additional constraints introduced by Cieliebak et al (2004).
These constraints relate the time window lengths of jobs to
their processing times:
Looseness If all jobs j ∈ J satisfy |d j − t j | ≤ λ p j for some
number λ ∈ R, then the instance has looseness λ . By λ L OOSE I NTERVAL -C ONSTRAINED S CHEDULING we denote
the problem restricted to instances of looseness λ .
Slack If all jobs j ∈ J satisfy |d j − t j | ≤ p j + σ for some
number σ ∈ R, then the instance has slack σ . By σ -S LACK
I NTERVAL -C ONSTRAINED S CHEDULING we denote the
problem restricted to instances of slack σ .
Both constraints on I NTERVAL -C ONSTRAINED S CHEDUL ING are very natural: clients may accept some small deviation
of at most σ from the desired start times of their jobs. Moreover, it is conceivable that clients allow for a larger deviation
2
for jobs that take long to process anyway, leading to the case
of bounded looseness λ .
Cieliebak et al (2004) showed that, even for constant λ >
1 and constant σ ≥ 2, the problems λ -L OOSE I NTERVAL C ONSTRAINED S CHEDULING and σ -S LACK I NTERVAL C ONSTRAINED S CHEDULING are strongly NP-hard.
Instead of giving up on finding optimal solutions and
resorting to approximation algorithms (Chuzhoy et al 2004;
Cieliebak et al 2004), we conduct a more fine-grained complexity analysis of these problems employing the framework of parameterized complexity theory (Cygan et al 2015;
Downey and Fellows 2013; Flum and Grohe 2006; Niedermeier 2006), which so far received comparatively little attention in the field of scheduling with seemingly only a handful
of publications (van Bevern et al 2015a,b; Bodlaender and
Fellows 1995; Cieliebak et al 2004; Fellows and McCartin
2003; Halldórsson and Karlsson 2006; Hermelin et al 2015;
Mnich and Wiese 2015). In particular, we investigate the
effect of the parameter m of available machines on the parameterized complexity of interval-constrained scheduling
without preemption.
Related work I NTERVAL -C ONSTRAINED S CHEDULING is
a classical scheduling problem and strongly NP-hard already
on one machine (Garey and Johnson 1979, problem SS1).
Besides the task of scheduling all jobs on a minimum number
of machines, the literature contains a wide body of work
concerning the maximization of the number of scheduled
jobs on a bounded number of machines (Kolen et al 2007).
For the objective of minimizing the number
p of machines,
Chuzhoy et al (2004) developed a factor-O( log n/ log log n)approximation algorithm. Malucelli and Nicoloso (2007) formalized machine minimization and other objectives in terms
of optimization problems in shiftable interval graphs. Online
algorithms for minimizing the number of machines have been
studied as well and we refer to recent work by Chen et al
(2016) for an overview.
Our work refines the following results of Cieliebak et al
(2004), who considered I NTERVAL -C ONSTRAINED S CHED ULING with bounds on the looseness and the slack. They
showed that I NTERVAL -C ONSTRAINED S CHEDULING is
strongly NP-hard for any looseness λ > 1 and any slack σ ≥
2. Besides giving approximation algorithms for various special cases, they give a polynomial-time algorithm for σ = 1
and a fixed-parameter tractability result for the combined
parameter σ and h, where h is the maximum number of time
windows overlapping in any point in time.
Our contributions We analyze the parameterized complexity
of I NTERVAL -C ONSTRAINED S CHEDULING with respect
to three parameters: the number m of machines, the looseness λ , and the slack σ . More specifically, we refine known
René van Bevern et al.
results of Cieliebak et al (2004) using tools of parameterized
complexity analysis. An overview is given in Table 1.1.
In Section 3, we show that, for any λ > 1, λ -L OOSE
I NTERVAL -C ONSTRAINED S CHEDULING remains weakly
NP-hard even on m = 2 machines and that it is strongly W[1]hard when parameterized by the number m of machines. In
Section 4, we give a pseudo-polynomial-time algorithm for λ L OOSE I NTERVAL -C ONSTRAINED S CHEDULING for each
fixed λ and m. Finally, in Section 5, we give a fixed-parameter algorithm for σ -S LACK I NTERVAL -C ONSTRAINED
S CHEDULING when parameterized by m and σ . This is in
contrast to our result from Section 3 that the parameter combination m and λ presumably does not give fixed-parameter
tractability results for λ -L OOSE I NTERVAL -C ONSTRAINED
S CHEDULING.
2 Preliminaries
Basic notation We assume that 0 ∈ N. For two vectors u =
(u1 , . . . , uk ) and v = (v1 , . . . , vk ), we write u ≤ v if ui ≤ vi
for all i ∈ {1, . . . , k}. Moreover, we write u v if u ≤ v and
u 6= v, that is, u and v differ in at least one component. Finally,
1k is the k-dimensional vector consisting of k 1-entries.
Computational complexity We assume familiarity with the
basic concepts of NP-hardness and polynomial-time manyone reductions (Garey and Johnson 1979). We say that a
problem is (strongly) C-hard for some complexity class C if it
is C-hard even if all integers in the input instance are bounded
from above by a polynomial in the input size. Otherwise, we
call it weakly C-hard.
In the following, we introduce the basic concepts of parameterized complexity theory, which are in more detail
discussed in corresponding text books (Cygan et al 2015;
Downey and Fellows 2013; Flum and Grohe 2006; Niedermeier 2006).
Fixed-parameter algorithms The idea in fixed-parameter algorithmics is to accept exponential running times, which are
seemingly inevitable in solving NP-hard problems, but to
restrict them to one aspect of the problem, the parameter.
Thus, formally, an instance of a parameterized problem Π is a pair (x, k) consisting of the input x and the
parameter k. A parameterized problem Π is fixed-parameter tractable (FPT) with respect to a parameter k if there
is an algorithm solving any instance of Π with size n in
f (k) · poly(n) time for some computable function f . Such
an algorithm is called a fixed-parameter algorithm. It is potentially efficient for small values of k, in contrast to an
algorithm that is merely running in polynomial time for each
fixed k (thus allowing the degree of the polynomial to depend on k). FPT is the complexity class of fixed-parameter
tractable parameterized problems.
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
3
Table 1.1 Overview of results on I NTERVAL -C ONSTRAINED S CHEDULING for various parameter combinations. The parameterized complexity
with respect to the combined parameter λ + σ remains open.
Combined
with
λ
Parameter
looseness λ
slack σ
NP-hard for any λ > 1
(Cieliebak et al 2004)
σ
number m of machines
?
NP-hard for any σ ≥ 2
(Cieliebak et al 2004)
m
We refer to the sum of parameters k1 +k2 as the combined
parameter k1 and k2 .
Parameterized intractability To show that a problem is presumably not fixed-parameter tractable, there is a parameterized analog of NP-hardness theory. The parameterized analog
of NP is the complexity class W[1] ⊇ FPT, where it is conjectured that FPT 6= W[1]. A parameterized problem Π with
parameter k is called W[1]-hard if Π being fixed-parameter
tractable implies W[1] = FPT. W[1]-hardness can be shown
using a parameterized reduction from a known W[1]-hard
problem: a parameterized reduction from a parameterized
problem Π1 to a parameterized problem Π2 is an algorithm
mapping an instance I with parameter k to an instance I 0 with
parameter k0 in time f (k) · poly(|I|) such that k0 ≤ g(k) and
I 0 is a yes-instance for Π1 if and only if I is a yes-instance
for Π2 , where f and g are arbitrary computable functions.
3 A strengthened hardness result
In this section, we strengthen a hardness result of Cieliebak
et al (2004), who showed that λ -L OOSE I NTERVAL -C ON STRAINED S CHEDULING is NP-hard for any λ > 1. This
section proves the following theorem:
Theorem 3.1 Let λ : N → R be such that λ (n) ≥ 1 + n−c
for some integer c ≥ 1 and all n ≥ 2.
Then λ (n)-L OOSE I NTERVAL -C ONSTRAINED S CHED ULING of n jobs on m machines is
(i) weakly NP-hard for m = 2, and
(ii) strongly W[1]-hard for parameter m.
Note that Theorem 3.1, in particular, holds for any constant
function λ (n) > 1.
We remark that Theorem 3.1 cannot be proved using the
NP-hardness reduction given by Cieliebak et al (2004), which
reduces 3-S AT instances with k clauses to I NTERVAL -C ON STRAINED S CHEDULING instances with m = 3k machines.
Since 3-S AT is trivially fixed-parameter tractable for the
W[1]-hard for parameter m for any λ > 1 (Theorem 3.1),
weakly NP-hard for m = 2 and any λ > 1 (Theorem 3.1),
pseudo-polynomial time for fixed m and λ (Theorem 4.1)
fixed-parameter tractable for parameter σ + m (Theorem 5.1)
NP-hard for m = 1 (Garey and Johnson 1979)
parameter number k of clauses, the reduction of Cieliebak
et al (2004) cannot yield Theorem 3.1.
Instead, to prove Theorem 3.1, we give a parameterized
polynomial-time many-one reduction from B IN PACKING
with m bins and n items to λ (mn)-L OOSE I NTERVAL -C ON STRAINED S CHEDULING with m machines and mn jobs.
B IN PACKING
Input: A bin volume V ∈ N, a list a1 , . . . , an ∈ N of items,
and a number m ≤ n of bins.
Question: Is there a partition S1 ] · · · ] Sm = {1, . . . , n} such
that ∑i∈Sk ai ≤ V for all 1 ≤ k ≤ m?
Since B IN PACKING is weakly NP-hard for m = 2 bins and
W[1]-hard parameterized by m even if all input numbers are
polynomial in n (Jansen et al 2013), Theorem 3.1 will follow.
Our reduction, intuitively, works as follows: for each
of the n items ai in a B IN PACKING instance with m bins
of volume V , we create a set Ji := { ji1 , . . . , jim } of m jobs
that have to be scheduled on m mutually distinct machines.
Each machine represents one of the m bins in the B IN PACK ING instance. Scheduling job ji1 on a machine k corresponds
to putting item ai into bin k and will take B + ai time of
machine k, where B is some large integer chosen by the
reduction. If ji1 is not scheduled on machine k, then a job
in Ji \ { ji1 } has to be scheduled on machine k, which will
take only B time of machine k. Finally, we choose the latest
deadline of any job as nB +V . Thus, since all jobs have to
be finished by time nB +V and since there are n items, for
each machine k, the items ai for which ji1 is scheduled on
machine k must sum up to at most V in a feasible schedule.
This corresponds to satisfying the capacity constraint of V of
each bin.
Formally, the reduction works as follows and is illustrated
in Figure 3.1.
Construction 3.2 Given a B IN PACKING instance I with
n ≥ 2 items a1 , . . . , an and m ≤ n bins, and λ : N → R such
that λ (n) ≥ 1 + n−c for some integer c ≥ 1 and all n ≥ 2,
we construct an I NTERVAL -C ONSTRAINED S CHEDULING
4
René van Bevern et al.
created instance
B + a1
B
0
4B + a4
j11
j31
j12
j32
j13
j33
2B
B+A
j11
feasible schedule
3B + a3
2B + a2
j12
j13
j43
3B
j32
j33
j23
M3
j42
j31
j21
M2
j41
j22
j23
2B + A
j22
M1
j21
2B + a1
3B + a2
3B + A
4B
4B +V
j42
j43
j41
3B + a1 + a3
4B + a1 + a3
= 4B + a4
Fig. 3.1 Reduction from B IN PACKING with four items a1 = 1, a2 = a3 = 2, a4 = 3, bin volume V = 3, and m = 3 bins to 3/2-L OOSE I NTERVAL C ONSTRAINED S CHEDULING. That is, Construction 3.2 applies with c = 1, A = 8, and B = 3 · 4 · 8 = 96. The top diagram shows (not to scale) the
jobs created by Construction 3.2. Herein, the processing time of each job is drawn as a rectangle of corresponding length in an interval being the
job’s time window. The bottom diagram shows a feasible schedule for three machines M1 , M2 , and M3 that corresponds to putting items a1 and a3
into the first bin, item a2 into the second bin, and a4 into the third bin.
instance with m machines and mn jobs as follows. First, let
n
A := ∑ ai
and
B := (mn)c · A ≥ 2A.
i=1
If V > A, then I is a yes-instance of B IN PACKING and we
return a trivial yes-instance of I NTERVAL -C ONSTRAINED
S CHEDULING.
Otherwise, we have V ≤ A and construct an instance of
I NTERVAL -C ONSTRAINED S CHEDULING as follows: for
each i ∈ {1, . . . , n}, we introduce a set Ji := { ji1 , . . . , jim } of
jobs. For each job j ∈ Ji , we choose the release time
t j := (i − 1)B,
the processing time
p j :=
(
B + ai
B
if j = ji1 ,
if j 6= ji1 ,
(3.1)
d j :=
if i < n,
iB +V
if i = n.
This concludes the construction.
In the remainder of this section, we show that Construction 3.2 is correct and satisfies all structural properties that
allow us to derive Theorem 3.1.
First, we show that Construction 3.2 indeed creates an I N TERVAL -C ONSTRAINED S CHEDULING instance with small
looseness.
Lemma 3.4 Given a B IN PACKING instance with n ≥ 2 items
and m bins, Construction 3.2 outputs an I NTERVAL -C ON STRAINED S CHEDULING instance with
and the deadline
(
iB + A
Remark 3.3 Construction 3.2 outputs an I NTERVAL -C ON STRAINED S CHEDULING instance with agreeable deadlines,
that is, the deadlines of the jobs have the same relative order
as their release times. Thus, in the offline scenario, all hardness results of Theorem 3.1 will also hold for instances with
agreeable deadlines.
In contrast, agreeable deadlines make the problem significantly easier in the online scenario: Chen et al (2016)
showed an online-algorithm with constant competitive ratio
for I NTERVAL -C ONSTRAINED S CHEDULING with agreeable deadlines, whereas there is a lower bound of n on the
competitive ratio for general instances (Saha 2013).
(i) at most m machines and mn jobs and
(ii) looseness λ (mn).
t
u
Proof It is obvious that the output instance has at most
mn jobs and m machines and, thus, (i) holds.
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
Towards (ii), observe that mn ≥ n ≥ 2, and hence, for
each i ∈ {1, . . . , n} and each job j ∈ Ji , (3.1) yields
|d j − t j | (iB + A) − (i − 1)B B + A
A
≤
=
= 1+
pj
B
B
B
A
= 1+
= 1 + (mn)−c ≤ λ (mn).
(mn)c · A
t
u
We now show that Construction 3.2 runs in polynomial time
and that, if the input B IN PACKING instance has polynomially
bounded integers, then so has the output I NTERVAL -C ON STRAINED S CHEDULING instance.
Lemma 3.5 Let I be a B IN PACKING instance with n ≥
2 items a1 , . . . , an and let amax := max1≤i≤n ai . Construction 3.2 applied to I
(i) runs in time polynomial in |I| and
(ii) outputs an I NTERVAL -C ONSTRAINED S CHEDULING
instance whose release times and deadlines are bounded
by a polynomial in n + amax .
Proof We first show (ii), thereafter we show (i).
(ii) It is sufficient to show that the numbers A and B
in Construction 3.2 are bounded polynomially in n + amax
since all release times and deadlines are computed as sums
and products of three numbers not larger than A, B, or n.
Clearly, A = ∑ni=1 ai ≤ n · max1≤i≤n ai , which is polynomially
bounded in n + amax . Since mn ≤ n2 , also B = (mn)c · A is
polynomially bounded in n + amax .
(i) The sum A = ∑n1=1 ai is clearly computable in time
polynomial in the input length. It follows that also B = (mn)c ·
A is computable in polynomial time.
t
u
5
It is easy to verify that this is indeed a feasible schedule.
(⇐) Assume that I 0 is a yes-instance for I NTERVAL -C ON STRAINED S CHEDULING. Then, there is a feasible schedule
for I 0 . We define a partition S1 ] · · · ] Sm = {1, . . . , n} for I
as follows. For each k ∈ {1, . . . , m}, let
Sk := {i ∈ {1, . . . , n} | ji1 is scheduled on machine k}. (3.2)
Since, for each i ∈ {1, . . . , n}, the job ji1 is scheduled on
exactly one machine, this is indeed a partition. We show
that ∑i∈Sk ai ≤ V for each k ∈ {1, . . . , m}. Assume, towards a
contradiction, that there is a k such that
∑ ai > V.
(3.3)
i∈Sk
By (3.1), for each i ∈ {1, . . . , n}, the jobs in Ji have the same
release time, each has processing time at least B, and the
length of the time window of each job is at most B + A ≤
B + B/2 < 2B. Thus, in any feasible schedule, the execution
times of the m jobs in Ji mutually intersect. Hence, the jobs
in Ji are scheduled on m mutually distinct machines. By the
pigeonhole principle, for each i ∈ {1, . . . , n}, exactly one
job ji∗ ∈ Ji is scheduled on machine k. We finish the proof by
showing that,
∀i ∈ {1, . . . , n}, job ji∗ is not finished before
time iB + ∑ a j .
(3.4)
j∈Sk , j≤i
This claim together with (3.3) then yields that job jn∗ is not
finished before
nB + ∑ a j = nB + ∑ a j > nB +V,
It remains to prove that Construction 3.2 maps yes-instances
j∈Sk , j≤n
j∈Sk
of B IN PACKING to yes-instances of I NTERVAL -C ONSTRAINED
which contradicts the schedule being feasible, since jobs in Jn
S CHEDULING, and no-instances to no-instances.
have deadline nB +V by (3.1). It remains to prove (3.4). We
Lemma 3.6 Given a B IN PACKING instance I with m bins
proceed by induction.
and the items a1 , . . . , an , Construction 3.2 outputs an I N The earliest possible execution time of j1∗ is, by (3.1),
0
TERVAL -C ONSTRAINED S CHEDULING instance I that is a
time 0. The processing time of j1∗ is B if j1∗ 6= j11 , and B + a1
yes-instance if and only if I is.
otherwise. By (3.2), 1 ∈ Sk if and only if j11 is scheduled
on machine k, that is, if and only if j1∗ = j11 . Thus, job j1∗ is
Proof (⇒) Assume that I is a yes-instance for B IN PACKING.
not finished before B + ∑ j∈Sk , j≤i a j and (3.4) holds for i = 1.
Then, there is a partition S1 ] · · · ] Sm = {1, . . . , n} such that
Now, assume that (3.4) holds for i−1. We prove it for i. Since
∑i∈Sk ai ≤ V for each k ∈ {1, . . . , m}. We construct a feasible
∗
ji−1
is not finished before (i − 1)B + ∑ j∈Sk , j≤i−1 a j , this is
schedule for I 0 as follows. For each i ∈ {1, . . . , n} and k such
the
earliest
possible execution time of ji∗ . The processing
that i ∈ Sk , we schedule ji1 on machine k in the interval
∗
time of ji is B if ji∗ 6= ji1 and B + ai otherwise. By (3.2),
"
!
i ∈ Sk if and only if ji∗ = ji1 . Thus, job ji∗ is not finished
(i − 1)B + ∑ a j , iB + ∑ a j + ai
before iB + ∑ j∈Sk , j≤i a j and (3.4) holds.
t
u
j∈Sk , j<i
j∈Sk , j<i
and each of the m − 1 jobs Ji \ { ji1 } on a distinct machine ` ∈
{1, . . . , m} \ {k} in the interval
"
!
(i − 1)B + ∑ a j
j∈S` , j<i
,
iB + ∑ a j .
j∈S` , j<i
We are now ready to finish the proof of Theorem 3.1.
Proof (of Theorem 3.1) By Lemmas 3.4 to 3.6, Construction 3.2 is a polynomial-time many-one reduction from B IN
PACKING with n ≥ 2 items and m bins to λ (mn)-L OOSE I N TERVAL -C ONSTRAINED S CHEDULING , where λ : N → R
6
René van Bevern et al.
such that λ (n) ≥ 1 + n−c for some integer c ≥ 1 and all n ≥ 2.
We now show the points (i) and (ii) of Theorem 3.1.
(i) follows since B IN PACKING is weakly NP-hard for m =
2 (Jansen et al 2013) and since, by Lemma 3.4(i), Construction 3.2 outputs instances of λ (mn)-L OOSE I NTERVAL -C ON STRAINED S CHEDULING with m machines.
(ii) follows since B IN PACKING is W[1]-hard parameterized by m even if the sizes of the n items are bounded by a
polynomial in n (Jansen et al 2013). In this case, Construction 3.2 generates λ (mn)-L OOSE I NTERVAL -C ONSTRAINED
S CHEDULING instances for which all numbers are bounded
polynomially in the number of jobs by Lemma 3.5(ii). Moreover, Construction 3.2 maps the m bins of the B IN PACKING
instance to the m machines of the output I NTERVAL -C ON STRAINED S CHEDULING instance.
t
u
Concluding this section, it is interesting to note that Theorem 3.1 also shows W[1]-hardness of λ -L OOSE I NTERVAL C ONSTRAINED S CHEDULING with respect to the height
parameter considered by Cieliebak et al (2004):
Definition 3.7 (Height) For an I NTERVAL -C ONSTRAINED
S CHEDULING instance and any time t ∈ N, let
St := { j ∈ J | t ∈ [t j , d j )}
denote the set of jobs whose time window contains time t.
The height of an instance is
h := max |St |.
t∈N
Proposition 3.8 Let λ : N → R be such that λ (n) ≥ 1 + n−c
for some integer c ≥ 1 and all n ≥ 2.
Then λ (n)-L OOSE I NTERVAL -C ONSTRAINED S CHED ULING of n jobs on m machines is W[1]-hard parameterized
by the height h.
Proof Proposition 3.8 follows in the same way as Theorem 3.1; one additionally has to prove that Construction 3.2
outputs I NTERVAL -C ONSTRAINED S CHEDULING instances
of height at most 2m. To this end, observe that, by (3.1), for
each i ∈ {1, . . . , n}, there are m jobs released at time (i − 1)B
whose deadline is no later than iB + A < (i + 1)B since
A ≤ B/2. These are all jobs created by Construction 3.2. Thus,
St contains only the m jobs released at time bt/Bc · B and the
m jobs released at time bt/B − 1c · B, which are 2m jobs in
total.
t
u
Remark 3.9 Proposition 3.8 complements findings of Cieliebak et al (2004), who provide a fixed-parameter tractability result for I NTERVAL -C ONSTRAINED S CHEDULING parameterized by h + σ : our result shows that their algorithm
presumably cannot be improved towards a fixed-parameter
tractability result for I NTERVAL -C ONSTRAINED S CHEDUL ING parameterized by h alone.
4 An algorithm for bounded looseness
In the previous section, we have seen that λ -L OOSE I N S CHEDULING for any λ > 1 is
strongly W[1]-hard parameterized by m and weakly NP-hard
for m = 2. We complement this result by the following theorem, which yields a pseudo-polynomial-time algorithm for
each constant m and λ .
TERVAL -C ONSTRAINED
Theorem 4.1 λ -L OOSE I NTERVAL -C ONSTRAINED S CHED ULING is solvable in `O(λ m) · n + O(n log n) time, where ` :=
max j∈J |d j − t j |.
The crucial observation for the proof of Theorem 4.1 is the
following lemma. It gives a logarithmic upper bound on the
height h of yes-instances (as defined in Definition 3.7). To
prove Theorem 4.1, we will thereafter present an algorithm
that has a running time that is single-exponential in h.
Lemma 4.2 Let I be a yes-instance of λ -L OOSE I NTER S CHEDULING with m machines and
` := max j∈J |d j − t j |. Then, I has height at most
log `
2m ·
+1 .
log λ − log(λ − 1)
VAL -C ONSTRAINED
Proof Recall from Definition 3.7 that the height of an I NTER VAL -C ONSTRAINED S CHEDULING instance is maxt∈N |St |.
We will show that, in any feasible schedule for I and at
any time t, there are at most N jobs in St that are active on
the first machine at some time t 0 ≥ t, where
N≤
log `
+ 1.
log λ − log(λ − 1)
(4.1)
By symmetry, there are at most N jobs in St that are active
on the first machine at some time t 0 ≤ t. Since there are
m machines, the total number of jobs in St at any time t, and
therefore the height, is at most 2mN.
It remains to show (4.1). To this end, fix an arbitrary
time t and an arbitrary feasible schedule for I. Then, for
any d ≥ 0, let J(t +d) ⊆ St be the set of jobs that are active on
the first machine at some time t 0 ≥ t but finished by time t + d.
We show by induction on d that
(
0
if d = 0,
|J(t + d)| ≤
(4.2)
log d
− log(1−1/λ
+
1
if
d ≥ 1.
)
If d = 0, then |J(t + 0)| = 0 and (4.2) holds. Now, consider
the case d ≥ 1. If no job in J(t + d) is active at time t + d − 1,
then J(t + d) = J(t + d − 1) and (4.2) holds by the induction
hypothesis. Now, assume that there is a job j ∈ J(t + d) that
is active at time t + d − 1. Then, d j ≥ t + d and, since j ∈ St ,
t j ≤ t. Hence,
pj ≥
|d j − t j | |t + d − t|
d
≥
= .
λ
λ
λ
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
It follows that
|J(t + d)| ≤ 1 + |J(t + d − dd/λ e)|.
(4.3)
Thus, if d − dd/λ e = 0, then |J(t + d)| ≤ 1 + |J(t)| ≤ 1 and
(4.2) holds. If d − dd/λ e > 0, then, by the induction hypothesis, the right-hand side of (4.3) is
log(d − dd/λ e)
+1
log(1 − 1/λ )
log(d(1 − 1/λ ))
+1
≤ 1−
log(1 − 1/λ )
log d + log(1 − 1/λ )
= 1−
+1
log(1 − 1/λ )
log d
=−
+ 1,
log(1 − 1/λ )
≤ 1−
and (4.2) holds. Finally, since ` = max1≤ j≤n |d j − t j |, no job
in St is active at time t + `. Hence, we can now prove (4.1)
using (4.2) by means of
log `
N ≤ |J(t + `)| ≤ −
+1
log(1 − 1/λ )
log `
=−
+1
log λ λ−1
log `
=−
+1
log(λ − 1) − log λ
log `
=
+ 1.
log λ − log(λ − 1)
t
u
Proof It is well-known that (1 − 1/λ )λ < 1/e for any λ ≥ 1.
Hence, λ logb (1 − 1/λ ) = logb (1 − 1/λ )λ < logb 1/e ≤ −1,
that is, −λ logb (1 − 1/λ ) ≥ 1. Thus,
1
≤ λ.
− logb (1 − 1/λ )
(a) If t ≥ 1 and S ⊆ St−1 , then set T [t, S, b] := T [t − 1, S0 , b0 ],
where
(b) Otherwise, set T [t, S, b] := 1 if and only if at least one of
the following two cases applies:
i) there is a machine i ∈ {1, . . . , m} such that bi > −`
and T [t, S, b0 ] = 1, where b0 := (b01 , . . . , b0m ) with
(
bi − 1 if i0 = i,
0
bi0 :=
bi0
if i0 6= i,
or
ii) there is a job j ∈ S and a machine i ∈ {1, . . . , m}
such that bi > 0, t + bi ≤ d j , t + bi − p j ≥ t j , and
T [t, S \ { j}, b0 ] = 1, where b0 := (b01 , . . . , b0m ) with
(
bi − p j if i0 = i,
0
bi0 :=
bi0
if i0 6= i.
Note that, since j ∈ St , one has t j ≥ t − ` by definition of `. Hence, b0i ≥ −` is within the allowed
range {−`, . . . , `}.
Finally,
1
1
=
− logb (1 − 1/λ ) − logb ( λ −1 )
λ
1
=
.
− logb (λ − 1) + logb λ
To compute T , first, set T [0, 0,
/ b] := 1 for every vector b ∈
{−`, . . . , `}m . Now we compute the other entries of T by
increasing t, for each t by increasing b, and for each b by S
with increasing cardinality. Herein, we distinguish two cases.
b0i := min{bi + 1, `} for each i ∈ {1, . . . , m}.
Proposition 4.3 For any λ ≥ 1 and any b ∈ (1, e], it holds
that
1
≤ λ.
logb λ − logb (λ − 1)
and
Algorithm 4.4 We solve I NTERVAL -C ONSTRAINED S CHED ULING using dynamic programming. First, for an I NTERVAL C ONSTRAINED S CHEDULING instance, let ` := max j∈J |d j −
t j |, let St be as defined in Definition 3.7, and let St< ⊆ J be
the set of jobs j with d j ≤ t, that is, that have to be finished
by time t.
We compute a table T that we will show to have the
following semantics. For a time t ∈ N, a subset S ⊆ St of jobs
and a vector b = (b1 , . . . , bm ) ∈ {−`, . . . , `}m ,
1 if all jobs in S ∪ St< can be scheduled so
that machine i is idle from time t + bi
T [t, S, b] =
for each i ∈ {1, . . . , m},
0 otherwise.
S0 := S ∪ (St−1 ∩ St< ) and b0 := (b01 , . . . , b0m ) with
The following proposition gives some intuition on how the
bound behaves for various λ .
1
≤1
−λ logb (1 − 1/λ )
7
Finally, we answer yes if and only if T [tmax , Stmax , 1m · `] = 1,
where tmax := max j∈J t j .
t
u
t
u
Towards our proof of Theorem 4.1, Lemma 4.2 provides a
logarithmic upper bound on the height h of yes-instances of
I NTERVAL -C ONSTRAINED S CHEDULING. Our second step
towards the proof of Theorem 4.1 is the following algorithm,
which runs in time that is single-exponential in h. We first
present the algorithm and, thereafter, prove its correctness
and running time.
Lemma 4.5 Algorithm 4.4 correctly decides I NTERVAL -C ON STRAINED S CHEDULING .
Proof We prove the following two claims: For any time 0 ≤
t ≤ tmax , any set S ⊆ St , and any vector b = (b1 , . . . , bm ) ∈
{−`, . . . , `}m ,
if T [t, S, b] = 1, then all jobs in S ∪St< can be scheduled so that machine i is idle from time t + bi for
each i ∈ {1, . . . , m},
(4.4)
8
René van Bevern et al.
and
if all jobs in S ∪ St< can be scheduled so that
machine i is idle from time t + bi for each i ∈
{1, . . . , m}, then T [t, S, b] = 1.
(4.5)
From (4.4) and (4.5), the correctness of the algorithm easily
follows: observe that, in any feasible schedule, all machines
are idle from time tmax + ` and all jobs J ⊆ Stmax ∪ St<max are
scheduled. Hence, there is a feasible schedule if and only if
T [tmax , Stmax , 1m · `] = 1. It remains to prove (4.4) and (4.5).
First, we prove (4.4) by induction. For T [0, 0,
/ b] = 1,
(4.4) holds since there are no jobs to schedule. We now
prove (4.4) for T [t, S, b] under the assumption that it is true
for all T [t 0 , S0 , b0 ] with t 0 < t or t 0 = t and b0 b.
If T [t, S, b] is set to 1 in Algorithm 4.4(a), then, for S0
and b0 as defined in Algorithm 4.4(a), T [t − 1, S0 , b0 ] = 1.
<
By the induction hypothesis, all jobs in S0 ∪ St−1
can be
scheduled so that machine i is idle from time t − 1 + b0i ≤
<
t +bi . Moreover, S∪St< = S0 ∪St−1
since S0 = S∪(St−1 ∩St< ).
Hence, (4.4) follows.
If T [t, S, b] is set to 1 in Algorithm 4.4(bi), then one has
T [t, S, b0 ] = 1 for b0 as defined in Algorithm 4.4(bi). By the
induction hypothesis, all jobs in S ∪ St< can be scheduled so
that machine i0 is idle from time t + b0i0 ≤ t + bi0 , and (4.4)
follows.
If T [t, S, b] is set to 1 in Algorithm 4.4(bii), then T [t, S \
{ j}, b0 ] = 1 for j and b0 as defined in Algorithm 4.4(bii).
By the induction hypothesis, all jobs in (S \ { j}) ∪ St< can
be scheduled so that machine i0 is idle from time t + b0i0 . It
remains to schedule job j on machine i in the interval [t +
b0i ,t + bi ), which is of length exactly p j by the definition of b0 .
Then, machine i is idle from time t +bi and any machine i0 6= i
is idle from time t + b0i0 = t + bi0 , and (4.4) follows.
It remains to prove (4.5). We use induction. Claim (4.5)
clearly holds for t = 0, S = 0,
/ and any b ∈ {−`, . . . , `}m by
the way Algorithm 4.4 initializes T . We now show (4.5)
provided that it is true for t 0 < t or t 0 = t and b0 b.
<
If S ⊆ St−1 , then S ∪ St< = S0 ∪ St−1
for S0 as defined in
<
Algorithm 4.4(a). Moreover, since no job in S0 ∪ St−1
can be
active from time t − 1 + ` by definition of `, each machine i
is idle from time t − 1 + min{bi + 1, `} = t − 1 + b0i , for b0 =
(b01 , . . . , b0m ) as defined in Algorithm 4.4(a). Hence, T [t −
1, S0 , b0 ] = 1 by the induction hypothesis, Algorithm 4.4(a)
applies, sets T [t, S, b] := T [t − 1, S0 , b0 ] = 1, and (4.5) holds.
If some machine i is idle from time t + bi − 1, then, by
the induction hypothesis, T [t, S, b0 ] = 1 in Algorithm 4.4(bi),
the algorithm sets T [t, S, b] := 1, and (4.5) holds.
In the remaining case, every machine i is busy at time t +
bi − 1 and K := S \ St−1 6= 0.
/ Thus, there is a machine i
executing a job from K. For each job j0 ∈ K, we have t j0 ≥ t.
Since machine i is idle from time t + bi and executes j0 , one
has bi > 0. Let j be the last job scheduled on machine i. Then,
since machine i is busy at time t + bi − 1, we have d j ≥ t +
bi > t and j ∈
/ St< . Hence, j ∈ St . Since machine i is idle from
time t + bi , we also have t + bi − p j ≥ t j . Now, if we remove j
from the schedule, then machine i is idle from time t +bi − p j
and each machine i0 6= i is idle from time t + b0i0 = t + bi0 .
Thus, by the induction hypothesis, T [t, S \ { j}, b0 ] = 1 in
Algorithm 4.4(bii), the algorithm sets T [t, S, b] := 1, and
(4.5) holds.
t
u
Lemma 4.6 Algorithm 4.4 can be implemented to run in
O(2h · (2` + 1)m · (h2 m + hm2 ) · n` + n log n) time, where ` :=
max j∈J |d j − t j | and h is the height of the input instance.
Proof Concerning the running time of Algorithm 4.4, we first
bound tmax . If tmax > n`, then there is a time t ∈ {0, . . . ,tmax }
such that St = 0/ (cf. Definition 3.7). Then, we can split the
instance into one instance with the jobs St< and into one
instance with the jobs J \ St< . We answer “yes” if and only if
both of them are yes-instances. Henceforth, we assume that
tmax ≤ n`.
In a preprocessing step, we compute the sets St and St−1 ∩
<
St , which can be done in O(n log n + hn + tmax ) time by
sorting the input jobs by deadlines and scanning over the
input time windows once: if no time window starts or ends at
time t, then St is simply stored as a pointer to the St 0 for the
last time t 0 where a time window starts or ends.
Now, the table T of Algorithm 4.4 has at most (tmax +
1) · 2h · (2` + 1)m ≤ (n` + 1) · 2h · (2` + 1)m entries. A table
entry T [t, S, b] can be accessed in O(m + h) time using a
carefully initialized trie data structure (van Bevern 2014)
since |S| ≤ h and since b is a vector of length m.
To compute an entry T [t, S, b], we first check, for each
job j ∈ S, whether j ∈ St−1 . If this is the case for each j, then
Algorithm 4.4(a) applies. We can prepare b0 in O(m) time
and S0 in O(h) time using the set St−1 ∩ St< computed in the
preprocessing step. Then, we access the entry T [t − 1, S0 , b0 ]
in O(h + m) time. Hence, (a) takes O(h + m) time.
If Algorithm 4.4(a) does not apply, then we check whether
Algorithm 4.4(bi) applies. To this end, for each i ∈ {1, . . . , m},
we prepare b0 in O(m) time and access T [t, S, b0 ] in O(h + m)
time. Hence, it takes O(m2 + hm) time to check (bi).
To check whether Algorithm 4.4(bii) applies, we try
each j ∈ S and each i ∈ {1, . . . , m} and, for each, prepare b0
in O(m) time and check T [t, S \ { j}, b0 ] in O(h + m) time.
Thus (bii) can be checked in O(h2 m + hm2 ) time.
t
u
With the logarithmic upper bound on the height h of yesinstances of I NTERVAL -C ONSTRAINED S CHEDULING given
by Lemma 4.2 and using Algorithm 4.4, which, by Lemma 4.6,
runs in time that is single-exponential in h for a fixed number m of machines, we can now prove Theorem 4.1.
Proof (of Theorem 4.1) We use the following algorithm. Let
log `
h := 2m ·
+1 .
log λ − log(λ − 1)
A parameterized complexity view on non-preemptively scheduling interval-constrained jobs
If, for any time t ∈ N, we have |St | > h, then we are facing
a no-instance by Lemma 4.2 and immediately answer “no”.
This can be checked in O(n log n) time: one uses the interval
graph coloring problem to check whether we can schedule
the time windows of all jobs (as intervals) onto h machines.
Otherwise, we conclude that our input instance has height
at most h. We now apply Algorithm 4.4, which, by Lemma 4.6,
runs in O(2h · (2` + 1)m · (h2 m + hm2 ) · n` + n log n) time.
Since, by Proposition 4.3, h ∈ O(λ m log `), this running time
is `O(λ m) h · n + O(n log n).
t
u
A natural question is whether Theorem 4.1 can be generalized
to λ = ∞, that is, to I NTERVAL -C ONSTRAINED S CHED ULING without looseness constraint. This question can be
easily answered negatively using a known reduction from
3-PARTITION to I NTERVAL -C ONSTRAINED S CHEDULING
given by Garey and Johnson (1979):
Proposition 4.7 If there is an `O(m) · poly(n)-time algorithm
for I NTERVAL -C ONSTRAINED S CHEDULING, where ` :=
max j∈J |d j − t j |, then P = NP.
Proof Garey and Johnson (1979, Theorem 4.5) showed that
I NTERVAL -C ONSTRAINED S CHEDULING is NP-hard even
on m = 1 machine. In their reduction, ` ∈ poly(n). A supposed `O(m) · poly(n)-time algorithm would solve such instances in polynomial time.
t
u
5 An algorithm for bounded slack
So far, we considered I NTERVAL -C ONSTRAINED S CHED ULING with bounded looseness λ . Cieliebak et al (2004)
additionally considered I NTERVAL -C ONSTRAINED S CHED ULING for any constant slack σ .
Recall that Cieliebak et al (2004) showed that λ -L OOSE
I NTERVAL -C ONSTRAINED S CHEDULING is NP-hard for
any constant λ > 1 and that Theorem 3.1 shows that having a small number m of machines does make the problem
significantly easier.
Similarly, Cieliebak et al (2004) showed that σ -S LACK
I NTERVAL -C ONSTRAINED S CHEDULING is NP-hard already for σ = 2. Now we contrast this result by showing
that σ -S LACK I NTERVAL -C ONSTRAINED S CHEDULING is
fixed-parameter tractable for parameter m + σ . More specifically, we show the following:
Theorem 5.1 σ -S LACK I NTERVAL -C ONSTRAINED S CHED ULING is solvable in time
O (σ + 1)(2σ +1)m · n · σ m · log σ m + n log n .
Similarly as in the proof of Theorem 4.1, we first give an
upper bound on the height of yes-instances of I NTERVAL C ONSTRAINED S CHEDULING as defined in Definition 3.7.
To this end, we first show that each job j ∈ St has to occupy
some of the (bounded) machine resources around time t.
9
Lemma 5.2 At any time t in any feasible schedule for σ S LACK I NTERVAL -C ONSTRAINED S CHEDULING, each job
j ∈ St is active at some time in the interval [t − σ ,t + σ ].
Proof If the time window of j is entirely contained in [t −
σ ,t + σ ], then, obviously, j is active at some time during the
interval [t − σ ,t + σ ].
Now, assume that the time window of j is not contained
in [t − σ ,t + σ ]. Then, since j ∈ St , its time window contains t by Definition 3.7 and, therefore, one of t − σ or t + σ .
Assume, for the sake of contradiction, that there is a schedule such that j is not active during [t − σ ,t + σ ]. Then j is
inactive for at least σ + 1 time units in its time window—
a contradiction.
t
u
Now that we know that each job in St has to occupy machine
resources around time t, we can bound the size of St in the
amount of resources available around that time.
Lemma 5.3 Any yes-instance of σ -S LACK I NTERVAL -C ON STRAINED S CHEDULING has height at most (2σ + 1)m.
Proof Fix any feasible schedule for an arbitrary yes-instance
of σ -S LACK I NTERVAL -C ONSTRAINED S CHEDULING and
any time t. By Lemma 5.2, each job in St is active at some
time in the interval [t −σ ,t +σ ]. This interval has length 2σ +
1. Thus, on m machines, there is at most (2σ + 1)m available
processing time in this time interval. Consequently, there can
be at most (2σ + 1)m jobs with time intervals in St .
t
u
We finally arrive at the algorithm to prove Theorem 5.1.
Proof (of Theorem 5.1) Let h := (2σ + 1)m. In the same
way as for Theorem 4.1, in O(n log n) time we discover that
we face a no-instance due to Lemma 5.3 or, otherwise, that
our input instance has height at most h. In the latter case,
we apply the O(n · (σ + 1)h · h log h)-time algorithm due to
Cieliebak et al (2004).
t
u
6 Conclusion
Despite the fact that there are comparatively few studies on
the parameterized complexity of scheduling problems, the
field of scheduling indeed offers many natural parameterizations and fruitful challenges for future research. Notably,
Marx (2011) saw one reason for the lack of results on “parameterized scheduling” in the fact that most scheduling
problems remain NP-hard even for a constant number of machines (a very obvious and natural parameter indeed), hence
destroying hope for fixed-parameter tractability results with
respect to this parameter. In scheduling interval-constrained
jobs with small looseness and small slack, we also have been
confronted with this fact, facing (weak) NP-hardness even
for two machines.
10
The natural way out of this misery, however, is to consider
parameter combinations, for instance combining the parameter number of machines with a second one. In our study, these
were combinations with looseness and with slack (see also
Table 1.1). In a more general perspective, this consideration
makes scheduling problems a prime candidate for offering
a rich set of research challenges in terms of a multivariate
complexity analysis (Fellows et al 2013; Niedermeier 2010).
Herein, for obtaining positive algorithmic results, research
has to go beyond canonical problem parameters, since basic
scheduling problems remain NP-hard even if canonical parameters are simultaneously bounded by small constants, as
demonstrated by Kononov et al (2012).1
Natural parameters to be studied in future research on I N TERVAL -C ONSTRAINED S CHEDULING are the combination
of slack and looseness—the open field in our Table 1.1—and
the maximum and minimum processing times, which were
found to play an important role in the online version of the
problem (Saha 2013).
Finally, we point out that our fixed-parameter algorithms
for I NTERVAL -C ONSTRAINED S CHEDULING are easy to
implement and may be practically applicable if the looseness,
slack, and number of machines is small (about three or four
each). Moreover, our algorithms are based on upper bounds
on the height of an instance in terms of its number of machines, its looseness, and slack. Obviously, this can also be
exploited to give lower bounds on the number of required machines based on the structure of the input instance, namely, on
its height, looseness, and slack. These lower bounds may be
of independent interest in exact branch and bound or approximation algorithms for the machine minimization problem.
References
van Bevern R (2014) Towards optimal and expressive kernelization for
d-Hitting Set. Algorithmica 70(1):129–147
van Bevern R, Chen J, Hüffner F, Kratsch S, Talmon N, Woeginger GJ
(2015a) Approximability and parameterized complexity of multicover by c-intervals. Information Processing Letters 115(10):744–
749
van Bevern R, Mnich M, Niedermeier R, Weller M (2015b) Interval
scheduling and colorful independent sets. Journal of Scheduling
18(5):449–469
Bodlaender HL, Fellows MR (1995) W[2]-hardness of precedence
constrained k-processor scheduling. Operations Research Letters
18(2):93–97
Chen L, Megow N, Schewior K (2016) An O(log m)-competitive algorithm for online machine minimization. In: Proceedings of the
27th Annual ACM-SIAM Symposium on Discrete Algorithms
(SODA’16), SIAM, pp 155–163
1 The results of Kononov et al (2012) were obtained in context of a
multivariate complexity analysis framework described by Sevastianov
(2005), which is independent of the framework of parameterized complexity theory considered in our work: it allows for systematic classification of problems as polynomial-time solvable or NP-hard given concrete
constraints on a set of instance parameters. It is plausible that this framework is applicable to classify problems as FPT or W[1]-hard as well.
René van Bevern et al.
Chuzhoy J, Guha S, Khanna S, Naor J (2004) Machine minimization
for scheduling jobs with interval constraints. In: Proceedings of
the 45th Annual Symposium on Foundations of Computer Science
(FOCS’04), pp 81–90
Cieliebak M, Erlebach T, Hennecke F, Weber B, Widmayer P (2004)
Scheduling with release times and deadlines on a minimum number
of machines. In: Exploring New Frontiers of Theoretical Informatics, IFIP International Federation for Information Processing, vol
155, Springer, pp 209–222
Cygan M, Fomin FV, Kowalik L, Lokshtanov D, Marx D, Pilipczuk M,
Pilipczuk M, Saurabh S (2015) Parameterized Algorithms. Springer
Downey RG, Fellows MR (2013) Fundamentals of Parameterized Complexity. Springer
Fellows MR, McCartin C (2003) On the parametric complexity of
schedules to minimize tardy tasks. Theoretical Computer Science
298(2):317–324
Fellows MR, Jansen BMP, Rosamond FA (2013) Towards fully multivariate algorithmics: Parameter ecology and the deconstruction
of computational complexity. European Journal of Combinatorics
34(3):541–566
Flum J, Grohe M (2006) Parameterized Complexity Theory. Springer
Garey MR, Johnson DS (1979) Computers and Intractability: A Guide
to the Theory of NP-Completeness. Freeman
Halldórsson MM, Karlsson RK (2006) Strip graphs: Recognition and
scheduling. In: Proceedings of the 32nd International Workshop on
Graph-Theoretic Concepts in Computer Science (WG’06), Springer,
LNCS, vol 4271, pp 137–146
Hermelin D, Kubitza JM, Shabtay D, Talmon N, Woeginger G (2015)
Scheduling two competing agents when one agent has significantly
fewer jobs. In: Proceedings of the 10th International Symposium
on Parameterized and Exact Computation (IPEC’15), Leibniz International Proceedings in Informatics (LIPIcs), vol 43, Schloss
Dagstuhl–Leibniz-Zentrum für Informatik, pp 55–65
Jansen K, Kratsch S, Marx D, Schlotter I (2013) Bin packing with fixed
number of bins revisited. Journal of Computer and System Sciences
79(1):39–49
Kolen AWJ, Lenstra JK, Papadimitriou CH, Spieksma FCR (2007)
Interval scheduling: A survey. Naval Research Logistics 54(5):530–
543
Kononov A, Sevastyanov S, Sviridenko M (2012) A complete 4parametric complexity classification of short shop scheduling problems. Journal of Scheduling 15(4):427–446
Malucelli F, Nicoloso S (2007) Shiftable intervals. Annals of Operations
Research 150(1):137–157
Marx D (2011) Fixed-parameter tractable scheduling problems. In:
Packing and Scheduling Algorithms for Information and Communication Services (Dagstuhl Seminar 11091)
Mnich M, Wiese A (2015) Scheduling and fixed-parameter tractability.
Mathematical Programming 154(1-2):533–562
Niedermeier R (2006) Invitation to Fixed-Parameter Algorithms. Oxford
University Press
Niedermeier R (2010) Reflections on multivariate algorithmics and
problem parameterization. In: Proceedings of the 27th International Symposium on Theoretical Aspects of Computer Science
(STACS’10), Schloss Dagstuhl–Leibniz-Zentrum für Informatik,
Leibniz International Proceedings in Informatics (LIPIcs), vol 5,
pp 17–32
Saha B (2013) Renting a cloud. In: Annual Conference on Foundations of Software Technology and Theoretical Computer Science
(FSTTCS) 2013, Schloss Dagstuhl–Leibniz-Zentrum für Informatik, Leibniz International Proceedings in Informatics (LIPIcs),
vol 24, pp 437–448
Sevastianov SV (2005) An introduction to multi-parameter complexity
analysis of discrete problems. European Journal of Operational
Research 165(2):387–397
| 8cs.DS
|
arXiv:1710.02330v1 [math.GR] 6 Oct 2017
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR
PRODUCTS
VALERIY G. BARDAKOV, ANDREI V. LAVRENOV, AND MIKHAIL V. NESHCHADIM
Abstract. In this paper we give an example of a linear group such that its tensor
square is not linear. Also, we formulate some sufficient conditions for the linearity
of non-abelian tensor products G ⊗ H and tensor squares G ⊗ G. Using these
results we prove that tensor squares of some groups with one relation and some
knot groups are linear. We prove that the Peiffer square of finitely generated linear
groups is linear. At the end we construct faithful linear representations for the
non-abelian tensor square of free group and free nilpotent group.
1. Introduction
R. Brown and J.-L. Loday [8, 9] introduced the non-abelian tensor product G ⊗ H
for a pair of groups G and H following works of Miller [5], and Lue [7]. They showed
that the third homotopy group of the suspension of an Eilenberg-MacLane space
K(G, 1) satisfies
π3 SK(G, 1) ∼
= J2 (G),
where J2 (G) is the kernel of the derived map κ : G ⊗ G −→ G′ , g ⊗ h 7→ [g, h] =
g−1 h−1 gh. Hence there exists the short exact sequence
0 −→ π3 SK(G, 1) −→ G ⊗ G −→ G′ −→ 1.
Also, the non-abelian tensor product is used to describe the third relative homotopy
group of a triad as a non-abelian tensor product of the second homotopy groups of
appropriate subspaces. More specifically, let a CW -complex X be the union X =
A∪ B of two path-connected CW -subspaces A and B whose intersection C = A∩ B is
path-connected. If the canonical homomorphisms π1 (C) −→ π1 (A), π1 (C) −→ π1 (B)
are surjective, then, according to [8],
π3 (X, A, B) ∼
= π2 (A, C) ⊗ π2 (B, C),
where the groups π2 (A, C) and π2 (B, C) act on one another via π1 (C).
The investigation of the non-abelian tensor product from a group theoretical point
of view started with a paper by Brown, Johnson, and Robertson [15]. They compute
the non-abelian tensor square of all non-abelian groups of order up to 30 using Tietze
transformations.
One of the topics of research on the non-abelian tensor products has been to determine which group properties are preserved by non-abelian tensor products. By using
Date: October 9, 2017.
2010 Mathematics Subject Classification. Primary 20E25; Secondary 20G20, 20E05.
Key words and phrases. non-abelian tensor product; linear group; faithful linear representation.
1
2
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
homological arguments, Ellis [10] showed that if G and H are finite groups, then
G ⊗ H is also finite. Visscher [12] proved that if G, H are solvable (nilpotent), then
G ⊗ H is solvable (nilpotent) and gives a bound on the nilpotency class of G ⊗ H. In
[17] was proved that the tensor product of groups of nilpotency class at most n is a
group of nilpotency class at most n, thereby improving the bound given by Visscher.
For other results in this direction see the survey of Nakaoka [13].
In this paper we study the linearity problem for non-abelian tensor products. Let
n be a positive integer and let P be a field. A group G is said to be linear of degree
n over P if it is isomorphic with a subgroup of GLn (P ), the group of all n × n nonsingular matrices over P or, equivalently, if it is isomorphic with a group of invertible
linear transformations of a vector space of dimension n over P (see [19]). We study
the following
Question 1. Let G and H be linear groups. Are the groups G ⊗ H, G ⊗ G linear?
We show that in general case the answer is negative. More accurately, we prove
that the tensor square SLn (Q) ⊗ SLn (Q) of special linear group SLn (Q) over the
field of rational numbers is not linear for n ≥ 3. On the other side we formulate some
sufficient conditions under which the groups G ⊗ H, G ⊗ G are linear. Using these
conditions, we prove that the non-abelian tensor squares of some groups with one
defining relation and groups of fibered knots are linear. If G is a finitely generated free
group or finitely generated free nilpotent group, then we construct concrete faithful
linear representations for G ⊗ G.
With the non-abelian tensor square G ⊗ G connect other group constructions:
exterior tensor square G ∧ G and Peiffert square G ⊲⊳ G. We prove that if G is finitely
generated, then G ⊲⊳ G is linear.
We note that the following problems are still open.
Problem 1. 1) Let G be a finitely generated linear group. Is the group G ⊗ G linear?
2) Let G be a linear group. Is the group G ⊲⊳ G linear?
Acknowledgement. The authors thank V. Thomas, S. O. Ivanov, V. Petrushenko,
A. Yu. Ol’shanskii, V. A. Roman’kov, M. Singh for useful discussions. They also acknowledge support from the Russian Science Foundation (project No. 16-41-02006).
2. Preliminaries
In this paper we shall use the following notations. For elements x, y in a group
G, the conjugation of x by y is xy = y −1 xy; and the commutator of x and y is
[x, y] = x−1 xy = x−1 y −1 xy. We write G′ for the derived subgroup of G, Gab for the
abelianized group G/G′ .
Recall the definition of the non-abelian tensor product G ⊗ H of groups G and H
(see [8, 9]). This tensor product is defined for any pair of groups G and H where
each one acts on the other (on the right)
G × H −→ G, (g, h) 7→ g h ; H × G −→ H, (h, g) 7→ hg
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR PRODUCTS
3
and on itself by conjugation, in such a way that for all g, g1 ∈ G and h, h1 ∈ H,
h g1
−1 g h1
h1
g1−1
(hg1 )
g
and h(g ) = hh1
.
g
=
In this situation we say that G and H act compatibly on each other. The non-abelian
tensor product G ⊗ H is the group generated by all symbols g ⊗ h, g ∈ G, h ∈ H,
subject to the relations
gg1 ⊗ h = (gg1 ⊗ hg1 )(g1 ⊗ h) and g ⊗ hh1 = (g ⊗ h1 )(gh1 ⊗ hh1 )
for all g, g1 ∈ G, h, h1 ∈ H.
In particular, as the conjugation action of a group G on itself is compatible, then
the tensor square G⊗G of a group G may always be defined. Also, the tensor product
G ⊗ H is defined if G and H are two normal subgroups of some group M and actions
are conjugations in M .
Recall the main diagram for the non-abelian tensor square (see [8, 9]). Let G be
a group. One of the main tools for studying of the non-abelian tensor square G ⊗ G
is the following diagram:
0
↓
ψ
H3 (G) → Γ(Gab ) →
q
q
J2 (G)
↓
0
↓
→ H2 (G) → 0
↓
ψ
H3 (G) → Γ(Gab ) → G ⊗ G → G ∧ G → 1
↓
↓
G′
=
G′
↓
↓
1
1
with exact rows and columns. Here
1) H2 (G), H3 (G) are the second and the third homology groups of G with the
coefficients in the trivial ZG-module Z. The second homology group H2 (G) for the
group G = F/R, where F is a free group, can be find by the Hopf formular:
H2 (G) ∼
= (F ′ ∩ R)/[F, R].
2) G ∧ G is the exterior product of G onto itself. For the group G = F/R it can
be presented in the form (see [2])
G∧G∼
= F ′ /[F, R].
In particular, if G is a free group, then G ∧ G ∼
= G′ .
3) J2 (G) = π3 SK(G, 1) is the kernel of the derived map κ : G ⊗ G −→ G′ , which
on the generators of G ⊗ G is defined by the rule:
g1 ⊗ g2 7→ [g1 , g2 ].
The group J2 (G) lies in the center Z(G ⊗ G) and its elements are invariant under the
action of G onto G ⊗ G, which is defined by the formula
(g1 ⊗ g2 )g = g1g ⊗ g2g .
4
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
In particular, if g2 = g1 , then
(g1 ⊗ g1 )g = g1 ⊗ g1
for any g, g1 ∈ G.
4) Γ(Gab ) is Whitehead’s quadratic functor. The group Γ(Gab ) is generated by
elements γ(gG′ ) and ψ is defined by the formula
γ(gG′ ) 7−→ g ⊗ g.
The image ψΓ(Gab ) is not equal in the general case to the group J2 (G) since
J2 (G)/ψΓ(Gab ) ∼
= H2 (G);
For the functor Γ : A 7−→ Γ(A), where A is an abelian group is known that
a) Γ(A× B) ∼
= Γ(A)× Γ(B)× (A⊗Z B), where A⊗Z B is the abelian tensor product
of abelian groups;
Zn n ≡ 1 (mod 2),
∼
b) Γ(Zn ) =
Z2n n ≡ 0 (mod 2).
c) Γ(Z) ∼
= Z.
n(n+1)
In particular, Γ(Zn ) ∼
=Z 2 .
3. Linearity problem
In this section we will use result of Malcev [19] (see also [20, Chapter 2]) on the
linearity of abelian group. To formulate it recall come definitions. If G is any group
τ (G) is the subgroup of G generated by all the periodic normal subgroups of G; that
is τ (G) is the maximum periodic normal subgroup of G. G has finite rank at most
n if every finite subset of G is contained in an n-generator subgroup of G. If G is
abelian and periodic then G has finite rank at most n if and only if for each prime p
the Sylow p-subgroup of G is a direct product of at most n cyclic and Prüfer p-groups
(a Prüfer p-group is a Cp∞ -group). If π is any set of primes and G is a group with a
unique maximal π-subgroup we denote this maximal π-subgroup by Gπ .
Malcev proved:
i) An abelian group A has a faithful representation of degree n ≥ 1 over some field
of characteristic zero if and only if τ (G) has rank at most n.
ii) An abelian group A has a faithful representation of degree n ≥ 1 over some
field of characteristic p > 0 if and only if τ (G)p′ has finite rank r and τ (A)p has finite
exponent pe satisfying
pe−1 + max{1, r} < n + 1.
We are ready to prove the following
Proposition 3.1. There is a linear group G such that G ⊗ G = G ∧ G is not linear.
Proof. For a perfect group G = G′ it follows from the main diagram (see section 2)
that G ⊗ G = G ∧ G and the sequence
0 −→ H2 (G, Z) −→ G ⊗ G −→ G −→ 0
is exact.
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR PRODUCTS
5
For n ≥ 3 the group SLn (Q) is perfect and its second homology group coincides
with K2 -group of a field Q,
H2 (SLn (Q), Z) = K2 (Q),
see [6, Corollary 11.2].
Q
Next, K2 (Q) = {±1}×
(Z/p)× by [6, Theorem 11.6], so that K2 (Q) contains an
p prime
.
p6 .. 2
abelian 2-group of infinite rank and unbounded exponent. Using Maltsev’s criterion
we conclude that such a group can not be linear. Therefore the group
SLn (Q) ⊗ SLn (Q) = SLn (Q) ∧ SLn (Q)
is not linear as well.
To study the linearity problem for the non-abelian tensor product we can use a
presentation of a tensor product as a central extension (see, for example, [17].) The
derivative subgroup of G by H is called the following subgroup
DH (G) = [G, H] = hg −1 g h | g ∈ G, h ∈ Hi.
The map κ : G ⊗ H −→ DH (G) defined by κ(g ⊗ h) = g −1 gh is a homomorphism, its
kernel is the central subgroup of G ⊗ H and G acts on G ⊗ H by the rule (g ⊗ h)x =
gx ⊗ hx , x ∈ G. There exists the short exact sequence
1 −→ A −→ G ⊗ H −→ DH (G) −→ 1.
In this case A can be viewed as Z[DH (G)]-module via conjugation in G ⊗ H, i. e.
under the action induced by setting
a · g = x−1 ax, a ∈ A, x ∈ G ⊗ H, κ(x) = g.
We can formulate some sufficient conditions for the linearity of G ⊗ H. Well
known that tensor product G ⊗ H with trivial actions isomorphic to the abelian
tensor product Gab ⊗Z H ab . Hence, in this case the question on the linearity of G ⊗ H
is equivalent to the question on the linearity of abelian tensor product and the answer
follows from the Malcev theorem.
Further we will assume that the action of G on H or the action of H on G is
non-trivial. We have the following short exact sequence
0 −→ A −→ G ⊗ H −→ DH (G) −→ 1.
(1)
Note that A is the kernel of the natural map G ⊗ H −→ DH (G), g ⊗ h −→ g−1 gh ,
g ∈ G, h ∈ H, and is a central subgroup of G ⊗ H.
Proposition 3.2. Let the following conditions hold
1) A, DH (G) are linear groups;
2) H 2 (DH (G), A) = 0, in particular, this condition holds if A is divisible or DH (G)
is a free group.
Then G ⊗ H = A × DH (G) is the direct product and is a linear group.
6
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
Proof. It is well known that if H 2 (DH (G), A) = 0, then the sequence
0 −→ A −→ G ⊗ H −→ DH (G) −→ 1,
splits. In particular, this condition holds if A is divisible or DH (G) is a free group.
Since A is a central subgroup, then G ⊗ H ∼
= A × DH (G) and is a linear group as
a direct product of linear groups.
The main problem in the using of this theorem is the description of the central
subgroup A. For the tensor square we can use another approach.
Let us formulate some sufficient conditions under which G ⊗ G is a direct product
of the commutator subgroup G′ and the Whitehead group Γ(Gab ).
Theorem 3.3. Let H2 (G) = H3 (G) = H2 (G′ ) = 0 and one of the following conditions
hold
1) H 2 (G′ , Γ(Gab )) = 0;
2) Γ(Gab ) is a divisible group;
3) G′ /G′′ does not have torsion.
Then
G⊗G∼
= Γ(Gab ) × G′ .
If moreover G is finitely generated and G′ is linear, then G ⊗ G is linear.
Proof. Since, H2 (G) = H3 (G) = 0, then the main diagram has the form
0
↓
ψ
0 → Γ(Gab ) →
q
J2 (G)
↓
0
↓
→
0
↓
→ 0
ψ
0 → Γ(Gab ) → G ⊗ G → G ∧ G → 1
↓
↓
G′
=
G′
↓
↓
1
1
From this diagram J2 (G) = Γ(Gab ) and G ∧ G = G′ . Hence we have the short
exact sequence
0 −→ Γ(Gab ) −→ G ⊗ G −→ G′ −→ 1.
If H 2 (G′ , Γ(Gab )) = 0, then this sequence is splittable:
G ⊗ G = Γ(Gab ) ⋋ G′ .
As we know if Γ(Gab ) is divisible or G′ is free, then H 2 (G′ , Γ(Gab )) = 0. Let us
show that if G′ /G′′ does not have torsion, then H 2 (G′ , Γ(Gab )) = 0. Indeed, by the
universal coefficient theorem there is the following short exact sequence
0 −→ ExtZ (H1 (G′ ), Γ(Gab )) −→ H 2 (G′ , Γ(Gab )) −→ HomZ (H2 (G′ ), Γ(Gab )) −→ 0.
Since H2 (G′ ) = 0 we have the short exact sequence
0 −→ ExtZ (H1 (G′ ), Γ(Gab )) −→ H 2 (G′ , Γ(Gab )) −→ 0.
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR PRODUCTS
7
Hence H 2 (G′ , Γ(Gab )) = 0 if and only if ExtZ (H1 (G′ ), Γ(Gab )) = 0. It is known that
if H1 (G′ ) is free abelian, then ExtZ (H1 (G′ ), Γ(Gab )) = 0.
Since Γ(Gab ) is a central subgroup, this product is the direct product:
G ⊗ G = Γ(Gab ) × G′ .
If G is finitely generated, then Gab is finitely generated abelian group and Γ(Gab )
is also finitely generated abelian group. Then G ⊗ G is linear as direct product of
two linear groups.
As consequence we get the following result
Corollary 3.4. [2] Let Fn be a free group of rank n. Then
Fn ⊗ Fn ∼
= Zn(n+1)/2 × (Fn )′ .
4. Groups with one defining relation and knot groups
Let G be a group with one defining relation:
G = hX kr = 1i,
where r 6∈ F ′ , F = h X i. Then Hk (G) = 0, k ≥ 2 (see [3, p. 49]). Hence, there exist
the following short exact sequence:
0 −→ Γ(Gab ) −→ G ⊗ G −→ G′ −→ 1.
If Gab does not have torsion, then Gab is a free abelian group and Γ(Gab ) is a free
abelian group. Then, if H 2 (G′ ) = 0, then H 2 (G′ , Γ(Gab )) = 0, what follows from the
decomposition
H k (S, A ⊕ B) = H k (S, A) ⊕ H k (S, B)
for every group S and every S-modules A and B.
From Theorem 3.3 follows
Proposition 4.1. Let G be a group with one defining relation:
G = hX kr = 1i,
where r 6∈ F ′ , F = h X i. If one from the following conditions hold:
1) Gab does not have torsion and H 2 (G′ ) = 0;
2) G′ /G′′ does not have torsion.
Then G ⊗ G = Γ(Gab ) × G′ . If moreover G is finitely generated and G is linear,
then G ⊗ G is linear.
It is well known that if K is a tame knot in 3-sphere S3 and GK = π1 (S3 \ K) its
group, then Hn (GK ) = 0 for n > 1 (see, for example [18, p. 5]). Recall that a knot
K is called fibered if there is a 1-parameter family Ft of Seifert surfaces for K, where
the parameter t runs through the points of the unit circle S 1 , such that if s is not
equal to t then the intersection of Fs and Ft is exactly K. The commutator subgroup
G′K for the fibered knot K is a free of finite rank [14] and GK is linear [1].
Proposition 4.2. Let K be a tame fibered knot in 3-sphere S3 , then GK ⊗ GK =
G′K × Z and has faithful linear representation into GL2 (Z[t, t−1 ]).
8
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
ab
Proof. It is well known that Gab
K = Z and then Γ(GK ) = Z. From Theorem 3.3
′
follows that GK ⊗ GK = Z × GK .
To construct a linear representation, use the fact that G′K is a free group of finite
rank and by Sanov’s theorem [11, Chapter 5] it has a faithful linear representation
into SL2 (Z) ≤ GL2 (Z[t, t−1 ]). Define a linear representation of Gab
K = Z = hγi into
−1
GL2 (Z[t, t ]) by the rule
t 0
γ 7−→
.
0 t
Since the image of γ is a scalar matrix, i.e. lies in the center of GL2 (Z[t, t−1 ]), we
constructed a faithful linear representation of GK ⊗ GK .
Example 4.3. 1) The braid group B3 on 3 strings has presentation
B3 = hσ1 , σ2 || σ1 σ2 σ1 = σ2 σ1 σ2 i.
and is the group of trefoil knot. The commutator subgroup B3′ is a free group of rank
2. Hence the tensor square B3 ⊗ B3 = Z × F2 has a faithful linear representation into
GL2 (Z[t, t−1 ]).
2) It is known that the group of the figure eight knot has a presentation
G = hx, y || yx−1 yxy −1 = x−1 yxy −1 xi
and is a fibered knot. Hence the tensor square G ⊗ G = Z × G′ has a faithful linear
representation into GL2 (Z[t, t−1 ]).
In the first example we shown that B3 ⊗B3 = Z×F2 . On the other side B3 contains
the pure braid group P3 , which is normal in B3 , has index 6 and is the direct product
of the center, which is isomorphic to Z, and a free group of rank 2. Hence, B3 ⊗ B3
is isomorphic to P3 and we proved
Proposition 4.4. There is a non-trivial non-abelian group G such that the tensor
square G ⊗ G is isomorphic to a proper subgroup of G.
Question 2. 1) Is it true that Bn ⊗ Bn , n > 3, is linear?
2) Is it true that for arbitrary tame knot K the group G(K) ⊗ G(K) is linear?
5. On the linearity of the Peiffer product
Recall the definition of the Peiffer product. Given G and H acting compatibly on
each other, in [21] it was defined their Peiffer product G ⊲⊳ H as the quotient of the
free product G ∗ H by the normal closure K of all elements of the form
h−1 g −1 hgh or g −1 h−1 ghg
where g ∈ G and h ∈ H. Whitehead [21] posed question on the asphericity of
subcomplexes of aspherical 2-complexes and reformulated it as part of the wider
problem of finding conditions under which the groups G and H are embedded in
G ⊲⊳ H.
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR PRODUCTS
9
In [4] it was proved that if ϕ : G ∗ H → G ⊲⊳ H, then modulo K = Ker(ϕ),
hg ≡ ghg , so that every element of G ⊲⊳ H can be written as ϕ(g)ϕ(h) for suitable g,
h. Denote ϕ(g)ϕ(h) as hg, hi. The relations
−1
hg, hihg1 , h1 i = hgg1 , hg1 h1 i = hgg1h , hh1 i
are defining relations for G ⊲⊳ H on the generators hg, hi and so G ⊲⊳ H is a homomorphic image of both the semidirect products G ⋉ H and G ⋊ H. The group G ⊲⊳ H
is obtained from G ⋉ H (or from G ⋊ H) by imposing the relations
(g−1 gh , 1) = (1, h−g h).
If G and H act on one another trivially, then G ⊲⊳ H is just the direct product
G × H and K = GH, where GH is the Cartesian subgroup of G ∗ H (the kernel
of the canonical homomorphism G ∗ H −→ G × H.
From [4, Proposition 2.1] follows
G ⊲⊳ G ∼
= Gab × Gab .
Using this isomorphism one can proves
Proposition 5.1. Let G be a linear group and G is finitely generated or G = G′ ,
then G ⊲⊳ G is linear.
From this proposition follows that if G = SLn (Q), n ≥ 2, then G ⊲⊳ G is linear.
On the other side, we know that SLn (Q) ⊗ SLn (Q) and SLn (Q) ∧ SLn (Q) are not
linear for n ≥ 3.
But we don’t know is this proposition true for arbitrary linear group. In particular,
we do not know an answer on the following
Question 3. Let G be a linear group. Is it true that Gab is also linear?
It is evident that the following short exact sequence holds
1 −→ 1 × G′ −→ Gab × G −→ Gab × Gab −→ 1.
Since Gab × G ∼
= G ⊲⊳ G we can add in the main diagram new terms.
Proposition 5.2. The following diagram holds
0
↓
ψ
H3 (G) → Γ(Gab ) →
q
q
ψ
H3 (G) → Γ(Gab ) →
J2 (G)
↓
0
↓
→
H2 (G)
↓
→
1
G⊗G
→
G∧G
→
1
↓
↓
G ⊲⊳ G
=
G ⊲⊳ G
↓
↓
ab
ab
ab
G ×G
= G × Gab
= H1 (G) × H1 (G)
↓
↓
1
1
10
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
6. Faithful linear representations
In the paper [2] it was proved:
1) If Fn is the free group of rank n, then
Fn ⊗ Fn ∼
= Zn(n+1)/2 × (Fn )′ .
2) If Nn,c = Fn /γc Fn is the free nilpotent group of rank n > 1 and class c ≥ 1,
then
Nn,c ⊗ Nn,c ∼
= Zn(n+1)/2 × (Nn,c+1 )′ .
Proposition 6.1. Let G be a free countable group. Then the exterior square G ∧ G
has a faithful representation into SL2 (Z) and the tensor square G ⊗ G has a faithful
representation into GL2 (C).
Proof. As was proven in [9], for the free group G there are isomorphisms
G∧G∼
= G′ , G ⊗ G ∼
= Γ(Gab ) × G′ .
Since G is free, its commutator subgroup G′ is free. Hence, by the Sanov result [11,
Chapter 5] there is a faithful representation of G′ into SL2 (Z) and the first part of
the proposition holds.
Further, Γ(Gab ) is a free abelian group. Let ak , k ∈ I be its free generators. Take
transcendental elements tk , k ∈ I in the field C, which are algebraically independent
other Q. Then the matrix group
tk 0
T =
kk∈I
0 tk
is isomorphic to the group Γ(Gab ). If ϕ : G′ −→ GL2 (Z) is an embedding, then
ϕG′ , T ∼
= ϕG′ × T.
Hence, the group G ⊗ G has a faithful representation in the matrix group over the
ring Z[t±1
k , k ∈ I].
If G = F∞ is countably generated and has a faithful representation into SL2 (Z).
ab ) is linear we use the following property
To prove that Γ(F∞
ab
Γ(F∞
) = Γ(lim Fnab ) = lim(ΓFnab ).
For the finitely generated free groups from this theorem follows
±1
Corollary 6.2. The tensor square Fn ⊗Fn has a faithful representation into GL2 (Z[t±1
1 , . . . , tm ]),
n(n+1)
where m = 2 .
For the free nilpotent groups we can prove
Proposition 6.3. There is a faithful representation
Nn,c ⊗ Nn,c −→ Tc+2 (C)
into the group of triangular matrices Tc+2 (C).
LINEARITY PROBLEM FOR NON-ABELIAN TENSOR PRODUCTS
11
Proof. We noted that
Nn,c ⊗ Nn,c ∼
= Zn(n+1)/2 × (Nn,c+1 )′ .
Hence, we have to define faithful linear representations for Zn(n+1)/2 = ha1 , a2 , . . . , am i,
m = n(n + 1)/2 and for (Nn,c+1 )′ , where Nn,c+1 = hx1 , x2 , . . . , xn i. Let
τ1 , τ2 , . . . , τm , tij , i = 1, 2, . . . , n, j = 1, 2, . . . , c + 1,
be the complex numbers, which are algebraically independent over Q. Define the
following maps
ak 7→ τk E ∈ Tc+2 (C), k = 1, 2, . . . , m,
which is defined a faithful representation of
1 ti1 0
0 ... 0
0
1
t
0 ... 0
i2
.
.
.
.
.
.
.
.
.
.
.. ... ...
xi 7→
0
0
0
0 ... 1
0
0
0
0 ... 0
Zn(n+1)/2 into Tc+2 (C), and
0
0
...
∈ U Tc+2 (C) i = 1, 2, . . . , n.
tic
1
As Romanovskii proved [16] the map, defined on xi is a faithful representation of
Nn,c+1 into U Tc+2 (C). Hence we have a faithful representation of Zn(n+1)/2 × Nn,c+1
into Tc+2 (C). Since Zn(n+1)/2 × (Nn,c+1 )′ is a subgroup of Zn(n+1)/2 × Nn,c+1 , we have
the need representation.
References
[1] M. Aschenbrenner, S. Friedl and H. Wilton, 3-Manifold Groups, EMS Series of Lectures in
Mathematics, vol. 20, European Mathematical Society (EMS), Zurich, 2015.
[2] R. D. Blyth, F. Fumagalli, M. Morigi, Some structural results of the non-abelian tensor square
of groups, J. Group Theory, 2010, 13, 83–94.
[3] K. S. Brown, Cohomology of groups. Graduate Texts in Mathematics, 87. Springer-Verlag, New
York, 1994. x+306 pp.
[4] N. D. Gilbert, P. J. Higgins, The non-abelian tensor product of groups and related constructions,
Glasgow Math. J., 1989, 31, 17–29.
[5] C. Miller, The second homology group of a group; relations among commutators, Proceedings
AMS 3, 588–595, (1952).
[6] J. Milnor, Introduction to Algebraic K-Theory, Princeton University Press, Princeton New
Jersey, 1971, pp. 1–184
[7] A. S.-T. Lue, The Ganea map for nilpotent groups, J. London Math. Soc. 14, 309–312, (1976).
[8] R. Brown, J.-L. Loday, Excision homotopique en basse dimension, C. R. Acad. Sci. Paris Ser.
I Math. 298 (15) (1984), 353–356.
[9] R. Brown, J.-L. Loday, Van Kampen theorems for diagrams of spaces, Topology 26 (3) (1987),
311–335, with an appendix by M. Zisman.
[10] G. J. Ellis, The non-abelian tensor product of finite groups is finite, J. Algebra 111 (1) (1987),
203–205.
[11] M. I. Kargapolov, Yu. I. Merzlyakov, Osnovy teorii grupp. (Russian) [Fundamentals of group
theory] Fourth edition. Fizmatlit “Nauka”, Moscow, 1996. 288 pp.
[12] M. P. Visscher, On the nilpotency class and solvability length of nonabelian tensor products of
groups, Arch. Math. (Basel) 73 (3) (1999), 161–171.
[13] I. N. Nakaoka, Non-abelian tensor products of solvable groups, J. Group Theory 3 (2) (2000),
157-167.
12
V. G. BARDAKOV, A. V. LAVRENOV, AND M. V. NESHCHADIM
[14] D. J. Collins, H. Zieschang, Combinatorial group theory and fundamental groups. Algebra, VII,
1-166, 233-240, Encyclopaedia Math. Sci., 58, Springer, Berlin, 1993.
[15] R. Brown, D. L. Johnson, E. F. Robertson, Some computations of non-abelian tensor products
of groups, J. Algebra, 1987, 111, 177–202.
[16] N. S. Romanovskii, Identity bases for certain matrix groups (Russian), Algebra i Logika, 1971,
10, 401–406.
[17] G. Donadze, M. Ladra, V. Thomas, On some closure properties of the non-abelian tensor
product, J. Algebra, 2017, 472, 399–413
[18] A. Kawauchi, A survey of knot theory. Translated and revised from the 1990 Japanese original
by the author. Birkhuser Verlag, Basel, 1996.
[19] A. I. Malcev, On isomorphic matrix representations of infinite groups. (Russian) Rec. Math.
[Mat. Sbornik] N.S. 8 (50), (1940), 405–422.
[20] B. A. F. Wehrfritz, Infinite linear groups, Berlin, 1973.
[21] J. H. C. Whitehead, On adding relations to homotopy groups, Ann. of Math., 1941, 42, 409–428.
Sobolev Institute of Mathematics, Novosibirsk 630090, Russia,
Novosibirsk State University, Novosibirsk 630090, Russia,
Novosibirsk State Agrarian University, Dobrolyubova street, 160, Novosibirsk, 630039,
Russia,
E-mail address: [email protected]
Saint-Petersburg State University, University Embankment, 7–9, Saint-Petersburg,
199034, Russia,
E-mail address: [email protected]
Sobolev Institute of Mathematics and Novosibirsk State University, Novosibirsk
630090, Russia,
E-mail address: [email protected]
| 4math.GR
|
“Zero-Shot” Super-Resolution using Deep Internal Learning
Assaf Shocher∗
arXiv:1712.06087v1 [cs.CV] 17 Dec 2017
∗
Nadav Cohen†
Michal Irani∗
Dept. of Computer Science and Applied Math, The Weizmann Institute of Science, Israel
†
School of Mathematics, Institute for Advanced Study, Princeton, New Jersey
Project Website: http://www.wisdom.weizmann.ac.il/∼vision/zssr/
nally supervised1 methods perform extremely well on data
satisfying the conditions they were trained on, their performance deteriorates significantly once these conditions are
not satisfied.
For example, SR CNNs are typically trained on highquality natural images, from which the low-resolution (LR)
images were generated with a specific predefined downscaling kernel (usually a Bicubic kernel with antialiasing
– MATLAB’s default imresize command), without any distracting artifacts (sensor noise, non-ideal PSF, image compression, etc.), and for a predefined SR scaling-factor (usually ×2, ×3 or ×4; assumed equal in both dimensions).
Fig. 2 shows what happens when these conditions are not
satisfied, e.g., when the LR image is generated with a nonideal (non-bicubic) downscaling kernel, or contains aliasing
effects, or simply contains sensor noise or compression artifacts. Fig. 1 further shows that these are not contrived cases,
but rather occur often when dealing with real LR images –
images downloaded from the internet, images taken by an
iPhone, old historic images, etc. In those ‘non-ideal’ cases,
SotA SR methods often produce poor results.
In this paper we introduce “Zero-Shot” SR (ZSSR),
which exploits the power of Deep Learning, without relying on any prior image examples or prior training. We exploit the internal recurrence of information within a single
image and train a small image-specific CNN at test time,
on examples extracted solely from the LR input image itself
(i.e., internal self-supervision). As such, the CNN can be
adapted to different settings per image. This allows to perform SR on real images where the acquisition process is unknown and non-ideal (see example results in Figs. 1 and 2).
On ‘non-ideal’ images, our method outperforms externallytrained SotA SR methods by a large margin.
The recurrence of small pieces of information (e.g.,
small image patches) across scales of a single image,
Abstract
Deep Learning has led to a dramatic leap in SuperResolution (SR) performance in the past few years. However, being supervised, these SR methods are restricted to
specific training data, where the acquisition of the lowresolution (LR) images from their high-resolution (HR)
counterparts is predetermined (e.g., bicubic downscaling),
without any distracting artifacts (e.g., sensor noise, image
compression, non-ideal PSF, etc). Real LR images, however, rarely obey these restrictions, resulting in poor SR results by SotA (State of the Art) methods. In this paper we introduce “Zero-Shot” SR, which exploits the power of Deep
Learning, but does not rely on prior training. We exploit
the internal recurrence of information inside a single image, and train a small image-specific CNN at test time, on
examples extracted solely from the input image itself. As
such, it can adapt itself to different settings per image. This
allows to perform SR of real old photos, noisy images, biological data, and other images where the acquisition process is unknown or non-ideal. On such images, our method
outperforms SotA CNN-based SR methods, as well as previous unsupervised SR methods. To the best of our knowledge,
this is the first unsupervised CNN-based SR method.
1. Introduction
Super-Resolution (SR) from a single image has recently
received a huge boost in performance using Deep-Learning
based methods [3, 9, 8, 11, 12]. The recent SotA (State of
the Art) method [12] exceeds previous non-Deep SR methods (supervised [21] or unsupervised [4, 5, 6]) by a few dBs
– a huge margin! This boost in performance was obtained
with very deep and well engineered CNNs, which were
trained exhaustively on external databases, for lengthy periods of time (days or weeks). However, while these exter-
1 We use the term “supervised” for any method that trains on externally supplied examples (even if their generation does not require manual
labelling).
1
(a) Historic image: Check-point Charlie (end of World-War II) – SR × 2
ZSSR (ours)
EDSR [12]
(b) iPhone image – SR × 3
(c) Historic image: JFK funeral – SR × 2
EDSR [12]
ZSSR (ours)
(d) Outdoor image downloaded from the Internet – SR × 2
EDSR [12]
ZSSR (ours)
EDSR [12]
ZSSR (ours)
Figure 1: SR of real images (unknown LR acquisition process). Real-world images rarely obey the ‘ideal conditions’
assumed by supervised SR methods. For example, old historic photos (a,c), images taken by smartphones (b), random images
on the Internet (d), etc. Since ZSSR trains at test time on examples extracted from the test image, it is better at performing SR
‘In-the-Wild’ (i.e., in unconstrained and unknown settings). Full sized images can be found on our project website.
2
(a) SR under aliasing:
Ground truth
(PSNR /SSIM)
EDSR+ [12]
( 21.64 / 0.6641)
ZSSR (ours)
(25.02 / 0.7658)
EDSR+ [12]
(24.44 / 0.7006)
ZSSR (ours)
(27.62 / 0.8367)
(b) SR under unknown non-ideal downscaling kernel:
Ground truth
(PSNR /SSIM)
Figure 2: SR of ‘non-ideal’ LR images – a controlled experiment. (a) LR image generated with aliasing (donwscaling
kernel is a delta function). (b) LR image generated with a non-ideal downscaling kernel. The unknown image-specific kernel
is estimated directly from the LR test image using [14], and fed into our image-specific CNN as the downscaling kernel (note
that externally-trained networks cannot make use of such image-specific information at test-time). Full sized images can be
found on our project website. Quantitative evaluation on hundreds of ‘non-ideal’ LR images can be found in Sec. 4.2.
was shown to be a very strong property of natural images [4, 23]. This formed the basis for many unsupervised image enhancement methods, including unsupervised SR [4, 5, 6], Blind-SR [14] (when the downscaling kernel is unknown), Blind-Deblurring [15, 1], BlindDehazing [2], and more. While such unsupervised methods
can exploit image-specific information (hence are less subject to the above-mentioned supervised restrictions), they
typically rely on simple Eucledian similarity of small image patches, of predefined size (typically 5 × 5), using Knearest-neighbours search. As such, they do not generalize
well to patches that do not exist in the LR image, nor to new
implicitly learned similarity measures, nor can they adapt to
non-uniform sizes of repeating structures inside the image.
Since the visual entropy inside a single image is much
smaller than in a general external collection of images [23],
a small and simple CNN suffices for this image-specific
task. Hence, even though our network is trained at test time,
its train+test runtime is comparable to the test runtime of
SotA supervised CNNs. Interestingly, our image-specific
CNN produces impressive results (although not SotA) on
the ‘ideal’ benchmark datasets used by the SotA supervised
methods (even though our CNN is small and has not been
pretrained), and surpasses SotA supervised SR by a large
margin on ‘non-ideal’ images. We provide both visual and
empirical evidence of these statements.
The term “Zero-Shot” used here, is borrowed from the
domains of recognition/classification. Note however, that
unlike these approaches for Zero-Shot Learning [22] or
One-shot Learning [18], our approach does not require any
side information/attributes or any additional images. We
may only have a single test image at hand, one of a kind,
and nothing else. Nevertheless, when additional information is available and provided (e.g., the downscaling kernel
can be estimated directly from the test image using [14]),
our image-specific CNN can make good use of this at test
time, to further improve the results.
Our image-specific CNN leverages on the power of the
cross-scale internal recurrence of image-specific information, without being restricted by the above-mentioned limitations of patch-based methods. We train a CNN to infer
complex image-specific HR-LR relations from the LR image and its downscaled versions (self-supervision). We then
apply those learned relations on the LR input image to produce the HR output. This outperforms unsupervised patchbased SR by a large margin.
3
Our contributions are therefore several-fold:
(i) To our best knowledge, this is the first unsupervised
CNN-based SR method.
(ii) It can handle non-ideal imaging conditions, and a wide
variety of images and data types (even if encountered for
the first time).
(iii) It does not require pretraining and can be run with modest amounts of computational resources.
(iv) It can be applied for SR to any size and theoretically
also with any aspect-ratio.
(v) It can be adapted to known as well as unknown imaging
conditions (at test time).
(v) It provides SotA SR results on images taken with ‘nonideal’ conditions, and competitive results on ‘ideal’ conditions for which SotA supervised methods were trained on.
Figure 3: Internal predictive power of image-specific information. Simple unsupervised internal-SR [4] is able to
reconstruct the tiny handrail in the tiny balconies, whereas
externally-trained SotA SR methods fail to do so. Evidence
to the existence of those tiny handrails exists only internally,
inside this image, at a different location and scale (in one
of the larger balconies). Such evidence is not found in any
external database of images (no matter how large it is).
2. The Power of Internal Image Statistics
Fundamental to our approach is the fact that natural images have strong internal data repetition. For example,
small image patches (e.g., 5×5, 7×7) were shown to repeat many times inside a single image, both within the same
scale, as well as across different image scales. This observation was empirically verified by [4, 23] using hundreds
of natural images, and was shown to be true for almost any
small patch in almost any natural image.
3. Image-Specific CNN
Our image-specific CNN combines the predictive power
and low entropy of internal image-specific information,
with the generalization capabilities of Deep-Learning.
Given a test image I, with no external examples available
to train on, we construct an Image-Specific CNN tailored
to solve the SR task for this specific image. We train our
CNN on examples extracted from the test image itself. Such
examples are obtained by downscaling the LR image I, to
generate a lower-resolution version of itself, I ↓ s (where
s is the desired SR scale factor). We use a relatively light
CNN, and train it to reconstruct the test image I from its
lower-resolution version I ↓ s (top part of Fig. 4(b)). We
then apply the resulting trained CNN to the test image I,
now using I as the LR input to the network, in order to construct the desired HR output I ↑ s (bottom of Fig. 4(b)).
Note that the trained CNN is fully convolutional, hence can
be applied to images of different sizes.
Since our “training set” consists of one instance only (the
test image), we employ data augmentation on I to extract
more LR-HR example-pairs to train on. The augmentation
is done by downscaling the test image I to many smaller
versions of itself (I = I0 , I1 , I2 , ..., In ). These play the role
of the HR supervision and are called “HR fathers”. Each
of the HR fathers is then downscaled by the desired SR
scale-factor s to obtain the “LR sons”, which form the input training instances. The resulting training set consists of
many image-specific LR-HR example pairs. The network
can then stochastically train over these pairs.
We further enrich the training set by transforming each
Fig. 3 shows an example of a simple single-image SR
based on internal patch recurrence (courtesy of [4]). Note
that it is able to recover the tiny handrails in the tiny balconies, since evidence to their existence is found elsewhere
inside this image, in one of the larger balconies. In fact,
the only evidence to the existence of these tiny handrails
exists internally, inside this image, at a different location
and different scale. It cannot be found in any external
database of examples, no matter how large this dataset is!
As can be seen, SotA SR methods fail to recover this imagespecific information when relying on externally trained images. While the strong internal predictive-power is exemplified here using a ‘fractal-like’ image, the internal predictivepower was analyzed and shown to be strong for almost any
natural image [4].
In fact, it was empirically shown by [23] that the internal entropy of patches inside a single image is much
smaller than the external entropy of patches in a general
collection of natural images. This further gave rise to
the observation that internal image statistics often provides
stronger predictive-power than external statistics obtained
from a general image collection. This preference was further shown to be particularly strong under growing uncertainty and image degradations (see [23, 16] for details).
4
Figure 4: Image-Specific CNN – “Zero-Shot” SR. (a) Externally-supervised SR CNNs are pre-trained on large external
databases of images. The resulting very deep network is then applied to the test image I. (b) Our proposed method (ZSSR): a
small image-specific CNN is trained on examples extracted internally, from the test image itself. It learns how to recover the
test image I from its coarser resolutions. The resulting self-supervised network is then applied to the LR image I to produce
its HR output.
LR-HR pair using 4 rotations (0◦ , 90◦ , 180◦ , 270◦ ) and their
mirror reflections in the vertical and horizontal directions.
This adds ×8 more image-specific training examples.
For the sake of robustness, as well as to allow large SR
scale factors s even from very small LR images, the SR is
performed gradually [4, 20]. Our algorithm is applied for
several intermediate scale-factors (s1 , s2 , ..., sm = s). At
each intemediate scale si , we add the generated SR image
HRi and its downscaled/rotated versions to our gradually
growing training-set, as new HR fathers. We downscale
those (as well as the previous smaller ‘HR examples’) by
the next gradual scale factor si+1 , to generate the new LRHR training example pairs. This is repeated until reaching
the full desired resolution increase s.
polated LR and its HR parent. We use L1 loss with ADAM
optimizer [10]. We start with a learning rate of 0.001. We
periodically take a linear fit of the reconstruction error and
if the standard deviation is greater by a factor than the slope
of the linear fit we divide the learning rate by 10. We stop
when we get to a learning rate of 10−6 .
To accelerate the training stage and make the runtime independent of the size of the test image I, at each iteration we
take a random crop of fixed size from a randomly-selected
father-son example pair. The crop is typically 128×128 (unless the sampled image-pair is smaller). The probability of
sampling a LR-HR example pair at each training iteration is
set to be non-uniform and proportional to the size of the HRfather. The closer the size-ratio (between the HR-father and
the test image I) is to 1, the higher its probability to be sampled. This reflects the higher reliability of non-synthesized
HR examples over synthesize ones.
Lastly, we use a method similar to the geometric
self-ensemble proposed in [12] (which generates 8 different
outputs for the 8 rotations+flips of the test image I, and
then combines them). We take the median of these 8
outputs rather than their mean. We further combine it
with the back-projection technique of [7, 4], so that each
of the 8 output images undergoes several iterations of
back-projection and finally the median image is corrected
by back-projection as well.
3.1. Architecture & Optimization
Supervised CNNs, which train on a large and diverse external collection of LR-HR image examples, must capture
in their learned weights the large diversity of all possible
LR-HR relations. As such, these networks tend to be extremely deep and very complex. In contrast, the diversity of
the LR-HR relations within a single image is significantly
smaller, hence can be encoded by a much smaller and simpler image-specific network.
We use a simple, fully convolutional network, with 8
hidden layers, each has 64 channels. We use ReLU activations on each layer. The network input is interpolated to
the output size. As done in previous CNN-based SR methods [9, 8, 3], we only learn the residual between the inter-
Runtime: Although training is done at test time, the average runtime per image is 54 sec for a single increase in SR
5
scale-factor (on a Tesla K-80 GPU). This runtime is independent of the image size or the relative SR scale-factor s
(this is a result of the equally sized random crops used in
training; the final test runtime is negligible with respect to
training iterations). Nevertheless, better SR results are obtained when using a gradual increase in resolution. For example, a gradual increase using 6 intermediate scale-factors
typically improves the PSNR by ∼0.2dB, but increases the
runtime to ∼5min per image. There is therefore a tradeoff
between runtime and the output quality, which is up to the
user to choose. Most of the results reported in this paper
were produced using 6 intermediate scale-factors.
For comparison, the test-time of the leading EDSR+ [12]
(on the same platform) is ∼20 sec for SR×2 on a 200×200
image. However, EDSR’s run time grows quadratically
with the image size, and reaches 5min for an 800 × 800
image. Beyond that size, our network is faster than EDSR+.
(whether due to sensor noise in the image, JPEG compression artifacts, etc.) We found that adding a small amount
of Gaussian noise (with zero mean and a small standarddeviation of ∼5 grayscales), improves the performance for a
wide variety of degradations (Gaussian noise, speckle noise,
JPEG artifacts, and more). We attribute this phenomenon
to the fact that image-specific information tends to repeat
across scales, whereas noise artifacts do not [24]. Adding
a bit of synthetic noise to the LR sons (but not to their HR
fathers) teaches the network to ignore uncorrelated crossscale information (the noise), while learning to increase the
resolution of correlated information (the signal details).
Indeed, our experiments show that for low-quality LR
images, and for a wide variety of degradation types, the
image-specific CNN obtains significantly better SR results
than SotA EDSR+ [12] (see Sec. 4). Similarly, in the case
of non-ideal downscaling kernels, the image-specific CNN
obtains a significant improvement over SotA (even in the
absence of any noise). When the downscaling kernel is
known (e.g., a sensor with a known PSF), it can be provided
to our network. When the downscaling kernel is unknown
(which is usually the case), a rough estimate of the kernel
can be computed directly from the test image itself (e.g.,
using the method of [14]). Such rough kernel estimations
suffice to obtain +1dB improvement over EDSR+ on nonideal kernels (see examples in Figs. 1 and 2, and empirical
evaluations in Sec. 4).
Note that providing the estimated downscaling kernel to
externally-supervised SotA SR methods at test time, would
be of no use. They would need to exhaustively re-train a
new network on a new collection of LR-HR pairs, generated
with this specific (non-parametric) downscaling kernel.
3.2. Adapting to the Test Image
When the acquisition parameters of the LR images from
their HR ones are fixed for all images (e.g., same downscaling kernel, high-quality imaging conditions), current supervised SR methods achieve an incredible performance [19].
In practice, however, the acquisition process tends to change
from image to image, since cameras/sensors differ (e.g., different lens types and PSFs), as well as the individual imaging conditions (e.g., subtle involuntary camera shake when
taking the photo, poor visibility conditions, etc). This results in different downscaling kernels, different noise characteristics, various compression artifacts, etc. One could
not practically train for all possible image acquisition configurations/settings. Moreover, a single supervised CNN is
unlikely to perform well for all possible types of degradations/settings. To obtain good performance, one would need
many different specialized SR networks, each trained (for
days or weeks) on different types of degradations/settings.
This is where the advantage of an image-specific network
comes in. Our network can be adapted to the specific degradations/settings of the test image at hand, at test time. Our
network can receive from the user, at test time, any of the
following parameters:
(i) The desired downscaling kernel (when no kernel is provided, the bicubic kernel serves as a default).
(ii) The desired SR scale-factor s.
(iii) The desired number of gradual scale increases (a tradeoff between speed and quality – the default is 6).
(iv) Whether to enforce Backprojection between the LR and
HR image (the default is ‘Yes’).
(v) Whether to add ‘noise’ to the LR sons in each LR-HR
example pair extracted from the LR test image (the default
is ‘No’).
The last 2 parameters (cancelling the Backprojection and
adding noise) allow to handle SR of poor-quality LR images
4. Experiments & Results
Our method (ZSSR - ‘Zero-Shot SR’) is primarily aimed
at real LR images obtained with realistic (unknown and
varying) acquisition setting. These usually have no HR
ground truth, hence are evaluated visually (as in Fig. 1).
In order to quantitatively evaluate ZSSR’s performance, we
ran several controlled experiments, on a variety of settings.
Interestingly, ZSSR produces competitive results (although
not SotA) on the ‘ideal’ benchmark datasets for which the
SotA supervised methods train and specialize (even though
our CNN is small, and has not been pretrained). However,
on ‘non-ideal’ datasets, ZSSR surpasses SotA SR by a large
margin. All reported numerical results were produced using
the evaluation script of [8, 9].
4.1. The ‘Ideal’ Case
While this is not the aim of ZSSR, we tested it also
on the standard SR benchmarks of ‘ideal’ LR images. In
these benchmarks, the LR images are ideally downscaled
from their HR versions using MATLAB’s imresize com6
Ground Truth
(PSNR, SSIM)
VDSR [8]
(20.11, 0.9136)
EDSR+ [12]
(25.29 / 0.9627)
ZSSR (ours)
(25.68 / 0.9546)
Figure 5: In images with strong internal repetitive structures, ZSSR tends to surpass VDSR, and sometimes also EDSR+,
even though the LR image was generated using the ‘ideal’ supervised setting (i.e., bicubic downscaling).
Dataset
Set5
Set14
BSD100
Scale
×2
×3
×4
×2
×3
×4
×2
×3
×4
SRCNN [3]
36.66 / 0.9542
32.75 / 0.9090
30.48 / 0.8628
32.42 / 0.9063
29.28 / 0.8209
27.49 / 0.7503
31.36 / 0.8879
28.41 / 0.7863
26.90 / 0.7101
Supervised
VDSR [8]
37.53 / 0.9587
33.66 / 0.9213
31.35 / 0.8838
33.03 / 0.9124
29.77 / 0.8314
28.01 / 0.7674
31.90 / 0.8960
28.82 / 0.7976
27.29 / 0.7251
EDSR+ [12]
38.20 / 0.9606
34.76 / 0.9290
32.62 / 0.8984
34.02 / 0.9204
30.66 / 0.8481
28.94 / 0.7901
32.37 / 0.9018
29.32 / 0.8104
27.79 / 0.7437
Unsupervised
SelfExSR [6]
ZSSR (ours)
36.49 / 0.9537 37.37 / 0.9570
32.58 / 0.9093 33.42 / 0.9188
30.31 / 0.8619 31.13 / 0.8796
32.22 / 0.9034 33.00 / 0.9108
29.16 / 0.8196 29.80 / 0.8304
27.40 / 0.7518 28.01 / 0.7651
31.18 / 0.8855 31.65 / 0.8920
28.29 / 0.7840 28.67 / 0.7945
26.84 / 0.7106 27.12 / 0.7211
Table 1: Comparison of SR results for the ’ideal’ case (bicubic downscaling).
mand (a bicubic kernel downsampling with antialiasing).
Table 1 shows that our image-specific ZSSR achieves competitive results against externally-supervised methods that
were exhaustively trained for these conditions. In fact,
ZSSR is significantly better than the older SRCNN [3], and
in some cases achieves comparable or better results than
VDSR [8] (which was the SotA until a year ago). Within
the unsupervised-SR regime, ZSSR outperforms the leading method SelfExSR [6] by a large margin.
Moreover, in images with very strong internal repetitive
structures, ZSSR tends to surpass VDSR, and sometimes
also EDSR+, even though these LR images were generated using the ‘ideal’ supervised setting. One such example is shown in Fig. 5. Although this image is not a typical natural image, further analysis shows that the preference for internal learning (via ZSSR) exhibited in Fig. 5
exists not only in ‘fractal-like’ images, but is also found in
general natural images. Several such examples are shown
in Fig. 6. As can be seen, some of the pixels in the image
(those marked in green) benefit more from exploiting internally learned data recurrence (ZSSR) over deeply learned
external information, whereas other pixels (those marked in
red) benefit more from externally learned data (EDSR+).
As expected, the internal approach (ZSSR) is mostly advantageous in image area with high recurrence of informa-
tion, especially in areas where these patterns are extremely
small (of extremely low resolution), like the small windows
in the top of the building. Such tiny patters find larger (highres) examples of themselves elsewhere inside the same image (at a different location/scale). This indicates that there
may be potential for further SR improvement (even in the
‘ideal’ bicubic case), by combining the power of InternalLearning with External-Learning in a single computational
framework. This remains part of our future work.
4.2. The ‘Non-ideal’ Case
Real LR images do not tend to be ideally generated. We
have experimented with non-ideal cases that result from
either: (i) non-ideal downscaling kernels (that deviate from
the bicubic kernel), and (ii) low-quality LR images (e.g.,
due to noise, compression artifacts, etc.) In such non-ideal
cases, the image-specific ZSSR provides significantly
better results than SotA SR methods (by 1 − 2dB). These
quantities experiments are described next. Fig. 2 shows a
few such visual results. Additional visual results and full
images can be found in our project website.
(A) Non-ideal downscaling kernels: The purpose of this
experiment is to test more realistic blur kernels with the
ability to numerically evaluate the results. For this purpose
7
VDSR [8]
27.7212 / 0.7635
EDSR+ [12]
27.7826 / 0.7660
Blind-SR [14]
28.420 / 0.7834
ZSSR [estimated kernel] (ours)
28.8118 / 0.8306
ZSSR [true kernel] (ours)
29.6814 / 0.8414
Table 2: SR in the presence of unknown downscaling kernels. LR images were generated from the BSD100 dataset using random
downscaling kernels (of reasonable size). SR×2 was then applied to those images. Please see text for more details.
Bicubic interpolation
27.9216 / 0.7504
EDSR+ [12]
27.5600 / 0.7135
ZSSR (ours)
28.6148 / 0.7809
Table 3: SR in the presence of unknown image degradation. Each LR image from the BSD100 dataset was randomly degraded using one of 3 types of degradations: (i) Gaussian noise,
(ii) Speckle noise, (iii) JPEG compression. SR×2 was then applied to those images, without knowing the type of degradation.
ZSSR shows robustness to unknown degradations, whereas SotA
SR methods are not. In fact, under such conditions, bicubic interpolation outperforms current SotA SR methods.
we created a new dataset from BSD100 [13] by downscaling
the HR images using random (but reasonably sized) Gaussian kernels. For each image, the covariance matrix Σ of its
downscaling kernel was chosen to have a random angle θ
and random lengths λ1 , λ2 in each axis:
Figure 6: Internal vs. External preference. Green: pixels that favor Internal-SR (i.e., pixels where ZSSR obtains
lower error with respect to the ground-truth HR image);
Red: pixels that favour External-SR (EDSR+). Notice
that edges (which are frequent in natural images) prefer
External-SR, whereas unique image structures (e.g., the tiny
windows at the top of the high building) prefer Internal-SR.
Such tiny patters find larger (high-res) examples of themselves elsewhere inside the same image (at a different location/scale).
λ1 , λ2 v U [0, s2 ]
θ v U [0, π]
Λ = diag(λ1 , λ2 )
cos(θ) − sin(θ)
U=
sin(θ) cos(θ)
Σ = UΛUt
(1)
generated by SotA methods are very blurry (see Fig. 2, and
project website) . Interestingly, the unsupervised Blind-SR
method of [14], which does not use any deep learning, also
outperforms SotA SR methods. This further supports the
analysis and observations of [17], that (i) an accurate donwscaling model is more important than sophisticated image
priors, and (ii) using the wrong donwscaling kernel leads to
oversmoothed SR results.
A special case of a non-ideal kernel is the δ kernel,
which results in aliasing. This case too, is not handled well
by SotA methods (see example in Fig. 2).
where s is the HR-LR downscaling factor. Thus, each
LR image was subsampled by a different random kernel. Table 2 compares our performance against the leading externally-supervised SR methods [12, 8]. We also
compared our performance to the unsupervised Blind-SR
method of [14]. We considered two cases for applying
ZSSR: (i) The more realistic scenario of unknown downscaling kernel. For this mode we used [14] to evaluate the
kernel directly from the test image and fed it to ZSSR. The
unknown SR kernel is estimated in [14] by seeking a nonparametric downscaling kernel which maximizes the similarity of patches across scales in the LR test image. (ii) We
applied ZSSR with the true downscaling kernel used to create the LR image. Such a scenario is potentially useful for
images obtained by sensors with known specs.
Note that none of the externally-supervised methods are
able to benefit from knowing the blur kernel of the test image (whether estimated or real), since they were trained
and optimized exhaustively for a specific kernel. Table 2
shows that ZSSR outperforms SotA methods by a very large
margin: +1db for unknown (estimated) kernels, and +2db
when provided the true kernels. Visually, the SR images
(B) Poor-quality LR images:
In this experiment, we
tested images with different types of quality degradation.
To test the robustness of ZSSR in coping with unknown
damage, we chose for each image from BSD100 [13] a random type of degradation out of 3 degradations: (i) Gaussian
noise [σ = 0.05], (ii) Speckle noise [σ = 0.05], (iii) JPEG
compression [quality = 45 (By MATLAB standard)]. Table 3 shows that ZSSR is robust to unknown degradation
types, while these typically damage SR supervised methods
to the point where bicubic interpolation outperforms current SotA SR methods!
8
5. Conclusion
We introduce the concept of “Zero-Shot” SR, which exploits the power of Deep Learning, without relying on any
external examples or prior training. This is obtained via a
small image-specific CNN, which is trained at test time on
internal examples extracted solely from the LR test image.
This yields SR of real-world images, whose acquisition process is non-ideal, unknown, and changes from image to image (i.e., image-specific settings). In such real-world ‘nonideal’ settings, our method substantially outperforms SotA
SR methods, both qualitatively and quantitatively. To our
best knowledge, this is the first unsupervised CNN-based
SR method.
[13]
[14]
[15]
[16]
[17]
References
[1] Y. Bahat, N. Efrat, and M. Irani. Non-uniform blind deblurring by reblurring. In ICCV, 2017. 3
[2] Y. Bahat and M. Irani. Blind dehazing using internal patch
recurrence. In ICCP, 2016. 3
[3] K. H. X. T. Chao Dong, Chen Change Loy. Learning a deep
convolutional network for image super-resolution. In European Conference on Computer Vision (ECCV), 2014. 1, 5,
7
[4] S. B. D. Glasner and M. Irani. Super-resolution from a single image. In International Conference on Computer Vision
(ICCV), 2009. 1, 3, 4, 5
[5] G. Freedman and R. Fattal. Image and video upscaling
from local self-examples. ACM Transactions on Graphics,
30(2):12, 2011. 1, 3
[6] J.-B. Huang, A. Singh, and N. Ahuja. Single image superresolution from transformed self-exemplars. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 5197–5206, 2015. 1, 3, 7
[7] M. Irani and S. Peleg. Improving resolution by image registration. CVGIP: Graphical Model and Image Processing,
53(3):231–239, 1991. 5
[8] J. Kim, J. Kwon Lee, and K. Mu Lee. Accurate image superresolution using very deep convolutional networks. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1646–1654, 06 2016. 1, 5, 6,
7, 8
[9] J. Kim, J. K. Lee, and K. M. Lee. Deeply-recursive convolutional network for image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR
Oral), June 2016. 1, 5, 6
[10] D. P. Kingma and J. Ba. Adam: A method for stochastic
optimization. CoRR, abs/1412.6980, 2014. 5
[11] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and
W. Shi. Photo-realistic single image super-resolution using a
generative adversarial network. arXiv:1609.04802, 2016. 1
[12] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee. Enhanced
deep residual networks for single image super-resolution.
In The IEEE Conference on Computer Vision and Pattern
[18]
[19]
[20]
[21]
[22]
[23]
[24]
9
Recognition (CVPR) Workshops, July 2017. 1, 2, 3, 5, 6, 7,
8
D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database
of human segmented natural images and its application to
evaluating segmentation algorithms and measuring ecological statistics. In Proc. 8th Int’l Conf. Computer Vision, volume 2, pages 416–423, July 2001. 8
T. Michaeli and M. Irani. Nonparametric blind superresolution. In International Conference on Computer Vision
(ICCV), 2013. 3, 6, 8
T. Michaeli and M. Irani. Blind deblurring using internal
patch recurrence. In European Conference on Computer Vision (ECCV). 2014. 3
I. Mosseri, M. Zontak, and M. Irani. Combining the power
of internal and external denoising. In ICCP, 2013. 4
A. A. B. N. A. l. Netalee Efrat, Daniel Glasner. Accurate blur
models vs. image priors in single image super-resolution. In
ICCV, 2013. 8
A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. P.
Lillicrap. Meta-learning with memory-augmented neural
networks. In International Conference on Machine Learning (ICML). 3
R. Timofte, E. Agustsson, M.-H. Van Gool, Luc Yang,
L. Zhang, et al. Ntire 2017 challenge on single image superresolution: Methods and results. In The IEEE Conference
on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017. 6
R. Timofte, R. Rothe, and L. Van Gool. Seven ways to improve example-based single image super resolution. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 5
R. Timofte, V. D. Smet, and L. V. Gool. A+: Adjusted anchored neighborhood regression for fast super-resolution. In
ACCV, 2014. 1
Y. Xian, B. Schiele, and Z. Akata. Zero-shot learning – the
Good, the Bad and the Ugly. In CVPR, 2017. 3
M. Zontak and M. Irani. Internal statistics of a single natural
image. In IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2011. 3, 4
M. Zontak, I. Mosseri, and M. Irani. Separating signal from
noise using patch recurrence across scales. In CVPR, 2013.
6
| 9cs.NE
|
arXiv:1307.5076v1 [cs.CE] 18 Jul 2013
Computational Science
Laboratory Technical Report
CSL-TR-3-2013
July 22, 2013
Alexandru Cioaca,
Adrian Sandu, and Eric de Sturler
Efficient methods for computing
observation impact
in 4D-Var data assimilation
Computational Science Laboratory
Computer Science Department
Virginia Polytechnic Institute and State University
Blacksburg, VA 24060
Phone: (540)-231-2193
Fax: (540)-231-6075
Email: [email protected]
Web: http://csl.cs.vt.edu
Innovative Computational Solutions
Low-rank Approximations
for Computing Observation Impact
in 4D-Var Data Assimilation
Alexandru Cioaca, Adrian Sandu
Abstract
We present an efficient computational framework to quantify the impact of
individual observations in four dimensional variational data assimilation. The
proposed methodology uses first and second order adjoint sensitivity analysis,
together with matrix-free algorithms to obtain low-rank approximations of observation impact matrix. We illustrate the application of this methodology to
important applications such as data pruning and the identification of faulty
sensors for a two dimensional shallow water test system.
Preprint submitted to Elsevier
July 22, 2013
Contents
1 Introduction
1
2 4D-Var Data Assimilation
2.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Computational aspects . . . . . . . . . . . . . . . . . . . . . . . .
2
2
3
3 Observation Impact
3.1 4D-Var sensitivity to observations .
3.2 Observation impact matrix . . . .
3.2.1 Main features . . . . . . . .
3.2.2 Computational ingredients .
3
3
5
5
6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Efficient Computation of Observation Sensitivity and Impact
4.1 An iterative (serial) approach for matrix-free low-rank approximations with SVD . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 An ensemble-based (parallel) approach for matrix-free low-rank
approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
7
9
5 Applications
10
5.1 Test problem: shallow water equations . . . . . . . . . . . . . . . 10
5.2 Data assimilation setup . . . . . . . . . . . . . . . . . . . . . . . 11
5.3 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.3.1 Validating the low-rank approximations of the observation
impact matrix . . . . . . . . . . . . . . . . . . . . . . . . 12
5.3.2 Observation impact . . . . . . . . . . . . . . . . . . . . . 13
5.3.3 Pruning observations based on sensitivity values . . . . . 17
5.3.4 Identifying faulty data using observation sensitivity . . . . 18
6 Conclusions and future work
19
2
1. Introduction
Data assimilation is a dynamic-data driven application that integrates information from physical observations with numerical model predictions. This
paper describes a systematic approach to quantify the contribution of each observation data point in improving the model state estimates. The focus is on
the four-dimensional variational (4D-Var) data assimilation approach, which fits
a model trajectory against time-distributed observations, to obtain maximum
likelihood estimates of the initial state and model parameters.
Computing the “observation impact” (OI) means quantifying the individual
(or group) contributions of “assimilated” data points to reduce the uncertainty
in state or model parameters. OI provides a numerical measure to distinguish
between crucial and redundant information, and benefits applications such as
big data pruning, error detection/correction, intelligent sensor placement, and
other decision-making processes.
There are several approaches available in the scientific literature to compute
OI. In early studies OI was associated with the energy of perturbations along
the dominant directions of maximum error growth [1, 2]. Other approaches
study observation impact using metrics from information theory [3, 4], statistical design [5], and robust control [6]. More recently, observation impact has
been assessed through adjoint-sensitivity analysis, as a counterpart measure to
observation sensitivity [7–9]. The current state-of-the-art uses the second-order
sensitivity equations of 4D-Var system[10–12], a powerful approach followed by
the present research study. The efficient implementation of this methodology
is hindered by several computational limitations related to the calculation of
second-order derivative approximations.
This work develops a systematic and efficient approach to compute OI in a
sensitivity analysis framework. Observation impact is formalized as a sensitivity matrix which maps changes in observation space to associated changes in
solution space. This matrix is usually not available in explicit (full) form. Our
computational approach makes smart use of tangent linear and adjoint models
[13–15] to propagate first and second order sensitivities through the data assimilation system, and obtain products of the OI matrix with user defined vectors.
Matrix-free linear algebra tools are necessary for this. Eigenvalue and singular
value decompositions are used to obtain low-rank approximations that capture
the main features of observation impact, while significantly reducing the computational burden. As of recently, low-rank approximations of matrices have
become very popular in image processing [16], information retrieval [17], and
machine learning [18] to extract correlations and remove noise from data. Two
alternative ways to compute a low-rank approximation for observation impact
are given, one of serial nature and one highly parallel. Several applications of OI
are exemplified using the two dimensional shallow water equations test problem.
The remaining part of the paper is structured as follows. Section 2 reviews
the formulation of 4D-Var data assimilation. Section 3 reviews the 4D-Var
sensitivity equations necessary for the observation impact matrix (following the
original derivation [11]). We discuss the structural and functional properties of
1
the impact matrix and the required numerical tools for evaluating its action.
Section 4 introduces low-rank approximations of the observation impact matrix.
Numerical experiments to exemplify and validate the observation impact matrix
are presented in Section 5. Conclusions and future directions of research are
outlined in Section 6.
2. 4D-Var Data Assimilation
2.1. Formulation
Data assimilation is the process by which model predictions are constrained
with real measurements [19, 20]. It combines three sources of information: an
a priori estimate of the initial state of the system (“background”), knowledge
of the physical laws governing the behavior of the system (captured by a numerical model), and measurements of the real system (“observations”). Data
assimilation solves the inverse problem for improving estimates of model states,
initial and boundary conditions, or various system parameters.
Four-dimensional variational (4D-Var) assimilation is formulated as a PDEconstrained nonlinear optimization. The improved initial state xa0 (“analysis”)
is obtained by minimizing the following cost function:
T
1
b
x0 − xb0 · B−1
(1a)
J (x0 ) =
0 · (x0 − x0 )
2
N
1X
T
(Hk (xk ) − yk ) · R−1
+
k · (Hk (xk ) − yk ) ,
2
k=1
xa0
=
arg min J (x0 )
x0
subject to xk = Mt0 →tk (x0 ) .
(1b)
Here M denotes the numerical model used to evolve the initial state vector x0
in time. Hk denotes the observation operator at assimilation time tk , and maps
the model state xk ≈ x(tk ) to the observation space. Rk is the observation
error covariance matrix. A practical assumption is that observations are independent, hence uncorrelated, which reduces Rk to a diagonal matrix containing
observation variances. The covariance matrices B0 and Rk are predefined, and
the quality of their approximation influences the resulting analysis.
The first term of the 4D-Var cost functional (1a) quantifies the mismatch
between the initial solution (x0 ) and the background state (xb0 ) at the initial
time (t0 ). This mismatch is computed in a least-squares sense, scaled by the
inverse background error covariance matrix B0 . The second term measures the
mismatch between the model trajectory (initialized from x0 ) and observations
yk taken at times tk , k = 0, . . . , N scaled by the inverse observation error
covariances Rk . Through B0 and Rk the inverse problem takes into account
the uncertainty in the data and model predictions. 4D-Var can be viewed as a
method of Bayesian inference, which computes the maximum likelihood initial
solution conditioned by the observations.
When assimilating observations only at the initial time t0 the method is
known as three-dimensional variational (3D-Var), as the additional time dimension is not present.
2
2.2. Computational aspects
The minimization problem (1b) is computed numerically using gradientbased iterative nonlinear solvers such as quasi-Newton, nonlinear conjugate gradients, and truncated Newton methods.
The iterative solver starts from an initial guess and advances to a minimizer,
along directions of descent computed using the gradients of J . The descent
directions depend on all three available sources of information; intermediate
iterations can be viewed as the result of partially assimilating information given
by observations.
The iterative nature of the solution approach has several important consequences. The resulting analysis is not the global minimizer, but rather is a local
one. When the computational cost per iteration is high, the available computational resources constrain the number of iterations that can be performed. Thus
in practice analyses are almost always suboptimal.
We will call the numerical model M (1b) the forward model (fwd) or forecast model. Applying gradient-based methods to minimize (1b) requires the
derivatives of J with respect to the initial model states. This can be achieved
using adjoint modeling [13, 14], a methodology that has been successfully implemented in optimization, sensitivity analysis and uncertainty quantification
[15, 21, 22]. The adjoint models (adj) can be based either on the linearization
of the original differential equations (“continuous” adjoint approach), or the linearization of the numerical method (“discrete” adjoint approach). A convenient
approach to generate discrete adjoint models is automatic differentiation [23],
which takes as input the source code of the fwd model, performs line by line
differentiation, and returns the source code of the adjoint models.
3. Observation Impact
3.1. 4D-Var sensitivity to observations
We now establish a relation between the analysis xa0 and the observations
yk . This is done using sensitivity analysis, and expresses how slight changes
in the observational data translate into changes in the resulting analysis. This
section follows the 4D-Var sensitivity approach of Daescu [11].
Consider the problem of finding a vector x = (x1 , x2 , ..., xn )T ∈ Rn that
minimizes the twice continuously differentiable cost function:
xa = arg min J (x, u) .
x
(2)
The function depends on the state vector x and parameter vector u ∈ Rm .
For any value ū of the parameter, the solution obeys the first order optimality
condition
∇x J (xa , ū) = 0 .
(3)
Assume that the Hessian of the function is positive definite at the optimum,
∇2x,x J (xa , ū) > 0. The implicit function theorem applied to (3) guarantees
3
there exists a vicinity of ū where the optimal solution is a smooth function of
the parameters, xa = xa (u), and has the sensitivity
−1
∇u xa (u) = −∇2u,x J (u, xa ) · ∇2x,x J (u, xa )
.
We apply this procedure to the 4D-Var cost function (1a) with the parameters
u being the observational data yk . The first-order necessary condition reads:
∇x0 J (xa0 )
=
B−1
0
a
x −x
b
+
N
X
MT0,k HTk R−1
k (Hk (xk ) − yk ) = 0 ,
(4)
k=1
where M0,k = M0t0 →tk (x0 ) is the tangent linear propagator associated with the
numerical model M, and Hk = Hk0 (xk ) is the linearized observation operator
at time tk . Differentiating (4) with respect to the observations yk
∇2yk ,x0 J (xa0 ) = −Rk Hk M0,k
leads to the following expression for the sensitivity of the analysis to observations:
a T
∂x0
a
= R−1
(5)
∇yk x0 =
k Hk M0,k A0 ,
∂yk
−1
A0 = ∇2x0 ,x0 J (xa0 )
.
(6)
Consider a verification cost functional Ψ : Rn → R that measures the discrepancy between the analysis and a verification solution xv0 :
Ψ(xa0 ) =
1 a
(x − xv0 )T C (xa0 − xv0 ) .
2 0
(7)
Here the verification solution xv0 is also defined at t0 . The matrix C is a weighting matrix corresponding to a particular norm or restricting the verification to
a subdomain of the solution space. Using chain-rule differentiation and (5), the
sensitivity to observations (the gradient of Ψ with respect to observations yk )
is:
∇yk Ψ(xa0 ) = ∇yk xa0 · ∇xa0 Ψ(xa0 )
= R−1
k
Hk M0,k A0 C (xa0
(8)
−
xv0 ) .
The sensitivity (8) can be computed via the following steps:
• The sensitivity of the verification function with respect to the analysis is
∇xa0 Ψ(xa0 ) = C (xa0 − xv0 ) .
• A “supersensitivity” is obtained through solving a linear system with the
matrix the 4D-Var Hessian eqn:inverse-hessian
µ = A0 · ∇xa0 Ψ(xa0 ) .
(9)
This step dominates the computational cost of the sensitivity calculation.
4
• Finally, the verification sensitivity to observations yk is the supersensitivity vector propagated to time tk through the tangent linear model,
mapped through the observation selection operator, and scaled with the
inverse error covariance:
∇yk Ψ(xa0 ) = R−1
k Hk M0,k · µ .
(10)
The sensitivities of Ψ to the background xb0 and to the error covariance matrices
B and R can be derived in a similar fashion [11].
3.2. Observation impact matrix
3.2.1. Main features
Equation (5) defines a matrix whose elements are the sensitivities of each
component of the analysis vector xa0 to each component of the observation vector
yk assimilated at time tk . The observation impact matrix collects the sensitivities to all observations as follows:
−1
R1 H1 M0,1
∇y1 xa0
H2 M0,2
a T ∇y2 xa0 R−1
2
∂x
A0 .
=
=
(11)
T=
..
..
∂y
.
.
∇yN xa0
R−1
H
M
N
0,N
N
Each of the n columns of T represents the sensitivities of one particular
model state to all observations. For real applications n ∼ 107 − 1010 . Each row
of T contains the sensitivities of each state to one particular observation. In
typical data assimilation applications the number of observations (rows) is twothree orders of magnitude smaller than the number of model states (columns).
We now seek to understand the structure of the observation impact matrix
(11), which is the transpose of the sensitivity matrix ∂xa /∂y. The impact matrix is the product of submatrices of type R−1
k Hk M0,k , with the inverse of
the 4D-Var Hessian (6). The Hessian is symmetric and positive-definite when
evaluated at the minimum, but can lose positive-definiteness when evaluated at
an inexact analysis, such as when the minimization of 4D-Var was incomplete.
The inverse of the 4D-Var Hessian at the minimum approximates the analysis
error covariance [24]. This aposteriori error covariance characterizes the 4DVar process and quantifies the uncertainty reduction due to assimilating the
observations. Each column of this matrix represents the error covariance corresponding to a certain model state. The tangent linear model M0,k propagates
this error covariance to the observational time tk . The propagated perturbations
are mapped to observation space through Hk , and then scaled with R−1
k . The
inverse observation error covariance can be interpreted as a measure of trust assigned to each data point. Large values of the result are associated to points in
space and time where better measurements can benefit the assimilation process.
5
This metric can be assessed not only at times when we have observations available, but also at intermediate times, and can prove to be useful for deploying
adaptive observations.
Small changes in observations ∆y lead to a change in the analysis ∆xa0 which,
to first order, can be calculated using the observation impact matrix (11) as
∆xa0 = TT · ∆y .
(12)
We call the change in analysis ∆xa0 the observation impact. It is computed via
the following steps:
• Each observation change ∆yk is scaled and pulled back to the initial time
via the adjoint model:
∆yk = MT0,k HTk R−1
k · ∆yk .
• The aggregated contribution of all observation changes is
∆y =
N
X
∆yk ,
k=0
and in practice can be computed via a single adjoint run.
• The observation impact is obtained by solving a linear system whose matrix is the 4D-Var cost function Hessian
∆xa0 = A0 · ∆y .
3.2.2. Computational ingredients
For real models the matrices that appear in (11) are very large. The calculations of observation impact rely on matrix-vector products. The linear operator
M0,k is applied by running the tangent linear model from t0 to tk . Application
of the linearized observation mapping operator Hk and scaling by R−1
k can be
performed directly as they are relatively inexpensive.
The inverse Hessian has been successfully used in practical applications to
compute the supersensitivity. This was done by solving iteratively a linear
system for the supersensitivity, the system matrix being the 4D-Var Hessian.
Iterative solvers of Krylov-type require only Hessian-vector products. These
products can be computed by running the second-order adjoint model. Furthermore, the linear solvers can be preconditioned for faster convergence and several
matrix-free methods are readily available [25]. When the second-order adjoint
is not available, several approximations can be used as follows.
1. The finite difference of gradients
∇2x0 ,x0 J (xa0 ) · u ≈
∇x0 J (xa0 + · u)T − ∇x0 J (xa0 )T
.
requires two first-order adjoint runs, which can be performed in parallel.
The accuracy of this approximation is typically low.
6
2. The Gauss-Newton approximation of the Hessian
∇2x0 ,x0 J (xa0 ) · u ≈ B−1
0 ·u+
N
X
MT0,k HTk R−1
k Hk M0,k · u
k=1
discards second order terms which contain the observation residual, so
this approximation is independent of the observation data. It requires one
tangent-linear model run followed by one first-order adjoint run.
3. Limited memory quasi-Newton approximations are based on the sequence
of solutions and gradients generated during the numerical optimization
procedure. An example of such an approximation is L-BFGS [26].
4. Efficient Computation of Observation Sensitivity and Impact
The computation of observation sensitivity or impact is a non-trivial task
since for practical problems of interest it is not feasible to build the full matrices.
We seek to develop computationally inexpensive approximations of the impact
matrix that capture the most important features, and whose accuracy can be
improved with an increased computational effort. Our approach is based on
the Singular Value Decomposition (SVD), a powerful tool to generate low-rank
approximations of large matrices. We present two algorithms, one iterative
(inherently serial), and one ensembled-based (inherently parallel).
Consider the SVD of (11) and the corresponding low-rank approximations
T = U S VT ,
T(p) = U(p) S(p) V(p) ,
where U and V are orthogonal, S is diagonal, U(p) and V(p) are the right and
the left singular vectors associated with the largest p singular values, and S(p)
has these dominant singular values on the diagonal formed.
T(p) has the smallest ”reconstruction error” in both Frobenius kT − T(p) kF
and L2 norms kT − T(p) k2 among all the rank p approximations of T. The
accuracy increases as more dominant singular modes are added. The cut-off
threshold is particular to the problem under study and can be determined from
the singular value spectrum decay.
4.1. An iterative (serial) approach for matrix-free low-rank approximations with
SVD
An ideal iterative SVD algorithm for our problem uses one matrix-vector
product per iteration, and reveals one new singular pair with each iteration.
The singular vectors are discovered in decreasing order of the magnitude of their
singular values. Thus, running the algorithm for p iterations would generate the
leading p singular values and their associated singular vectors.
There are no classic algorithms to compute iteratively the SVD of matrix
available only in operator form. We change our problem to computing the
7
leading eigenvectors of the product between the observation impact matrix and
its transpose, T∗ T = US2 U∗ , where
T∗ T
M0,k
= A∗0
=
N
X
∗
M0,k M0,k A0 =
k=1
R−1
k Hk
N
X
T∗k Tk ,
(13)
k=1
M0,k ,
Tk = M0,k A0 .
This problem is solved using Krylov-based approaches (Lanczos [27], Arnoldi
[28]), e.g., by the Jacobi-Davidson algorithm [29], available in the JDQZ software
library [30].
Our algorithm computes a low-rank approximation of A0 , the inverse 4D-Var
Hessian. An iterative procedure can be used to compute the smallest eigenvalues and the corresponding eigenvectors of the Hessian matrix, which are the
dominant eigenpairs of A0 . The low-rank approximation of the inverse Hessian
reads:
−1
∗
A0 = (V D V∗ ) = V D−1 V∗ ≈ V(p) D−1
V
.
(14)
(p)
(p)
We replace the inverse Hessian in (13) with its low-rank approximation:
∗
−1
∗
∗
M
M
V
D
V
T∗k Tk ≈
V(p) D−1
V
0,k
(p)
0,k
(p)
(p)
(p)
(p)
Wk
=
−1
∗
∗
V(p) D−1
(p) Wk Wk D(p) V(p) ,
=
M0,k V(p) .
(15)
The columns of Wk are the Hessian eigenvectors V(p) propagated forward
and scaled by the tangent linear model M0,k . The p tangent linear models can
be performed in parallel. This allows us to approximate (13) as:
T∗ T ≈
W
=
−1
∗
V(p) D−1
(p) W D(p) V(p) ,
N
X
Wk∗ Wk .
k=1
With the eigendecomposition:
−1
∗
D−1
(p) W D(p) = Vred Dred Vred ,
(16)
the approximation of (13) becomes
∗
T∗ T ≈ V(p) Vred Dred V(p) Vred .
(17)
This represents a truncated singular vector decomposition of T, with Dred the
matrix of dominant singular values, and Vp Vred the matrix of left singular
vectors.
Algorithm 1 summarizes the main computational steps. The computational
cost is dominated by the first step, where the expensive second-order adjoint
model is run repeatedly to generate the Hessian eigenpairs. The iterative approach is suited for applications that benefit from an iterative improvement of
the low-rank approximation. The methodology can be applied to any data assimilation system for which first and second order adjoint models are available.
8
Algorithm 1 Iterative algorithm for low-rank approximations
Solve iteratively the eigenvalue problem for the 4D-Var Hessian (14)
Map newly generated eigenvectors through the tangent linear model (15)
3: Compute the truncated SVD of the resulting matrix (16)
4: Project the left singular vectors onto the eigenvector base of the 4D-Var
Hessian and build the low-rank approximation of T (17)
1:
2:
4.2. An ensemble-based (parallel) approach for matrix-free low-rank approximations
This approach uses a “randomized SVD” algorithm [31] to compute the
Moore-Penrose pseudoinverse [32] of the Hessian. The 4D-Var Hessian matrix
A−1
0 is available only in operator form, i.e., only matrix vector products can be
evaluated by running the second order adjoint. The randomized algorithm is as
follows:
1: Draw p random vectors and form a matrix Ω.
−1
2: Compute the product Y = A0 Ω using Hessian-vector multiplications, i.e.,
running the second order order adjoint model for each column.
3: Construct the QR decomposition Y = QR.
Each of the above steps can be performed in parallel.
The columns of Q form an orthonormal basis for the range of Y. Randomized
SVD uses a series of algebraic manipulations, starting from the observation that
Q is also the orthonormal factor in the QR decomposition of A−1
0 :
A−1
0 = QB,
B = Q∗ A−1
0 ,
B∗ = A−1
0 Q.
(18)
Next, compute an SVD of B:
∗
B = UB ΣB VB
(19)
and replace (19) in (18) to obtain the SVD of A−1
0 :
A−1
0
∗
∗
= Q UB ΣB VB
= U A ΣB V B
.
(20)
The left singular vectors of A represent the projections of the left singular
vectors of B onto the columns of Q. The singular values and right singular
vectors of A are the same as those of B. The pseudoinverse of the 4D-Var
Hessian A−1
0 reads:
+
∗
A+
(21)
0 ≈ V B ΣB U A .
The observation impact matrix is approximated using the tangent linear
model M0,k and the pseudoinverse A+
0:
T≈
N
X
M0,k A+
0 .
k=1
The computational flow is summarized in Algorithm 2.
9
(22)
Algorithm 2 Sampling algorithm for low-rank approximations
Build the matrix B, through parallel second-adjoint runs (18)
Compute a full SVD of B (19)
−1
3: Project the left singular vectors of B in Q and form the SVD of A0 (20)
+
4: Compute the Hessian pseudoinverse A0 (21)
5: Build the impact matrix T through parallel tangent linear runs (22)
1:
2:
Computing the rows of B (18) is done as matrix-vector products through
second-order adjoint runs that can be performed in parallel. The last step,
which propagates the components of the pseudoinverse through the linearization
of the model, is achieved by multiple tangent linear model runs in parallel. The
tangent model results are checkpointed at each of the observation times, so that
only one run across the entire time horizon is necessary for each input vector.
5. Applications
We illustrate several applications of the observation impact matrix in 4DVar using the two-dimensional shallow water equations. We first describe the
system and its numerical discretization, then present the 4D-Var implementation and the experimental setting for data assimilation. Observation sensitivity
is computed both in full and using low-rank approximation to assess how well
the latter captures the essential features. We apply this analysis to three applications, namely, detecting the change in impact from perfect data to noisy
data, pruning the least important observations, and detecting faulty sensors.
5.1. Test problem: shallow water equations
The two-dimensional shallow-water equations (2D swe) [33] approximate
the movement of a thin layer of fluid inside a basin:
∂
∂
∂
h+
(uh) +
(vh) = 0
∂t
∂x
∂y
∂
∂
∂
1
(uh) +
(uvh) = 0
u2 h + gh2 +
∂t
∂x
2
∂y
∂
∂
∂
1 2
2
(vh) +
(uvh) +
v h + gh
= 0.
∂t
∂x
∂y
2
(23)
Here h(t, x, y) is the fluid layer thickness, and u(t, x, y) and v(t, x, y) are the
components of the velocity field of the fluid. The gravitational acceleration is
denoted by g.
We consider a spatial domain Ω = [−3, 3]2 (spatial units), and an integration window is t0 = 0 ≤ t ≤ tf = 0.1 (time units). Boundary conditions are
considered periodic. The space discretization is realized using a finite volume
scheme, and the time integration uses a fourth Runge-Kutta scheme, following
10
Table 1: Normalized CPU times of different sensitivity models. The forward
model takes one time unit to run.
fwd
tlm
foa
soa
1
2.5
3.7
12.8
fwd + tlm
fwd + foa
fwd + tlm + foa + soa
3.5
4.7
20
the method Lax-Wendroff [34]. The model uses a square -q × q uniform spatial discretization grid, which brings the number of model (state) variables to
n = 3 q2 .
We use the automatic differentiation tool TAMC [35, 36] to build various
sensitivity models, as follows. The tangent-linear model (tlm) propagates perturbations forward in time. The first-order adjoint model (foa) propagates
perturbations backwards in time, and efficiently computes the gradient of a
scalar cost functional defined on the model states. The second-order adjoint
model (soa) computes the product between the Hessian of the cost function
and a user-defined vector [22].
The overhead introduced by the sensitivity models is considerable. Table 1
presents the CPU times of tlm, foa, and soa models (normalized with respect
to that of one fwd run). One soa integration is about 3.5 times more expensive
than a single first-order adjoint run, while the foa takes 3.7 times longer than
the forward run. These relative costs depend on the discretization methodology
and the implementation of sensitivities. Our previous research showed how to
build efficient adjoint models by reusing computations performed in the forward
run [22]. For example, for the shallow water model, the alternative continuous
adjoints we built required a fraction of the forward model CPU time to run.
5.2. Data assimilation setup
The 4D-Var system is set up for a simple version of the “circular dam”
problem [37]. The reference initial height field h is a Gaussian bell of a width
equal to 1 length units centered on the grid, and the reference initial velocity
vector components are constant u = v = 0. The physical interpretation is that
the front of water falls to the ground (h decreases) under the effect of gravity
and creates concentric ripples which propagate towards the boundaries. Figures
1 represent snapshots of the reference trajectory at initial and final time.
The computational grid is square and regular with q = 40 grid points in
each direction, for a total of 4800 model variables (states). The simulation time
interval is set to 0.01 seconds, using N = 100 timesteps of size 0.0001 (time
units).
The h component of the a priori estimate (background) xb is generated by
adding a correlated perturbation to the h reference solution at initial time. The
background error covariance B0 corresponds to a standard deviation of 5% of
the reference field values. For the u and v components we use white noise to
11
40
130
40
130
35
125
35
125
30
120
30
120
25
115
25
115
20
110
20
110
15
105
15
105
10
100
10
100
5
95
5
95
0
0
90
0
0
10
20
30
40
(a) Initial time
10
20
30
40
90
(b) Final time
Figure 1: The height field h at the beginning and at the end of the reference
trajectory.
prescribe perturbations. The spatial error correlation uses a Gaussian decay
model, with a correlation distance of five grid points.
Synthetic observations are generated at the final time, by adding random
noise to the reference trajectory. Since the observation errors are assumed uncorrelated, the observation error covariance matrix R is diagonal. The standard
deviation for observation noise is 1% of the largest absolute value of the observations for each variable. The observation operator H is linear and selects observed
variables at specified grid points. For the following experiments, we consider
observations of all variables at each grid point.
The minimization of the 4D-Var cost function is performed with the L-BFGSB solver [26] using a fixed number of 100 iterations.
5.3. Experimental results
5.3.1. Validating the low-rank approximations of the observation impact matrix
In the first experiment, we study the effect of two possible sources of errors:
data noise and SVD truncation errors. These issues are inherent to performing
the data assimilation and the observation impact analysis.
We apply our computational methodology to the data assimilation scenario
introduced in Section 5.2. To assess how the reanalysis is affected by the presence of noise, we first assimilate perfect observations (i.e., reference model values), then assimilate the same observations with added small noise. Figure 2
reveals that the convergence of the numerical optimization procedure is similar
in the two scenarios, and leads to a similar decrease in the root mean square
(RMS) error for each variable.
We now apply the systematic approach presented in Section 4 to each one
of the two assimilation scenarios, in order to compute the sensitivity to observations (8) when the verification solution xv0 is the background xb0 and C is the
identity matrix:
∇yk Ψ(xa0 ) = T (xa0 − xb0 ) .
12
(24)
10
RMSE
RMSE
1
PERFECT DATA
NOISY DATA
1.3
PERFECT DATA
NOISY DATA
PERFECT DATA
NOISY DATA
1.3
10
RMSE
2
10
1.2
10
10
1.2
10
0
50
BFGS iterations
(a) h
100
0
50
BFGS iterations
(b) u
100
0
50
BFGS iterations
100
(c) v
Figure 2: Root mean square error decrease for each variable versus the number
of L-BFGS iterations.
This procedure provides the sensitivity of the 4D-Var increment norm xa0 − xb0
to data. Large sensitivity values will be associated to observations that had a
more important contribution to the 4D-Var correction.
Figure 3 plots the observation sensitivity (24) for each one of h, u and v, for
both perfect and noisy scenarios. The standard deviation for observation noise
is equal to 1% of the largest absolute value of the observations for each variable.
We notice the observation sensitivity is locally correlated. For the h observations
∇yk Ψ(xa0 ) is almost symmetric, while for u and v is aligned along the East-West
and South-North directions. Observation sensitivities for perfect and for noisy
data exhibit similar spatial features and numerical values. Reasonably small
noise in data does not significantly affect the sensitivity fields.
The results in Figure 3 are obtained with the full sensitivity matrix, built
offline. We compare these results with a low-rank approximation of the sensitivity matrix computed using Algorithm 1. The low rank approximation of the
sensitivity is
(∇yk Ψ(xa0 ))(p) = T(p) · xa0 − xb0 .
(25)
Figure 4 displays the singular value spectrum of the observation impact matrix
and the reconstruction error of the low-rank truncated values for observation
sensitivity. It can be noticed that the decay in singular value spectrum resembles
that of the truncation error. Based on this, a good choice for the rank of the
approximation is p ≈ 1600, which corresponds to one third of the full workload.
In Figure 5 we plot the low-rank approximation (25) of sensitivity to h
observations and the corresponding truncation error. A visual comparison with
the full-rank observation sensitivity plotted in Figure 3(a) reveals that the lowrank approximation captures well the main features of the sensitivity field.
5.3.2. Observation impact
The computational methodology developed in Section 4 is now applied to
compute the forward observation impact. Specifically, we map changes in the
observations (innovation vector ∆y = y − H(x)) to changes in the analysis
(∆xa0 ) using the relation (12).
13
40
40
5
4
30
30
4
20
2
y
y
3
20
2
1
10
10
0
0
0
0
20
x
0
0
40
(a) perfect h observations
20
x
40
(b) noisy h observations
40
40
1
0.5
0
20
−0.5
10
1
30
0.5
y
y
30
20
0
10
−0.5
−1
0
0
20
x
−1
0
0
40
(c) perfect u observations
40
(d) noisy u observations
40
40
30
1
30
1
20
0
20
0
y
y
20
x
10
10
−1
−1
0
0
20
x
0
0
40
(e) perfect v observations
20
x
40
(f) noisy v observations
Figure 3: Sensitivity field (24) for perfect and for noisy observations of h, u,
and v.
14
5
10
2
10
RMS
Value
0
10
0
10
−2
10
−5
10
−10
0
1600
3200
Mode
10
4800
(a) Singular value spectrum
0
1600
3200
Rank
4800
(b) Truncation error
Figure 4: Singular value decay for the observation impact matrix T (11) and
the corresponding truncation error norm.
40
40
5
4
30
0.4
30
0.2
20
0
10
−0.2
20
y
y
3
2
1
10
0
0
0
20
x
−0.4
0
0
40
(a) Low-rank estimate
20
x
40
(b) Truncation error field
Figure 5: Low-rank approximation of observation sensitivity (25) for h data and
the associated truncation error field for 1600 modes.
15
40
40
0.5
30
20
y
y
30
0
10
0
0
0.5
20
0
10
20
x
40
−0.5
0
0
−0.5
20
x
40
(a) Full-rank impact for center obser- (b) Low-rank impact for center observation
vation
40
0.05
30
0.05
30
20
0
20
10
0
0
y
y
40
−0.05
20
x
10
−0.05
0
0
40
0
20
x
40
(c) Full-rank impact for corner obser- (d) Low-rank impact for corner observation
vation
Figure 6: The impact of two h observations from assimilation time t1 00, placed
in the center and in the corner of the grid. Both full rank and reduced rank
solutions are shown. 500 modes are used for the reduced rank approximation.
We compute the observation impact in full-rank and low-rank corresponding
to two observations of h, one in the center (grid coordinates (20, 20)) and one in
the corner (at (5, 5)) at the locations represented with white markers. This is
achieved by multiplying TT with a vector containing just the innovation brought
by the observation whose impact we want to evaluate (all other vector entries
have value zero). The resulting impact fields are plot in Figure 6. The spatial
features of the observation impact have a local radial correlation in both cases.
This means the information carried by the observation is spread in its proximity
by the 4D-Var process; this is also true across the periodic boundaries of the
shallow water system. Moreover, the low-rank approximations are able to pick
up the important features of the full-rank calculations, and provide impacts of
a similar magnitude.
Having already computed the SVD of the observation impact matrix, we also
look at the directions in the data space ∆y that have the largest impact on the
analysis, and the directions in the 4D-Var correction space ∆x that benefit most
from the assimilation. These directions are given by the dominant left and right
singular vectors of TT , respectively. Figure 7 plots the first dominant left (and
16
8
40
1000
30
0
20
2
30
y
y
x 10
40
20
0
−1000
10
10
−2
−2000
0
0
20
x
0
0
40
20
x
40
(a) First dominant left singular vector (b) First 500 dominant left singular
(model space)
vectors (model space)
6
40
x 10
40
3
30
10
30
20
0
20
2
10
0
0
y
y
1
−1
10
−10
20
x
0
−2
0
0
40
20
x
40
(c) First dominant right singular vector (d) First 500 dominant right singular
(observation space)
vectors (observation space)
Figure 7: Principal directions of growth in observation and solution space, as
defined by the dominant left and right singular vectors of the observation impact
matrix.
right) singular vectors corresponding to the h variable, and also the composition
of the most important 500 directions for the observation and solution space using
the formula
vdomin =
500
X
s2i ∗ vi ,
(26)
i=1
where si and vi are the singular pair corresponding to the i-th mode.
5.3.3. Pruning observations based on sensitivity values
In this experiment we illustrate how sensitivity analysis can be used to select
the most useful data points.
For this, we compute the 4D-Var reanalysis and then apply the observation
impact methodology to compute the sensitivity of the cost function Ψ to each
one of the 4800 observations (24). We then split our observation set in three
subsets of 1600 observations, corresponding to each one of h, u and v. Within
each subset, we order the observations by their sensitivity; the top 800 observations form the high sensitivity set, and the ones in the bottom 800 form the
17
RMSE [log scale]
RMSE [log scale]
100
FULL
HIGH
LOW
10
1
1.3
10
FULL
HIGH
LOW
1.2
10
10
100
BFGS iterations [log scale]
1
10
100
BFGS iterations [log scale]
(a) h RMS error decrease versus the (b) u RMS error decrease versus the
number of L-BFGS iterations.
number of L-BFGS iterations.
30
1.3
10
y
RMSE [log scale]
40
1.2
10
FULL
HIGH
LOW
20
10
0
0
1
10
100
BFGS iterations [log scale]
10
20
x
30
40
(c) v RMS error decrease versus the (d) Location of high (red) and low imnumber of L-BFGS iterations.
pact observations
Figure 8: Location of high and low sensitivity observations, and their use in
data assimilation.
low sensitivity set. This procedure partitions our initial 4800 observations in
two halves with respect to observation sensitivity, taken variable-wise. The 4DVar data assimilation process is repeated using either one of the high or low
data sets.
Figure 8(a) shows the distribution of high versus low observations on our
computational grid. We run 4D-Var to assimilate only the high, followed by
assimilating only the low data points, and compare the decrease in the RMS
true error in h in Figure 8(b). The convergence results show that assimilating
observations of larger sensitivity yields slightly better results. When the number
of optimizer iterations increases, the two scenarios exhibit similar performance,
comparable to using all the observations.
5.3.4. Identifying faulty data using observation sensitivity
Finally, we consider the problem of detecting isolated faulty sensors and
try to solve it using our new computational capabilities. We prescribe certain
sensors to exhibit the wrong behavior and expect the observation sensitivity
analysis to reflect this.
18
We perform the 4D-Var scenario above again, using perfect observations at
each grid point, with the exception of two points where the observation values
are increased by a factor of 10 (thus simulating two faulty instruments). The two
points are located at coordinates (20, 20), the grid center where fluid height field
has large values initially, and at coordinates (10, 10), which is initially outside
the water height profile.
The 4D-Var data assimilation uses the faulty data together with the correct
values. Figure 9(a) plots the correction in the variable h, i.e., the quantity
(xa0 − xb0 ). The correction field is smooth and there is no indication that there
are very large errors in some data points. Figure 9(b) plots the supersensitivity
vector µ obtained by mapping the increment through the inverse 4D-Var Hessian
(9). The supersensitivity field exhibits a clear structure consisting of large
values in concentric circles around the two locations with faulty observations.
Figure 9(c) shows the observation sensitivity y, obtained by propagating the
supersensitivity through the tangent linear model (10). The observation impact
methodology clearly identifies two points of abnormally large sensitivity. These
are exactly the locations of the faulty sensors, and because their values are very
different than the surrounding field, they have an abnormally large impact on
the analysis. This is precisely identified by the observation impact metrics.
6. Conclusions and future work
This research is motivated by dynamic data driven applications which use
measurements of the real system state to constrain computer model predictions.
We develop efficient approaches to compute the observation impact in 4D-Var
data assimilation, i.e., the contribution of each data point to the final analysis. Quantification of observation impact is important for applications such
as data pruning, tuning the parameters of data assimilation, identification of
faulty sensors, and optimal configuration of sensor networks. Immediate applications include numerical weather predicition, climate and air-quality forecast,
hydrology, renewable energy systems and biohazard proliferation.
While 4D-Var is the state of the art data assimilation methodology, and is
widely used in many operational centers, it is one of the most computationally
challenging approaches to data assimilation. The computation of observation
impact adds to the cost of performing analyses. We review the derivation of
the sensitivity equations that lead to the observation impact matrix, and make
in-depth comments about its structure, functionality and computation requirements. One contribution of this work is to develop highly efficient implementation strategies for the 4D-Var sensitivity equations via a smart use of adjoint
models. Two matrix free algorithms are proposed, one serial and one parallel, to
compute approximations of observation impact matrix. These algorithms compute SVD-based low-rank approximations, which capture the most important
features of observation impact matrix. The accuracy of the generated approximations scales with the computational cost.
A second contribution of this work is to illustrate the use of observation impact in solving two practical applications: pruning observations of small impact,
19
40
40
10
30
5
20
0
10
−5
y
y
30
10
20
0
10
−10
0
0
−10
20
x
0
0
40
(a) 4D-Var increment
20
x
40
(b) Supersensitivity field
40
35
0
30
−10
25
−20
20
15
−30
10
−40
5
0
0
−50
10
20
30
40
(c) Sensitivity to observations
Figure 9: Observation sensitivity field when the assimilated data is corrupted at
two locations with coordinates (10,10) and (20,20). The location of the faulty
sensors is unknown to the data assimilation system, but is retrieved via the
observation impact methodology.
20
and detecting faulty sensors. Numerical experiments that validate the proposed
computational framework are carried out with a two dimensional shallow water
equations system.
Several future research directions emerge from this study. On the computational side, the impact approximation algorithms can be further developed to
achieve superior performance. A rigorous analysis of approximation errors will
be developed to guide the choice of the number of iterations and the truncation
level. On the application side, our will be integrated in real large scale applications to provide a measure of importance of different measurements in real time
assimilation. In hindsight, the impact can be used to prune the data subsets and
to detect erroneous data points. In foresight, it can be used to design efficient
strategies of sensor placement for targeted and adaptive observations.
Acknowledgements
This work was supported by the National Science Foundation through the
awards NSF DMS-0915047, NSF CCF-0635194, NSF CCF-0916493 and NSF
OCI-0904397, and by AFOSR DDDAS program through the awards FA9550–
12–1–0293–DEF and AFOSR 12-2640-06.
References
[1] T. Palmer, T. Gelaro, J. Barkmeijer, and R. Buizza. Singular vectors,
metrics, and adaptive observations. Journal of the Atmospheric Sciences,
55(4):633–653, 1998.
[2] W. Liao, A. Sandu, and G. Carmichael. Total energy singular vector analysis for atmospheric chemical transport models. Monthly Weather Review,
134(9):2443–2465, 2006.
[3] D. Zupanski, A.Y. Hou, S.Q. Zhang, M. Zupanski, C.D. Kummerow, and
S.H. Cheung. Applications of information theory in ensemble data assimilation. Quarterly Journal of the Royal Meteorological Society, 133(627):1533–
1545, 2007.
[4] K. Singh, A. Sandu, and M. Jardak. Information theoretic metrics to
characterize observations in variational data assimilation. Proceedings of
the International Conference on Computational Science, 9(1):1047–1055,
2012.
[5] L.M. Berliner, Z.Q. Lu, and C. Snyder. Statistical design for adaptive
weather observations. Journal of the Atmospheric Sciences, 56(15):2536–
2552, 1999.
[6] W. Kang and L. Xu. Optimal placement of mobile sensors for data assimilations. Tellus A, 64(0), 2012.
21
[7] R. Gelaro, Y. Zhu, and R.M. Errico. Examination of various-order adjointbased approximations of observation impact. Meteorologische Zeitschrift,
16(6):685–692, 2007.
[8] R. Gelaro and Y. Zhu. Examination of observation impacts derived from
observing system experiments (OSEs) and adjoint models. Tellus A,
61(2):179–193, 2009.
[9] Y. Tremolet. Computation of observation sensitivity and observation impact in incremental variational data assimilation. Tellus A, 60(5):964–978,
2008.
[10] F.X. Le Dimet and H.E. Ngodock. Sensitivity analysis in variational data
assimilation. Journal of the Meteorological Society of Japan, 75(1):245–255,
1997.
[11] D.N. Daescu. On the sensitivity equations of four-dimensional variational
(4D-Var) data assimilation. Monthly Weather Review, 136(8):3050–3065,
2008.
[12] D.N. Daescu and R. Todling. Adjoint sensitivity of the model forecast to
data assimilation system error covariance parameters. Quarterly Journal
of the Royal Meteorological Society, 136(653):2000–2012, 2010.
[13] D.G. Cacuci. Sensitivity theory for nonlinear systems. I: Nonlinear functional analysis approach. Journal of Mathematical Physics, 22:2794, 1981.
[14] Z. Wang, I.M. Navon, F.X. Le Dimet, and X. Zou. The second order adjoint
analysis: theory and applications. Meteorology and Atmospheric Physics,
50(3):3–20, 1992.
[15] A. Sandu and L. Zhang. Discrete second order adjoints in atmospheric chemical transport modeling. Journal of Computational Physics,
227(12):5949–5983, 2008.
[16] J. Ye. Generalized low rank approximations of matrices. Machine Learning,
61(1-3):167–191, 2005.
[17] M.W. Berry and R.D. Fierro. Low-rank orthogonal decompositions for
information retrieval applications. Numerical Linear Algebra with Applications, 3(4):301–327, 1996.
[18] I.S. Dhillon and D.S. Modha. Concept decompositions for large sparse text
data using clustering. Machine Learning, 42(1-2):143–175, 2001.
[19] R. Daley. Atmospheric data analysis. Cambridge University Press, Cambridge, 1991.
[20] E. Kalnay. Atmospheric modeling, data assimilation and predictability.
Cambridge University Press, Cambridge, 2002.
22
[21] A. Sandu, D.N. Daescu, G.R. Carmichael, and T. Chai. Adjoint sensitivity
analysis of regional air quality models. Journal of Computational Physics,
204(1):222–252, 2005.
[22] A. Cioaca, A. Sandu, and M. Alexe. Second-order adjoints for solving PDEconstrained optimization problems. Optimization Methods and Software,
27(4-5):625–653, 2011.
[23] A. Griewank. On automatic differentiation. Mathematical Programming:
recent developments and applications, 6:83–107, 1989-.
[24] I.Y. Gejadze, F.X. Le Dimet, and V. Shutyaev. On analysis error covariances in variational data assimilation. SIAM Journal on Scientific Computing, 30(4):1847–1874, 2008.
[25] A. Cioaca, A. Sandu, and E. de Sturler. Efficient methods for computing
observation impact in 4d-var data assimilation. Technical report, 2012.
[26] C. Zhu, R.H. Byrd, P. Lu, and J. Nocedal. Algorithm 778: L-BFGS-B:
Fortran subroutines for large-scale bound-constrained optimization. ACM
Transactions on Mathematical Software (TOMS), 23(4):550–560, 1997.
[27] C. Lanczos. An iteration method for the solution of the eigenvalue problem
of linear differential and integral operators. United States Government
Press Office, 1950.
[28] W.E. Arnoldi. The principle of minimized iterations in the solution of the
matrix eigenvalue problem. Quarterly of Applied Mathematics, 9(1):17–29,
1951.
[29] G.L.G. Sleijpen and H.A. Van der Vorst. A Jacobi–Davidson iteration
method for linear eigenvalue problems. SIAM Review, 42(2):267–293, 2000.
[30] D.R. Fokkema and M.B. van Gijzen. Short manual for the JDQZ-package,
1999.
[31] E. Liberty, F. Woolfe, P.G. Martinsson, V. Rokhlin, and M. Tygert. Randomized algorithms for the low-rank approximation of matrices. Proceedings of the National Academy of Sciences, 104(51):20167–20172, 2007.
[32] T.O. Lewis and T.G. Newman. Pseudoinverss of postive semidefeinite matrices. SIAM Journal on Applied Mathematics, 16(4):701–703, 1968.
[33] D.D. Houghton and A. Kasahara. Nonlinear shallow fluid flow over an
isolated ridge. Communications on Pure and Applied Mathematics, 21(1):1–
23, 1968.
[34] R. Liska and B. Wendroff. Composite schemes for conservation laws. SIAM
Journal of Numerical Analysis, 35(6):2250–2271, 1998.
23
[35] R. Giering. Tangent linear and adjoint model compiler. User Manual,
TAMC Version, 4, 1997.
[36] R. Giering and T. Kaminski. Recipes for adjoint code construction. ACM
Transactions on Mathematical Software, 24(4):437–474, 1998.
[37] K. Anastasiou and C.T. Chan. Solution of the 2d shallow water equations using the finite volume method on unstructured triangular meshes.
International Journal for Numerical Methods in Fluids, 24(11):1225–1245,
1997.
24
| 5cs.CE
|
arXiv:1403.1512v4 [cs.DS] 12 Sep 2016
Parameterized Complexity of the k-Arc Chinese
Postman Problem∗
Gregory Gutin
Mark Jones
Bin Sheng
Department of Computer Science
Royal Holloway, University of London
Egham, Surrey TW20 0EX, UK
September 13, 2016
Abstract
In the Mixed Chinese Postman Problem (MCPP), given an edgeweighted mixed graph G (G may have both edges and arcs), our aim
is to find a minimum weight closed walk traversing each edge and arc at
least once. The MCPP parameterized by the number of edges was known
to be fixed-parameter tractable using a simple argument. Solving an open
question of van Bevern et al., we prove that the MCPP parameterized by
the number of arcs is also fixed-parameter tractable. Our proof is more
involved and, in particular, uses a well-known result of Marx, O’Sullivan
and Razgon (2013) on the treewidth of torso graphs with respect to small
separators. We obtain a small cut analog of this result, and use it to
construct a tree decomposition which, despite not having bounded width,
has other properties allowing us to design a fixed-parameter algorithm.
1
Introduction
A mixed graph is a graph that may contain both edges and arcs (i.e., directed
edges). A mixed graph G is strongly connected if for each ordered pair x, y of
vertices in G there is a path from x to y that traverses each arc in its direction.
We provide further definitions and notation on (mainly) directed graphs in the
next section.
In this paper, we will study the following problem.
Mixed Chinese Postman Problem (MCPP)
Instance: A strongly connected mixed graph G = (V, E∪A), with vertex
set V , edge set E and arc set A; a weight function w : E ∪ A → N0 .
Output: A closed walk of G that traverses each edge and arc at least
once, of minimum weight.
∗ A prelimary version of this paper was accepted for publication in the Proceedings of ESA
2014.
1
1 INTRODUCTION
2
There is numerous literature on various algorithms and heuristics for MCPP;
for informative surveys, see [2, 5, 9, 14, 17]. We call the problem the Undirected Chinese Postman Problem (UCPP) when A = ∅, and the Directed Chinese Postman Problem (DCPP) when E = ∅. It is well-known
that UCPP is polynomial-time solvable [8] and so is DCPP [3, 6, 8], but MCPP
is NP-complete, even when G is planar with each vertex having total degree 3
and all edges and arcs having weight 1 [16]. It is therefore reasonable to believe
that MCPP may become easier the closer it gets to UCPP or DCPP.
Van Bevern et al. [2] considered two natural parameters for MCPP: the
number of edges and the number of arcs. They showed that MCPP is fixedparameter tractable (FPT) when parameterized by the number k of edges. That
is, MCPP can be solved in time f (k)nO(1) , where f is a function only depending
on k, and n is the number of vertices in G. For background and terminology
on parameterized complexity we refer the reader to the monographs [7, 11, 15].
Van Bevern et al.’s algorithm is as follows. Replace every undirected edge uv
→ or arc −
→ or the pair −
→ and −
→ (all arcs have the same
by either the arc −
uv
vu
uv
vu
weight as uv) and solve the resulting DCPP. Thus, the MCPP can be solved
in time O(3k n3 ), where n is the number of the number of vertices in G.
We describe a faster algorithm here. Replace every undirected edge uv by
→ and −
→ Now construct a network N from the resulting digraph D
the pair −
uv
vu.
as follows: the cost of every arc of G is the same as its weight, the cost of every
→ in D, which is not in G, is the weight the undirected edge xy of G, the
arc −
xy
→ and −
→ of arcs that
lower bound of every arc of G is 1, and for each pair −
uv
vu
replaced an undirected edge uv, we assign lower bound 0 to one of the edges and
1 to the other. All upper bounds are ∞. Find a minimum-cost circulation (i.e.,
a flow of value 0) in N . This will correspond to a closed walk in D in which
→ and −
→
all arcs of G are traversed at least once and at least one of the arcs −
uv
vu
corresponding to an undirected edge uv of G is traversed at least once (the arc
whose lower bound is 1 in N ). As there are 2k ways to assign lower bounds to
the pairs of arcs in N , we obtain a running time of O(2k n3 ).
Van Bevern et al. [2] and Sorge [18] left it as an open question whether the
MCPP is fixed-parameter tractable when parameterized by the number of arcs.
This is the parameterization we consider in this paper.
k-arc Chinese Postman Problem (k-arc CPP)
Instance: A strongly connected weighted mixed graph G = (V, E ∪ A),
with vertex set V , edge set E and arc set A; a weight function w :
E ∪ A → N0 .
Parameter: k = |A|.
Output: A closed walk of G that traverses each edge and arc at least
once, of minimum weight.
This parameterized problem is of practical interest, for example, if we view
the mixed graph as a network of streets in a city: while edges represent two-way
streets, arcs are for one-way streets. Many cities have a relatively small number
of one-way streets and so the number of arcs appears to be a good parameter
1 INTRODUCTION
3
for optimizing, say, police patrol in such cities [2].
We will assume for convenience that the input G of k-arc CPP is a simple
graph, i.e. there is at most one edge or one arc (but not both) between any pair
of vertices. The multigraph version of the problem may be reduced to the simple
graph version by subdividing arcs and edges. As the number of arcs and edges
is at most doubled by this reduction, this does not affect the parameterized
complexity of the problem.
We will show that k-arc CPP is fixed-parameter tractable. Our proof is
significantly more complicated than the ones described above for the MCPP
parameterized by the number of edges. For that problem, as we saw, we can
replace the undirected edges with arcs. However a similar approach for MCPP
parameterized by the number of arcs (replacing arcs with edges) does not work.
Instead, in FPT time, we reduce the problem to the Balanced Chinese Postman Problem (BCPP), in which there are no arcs, but instead a demand
function on the imbalance of the vertices is introduced (the parameter for the
BCPP is based on the values of the demand function). This reduction is only
the first step of our proof, as unfortunately the BCPP is still NP-hard, unlike
the DCPP.
The BCPP turns out to be polynomial time solvable as long as a certain
connectivity property holds. Solving the problem in general requires making
some guesses on the edges in certain small cuts in the graph. To keep the
running time fixed-parameter, we require a structure on the graph that allows
us to only consider a few edges from small cuts at a time. To acheive this, we
make use of a recent result of Marx, O’Sullivan and Razgon [13] on the treewidth
of torso graphs with respect to small separators.
Marx, O’Sullivan and Razgon [13] use the following notion of a graph torso.
Let G = (V, E) be a graph and S ⊆ V. The graph torso(G, S) has vertex set S
and vertices a, b ∈ S are connected by an edge ab if ab ∈ E or there is a path in
G connecting a and b whose internal vertices are not in S.
Marx et al. [13] show that for a number of graph separation problems, it is
possible to derive a graph closely related to a torso graph, which has the same
separators as the original input graph. The separation problem can then be
solved on this new graph, which has bounded treewidth. By contrast, we use
the torso graph as a tool to construct a tree decomposition of the original graph,
which does not have bounded width, but has enough other structural restrictions
to make a dynamic programming algorithm possible. So, our application of
Marx et al.’s result is quite different from its use in [13], and we believe it
may be used for designing fixed-parameter algorithms for other problems on
graphs. Note that Marx et al. are interested in small separators (i.e. sets of
vertices whose removal disconnects a graph), whereas we are interested in small
cuts (sets of edges whose removal disconnects a graph). We therefore prove an
analog of Marx et al.’s result for cuts.
The rest of the paper is organized as follows. The next section contains
further terminology and notation. In Section 3, we reduce k-arc CPP to
Balanced Chinese Postman Problem (BCPP). In Section 4, we introduce
and study two key notions that we use to solve BCPP: t-roads, which witness
2 FURTHER TERMINOLOGY AND NOTATION
4
a connectivity property of the graph that makes the BCPP easy to solve; and
small t-cuts, which witness the fact that a t-road does not exist. In Section 5,
we investigate a special tree decomposition of the input graph of BCPP. This
decomposition is used in a dynamic programming algorithm given in Section 6.
The last section contains some conclusions and open problems.
2
Further Terminology and Notation
For a positive integer p and an integer q, q < p, [q, p] will denote the set
{q, q + 1, . . . , p} and [p] the set [1, p]. To avoid confusion, we denote an edge
→
between two vertices u, v as uv, and an arc from u to v as −
uv.
Although we will shortly reduce the the k-arc CPP to a problem on undirected graphs, we will still be interested in directed graphs as a way of expressing
solutions. For example, a walk which is a solution to an instance of the k-arc
CPP can be represented by a directed multigraph, with one copy of an arc
uv for each time the walk passes from u to v. This motivates the following
definitions.
→ denotes the number of arcs of the form
For a mixed multigraph G, µG (−
uv)
−
→
uv in G, and µG (uv) denotes the number of edges of the form uv. For a mixed
multigraph G, let D be a directed multigraph derived from G by replacing
→ of G with multiple copies of −
→ (at least one), and replacing each
each arc −
uv
uv
→ and −
→ (such that there is at
edge uv in G with multiple copies of the arcs −
uv
vu
−
→
−
→
least one copy of uv or at least one copy of vu). Then we say D is a multi→ + µ (−
→
orientation of G. If D is a multi-orientation of G and µD (−
uv)
D vu) =
−
→
−
→
µG (uv) + µG (uv) + µG (vu) for each u, v ∈ V (i.e. D is derived from G by
keeping every arc of G and replacing every edge of G with a single arc), we say
D is an orientation of G. If D is an orientation of G and G is undirected, we say
that G is the undirected version of D. The underlying graph G of an undirected
multigraph H can be obtained from H by deleting all but one edge among all
edges between u and v for every pair u, v of vertices of H.
For a simple weighted graph G and a multi-orientation D of G, the weight
of D is the sum of the weights of all its arcs, where the weight of an arc in D is
the weight of the corresponding edge or arc in G.
−
For a directed multigraph D = (V, A) and v ∈ V , d+
D (v) and dD (v) denote
the out-degree and in-degree of v in D, respectively. Let t : V → Z be a function
and Vt+ = {u ∈ V, t(u) > 0}, Vt− = {u ∈ V, t(u) < 0}. We say that a vertex
−
u in D is t-balanced if d+
is t-balanced if
D (u) − dD (u) = t(u). We say that DP
every vertex is t-balanced. Note that if D is t-balanced then v∈V t(v) = 0.
−
We say that a vertex u in D is balanced if d+
D (u) = dD (u), and we say that D
is balanced if every vertex is balanced.
In directed multigraphs, all walks (in particular, paths and cycles) that we
consider are directed. A directed multigraph D is Eulerian if there is a closed
walk of D traversing every arc exactly once. It is well-known that a directed
multigraph D is Eulerian if and only if D is balanced and the undirected version
of D is connected [1].
3 REDUCTION TO BALANCED CPP
5
For an undirected graph G = (V, E), and two vertex sets X, Y ⊆ V (G),
an (X, Y )-separator ( (X, Y )-cut, respectively) is a set S ⊆ V \ (X ∪ Y ) (a set
S ⊆ E, respectively) such that there is no path between vertices in X and Y in
graph G − S. When X = {x} and Y = {y}, we speak of (x, y)-separators and
(x, y)-cuts.
Observe that the following is an equivalent formulation of the k-arc CPP.
k-arc Chinese Postman Problem (k-arc CPP)
Instance: A strongly connected mixed graph G = (V, E∪A), with vertex
set V , edge set E and arc set A; weight function w : E ∪ A → N0 .
Parameter: k = |A|.
Output: A directed multigraph D of minimum weight such that D is a
multi-orientation of G and D is Eulerian.
3
Reduction to Balanced CPP
Our first step is to reduce k-arc CPP to a problem on a graph without arcs.
Essentially, given a graph G = (V, E ∪ A) we will “guess” the number of times
each arc in A is traversed in an optimal solution. This then leaves us with
a problem on G′ = (V, E). Rather than trying to find an Eulerian multiorientation of G, we now try to find a multi-orientation of G′ in which the
imbalance between the in- and out-degrees of each vertex depends on the guesses
for the arcs in A incident with that vertex.
More formally, we will provide a Turing reduction to the following problem:
Balanced Chinese Postman Problem (BCPP)
Instance: An undirected graph G = (V, E); a weight
function
P
w : E → N0 ; aPdemand function t : V → Z such that v∈V t(v) = 0.
Parameter: p = v∈V + t(v).
t
Output: A minimum weight t-balanced multi-orientation D of G.
P
Henceforth, any demand function t : V → Z will be such that v∈V t(v) = 0.
Observe that when t(v) = 0 for all v ∈ V , BCPP is equivalent to UCPP.
BCPP was studied by Zaragoza Martı́nez [19] who proved that the problem
is NP-hard. We will reduce k-arc CPP to BCPP by guessing the number
of times each arc is traversed. In order to ensure a fixed-parameter aglorithm,
we need a bound (in terms of |A|) on the number of guesses. We will do this
by bounding the total number of times any arc can be traversed in an optimal
solution.
Lemma 1. Let G = (V, A ∪ E) be a mixed graph, and let k = |A|. Then for
any optimal
to k-arc CPP on G with minimal number of arcs, we
P solution D
→ ≤ k 2 /2 + 2k.
have that uv∈A
−
→
µD (−
uv)
→ : −
→ ∈ A and µ (−
→
Proof. Let A = A1 ∪ A2 where A1 = {−
uv
uv
D uv) ≥ 3} and
A2 = A \ A1 . Let |A1 | = p and |A2 | = k − p = q.
3 REDUCTION TO BALANCED CPP
6
→ ∈ A. Since D is balanced, we have that D has µ (−
→
Consider an arc −
uv
D uv)
−
→
arc-disjoint directed cycles, each containing exactly one copy of uv. We claim
that each such cycle must contain at least one copy of an arc in A2 . Indeed,
→ that does not contain any arc in A ,
otherwise, there is a cycle C containing −
uv
2
which means that C consists of arcs in A1 and arcs corresponding to (undirected)
edges in G. We may construct a directed multigraph D′ as follows: Remove from
D two copies of each arc in A1 that appears in C, and reverse the arcs in C
that correspond to undirected edges in G. Observe that D′ is Eulerian and is
also a multi-orientation of G, and so D′ is a solution with smaller weight than
D or an optimal solution with fewer arcs than D, contradicting the minimality
of D.
→ cycles contains at least one copy of an arc in A .
So each of the µD (−
uv)
2
−
→ ≤ 2q. Thus,
Observe that
D
has
at
most
2q
copies
of
arcs
in
A
,
and
so
µ
(
uv)
2
D
P
→ ≤ p · 2q + 2q ≤
→ + P−
→ = P−
µD (−
uv)
µD (−
uv)
→
we have −
→
µD (−
uv)
→
uv∈A
uv∈A
uv∈A
2
1
p+q 2
2 · ( 2 ) + 2k = k 2 /2 + 2k.
Now we may prove the following:
Lemma 2. Suppose that there exists an algorithm which finds the optimal solution to an instance of BCPP on (G′ , w′ , t′ ) with parameter p in time f (p)|V (G′ )|O(1) .
Then there exists an algorithm which finds the optimal solution to an instance
of k-arc CPP on (G = (V, A ∪ E), w) with parameter k, which runs in time
⌊k2 /2+2k⌋
· f (⌊k 2 /2 + 2k⌋) · |V |O(1) .
k
Thus, if BCPP is FPT then so is k-arc CPP.
Proof. Let (G = (V, A ∪ E), w) be an
of k-arc CPP, and let k = |A|.
P instance −
→ ≤ κ for any optimal solution
Let κ = ⌊k 2 /2 + 2k⌋. By Lemma 1, −
→
µ
(
uv)
D
uv∈A
D to k-arc CPP on (G, w) with minimal number of arcs.
Let G′ = (V, E)P
and let w′ be w restricted to E. Given a function φ :
→ ≤ κ, let t : V → [−κ, κ] be the function
A → [κ] such thatP −
→
φ(−
uv)
φ
uv∈A
P
−
→
−
→ for all v ∈ V . Observe that
such
that
t
(v)
=
−
→
φ(
uv)
−
−
→
φ(
vu)
φ
vu∈A
P
P uv∈A−
→
−
→
φ(uv) ≤ κ, and thus BCPP on (G′ , w′ , tφ ) has parameuv∈A
v∈V + tφ (v) ≤
tφ
ter pφ ≤ κ.
Observe that for any given solution Dφ to BCPP on (G′ , w′ , tφ ), Dφ is
−
→
−
tφ -balanced, thus d+
Dφ (u) − dDφ (u) = tφ (u). If we add φ(uv) copies of each
→ ∈ A to D , and denote the resulting graph with D, then graph D is
arc −
uv
φ
P
+
→ d− (v) =
balanced. For any vertex v ∈ V (G), d+
−
→
φ(−
vu),
D (v) = dDφ (v) +
vu∈A
D
P
P
−
→ thus d+ (v) − d− (v) = d+ (v) − d− (v) + −
−
→ −
d−
(v)
+
−
→
φ(
uv),
→
φ(
vu)
D
D
D
D
D
uv∈A
vu∈A
φ
φ
Pφ
−
→
−
→
φ(uv) = tφ (v) − tφ (v) = 0. So D is a connected balanced graph (and
uv∈A
thus Eulerian), which is also a multi-orientation
thus is a solution to
P of G, and
−
→
−
→
k-arc CPP on (G, w) with weight w′ (Dφ ) + −
→
φ(
uv)w(
uv).
uv∈A
Furthermore for any solution D to k-arc CPP on (G, w), we know that D
−
−
→
−
→
is balanced, so for any vertex v ∈ V (G), d+
D (v) = dD (v). Let φ(uv) = µD (uv)
−
→
for each uv ∈ A and letPDφ be D restricted to E. For any
vertex v ∈ V (G), we
−
→ d− (v) = d− (v)− P−
−
→
+
have d+
(v)
=
d
(v)−
−
→
φ(
vu),
→
Dφ
D
vu∈A
Dφ
D
uv∈A φ(uv), therefore,
P
P
→
→ = t (v). So D is a t -balanced
−
d+
−
→
φ(−
uv)−
−
→
φ(−
vu)
φ
φ
φ
Dφ (v)−dDφ (v) =
uv∈A
vu∈A
4 EXPRESSING CONNECTIVITY: T -ROADS AND T -CUTS
7
multi-orientation P
of G′ and thus a solution to BCPP on (G′ , w′ , tφ ) and D has
−
→
−
→
′
weight w (Dφ ) + −
→
uv∈A
φ(uv)w(uv).
q
There are at most k ways of choosing positive integers x1 , . . . , xk such that
Pi
P
i∈[k] xi ≤ q. Indeed, for each i ∈ [k] let yi =
j=1 xj . Then yi < yj for i < j
and yi ∈ [q] for all i, and for any such choice of y1 , . . . , yk there is corresponding
Pk
choice of x1 , . . . , xk satisfying
i=1 xi ≤ q. Therefore the number of valid
choices for x1 , . . . , xk is the number of ways of choosing y1 , . . . , yk , which is the
number of ways of choosing k elements
from a set of q elements.
κ
Therefore
there
are
at
most
choices
for a function φ : A → [κ] such that
k
P
→ ≤ κ. Each choice leads to an instance of BCPP with parameter
−
→
φ(−
uv)
uv∈A
at most κ. Therefore in time κk f (κ) · |V |O(1) we can find, for every valid choice
of φ, the optimal solution Dφ to BCPP on (G′ , w′ , tφ ).
P
→
−
→
It then remains to choose the function φ that minimizes w′ (Dφ )+ −
→
φ(−
uv)w(
uv),
uv∈A
−
→
−
→
and return the graph Dφ together with φ(uv) copies of each arc uv ∈ A.
Due to Lemma 2, we may now focus on BCPP.
4
Expressing Connectivity: t-roads and t-cuts
Although we will not need the result until later, now is a good time to prove a
bound for BCPP somewhat similar to that in Lemma 1.
P
Lemma 3. Let (G, w, t) be an instance of BCPP, with p = v∈V + t(v). Then
t
for any optimal solution D to BCPP on (G, w, t) with minimal number of arcs,
→ + µ (−
→
we have that µD (−
uv)
D vu) ≤ max{p, 2} for each edge uv in G.
→
→ + µ (−
uv)
Proof. Suppose that µD (−
D vu) > max{p, 2} for some edge uv in G.
→
→ ≥ 1, then by removing one copy of −
−
→
uv
vu)
Observe that if µD (uv) ≥ 1 and µD (−
−
→
and one copy of vu, we obtain a solution to BCPP on (G, w, t) with weight at
→ − 1 + µ (−
→ −1 >
most that of D but with fewer arcs. (Note that µ (−
uv)
vu)
D
D
2 − 2 = 0, and so we still have a solution). Therefore, we may assume that
→ > max{p, 2} and µ (−
→
µD (−
uv)
D vu) = 0.
→
We now show that there must exist a cycle in D containing a copy of −
uv.
Modify D by adding a new vertex x, with t(v) arcs from x to v for each
v ∈ Vt+ , and −t(v) arcs from v to x for each v ∈ Vt− . Let D∗ be the resulting
→
directed graph. Then observe that D∗ is balanced, and therefore D∗ has µD (−
uv)
−
→
arc-disjoint cycles, each containing exactly one copy of uv. At most p of these
→
uv
cycles can pass through x. Therefore there is at least one cycle containing −
which is a cycle in D.
So now let v = v1 , v2 , . . . , vl = u be a sequence of vertices such that
→
−−−→
µD (−
vi−
v−
i+1 ) ≥ 1 for each i ∈ [l − 1]. Replace one copy of each arc vi vi+1 with
−
−
−
→
−
→
a copy of vi+1 vi and remove 2 copies of uv. Observe that the resulting graph
covers every edge of G, and the imbalance of each vertex is the same as in D.
Therefore, we have a solution to BCPP on (G, w, t) with weight at most that
of D but with fewer arcs. This contradiction proves the lemma.
4 EXPRESSING CONNECTIVITY: T -ROADS AND T -CUTS
8
In order to solve the BCPP on a graph G, we first add copies of edges to
G to produce a multigraph H, and then find an orientation of H which is a
solution to the BCPP on G. Thus, H is the undirected version of a solution
to the BCPP on G. The lemma below gives a connectivity condition which
must be satisfied by any undirected version of a solution. Furthermore, any
multigraph that satisfies this condition has an orientation which is a solution to
BCPP, and such a solution can be found in polynomial time. We will then be
able to solve the BCPP on G by searching for the minimum weight graph H
that satisfies this condition.
Definition 1. Let H = (V, E) be an undirected multigraph and t a demand
function V → Z. A t-road is a directed multigraph T with vertex set V such
−
that for each vertex v ∈ V , d+
T (v) − dT (v) = t(v). We say H has a t-road T if
′
there is a subgraph H of H such that T is an orientation of H ′ .
For an instance (G, w, t) of the BCPP with parameter p, it may be useful to
think of a t-road as a set of p arc-disjoint paths from vertices in Vt+ to vertices in
Vt− , although a t-road does not necessarily have to have such a simple structure.
Indeed, a t-road may also contain several closed walks. In particular, we note
that any solution to the BCPP on (G, w, t) is itself a t-road.
The following lemma and corollary show the relevance of t-roads to the
BCPP. Formally an input of BCPP is a simple graph, but to show Corollary
1 we will abuse this formality and allow multigraphs.
Lemma 4. Let H be an undirected multigraph and let (H, w, t) be an instance
of the BCPP. Then (H, w, t) has a solution which is an orientation of H (which
is necessarily an optimal solution) if and only if H has a t-road and for every
vertex v of H, dH (v) − t(v) is even. Furthermore, such a solution can be found
in polynomial time.
Proof. Suppose first that (H, w, t) has a solution of weight w(H). Then there
is a directed multigraph D with vertex set V (H) such that D is an orientation
−
of H, and d+
D (v) − dD (v) = t(v) for every vertex v ∈ V (H). Thus, D itself is
a t-road which is an orientation of a subgraph of H, and so H has a t-road.
−
Furthermore, for every vertex v of H, dH (v) − t(v) = d+
D (v) + dD (v) − t(v) =
−
−
−
+
dD (v) − dD (v) − t(v) + 2dD (v) = 2dD (v), which is even.
Conversely, suppose that H has a t-road and for every vertex v of H, dH (v)−
t(v) is even. Let T be a t-road in H. Delete the edges corresponding to T
from H, and observe that in the remaining graph every vertex v has degree
−
+
−
−
−
dH (v)−d+
T (v)−dT (v) = dH (v)−dT (v)+dT (v)−2dT (v) = dH (v)−t(v)−2dT (v),
which is even. Thus in this remaining graph every vertex is of even degree, and
so we may decompose the remaining edges into cycles. Orient each of these
cycles arbitrarily, and finally add the arcs of T . Let D be the resulting digraph.
−
+
−
Then for each vetex v ∈ V (H), d+
D (v) − dD (v) = 0 + dT (v) − dT (v) = t(v). Thus
D is t-balanced and is an orientation of H, as required.
By letting H be the undirected version of an optimal solution to an instance
(G, w, t), we get the following corollary.
4 EXPRESSING CONNECTIVITY: T -ROADS AND T -CUTS
9
Corollary 1. Given an instance (G, w, t) of the BCPP, let H be an undirected
multigraph of minimum weight, such that the underlying graph of H is G, H
has a t-road, and dH (v) − t(v) is even for every vertex v. Then there exists an
optimal solution to (G, w, t) which is an orientation of H, which can be found
in polynomial time.
Suppose that G has a t-road. Then by the above corollary, it is enough
to find a minimum weight multigraph H with underlying graph G, such that
dH (v) − t(v) is even for every vertex v. This can be done in polynomial time as
follows.
Given a graph G = (V, E) and set X ⊆ V of vertices, an X-join is a set
J ⊆ E such that |J ∪ E(v)| is odd if and only if v ∈ X, where E(v) is the set of
edges incident to v. Let X be the set of vertices such that dH (v) − t(v) is odd.
Note that if J is an X-join of minimum weight, the mutigraph H = (V, E ∪J) is
a minimum weight multigraph with underlying graph G, such that dH (v) − t(v)
is even for every vertex v.
Thus, to solve the BCPP on (G, w, t), where G has a t-road, it is enough
to find a minimum weight X-join. This problem is known as the Minimum
Weight X-Join Problem (traditionally, it is called the Minimum Weight
T -Join Problem, but we use T for t-roads) and can be solved in polynomial
time:
Lemma 5. [8] The Minimum Weight X-Join Problem can be solved in time
O(n3 ).
Let us briefly recall the proof of Lemma 5. Create a graph with vertex set
X. For any two vertices u, v ∈ X, create an edge uv of weight equal to the
minimum weight of a path between u and v in G. Find the minimum weight
perfect matching in this graph. Then the weight of this matching is the weight
of an X-join, and an X-join can be found by taking the paths corresponding to
edges in the matching.
The above remark shows that if G has a t-road, then we can solve the
BCPP in polynomial time. In general, G may not have a t-road. However,
given a solution D to the BCPP on (G, w, t), the undirected version of D must
have a t-road (indeed, D itself is a t-road). Therefore if we can correctly guess
the part of a solution corresponding to a t-road, and amend G using this partial
solution, the rest of the problem becomes easy. The following definition and
lemmas allow us to restrict such a guess to the places where there are small cuts
that prevent a t-road from existing.
Definition 2. Let H = (V, P
E(H)) be an undirected multigraph
and t : V → Z a
P
demand function such that v∈V t(v) = 0. Let p = v∈V + t(v). Then a small
t
t-cut of H is a minimal (Vt+ , Vt− )-cut F such that |F | < p.
A t-road in H, if one exists, can be found in polynomial time by computing
a flow of value p from a to b in the unit capacity network N with underlying
multigraph H ∗ , where H ∗ is the multigraph derived from H by creating two
4 EXPRESSING CONNECTIVITY: T -ROADS AND T -CUTS
10
new vertices a, b, with t(v) edges between a and v for each v ∈ Vt+ , and −t(v)
edges between b and v for each v ∈ Vt− . The next lemma follows from the
well-known max-flow-min-cut theorem for N .
Lemma 6. An undirected multigraph H has a t-road if and only if H does not
have a small t-cut.
The next lemma shows that if we want to decide where to duplicate edges
to get a t-road, we can restrict our attention to the edges in small t-cuts.
Definition 3. Let G = (V, E) be an undirected graph and t : V → Z a demand
function. Let F (G) be the union of all small t-cuts in G. Then a directed
→ = 0 for all uv ∈
→ +
multigraph T is well-(G, t)-behaved if µT (−
uv)
/ E and µT (−
uv)
−
→
µT (vu) ≤ 1 for all uv ∈ E \ F (G).
Lemma 7. Let D be an optimal solution to BCPP on (G = (V (G), E(G)), w, t),
and let H be the underlying graph of D. Then H has a well-(G, t)-behaved troad.
Proof. Let F (G) ⊆ E(G) be the union of all small t-cuts in G. Let J be the
undirected multigraph derived from H by removing all but one copy of every
edge in E(G) \ F (G). Note that V (G) = V (H) = V (J), E(G) ⊆ E(J) ⊆ E(H),
moreover, H and J have the same weight function w as G. Observe that every
(well-(G, t)-behaved) t-road in J is also a (well-(G, t)-behaved) t-road in H and
every t-road in J is well-(G, t)-behaved. So, it is sufficient to show that J has a
t-road.
Note that if J does not have a t-road, then by Lemma 6, J has a small t-cut.
Note also that by construction, D is a t-road, thus H has a t-road and therefore
does not have a small t-cut. Consider a small t-cut S in J and suppose that
every edge in S is a copy of an edge in F (G). As J contains the same number
of copies of each edge in F (G) as H does, it follows that any path from u ∈ Vt+
to v ∈ Vt− in H \ S is also a path from u to v in J \ S. But as H does not have
a small t-cut, such a path must exist, contradicting the assumption that S is a
small t-cut in J. Therefore every small t-cut in J contains a copy of an edge
not in F (G). If J has a small t-cut, then as G is a subgraph of J, every small
t-cut in J is also a small t-cut in G, it follows that there is a small t-cut in G
containing edges not in F . This is a contradiction by definition of F . Therefore
we may conclude that J does not have a small t-cut, and so J has a t-road, as
required.
If |F (G)|, the number of edges of G in small t-cuts, is bounded by a function
on p then, using Lemma 3 and Lemma 7 we can solve BCPP in FPT time
by guessing the multiplicities of each edge in F for an optimal solution D.
Unfortunately, |F (G)| may be larger than any function of p in general. It is also
possible to solve the problem on graphs of bounded treewidth using dynamic
programming techniques, but in general the treewidth may be unbounded. In
Section 5 we give a tree decomposition of G in which the number of edges from
F (G) in each bag is bounded by a function of p. This allows us to combine both
5 TREE DECOMPOSITION
11
techniques. In Section 6 we give a dynamic programming algorithm utilizing
Lemma 7 that runs in FPT time.
5
Tree Decomposition
Definition 4. Given an undirected graph G = (V, E), a tree decomposition of
G is a pair (T , β), where T is a tree and β : V (T ) → 2V such that
S
1. x∈V (T ) β(x) = V ;
2. for each edge uv ∈ E, there exists a node x ∈ V (T ) such that u, v ∈ β(x);
and
3. for each v ∈ V , the set β −1 (v) of nodes form a connected subgraph in T .
The width of (T , β) is maxx∈V (T ) (|β(x)| − 1). The treewidth of G (denoted
tw(G)) is the minimum width of all tree decompositions of G.
In this section, we provide a tree decomposition of G which we will use for
our dynamic programming algorithm. The tree decomposition does not have
bounded treewidth (i.e. the bags do not have bounded size), but the intersection
between bags is small, and each bag has a bounded number of vertices from small
t-cuts. This will turn out to be enough to develop a fixed-parameter algorithm,
as in some sense the hardness of BCPP comes from the small t-cuts.
Our tree decomposition is based on a result by Marx, O’Sullivan and Razgon
[13], in which they show that the minimal small separators of a graph “live in
a part of the graph that has bounded treewidth”[13].
Definition 5. Let G be a graph and C ⊆ V (G). The graph torso(G,C) has
vertex set C and vertices a, b ∈ C are connected by an edge if ab ∈ E(G) or
there is a path P in G connecting a and b whose internal vertices are not in C.
Lemma 8. [13, Lemma 2.11] Let a, b be vertices of a graph G = (V, E) and
let l be the minimum size of an (a, b)-separator. For some e ≥ 0, let S be the
union of all minimal (a, b)-separators of size at most l + e. Then there is an
f (l, e) · (|E| + |V |) time algorithm that returns a set S ′ ⊇ S disjoint from {a, b}
such that tw(torso(G, S ′ )) ≤ g(l, e), for some functions f and g depending only
on l and e.
Marx et al.’s result concerns small separators (i.e. sets of vertices whose
removal disconnects a graph), whereas we are interested in small cuts (sets of
edges whose removal disconnects a graph). For this reason, we prove the “edge
version” of Marx et al.’s result, which follows directly from their version.
Lemma 9. Let a, b be vertices of a graph G = (V, E) and let l be the minimum
size of an (a, b)-cut. For some e ≥ 0, let D be the union of all minimal (a, b)cuts of size at most l + e, and let C = V (D) \ {a, b}. Then there is an f (l, e) ·
(|E| + |V |) time algorithm that returns a set C ′ ⊇ C disjoint from {a, b} such
that tw(torso(G, C ′ )) ≤ g(l, e), for some functions f and g depending only on l
and e.
5 TREE DECOMPOSITION
12
Proof. We may assume that ab ∈
/ E, as otherwise all minimal (a, b)-cuts must
contain ab and deleting ab from G will not change the set C as it is disjoint
from {a, b}.
The main idea is to augment G to produce a graph G∗ such that every vertex
in C is part of a minimal (a, b)-separator in G∗ . We then apply Lemma 8 to
get a set S ′ ⊇ S and tree decomposition of torso(G∗ , S ′ ) of bounded width, and
then use this to produce a set C ′ ⊇ C and tree decomposition of torso(G, C ′ )
of bounded width.
We first produce the graph G∗ by subdividing each edge f in G with a new
vertex vf . Let S be the union of all minimal (a, b)-separators in G∗ of size at
most l + e. Let l′ be the minimum size of an (a, b)-separator in G∗ (note that l′
may be different from l).
Observe that for any minimal (a, b)-cut F of size at most l + e in G, the
set {vf : f ∈ F } is a minimal (a, b)-separator in G∗ . This implies that l′ ≤ l.
Furthermore, given any edge f ′ = uv such that f ′ ∈ F , assuming u ∈
/ {a, b},
the set ({vf : f ∈ F } \ {vf ′ }) ∪ {u} is an (a, b)-separator in G∗ and {vf : f ∈
F } \ {vf ′ } is not. So, X ∪ {u} is a minimal (a, b)-separator in G∗ for some
X ⊆ {vf : f ∈ F } \ {vf ′ }. Therefore, u is in a minimal (a, b)-separator in
G∗ of size less than l + e. A similar argument holds for v. It follows that
{u, v} \ {a, b} ⊆ S for any edge uv ∈ D, and so C ⊆ S.
Let e′ = (l − l′ ) + e, so we have l′ + e′ = l + e. Now apply Lemma 8 to get
a set S ′ disjoint from {a, b} such that S ⊆ S ′ , and a tree decomposition (T , β ′ )
of torso(G∗ , S ′ ) with treewidth at most g(l′ , e′ ). As l′ ≤ l and e′ ≤ l + e, this
treewidth is bounded by a function depending only on l and e.
Define a function h : S ′ → V (G) as follows. For each edge f ∈ E(G) such
that vf ∈ S ′ , if a or b is an endpoint of f , set h(vf ) to be the other endpoint,
and otherwise let h(vf ) be an arbitrary endpoint of f . For every other v ∈ S,
let h(v) = v. Now let C ′ = {h(v) : v ∈ S ′ }. Observe that C ⊆ S ′ ∩ V (G) ⊆ C ′ .
We produce a tree decomposition of torso(G, C ′ ) as follows. Given the
tree decomposition (T , β ′ ) of torso(G∗ , S ′ ), define β : V (T ) → C ′ by β(x) =
h(β ′ (x)) = {h(v) : v ∈ β ′ (x)}. We now show that (T , β) is indeed a tree
decomposition of torso(G, C ′ ).
S
It follows from construction that x∈V (T ) β(x) = C ′ = V (torso(G, C ′ )).
Now consider an edge uw in torso(G, C ′ ). We will show that there is an
edge st in torso(G∗ , S ′ ) with h(s) = u, h(t) = w. It follows that s, t ∈ β ′ (x) for
some node x ∈ V (T ), and consequently u, w ∈ β(x) for the same node x. This
satisfies the second condition of the tree decomposition.
As u, w are adjacent in torso(G, C ′ ), there must be a path between them
which has no internal vertices in C ′ . By subdividing each edge f in this path
with the vertex vf , we get a path P between u and w in G∗ which has no internal
vertices in C ′ . Suppose P contains an internal vertex v with v ∈ S ′ . Observe
that P must also contain h(v) (if h(v) = v then this is obvious, and otherwise v
has only two neighbours, both of which must be in P and one of which is h(v)).
If h(v) 6= u and h(v) 6= w, then h(v) is also an internal vertex of P , and P has
an internal vertex in C ′ , a contradiction. Therefore the only internal vertices v
5 TREE DECOMPOSITION
13
of P which are in S ′ are those for which h(v) = u or h(v) = w.
If P does not have any vertices in h−1 (u) (which may happen if u ∈
/ S ′ ), then
u must have a neighbour vf with h(vf ) = u. Then by adding such a neighbour
to P , we may assume that P contains at least one vertex in h−1 (u). Similarly
we may assume P contains at least one vertex in h−1 (w). By considering the
shortest subpath of P containing vertices in both h−1 (u) and h−1 (w), we have
that there is a path in G∗ with endpoints s, t ∈ S ′ , with no internal vertices in S ′ ,
such that h(s) = u, h(t) = w. It follows that s, t are adjacent in torso(G∗ , S ′ ).
Now consider β −1 (u) for some vertex u ∈ C ′ . SWe wish to show that β −1 (u)
forms a connected subgraph in T . As β −1 (u) = {β ′−1 (v) : v ∈ h−1 (u)}, each
β ′−1 (v) forms a connected subgraph in T , and β ′−1 (v1 ) ∩ β ′−1 (v2 ) 6= ∅ for adjacent v1 , v2 in torso(G∗ , S ′ ), it will be sufficient to show that h−1 (u) induces a
connected subgraph in torso(G∗ , S ′ ). If u ∈ S ′ , then all vertices in h−1 (u) \ {u}
are adjacent to u in torso(G∗ , S ∗ ), and therefore h−1 (u) induces a graph that
contains a star rooted at u as a subgraph. On the other hand if u ∈
/ S ′ , then for
−1
∗
any v1 , v2 ∈ h (u), there is a path v1 uv2 in G , which contains no internal vertices in S ′ , and so v1 , v2 are adjacent in torso(G∗ , S ′ ). Therefore h−1 (u) induces
a clique in torso(G∗ , S ′ ). In either case, h−1 (u) induces a connected subgraph
in torso(G∗ , S ′ ). This satisfies the third condition of the tree decomposition,
which completes the proof that (T , β) is a tree decomposition of torso(G, C ′ ).
Finally, note that by construction maxx∈V (T ) (|β(x)|−1) ≤ maxx∈V (T ) (|β ′ (x)|−
1), and so (T , β) has width at most g(l′ , e′ ), which as previously discussed is
bounded by a function depending only on l and e.
It remains to analyse the running time. Construction of G∗ can be done
in linear time as we need to process each edge of G once. G∗ has 2|E(G)|
edges and |V (G)| + |E(G)| vertices, and therefore the algorithm of Lemma 8
takes time f (l′ , e′ ) · (|E(G∗ )| + |V (G∗ )|) ≤ f (l, l + e) · (3|E(G)| + |V (G)|) ≤
3f (l, l + e) · (|E(G)| + |V (G)|). Finally, transforming the decomposition (T , β ′ )
into (T , β) takes time O(|V (T )| · maxx∈V (T ) |β(x)| = O(|V (T )| · g(l, l + e), and
we may assume |V (T )| is linear in |E(G)| + |V (G)| as it took linear time to
construct. Therefore the total running time is linear in |E(G)| + |V (G)|.
We will now use the treewidth result on torso graphs to construct a tree
decomposition of the original graph, in which the width may not be bounded,
but the intersection between bags and the number of edges in small cuts in each
bag is bounded by a function of p. In order to make our dynamic programming
simpler, it is useful to place further restrictions on the structure of a tree decomposition. The notion of a nice tree decomposition is often used in dynamic
programming, as it can impose a simple structure and can be found whenever
we have a tree decomposition.
Definition 6. Given an undirected graph G = (V, E), a nice tree decomposition
(T , β) is a tree decomposition such that T is a rooted tree, and each of the nodes
x ∈ V (T ) falls under one of the following classes:
• x is a Leaf node: Then x has no children in T ;
14
5 TREE DECOMPOSITION
• x is an Introduce node: Then x has a single child y in T , and there
exists a vertex v ∈
/ β(y) such that β(x) = β(y) ∪ {v};
• x is a Forget node: Then x has a single child y in T , and there exists
a vertex v ∈ β(y) such that β(x) = β(y) \ {v};
• x is a Join node: Then x has two children y and z, and β(x) = β(y) =
β(z).
(Note that sometimes it is also required that |β(x)| = 1 for every leaf node
x, but for our purposes we allow β(x) to be undbounded.)
It is well-known that given a tree decomposition of a graph, it can be transformed into a nice tree decomposition of the same width in polynomial time
[12].
Lemma 10 (Lemma 13.1.3, [12]). For constant k, given a tree decomposition
of a graph G of width k and O(n) nodes, where n is the number of vertices of
G, one can find a nice tree decomposition of G of width k and with at most 4n
nodes in O(n) time.
It is also known that a tree decomposition of a graph can be found in fixedparameter time.
Lemma 11. [4] There exists an algorithm that, given an n-vertex graph G and
3
integer k, runs in time k O(k ) · n and either constructs a tree decomposition of
G of width at most k, or concludes that G has treewidth greater than k.
3
Observe that as the running time in Lemma 11 is k O(k ) · n, we may assume
3
the tree decomposition has at most k O(k ) · n nodes. Then applying Lemma
10, we have that for any graph G with treewdith k, we can find a nice tree
decomposition of G with at most 4|V (G)| nodes in time fixed-parameter with
respect to k.
Our tree decomposition will be similar but not identical to a nice tree decomposition. We are now ready to give our tree decomposition, which is the
main result of this section.
We believe this lemma may be useful for other problems in which the “difficult” parts of a graph are the small cuts or separators.
Lemma 12. Let a, b be vertices of a graph G = (V, E) and let l be the minimum size of an (a, b)-cut (respectively, let l be the minimum size of an (a, b)separator). For some e ≥ 0, let D be the union of all minimal (a, b)-cuts of size
at most l + e, and let C = V (D) \ (a, b) (respectively, let C be the union of all
minimal (a, b)-separators of size at most l + e).
Then there is an f (l, e) · (|E| + |V |) time algorithm that returns a set C ′
disjoint from {a, b} and a (binary) tree decomposition (T , β) of G such that:
1. C ⊆ C ′ ;
5 TREE DECOMPOSITION
15
2. β(x) ⊆ C ′ for any node x in T which is not a leaf node (in particular, the
intersection between any two bags of adjacent nodes of T is contained in
C ′ );
3. For any node x in T , |β(x) ∩ C ′ | ≤ g(l, e);
4. (T , β) restricted to C ′ (i.e. (T , β ′ ), where β ′ (x) = β(x) ∩ C ′ ) is a nice
tree decomposition;
for some functions f and g depending only on l and e.
Proof. If C is the union of all vertices appearing in the set D of all minimal (a, b)cuts of size at most l + e, then apply Lemma 9. If C is the union of all minimal
(a, b)-separators of size at most l + e, then apply Lemma 8. In either case, we
get a set C ′ ⊇ C disjoint from {a, b} such that tw(torso(G, C ′ )) ≤ g(l, e), for a
function g depending only on l and e. From here on the proof is identical for
the two cases.
Using Lemmas 10 and 11, we may find a nice tree decomposition of torso(G, C ′ )
of width at most g(l, e) in time f (g(l, e)) · (|E| + |V |), for some function f depending only on g(l, e). Let (T ′ , β ′ ) be the resulting tree decomposition of
torso(G, C ′ ).
We now add the vertices of G which are not in C ′ to this decomposition.
Consider any component X of G − C ′ . Then N (X) ⊆ C ′ . Furthermore, by definition of torso(G, C ′ ), any pair of vertices in N (X) are adjacent in torso(G, C ′ ).
It is well-known that the vertices of a clique in a graph are fully contained in
a single bag in any tree decomposition of the graph. Therefore, N (X) ⊆ β ′ (x)
for some node x in T .
If x is a Leaf node then modify β ′ (x) by adding X to it. Otherwise, modify
′
(T , β ′ ) by inserting (in the edge of T between x and its parent) a new Join node
y as the parent of x, with another child node z of y, such that β ′ (y) = β ′ (x),
and β ′ (z) = β ′ (x) ∪ X. Thus, X is still added to a Leaf node.
Let (T , β) be the resulting tree decomposition. As every component
of G−C ′
S
′
was added to a bag in a tree decomposition of torso(G, C ), x∈V (T ) β(x) =
V (G). Every edge between vertices in C ′ is in a bag due to the tree decomposition of torso(G, C ′ ), and for every v ∈
/ C ′ , N (v) is contained in the same bag
as v. Therefore for every edge uv in G, u and v appear in the same bag. For
any vertex v, β −1 (v) consists of a single node if v ∈
/ C ′ , and otherwise β −1 (v)
is connected by the tree decomposition of torso(G, C ′ ). Thus, (T , β) is a tree
decomposition of G.
Furthermore, by construction β(x) ⊆ C ′ for every non-leaf node x, |β(x) ∩
′
C | ≤ g(l, e) for every node x, and (T , β) restricted to C ′ is a nice tree decomposition.
We now modify this approach slightly to get the desired tree decomposition
when C is the union of all edges in small t-cuts.
Lemma 13. Let (G = (V, E), w, t) be an instance of BCPP, let C be the
non-empty set of vertices appearing in edges in small t-cuts. Then there is
6 DYNAMIC PROGRAMMING
16
an f (p) · (|E| + |V |) time algorithm that returns a set C ′ and a (binary) tree
decomposition (T , β) of G such that:
1. C ⊆ C ′ ;
2. β(x) ⊆ C ′ for any node x in T which is not a leaf node (in particular, the
intersection between any two bags of adjacent nodes of T is contained in
C ′ );
3. For any node x in T , |β(x) ∩ C ′ | ≤ g(p);
4. (T , β) restricted to C ′ (i.e. (T , β ′ ), where β ′ (x) = β(x) ∩ S ′ ) is a nice tree
decomposition;
for some functions f and g depending only on p.
Proof. First construct the multigraph G∗ from G by creating two new vertices
a, b, with t(v) edges between a and v for each v ∈ Vt+ , and −t(v) edges between
b and v for each v ∈ Vt− . Then by definition, C is the set of vertices appearing
in an edge e ∈ E(G) such that e is part of a minimal (a, b)-cut in G∗ of size less
than p.
Now apply Lemma 12 to get a set C ′ disjoint from {a, b} such that C ⊆ C ′
and a tree decomposition of torso(G∗ , C ′ ) with treewidth at most g(l, e), where
l + e = p − 1 and so g(l, e) is bounded by a function of p. It follows from
the definition of a torso graph that torso(G∗ \ {a, b}, C ′) is a subgraph of
torso(G∗ , C ′ ) \ {a, b} [13, Lemma 2.6], and so we can get a tree decomposition (T , β) of torso(G, C ′ ) by removing a and b from every bag in the tree
decomposition of torso(G∗ , C ′ ). As a, b ∈
/ C ′ , the resulting tree decomposition
is still a nice tree decomposition when restricted to C ′ .
6
Dynamic Programming
Let (G, w, t) be an instance of BCPP. Let (T , β) be the tree decomposition of
G and C ′ the set of vertices containing all vertices of every small t-cut given by
Lemma 13. In this section we give a dynamic programming algorithm based on
this decomposition.
Before describing the algorithm, we give some notation that we use in this
section. Let α(x) = β(x) ∩ C ′ . Thus β(x) ∩ β(y) ⊆ α(x) for all nodes x 6= y,
and α(x) = β(x) for every non-leaf x. Furthermore, for any Join node x with
two children y and z, we have that α(x) = α(y) = α(z), even if one or both of
the children of x is a Leaf node whose bag contains vertices not in C ′ . Let γ(x)
be the union of the bags of all predecessors of x including x itself. Thus, if r is
the root node of T , then γ(r) = V (G).
Let h : V (G) → {odd, even} be the function such that h(v) = odd if t(v)
is odd and h(v) = even if t(v) is even. Observe that in the undirected version
of any solution to BCPP on (G, w, t), each vertex v will have odd degree if and
6 DYNAMIC PROGRAMMING
17
only if h(v) = odd. Thus, h and similar functions will be used to tell us whether
a vertex should have odd or even degree.
To simplify some expressions, we adopt the convention that odd + odd =
even, even + even = even, and odd + even = odd. We say a vertex v
is h-balanced if it has odd degree if and only if h(v) = odd. An undirected
multigraph H is h-balanced if every vertex is h-balanced.
We now give an outline of our algorithm.
By Corollary 1, in order to solve an instance (G, w, t) of the BCPP, it is
enough to find an undirected multigraph H of minimum weight, such that the
underlying graph of H is G, H has a t-road, and dH (v) − t(v) is even for every
vertex v. Our algorithm will therefore focus on solving this problem, rather than
finding an optimal multi-orientation of G directly. By Lemma 7, we may assume
H has a well-(G, t)-behaved t-road T . Our dynamic programming algorithm will
give a way to find H and T .
For each node x in T , our dynamic programming algorithm will calculate a
value ψ(x, H ′ , T ′ , t′ , h′ ), for a particular range of graphs H ′ and T ′ and functions
t′ and h′ . Informally, ψ(x, H ′ , T ′ , t′ , h′ ) denotes a potential solution of minimum
weight, restricted to γ(x). Let H denote this restricted solution and T its “troad”, similarly restricted to γ(x). The subgraph H ′ tells us how H should
look when restricted to α(x), and similarly T ′ tells us how T should look when
restricted to α(x). The function t′ tells us what the imbalance of each vertex
should be within T . Roughly speaking, it is the function for which T is a t′ -road.
(Note that T will not necessarily be a t-road itself, as it is only a restriction of a
potential t-road to γ(x).) In a similar way, the function h′ , which maps vertices
to either odd or even, tells us whether the degree of each vertex within H
should be odd or even. (We note that in a full solution, the parity of the degree
of each vertex v will be defined by h(v), but as H is only a partial solution it
may be that h′ (v) 6= h(v) for some v.)
More formally, let x be a node of T , let H ′ be an undirected multigraph
with underlying graph G[α(x)], such that µH ′ (uv) ≤ max{p, 2} for all edges uv.
→
→ + µ ′ (−
uv)
Let T ′ be a directed graph with vertex set α(x), such that µT ′ (−
T vu) ≤
′
µH ′ (uv) for all edges uv. Let t be a function α(x) → [−p, p] and let h′ be
a function α(x) → {odd, even}. Then let ψ(x, H ′ , T ′ , t′ , h′ ) be an undirected
multigraph H with underlying graph G[γ(x)], of minimum weight such that
1. H[α(x)] = H ′ .
2. H has a well-(G, t)-behaved t∗ -road T such that T restricted to α(x) is
T ′ , where t∗ : γ(x) → [−p, p] is the function such that t∗ (v) = t′ (v) for
v ∈ α(x) and t∗ (v) = t(v), otherwise.
3. H is h∗ -balanced, where h∗ : γ(x) → {odd, even} is the function such
that h∗ (v) = h′ (v) if v ∈ α(x) and h∗ (v) = h(v), otherwise.
If no such H exists, let ψ(x, H ′ , T ′ , t′ , h′ ) = null.
The following lemma shows that to solve the BCPP, it is enough to calculate
ψ(x, H ′ , T ′ , t′ , h′ ) for every choice of x, H ′ , T ′ , t′ , h′ .
6 DYNAMIC PROGRAMMING
18
Lemma 14. Let r be the root node of T . Let t′ be t restricted to α(r), and
let h′ be h restricted to α(r). Let H ′ and T ′ be chosen such that the weight of
H = ψ(r, H ′ , T ′ , t′ , h′ ) is minimized. Then the weight of H is the weight of an
optimal solution to the BCPP on (G, w, t), and given H we may construct an
optimal solution to BCPP on (G, w, t) in polynomial time.
Proof. Observe that by construction of t′ and h′ , t∗ and h∗ in the definition of
ψ(r, H ′ , T ′ , t′ , h′ ) are t and h, respectively. Also observe that for a graph H,
d∗H (v) − t(v) is even for each vertex v if and only if H is h-balanced.
Let D be an optimal solution to the BCPP on (G, w, t), and let H be the
undirected version of D. By Lemma 4, H is h-balanced and has a t-road.
Furthermore by Lemma 7, H has a well-(G, t)-behaved t-road T . By Lemma
3, we may assume that µH (uv) ≤ max{p, 2} for each edge uv. H clearly has
underlying graph G = γ(r). So by letting H ′ be H[α(r)] and letting T ′ be
T [α(r)], we have that H satisfies all the conditions of ψ(x, H ′ , T ′ , t′ , h′ ) (except
possibly for minimality).
On the other hand, suppose H satisfies all these conditions. Then in particular, H has underlying graph γ(r) = G, H is h-balanced, and H has a t-road.
It follows by Lemma 4 that there exists a solution to the BCPP on (G, w, t)
which is an orientation of H.
It follows that the minimum weight solution to the BCPP on (G, w, t) has
the same weight as H = ψ(r, H ′ , T ′ , t′ , h′ ), when H ′ and T ′ are chosen such that
the weight of H is minimized.
When using dynamic programming algorithms based on tree decompositions,
the most commonly used approach is to consider the restriction of possible
solutions to each bag, and combine information about the possible restrictions on
each bag to construct a full solution. However this approach only works when the
tree decomposition is of bounded width, as the number of restrictions to consider
on each bag is bounded. In our case, some of the bags in the decomposition
may be arbitrarily large, so we cannot consider all possible solutions on a bag.
However, we do have that each bag contains a bounded number of vertices from
C ′ , where C ′ contains all vertices that appear in edges in small t-cuts. It will
turn out to be enough to make a guess based on the edges between vertices in
C ′ , after which the rest of the problem can be solved efficiently.
Finally, we show how to calculate ψ(x, H ′ , T ′ , t′ , h′ ) for every choice of x, H ′ , T ′ , t′ , h′ .
Lemma 15. ψ(x, H ′ , T ′ , t′ , h′ ) can be calculated in FPT time, for all choices
of x, H ′ , T ′ , t′ , h′ .
Proof. Consider some node x, and assume that we have already calculated
ψ(y, H ′′ , T ′′ , t′′ , h′′ ), for all descendants y of x and all choices of H ′′ , T ′′ , t′′ , h′′ .
We consider the possible types of nodes separately.
x is a Leaf node: Consider the multigraph Gx with vertex set β(x) =
γ(x) such that Gx [α(x)] = H ′ , and Gx has exactly one copy of each edge
in G[β(x)] not contained in α(x). Note that Gx is necessarily a subgraph of
6 DYNAMIC PROGRAMMING
19
H = ψ(x, H ′ , T ′ , t′ , h′ ). Moreover, as every edge from G[β(x)] in a small tcut of G is contained in α(x), any well-(G, t)-behaved t∗ -road in H is also a
well-(G, t)-behaved t∗ -road in Gx .
It follows that if there exists any H satisfying the first two conditions of
ψ(x, H ′ , T ′ , t′ , h′ ) then Gx has a well-(G, t)-behaved t∗ -road. So we first check
whether Gx has a t∗ -road (we note that any t∗ -road in Gx is well-(G, t)-behaved
by construction of Gx ).
If Gx has a t∗ -road, it remains to find a minimum weight (multi)set of edges
to add to Gx to make it h∗ -balanced. This can be done by solving the Minimum
Weight X-Join Problem, where X is the set of all vertices in β(x) that are
not h∗ -balanced in Gx . By Lemma 5, this can be done in polynomial time.
x is an Introduce node:
Let y be the child node of x, and let v be the single vertex in β(x)\α(y). Then
no vertices in γ(x) are adjacent with v, except for those in α(x). In particular
for any H = ψ(x, H ′ , T ′ , t′ , h′ ), the only edges of H incident with v are those
in H ′ . Thus, if v is not h′ -balanced in H ′ , then ψ(x, H ′ , T ′ , t′ , h′ ) = null.
Similarly, if v is not t′ -balanced in T ′ , then ψ(x, H ′ , T ′ , t′ , h′ ) = null.
Otherwise, suppose that H = ψ(x, H ′ , T ′ , t′ , h′ ), and let H ∗ be H restricted
to γ(y). We now construct the values H ′′ , T ′′ , t′′ , h′′ for which H ∗ = ψ(y, H ′′ , T ′′ , t′′ , h′′ )
must hold. Observe that H ∗ [α(y)] = H[α(y)] = H ′ [α(y)]. Thus we will set
H ′′ = H ′ [α(y)]. For the well-(G, t)-behaved t′ -road T that H must have, let
T ∗ be T restricted to γ(y), and observe that T ∗ [α(y)] = T [α(y)] = T ′ [α(y)].
Thus we will set T ′′ = T ′ [α(y)]. As T ∗ is equal to T with the arcs incident to u removed, we have that the imbalance of a vertex u ∈ α(y) in T ∗
→
→ + µ ′ (−
uv)
is t′ (u) − µT ′ (−
T vu). (Note also that for u ∈ γ(y) \ α(u), the imbalance of u remains unchanged i.e. is still t(u).) Thus we let t′′ : α(y) → [−p, p]
→
→ + µ ′ (−
uv)
be the function such that t′′ (u) = t′ (u) − µT ′ (−
T vu). By a similar
argument, the parity of the degree of any vertex u ∈ α(y) in H ∗ is equal to
h′′ (u), where h′′ : α(y) → {odd, even} is such that if µH ′ (uv) is odd then
h′′ (u) = h′ (u) + odd, and otherwise h′′ (u) = h′ (u). Thus, we have that
H ∗ = ψ(y, H ′′ , T ′′ , t′′ , h′′ ) (where the minimality property follows from the fact
that any improvement on H ∗ would give a corresponding improvement on H),
Thus, we may set ψ(x, H ′ , T ′ , t′ , h′ ) to be ψ(y, H ′′ , T ′′ , t′′ , h′′ ) together with
the edges of H ′ incident with v.
x is a Forget node:
Let y be the child node of x, and let v be the single vertex in α(y) \ β(x).
Note that γ(y) = γ(x). Let t′′ : α(y) → [−p, p] be the function t∗ restricted to
α(y) (i.e. t′′ extends t′ and assigns v to t(v)). Let h′′ : α(y) → {odd, even}
be the function h∗ restricted to α(y) (i.e. extends h′ and assings v to h(v)). If
H = ψ(x, H ′ , T ′ , t′ , h′ ) then by construction of t′′ and h′′ , we also have H =
ψ(y, H ′′ , T ′′ , t′′ , h′′ ), for some values of H ′′ and T ′′ . Note that the possible
values of H ′′ are those for which H ′′ [α(x)] = H ′ , and so the only choice is the
multiplicity of each edge in H ′′ incident with v. By Lemma 3 we may assume
the multiplicity of each such edge is at most p, and therefore we have at most
(p + 1)|α(x)| ≤ (p + 1)g(p) possible values of H ′′ . Similarly, we have at most
(p + 1)|α(x)| ≤ (p + 1)2g(p) possible values of T ′′ .
6 DYNAMIC PROGRAMMING
20
We therefore may set ψ(x, H ′ , T ′ , t′ , h′ ) to be the minimum weight ψ(y, H ′′ , T ′′ , t′′ , h′′ )
over all possible values of H ′′ and T ′′ .
x is a Join node:
Let y and z be the children of x, and recall that α(x) = α(y) = α(z), and
furthermore γ(x) = γ(y) ∪ γ(z) and γ(y) ∩ γ(z) = α(x).
Suppose that H = ψ(x, H ′ , T ′ , t′ , h′ ), and consider the graphs Hy = H[γ(y)]
and Hz = H[γ(z)].
Observe that the weight of H is equal to w(Hy ) + w(Hz ) − w(H ′ ) (as the
only edges contained in both Hy and Hz are those in H ′ ). Since α(x) = α(y),
it must be the case that Hy = ψ(y, H ′ , T ′ , t′′ , h′′ ) for some choice of t′′ and h′′ .
Similarly Hz = ψ(z, H ′ , T ′ , t′′′ , h′′′ ) for some choice of t′′′ and h′′′ . It remains to
determine the possible choices of t′′ , h′′ , t′′′ , h′′′ .
Consider a well-(G, t)-behaved directed multigraph T , and let Ty = T [γ(y)]
and Tz = T [γ(z)]. Let t∗ ′′ be the function such that Ty is a t∗ ′′ -road, and let t∗ ′′′
be the function such that Tz is a P
t∗ ′′′ -road. For any vP∈ α(x), the imbalance of v
−
→
→
∗ ′′′
∗ ′′
vu)+
in T is equal to t (v)+t (v)− u∈α(x) µT ′ (−
u∈α(x) µT ′ (uv) (where the
last two terms come from the fact that arcs in T ′ are counted twice in t∗ ′′ (v) +
∗ ′′′
(v)). Thus, v isPt∗ -balanced in T if and only if t∗ (v) = t∗ ′′ (v) + t∗ ′′′ (v) −
tP
−
→
−
→
u∈α(x) µT ′ (uv). We also note that for v ∈ γ(y) \ α(x), v is
u∈α(x) µT ′ (vu) +
t∗ -balanced in T if and only if t∗ ′′ (v) = t∗ (v) = t(v), and for v ∈ γ(z) \ α(x), v
is t∗ -balanced in T if and only if t∗ ′′′ (v) = t∗ (v) = t(v).
Let h∗ ′′ : γ(y) → {odd, even} be the function such that Hy is h∗ ′′ -balanced,
and let h∗ ′′′ : γ(z) → {odd, even} be the function such that Hz is h∗ ′′′ -balanced
Then by a similar argument, a vertex v ∈ α(x) is h∗ -balanced in H if and only
if h∗ (v) = h∗ ′′ (v) + h∗ ′′′ (v) + c, where c = even if v has even degree in H, and
c = odd otherwise.
The above implies that ψ(x, H ′ , T ′ , t′ , h′ ) is the union of ψ(y, H ′ , T ′ , t′′ , h′′ )
and ψ(z, H ′ , T ′ , t′′′ , h′′ ), where t′′ , h′′ , t′′′ , h′′′ are chosen to minimize the the
total weight of ψ(y, H ′ , T ′ , t′′ , h′′ ) and ψ(z, H ′ , T ′ , t′′′ , h′′ ) and such that
1. t′′ (v), t′′′ (v) ∈ [−p, p] for all v ∈ α(x);
P
→ P
−
→
2. t′ (v) = t′′ (v)+t′′′ (v)− u∈α(x) µT ′ (−
vu)+
u∈α(x) µT ′ (uv) for all v ∈ α(v);
3. h′ (v) = h′′ (v) + h′′′ (v) if v has even degree in H, and h′ (v) = odd +
h′′ (v) + h′′′ (v) otherwise.
Observe that in the case of a Join node, there is only one possible choice of
t′′′ for each choice of t′′ and only one possible choice of h′′′ for each choice of
h′′ . Therefore there are at most [2(2p + 1)]g(p) possible choices for t′′ , t′′′ , h′′ , h′′′ .
Therefore it is possible to calculate ψ(x, H ′ , T ′ , t′ , h′ ) in fixed-parameter time,
as long as we have already calculated ψ(y, H ′′ , T ′′ , t′′ , h′′ ), for all descendants y
of x and all choices of H ′′ , T ′′ , t′′ , h′′ .
It remains to show that the number of graphs ψ(x, H ′ , T ′ , t′ , h′ ) to calculate
is bounded by a function of p times a polynomial in |V (G)|. We may assume
the number of nodes x in T is bounded by |V (G)|. As |α(x)| is bounded by a
7 RELATED OPEN PROBLEM
21
function of p, and H ′ has at most max{p, 2} edges for each edge within α(x),
and T ′ has at most max{p, 2} arcs for each edge within α(x), the number of
possible graphs H ′ and T ′ is bounded by a function of p. Finally, as |α(x)| is
bounded by a function of p, the number of possible functions t : α(x) → [−p, p]
and h′ : α(x) → {odd, even} is also bounded by a function of p.
Lemmas 14 and 15 imply the following:
Theorem 1. BCPP is fixed-parameter tractable.
Theorem 1 and Lemma 2 imply the following:
Theorem 2. k-arc CPP is fixed-parameter tractable.
7
Related Open Problem
Van Bevern et al. [2] mention two other parameterizations of MCPP. One of
them is by tw(G). It was proved by Fernandes et al. [10] that this parameterisation of MCPP is in XP, but it is unknown whether it is FPT [2]. A vertex
v of G is called even if the number of arcs and edges incident to v is even.
Edmonds and Johnson [8] proved that if all vertices of G are even then MCPP
is polynomial time solvable. So, the number of odd (not even) vertices is a
natural parameter. It is unknown whether the corresponding parameterization
of MCPP is FPT [2].
Acknowledgement We are grateful to both referees for several very useful
suggestions for improvement of the paper. Research of GG was supported by
Royal Society Wolfson Research Merit Award. Research of BS was supported
by China Scholarship Council.
References
[1] J. Bang-Jensen and G. Gutin, Digraphs: Theory, Algorithms and Applications, 2nd Ed., Springer, 2009.
[2] R. van Bevern, R. Niedermeier, M. Sorge, and M. Weller, Complexity of
Arc Rooting Problems. Chapter 2 in A. Corberán and G. Laporte (eds.),
Arc Routing: Problems, Methods and Applications, SIAM, Phil., in press.
[3] E. J. Beltrami and L. D. Bodin. Networks and vehicle routing for municipal
waste collection. Networks, 4(1):65–94, 1974.
[4] Bodlaender, H.L. A linear-time algorithm for finding tree-decompositions
of small treewidth. SIAM J. Computing 25(6), 1305–1317 (1996)
[5] P. Brucker. The Chinese postman problem for mixed graphs. Lect. Notes
Comput. Sci. 100 (1981) 354–366.
REFERENCES
22
[6] N. Christofides. The optimum traversal of a graph. Omega 1(1973) 719–732.
[7] R. G. Downey and M. R. Fellows, Fundamentals of Parameterized Complexity, Springer, 2013.
[8] J. Edmonds and E. L. Johnson. Matching, Euler tours and the Chinese
postman. Mathematical Programming, 5 (1973) 88–124.
[9] H. A. Eiselt, M. Gendreau, and G. Laporte. Arc routing problems. I. The
Chinese postman problem. Oper. Res. 43 (1995) 231–242.
[10] C. G. Fernandes, O. Lee, and Y. Wakabayashi. Minimum cycle cover and
Chinese postman problems on mixed graphs with bounded tree-width. Discrete Applied Mathematics, 157(2):272–279, 2009.
[11] J. Flum and M. Grohe, Parameterized Complexity Theory, Springer, 2006.
[12] T. Kloks, Treewidth: Computations and Approximations. LNCS, vol 842,
Springer, Heidelber (1994)
[13] D. Marx, B. O’Sullivan and I. Razgon, Finding small separators in linear
time via treewidth reduction. ACM Trans. Algorithms 9 (2013) article 30.
[14] E. Minieka. The Chinese postman problem for mixed networks. Management Sci. 25 (1979/80) 643–648.
[15] R. Niedermeier, Invitation to Fixed-Parameter Algorithms, Oxford UP,
2006.
[16] C. H. Papadimitriou. On the complexity of edge traversing. J. ACM 23
(1976) 544–554.
[17] Y. Peng. Approximation algorithms for some postman problems over mixed
graphs. Chinese J. Oper. Res. 8 (1989) 76–80.
[18] M. Sorge, Some Algorithmic Challenges in Arc Routing. Talk at NII Shonan
Seminar no. 18, May 2013.
[19] F.J. Zaragoza Martı́nez, Postman Problems on Mixed Graphs. PhD thesis,
University of Waterloo, 2003.
| 8cs.DS
|
ON INVARIANTS AND ENDOMORPHISM RINGS OF CERTAIN
LOCAL COHOMOLOGY MODULES
arXiv:1310.0920v2 [math.AC] 7 Oct 2013
WAQAS MAHMOOD AND
PETER SCHENZEL
Abstract. Let (R, m) denote an n-dimensional Gorenstein ring. For an ideal I ⊂ R
with grade I = c we define new numerical invariants τi,j (I) as the socle dimensions of
i
Hm
(HIn−j (R)). In case of a regular local ring containing a field these numbers coincide
with the Lyubeznik numbers λi,j (R/I). We use τd,d (I), d = dim R/I, to characterize
the surjectivity of the natural homomorphism f : R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)). As a
technical tool we study several natural homomorphisms. Moreover we prove a few results
on τi,j (I).
1. Introduction
Let (R, m) denote a local ring. For an ideal I ⊂ R let HIi (R), i ∈ Z, denote the local
cohomology modules of R with respect to I (see [4] and [2] for the definitions). In recent
research there is an interest in the study of the endomorphism rings HomR (HIi (R), HIi (R))
for certain i and I. This was started by Hochster and Huneke (see [8]) for i = dim R and
I = m. See also [3] for a generalization to an arbitrary ideal I and i = dim R.
The case of a cohomologically complete intersection with i = grade I was investigated
at first by Hellus and Stückrad (see [7]). They have shown that the natural map
f : R̂ → HomR̂ (HIi R̂ (R̂), HIi R̂ (R̂))
is an isomorphism. In the case of a Gorenstein ring (R, m) it follows (see [17]) that
HomR̂ (HIi R̂ (R̂), HIi R̂ (R̂)) is a commutative ring. Here we continue with the investigations
of f .
Theorem 1.1. Let (R, m) denote an n-dimensional Gorenstein ring. For an ideal I ⊂ R
with grade I = c and d = dim R/I the following conditions are equivalent:
(i) dimk HomR (k, Hmd (HIc (R))) = 1.
(ii) The natural map f : R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)) is surjective.
Let S = ∩(R \ p) where the intersection is taken over all associated prime ideal p ∈
AssR R/I with dim R/p = dim R/I. Then ker f = 0RS ∩ R. That is, if R̂ is a domain
then Ann HIcR̂ (R̂) = 0.
As a consequence of Theorem 1.1 there is a numerical condition for f being an isomorphism. In general the socle dimension of Hmd (HIc (R)) is not finite. In case of (R, m) a
regular local ring containing a field it is known that dimk HomR (k, Hmd (H c (R))) is equal
2000 Mathematics Subject Classification. Primary: 13D45.
Key words and phrases. Local cohomology, Bass number, Lyubeznik number, endomorphism ring.
This research was partially supported by the Higher Education Commission, Pakistan.
1
2
W. MAHMOOD AND P. SCHENZEL
to the Lyubeznik number dimk ExtdR (k, HIc (R)). Another part of these investigations
is the relation of the Bass-Lyubeznik number to the socle dimension of Hmd (HIc (R)).
To this end we consider several re-interpretations of the natural mapf by Local and
Matlis Duality resp. This is related to the problem whether the natural homomorphism
ExtdR (k, HIc(R)) → k is non-zero, a question originally posed by the second author and
Hellus (see [6]). Under the additional assumption of inj dimR HIc (R) ≤ d we are able
to prove that if ExtdR (k, HIc(R)) → k is non-zero (resp. isomorphic), then f is injective (resp. isomorphic). Note that if (R, m) is a regular local ring containing a field
inj dimR HIc (R) ≤ d holds by the results of Huneke and Sharp resp. Lyubeznik (see [10]
resp. [11]). It is unknown to us whether this is true for any regular local ring. In the final
section we discuss the τ -numbers in more detail.
2. Preliminaries
Let (R, m) denote a (local) Gorenstein ring of dimension n. Let I ⊂ R be an ideal of
grade(I) = c. Then we have that height(I) = c and d := dimR (R/I) = n − c. For the
definition and basic results on local cohomology theory we refer to [4] and the textbook
[2]. In the following we need a slight extension of the Local Duality Theorem as its was
proved at first by Grothendieck (see [4]).
For an arbitrary local ring (R, m) let R̂ denote the m-adic completion. Moreover E =
ER (k), k = R/m, denotes the injective hull of the residue field k.
Let x = x1 , . . . , xr ∈ I denote a system of elements of R such that Rad I = Rad(x)R.
We consider the Čech complex Čx with respect to x = x1 , ..., xr (see [2] and also [13] for
its definition). Then
H i (Čx ⊗R M) ≃ HIi (M)
for i ∈ Z and an R-module M, where HIi (M), i ∈ Z, denote the local cohomology of M
with support in I.
Lemma 2.1. Let (R, m) denote a Gorenstein ring. Let M denote an arbitrary R-module.
Then there are the following natural isomorphisms
i
(a) TorR
n−i (M, E) ≃ Hm (M) and
n−i
(b) HomR (Hmi (M), E) ≃ ExtR
(M, R̂)
for all i ∈ Z.
Proof. Let x = x1 , . . . , xn denote a system of parameters of R. Then Hmi (R) ≃ H i (Čx )
and therefore
Hmi (R) ≃ H i (Čx ) = 0
for all i 6= n and Hmn (R) ≃ E. Note that R is a Gorenstein ring. That is, Čx [−n] provides
a flat resolution of E. Whence by the definitions
i
i
TorR
n−i (M, E) ≃ H (Čx ⊗R M) ≃ Hm (M)
for all i ∈ Z and any R-module M. This proves the first statement. By applying the
Matlis duality functor it follows that
i
HomR (TorR
n−i (M, E), E) ≃ HomR (Hm (M), E)
ON INVARIANTS OF LOCAL COHOMOLOGY
3
Therefore the duality beween ”Ext” and ”Tor” yields the isomorphism
n−i
HomR (TorR
n−i (M, E), E) ≃ ExtR (M, HomR (E, E)).
By Matlis duality we get HomR (E, E) ≃ R̂, which completes the proof.
Note that the proof of Lemma 2.1 is well-known (see for example [9]). The new aspect
- we need in the following - is its validity for an arbitrary R-module M.
Now let us recall the definition of the truncation complex as it was introduced in [15].
Assume that (R, m) is an n-dimensional Gorenstein ring. Let ER· (R) denote the minimal
injective resolution of R. Let ΓI (−) denote the section functor with support in I. Then
ΓI (ER· (R))i = 0 for all i < c (see [15] for more details).
Definition 2.2. Let CR· (I) be the cokernel of the embedding HIc (R)[−c] → ΓI (ER· (R))
considered as a morphism of complexes. It is called the truncation complex of R with
respect to I. Then there ia a short exact sequence of complexes
0 → HIc (R)[−c] → ΓI (ER· (R)) → CR· (I) → 0.
In particular H i (CR· (I)) = 0 for all i ≤ c or i > n and H i (CR· (I)) ≃ HIi (R) for all
c < i ≤ n.
The advantage of the truncation complex is the separation of the properties of the
cohomology module HIc (R) of those of the other cohomology modules HIi (R) for i 6= c.
The truncation complex is helpful in order to construct a few natural homomorphisms.
But as a first application of the truncation complex we have the following result.
Lemma 2.3. Let (R, m) be a local ring. Let X denote an R-module such that SuppR (X) ⊆
V (I) for an ideal I of grade I = c. Then there is a natural isomorphism
HomR (X, HIc (R)) ≃ ExtcR (X, R)
and ExtiR (X, R) = 0 for all i < c.
Proof. For the proof see [16, Theorem 2.3].
The truncation complex provides a few natural maps that arise also as edge homomorphisms of certain spectral sequences.
Lemma 2.4. With the previous notation and Definition 2.2 there are the following natural
homomorphisms
(a) Hmd (HIc (R)) → E,
(b) R̂ → ExtcR (HIc (R), R̂),
c
(c) TorR
c (E, HI (R)) → E,
(d) ExtdR (k, HIc (R)) → k, and
c
(e) TorR
c (k, HI (R)) → k.
Proof. Let x = x1 , . . . , xn ∈ m denote a system of parameters of R. Let Čx denote the
Čech complex with respect to x. It induces a homomorphism of complexes of R-modules
Čx ⊗R HIc (R)[−c] → Čx ⊗R ΓI (ER· (R))
4
W. MAHMOOD AND P. SCHENZEL
By view of [14, Theorem 3.2] we have a morphism of complexes
Γm (ΓI (ER· (R))) → Čx ⊗R ΓI (ER· (R))
that induces an isomorphism in cohomology. Moreover there is an identity of functors
Γm (ΓI (−)) = Γm (−). Because R is a Gorenstein ring it follows that Γm (ER· (R)) ≃ E[−n].
Therefore the above homomorphism considered in degree n induces the homomorphism
in (a).
By view of the Local Duality (see 2.1) we have the isomorphism
c
Hmd (HIc (R)) ≃ TorR
c (E, HI (R)).
Therefore the homomorphism in (a) implies the homomorphism in (c).
Again by Local Duality (see 2.1) there is the isomorphism
ExtcR (HIc (R), R̂) ≃ HomR (Hmd (HIc (R)), E)
By the Matlis Duality it follows that the homomorphism in (a) provides the homomorphism in (b).
Now let F· denote a free resolution of the residue field k. Then the truncation complex
induces – by tensoring with F· – the following morphism of complexes of R-modules
(F· ⊗R HIc (R))[−c] → F· ⊗R ΓI (ER· (R)).
Now let y = y1 , . . . , yr a generating set for the ideal I. Let Čy denote the Čech complex
with respect to y. Because of the natural morphisms ΓI (ER· (R)) → Čy ⊗R ER· (R) (see
[14]) induces an isomorphism in cohomology. We get the quasi-isomorphism of complexes
∼
F· ⊗R ΓI (ER· (R)) −→ F· ⊗R Čy ⊗R ER· (R)
∼
But now we have a quasi-isomorphism F· ⊗R Čy −→ F· . Then the homology in degree
0 induces the homomorphism in (e). A similar argument with HomR (F· , −) implies the
homomorphism in (d). This completes the proof.
3. Natural homomorphisms
Let M denote an R-module. Then there is the natural homomorphism
Φ : R → HomR (M, M), r → fr
where fr : M → M, m → rm, for all m ∈ M and r ∈ R. This homomorphism of R
into the endomorphism ring is in general neither injective nor surjective. Of course the
endomorphism ring is in general not a commutative ring.
Remark 3.1. (A) For a local Gorenstein ring (R, m) we use the above notations and
conventions. By view of Lemma 2.4 there are the following natural homomorphisms
ϕ1 : Hmd (HIc (R)) → E,
c
ϕ2 : TorR
c (E, HI (R)) → E,
ϕ3 : R̂ → ExtcR (HIc (R), R̂), and
ϕ4 : R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)).
(B) It is known (see [17, Theorem 3.2 (a)]) that S = HomR (HIc (R), HIc (R)) is a commutative ring for an ideal I of a Gorenstein ring (R, m) and c = grade I.
ON INVARIANTS OF LOCAL COHOMOLOGY
5
In the following it will be our intention to investigate the homomorphisms of Remark
3.1 in more detail. In particular we have the following result:
Theorem 3.2. Let I ⊂ R denote an ideal of height c of a Gorenstein ring (R, m). With
the previous notation we have:
(a) ϕi , i = 1, . . . , 4, are all non-zero homomorphisms.
(b) The following conditions are equivalent:
(i) ϕ1 is injective (resp. surjective).
(ii) ϕ2 is injective (resp. surjective).
(iii) ϕ3 is surjective (resp. injective).
(iv) ϕ4 is surjective (resp. injective).
(c) The natural map ψ : HomR (k, Hmd (HIc (R))) → k is non-zero and therefore onto.
Proof. First of all we prove that there is a natural isomorphism
ExtcR (HIc (R), R̂) ≃ HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂))
This follows from Lemma 2.3 since
ExtcR (HIc (R), R̂) ≃ ExtcR̂ (HIcR̂ (R̂), R̂)
and R̂ is a flat R-module. Moreover we have c = grade I R̂.
Because ϕ4 is non-zero and the previous isomorphism is natural we get that ϕ3 is nonzero too. By Lemma 2.1 the module ExtcR (HIc (R), R̂) is the Matlis dual of Hmd (HIc (R)). So
ϕ1 is also a non-zero homomorphism. By Local Duality (see Lemma 2.1) the last module
c
is isomorphic to TorR
c (HI (R), E). This finally proves also that ϕ2 6= 0. Therefore the
statement in (a) is shown to be true. With the same arguments the equivalence of the
conditions in the statement (b) follows by Matlis Duality.
For the proof of (c) first note that ϕ4 is non-zero and 1 7→ id. So the reduction modulo
m induces a non-zero homomorphism k → k ⊗ HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)). But this is the
Matlis dual of ψ as it follows by Local Duality.
In order to investigate the natural homomorphisms (d), (e) of Lemma 2.4 we need a
few more preliminaries. First note that by Lemma 2.4 (see also [17, Section 6]) there is a
commutative diagram
ExtdR (k, HIc (R)) → k
↓
↓
d
c
Hm (HI (R))
→ E
here the right vertical morphism is - by construction - the natural inclusion and the left
vertical morphism is the direct limit of the natural homomorphisms
λα : ExtdR (k, K(R/I α )) → Hmd (K(R/I α ))
for all α ∈ N, where K(M) denote the canonical module of M. Then it induces the
following commutative diagram
ExtdR (k, HIc (R))
λ↓
ϕ
→ k
k
ψ
HomR (k, Hmd (HIc (R))) → k
6
W. MAHMOOD AND P. SCHENZEL
It is of some interest to investigate the homomorphism ϕ. In fact, it was conjectured that
ϕ is in general non-zero. This was shown to be true for R a regular local ring containing
a field (see [17, Theorem 1.3]) while ψ is non-zero in a Gorenstein ring as shown above
(see Theorem 3.2 (c)).
For our purposes here let us summarize the main results of Huneke and Sharp resp.
Lyubeznik (see [10] and [11]).
Theorem 3.3. Let (R, m) be a regular local ring containing a field. Let I ⊂ R be an ideal
of grade c. Then for all i, j ∈ Z the following results are true:
(a) Hmj (HIi (R)) is an injective R-module.
(b) inj dimR (HIi (R)) ≤ dimR (HIi (R)) ≤ dim R − i.
(c) ExtjR (k, HIi (R)) ≃ HomR (k, Hmj (HIi (R))) and its dimension is finite.
Proof. For the proof we refer to [10] and [11].
In particular the result of Theorem 3.3 (c) is a motivation for the following Definition.
Definition 3.4. If I ⊂ R is an ideal of grade c in (R, m) a Gorenstein ring. Then
τi,j (I) := dimk HomR (k, Hmi (HIn−j (R)))
is called τ -number of type (i, j) of I.
Remark 3.5. Note that τi,j (I) is the socle dimension of the local cohomology module
Hmi (HIn−j (R)). In the case of a regular local ring (R, m) containing a field it was shown
by Huneke and Sharp resp. Lyubeznik (see Theorem 3.3) that
ExtiR (k, HIn−j (R)) ≃ HomR (k, Hmi (HIn−j (R)))
for all i, j ∈ Z. These Bass numbers are also called Lyubeznik numbers. In fact, Lyubeznik
showed that they are invariants of R/I and finite.
By the example of Hartshorne (see Example 5.3) it turns out that for a Gorenstein
(R, m) the integers τi,j (I) need not to be finite.
As a first result on the integers τi,j (I) let us consider its vanishing resp. non-vanishing.
Theorem 3.6. Let I ⊂ R denote an ideal of the local Gorenstein ring R and c = grade I.
(a) τd,d (I) 6= 0, where d = dim R/I.
(b) τi,j (I) = 0 if and only if Hmi (HIn−j (R)) = 0, where n = dim R.
(c) τi,j (I) < ∞ if and only if Hmi (HIn−j (R)) is an Artinian R-module.
Proof. The statement in (a) is a consequence of Theorem 3.2 (c). For the proof of (b)
first note that any element of Hmi (HIn−j (R)) is annihilated by a power of m. Then the
statement is true by the well-known fact that the module is zero if and only if its socle
vanishes.
In order to prove (c) recall again that Supp Hmi (HIn−j (R)) ⊆ V (m). Then the finiteness
of the socle dimension is equivalent to the Artinianness of a module.
ON INVARIANTS OF LOCAL COHOMOLOGY
7
4. The endomorphism ring
As above let I ⊂ R denote an ideal of grade c in the Gorenstein ring (R, m). In this
section we shall investigate the natural homomorphism
ϕ4 : R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)).
As a first step we want to characterize when it will be surjective. To this end we need a
few preparations. For an ideal I let
Assh R/I = {p ∈ Ass R/I| dim R/p = dim R/I}.
Then we define the multiplicatively closed set S = ∩(R \ p), where the intersection is
taken over all p ∈ Assh R/I. For n ∈ N the n-th symbolic power I (n) of I is defined as
I (n) = I n RS ∩ R.
Definition 4.1. With the previous notation put
u(I) = ∩n≥1 I (n) = 0RS ∩ R.
Note that u(I) is equal to the intersection of all p-primary components q of a minimal
primary decomposition of the zero ideal 0 in R satisfying S ∩ p = ∅ .
Lemma 4.2. Let I ⊂ R is an ideal of grade c in (R, m) a Gorenstein ring. Then ker ϕ4 =
u(I R̂) and AnnR̂ HIcR̂ (R̂) = u(I R̂).
Proof. It is known (see [17, Theorem 3.2]) that ker ϕ4 = u(I R̂). Then cleary u(I R̂) =
AnnR̂ HIcR̂ (R̂).
It follows that HIcR̂ (R̂) is a torsion-free R̂-module if R̂ is a domain. In the next we
investigate when ϕ4 is surjective.
Theorem 4.3. Let (R, m) denote an n-dimensional Gorenstein ring. let I ⊂ R be an
ideal and grade(I) = c. Set d := n − c then the following conditions are equivalent:
(i) τd,d (I) = 1.
(ii) The natural homomorphism HomR (k, Hmd (HIc (R))) → k is an isomorphism.
(iii) The natural homomorphism R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)) is surjective.
Proof. First of all note that we may assume that R is complete without loss of generality.
(iii)⇒(ii): Fix the above notation. If ϕ4 is surjective it follows (see Theorem 3.2) that
ϕ1 : Hmd (HIc (R)) → E
is injective. Since ψ is non-zero it provides (by applying HomR (k, −)) that
ψ : HomR (k, Hmd (HIc (R))) → k
is an isomorphism.
(ii)⇒(i): This is obviously true by the definition of τd,d .
(i)⇒(iii): In order to simplify notation we put S = HomR (HIc (R), HIc (R)). Then we
have to show that the natural homomorphism R → S is surjective. Note that ϕ4 is
non-zero since 1 → id and HIc (R) 6= 0. Moreover
S ≃ HomR (Hmd (HIc (R)), E)
8
W. MAHMOOD AND P. SCHENZEL
as follows because of S ≃ ExtcR (HIc (R), R) (see Lemma 2.3) and by the Local Duality
Theorem. Therefore we have isomorphisms
S/mS ≃ k ⊗R HomR (Hmd (HIc (R)), E) ≃ HomR (HomR (k, Hmd (HIc (R))), E)
By the assumption and Matlis Duality this implies that dimk S/mS = 1.
Now we claim that S is m-adically complete. With the same duality argument as above
there are isomorphisms
S/mα S ≃ HomR (HomR (R/mα R, Hmd (HIc (R))), E).
for all α ∈ N. Now both sides form an inverse system of modules. Since inverse limits
commutes with direct limits into HomR (−, −) at the first place. It provides an isomorphism
lim S/mα S ≃ HomR (Hm0 (Hmd (HIc (R))), E)
←−
as follows by passing to the inverse limit. But Hm0 (Hmd (HIc (R))) ≃ Hmd (HIc (R)) since
Hmd (HIc (R)) is an R-module whose support is contained in V (m). Finally this provides an
isomorphism
lim S/mα S ≃ S
←−
and hence S is m-adically complete.
By virtue of [12, Theorem 8.4] S is a finitely generated R-module, in fact a cyclic
module S ≃ R/J. This proves the statement in (iii).
As shown above there is the following commutative diagram
ExtdR (k, HIc (R))
λ↓
ϕ
→ k
k
ψ
HomR (k, Hmd (HIc (R))) → k.
Whence there is a relation between the Bass number µd (m, HIc (R)) and τd,d (I). In fact,
if (R, m) is regular local ring containing a field, then λ is an isomorphism (see [17]). It is
conjectured ([6]) that ϕ is non-zero. Note that if ϕ is non-zero then ψ is non-zero too and
both maps are surjective. In the next we want to continue with the investigations of ϕ.
Theorem 4.4. Let (R, m) be Gorenstein ring. Suppose that I ⊂ R is an ideal of grade c
such that inj dimR (HIc (R)) ≤ d = n − c. Suppose that the natural homomorphism
ϕ : ExtdR (k, HIc (R)) → k
is non-zero (resp. isomorphic). Then
ϕ4 : R̂ → HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂))
is injective (resp. isomorphic).
Proof. Since R̂ is a flat R-module, the injective dimension of HIc (R) is finite if and only
if the injective dimension of HIcR̂ (R̂) is finite. Moreover ϕ is non-zero (resp. isomorphic)
if and only if ExtdR̂ (k, HIcR̂ (R̂)) → k is non-zero (resp. isomorphic). So without loss of
generality we may assume that R is complete.
It is a consequence of the truncation complex that there are natural homomorphisms
ϕα : ExtdR (R/mα , HIc (R)) → HomR (R/mα , E)
ON INVARIANTS OF LOCAL COHOMOLOGY
9
for all α ∈ N.
Put H = HIc (R). Then the short exact sequence
0 → mα /mα+1 → R/mα+1 → R/mα → 0
induces a commutative diagram with exact rows
ExtdR (R/mα , H) → ExtdR (R/mα+1 , H) → ExtdR (mα /mα+1 , H) → 0
↓ ϕα
↓ ϕα+1
↓f
α
α+1
0 → HomR (R/m , E) → HomR (R/m , E) → HomR (mα /mα+1 , E) → 0
Recall that the homomorphism ExtdR (R/mα+1 , H) → ExtdR (mα /mα+1 , H) is onto since
inj dim H ≤ d.
Now suppose that ϕ is non-zero, that is, it is surjective. So assume that ϕ is surjective
(resp. isomorphic). Then the natural homomorphism f is surjective (resp. isomorphic).
The snake lemma provides by induction on α ∈ N that ϕα is surjective (resp. isomorphic).
By passing to the direct limit of the direct systems it induces the homomorphism
ϕ1 : Hmd (HIc (R)) → E
is surjective (resp. isomorphic). Note that the direct limit is exact on direct systems. So
the statement follows by virtue of Theorem 3.2
Remark 4.5. (A) The assumption inj dim HIc (R) ≤ d implies that inj dim HIc (R) = d.
This follows since d = dim HIc (R) ≤ inj dim HIc (R) as it is easily seen.
(B) In the case of a regular local ring (R, m) containing a field it is shown (see Theorem
3.3) that inj dim HIc (R) ≤ d. We do not know whether this holds in any regular local ring.
5. The τ -numbers
In the following we shall give two applications concerning the τ -numbers of the ideal
I ⊂ R of a Gorenstein ring (R, m). To this end let us recall the following result:
Lemma 5.1. Let (R, m) denote an n-dimensional Gorenstein ring. Let I ⊂ R denote an
ideal with c = grade I. Let CR· (I) denote the truncation complex of I (see Definition 2.2).
Then there is a short exact sequence
0 → Hmn−1 (CR· (I)) → Hmd (HIc (R)) → E → Hmn (CR· (I)) → 0,
isomorphisms Hmi−c (HIc (R)) ≃ Hmi−1 (CR· (I)) for i < n and the vanishing Hmi−c (HIc (R)) = 0
for i > n.
Proof. For the proof we refer to [6, Lemma 2.2]. It is easily seen a consequence of the
truncation complex.
Theorem 5.2. Let I ⊂ R denote a 1-dimensional ideal in the Gorenstein ring (R, m).
(a) τi,j (I) = 0 for all (i, j) 6∈ {(0, 0), (1, 1)}.
(b) τ0,0 (I) = dimk Ext1R (k, Hm1 (HIn−1 (R))) < ∞ and τ1,1 (I) = 1.
(c) HomR̂ (HIcR̂ (R̂), HIcR̂ (R̂)) ≃ R̂/u(I R̂).
10
W. MAHMOOD AND P. SCHENZEL
Proof. Since H i (CR· (I)) ≃ HIi (R) for all c < i ≤ n and zero otherwise there is a naturally
defined morphism CR· (I) → HIn (R)[−n] that induces an isomorphism in cohomology.
Therefore Hmi (CR· (I)) ≃ Hmi−n (HIn (R)) for all i ∈ Z. Whence there are an exact sequence
0 → Hm−1 (HIn (R)) → Hm1 (HIn−1 (R)) → E → Hm0 (HIn (R)) → 0
and isomorphisms Hmi−n (HIn (R)) ≃ Hmi+2−n (HIn−1 (R)) for all i 6= n, n − 1 as it is a consequence of Lemma 5.1. This provides all the vanishing of the τ -numbers as claimed in
the statement. To this end recall that dim HIi (R) ≤ n − i for i = n − 1, n. Moreover it
implies a short exact sequence
0 → Hm1 (HIn−1(R)) → E → Hm0 (HIn (R)) → 0.
By applying Ext·R (k, −) it induces isomorphisms
HomR (k, Hm1 (HIn−1(R))) ≃ k and HomR (k, Hm0 (HIn (R))) ≃ Ext1R (k, Hm1 (HIn−1(R))).
To this end recall the the natural homomorphism HomR (k, Hm1 (HIn−1(R))) → k is not
zero (see Theorem 3.3 (c)). Because HIn (R) is an Artinian R-module (see e.g [3]) we have
Hm0 (HIn (R)) = HIn (R) and its socle dimension is finite. This proves the statements in (a)
and (b). The claim in (c) is a particular case of Theorem 4.3.
In his paper (see [l]) Blickle has shown a certain duality statement for Lyubeznik numbers in the case of cohomologically isolated singularities in a regular local ring containing
a field (see also [17, Corollary 5.4] for a slight extension). Here we prove a corresponding
result for the τ -numbers for an ideal in a Gorenstein ring.
Theorem 5.3. Let I ⊂ R denote an ideal of grade c in the Gorenstein ring (R, m).
Suppose that c < n − 1 and Supp HIi (R) ⊆ V (m) for all i 6= c. Then
τ0,j (I) = τd−j+1,d (I) − δd−j+1,d ,
where d = dim R/I.
Proof. Under the assumption of Supp HIi (R) ⊆ V (m) for all i 6= c it follows that there
are isomorphisms Hmi (CR· (I)) ≃ H i (CR· (I)) for all i ∈ Z. The exact sequence and the
isomorphisms shown in Lemma 5.1 provide the statement of the claim. Note that we have
an exact sequence
0 → HIn−1(R) → Hmd (HIc (R)) → E.
This in particular provides (as above) that τd,d (I) = τ0,1 (I) + 1.
Even under the assumptions of Theorem 5.3 it may happen that τd,d is infinite. To this
end consider the following example of Hartshorne (see [5]).
Example 5.4. (cf. [5, Section 2] and [17, Example 3.5]) Let k be a field and let A =
k[|u, v, x, y|] be the formal power series ring in four variables. Consider the complete
Gorenstein ring R = A/f A, where f = xv − yu. Let I = (x, y)R. Then n = 3, d = 2, c = 1
and Supp HI2 (R) = V (m), HI3 (R) = 0. It was shown by Hartshorne (see [5, Section 2])
that
dimk HomR (k, HI2 (R)) = τ0,1 (I) = τ2,2 (I) − 1
is not finite.
ON INVARIANTS OF LOCAL COHOMOLOGY
11
Problem 5.5. It would be of some interest to get a criterion for τd,d (I) to be finite for
an ideal I ⊂ R with d = dim R/I. By Theorem 3.6 this is equivalent to the Artinianness
of Hmd (HIc (R)).
Acknowledgement. The authors are grateful to the reviewer for suggestions to improve
the manuscript.
References
[l]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
M. Blickle: Lyubeznik’s invariants for cohomologically isolated singularities, J. Algebra 308
(2007), 118-123.
M. Brodmann, R. Sharp: Local Cohomology. An Algebraic Introduction with Geometric Applications. Cambr. Stud. in Advanced Math., No. 60. Cambridge University Press, (1998).
M. Eghbali, P. Schenzel: On an endomorphism ring of local cohomology, Comm. Algebra, to
appear.
A. Grothendieck: Local Cohomology, Notes by R. Hartshorne, Lecture Notes in Math. vol. 41,
Springer, 1967.
R. Hartshorne: Affine duality and cofiniteness, Invent. Math. 9, 145-164 (1970).
M. Hellus, P. Schenzel: On cohomologically complete intersections, J. Algebra 320 (2008)
3733-3748.
M. Hellus, J. Stückrad: On endomorphism rings of local cohomology modules. Proc. Am.
Math. Soc. 136 (2008) 2333-2341.
M. Hochster, C. Huneke: Indecomposable canonical modules and connectedness, In: Commutative Algebra: Syzygies, Multiplicities, and Birational Algebra, (Eds.: W. Heinzer, C. Huneke,
J. Sally), Contemporary Math. 159, pp. 197-208, Providence, 1994.
C. Huneke : Lectures on local cohomology (with an Appendix by Amelia Taylor), Contemp.
Math. 436, pp. 51-100, Providence, 2007.
C. Huneke, R.Y. Sharp: Bass numbers of local cohomology modules, Trans. Amer. Math. Soc.
339 (1993) 765779.
G. Lyubeznik: Finiteness properties of local cohomology modules (an Application of D-modules
to Commutative Algebra), Invent. Math. 113 (1993) 41-55.
H. Matsumura: Commutative Ring Theory, Cambridge Studies in Advanced Mathematics 8,
Cambridge University Press, Cambridge (1986).
P. Schenzel: On the use of local cohomology in Algebra and Geometry. In: Six Lectures in
Commutative Algebra, Proceed. Summer School on Commutative Algebra at Centre de Recerca
Matemàtica, (Ed.: J. Elias, J. M. Giral, R. M. Miró-Roig, S. Zarzuela), Progr. Math. 166, pp.
241-292, Birkhäuser, 1998.
P. Schenzel: Proregular sequences, local cohomology, and completion, Math. Scand 92 (2003)
161-180.
P. Schenzel: On birational Macaulayfications and Cohen-Macaulay canonical modules. J. Algebra 275 (2004), 751-770.
P. Schenzel: Matlis duals of local cohomology modules and their endomorphism rings. Arch.
Math. (Basel) 95 (2010) 115-123.
P. Schenzel: On the structure of the endomorphism ring of a certain local cohomology module.
J. Algebra 344 (2011), 229-245.
Abdus Salam School of Mathematical Sciences, GCU, Lahore Pakistan
E-mail address:
[email protected]
Martin-Luther-Universität Halle-Wittenberg, Institut für Informatik, D — 06 099
Halle (Saale), Germany
E-mail address: [email protected]
| 0math.AC
|
CONSTRUCTION OF ALMOST REVLEX IDEALS WITH HILBERT
FUNCTION OF SOME COMPLETE INTERSECTIONS
arXiv:1803.02330v1 [math.AC] 6 Mar 2018
CRISTINA BERTONE AND FRANCESCA CIOFFI
Abstract. We prove the existence of the almost revlex ideal J ⊂ K[x1 , . . . , xn ] with the
same Hilbert function of a complete intersection ideal I generated by n forms of degrees
d1 ≤ · ·P· ≤ dn , when for every i ≥ 4 the degrees satisfy the condition di ≥ ūi−1 + 1 =
o
nj i−1 d −i+1 k P
i−2
j=1 j
, j=1 dj − i + 2 + 1. We also show that, for every t ≥ ūn + 1, all
min
2
terms of degree t outside J are divisible by the last variable. This further property of J
has an important role in our inductive and constructive proof. We explicitly construct
Jt , for every t up to the regularity of I, by induction. Thanks to a theorem of Cho and
Park (2008), our result gives a new partial solution to the Moreno-Socias conjecture over
a field of characteristic 0, in the case di ≥ ūi−1 + 1 for every i ≥ 4.
Introduction
Given an Artinian strongly stable ideal J ⊂ K[x1 , . . . , xn ] in n variables, over a field
K, it is straightforward that there exists an integer ℓ such that for every t ≥ ℓ the terms
of degree t outside J are divisible by the last variable. As a consequence, the Hilbert
function of K[x1 , . . . , xn ]/J is decreasing from ℓ on.
We observe that, in case we deal with the Hilbert function H of a complete intersection,
the minimal integer ℓ with the above property is related to the explicit construction of
the almost reverse lexicographic ideal with Hilbert function H.
More precisely, let H be the Hilbert function of a complete intersection defined by n
forms of degrees d1 ≤ · · · ≤ dn in n variables over a field K. We prove that, if the following
condition is satisfied for every i ≥ 4,
i−2
nj Pi−1 dj − i + 1 k X
o
j=1
(⋆)
di ≥ min
,
dj − i + 2 + 1,
2
j=1
then there exists the almost reverse lexicographic ideal J with Hilbert function H; further,njevery
term
of degree t outside
Pn
k P
o J is divisible by the last variable, for every t ≥
n−1
j=1 dj −n
min
, j=1 dj − n + 1 + 1 (see Theorem 3.8).
2
The further property that, for suitable t, the terms of degree t outside J are divisible
by the last variable is crucial in the inductive and constructive proof we provide for our
result, together with the combinatorial properties of the first expansion of the sous-escalier
of a strongly stable ideal [14, 15, 16] and the particular structure of the Hilbert function
of a complete intersection [21].
Thanks to the interesting connection between the generic initial ideal of generic forms
and almost reverse lexicographic ideals that is given by Y. H. Cho and J. P. Park in [3,
Theorem 2.11] over a field of characteristic 0, we can show that our result has consequences
in the study of the generic initial ideal (with respect to the reverse lexicographic term
1
2
CRISTINA BERTONE AND FRANCESCA CIOFFI
order) of an ideal generated by generic forms. To this aim, we recall the Moreno-Socias
conjecture (see [18, Conjecture 4.1]):
Let K be an infinite field. If I is a homogeneous ideal generated by polynomial
generic forms over K, then the generic initial ideal gin(I) of I is almost reverse
lexicographic.
Thus, thanks to [3, Theorem 2.11], the result we present in this paper proves the MorenoSocias conjecture over a field of characteristic 0 in the case condition (⋆) is satisfied, for
every i ≥ 4 (see Corollary 3.9). Indeed, discussions about generic properties of a sequence
of forms in the polynomial ring K[x1 , . . . , xn ] show that, in order to prove the MorenoSocias conjecture, it is enough to consider Artinian ideals, hence n generic forms, which
generate a complete intersection (see [20, Section 4]).
It is noteworthy that the Moreno-Socias conjecture implies the Fröberg conjecture (see
[9, 20] and the references therein, and [19] for a recent contribution) and the Pardue
conjecture (see [20, 3]). For some main results about these conjectures and connections
with the Stanley and Lefschetz properties, we refer to [1, 18, 3, 20, 17, 2], and to the
references therein.
As far as we know, the Moreno-Socias conjecture has already been proved, mostly in
characteristic 0, for n = 2 in [1], for n ≤ 3 in [18, Proposition 4.2], for n = 4 and d1 = 2 in
P
Pi−1
[10], if di > ( i−1
j=1 dj ) − i + 1 for every i ≥ 4 in [3, Corollary 2.12], if di ≥ (
j=1 dj ) − i − 1
for every i in [2, Theorem 3.19], and also in other particular cases in [4].
1. Background
Let R := K[x1 , . . . , xn ] be the polynomial ring over an infinite field K in n variables
endowed with the degree reverse lexicographic term order ≻ such that
P x1 ≻ · · · ≻ xn . In
this setting, for every term τ := xα1 1 . . . xαnn 6= 1 we let deg(τ ) := ni=1 αi be its degree
and we define min(τ ) := max{i | αi 6= 0}, which is the index of the minimal variable that
appears in τ with a non-null exponent. If min(τ ) = i, we will say that “τ has minimal
variable xi ”. We denote by T the set of the terms of R.
For every integer t, we denote by Rt the K-vector space of homogeneous polynomials
of R of degree t, for every subset Γ ⊂ R, we let Γt := Γ ∩ Rt and denote by hΓt i the
K-vector space generated by Γt .
Definition 1.1. A subset L ⊂ Tt is a reverse lexicographic segment (revlex segment, for
short) if, for every τ ∈ L and τ ′ ∈ Tt , τ ′ ≻ τ implies that τ ′ belongs to L (see the paper
[5] for other types of segments).
We refer to [22, 23] for definitions and results about Hilbert functions of standard graded
K-algebras. When we consider a homogeneous ideal I ⊂ R and the Hilbert function HR/I
of R/I, we equivalently say “the Hilbert function of I”.
Given a monomial ideal J ⊂ R, we denote by BJ the minimal monomial basis of J
and by N (J) the sous-escalier of J, that is the set T \ J of terms of R outside J. Recall
that, for every integer t, the cardinality of N (J)t coincides with the value of the Hilbert
function of R/J at the degree t.
Definition 1.2. A monomial ideal J ⊂ R is a reverse lexicographic ideal (revlex ideal,
for short) if Jt ∩ T is a revlex segment, for every degree t. A monomial ideal J ⊂ R is an
almost reverse lexicographic ideal (or weakly reverse lexicographic ideal or almost revlex
3
ideal, for short) if, for every minimal generator τ ∈ BJ of J, a term τ ′ ∈ Tdeg(τ ) belongs
to J if τ ′ ≻ τ (see [7]).
Example 1.3. Let H be the Hilbert function of the general rational curve of degree 5 in
P3K . The following almost revlex ideal J = (x32 , x22 x1 , x2 x21 , x31 , x23 x21 ) ⊂ K[x1 , . . . , x4 ] has
Hilbert function H and is not a revlex ideal because J4 ∩ T is not a revlex segment.
Remark 1.4. If J is an almost revlex ideal in the polynomial ring in n − 1 variables
K[x1 , . . . , xn−1 ] ⊂ R, then the ideal JR is an almost revlex ideal too.
Definition 1.5. A monomial ideal J ⊂ R is strongly stable if, for all τ ∈ J and for every
x τ
variable xi such that τ is divisible by xi , the term xji belongs to J for every xj ≻ xi .
Remark 1.6. (See [11, Definitions 8 and 10, and Remark 11], for example). A revlex ideal
is an almost revlex ideal. If J, J ′ ⊂ S are almost revlex ideals such that HS/J (t) = HS/J ′ (t)
for every t, then J = J ′ . An almost revlex ideal is strongly stable, in particular a revlex
ideal is strongly stable and an ideal generated by a revlex segment is strongly stable too.
If I ⊂ R is a homogeneous ideal of R, we denote by I≤t the ideal generated in R by the
homogeneous polynomials of I of degree ≤ t and by αI := min{t : HR/I (t) < t+n−1
}=
t
min{t : It 6= 0} the initial degree of I. Moreover, we denote by reg(I) the regularity of
I, that is the minimal integer m such that the h-th syzygy module of I is generated in
degrees ≤ m + h, for every h ≥ 0.
If J ⊂ R is a strongly stable ideal, then the first expansion of N (J)t is E(N (J)t) :=
Tt+1 \ ({x1 , . . . , xn } · Jt ).
Remark 1.7. [14, 15, 16] If J ⊂ R is a strongly stable ideal, then it is well-known that,
for every degree t, the first expansion E(N (J)t) of N (J)t is equal to N (J≤t )t+1 . The
first expansion can be directly computed without repetitions and in increasing order with
respect to the reverse lexicographic order as follows, where in square brackets we denote
a list of terms of Tt that is increasingly ordered with respect to ≻:
(1.1)
E(N (J)t ) =
n−1
G
xn−i · [τ ∈ N (J)t : min(τ ) ≤ n − i].
i=0
Thus, if ℓ is an integer such that N (J)ℓ ∩K[x1 , . . . , xn−1 ] = ∅ and H is the Hilbert function
of R/J, then for every t ≥ ℓ we immediately obtain:
(i) N (J)t ∩ K[x1 , . . . , xn−1 ] = ∅,
(ii) H(t) ≥ H(t + 1).
This is a weaker situation than that considered in [15, Proposition 2.13].
Given a homogeneous ideal I ⊂ R, we denote by in(I) its initial ideal with respect to
the degree revlex term order. The ideals I and in(I) have the same Hilbert function.
Referring to [8, Section 15.9], we recall the well-known result that there is a non-empty
open subset A of the general linear group GLK (n) of the invertible matrices of order n
with entries in K such that in(g(I)) is a constant monomial ideal, for every g ∈ A. This
constant ideal in(g(I)) is denoted by gin(I) and called the generic initial ideal of I. In
our setting, gin(I) is invariant under the action of the Borel subgroup of lower-triangular
invertible matrices, hence it is a Borel-fixed ideal. A strongly stable ideal is Borel-fixed
and the converse holds in characteristic 0.
4
CRISTINA BERTONE AND FRANCESCA CIOFFI
Q
Following [21, 20], the set ri=1 Rdi of sequences f1 , . . . , fr of r homogeneous polynomials
of degrees d1 , . . . , dr can be considered as an affine space in which the coordinates of the
points are the coefficients of the polynomials in the sequence. Thus, we can say that a
property of a sequence of homogeneous
polynomials f1 , . . . , fr is generic if it holds on a
Q
nonempty Zariski-open subset of ri=1 Rdi and that a generic property holds for a generic
sequence of forms or for generic forms. We have that n generic forms in R generate a
complete intersection ideal I ⊂ R (see [21, page 531] or [20, page 582]).
Remark 1.8. Referring to [13], a proper ideal I in a Noetherian ring is called a complete
intersection if the length µ(I) of the shortest system of minimal generators of I is equal to
the height h(I) of I. A proper ideal I that is generated by a regular sequence in a Noetherian ring is a complete intersection and the converse holds if the ring is Cohen-Macaulay,
like a polynomial ring over a field is [13, Definition 3.11, Corollary 5.11, Theorem 3.14].
We end this section with a result due to Y. H. Cho and J. P. Park, which will be used
in Corollary 3.9 because it gives a crucial connection between the generic initial ideal of
generic forms and almost revlex ideals. The proof of this result is based on the description
of generic initial ideals given in [8] by means of extensors.
Theorem 1.9. [3, Theorem 2.11] Assume that the field K has characteristic 0. Let J be an
almost revlex ideal in R = K[x1 , . . . , xn ]. Suppose that I is a homogeneous ideal generated
by generic forms in R. If J and I have the same Hilbert function, then gin(I) = J.
2. Preliminary results
In this section, we formulate in terms of generic forms some results that are well-known
for complete intersections and which have some important consequences for our aims.
Proposition 2.1. The Hilbert function H of an ideal I ⊂ R generated byP
n generic forms
of degrees d1 ≤ · · · ≤ dn is symmetric and the regularity of I is reg(I) = nj=1 dj − n + 1.
Proof. We know that n generic forms in R generate an Artinian complete intersection
and a complete intersection is Gorenstein. Then, the first assertion is a consequence of [6,
Theorem (Hilbert Functions under Liaison)]. The second assertion directly follows from
the shape of the minimal free resolution of a complete intersection, which is obtained from
a Koszul complex [12, Example 13.16].
Example 2.2.
(a) The Hilbert function H of the ideal I generated by 3 generic forms of degree 2 in
K[x1 , x2 , x3 ] assumes the values H(0) = 1, H(1) = 3, H(2) = 3, H(3) = 1, and
H(t) = 0 for every t ≥ 4. Indeed, H is symmetric and reg(I) = 4. The function
H also is the Hilbert function of the revlex ideal J = (x21 , x1 x2 , x22 , x1 x23 , x2 x23 , x43 ).
We have gin(J) = J because J is strongly stable, and hence Borel-fixed. Thus, by
Theorem 1.9 the ideal J is the generic initial ideal of any ideal generated by 3 generic
forms of degree 2 in K[x1 , . . . , x3 ], over a field K of characteristic 0.
(b) The Hilbert function H of the ideal generated by 4 generic forms of degree 2 in
K[x1 , . . . , x4 ] assumes the values H(0) = 1, H(1) = 4, H(2) = 6, H(3) = 4, H(4) = 1,
and H(t) = 0 for every t ≥ 5. Indeed, H is symmetric and reg(I) = 5. The function H
also is the Hilbert function of the revlex ideal J = (x21 , x1 x2 , x22 , x1 x3 , x2 x23 , x33 , x2 x3 x4 ,
x23 x4 , x1 x34 , x2 x34 , x3 x34 , x54 ). As before, by Theorem 1.9 the ideal J is the generic initial
5
ideal of any ideal generated by 4 generic forms of degree 2 in K[x1 , . . . , x4 ], over a
field K of characteristic 0.
The following very classical result connects the Hilbert function of a K-algebra with
the Hilbert function of a general hypersurface section (for example, see [12, Lecture 13]).
Lemma 2.3. If M ⊂ R is a homogeneous ideal and F ∈ R is a form of degree d that is
not a zero-divisor in R/M, then HR/M (t) − HR/M (t − d) = HR/(I,F ) (t), for every t.
Proof. It is enough
apply the
of a Hilbert function on the short exact
property
additive
to
·F
R
R
R
−→ M −→ (M,F ) −→ 0.
sequence 0 −→ M
t−d
t
t
Proposition 2.4. Let H the Hilbert function of the ideal I ⊂ R generated by n generic
forms of degrees d1 ≤ · · · ≤ dn−1 ≤ dn and H ′ the Hilbert function of the ideal I ′ ⊂
K[x1 , . . . , xn−1 ] generated by n − 1 generic forms of degrees d1 ≤ · · · ≤ dn−1 . For every t
(2.1)
H(t) =
t
X
j=0
H ′ (j) −
t−d
Xn
H ′ (j).
j=0
P
Proof. The Hilbert function of J ′ := gin(I ′ )R is tj=0 H ′(j), for every t. We can also
P
observe that H(t) ≤ tj=0 H ′ (j), for every t ≥ 0, because I contains the ideal I¯ generated
by the given first n − 1 generic forms of I and I¯ has the same Hilbert function as J ′ . By
applying Lemma 2.3 to M = I¯ and I = (M, F ), we obtain formula (2.1).
Notation 2.5. Let d1 ≤ · · · ≤ dn be n positive integers. For every 1 ≤ i ≤ n, we will
need the following notation:
P
• mi := ( ij=1 dj ) − i;
o
n
mi
• ū1 := 0 and ūi := min ⌊ 2 ⌋, mi−1 .
The following result is a weaker version of [21, Theorem 1 and Corollary 2], where the
behavior of the Hilbert function H of a complete intersection is precisely described. Here,
we only recall the conditions we need in order to prove our main result.
Theorem 2.6. The Hilbert function H of an ideal I ⊂ R generated by n generic forms
of degrees d1 ≤ · · · ≤ dn is strictly increasing in the range [0, ūn ] and is decreasing in the
range [ūn , reg(I)].
Proof. We can refer to [21, Theorem 1 and Corollary 2], because n generic forms in R
generate a complete intersection.
Corollary 2.7. In the same setting of Proposition 2.4, we have:
(i) H(t + 1) + H ′ (t) − H ′ (t + 1) + H ′ (t − dn + 1) − H ′ (t − dn ) =
H(t) + H ′ (t) − H ′ (t − dn ), for every t;
(ii) if dn ≥ ūn , then H ′(t + 1 − dn ) − H ′ (t + 1) ≥ 0, for every t ≥ dn ;
(iii) if dn < ūn , then H ′ (t + 1) − H ′(t + 1 − dn ) > 0, for every dn ≤ t < ūn .
Proof. Item (i) is an immediate consequence of Proposition 2.4. Items (ii) and (iii) follow
from item (i) and Theorem 2.6.
Example 2.8. If n = 4 and d1 = 4, d2 = 5, d3 = 7, d4 = 8, then m1 = 3, m2 = 7, m3 = 13,
m4 = 20, ū2 = 3, ū3 = 6, ū4 = 10 and the Hilbert functions of Proposition 2.4 are:
6
CRISTINA BERTONE AND FRANCESCA CIOFFI
t
H(t)
H ′ (t)
t
Σj=0 H ′ (j)
0
1
1
1
1 2 3 4 5 6 7
8
9 10 11 12 13 14 15
4 10 20 34 51 70 89 105 116 120 116 105 89 70 . . .
3 6 10 14 17 19 19 17 14 10
6
3
1
0 ...
4 10 20 34 51 70 89 106 130 140 146 149 150 150 . . .
3. Main result
Let d1 ≤ · · · ≤ dn−1 ≤ dn be n positive integers. In this section, we continue to denote
by H the Hilbert function of the ideal I generated by n generic forms in R of degrees
d1 ≤ · · · ≤ dn−1 ≤ dn and by H ′ the Hilbert function of the ideal I ′ generated by n − 1
generic forms in K[x1 , . . . , xn−1 ] of degrees d1 ≤ · · · ≤ dn−1 .
We will prove that under the assumption (⋆), i.e. di ≥ ūi−1 + 1 for every i ≥ 4, there
exists the almost revlex ideal J with Hilbert function H and the terms of degree t outside
J are divisible by xn , for every t ≥ ūn + 1. We will repeatedly use Remark 1.7.
Proposition 3.1. (Base Case) If n = 2, there exists the revlex ideal in K[x1 , x2 ] with
Hilbert function H. In particular, αJ [2] = d1 = ū2 + 1.
Proof. It is enough to recall that the following revlex ideal in K[x1 , x2 ]
(xd11 , xd11 −1 xd22 −d1 +1 , xd11 −2 xd22 −d1 +3 , . . . , xd11 −k xd22 −d1 +2k−1 , . . . , x1 xd21 +d2 −3 , xd21 +d2 −1 )
has the same Hilbert function of the ideal generated by two generic forms of degrees
d1 ≤ d2 (see [1, Corollary 2.2] and [18, Proposition 4.2]).
Lemma 3.2. Let 1 < d1 ≤ · · · ≤ dn−1 ≤ dn be n ≥ 3 positive integers. Then:
(i) d3 ≥ d2 ≥ ū2 + 1.
(ii) If ūn ≥ t ≥ dn , then ūn−1 > t − dn .
(iii) If ūn ≤ ūn−1 then d1 = dn .
Proof. For item (i), it is enough to observe that ū2 ≤ d2 − 1 by the definition.
For item (ii), the hypotheses are ⌊ m2n ⌋ ≥ t ≥ dn and mn−1 ≥ t ≥ dn . Then, ⌊ mn−1
⌋=
2
d1 +···+dn−1 −n+1
mn
⌊
⌋ > ⌊ 2 ⌋ − dn ≥ t − dn and mn−2 = d1 + · · · + dn−2 − n + 2 > mn−1 − dn ≥
2
t − dn .
For item (iii), on the contrary assume d1 < dn . We have four possible cases.
If ūn = ⌊ m2n ⌋ ≤ mn−1 and ūn−1 +1 = ⌊ mn−1
⌋+1, then ūn−1 +1 = ⌊ d1 +···+d2n−1 −n+1 ⌋+1 ≤
2
⌊ m2n ⌋, because 1 − dn ≤ −2 being 1 < d1 < dn .
⌋ + 1, then ūn−1 + 1 ≤
If ūn = ⌊ m2n ⌋ ≤ mn−1 and ūn−1 + 1 = mn−2 + 1 ≤ ⌊ mn−1
2
mn
⌋
+
1
≤
⌊
⌋,
as
before.
⌊ mn−1
2
2
If ūn = mn−1 ≤ ⌊ m2n ⌋ and ūn−1 + 1 = mn−2 + 1 ≤ ⌊ mn−1
⌋ + 1, then mn−2 + 1 =
2
d1 + · · · + dn−2 − n + 2 + 1 ≤ mn−1 , because 2 − dn−1 ≤ 0.
⌋ + 1 < mn−2 + 1, we have ūn ≤ ūn−1 only
If ūn = mn−1 < ⌊ m2n ⌋ and ūn−1 + 1 = ⌊ mn−1
2
if d1 + · · · + dn ≤ n − 1, that is impossible.
Example 3.3. In order to have an example in which the hypothesis of the statement of
Lemma 3.2(iii) holds, we can consider n = 3, and then ū3 = 1 = ū2 and 2 = d1 = d2 = d3 .
Proposition 3.4. (Core of the Inductive Step) Let ūn ≥ t ≥ dn ≥ ūn−1 + 1 and Je =
et
Je≤t ⊂ R be an almost revlex ideal with HR/Je(h) = H(h), for every h ≤ t, and N (J)
consisting of
H(t − 1) terms with minimal variable xn ,
7
H ′(t) − H ′ (t − dn ) terms with minimal variable xn−1 .
Then, we can construct an almost revlex ideal J = J≤t+1 so that HR/J (h) = H(h) for
every h ≤ t + 1 and
(1) if ūn > t, then N (J)t+1 consists of
H(t) terms with minimal variable xn ,
H ′ (t + 1) − H ′ (t + 1 − dn ) terms with minimal variable xn−1 ;
(2) if ūn = t, then N (J)ūn +1 consists of terms with minimal variable xn .
e t ) consists of
Proof. In both cases the first expansion E(N (J)
′
′
H(t) = H(t − 1) + H (t) − H (t − dn ) terms with minimal variable xn ,
H ′(t) − H ′ (t − dn ) terms with minimal variable xn−1 ,
which are N := H(t + 1) + H ′ (t) − H ′ (t + 1) + H ′ (t + 1 − dn ) − H ′ (t − dn ) terms of degree
t + 1, by Corollary 2.7(i).
If t < ūn , then N ≥ H(t + 1), because H ′ (t) − H ′ (t + 1) ≥ 0 by Theorem 2.6, being
t ≥ dn ≥ ūn−1 +1, and H ′ (t+1−dn )−H ′ (t−dn ) > 0 by Theorem 2.6, being t−dn +1 ≤ ūn−1
e t ) exactly s := N − H(t + 1)
thanks to Lemma 3.2(ii). Hence, we can remove from E(N (J)
terms that by construction we can choose to be the highest possible τ1 , . . . , τs divisible by
xn−1 because N −H(t+1) = H ′(t)−H ′ (t+1)+H ′(t+1−dn )−H ′(t−dn ) ≤ H ′ (t)−H ′ (t−dn )
by Corollary 2.7(iii). Thus, we define J so that J≤t = Je≤t , Jt+1 := hJet+1 ∪ {τ1 , . . . , τs }i,
hence J is almost revlex and N (J)t+1 consists of
H(t) terms with minimal variable xn .
H ′(t + 1) − H ′ (t + 1 − dn ) terms with minimal variable xn−1 .
If t = ūn , we now show that the number H ′(ūn ) − H ′ (ūn − dn ) of terms with minimal
e ūn ) is lower than or equal to the number s of terms of degree
variable xn−1 in E(N (J)
e ūn ). By Corollary 2.7(i)
ūn + 1 that we need to remove from the first expansion E(N (J)
we have
H(ūn + 1) + H ′(ūn ) − H ′(ūn + 1) + H ′ (ūn + 1 − dn ) − H ′ (ūn − dn ) =
= H(ūn ) + H ′ (ūn ) − H ′ (ūn − dn )
and, being H(ūn ) ≥ H(ūn + 1) due to Theorem 2.6, we obtain the claim. Denoting by
e ūn ), we define J so that J≤ūn = Je≤ūn , Jūn +1 :=
τ1 , . . . , τs the greatest terms in E(N (J)
hJeūn +1 ∪ {τ1 , . . . , τs }i, hence J is almost revlex and every term in N (J)ūn +1 has minimal
variable xn .
From now, we denote by J [i] the generic initial ideal of the ideal generated by i generic
forms of degrees d1 ≤ · · · ≤ di in K[x1 , . . . , xi ], for every 1 ≤ i ≤ n.
Proposition 3.5. (Ending of the Inductive Step) Let t > ūn and Je = Je≤t ⊂ R be
e t consisting of
an almost revlex ideal with HR/Je(h) = H(h), for every h ≤ t, and N (J)
terms with minimal variable xn . Then, we can construct the almost revlex ideal J so that
HR/J (v) = H(v) and, for every v ≥ t, every term in N (J)v is divisible by xn .
e t consists of H(t) terms with minimal variable xn . Then,
Proof. By the hypothesis, N (J)
e t ) exactly contains H(t) terms of degree v := t+1 with minimal
the first expansion E(N (J)
variable xn , by Remark 1.7. Being v > ūn , we have H(v − 1) ≥ H(v), by Theorem 2.6.
e t ) and define
Letting s := H(v−1)−H(v), we take the greatest s terms τ1 , . . . , τs in E(N (J)
J so that J≤t = Je≤t , Jv := hJev ∪ {τ1 , . . . , τs }i, hence J is almost revlex and N (J)v consists
8
CRISTINA BERTONE AND FRANCESCA CIOFFI
of terms with minimal variable xn . We repeat this construction until v = reg(J [n] ), so
that H(v) = 0.
Lemma 3.6. Assume 1 < d = d1 = d2 = · · · = dn and n ≥ 3.
(i) If d ≥ ūi−1 + 1 for every i ≥ 4, then either n = 3 or n = 4 and d = 2.
(ii) If ūi ≥ d ≥ ūi−1 + 1 for every i ≥ 3, then n = 3 and d ≥ 3.
(iii) If n = 3 and ū3 ≤ ū2 , then d = 2.
Proof. For item (i), from Lemma 3.2(i) we have d3 = d ≥ ū2 + 1 = d and, if d4 =
⌋ + 1 = ⌊ 3d−1
⌋, then d = 2. Moreover, we have 2 = d5 < ū4 + 1.
d ≥ ū3 + 1 = ⌊ 3d−3
2
2
⌋ ≥ d only if d ≥ 3. For item (iii), we have
For item (ii), we observe that ū3 = ⌊ 3d−3
2
ū3 + 1 ≤ ū2 + 1 = d. If m2 + 1 = ū3 + 1 ≤ d, then d2 ≤ 1, that is impossible in the given
setting. If ⌊ m23 ⌋ + 1 = ū3 + 1 ≤ d, then d ≤ 2.
Proposition 3.7. (Inductive Step) Let n ≥ 3 and ūn ≥ dn ≥ ūn−1 + 1. If there is the
¯t
almost revlex ideal J¯ ⊂ K[x1 , . . . , xn−1 ] with Hilbert function H ′ and all terms in N (J)
divisible by xn−1 , for every t ≥ ūn−1 + 1, then there is the almost revlex ideal J ⊂ R with
Hilbert function H and all terms in N (J)t divisible by xn , for every t ≥ ūn + 1.
Proof. Excluding the case d1 = · · · = dn = 1, which is obvious, we have 1 < d1 ≤ · · · ≤ dn .
Being n ≥ 3, by Lemma 3.6(ii), we have either dn > d1 = αJ > 1 or n = 3 and d3 ≥ 3.
Recall that H ′(t) is the Hilbert function of the ideal I ′ ⊂ K[x1 , . . . , xn−1 ] generated by
n − 1 generic forms in K[x1 , . . . , xn−1 ]. By the hypothesis, there exists the almost revlex
¯ t divisible by
ideal J¯ ⊂ K[x1 , . . . , xn−1 ] with Hilbert function H ′ (t) and all terms in N (J)
¯ ⊂ R.
xn−1 , for every t ≥ ūn−1 + 1. Consider the ideal Je := JR
Recall that H(t) denotes the Hilbert function of an ideal generated by n generic forms
e we now construct the almost revlex ideal
in R of degrees d1 ≤ · · · ≤ dn . Starting from J,
J with Hilbert function H(t) such that all terms in N (J)t are divisible by xn , for every
t ≥ ūn + 1. We transform Je by removing suitable terms from the sous-escalier of Je at
every degree t ≥ dn , in order to decrease its Hilbert function, which is greater than H(t)
by Proposition 2.4.
e t consists of H ′(t) terms
From the hypothesis, for every t ≥ dn ≥ ūn−1 + 1 > 1, N (J)
Pt−1 ′
with minimal variable xn−1 and j=0 H (j) terms with minimal variable xn .
We begin
by setting Jt = Jet , for every t < dn , and observing that
Pdnour construction
H(dn ) = j=0 H ′ (j) − H ′ (0). Hence, at degree t = dn we need to remove only 1 = H ′ (0)
e dn . From the hypothesis we have dn ≤ ūn ≤ mn−1 = reg(J [n−1] ) − 1.
term from N (J)
Then, we can choose the highest possible term τ in N (J ′)dn divisible by xn−1 , because
H ′ (dn ) 6= 0. So, we define Jdn := hJedn ∪ {τ }i and obtain N (J)dn consisting of
H(dn − 1) terms with minimal variable xn ,
H ′(dn ) − H ′ (0) terms with minimal variable xn−1 .
Now, we can assume to have J≤t satisfying the hypotheses of Proposition 3.4.
If t < ūn , then we are in the case (1) of Proposition 3.4, and if t + 1 < ūn then we can
apply again Proposition 3.4.
If t = ūn , then we are in the case (2) of Proposition 3.4 and, for every t ≥ ūn + 1,
we obtain J≤t with N (J)t only consisting of terms with minimal variable xn , applying
Proposition 3.5.
9
Theorem 3.8. (Main Result) Assume di ≥ ūi−1 + 1, for every i ≥ 4. Then, there exists
the almost revlex ideal J ⊂ R with Hilbert function H(t) and all terms in N (J)t divisible
by xn , for every t ≥ ūn + 1.
Proof. By induction on n, if n = 2 the statement holds by Proposition 3.1. Assume n ≥ 3.
By Lemma 3.6, we have either dn > d1 or n = 3 or n = 4 and 2 = d1 = d2 = d3 = d4 . For
this last case, we refer to Example 2.2 where ū4 + 1 = d4 = 2. So, we can assume dn > d1
or n = 3. By the hypothesis, there exists the almost revlex ideal J¯ ⊂ K[x1 , . . . , xn−1 ] with
¯ t divisible by xn−1 , for every t ≥ ūn−1 + 1.
Hilbert function H ′ (t) and all terms in N (J)
¯ ⊂ R and let Jt := Jet , for every t < dn . We distinguish four
Consider the ideal Je := JR
cases.
If ūn ≥ dn ≥ ūn−1 + 1, then we argue as in Proposition 3.7.
If dn ≥ ūn + 1 and ūn + 1 ≤ ūn−1 + 1, then dn = d1 by Lemma 3.2, and then n = 3, 4
and d1 = 2 by Lemma 3.6(i) and (iii). In this case we have ū3 = ū2 = 1 and ū4 = 2, so
we have only the case n = 3 and d1 = 2 for which we refer to Example 2.2.
e dn = N (Je≤dn −1 )dn only
If dn ≥ ūn + 1 > ūn−1 + 1 and dn ≥ reg(J [n−1] ), then N (J)
consists of termsP
with minimal variable xn and we can apply Proposition 3.5. Moreover,
we have H(t) = tj=0 H ′ (t), for every t < dn , and then reg(J [n−1] ) = ūn + 1 = dn .
If dn ≥ ūn + 1 > ūn−1 + 1 and dn < reg(J [n−1] ) then we can begin our construction
Pn
as in the proof of Proposition 3.7, as follows. We have H(dn ) = dj=0
H ′ (j) − H ′ (0) so
e dn . Assuming
that at degree t = dn we need to remove only 1 = H ′ (0) term from N (J)
e dn divisible
dn ≤ reg(J [n−1] ) − 1, then we can choose the highest possible term τ in N (J)
′
by xn−1 , because H (dn ) 6= 0. So, we define Jdn := hJedn ∪ {τ }i and obtain J almost
degrevlex with Hilbert function H(t) for every t ≤ dn and N (J)dn consisting of
H(dn − 1) terms with minimal variable xn ,
H ′(dn ) − H ′ (0) terms with minimal variable xn−1 .
By Corollary 2.7(i) we have:
H(dn ) + H ′ (dn − 1) − H ′ (dn ) + H ′(0) − H ′ (−1) = H(dn − 1) + H ′ (dn − 1) − H ′ (−1)
hence H(dn ) − H ′ (dn ) + H ′ (0) = H(dn − 1) and H ′ (dn ) − H ′(0) ≤ 0 because H(dn − 1) ≥
H(dn ) by Theorem 2.6. In particular, we obtain H ′ (dn ) − H ′ (0) = 0 because H ′ (dn ) > 0
and H ′ (0) = 1. Thus, we can repeatedly apply Proposition 3.5 obtaining that all the terms
of degree t outside J have minimal variable xn , for every t ≥ dn = reg(J [n−1]P
)−1. Further,
dn = reg(J [n−1] ) − 1 and H(dn ) = H(dn − 1). Moreover, we have H(t) = tj=0 H ′ (t), for
every t < dn = reg(J [n−1] ) − 1, and then reg(J [n−1] − 1) = ūn + 1 = dn by the property of
ūn described in Theorem 2.6.
Corollary 3.9. Let K be a field of characteristic 0. If d1 ≤ · · · ≤ dn are n positive
integers so that
i−2
nj Pi−1 dj − i + 1 k X
o
j=1
di ≥ ūi−1 + 1 = min
,
dj − i + 2 + 1
2
j=1
for every i ≥ 4, then the generic initial ideal gin(I) of the ideal I generated by n generic
forms of degrees d1 , . . . , dn in R is almost revlex and all terms of degree t outside gin(I)
are divisible by xn , for every t ≥ ūn + 1.
Proof. The statement is a consequence of Theorems 1.9 and 3.8.
10
CRISTINA BERTONE AND FRANCESCA CIOFFI
Remark 3.10. We highlight that, given a degree t, the condition that the terms in N (J)t
are divisible by xn is weaker than the condition that Jt ∩T is a revlex segment. Indeed, we
can construct the almost revlex ideal J with the same Hilbert function of a complete intersection generated by forms of degrees d1 = d2 = d3 = 4, d4 = 5, d5 = 6 in K[x1 , . . . , x5 ],
although Theorem 3.8 does not apply to this case. We can see that N (J)10 only contains
terms divisible by x5 , but J10 ∩ T is not a revlex segment. Instead, in next Example
3.11, we construct an almost revlex ideal which becomes a revlex segment at the minimal
degree at which its sous-escalier only contains terms divisible by the last variable.
Example 3.11. The Hilbert function H of a complete intersection generated by 3 forms of
degrees d1 = 3, d2 = 4, d3 = 4 in K[x1 , x2 , x3 ] is:
t 0 1 2 3 4 5 6 7 8
H(t) 1 3 6 9 10 9 6 3 1
We have ū2 = 2 and ū3 = 4 and, as expected, there exists the following almost revlex
ideal J ⊂ K[x1 , x2 , x3 ] with Hilbert function H:
J = (x31 , x22 x21 , x32 x1 , x52 , x3 x42 , x33 x2 x21 , x33 x22 x1 , x33 x32 , x53 x21 , x53 x2 x1 , x53 x22 , x73 x1 , x73 x2 , x93 ).
We can observe that the term x42 belongs to N (J)4 . Hence, there is a term of degree ū3 = 4
outside J which is not divisible by the last variable x3 . Thus, the result of Theorem 3.8
concerning the divisibility of the terms in N (J) by the last variable is sharp.
In order to give a direct experience of computation, let us show how we obtain J
from the almost revlex ideal J ′ = (x31 , x22 x21 , x42 x1 , x62 )K[x1 , x2 , x3 ], where note that BJ ′
generates in K[x1 , x2 ] the almost revlex ideal having the same Hilbert function as the
complete intersection generated by two forms of degrees d1 = 3, d2 = 4. We have
t
H(t)
H ′ (t)
t
Σj=0 H ′ (j)
0
1
1
1
1
3
2
3
2
6
3
6
3 4 5 6 7 8 9 10
9 10 9 6 3 1 0 . . .
3 2 1 0 0 0 0 ...
9 11 12 12 12 12 12 . . .
′
and d3 = 4 ≥ ū2 + 1 = 3. We let J≤3 = J≤3
and observe that the first expansion E(N (J)3 )
contains 9 terms with minimal variable x3 and 3 with minimal variable x2 :
x43 , x33 x2 , x33 x1 , x23 x22 , x23 x2 x1 , x23 x21 , x3 x32 , x3 x22 x1 , x3 x2 x21 , x42 , x32 x1 , x22 x21 .
So, we need to remove the two greatest terms x32 x1 , x22 x21 from this expansion obtaining
J≤4 = J≤3 + (x22 x21 , x32 x1 ). The first expansion E(N (J)4) contains 10 terms with minimal
variable x3 and 1 with minimal variable x2 :
x53 , x43 x2 , x43 x1 , x33 x22 , x33 x2 x1 , x34 x21 , x23 x32 , x23 x22 x1 , x23 x2 x21 , x3 x42 , x52 .
So, we need to remove the 2 greatest terms x3 x42 , x52 from this expansion obtaining J≤5 =
J≤4 + (x3 x42 , x52 ). The first expansion E(N (J)5) contains 9 terms with minimal variable
x3 :
x63 , x53 x2 , x53 x1 , x43 x22 , x43 x2 x1 , x43 x21 , x33 x32 , x33 x22 x1 , x33 x2 x21 .
So, we need to remove the greatest 3 terms x33 x32 , x33 x22 x1 , x33 x2 x21 from this expansion
obtaining J≤6 = J≤5 + (x33 x2 x21 , x33 x22 x1 , x33 x32 ). The first expansion E(N (J)6) contains 6
terms with minimal variable x3 :
x73 , x63 x2 , x63 x1 , x53 x22 , x53 x2 x1 , x53 x21 .
11
So, we need to remove the greatest 3 terms x53 x22 , x53 x2 x1 , x53 x21 from this expansion obtaining J≤7 = J≤6 + (x53 x22 , x53 x2 x1 , x53 x21 ). The first expansion E(N (J)7) contains 3 terms with
minimal variable x3 :
x83 , x73 x2 , x73 x1 .
So, we need to remove the 2 greatest terms x73 x2 , x73 x1 from this expansion obtaining
J≤8 = J≤7 + (x73 x2 , x73 x1 ). The first expansion E(N (J)8) contains 1 term with minimal
variable x3 . So we remove this last term x93 from this expansion obtaining J = J≤9 .
Example 3.12. Theorem 3.8 and, hence, Corollary 3.9 can be applied to the cases that
have been considered in Examples 2.8 and 3.11 and also in the following cases.
If n = 4, d1 = 3, d2 = 4, d3 = 4, d4 = 14, then we have ū2 = 2, ū3 = 4, ū4 = 8 and the
Hilbert functions of Proposition 2.4 are:
t
H(t)
H ′ (t)
Σtj=0 H ′ (j)
0
1
1
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 10 19 29 38 44 47 48 48 48 48 48 48 47 44
3 6 9 10 9 6 3 1 0 0 0 0 0 0 . . .
4 10 19 29 38 44 47 48 48 48 48 48 48 48 . . .
In this case, we have d4 ≥ ū4 + 1 > ū3 + 1, d4 > reg(J [n−1]) = ū4 + 1 = 9.
If we take d4 = 8 instead of 14, then we have ū4 + 1 = reg(J [n−1 ]) − 1 = d4 = 8 and
t 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
H(t) 1 4 10 19 29 38 44 47 47 47 44 38 29 19 10 4 1
4. Conclusions
In this paper we prove that there exists the almost revlex ideal J ⊂ K[x1 , . . . , xn ] with
the same Hilbert function of an Artinian complete intersection generated by n forms of
degrees d1 ≤ · · · ≤ dn , if the degrees d1 , . . . , dn satisfy the condition
i−2
nj Pi−1 dj − i + 1 k X
o
j=1
di ≥ ūi−1 + 1 = min
,
dj − i + 2 + 1,
2
j=1
for every i ≥ 4. Moreover, we show that all terms of degree t outside J are divisible by the
last variable, for every t ≥ ūn + 1 (see Theorem 3.8). This further property of J has an
important role in our inductive and constructive proof. Thanks to a theorem of Cho and
Park [3, Theorem 2.11], our result can be applied in order to state that the Moreno-Socias
conjecture holds over a field of characteristic 0, in the case di ≥ ūi−1 + 1 for every i ≥ 4,
which is considered in this context for the first time (see Corollary 3.9).
Acknowledgment
The authors are members of “National Group for Algebraic and Geometric Structures,
and their Applications” (GNSAGA - INDAM).
References
[1] Edith Aguirre, Abdul Salam Jarrah, and Reinhard Laubenbacher, Generic ideals and Moreno-Socı́as
conjecture, Proceedings of the 2001 International Symposium on Symbolic and Algebraic Computation, ACM, New York, 2001, pp. 21–23.
[2] Juliane Capaverde and Shuhong Gao, Gröbner Bases of Generic Ideals, Available at
http://arxiv.org/abs/1711.05309, 2017, Preprint.
12
CRISTINA BERTONE AND FRANCESCA CIOFFI
[3] Young Hyun Cho and Jung Pil Park, Conditions for generic initial ideals to be almost reverse
lexicographic, J. Algebra 319 (2008), no. 7, 2761–2771.
[4] Mircea Cimpoea¸s, A note on the generic initial ideal for complete intersections, Bull. Math. Soc.
Sci. Math. Roumanie (N.S.) 50(98) (2007), no. 2, 119–130.
[5] Francesca Cioffi, Paolo Lella, Maria Grazia Marinari, and Margherita Roggero, Segments and Hilbert
schemes of points, Discrete Math. 311 (2011), no. 20, 2238–2252.
[6] E. D. Davis, A. V. Geramita, and F. Orecchia, Gorenstein algebras and the Cayley-Bacharach theorem, Proc. Amer. Math. Soc. 93 (1985), no. 4, 593–597.
[7] Todd Deery, Rev-lex segment ideals and minimal Betti numbers, The Curves Seminar at Queen’s,
Vol. X (Kingston, ON, 1995), Queen’s Papers in Pure and Appl. Math., vol. 102, Queen’s Univ.,
Kingston, ON, 1996, pp. 193–219.
[8] David Eisenbud, Commutative algebra, Graduate Texts in Mathematics, vol. 150, Springer-Verlag,
New York, 1995, With a view toward algebraic geometry.
[9] Ralf Fröberg, An inequality for Hilbert series of graded algebras, Math. Scand. 56 (1985), no. 2,
117–144.
[10] Tadahito Harima, Sho Sakaki, and Akihito Wachi, Generic initial ideals of some monomial complete
intersections in four variables, Arch. Math. (Basel) 94 (2010), no. 2, 129–137. MR 2592759
[11] Tadahito Harima and Akihito Wachi, Generic initial ideals, graded Betti numbers, and k-Lefschetz
properties, Comm. Algebra 37 (2009), no. 11, 4012–4025.
[12] Joe Harris, Algebraic geometry, Graduate Texts in Mathematics, vol. 133, Springer-Verlag, New
York, 1992, A first course.
[13] Ernst Kunz, Introduction to commutative algebra and algebraic geometry, Birkhäuser Boston, Inc.,
Boston, MA, 1985, Translated from the German by Michael Ackerman, With a preface by David
Mumford.
[14] F. S. Macaulay, Some properties of enumeration in the theory of modular systems, Proc. London
Math. Soc. (1926), no. 26, 531–555.
[15] Maria Grazia Marinari and Luciana Ramella, Some properties of Borel ideals, J. Pure Appl. Algebra
139 (1999), no. 1-3, 183–200, Effective methods in algebraic geometry (Saint-Malo, 1998).
[16] Maria Grazia Marinari and Luciana Ramella, Borel ideals in three variables, Beiträge Algebra Geom.
47 (2006), no. 1, 195–209.
[17] J. Migliore and R. M. Miró-Roig, Ideals of general forms and the ubiquity of the weak Lefschetz
property, J. Pure Appl. Algebra 182 (2003), no. 1, 79–107.
[18] Guillermo Moreno-Socı́as, Degrevlex Gröbner bases of generic complete intersections, J. Pure Appl.
Algebra 180 (2003), no. 3, 263–283.
[19] Gleb Nenashev, A note on Fröberg’s conjecture for forms of equal degrees, C. R. Math. Acad. Sci.
Paris 355 (2017), no. 3, 272–276.
[20] Keith Pardue, Generic sequences of polynomials, J. Algebra 324 (2010), no. 4, 579–590.
[21] Les Reid, Leslie G. Roberts, and Moshe Roitman, On complete intersections and their Hilbert functions, Canad. Math. Bull. 34 (1991), no. 4, 525–535.
[22] Richard P. Stanley, Hilbert functions of graded algebras, Advances in Math. 28 (1978), no. 1, 57–83.
[23] Giuseppe Valla, Problems and results on Hilbert functions of graded algebras, Six lectures on commutative algebra (Bellaterra, 1996), Progr. Math., vol. 166, Birkhäuser, Basel, 1998, pp. 293–344.
Dipartimento di Matematica “G. Peano”, Università degli Studi di Torino, Italy
Dipartimento di Matematica e Applicazioni “R. Caccioppoli”, Università degli Studi di
Napoli Federico II, Italy
| 0math.AC
|
Using CMA-ES for Tuning Coupled PID Controllers
within Models of Combustion Engines
arXiv:1609.06741v4 [cs.SY] 5 Jun 2017
Kateřina Henclová
1
Regional Innovation Centre for Electrical Engineering,
University of West Bohemia,
Univerzitnı́ 26, 306 14 Pilsen,
Czech Republic
Abstract
Proportional integral derivative (PID) controllers are important and widely
used tools in system control. Tuning of the controller gains is a laborious
task, especially for complex systems such as combustion engines. To minimize
the time of an engineer for tuning of the gains in a simulation software, we
propose to formulate a part of the problem as a black-box optimization task.
In this paper, we summarize the properties and practical limitations of
tuning of the gains in this particular application. We investigate the latest
methods of black-box optimization and conclude that the Covariance Matrix
Adaptation Evolution Strategy (CMA-ES) with bi-population restart strategy, elitist parent selection and active covariance matrix adaptation is best
suited for this task. Details of the algorithm’s experiment-based calibration
are explained as well as derivation of a suitable objective function. The
method’s performance is compared with that of PSO and SHADE. Finally,
its usability is verified on six models of real engines.
Keywords: PID controller, tuning coupled controllers, CMA-ES, black-box
optimization
1. Introduction
In a running combustion engine, one or more PID controllers ensure that
certain quantities (e.g. intake pressure or exhaust gas temperature) remain
1
Email address: [email protected]
Preprint submitted to Applied Soft Computing
June 7, 2017
constant or within given range. Since these quantities may be naturally
related and affect each other, controllers are often coupled and their gains
cannot be tuned independently of each other.
When engines are modeled, as in simulation software WAVE2 , controllers
may and need to be tuned using simulations with little or no knowledge about
the transfer function that describes the system. The complicated finiteelement model cannot be described by a simple formula and, moreover, it
comes with its own modeling and discretization errors.
Presently, manual work makes up a major part of the controller tuning
process. This lengthy procedure is based on trial and error and requires a
knowledgeable and experienced control engineer. For systems with a single
controller (or multiple but decoupled controllers), simple rules of thumb can
be employed (e.g. Ziegler-Nichols [43]). Similar, already-solved problems can
also provide a guideline. However, when having a complicated or unique system of coupled controllers, the complexity of the task makes it very difficult
to solve even for an experienced control engineer. Moreover, in our application of PID controllers within combustion engine models, other professionals
need to tune the controllers as well, creating the need for a simple-to-use,
robust tool.
Our goal is to formulate at least a part of the problem and deliver a
method that would eliminate or significantly lower the need for manual tuning. It should find a solution of a given problem within acceptable time
and with as little user interaction as possible. When combined with simple
tuning rules or educated guesses, our method is to use the provided solution
approximation as a starting point and quickly find a more refined solution.
Further testing of solutions found by our method (such as verifying their
robustness) is, however, still left for manual post-processing. In principle,
it is simple to replace the proposed objective function by its expected value
over parametric uncertainties. However, the uncertainties on the parametric
space are not available for the considered application and thus it is not a
part of this work. Moreover, such approach would significantly increase the
computational load.
The proposed method is used to design the controller gains for a predefined trajectory of the controller. For more complex system that requires the
2
WAVE is 1D engine and gas dynamics simulation software package developed by Ricardo Software [36] and used by the author of this paper.
2
use of gain scheduling [1] or fuzzy PID [33], the proposed method will be
used to design the gain in the operational point designed by the supervising engineer. The engineer will also be responsible for selection of the gain
scheduling or fuzzy rules.
The PID tuning problem with either one controller or multiple but decoupled or symmetric controllers can be and has been reformulated as a
black-box optimization problem and solved with an appropriate method.
Evolutionary algorithms have been successfully applied to many engineering problems [6] and they have also been used to tune PID controllers. The
genetic algorithm [10, 24] was one of the first to be used [30], and many others followed such as the differential evolution (DE) [35, 37, 38] in [5, 25] or
the particle swarm optimization (PSO) [28] and its combinations (e.g. with
bacterial foraging algorithm) in [8, 9, 25, 29].
The tuning problem with multiple coupled controllers can be formulated
as an optimization problem as well but it adds another level of complexity to
the task. (To the author’s knowledge, there have not been any papers dealing
with such problems.) Even though there are naturally multiple objectives
(one for each controlled quantity), we can use a simple trick to carefully
transform them into a single objective and thus enable the problem to be
solvable by usual means. This way, however, the problem becomes harder to
solve, while the time budget remains small. With simulations taking up to
several minutes each, we aim for an overnight or a one-day computation on a
regular PC, i.e. a few thousand simulation runs at most. This imposes high
expectations upon efficiency of the method used.
Considering properties of the problem, we choose to use a variant of the
Covariance Matrix Adaptation Evolution Strategy (CMA-ES) [17, 21, 22,
27], an evolutionary algorithm founded deep in probability theory. It has
proven to be very effective and robust method in the extensive testing of
Black-Box Optimization Benchmarking (e.g. [20, 4]), surpassing the above
mentioned algorithms and many others (on the relevant sort of problems).
Despite its fame in the optimization community and large number of practical
applications, it has so far been little used for tuning PID controllers [25, 26,
42] or similar problems [19, 18].
In this paper, we experiment with the algorithm’s settings and come up
with a method that meets practical requirements, such as tolerable runtime,
and is thus fit for common use by engineers working with complicated simulations (e.g. users of WAVE). Moreover, we demonstrate its applicability on
six models of real-world engines with one, two and three controllers.
3
For comparison with CMA-ES, we use PSO (as implemented in the software package DEAP [7]) and a restarted version of SHADE (Success-History
based Adaptative DE [39]) with an implementation based on that by Tanabe
[40]. We show that they do not perform by far so well as CMA-ES.
2. Formulation of the problem
In this section, we formulate the controller tuning problem as a problem
of numerical optimization. This part is essential as it determines whether the
problem will be solvable (with the given practical limits) and what method
should be used. Compared to other research on controller tuning [5, 8, 9, 25,
29, 30], dealing with coupled controllers requires an extra level of complexity.
2.1. PID controllers in the context of engine simulations
PID controllers are well known and powerful tools in system control [11,
32]. To describe their inner workings, we will use the traditional, though
perhaps old-fashioned, notation. A controller’s input is the error function
e(t) = |actual(t) − target(t)| ,
(2.1)
i.e. the time-dependent absolute difference between the desired target value
and the actual value of a quantity (as measured by a sensor or computed
by a model). The output control signal that defines the system’s subsequent
reaction is given as
Z t
d
(2.2)
e(τ )dτ + D e(t),
C(t) = P e(t) + I
dt
0
where P , I and D are the proportional, integral and derivative gains, respectively.
The controllers’ implementation is provided within the simulation software WAVE. All the computations as well as the whole engine model are
given to us as a black box and we do not analyze them. The software as
well as the models are carefully calibrated to comply with the real world.
The simulation is fully deterministic and we do not introduce any random
perturbations of environment variables during the tuning process.
We input the controllers’ gains into the simulation software and obtain
computed trajectory of the controlled quantities over time. Our goal is to find
such constant gains P , I and D for each controller, that the corresponding
4
controlled quantities converge to the target value and do so as quickly as
possible. This provides us with the key how to compare solutions’ quality.
The entire tuning process is offline but, due to nonlinearities in the system, the curve is not as smooth as it is common for simple systems.
In this work, we are always provided with initial values and with (nonzero)
constant target values of the controlled quantities. The values are given by
the context of combustion engines.
2.2. The optimization problem
With k controllers within a system, each determined by three constant
gains P , I and D, there are 3k gains to be tuned: x = (P1 , I1 , D1 , . . . , Pk , Ik ,
Dk ). Let us further suppose that we are given an objective function M :
R3k → R. It returns a single number for every vector of controllers’ gains
and describes how well the whole system is controlled by controllers with such
gains. Assuming that higher quality inputs have lower function values, we
search for the objective function’s minimizer. The numerical optimization
problem derived from the controller tuning problem can be formulated as
looking for x∗ such that M (x∗ ) = essinf x∈R3k M (x). I.e. x∗ is a minimizer
of M that is reachable by numerical means. Naturally, we do not strive to
find the perfect optimizer x∗ but rather any good-enough solution x∗∗ for
which it holds M (x∗∗ ) − M (x∗ ) < , where > 0 is a given tolerance. Such
“minimizer” x∗∗ is not unique but, for practical purposes, it need not be.
In our case, function M must incorporate a finite-element model of an
engine and process its results. The model’s input is the vector of controllers’
gains (a candidate solution) and it outputs development of the controlled
quantities’ values over time (the controllers are a part of the model). These
outputs then need to be processed in order to obtain a single number. Here,
it is essential to carefully encode information about the candidate solutions’
quality, so that their final values rank them well.
2.3. Definition of the objective function
When the controllers’ gains are set and the simulation is run, it outputs
the error functions ei (t) = ei (x, t), i = 1, . . . , n, described by (2.1). We
shall now focus on how to process ei so that the final function value contains
all information about quality of the input. In this section, the vector of
controllers’ gains x is arbitrary but fixed. So, for simplicity of notation, it is
left out from the functions’ arguments: we write ei (t) instead of ei (x, t) etc.
5
2.3.1. Controlling multiple objectives
Let us assume for a moment that once controllers’ gains are set, quality of
the setting is characterized by functions Fi (t) : R3k → R, i = 1, . . . , n, that
describe the n controlled quantities’ development (simulation time t being
given). That provides us with n objectives to be minimized. In our case,
the number of controllers is the same as the number of controlled quantities
(i.e. k = n) but, in general, this is not required. We expect k to be small,
typically k ≤ 5.
We shall combine these objectives into a single one. Using a weighted sum
(see e.g. [31]), we can define the
Pnobjective function describing the response of
the whole model as F (t) =
i=1 wi Fi (t), where wi is the weight constant
corresponding to the i-th controlled quantity. The purpose of constants wi is
to express our priorities and, more importantly, make all the objective functions Fi comparable, as they may significantly differ in range depending on
the corresponding units. Therefore, we set wi = pi /scalei , where pi describes
priority of the i-th objective and scalei contains knowledge of its typical value
or range. In our case, we set scalei = targeti > 0 to be the (constant and
nonzero) target value of the i-th controlled quantity. The targeti element is
the remainder of the integral under its (constant) curve, considering that all
the objectives are measured for the same length of time. In our experiments,
we shall always set priorities pi = 1, i = 1, . . . , n.
Thus we have obtained the way of combining multiple objectives into one:
F (t) =
n
X
i=1
1
Fi (t).
|targeti |
(2.3)
2.3.2. Step-response based objective function
It remains to define Fi , the function corresponding to the i-th objective
(controlled quantity), i = 1, . . . , n. The general framework is based on how
the step response looks like, see the idealized step response in figure 1.
The step response describes how the controlled quantity behaves after an
engine is started. The abrupt step change of the target value is the most
difficult part for the controllers to deal with. Once the engine is running, the
changes in the environment and the target values are not so sudden. Thus,
when we tune the controllers for the “worst case scenario”, there is a good
reason to believe (supported by empirical evidence) that they will be able to
control other cases as well. When the environment changes too much, e.g.
6
Figure 1: For a given setting of controllers, the error at time t is given as e(t) =
|actual(t) − target(t)|.
hitting the extremes of RPM (rotations per minute), the controller gains may
be changed using gain scheduling [1].
Functions Fi are naturally given by formula
Z t
Fi (t) =
T (τ ) E(ei (τ )) dτ, t0 ≥ 0,
(2.4)
t0
where t0 is the initial time (typically t0 = 0), T (τ ) is a function of time,
E(ei (τ )) is a function of error, ei (τ ) is the error as defined in (2.1). In our
case, t is given by the end of the simulation.
Function E(ei (τ )) describes our concern over the actual error size. On
the other hand, function T (τ ) characterizes how much we care about when
the particular error occurs. Its value at a given time can also be perceived as
weight assigned to the corresponding error function value. There are multiple
commonly used criteria of step response quality (IAE, ITAE, ITSE or ISE),
which all fit the framework described above, giving us common choices of
functions E(ei (τ )) and T (τ ). The well known ITAE criterion [12]
Z t
IT AE(t) =
τ |e(τ )| dτ,
(2.5)
0
fits our purpose the best, as we are highly concerned about when the error
happens: it is not so important in the beginning, while at the end, error
means nonconvergence.
This information must not get lost even after combining multiple objectives in (2.3). Therefore, to sort the candidate solutions better, we slightly
modify (2.5). First, we shift the time, so that the value of one second loses
its importance as a factor. Then, we can choose to penalize nonconvergent
7
solutions. We can choose to stress the importance of time by increasing the
degree of its polynomial from linear to quadratic. Or, instead, we can leave
out the beginning of the time interval as the large error that can be expected
there is – from the optimization viewpoint – only noise. Various possibilities
have been experimented with and the best option found was to set
Z t
(τ + 1)|ei (τ )| dτ t0 ≥ 0.
(2.6)
Fi (t) =
t0
Here, the beginning of integration may be shifted from 0 to arbitrary t0 .
Selection of the shift t0 along with its justification and influence upon performance is discussed later. The simulation length t is set appropriately and
is same for all controlled quantities.
2.3.3. The objective function
In previous steps, we have constructed the objective function
M (x) = F (x, t) =
n
X
i=1
1
|targeti |
Z
t
(τ + 1)|ei (x, τ )| dτ,
(2.7)
t0
where n, targeti and 0 ≤ t0 < t are given and the error functions ei are
computed by the simulation software.
2.4. Character of the problem and practical limitations
Character of the problem plays a crucial role when choosing a fitting
method to solve it. For any set of controllers’ gains, the simulation provides
us with all data necessary to compute the objective function value (equation (2.7)). However, we cannot analyze the model itself and that forces us
to take it as a black box. We can only make a few careful assumptions.
In general, objective function M is non-convex, non-linear, non-quadratic
and highly multimodal (i.e. having multiple local optima). It is probably
continuous, but we know nothing of its conditioning (i.e. when a function
is ill-conditioned, it may be very steep in some places), and there are no
derivatives available. We cannot presume that M is smooth, so neither the
derivative approximations would necessarily be sensible. However, we do
not expect M to be noisy due to its heart made of a FEM model. We also
know that the objective function is non-separable, i.e. we cannot tune one
parameter independently of the others and expect to find the optimum that
way. One controller’s parameter’s value strongly affects other parameters
8
and one controller’s setting may affect the behavior of other controllers (the
controllers may be coupled). Since the relationships between the variables
are far from random, we would like the algorithm to mine and use this information to improve its search.
Beside this, we have to consider other practical limitations. Engine simulations, as in WAVE, take up to several minutes (per case – when parameters
like rotations per minute etc. are fixed), but the time budget is limited. Usually aiming for an overnight computation, we are allowed only a few thousand
evaluations of the objective function. That is, when parallelization is used,
and therefore we require the optimization algorithm to be parallelizable. Last
but not least, the algorithm must be robust – it must reliably produce decent
results, even at the cost of somewhat slower performance.
3. The optimization method
Metaheuristic and evolutionary methods have been extremely successful when tackling hard black-box optimization problems [6]. In order to
choose a method appropriate for our problem, we consider the extensive
Black-Box Optimization Benchmarking (BBOB) results, e.g. [20, 3, 4, 23],
which compare many algorithms on a wide range of test functions. Following the relevant data – tests on multimodal, non-separable, noiseless functions – we choose the Covariance Matrix Adaptation Evolution Strategy with
bi-population restart scheme (BIPOP-CMA-ES). Despite its efficiency and
growing number of practical applications, its use for tuning of PID controllers
is scarce [42, 25, 26].
3.1. CMA-ES: idea of the basic algorithm
The Covariance Matrix Adaptation Evolution Strategy (CMA-ES) [17,
16] is an evolutionary algorithm that uses stochastic and algebraic tools to
define optimally diverse population of candidate solutions in an area that
seems to be most promising. The size of the area and its location are determined based on the algorithm’s previous experience with the objective
function. New candidate solutions are sampled from a multivariate normal
distribution:
xk ∼ m + σN (0, C), k = 1, . . . , λ,
(3.1)
whose mean m and covariance matrix C are adapted in each generation along
with the general step size σ. The number of sampled candidate solutions λ
is called the population size.
9
The mean in the new generation is defined as weighted average of several best-ranking individuals (the parent set) of the last generation. The
weights are constant and depend only on the individuals’ relative ranking.
The covariance matrix is supposed to capture information about the objective
function’s features (curvature) around the mean. Detailed description of the
sophisticated updating of the covariance matrix can be found in the original
paper [17] or the informal tutorial [15]. In short, the matrix of the previous generation is weighted and further improved by exploiting the maximum
likelihood principle (when estimating the distribution’s parameters from the
data) and by adding information of the overall progress across generations
(the evolution path). The latter technique is also used for controlling the
overall step length σ.
Authors of CMA-ES have designed the method so that minimal user
interaction or manual setting of parameters is necessary. The starting point is
needed in order to initiate the first mean m(0) but otherwise all the parameters
are assigned usable default values. A user might want to change only the
initial step length σ (0) and the population size λ. The parent set size is given
as µ = 12 λ. The population size is by default very small, given by the formula
λdef = 4 + b3 log(d)c , where d is the problem dimension [17]. For a given
budget, smaller population size means faster adaptation. More often than
not, this is a desirable property. Choice of m(0) and σ (0) is described further.
3.2. Extensions of the basic algorithm
The basic algorithm described above can be further extended and upgraded. In our application, we shall use the elitist BIPOP-aCMA-ES version, i.e. Covariance Matrix Adaptation Evolution Strategy [17] with active
covariance matrix updates [27], elitist scheme of parent selection [21] and bipopulation restart strategy [22]. The choice of this particular variant, as well
as setting the algorithm’s parameters, is supported by numerical experiments
described in section 4.
3.2.1. Restart strategies
Restart strategies enable the basic algorithm to become more robust by
outweighing premature convergence. They are based on changing the population size, which also helps to balance exploration (exploring as much of the
vector space as possible) and exploitation (thorough investigation of promising areas) better.
10
The default population size λdef is small and the optimal population size
may be much larger. The IPOP (increasing population) strategy doubles
the population size (the factor of 2 is empirical) every time a new restart is
launched, while other parameters remain unchanged [2].
The more advanced BIPOP (bi-population) restart strategy [22] makes
use of two interlacing regimes. The first one uses the IPOP restart strategy,
while the second uses varying small populations. After each restart, it is
decided which of the two regimes is to be applied next depending on whose
count of conducted function evaluations is lower. The maximal population
size is limited by the number of restarts under the first regime.
There are two prominent restart criteria that need to be set carefully and
adjusted for each model, because they depend on its numerical values. The
first one is TolFunHist, which gives tolerance in function value history (range
of the best objective function values of several last generation is almost zero).
The second criterion is TolFun, which gives tolerance in function values of
the current generation (stop, if the difference between the current best and
worst candidate solution is TolFun-negligible) and it also checks the history
of function values in a very similar way to TolFunHist. Their default values
are too small for our use and better values can be guessed from the objective
function’s numerical values.
It was observed that it is generally good to try larger values of TolFunHist
first and diminish it if it invokes restarts too often. When logging the algorithm’s run, occurrence of premature restarts is usually obvious. Criterion
TolFun can save some runtime too, when it is set up correctly (in the same
way as the first criterion). Or we can set it to a small value, preferably a
fraction of TolFunHist. That way, most restarts are called by TolFunHist
and we need to tune only this one restart criterion.
3.2.2. Elitist selection
There are multiple options of how to choose parents of a new generation
[21]. The basic algorithm selects all parents from the current generation
(non-elitist selection). However, this scheme does not exploit very good solutions found early in the search. They are used as parents within the one
generation that produced them and then are lost. Selecting the parent set
among the individuals of the current generation and their parents as well
(elitist selection) helps to preserve the exceptionally good individuals until
they are superseded and thus amplify their influence. This approach speeds
up the convergence and is advantageous in our problem, where good solu11
tions are scarce. The downside is that it may, and sometimes it does, lead
to premature convergence to a local optimum.
3.2.3. Active covariance matrix adaptation
In the original algorithm, the successful individuals are used for the covariance matrix adaptation. Variance in directions that have proven to be
beneficial is increased and thus they are preferred when sampling next generation. With the so-called active updates, information hidden in the unsuccessful individuals is exploited as well [27]. As opposed to passive decay over
time, variance in detrimental directions is actively decreased. Simply put,
besides telling the method where to go, we also tell it where not to go.
3.3. Important properties of CMA-ES and their use
CMA-ES does not use gradients and it does not even presume their existence. Moreover, it does not even use the actual values of the objective
function once relative ranking has been assigned to the candidate solutions
(except for some stopping criteria). As a result, it is invariant to strictly
monotonic transformations of the objective function. I.e. those transformations of the objective function that have no effect upon the relative ranking of
individuals do not effect the method’s performance, making it more robust.
Further, the method exhibits invariance to invertible linear transformations of the search space. In particular, CMA-ES is invariant to scaling of
variables (coordinate axes), which is the key property that makes it wellsuited for tuning multiple controllers: parameters of one controller are usually of roughly the same scale, but with multiple controllers, the scaling
may differ by many orders. Scaling of variables is used in the following
way. The algorithm is given an initial approximation of the solution (hereafter called the reference point): s = (s1 , . . . , sN ). Then, when the algorithm wants to evaluate vector v = (v1 , . . . , vN ), it scales it by the reference point’s elements’ magnitudes and inputs vector w in the engine model:
w = (|s1 | · v1 , |s2 | · v2 , . . . , |sN | · vN ). Because the scale coefficients are all
positive and signs of the variables are essential, the algorithm must always
have a zero vector as its starting point. Also, the initial step length is set to
12
σ (0) = 1.0.
Algorithm 1: Elitist BIPOP-aCMA-ES
set λ, µ
initialize m, σ, C = I, pσ = 0, pc = 0
initialize restart regime = 1, count1 = 0, count2 = 0
while termination criteria not met do
while restart criteria not met do
if not first generation in a restart then
for i = 1, . . . , µ do
xi+µ = xi
// relabel parents of previous generation
fi+µ = fi
// relabel parents’ objective function values
for i = 1, . . . , λ do
xi ∼ N (m, σ 2 C)
fi = evaluate(xi )
// sample new population from normal distribution
// evaluate xi with objective function
sort xi , i = 1, . . . , λ + µ acc. to fi
// assign relative (descending) ranking
m∗ = m
m = update m(xi , . . . , xµ )
// move the mean utilizing the parents
// the evolution paths contain information about past progress
pσ = update pσ (pσ , σ −1 C −1/2 (m − m∗ ))
// isotropic evolution path update
pc = update pc (pc , σ −1 (m − m∗ ), kpσ k)
// anisotropic evolution path update
C = update C(C, pc , (x1 − m∗ )/σ, . . . , (xλ+µ − m∗ )/σ) // covariance matrix update
σ = update σ(σ, kpσ k)
// step size update
if restart regime = 1 then
count1 = count1 + λ
else
count2 = count2 + λ
if count1 < count2 then
restart regime = 1
else
restart regime = 2
reinitialize parameters and variables acc. to selected restart regime
4. Experiment-based calibration of the optimization method
For the experiments, the 1D engine simulation software package WAVE
by Ricardo Software was used [36], as the primary incentive was to develop
a working automated tuner of PID controllers within WAVE. For CMA-ES
implementation, we take the Python code distributed by Hansen [13].
Since models of engines take several minutes to run, while the CMA-ES
13
Figure 2: The basic testing model in WAVE.
runtime is negligible, we shall measure time in number of evaluations of the
objective function (= number of simulation runs).
4.1. The basic testing model
For the algorithms’ calibration and basic testing, a quick-to-run model
was used. It has three strongly coupled controllers and, despite its relatively
simple mechanics, it proved hard to tune.
In this model, we can see a single cylinder (orange circle) engine. The blue
“clouds” contain information about the surroundings (e.g. ambient pressure
and temperature or initial fluid composition). The thick black lines depict
the ducts. The green element is an orifice – an opening of variable diameter.
The yellow PID elements are the controllers, whose gains we want to tune.
From left to right, the first PID controls the orifice diameter, the second
one controls the fuel-air ratio by manipulating the fuel injector, and the
third controls the compression ratio (the ratio of the maximum to minimum
volume in the cylinder). The arrow-like elements are actuators that perform
the actual mechanical control based on the control signal outputted by the
corresponding controllers.
14
Figure 3: The basic testing WAVE model: comparison of good solutions. Pink: the target
value, dark blue: tuned by engineer, light blue: tuned by the algorithm.
The controlled quantities measured by sensors (depicted as gray circles)
are: indicated mean effective pressure (IMEP; the average pressure acting
upon the piston during its cycle; controlled by adjusting the fuel-air ratio),
exhaust gas temperature (controlled by the compression ratio, i.e. the ratio
of largest and smallest possible capacity of the combustion chamber), volumetric efficiency (the ratio of the volume of fluid actually displaced by a
piston; controlled by opening of the orifice). Clearly, they have nontrivial
influence on each other, so the controllers are coupled and cannot be tuned
independently of each other.
We then simulate the start of this engine, when the controllers face a
simultaneous step change of all three target values. The target values of the
controlled quantities are given to us (appropriately to the context). In figure
3, each plot contains two step responses. Controllers tuned by an engineer
(hereafter called the baseline solution) yield the dark blue curves that describe the development of the controlled quantities’ values. This solution is
sufficient but we can see that a solution found by our method (light blue)
is even better. More importantly, it can be reached with very little user
interaction, saving a lot of expert manual work.
However, most candidate solutions found in the tuning process are not
acceptable: oscillatory, convergent to a different value in one or more criteria,
or even divergent (see figure 4). In either case, it is important to observe the
development of the quantities’ values in order to determine the shift t0 used in
the objective function: we choose a time point just before the initially erratic
curve starts to follow a trend. For this model of overall simulation time 6
15
Figure 4: The basic testing WAVE model: unsatisfactory solutions still provide enough
information to estimate a fitting shift.
seconds, the best time shift found is t0 = 1.2 second. It seems fitting to set
the total simulation time such that the shift makes up the first 20%. That is,
when all the controlled quantities react with approximately the same speed
as in this case. In general, we need to compromise between the slowest and
quickest responses: cut out the uninteresting information but leave enough
to guide the search. Larger t0 means greater risk of leaving out important
information but possibly faster search.
Setting t0 > 0 is specific to our problem and certainly cannot be used in all
applications of PID controllers. Whenever large overshooting is an issue, e.g.
due to circuit breakers, this approach is not applicable. In our case, however,
we have an empirical guarantee that leaving out most of the transient part
(i.e. when 0 < τ < t0 ) does not lead to unacceptable solutions. For the
purposes of optimization, it is even desirable to ignore it as that reduces the
“noise” in the objective function caused by this information. The problem
then becomes easier to solve for the optimization method.
4.2. Tuning CMA-ES
The basic testing model was used to tune CMA-ES and other attributes.
For that purpose, we consider all three controllers to be only PI controllers
(i.e. with derivative gains set to zero) as is common in our application. Because the method uses randomness and each run is different, each test is run
ten times, with no change to the given reference point. Then the minimum,
maximum and average run times are computed. The same reference point
(i.e. scaling) of rather ‘poor quality” is used for all tests. When we compare
it to the vector of a good solution, the elements are off by two orders of
magnitude on average.
16
A run is terminated when the target objective function value is hit or the
budget of 12000 function evaluations is depleted. In each run, a different solution might be found, but setting the target value to 0.5 ensures high quality
of any of them (all the corresponding step responses are almost identical).
In practice, however, the algorithm will have to be stopped manually by the
user because the desired objective function value varies greatly from model
to model and the target value is difficult to estimate.
The results are summarized in table 1. The best setting found is always
compared to the setting, where one or two specified attributes were changed.
setting
CMA-ES variants
min
max
average
active, not elitist
644
10877
4145
elitist, not active
327
5361
2184
not active, not elitist
812
> 12000*
> 5334
adaptation method
TPA
926
> 12000*
> 4724
λ0 , µ0
λ0 = 2λdef , µ0 = λdef
1533
8812
3273
λ0 = 2λdef , µ0 = 1/2λdef
697
10973
4181
0.8 × 1.5
987
8980
3147
1.2 × 1.5
674
4990
2058
2.0 × 1.5
871
4960
2837
0s
919
9567
3687
0.6 s (10%)
354
11169
2963
268
2267
1098
TolFunHist
shift t0
BEST setting
Table 1: Experiments’ results comparing various settings of CMA-ES and the objective
function, where always only the specified attribute or two were changed from the BEST
setting. BEST setting: shift t0 = 1.2s (20%), CMA-ES variants = active and elitist,
adaptation method = CSA, population size λ0 = λdef , parent set size µ0 = 1/2λdef ,
restart parameter TolFunHist = 1.5. Entries marked *: number of evaluations once
exceeded the maximum number of iterations allowed. When computing the corresponding
average value, the value of 12000 was used in such cases, resulting in lower-bound for the
average run time.
When choosing a variant of CMA-ES, the results show that the best
combination is the elitist parent selection scheme with active updates of the
covariance matrix and CSA step size adaptation (the Cumulative Step-size
Adaptation [17] uses information of the algorithm’s overall progress across
generations). The alternative TPA (Two-Point step-size Adaptation, [14])
implements a line search along the direction of the latest mean shift.
17
Then we experiment with the initial population size λ0 (i.e. the population size before the first restart, which is further used to compute the
population sizes in the next ones). Using smaller than default populations is
discouraged by Hansen [15], while more exhaustive search with bigger populations motivated the development of the IPOP restart strategy. Therefore,
we try to double the initial population size: λ0 = 2λdef . The parent set size
µ0 , by default equal to half of the population size, is then either doubled as
well (i.e. µ0 = 12 λ0 ) or it stays the same (i.e. µ0 = 12 λdef ). For a given cost,
smaller populations enable more generations than large populations, causing
faster adaptation [17], which seems beneficial in our application.
The best working value of the influential restart parameter TolFunHist
(tolerance in function value history) for the basic testing model was found
to be n/2, where n = 3 is the number of controlled quantities. It was
further observed that setting n/2 seems to be almost universally usable in our
context, even though its optimal value differs for each model. The tolerance
in function value was set to TolFun = 0.1 for the basic testing model and
later roughly adjusted for other models, scaled in proportion to the “usual”
numerical values of the objective function and expected threshold.
In the objective function, shift t0 = 1.2 second proved to be most effective.
4.3. Testing robustness of the method
Now we compare the tuned method’s performance with various reference
points. We take the baseline solution and derive the other reference points
by multiplying each of its elements by factors 10−3 , 10−2 , 10−1 , 101 , and 102
(scaling by 103 proved to be too challenging for all tested methods). These
settings simulate user estimates of various quality.
In these tests, both PI and PID controllers are considered. Since we do
not have a PID baseline solution with nonzero D parameters, we estimate
them based on the corresponding magnitudes of P and I parameters.
Results are summarized in table 2 together with other methods’ robustness tests. We can see that the run times are very good: of 120 runs, 8
reached 2000 to 3000 evaluations and only two exceeded 3000 evaluations.
Surprisingly, there is no greater difference between the 6-dimensional PI setting and 9-dimensional PID setting. Higher problem dimension is balanced
by greater flexibility of the system. Closer look at the behavior of the algorithms suggests that too large search area (i.e. scaling) causes them to “get
lost” very far from the optimum. The run times are also influenced by the
fact that there exist good solutions at the “10−1 level”.
18
We can conclude that when the reference point (scaling) is within a reasonable range of two orders of magnitude in each coordinate, the average
run times are very usable. A user should be able to provide such a reference
point – an estimate of the solution. If in doubt, magnitudes of the reference
point’s coordinates should be chosen rather smaller than larger.
5. Comparison with PSO and SHADE
Performance of the above-described variant of CMA-ES on the problem
of tuning coupled controllers was further compared with performance of two
other prominent evolutionary algorithms: the particle swarm optimization
(PSO) and the differential evolution (DE) in its success-history based adaptive variant SHADE.
5.1. PSO
PSO [28, 34] is inspired by bird flocks or fish schools, where every individual moves by itself, yet the whole self-organized system acts as a single
organism. Each particle within a swarm moves in the search space as it is
assigned a different “velocity” vector in each generation. This vector is defined as a combination of the previous “velocity”, the individual’s best known
position so far and the swarm’s (or sub-swarm’s) best known position. This
way, the whole swarm moves towards historically best areas.
For PSO implementation, we use the default setting of PSO in the DEAP
(Distributed Evolutionary Algorithms in Python) optimization framework
[7]. We only add the essential scaling of variables same as in CMA-ES.
5.2. SHADE
The basic DE algorithm [35, 37, 38] maintains a population of candidate
solution, which are further combined and tested for better objective function
values. Each candidate solution X is tested against a new point Z, which
was obtained as a binary crossover of X and Y = A + f (B − C), where A, B
and C are three distinct candidate solutions in the population and f ∈ [0, 2]
is a parameter. The crossover probability of each vector element is given
by parameter cr ∈ [0, 1]. Both parameters and the population size greatly
influence efficiency of the search.
Considering the context of our problem and the performance-comparing
tests in [41], the SHADE variant of DE was chosen to compete with CMAES. The SHADE upgrade [39] introduces adaptation of the DE parameters
cr and f based on success history that is stored in an external archive.
19
20
56
893
102 PI b.
PI b.
10−1
10−2 PI b.
33
387
101 PID b.
102 PID b.
PID b.
10−1
10−2 PID b.
-
PID b.
calibration ref. p.
10−3
562
PID baseline
-
-
101 PI b.
PI b.
113
PI baseline
10−3
#1
reference p.
-
-
624
164
-
-
28
-
-
76
-
-
61
#2
-
1362
301
104
-
-
105
-
483
48
-
-
-
#3
280
-
-
71
-
226
39
-
-
-
#4
-
1734
1186
PSO runs
-
1680
-
43
-
-
28
1600
403
32
-
-
-
#5
-
1673
6524
239
-
-
428
4118
5155
342
-
-
130
#1
-
2125
1018
418
-
-
307
1090
1514
156
-
-
396
#2
-
1945
1150
209
-
-
497
2102
2783
1560
-
-
341
#3
-
1477
3297
437
-
-
441
1243
1437
260
-
-
2100
#4
SHADE runs
-
2163
1556
390
-
-
658
1422
2123
538
-
-
155
#5
268
416
202
141
953
62
32
259
222
61
821
540
12
min
2267
2324
941
782
4130
1462
256
1617
1334
816
5700
2064
168
max
CMA-ES runs
-
-
-
125
-
-
159
-
-
50
-
-
-
PSO
-
1877
2709
339
-
-
466
1995
2602
571
-
-
624
SHADE
1098
1138
580
343
2022
1102
67
812
592
317
2061
1049
76
CMA-ES
average run time
Table 2: Results of algorithm testing on the basic model. 5 PSO runs, 5 SHADE runs and 10 CMA-ES runs were performed
for each of 13 reference points. The value of “-” means that a satisfactory solution (i.e. solution with function value less than
0.5) was not found within the provided budget of 10000 function evaluations. Average run time was not computed if one or
more runs did not finish within the given budget.
Implementation of the SHADE algorithm by Tanabe was used [40] and
tuned in accordance with its author’s recommendations [41] and additional
experiments, setting, most notably, initial cr = 0.5, initial f = 0.5 and population size = 2N , where N is the problem dimension. One restart criterion
was added to the standard SHADE algorithm, based on recommendations
in [41]: if the number of objective function evaluations exceeds 1000, restart
can be launched if the best-so-far solution is not updated for 50 × N evaluations. The particular numbers were set based on experiments and the first
condition proved very important.
Unlike in CMA-ES and PSO, the reference point is not used for scaling
but to define boundaries of the area, from which the initial population is
sampled. Each coordinate of the reference vector is multiplied by −10 and
10 to define the boundaries. Candidate solutions in further generations are
not restricted.
5.3. PSO and SHADE experiment results
The experiments with PSO and SHADE were performed with same set of
13 reference points as in the previous tests of CMA-ES, with 5 runs of each
test. The stopping criteria were either hitting the target value of the objective
function or depleting the budget of 10000 objective function evaluations. All
test results can be found in table 2.
PSO was able to outperform CMA-ES but was not very reliable. Its good
performance was limited to very good reference points and the algorithm very
often (36 out of total 65 runs) did not converge within the given budget.
The performance of SHADE was more consistent but very slow compared
to CMA-ES. Same as PSO, SHADE never converged within the given budget
for the reference points of baselines multiplied by 101 and 102 or the one used
for CMA-ES calibration.
6. Verification of the method on models of real engines
We have shown that, on the testing model, the described version of CMAES performs very well and clearly outperforms PSO and DE. Now we verify
its usability on models of real engines provided by Ricardo. The testing set
consists of two models with 1 controller (labeled M1.1, M1.2), three models with 2 controllers (labeled M2.1, M2.2, M2.3) and one model with 3
controllers (labeled M3.1). For each, we are given a baseline setting of controllers: a setting tuned manually by an engineer, which is always in the PI
21
model
M1.2
min
max
aver.
28
PI baseline
11
66
35
101
244
280
255
4
32
21
min
max
aver.
PI baseline
2
68
101
M1.1
reference p.
reference p.
PI b.
35
153
79
102 PI b.
95
519
225
10−1 PI b.
20
120
66
10−2 PI b.
49
296
123
PI baseline
1
22
9
101
PI b.
4
28
11
102 PI b.
80
225
187
10−1 PI b.
34
100
51
10−2 PI b.
57
181
94
model
M2.1
PI b.
10−1 PI b.
M2.2
PI baseline
8
98
29
101 PI b.
60
770
364
10−1 PI b.
44
107
64
PI baseline
M2.3
M2.3
9
78
32
101 PI b.
250
757
629
10−1 PI b.
49
1188
347
PID baseline
10
91
57
101
274
857
522
82
1576
749
PID b.
10−1 PID b.
M3.1
PID baseline
41
331
152
101 PID b.
827
1763
1268
10−1 PID b.
179
3867
2476
Table 3: Real-world models
form (i.e. with D gain set to zero). When a baseline value of the D gain is
needed in our tests, we estimate its magnitude from the P and I gains.
The single-controller models are tested with 5 reference points: the baseline solution and its element-wise multiples by factors 10−2 , 10−1 , 101 and 102 .
The multi-controller models are tested with three reference points: the baseline solution and its multiples of 10−1 and 101 . The primary reason for this
restriction is the enormous time consumption. Each test is run five times. A
run is terminated upon reaching the (empirically set) target value specific to
the given model. Even though each run might produce a different solution,
as long as the target value is reached, their quality is guaranteed.
All results are summarized in table 3.
6.1. Single-controller models
Even though our primary objective is to tune multiple controllers, we
include also single-controller models as they are very important in practical
use and we want our method to work for them as well. Model M1.1 contains
a turbocharger and M1.2 contains a twin turbocharger. They represent the
typical use of a PI controller in a combustion engine.
22
With the model dimension being only 2, the run times for both M1.1
and M1.2 are very short. Good reference points for single controllers can be
obtained by commonly used rules of thumb (e.g. Ziegler-Nichols) or provided
by previous (personal or programmed) experience with similar models.
6.2. Multi-controller models
The two-controller models M2.1 and M2.2 do not pose a greater challenge
than the single-controller models. For model M2.3, we compare PI and PID
control. The extreme differences in minimal and maximal run time values
of “10−1 PI baseline” and “10−1 PID baseline” tests are caused by nonoptimal setting of the restart parameter tolerance in function value history
(tolhistfun) described above. After adjustment, the “10−1 PI baseline” and
“10−1 PID baseline” run times drop and the results of PI control resemble
those of M2.2.
Unlike the previous cases, the three-controller model M3.1 requires a full
PID control. The reference point was obtained by estimating the D gains
and adding them to the given (not very good) PI engineer-tuned baseline
“solution”. It can be seen that this model is considerably harder to tune
than the previous models. While the algorithm has no trouble finding the
same near-optimal “solutions” similar to the one given by an engineer, it was
hard to get to an actual global optimum, when all three controllers converge.
7. Conclusion
This paper has shown how the Covariance Matrix Adaptation Evolution
Strategy can be applied to the problem of tuning the gains of multiple coupled
PID controllers within combustion engine simulations. We have shown that
its version with bi-population restart scheme, elitist parent selection and
active covariance matrix updates is capable of finding good parameters of up
to three PID controllers through minimization of a fitting objective function.
The method has been calibrated on a testing model and verified on models of
real-world engines, showing its practical usability and tolerable computation
times even for poor-quality reference points. On the testing model, CMA-ES
clearly outperformed PSO and SHADE methods.
Acknowledgments
This research did not receive any specific grant from funding agencies in
the public, commercial, or not-for-profit sectors. The author would like to
23
thank Ricardo Prague, s.r.o. for their support and to anonymous reviewers,
whose comments and suggestions inspired significant improvements.
References
References
[1] K. J. Åström and B. Wittenmark. Adaptive Control: Second Edition.
Courier Corporation, 2013. ISBN 9780486319148.
[2] A. Auger and N. Hansen. A restart CMA evolution strategy with increasing population size. In B. McKay et al., editors, The 2005 IEEE
International Congress on Evolutionary Computation (CEC’05), vol. 2,
pp. 1769–1776, 2005.
[3] A. Auger, S. Finck, N. Hansen, and R. Ros. BBOB 2009: Comparison
Tables of All Algorithms on All Noiseless Functions. Technical Report
RT-0383, INRIA, April 2010.
[4] A. Auger, S. Finck, N. Hansen, and R. Ros. BBOB 2010: Comparison
Tables of All Algorithms on All Noiseless Functions. Technical Report
RT-388, INRIA, September 2010.
[5] Z. Bingul. A new pid tuning technique using differential evolution for unstable and integrating processes with time delay. In Neural Information
Processing: 11th International Conference, ICONIP 2004, Calcutta, India, November 22-25, pages 254–260, Berlin, Heidelberg, 2004. Springer
Berlin Heidelberg. ISBN 978-3-540-30499-9.
[6] J. Dréo, A. Petrowski, P. Siarry, and E. Taillard. Metaheuristics for hard
optimization: methods and case studies. Springer Science & Business
Media, 2006.
[7] F.-A. Fortin, F.-M. De Rainville, M.-A. Gardner, M. Parizeau, and C.
Gagné. DEAP: Evolutionary Algorithms Made Easy. Journal of Machine Learning Research, 13:2171–2175, jul 2012.
[8] Z.-L. Gaing. A particle swarm optimization approach for optimum design of PID controller in AVR system. IEEE Transactions on Energy
Conversion, 19(2):384–391, June 2004. ISSN 0885-8969.
24
[9] S. P. Ghoshal. Optimizations of PID gains by particle swarm optimizations in fuzzy based automatic generation control. Electric Power
Systems Research, 72(3):203–212, 2004.
[10] D.E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning. Artificial Intelligence. Addison-Wesley Publishing Company, 1989. ISBN 9780201157673.
[11] G. C. Goodwin, S. F. Graebe, and M. E. Salgado. Control System
Design. Prentice Hall PTR, Upper Saddle River, NJ, USA, 1st edition,
2000. ISBN 0139586539.
[12] D. Graham and R. C. Lathrop. The Synthesis of Optimum Transient
Response: Criteria and Standard Forms. Transactions of the American
Institute of Electrical Engineers, Part II: Applications and Industry, 72
(5):273–288, Nov 1953. ISSN 0097-2185.
[13] N. Hansen. cma 1.1.06: Python Package. https://pypi.python.org/
pypi/cma.
[14] N. Hansen. CMA-ES with Two-Point Step-Size Adaptation. Research
Report RR-6527, INRIA, 2008.
[15] N. Hansen. The CMA Evolution Strategy: A Tutorial. https://www.
lri.fr/~hansen/cmatutorial.pdf, 2015.
[16] N. Hansen and S. Kern. Evaluating the CMA evolution strategy on multimodal test functions. In International Conference on Parallel Problem
Solving from Nature, pages 282–291. Springer, 2004.
[17] N. Hansen and A. Ostermeier. Completely derandomized self-adaptation
in evolution strategies. Evolutionary Computation, 9(2):159–195, 2001.
[18] N. Hansen, A. S. P. Niederberger, L. Guzzella, and P. Koumoutsakos.
Evolutionary optimization of feedback controllers for thermoacoustic instabilities. In IUTAM Symposium on Flow Control and MEMS, pages
311–317. Springer, 2008.
[19] N. Hansen, A. S. P. Niederberger, L. Guzzella, and P. Koumoutsakos. A
method for handling uncertainty in evolutionary optimization with an
application to feedback control of combustion. IEEE Transactions on
Evolutionary Computation, 13(1):180–197, 2009.
25
[20] N. Hansen, A. Auger, R. Ros, S. Finck, and P. Posik. Comparing Results of 31 Algorithms from the Black-Box Optimization Benchmarking
BBOB-2009. Workshop Proceedings of the GECCO Genetic and Evolutionary Computation Conference 2010, pages 1689–1696, 2010.
[21] N. Hansen, D.V. Arnold, and A. Auger. Evolution Strategies. In J.
Kacprzyk and W. Pedrycz, editors, Springer Handbook of Computational Intelligence, chapter 44, pages 871–898. Springer Berlin Heidelberg, 2015. ISBN 9783662435052.
[22] N. Hansen. Benchmarking a BI-population CMA-ES on the BBOB2009 Function Testbed. In Proceedings of the 11th Annual Conference
Companion on Genetic and Evolutionary Computation Conference: Late
Breaking Papers, GECCO ’09, pages 2389–2396, 2009. ISBN 978-160558-505-5.
[23] N. Hansen and S. Finck and R. Ros. COCO - COmparing Continuous
Optimizers : The Documentation. Research Report RT-0409, INRIA,
May 2011.
[24] J. H. Holland. Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial
intelligence. University of Michigan Press, 1975.
[25] M. W. Iruthayarajan and S. Baskar. Evolutionary Algorithms Based
Design of Multivariable PID Controller. Expert Syst. Appl., 36(5):9159–
9167, July 2009. ISSN 0957-4174.
[26] W. M. Iruthayarajan and S. Baskar. Covariance Matrix Adaptation
Evolution Strategy Based Design of Centralized PID Controller. Expert
Syst. Appl., 37(8):5775–5781, August 2010. ISSN 0957-4174.
[27] G. A. Jastrebski and D. V. Arnold. Improving evolution strategies
through active covariance matrix adaptation. In IEEE Congress on
Evolutionary Computation – CEC 2006, pages 2814–2821, 2006.
[28] J. Kennedy, R. C. Eberhart and Y. Shi. Swarm Intelligence. Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA, 2001. ISBN 155860-595-9.
26
[29] W. M. Korani, H. T. Dorrah, and H. M. Emara. Bacterial foraging
oriented by Particle Swarm Optimization strategy for PID tuning. In
Computational Intelligence in Robotics and Automation (CIRA), 2009
IEEE International Symposium on, pages 445–450, Dec 2009.
[30] D.P. Kwok and F. Sheng. Genetic algorithm and simulated annealing
for optimal robot arm PID control. In Evolutionary Computation, 1994.
IEEE World Congress on Computational Intelligence., Proceedings of
the First IEEE Conference on, pages 707–713. IEEE, 1994.
[31] R. T. Marler and J. S. Arora. Survey of multi-objective optimization
methods for engineering. Structural and multidisciplinary optimization,
26(6):369–395, 2004.
[32] K. Ogata. Modern Control Engineering. Prentice Hall PTR, Upper
Saddle River, NJ, USA, 2nd edition, 1990. ISBN 0135891280.
[33] M. Petrov, I. Ganchev and A. Taneva. Fuzzy PID control of nonlinear
plants. Intelligent Systems, 2002. Proceedings. 2002 First International
IEEE Symposium, vol. 1, pp. 30–35. IEEE, 2002.
[34] R. Poli. Analysis of the publications on the applications of particle
swarm optimisation. Journal of Artificial Evolution and Applications,
2008:3, 2008.
[35] K. Price, R. M. Storn and J. A. Lampinen. Differential Evolution: A
Practical Approach to Global Optimization. Springer-Verlag New York,
Inc., Secaucus, NJ, USA, 2005. ISBN 3540209506.
[36] Ricardo Software. WAVE Manual, 2016.
[37] R. Storn. On the usage of differential evolution for function optimization.
In Fuzzy Information Processing Society, 1996. NAFIPS., 1996 Biennial
Conference of the North American, pages 519–523, Jun 1996.
[38] R. Storn and K. Price. Differential evolution – a simple and efficient
adaptive scheme for global optimization over continuous spaces. Technical Report TR-95–012, International Computer Science Institute, Berkeley, 1995.
27
[39] R. Tanabe and A. Fukunaga. Success-History Based Parameter Adaptation for Differential Evolution. IEEE CEC, 2013, pp. 71–78.
[40] R. Tanabe and A. Fukunaga. Evaluating the performance of SHADE
on CEC 2013 benchmark problems. IEEE CEC, 2013, pp. 1952–1959.
[41] R. Tanabe and A. Fukunaga. Tuning Differential Evolution for Cheap,
Medium, and Expensive Computational Budgets. IEEE CEC, 2015, pp.
2018-2025.
[42] Y. Wakasa, S. Kanagawa, K. Tanaka, and Y. Nishimura. Pid controller
tuning based on the covariance matrix adaptation evolution strategy.
IEEJ Transactions on Electronics, Information and Systems, 130(5):
737–742, 2010.
[43] J. G. Ziegler and N. B. Nichols. Optimum Setting for Automatic Controller. ASME transaction, pp. 759–768, 1942.
28
| 9cs.NE
|
High-Dimensional Joint Estimation
of Multiple Directed Gaussian Graphical Models
Yuhao Wang1,2
1
Santiago Segarra2
Caroline Uhler1,2
Laboratory for Information and Decision Systems, Massachusetts Institute of Technology
2 Institute for Data, Systems, and Society, Massachusetts Institute of Technology
arXiv:1804.00778v1 [math.ST] 3 Apr 2018
Abstract
We consider the problem of jointly estimating multiple related directed acyclic graph (DAG) models
based on high-dimensional data from each graph. This problem is motivated by the task of learning
gene regulatory networks based on gene expression data from different tissues, developmental stages or
disease states. We prove that under certain regularity conditions, the proposed `0 -penalized maximum
likelihood estimator converges in Frobenius norm to the adjacency matrices consistent with the datagenerating distributions and has the correct sparsity. In particular, we show that this joint estimation
procedure leads to a faster convergence rate than estimating each DAG model separately. As a corollary
we also obtain high-dimensional consistency results for causal inference from a mix of observational
and interventional data. For practical purposes, we propose jointGES consisting of Greedy Equivalence
Search (GES) to estimate the union of all DAG models followed by variable selection using lasso to
obtain the different DAGs, and we analyze its consistency guarantees. The proposed method is illustrated
through an analysis of simulated data as well as epithelial ovarian cancer gene expression data.
1
Introduction
Directed acyclic graph (DAG) models, also known as Bayesian networks, are widely used to model causal
relationships in complex systems across various fields such as computational biology, epidemiology, sociology, and environmental management [1, 11, 30, 35, 39]. In these applications we often encounter highdimensional datasets where the number of variables or nodes greatly exceeds the number of observations.
While the problem of structure identification for undirected graphical models in the high-dimensional setting is quite well understood [34, 25, 10, 4, 44], such results are just starting to become available for directed
graphical models. The difficulty in identifying DAG models can be attributed to the fact that searching over
the space of DAGs is NP-complete in general [5].
Methods for structure identification in directed graphical models can be divided into two categories
and hybrids of these categories. Constraint-based methods, such as the prominent PC algorithm, first
learn an undirected graph from conditional independence relations and in a second step orient some of
the edges [14, 39]. Score-based methods, on the other hand, posit a scoring criterion for each DAG model,
usually a penalized likelihood score, and then search for the network with the highest score given the observations. An example is the celebrated Greedy Equivalence Search (GES) algorithm, which can be used to
greedily optimize the `0 -penalized likelihood such as the Bayesian Information Criterion (BIC) [6]. Highdimensional consistency guarantees were recently obtained for the PC algorithm [19] and for score-based
methods [23, 28, 43].
Existing methods have focused on estimating a single directed graphical model. However, in many applications we have access to data from related classes, such as gene expression data from different tissues,
cell types or states [24, 37], different developmental stages [2], different disease states [41], or from different
perturbations such as knock-out experiments [8]. In all these applications one would expect that the underlying regulatory networks are similar to each other, since they stem from the same species, individual or cell
1
type, but also have important differences that drive differentiation, development or a certain disease. This
raises an important statistical question, namely how to jointly estimate related directed graphical models in
order to effectively make use of the available data.
Various methods have been proposed for jointly estimating undirected Gaussian graphical models. To
preserve the common structure, Guo et al. [15] suggested to use a hierarchical penalty and Danaher et
al. [7] suggested the use of a generalized fused lasso or group lasso penalty. While both approaches achieve
the same convergence rate as the individual estimators, Cai et al. [3] were able to improve the asymptotic
convergence rate of joint estimation using a weighted constrained `∞ /`1 minimization approach. Bayesian
methods have been proposed for this problem [32] as well. Related works also include [27], where it is
assumed that the networks differ only locally in a few nodes and [21, 38], where the assumption is that the
networks are ordered and related by continuously changing edge weights.
In this paper, we propose a framework based on `0 -penalized maximum likelihood estimation for jointly
estimating related directed Gaussian graphical models. We show that the joint `0 -penalized maximum likelihood estimator (MLE) achieves a faster asymptotic convergence rate as compared to the individual estimators. In addition, by viewing interventional data as data coming from a related network, we show that
the interventional BIC scoring function proposed in [16] can be obtained as a special case of the joint
`0 -penalized maximum likelihood approach presented here. Our theoretical consistency guarantees also explain the empirical findings of [16], namely that estimating a DAG model from interventional data usually
leads to better recovery rates as compared to estimating a DAG model from the same amount of purely
observational data. These theoretical results are based on the global optimum of `0 -penalized maximum
likelihood estimation. To overcome the computational bottleneck of this optimization problem we propose
a greedy approach (jointGES) for solving this problem by extending GES [6] to the joint estimation setting.
We analyze its properties from a theoretical point of view and test its performance on synthetic data and
gene expression data from epithelial ovarian cancer.
The remainder of this paper is structured as follows. In Section 2, we review some relevant background
related to DAG models and introduce notation for the joint DAG estimation problem studied in this paper. In
Section 3, we present the joint `0 -penalized maximum likelihood estimator and jointGES, an adaptation of
GES for solving this optimization problem. Section 4 establishes results regarding the statistical consistency
of the `0 -penalized MLE and jointGES whereas Section 5 presents the implications for learning DAG models
from a mix of observational and interventional data. In Section 6, we illustrate the performance of our
proposal in a simulation study and an application to the analysis of gene expression data. We conclude with
a short discussion in Section 7. The proofs of supporting results are contained in the Appendix.
2
Preliminaries
In Section 2.1 we introduce DAGs, their relation to linear structural equation models, and discuss statistical
features enjoyed by random vectors following these models. In Section 2.2 we briefly review existing
approaches for estimating a single directed graphical model from observational data. Finally, Section 2.3
describes a setting where multiple related directed graphical models exist.
2.1
Directed acyclic graphs and linear structural equation models
Let G = (V, E) denote a DAG with vertices V = [p] = {1, · · · , p} and directed edges E ⊆ V × V , where
|G| denotes the cardinality of E. Let A ∈ Rp×p be the adjacency matrix specifying the edge weights of the
underlying DAG G, i.e., Aij 6= 0 if and only if (i, j) ∈ E. Also, let ∼ N (0, Ω) denote a p-dimensional
2
multivariate Gaussian random variable with zero mean and diagonal covariance matrix Ω. In this work, we
assume that the observed random vector X = (X1 , · · · , Xp ) ∈ Rp is generated according to the following
linear structural equation model (SEM).
X = AT X + .
(1)
Hence X follows a multivariate Gaussian distribution with zero mean and covariance matrix Σ, where the
inverse covariance (or precision) matrix Θ = Σ−1 is given by
Θ = (I − A)Ω−1 (I − A)T .
(2)
Let Paj (G) denote the parents of node j in G, then it follows from (1) that the distribution of X can be
factorized as
P(X) =
p
Y
P(Xj |XPaj (G) ).
j=1
In addition, such a factorization of P according to G is equivalent to the Markov assumption with respect to
G [22, Theorem 3.27]. Formally, given i, j ∈ V and an arbitrary subset of nodes S ⊂ V \ {i, j}, then
j is d-separated from k | S in G
⇒
Xj ⊥
⊥ Xk |XS in P.
(3)
If the implication (3) holds bidirectionally, then P is said to be faithful [39] with respect to G. Note that there
exist DAGs G1 and G2 that encode the same d-separations and hence the same conditional independence
relations. Such DAGs are said to belong to the same Markov equivalence class.
A consequence of the acyclicity of G is that there exists at least one permutation π of [p] such that
Aij = 0 for all π(i) ≥ π(j). Putting it differently, if the rows and columns of A are reordered according
to π, then the resulting matrix is strictly upper triangular. Hence, if such a permutation π is known a priori,
one can obtain the SEM parameters (A, Ω) from Θ according to the following steps [cf. (2)]. First, we
reorder Θ according to π. Then, we perform on the reordered Θ an upper-triangular-plus-diagonal Cholesky
decomposition to obtain (A0 , Ω0 ). Finally, we revert the ordering by permuting the rows and columns of
A0 and Ω0 according to π −1 and obtain the sought (A, Ω). For an arbitrary permutation π and a given Θ,
we denote by (Aπ , Ωπ ) the Cholesky decomposition parameters obtained from the procedure just described.
Alternatively, one can obtain (Aπ , Ωπ ) by solving p linear regressions [cf. (1)]. More precisely, we can
obtain each column of Aπ by regressing Xj only on those Xi such that π(i) < π(j) for all j. Once Aπ is
obtained, one can estimate the variance of in (1) to get Ωπ . In the rest of the paper, we denote by (A0 , Ω0 )
and Θ0 the true parameters of the data-generating SEM and the associated precision matrix, respectively.
Moreover, we denote by (A0π , Ω0π ) the SEM parameters obtained from the described procedure when the
true precision matrix Θ0 is used. Notice that (A0 , Ω0 ) = (A0π , Ω0π ) if π is any permutation consistent
with the true underlying DAG G0 . The DAG Gπ corresponding to the non-zero entries of Aπ is known as
the minimal I-MAP (independence map) with respect to π. The minimal I-MAP with the fewest number of
edges is called minimal-edge I-MAP [43]. If P is faithful with respect to a DAG G, then G is a minimal-edge
I-MAP of P [33, 43].
Furthermore, it has been shown in [31] that all DAGs in a Markov equivalence class share the same
skeleton – i.e., the set of edges when directions are ignored – and v-structures. A v-structure is a triplet
(, k, `) ⊆ V such that (j, k), (`, k) ∈ E but j and ` are not connected in either direction. This motivates
the representation of a Markov equivalence class as a completely partially directed acyclic graph (CPDAG),
3
which is a graph containing both directed and undirected edges. A directed edge means that all DAGs in
the Markov equivalence class share the same direction for this edge whereas an undirected edge means that
both directions for that specific edge are present within the class. In the same way, one can represent a
subset of a Markov equivalence class via a partially directed acyclic graph (PDAG), where the directions
of the edges are only determined by the graphs within the subset. In particular, some undirected edges in
a CPDAG would become directed edges in a PDAG representing a subset of the class. Notice that both
DAGs and CPDAGs are special cases of PDAGs, where the former represents a single graph and the latter
represents the whole equivalence class.
2.2 `0 -penalized maximum likelihood estimation for a single DAG model
We denote by X̂ ∈ Rn×p our observed data, where each row of X̂ represents a realization of the random
vector X. We say that we are in the low-dimensional setting if asymptotically p remains a constant as n →
∞. By contrast, whenever p → ∞ as n → ∞, we say that we are in the high-dimensional setting. Assuming
faithfulness, Chickering [6] shows that GES achieves a consistent estimator in the low-dimensional setting
by optimizing the following objective – also known as the Bayesian information criterion (BIC) –
(Â, Ω̂) := argmax `n (X̂; A, Ω) − λ2 kAk0 ,
(4)
A∈A, Ω∈D+
where λ2 = 21 logn n , A denotes the set of all valid adjacency matrices associated with DAGs, D+ is the set
of non-negative diagonal matrices, and `n is the likelihood function
!
X̂ T X̂
(5)
· (I − A)Ω−1 (I − A)T + log det (I − A)Ω−1 (I − A)T .
`n (X̂; A, Ω) := −trace
n
In the high-dimensional setting, van de Geer and Bühlmann [43] give consistency guarantees for the global
optimum of (4) when the collection A is further constrained to contain only adjacency matrices with at most
d incoming edges for each node, where d = O(n/ log p). More precisely, they show that there exists some
parameter λ2 logn p such that the optimum (Â, Ω̂) in (4) converges in Frobenius norm to (A0π̂ , Ω0π̂ ) for
increasing n and p, where π̂ is a permutation consistent with Â, i.e.,
k − A0π̂ k2F + kΩ̂ − Ω0π̂ k2F = O λ2 |G0 | .
(6)
Notice, however, that (6) does not guarantee statistical consistency since π̂ need not be a permutation consistent with the true underlying DAG. Moreover, (6) does not hold for any permutation π̂ consistent with Â,
but [43] shows the existence of at least one such permutation. In addition, it is shown in [43] that the number
of non-zero elements in Â, A0π̂ , and A0 are all of the same order of magnitude, i.e., |Ĝ| |G0π̂ | |G0 |.
2.3
Collection of DAGs
Consider the setting where not all the observed data comes from the same DAG, but rather from a collection
of DAGs {G (k) = (V, E (k) )}K
k=1 that share the same node set V = [p]. In addition, we assume that all
DAGs in a collection are consistent with some permutation π. This precludes a scenario where (i, j) ∈ E (k)
0
and (j, i) ∈ E (k ) for some k 6= k 0 . This is a reasonable assumption in, e.g., the analysis of gene expression
data, where regulatory links may appear or disappear, but they in general do not change direction.
4
(k) }K
(k) }K
Denote by {(A(k) , Ω(k) )}K
k=1 a set of SEMs on the K DAGs {G
k=1 and by {X̂
k=1 the data
(k)
generated from each SEM, where we observe nk realizations for each DAG G . In this way, each row of
the data matrix X̂ (k) ∈ Rnk ×p corresponds to a realization of the random vector X (k) defined as
T
X (k) = A(k) X (k) + (k)
with (k) ∼ N (0, Ω(k) ).
Collections of DAGs arise for example naturally when considering data from perfect (also known as
hard) interventions [9]. Consider a non-intervened DAG G with SEM parameters (A, Ω) [cf. (1)]. Then a
perfect intervention on a subset of nodes Ik ⊂ V gives rise to the interventional distribution
T
with Ik ∼ N (0, ΩIk ),
X Ik = AIk X Ik + Ik
where AIijk = 0 if j ∈ Ik and AIijk = Aij otherwise, and the diagonal matrix ΩIk satisfies ΩIiik = Ωii if
i 6∈ Ik [16, 17]. We denote the DAG given by the non-zero entries of AIk by G Ik .
(k)
(k)
(k)
In accordance with the notation introduced in Section 2.1, we denote by G0 and (A0 , Ω0 ) the true
data-generating DAG and SEM parameters for class k, respectively, and by π0 a permutation that consistent
(k)
(k)
(k)
(k)
with A0 for all classes k ∈ [K]. Moreover, we denote by G0π and (A0π , Ω0π ) the DAG and SEM
(k)
parameters obtained from the Cholesky decomposition of the true precision matrix Θ0 when permuted by
(k)
(k)
π. We denote by Σ0 the true covariance matrix of the SEM for class k, i.e., the inverse of Θ0 . Finally,
(k)
(k)
union
we define G0union as the union of all G0 – i.e., an edge appears in G0union if it appears in any G0 – and G0π
(k)
Ik
Ik
as the union of G0π . For interventional data, we use (A0 , Ω0 ) to denote the true SEM parameters after
intervening on targets Ik .
3
Joint estimation of multiple DAGs
We first present a penalized maximum likelihood estimator that is the natural extension of (4) for the case
where a collection of DAGs is being estimated. Since this involves minimizing k·k0 , we then discuss a
greedy approach that alleviates the computational complexity of this estimator.
3.1
Joint `0 -penalized maximum likelihood estimator
With d denoting a pre-specified sparsity level and wk = nk /n indicating the proportion of observed data
from DAG k, we propose the following estimator:
n
o
π̂, {(Â(k) , Ω̂(k) )}K
k=1 :=
argmax
π,{(A(k) ,Ω(k) )}K
k=1
K
X
k=1
wk `nk (X̂
(k)
;A
(k)
(k)
,Ω
)−λ
2
K
X
k=1
|A(k) |
(7)
0
subject to A(k) ∈ Aπ , kA(k) k∞,0 ≤ d, Ω(k) ∈ D+ ∀k,
where Aπ is the set of all adjacency matrices consistent with permutation π and the matrix norm k · k∞,0
computes the maximum `0 -norm across the rows of the argument matrix. The optimization problem in (7)
seeks to maximize a weighted log-likelihood of the observations (where more weight is given to SEMs with
more realizations) penalized by the support of the union of all estimated DAGs. To see why this is true,
P
(k)
(k)
notice that k K
k=1 |A |k0 counts the number of (i, j) entries for which Aij 6= 0 for at least one graph k.
This penalization on the union of estimated DAGs promotes overlap in the supports of the different A(k) .
5
Algorithm 1 JointGES for joint `0 -penalized maximum likelihood estimation of multiple DAGs.
Input: Collection of observations X̂ (1) ∈ Rn1 ×p , · · · , X̂ (K) ∈ RnK ×p , sparsity constraint d, penalization
parameters λ1 and λ2
Output: Collection of weighted adjacency matrices Â(1) , · · · , Â(K)
1: Apply GES to find Ĝ union , an approximate solution to the following optimization problem
#
!
"
p
K
X
X
(k)
(k)
argmin
wk
min log kX̂j − X̂Paj (G) ak22 + λ21 |Paj (G)|
|Paj (G)|
G
a∈R
(8)
j=1 k=1
subject to
2:
max |Paj (G)| ≤ d
j
union by solving Kp sparse reEstimate the weighted adjacency matrices {Â(k) }K
k=1 consistent with Ĝ
gressions of the form
(k)
âj
=
1
(k)
kX̂j − X̂ (k) ak22 + λ22 kak1 .
n
union
k
a | supp(a)⊆Paj (Ĝ
)
argmin
Regarding the constraints in (7), the first constraint imposes that all estimated DAGs are consistent with
the same permutation π, which is itself an optimization variable. This constraint is in accordance with our
assumption in Section 2.3 and drastically reduces the search space of DAGs. The second constraint ensures
that the maximum in-degree in all graphs is at most d, and the last constraint imposes the natural requirement
that all noise covariances are diagonal and non-negative.
Notice that (7) is a natural extension of (4). Indeed, for the case K = 1 the objective in (7) immediately
boils down to that in (4). Moreover, when there is only one graph and π can be selected freely, the constraint
A(1) ∈ Aπ is effectively identical to A(1) ∈ A, i.e., the constraint in (4). Finally, observe that in (7)
we have included the additional maximum in-degree constraint required in the high-dimensional setting
[cf. discussion after (5)].
3.2
JointGES: Joint greedy equivalence search
The `0 norms as well as the optimization over all permutations π render the problem of (7) non-convex, thus,
hard to solve efficiently. In this section, we present a greedy approach to find a computationally tractable
approximation to a solution to (7). The algorithm, which we term JointGES, is succinctly presented in
Algorithm 1 and consists of two steps.
In the first step of Algorithm 1 we recover Ĝ union , our estimate of the union of all the DAGs to be
inferred. We do this by finding an approximate solution to (8) via the implementation of GES [6]. The
objective (scoring function) in (8) consists of two terms. The first term is given by the sum of the log(k)
likelihoods of the achievable residues when regressing the jth column of X (k) , denominated as Xj , on
(k)
XPaj (G) for each node j and DAG k. In [43], van de Geer and Bühlmann further show that if we keep
the underlying DAG G fixed,
estimator proposed in (4) is equivalent to optimizing
the maximum likelihood
Pp
2
. Thus, the first term in (8) corresponds to the first term
j=1 mina∈R|Paj (G)| log kX̂j − X̂Paj (G) ak2
in the objective of (7). The second term penalizes the size of the parent set of each node in the graph to
be recovered, effectively penalizing the number of edges in the graph. In this way, the scoring function in
6
(8) promotes a sparse G in the same way that the objective of (7) promotes the union of all K recovered
graphs to have a sparse support. Additionally, it is immediate to see that the scoring function in (8) is
decomposable [6], a key feature that enables the implementation of GES to find an approximate solution.
Once we have obtained the union of all sought DAGs Ĝ union from step 1, in the second step of our algorithm
we estimate the DAGs Ĝ (1) , · · · , Ĝ (K) by searching over the subDAGs of Ĝ union . More precisely, for each
(k)
(k)
node j we estimate its parents in Ĝ (k) by regressing Xj on X
using lasso, where the support of
union
Paj (Ĝ
(k)
âj
)
Ĝ (k) .
corresponds to the set of parents of j in
To summarize, Algorithm 1 recovers K DAGs by first estimating the union of all these DAGs Ĝ union
using standard GES and then inferring the specific weight adjacency matrices Â(k) via a lasso regression,
while ensuring consistency with the previously estimated Ĝ union .
4
Consistency guarantees
The main goal of this section is to provide theoretical guarantees on the consistency of the solution to
Problem (7) in the high-dimensional setting. Our main result is presented in Theorem 4.9. In Section 4.3
we present a laxer statement of consistency based on milder conditions.
4.1
Statistical consistency of the joint `0 -penalized MLE
A series of conditions must hold for our main result to be valid. We begin by stating these conditions
followed by the formal consistency result in Theorem 4.9. The rationale behind these conditions and their
implications are discussed after the theorem in Section 4.2.
(1)
(K)
Condition 4.1. All DAGs G0 , · · · , G0
are minimal-edge I-MAPs.
Condition 4.2. There exists a constant σ02 that bounds the variance of all the observed processes, i.e.,
(k)
maxk,i [Σ0 ]ii ≤ σ02 .
(k)
(k)
Condition 4.3. The smallest eigenvalues of all Σ0 are non-zero, i.e. mink Λmin (Σ0 ) = Λmin > 0.
Condition 4.4. There exists some constant α such that, for all k, the maximum allowable in-degree d is
bounded as d ≤ αnk / log p.
Condition 4.5. For all π and j there exist some constants α̃ and cs > 2 such that
)!
(
1
3
n
n
union
, 7
.
|Paj (G0π )| + cs ≤ α̃ min
K 7 (log p)3
K (log n)2 log p
Condition 4.6. The number of DAGs K satisfies K = o(log p) and the amount of data associated with each
DAG is comparable in the sense that n1 n2 · · · nK .
PK
(k)
union | ≤ c
Condition 4.7. There exists some constant ct > 0 such that |G0π
t
k=1 wk |G0π | for any permutation π.
Condition 4.8. There exist constants η1 and η2 such that 0 ≤ η1 < 1, 0 < η02 < (1 − η1 )/ct , and
(
p
q
)
X
log
p/n
(k)
(k)
1 [A0π ]i,j >
p/|G0union | ∨ 1
≥ (1 − η1 )|G0π |,
η0
i,j
7
(9)
for all permutations π and graphs k ∈ [K], where 1{·} denotes the indicator function and ct is as in
Condition 4.7.
With the above conditions in place, the following result can be shown.
Theorem 4.9. If Conditions 4.1-4.8 hold and λ is chosen such that λ2
log p
n
p
|G0union |
∨ 1 , then there exists
a constant c > 0 , that depends on cs , such that with probability 1 − exp(−cp) the solution to (7) satisfies
K
X
(k)
wk kÂ
−
(k)
A0π̂ k2F
+
k=1
K
X
(k)
wk kΩ̂(k) − Ω0π̂ k2F = O λ2 |G0union | .
(10)
k=1
Furthermore, denoting by Ĝ the union of the graphs Ĝ (k) associated with the K recovered adjacency matrices Â(k) , we have that
union
|Ĝ| |G0π̂
| |G0union |.
(11)
The proof of Theorem 4.9 is given in Appendix A.2. To intuitively grasp the result in the above theorem,
assume that the number of edges in G0union is proportional to the number of nodes p so that λ2 |G0union | → 0 for
increasing n as long as n > p log p. Hence, under these conditions, (10) guarantees that for the recovered
(k)
permutation π̂, the estimated adjacency matrix Â(k) converges to A0π̂ in Frobenius norm for all k. This not
only implies that both adjacency matrices have similar structure, but also that the edge weights are similar.
Moreover, from (11) it follows that the number of edges in the estimated graph Ĝ, i.e., |Ĝ| is similar to the
union |. More importantly,
number of edges in the union of all minimal I-MAPs with permutation π̂, i.e., |G0π̂
union
|Ĝ| is also similar to the number of edges in the true union graph |G0 |. Despite these guarantees, it should
be noted that the permutation π̂ need not coincide with the permutation π of the true graphs to be recovered.
We now assess the benefits of performing joint estimation of the K DAGs as opposed to estimating
them separately. To do so, we compare the guarantees in Theorem 4.9 to those developed in [43] for
separate estimation. More precisely, the application of the consistency bound reviewed in (6) yields that for
the separate estimation of K DAGs, when we are in the
setting where
all K DAGs are highly overlapping
log p
p
2
(cf. Condition 4.7), by choosing λ such that λ n
∨ 1 , one can guarantee that
|G union |
0
K
X
wk kÂ
(k)
−
(k)
A0π̂(k) kF
k=1
+
K
X
(k)
wk kΩ̂(k) − Ω0π̂(k) k2F = O Kλ2 |G0union | ,
(12)
k=1
where it should be noted that in the separate estimation the recovered permutation π̂ (k) can vary with k.
A direct comparison of (10) and (12) reveals that performing joint estimation improves the accuracy by
a factor of K from Ω(K logn p ) to Ω( logn p ). Hence, for joint estimation the accuracy scales with the total
number of samples n, showing that our procedure yields maximal gain from each observation, even if the
data is generated from K different DAGs. Moreover, the result in (10) holds under slightly milder conditions
than those needed for (12) to hold since Condition 4.8 is a relaxed version of the beta-min condition in [43].
A more detailed discussion about the conditions of Theorem 4.9 is given next.
4.2
Conditions for Theorem 4.9
It has been shown in [33] that if a data-generating distribution is faithful with respect to G, then G must
be a minimal-edge I-MAP. By enforcing the latter for every true graph, Condition 4.1 imposes a milder
8
requirement compared to the well-established faithfulness assumption [39]. Conditions 4.2-4.4 ensure that
we avoid overfitting and provide bounds for the noise variances. These are direct adaptations from Conditions 3.1-3.3 in [43]. Condition 4.5 is required to bound the difference between the sample variances of our
observations and the true variances, and is related to Condition 3.4 in [43] butadapted to our joint inference
n1/3
setting. Notice that Condition 4.5 is trivially satisfied when p = O K 7/3
. Condition 4.6 follows from
log n
the bounds for sample variances shown in [3]. Intuitively, we are imposing the natural restriction that the
number of DAGs is small compared to the number of nodes p in each DAG and the total number of observations n. Moreover, given that our objective is to draw estimation power from the joint inference of multiple
graphs, we require that each DAG is associated with a non-vanishing fraction of the total observations.
Condition 4.7 enforces that, for every permutation π, the number of edges in the union of all recovered
graphs is proportional to the weighted sum of the edges in every graph as K → ∞. In particular, this requires
P
(k)
(k)
the individual graphs G0π to be highly overlapping. To see why this is the case, notice that K
k=1 wk |G0π | is
(k)
upper bounded by the maximum number of edges across graphs G0π . Consequently, Condition 4.7 enforces
the number of edges in the union of graphs to be proportional to the number of edges in the single graph
with most edges, thus requiring a high level of overlap. Imposing high overlap for all permutations π might
seem too restrictive in some settings. Nonetheless, Condition 4.7 can sometimes be derived from apparently
less restrictive conditions as the following example illustrates.
P
(k)
Consider the more relaxed bound |G0union | ≤ ct K
k=1 wk |G0 |, which is equivalent to requiring Condition 4.7 to hold but only for permutations consistent with the true graph G0union . In the following example, we
show that this might be sufficient for Condition 4.7 to hold. Suppose that G0union consists of two connected
respectively defined on the subsets of nodes V1 and V2 . Moreover, assume
and G 00 union
components G 0 union
0
0
(k)
(k)
that the subDAGs of G0 over V1 (denoted by G 0 0 ) are identical for all k. Putting it differently, the differ(k)
ences between the DAGs G0 are limited to the second connected component. In addition, assume that for
0 union |. Then, for any permutation π,
all possible permutations π2 of nodes V2 we have that |G 00 union
0π2 | ≤ |G 0
where we denote by π1 (respectively π2 ) the restriction of π to the node set V1 (respectively V2 ), we have
that
union
union
union
|G0π
| = |G 0 0π1 | + |G 00 0π2 | ≤
K
X
(k)
wk |G 0 0π1 | +
K
X
k=1
k=1
(k)
wk |G 0 0 | ≤ 2
K
X
(k)
wk |G0π |,
k=1
from where Condition 4.7 is satisfied for ct = 2. This example shows that learning the structure of large
components that are common across the different DAGs is not affected by the changes in the smaller components of these DAGs. Despite the previous example, Condition 4.7 might still be too restrictive in some
settings. In this respect, a more relaxed requirement is stated in Section 4.3 under Condition 4.7’.
Condition 4.8 requires that, for every permutation π and every graph k, the value of at least a fixed
(k)
proportion (1 − η1 ) of the edges in G0π is above the ‘noise level’, i.e., the lower bound within the indicator
function in (9). Intuitively, if the true weight of many edges is close to zero then correct inference of the
graphs would be impossible since the true edges would be mistaken with spurious ones. Thus, it is expected
that the weights of a sufficiently large fraction of the edges have to be sufficiently large. Condition 4.8 is
the right formalization of this intuition. Moreover, notice that a straightforward replication
of the beta-min
p
condition introduced
in
[43]
would
have
required
the
‘noise
level’
to
scale
with
log
p/n
k , instead of the
p
smaller scaling of log p/n required in (9). In this sense, Condition 4.8 is a relaxed version of the extension
of the beta-min condition to the setting of joint graph estimation.
9
4.3
Consistency under milder conditions
As previously discussed, in some settings Condition 4.7 might be too restrictive. Hence, in this section we
present a consistency statement akin to Theorem 4.9 that holds for a milder version of Condition 4.7. More
precisely, consider the alternative condition stated next.
Condition 4.7’. Let ct (π) be some function of π that scales as a constant for permutations consistent with
PK
(k)
union | ≤ c (π)
G0union and scales as o(K) for all other permutations such that |G0π
t
k=1 wk |G0π | for all π.
Observe that for permutations π consistent with the true union graph G0union , Condition 4.7’ boils down
to the previously discussed Condition 4.7. However, for all other permutations, ct (π) need not be a constant
and is allowed to grow with K. Intuitively, for all permutations not consistent with G0union we are not
(k)
requiring a high level of overlap among all the graphs G0π . Nonetheless, since ct (π) = o(K) we do require
union to be ‘sparser’ than the extreme case in which all graphs G (k) are disjoint.
G0π
0π
In order to account for the fact that ct depends on the permutation π in Condition 4.7’, we have to modify
Condition 4.8 accordingly, resulting in the following alternative statement.
Condition 4.8’. Let Cmax := max ct (π), then there exist constants η1 and η2 such that 0 ≤ η1 < 1,
π
0 < η02 < (1 − η1 ), and
(
p
)
q
X
Cmax log p/n
(k)
(k)
≥ (1 − η1 )|G0π |,
1 [A0π ]i,j >
p/|G0union | ∨ 1
η0
i,j
for all permutations π and graphs k, where 1{·} denotes the indicator function.
The following consistency result holds for the alternative set of conditions.
Theorem 4.10. Under Conditions 4.1-4.6, 4.7’, and 4.8’ and with λ such that λ2 Cmax logn p
p
|G0union |
∨1 ,
then there exists a constant c > 0 that depends on cs such that with probability 1 − exp(−cp), the solution
to (7) satisfies that, at least for one k ∈ [K],
(k)
(k)
(k)
kÂ(k) − A0π̂ k2F + kΩ̂(k) − Ω0π̂ k2F = O λ2 |G0 | .
(13)
Furthermore, denoting by Ĝ (k) the graph associated with Â(k) for the k ∈ [K] satisfying (13), we have that
(k)
(k)
|Ĝ (k) | |G0π̂ | |G0 |.
(14)
The proof is given in Appendix A.3. Condition 4.7’ is milder than Condition 4.7 and this relaxation
entails a corresponding loss in the guarantees of recovery: Comparing (13) and (14) with (10) and (11)
immediately reveals that what could be guaranteed for the ensemble of graphs in Theorem 4.9 can only be
guaranteed for a single graph in Theorem 4.10, thereby explaining the trade-off in relaxing the conditions.
On the other hand, the result in Theorem 4.10 still draws inference power from the joint estimation of
multiple graphs since neither (13) nor (14) can be shown using existing results for separate estimation. To
be more precise, as discussed in Section 4.2, when performing separate estimation, theoretical guarantees
are based on thep
assumption that at least a fixed proportion of the edge weights are above the
p ‘noise level’,
which scales as log p/nk . However, Condition 4.8’ requires the noise level to scale with Cmax log p/n
which, given the fact that Cmax = o(K), is not large enough to achieve the guarantee needed for separate
estimation. In addition, the convergence rate of Ω(Cmax logn p ) in (13) is still faster than the corresponding
10
convergence rate of Ω(K logn p ) associated with separate estimation [cf. discussion after (12)]. We close
Section 4 with the following remark on the consistency guarantees of jointGES.
P
log nk
Remark 4.11 (Consistency of jointGES). In the low-dimensional setting, by choosing λ21 = K
k=1 wk 2nk
and assuming faithfulness, it follows from [6] that the first step in Algorithm 1 is guaranteed to recover the
true Markov equivalence class of G0union , denoted by M(G0union ), in the limit of large data. This allows us
(k)
union ) as inputs to the second step of
to recover {G0 }K
k=1 by successively considering all DAGs in M(G0
union
union
Algorithm 1, and then selecting the DAG G
∈ M(G0 ) whose output from step 2 is the sparsest.
For the high-dimensional setting, as even the global optimum of (7) is not guaranteed to recover the true
G0union (cf. Theorems 4.9 and 4.10), jointGES is not consistent in general. Recently, Maathuis et al. [28]
showed consistency of GES for single DAG estimation under more restrictive assumptions than the ones
here considered. Although of potential interest, further strengthening the presented conditions to guarantee
consistency of jointGES in the high-dimensional setting is not pursued in the current paper.
5
Extension to interventions
We show how our proposed method for joint estimation can be extended to learn DAGs from interventional
data. It is natural to consider learning from interventional data as a special case of joint estimation since
the DAGs associated with interventions are different but closely related. In this section we mimic some of
the developments of Sections 3 and 4 but specialized for the case of interventional data. More precisely,
we first propose an optimization problem akin to (7) and then state the consistency guarantees in the highdimensional setting of the associated optimal solution.
Recall from Section 2.3 that the true adjacency matrix AI0k of the SEM associated with an intervention
in the nodes Ik is identical to the true adjacency A0 of the non-intervened model except that [AI0k ]ij = 0
for all j ∈ Ik . In this way, our assumption that there exists a common permutation π consistent with all
DAGs under consideration (cf. Section 2.3) is automatically satisfied for interventional data. Additionally,
assuming that we observe samples X̂ Ik from K different models corresponding to the respective intervention
of the nodes in {Ik }K
k=1 , the knowledge of the intervened nodes can be incorporated into our optimization
problem as follows [cf. (7)].
o
n
π̂, Â, Ω̂, {(ÂIk , Ω̂Ik )}K
k=1 =
K
X
argmax
π,A,Ω,{(AIk ,ΩIk )}K
k=1
subject to
wk `nk (X̂ Ik ; AIk , ΩIk ) − λ2 kAk0
(15a)
k=1
A ∈ Aπ , kAk∞,0 ≤ d, Ω ∈ D+ ,
(15b)
AIijk
ΩIjjk
= 0 ∀j ∈ Ik ,
(15c)
∈ D+ .
(15d)
= Aij ∀j 6∈ Ik ,
AIijk
Ik
= Ωjj ∀j 6∈ Ik , Ω
From the solution of (15) we obtain an estimate for the non-intervened SEM (Â, Ω̂) as well as K estimates
for the corresponding intervened models (ÂIk , Ω̂Ik ). The objective in (15a) is equivalent to that in (7) where
we leverage the fact that the union of all intervened graphs results in the non-intervened one under the
implicit assumption that no single node has been intervened in every experiment. Alternatively, if some
nodes were intervened in all experiments, objective (15a) would still be valid since enforcing zeros in the
unobservable portions of A does not affect the recovery of the intervened adjacency matrices AIk . The
constraints in (15b) impose that A has to be consistent with permutation π and with bounded in-degree, and
Ω has to be a valid covariance matrix for uncorrelated noise. Putting it differently, (15b) enforces for the
11
non-intervened SEM what we impose separately for all SEMs in (7). The constraints in (15c) impose the
known relations between the intervened and the non-intervened adjacency matrices. Finally, (15d) constrains
the matrices ΩIk to be consistent with the base model on the non-intervened nodes while still being a valid
covariance on the intervened ones.
Even though it might seem that in (15) we are estimating K + 1 SEMs (the base case plus the K
intervened ones), from the previous reasoning it follows that the effective number of optimization variables
is significantly smaller. To be more specific, for a given π, once A is fixed then all the adjacency matrices AIk
are completely determined. Moreover, for a fixed Ω, the only freedom in ΩIk corresponds to the diagonal
entries associated with intervened nodes in Ik . In this way, it is expected that for a given number of samples,
the joint estimation of K SEMs obtained from interventional data [cf. (15)] outperforms the corresponding
estimation from purely observational data [cf. (7)].
Recalling that we denote by (AI0π̂k , ΩI0π̂k ) the parameters recovered from the Cholesky decomposition of
the true precision matrix ΘI0k under the assumption of consistency with permutation π̂, the following result
holds.
Corollary 5.1. If Conditions 4.1-4.8 hold and λ is chosen as λ2 logn p |Gp0 | ∨ 1 , then there exist constants
c1 , c2 > 0 such that with probability 1 − c1 exp(−c2 p), the solution to (15) satisfies
K
X
wk kÂIk − AI0π̂k k2F = O(λ2 |G0 |).
(16)
k=1
Furthermore, denoting by Ĝ the graph associated with the recovered adjacency matrix  for the nonintervened model, we have that
|Ĝ| |G0π̂ | |G0 |.
(17)
The proof is given in Appendix A.4. A quick comparison of Theorem 4.9 and Corollary 5.1 seems to
indicate that the consistency guarantees of observational and interventional data are very similar. Nonetheless, recovery from interventional data is strictly better as we argue next. As discussed after Theorem 4.9,
the results presented do not guarantee that the permutation π̂ recovered coincides with the true permutation
of the nodes. In principle, one could recover a spurious permutation π̂ (different from the true permutation
π) that correctly explains the observed data [cf. (10) and (16)] and leads to sparse graphs [cf. (11) and (17)].
However, the more interventions we have, the smaller the set of spurious permutations π̂ that can be recovered, as we illustrate in the following example. Figure 1 portrays the existence of a spurious permutation
that could be recovered from observational data but cannot be recovered from interventional data. More
precisely, Figure 1(a) presents the two true DAGs we aim to recover, where the second one is obtained
by intervening on node 2. By contrast, Figure 1(b) shows the DAGs that are obtained when performing
Cholesky decompositions on the true precision matrices under the spurious permutation π1 . Notice that the
sparsity levels of the DAGs in both figures are the same. In general, one could recover π1 instead of π0 from
observational data, but one would never recover π1 from interventional data. To see this, simply notice from
the figure that [AI0π2 1 ]32 6= 0 whereas for the interventional estimate [ÂI2 ]32 = 0 [cf. (15c)], thus, the error
terms in (16) cannot vanish for π̂ = π1 . This example also indicates that it is preferable to intervene on
multiple targets in the same experiment instead of doing interventions one at a time. This observation is in
accordance with new genetic perturbation techniques, such as Perturb-seq [8].
From a practical perspective, the objective in (15) corresponds to the same scoring function as GIES [16].
Therefore, GIES can be used to obtain an approximate solution to (15). A simulation study using GIES was
12
π0 = (1, 2, 3)
I1 = ;
I 2 = {2}
1
I
A 01 :
π1 = (3, 2, 1)
I1 = ;
3
1
I
A 02 :
2
3
I 2 = {2}
I
1
A 0π
:
1
2
3
3
I
2
A 0π
:
1
2
1
(a)
2
1
(b)
Figure 1: Interventional data can avoid the recovery of spurious permutations. (a) True DAGs to be recovered. (b) DAGs obtained from the Cholesky decomposition consistent with π1 . The spurious permutation
π1 does not satisfy (16) for cases where node 2 is intervened.
performed in [16, Section 5.2] showing that in line with the theoretical results obtained in this section, not
only identifiability increases, but also the estimates obtained using interventional data are better than with
the same amount of purely observational data.
6
Experiments
In this section, we present numerical experiments on both synthetic (Section 6.1) and real (Section 6.2) data
that support our theoretical findings.
6.1
Performance evaluation of joint causal inference
We analyze the performance of the joint recovery of K different DAGs where we vary K ∈ {3, 5}. For
all experiments, we set the number of nodes p = 100 and the total number of observations n = 600. In
addition, we selected the number of samples from each DAG to be the same, i.e., n1 = . . . = nK = n/K.
For each experiment, the true DAGs were constructed in two steps. First, we generated a core graph that is
shared among the K DAGs under consideration. We did this by generating a random graph from an ErdősRényi model with 100 edges in expectation, and then oriented the edges according to a random permutation
of the nodes. In the second step, we sampled 30 additional edges specific to each DAG uniformly at random.
(1)
(K)
This procedure results in the generation of a collection of true underlying DAGs G0 , . . . , G0 . Associated
(k)
(k)
with each DAG, we generated a true adjacency matrix A0 and a true diagonal error covariance Ω0 . For
the latter, we drew each error variance independently and uniformly from the interval [1, 2.25]. Regarding
the adjacency matrices, we drew the edge weights independently and uniformly from [−1, −0.1] ∪ [0.1, 1] to
ensure that they are bounded away from zero. Notice that we did not put any constraints on the edge weights
that are in the shared core structure for different DAGs: the same edge can change its weight in different
DAGs, or even flip sign.
For each setting, we randomly generated 100 collections of DAGs and data associated with them. We
then estimated the DAGs from the data via two different methods: a joint estimation procedure using the
jointGES algorithm presented in Section 3.2 and a separate estimation procedure using the well-established
GES method [6].
To assess performance of the two algorithms, we considered two standard measures, namely the structural Hamming distance (SHD) [42] and the receiver operating characteristic (ROC) curve. SHD is a com13
(a) Average SHD
(b) Average ROC curve
Figure 2: (a): Average SHD as a function of the scaling constant c for joint and separate GES; (b): Average
ROC curve obtained by varying the tuning parameters. JointGES consistently achieves a better performance
for both values of K.
monly used metric based on the number of operations needed to transform the estimated DAG into the true
one [19, 42]. Hence, a smaller SHD value indicates better performance. The ROC curve plots the true
positive rate against the false positive rate for different choices of tuning parameters.
In Figure 2 (a), we selected the `0 -penalization parameter λ21 = c logn p with scaling constant c ∈
{1, 2, 3, 4, 5} in both joint and separate estimation and then plotted average SHD as a function of the scaling
constant c averaged over the K DAGs to be recovered and the 100 realizations. The penalization parameter
λ2 in the second step of the joint estimation procedure was chosen based on 10-fold cross validation. In
Figure 2 (b) we plotted the average ROC curve where for each choice of tuning parameter, we computed the
true positive and false positive rates by averaging over the K DAGs to be recovered and the 100 realizations.
It is clear from the two figures that in general joint inference achieves better performance, which matches
our theoretical results in Section 4. However, Figure 2 (a) shows also that jointGES performs worse than
separate estimation for small scaling constants (c = 1). Note that this is in line with our theoretical findings
in Theorem 4.10, which imply that whenever Condition 4.7 – which sometimes is a restrictive assumption –
is violated, we need to choose a larger penalization parameter.
6.2
Gene regulatory networks of ovarian cancer subtypes
To assess the performance of the proposed joint `0 -penalized maximum likelihood method on real data, we
analyzed gene expression microarray data of patients with ovarian cancer [41]. Patients were divided into six
subtypes of ovarian cancer, labeled as C1-C6, where C1 is characterized by significant differential expression
of genes associated with stromal and immune cell types and with a lower survival rate as compared to the
other 5 subtypes. We hence grouped the subtypes C2-C6 together and our goal was to infer the differences
in terms of gene regulatory networks in ovarian cancer that could explain the different survival rates. The
gene expression data in [41] includes the expression profile of n = 83 patients with C1 subtype and n = 168
patients with other subtypes. We implemented our jointGES algorithm (Algorithm 1) to jointly learn two
gene regulatory networks: one corresponding to the C1 subtype GC1 and another corresponding to the
other five subtypes together GC2−6 . In addition, as in [3], we focused on a particular pathway, namely the
apoptosis pathway. Using the KEGG database [20, 29] we selected the genes in this pathway that were
associated with at most two microarray probes, resulting in a total of p = 76 genes.
14
Method
JointGES
GES
PC
|GC1 |
50
68
14
|GC2−C6 |
73
101
30
|GC1 ∩ GC2−C6 |
48
32
9
Table 1: Number of edges in the DAGs estimated by different methods for the gene regulatory network of
subtype C1 (|GC1 |) and all other subtypes (|GC2−6 |). The last column shows the number of edges shared
between both inferred graphs.
Table 1 lists the number of edges discovered by jointGES as well as for two separate estimation methods, namely using the GES [6] and PC [39] algorithms. All three methods were combined with stability
selection [26] in order to increase robustness of the output and provide a fair comparison. As expected, the
two graphs inferred using jointGES share a significant proportion of edges, whereas the overlap is markedly
smaller for the two separate estimation methods. Interestingly, under all estimation methods the network for
the C1 subtype contains fewer edges than the network of the other subtypes, thereby suggesting that GC1
could lack some important links that are associated with patient survival.
To obtain more insights into the relevance of the obtained networks, we analyzed the inferred hub nodes
in the three networks. For our analysis we defined as hub nodes those nodes, where the sum of the in- and
out-degree was larger than 5 in the union of the two DAGs. For jointGES, this union is given by Ĝ union , the
graph identified in the first step of Algorithm 1. The hub nodes identified by jointGES are CAPN1, CTSD,
LMNB1, CSF2RB, BIRC3. Among these, CAPN1 [12], CTSD [40], LMNB1 [36], and BIRC3 [18] have
been reported as being central to ovarian cancer in the existing literature. In addition, CSF2RB was also
discovered by joint estimation of undirected graphical models on this data set [3]. The hub nodes discovered
by GES are ATF4, BIRC2, CSF2RB, TUBA1C, MAPK3, while PC only discovered the hub node CSF2RB.
While we were not able to validate the relevance of any of these genes for ovarian cancer in the literature,
interestingly, CSF2RB has been identified as a hub node by all methods, thereby suggesting this gene as an
interesting candidate for future genetic intervention experiments.
7
Discussion
In this paper we presented jointGES, an algorithm for the joint estimation of multiple related DAG models
from independent realizations. Joint estimation is of particular interest in applications where data is collected
not from a single DAG, but rather multiple related DAGs, such as gene expression data from different tissues,
cell types or from different interventional experiments. JointGES first estimates the union of DAGs Ĝ union
by applying a greedy search to approximate the joint `0 -penalized maximum likelihood estimator and then
it uses variable selection to discover each DAG as a subDAG of Ĝ union . From an algorithmic perspective,
jointGES is to the best of our knowledge the first method to jointly estimate related DAG models in the
high-dimensional setting. From a theoretical perspective, we presented consistency guarantees on the joint
`0 -penalized maximum likelihood estimator. and showed that the accuracy bound scales with the total
number of samples, rather than the number of samples from each DAG. As a corollary to this result, we
obtained consistency guarantees for `0 -penalized maximum likelihood estimation of a causal graph from a
mix of observational and interventional data. Finally, we validated our results via numerical experiments on
simulated and real-world data, showing that the proposed jointGES algorithm for joint inference outperforms
separate-inference approaches using well-established algorithms such as PC and GES.
The present work serves as a platform for the potential development of multiple future directions. These
15
directions include: i) relaxing the assumption that all DAGs must be consistent with the same underlying
permutation; ii) extending jointGES to the setting where the samples come from K related DAGs but it
is unknown a priori which particular DAG each sample comes from; this is for example of interest in the
analysis of gene expression data from tumors or tissues that consist of a mix of cell types; iii) extending
jointGES to allow for latent confounders.
A
Appendix: Theoretical analysis of statistical consistency results
Throughout the appendix we develop the proofs of Theorems 4.9 and 4.10. To facilitate understanding, we
first provide a high-level explanation of the rationale behind the proof. If we have data generated from a
single DAG and we are given a permutation π consistent with the true DAG a priori, then we can estimate
(Â, Ω̂) by performing p regressions as explained in Section 2.1. By contrast, when the permutation is
unknown and we need to consider all the possible permutations, the total number of regressions to run
increases to p · p!. However, these regressions are not independent and, intuitively, by bounding the noise
level of a subset of these regressions, we can derive bounds for the noise on the other ones. We characterize
the ‘noise level’ of these regressions by analyzing the asymptotic properties of three random events. More
precisely, whenever these events hold – and we show that they hold with high probability –, the noise is small
enough so that the error of the `0 -penalized maximum likelihood estimator can also be bounded. Finally, we
use this upper bound in the error to show that the recovered graph converges to a minimal I-MAP that is as
sparse as the true DAG.
The remainder of this appendix is organized as follows. In Section A.1 we define the three random
events previously mentioned and show that each of them holds with high probability. Section A.2 then
leverages the definition of these events to show Theorem 4.9, our main result. In Section A.3 we prove
Theorem 4.10, which is based on relaxed conditions but similar proof techniques as in the previous section. Finally, Section A.4 fleshes out the proof of Corollary 5.1, our result applicable to the setting for
interventional data.
Throughout the appendix, we use the following notation. Let âj denote the j-th column of  and ω̂j
denote the j-th diagonal entry of Ω̂. Also, denote by a0jπ and ω0jπ the j-th column of A0π and the j-th
diagonal entry of Ω0π , respectively.
A.1
Random events
Our proofs of Theorems 4.9 and 4.10 are inspired by the main proof in [43] and, thus, also based on a
set of random events. However, the events considered here differ from those in [43] since, as explained in
Section 4.1, a naive application of the guarantees in [43] to the joint estimation scenario does not achieve
optimal learning rates [cf. discussion after (12)]. Intuitively, the rate gain achieved here comes from the
assumption that all the DAGs share the same permutation, allowing us to effectively reduce the size of the
search space.
In our proofs we consider three random events E1 , E2 , and E3 that are respectively stated – along with
proofs showing that they hold with high probability – in Sections A.1.1, A.1.2, and A.1.3.
Random event E1
A.1.1
(k)
(k)
Let jπ ∈ Rn denote the residual when regressing Xj
(k)
Xj
T (k)
(k)
(k)
(k)
on XS with S = {i | π(i) < π(j)}, i.e., jπ :=
− X (k) a0jπ . Similarly, let ˆjπ ∈ Rn denote the regression residual from the sampled data, i.e.,
16
(k)
(k)
ˆjπ := X̂j
(k)
− X̂ (k) a0jπ . Consider a generic set {A(k) }K
k=1 of adjacency matrices consistent with a given
(k)
(k)
permutation π, where the columns of A(k) are denoted by A(k) := (a1 , . . . , ap ), and let G union denote the
union of the support of A(1) , . . . , A(K) . Then, event E1 is defined as
( p K
p X
K
X X wk (k)T
X
2
wk
(k)
(k)
(k)
(k) (k)
E1 := 2
ˆjπ X̂ (aj − a0jπ ) ≤ δ1
X̂ (k) (aj − a0jπ )
n
n
2
(18)
j=1 k=1 k
j=1 k=1 k
o
union
+ λ21 |G union | + |G0π
| /δ1 , ∀ permutations π , and ∀ {A(k) }K
k=1 consistent with π ,
p
for some constant δ1 > 0 and some λ1 (log p)/n. On random event E1 a uniform inequality holds
(k)
across all K DAGs for the sample correlation between the regression residual jπ and any random variable
(k)
T
spanned by the random vector XS , i.e., X (k) v for any v ∈ Rp with vi = 0 for all i 6∈ S. Notice that
(k)
(k)
for convenience for further steps of the analysis, this generic vector v is written as aj − a0jπ in (18).
(k)
Furthermore, for simplicity in the rest of this appendix, we denominate the space spanned by XS as the
(k)
projection space of jπ . Intuitively, one could foresee E1 holding since the expected correlation between the
(k)
regression residual jπ and XS is equal to zero, and therefore the sample correlation can be upper bounded
by a sum of terms that converge to zero as n → ∞ as in (18).
We now show that random event E1 holds with high probability, a result stated in Theorem A.2. Essential
T (k)
(k)
towards proving this result is the observation that, since the random variable X (k) (aj − a0jπ ) lies within
(k)
the projection space of jπ , these two random variables are independent. We can therefore deal with the
(k)
(k)
randomness in ˆjπ and X̂S separately, one at a time. To formally leverage this observation, we rely on
Lemma 7.4 of [43], stated next for completeness.
Lemma A.1. [43, Lemma 7.4] Let Z be a fixed n × m matrix and e1 , · · · , en be independent N (0, σe2 )distributed random variables. Then for all t > 0
!
p
p
P
sup |eT Za|/n ≥ σe ( 2m/n + 2t/n) ≤ exp(−t).
kZak22 /n≤1
Based on the above lemma and recalling that Aπ denotes the set of adjacency matrices consistent with a
given permutation π, we can show the following result.
Theorem A.2. Assume that Conditions 4.2 and 4.6 hold, then for all t > 0 and all δ1 > 0,
P
max
π
sup
{A(k) }K
k=1 ∈Aπ
2
p X
K
X
wk
j=1 k=1
nk
(k)T
(k)
ˆjπ X̂ (k) (aj
≥
−
(k)
a0jπ )
16σ02 (t +
− δ1
p X
K
X
wk
(k)
n
j=1 k=1 k
union |)
2 log p)(|G union | + |G0π
nδ1
(1)T
(K)T
(k)
X̂ (k) (aj − a0jπ )
2
2
≤ exp(−t).
(1)T
(K)T
Proof. Let ˆjπ and a0jπ be the concatenated vectors ˆjπ := (ˆ
jπ , . . . , ˆjπ )T and a0jπ := (a0jπ , . . . , a0jπ )T .
Also, define the block diagonal matrix X̂ := diag(X̂ (1) , · · · , X̂ (K) ). We denote by Ajπ ⊂ RpK the set that
(k)
contains all vectors that can be formed by vertically concatenating the jth columns aj for all k and that
satisfy that
n
o
(k)
(k)
(k)
Ajπ := aj ∈ RpK | ∀i, if ∃k such that ai,j 6= 0, then Xi ⊥
⊥ jπ for all k .
17
Based on this, and recalling that Paj (·) denotes the set of parent nodes of j in the argument graph, we define
the random event Bjπ as
(
Bjπ := ∃aj ∈ Ajπ :
ˆTjπ X̂(aj − a0jπ ) /n
sup
(19)
kX̂(aj −a0jπ )k22 /n≤1
s
2K(|Paj
≥ σ0
(G union )|
+
union )|)
|Paj (G0π
n
s
2(t +
+
union )| log p
|Paj (G0π
n
+ 2 log p)
.
union )|) non-zero entries,
Combining the facts that: i) aj − a0jπ may have at most K(|Paj (G union )| + |Paj (G0π
2
and ii) the variance of each element of ˆjπ is upper bounded by σ0 (cf. Condition 4.2), we may apply
Lemma A.1 to show that
union
P(Bjπ ) ≤ exp −t − |Paj (G0π
)| log p − 2 log p .
(20)
union . Putting
As can be seen from (19), event Bjπ depends exclusively on the set of parents of node j in G0π
union
union , then
it differently, if for two permutations π1 and π2 node j has the same set of parents in G0π1 and G0π
2
the random events Bjπ1 and Bjπ2 coincide, since Ajπ , a0jπ and ˆjπ would all be the same for π ∈ {π1 , π2 }.
union
If we denote
by Πj (m) the set of permutations where node j has exactly m parents in G0π , then there are
p
at most m unique events Bjπ for all π ∈ Πj (m). We therefore obtain that
[
p
P
Bjπ ≤
exp (−t − m log p − 2 log p) ≤ exp (−t − 2 log p) .
m
π∈Πj (m)
Applying a union bound on the events Bjπ across all nodes j and permutations π yields that
p X
p−1
[[
X
[
P
Bjπ ≤
P
Bjπ ≤ p2 exp (−t − 2 log p) = exp(−t).
j
π
j=1 m=1
(21)
π∈Πj (m)
Combining (21) and (19) it follows that with probability at least 1 − exp(−t), for all j, π and all aj ∈ Ajπ
we have that
q
q
|ˆ
Tjπ X̂(aj − a0jπ )|
union
union
union
≤ σ0
2K(|Paj (G
)| + |Paj (G0π )|) + 2(t + |Paj (G0π )| log p + 2 log p) .
kX̂(aj − a0jπ )k2
0(k) }K
Based on the collection of adjacency matrices {A(k) }K
k=1 ∈ Aπ we define another collection {A
k=1
(k)
where each column a0 j is given by
( (k)
(k)T
(k)
(k)
aj
if ˆjπ X̂ (k) (aj − a0jπ ) ≥ 0,
0 (k)
aj =
(k)
(k)
2a0jπ − aj
otherwise.
(k)
(k)
(k)
(k)
Notice that the positions of the non-zero entries in a0 j − a0jπ coincide with those in aj − a0jπ . By also
using the fact that
kX̂(aj − a0jπ )k22 =
K
X
k=1
(k)
(k)
kX̂ (k) (aj − a0jπ )k22 =
K
X
k=1
18
(k)
(k)
kX̂ (k) (a0 j − a0jπ )k22 = kX̂(a0j − a0jπ )k22 ,
we have that for all j and π with probability at least 1 − exp(−t), it holds that
(k)
(k)T
jπ X̂ (k) (aj
k=1 |ˆ
PK
(k)
− a0jπ )|
(k)T
(k)
jπ X (k) (a0 j −
k=1 |ˆ
kX̂(a0j − a0jπ )k2
PK
=
(k)
a0jπ )|
=
|ˆ
Tjπ X̂(a0j − a0jπ )|
(22)
kX̂(aj − a0jπ )k2
kX̂(a0j − a0jπ )k2
q
q
union
union
union
2K(|Paj (G
)| + |Paj (G0π )|) + 2(t + |Paj (G0π )| log p + 2 log p) .
≤ σ0
In the above expression we may use that ∀ δ1 , a, b > 0, it follows ab ≤ δ1 a2 + b2 /δ1 in order to obtain
2|ˆ
Tjπ X̂(a0j − a0jπ )| ≤ δ1 kX̂(a0j − a0jπ )k22 +
(23)
2
q
q
union )|) +
union )| log p + 2 log p)
/δ1 .
4σ02
2K(|Paj (G union )| + |Paj (G0π
2(t + |Paj (G0π
By combining this with the fact that ∀a, b > 0, (a + b)2 ≤ 2(a2 + b2 ) and the fact that K = o(log p) from
Condition 4.6, we get that
q
2
q
union
union
union
2K(|Paj (G
)| + |Paj (G0π )|) + 2(t + |Paj (G0π )| log p + 2 log p)
(24)
union
≤ 4(t + 2 log p)(|Paj (G union )| + |Paj (G0π
)|).
By replacing (24) back in (23), we obtain that
2|ˆ
Tjπ X̂(a0j − a0jπ )|/n ≤ δ1 kX̂(a0j − a0jπ )k22 /n +
union )|)
16σ02 (t + 2 log p)(|Paj (G union )| + |Paj (G0π
.
nδ1
Rewriting the absolute value in the left hand side as the sum of the corresponding K absolute values and
adding the above inequality for j = 1, . . . , p we get that, with probability at least 1 − exp(−t),
2
p X
K
X
(k)T
(k)
(k)
wk |ˆ
jπ X̂ (k) (aj − a0 jπ )|/nk
j=1 k=1
≤
p
X
δ1 kX̂(aj −
a0jπ )k22 /n
j=1
union )|)
16σ02 (t + 2 log p)(|Paj (G union )| + |Paj (G0π
+
nδ1
Finally, by noticing that kX̂(aj − a0jπ )k22 /n =
of the theorem, concluding the proof.
A.1.2
PK
k=1 wk kX̂
(k) (a(k) −a(k) )k2 /n
k
j
0jπ 2
.
we recover the statement
Random event E2
(k)
(k)
(k)
Let ω0jπ denote the j-th diagonal entry of Ω0π , E2 holds whenever the empirical variances of all jπ , i.e.,
(k)
(k)
kˆ
jπ k22 /nk are close to the true variances ω0jπ . More formally,
E2 :=
p X
K
X
j=1 k=1
wk
(k)
kˆ
jπ k22 /nk −
(k)
ω0jπ
19
(k) !2
ω0jπ
≤ 4 λ22
union
p + |G0π
|
,
(25)
p
for some λ2 (log p)/n. Mimicking the development for event E1 , we now show that E2 also holds with
high probability. This result is stated in Theorem A.4. Similar to the proof of Theorem A.2, we first consider
the asymptotic property for a particular node j and permutation π, and then leverage this to get a uniform
bound across all permutations and nodes. For this proof, we use the following lemma stated in [3], which
also follows from [45]. After the lemma, we formally state our result.
Lemma A.3. [3, Lemma 2] Suppose X1 , · · · , Xn are K-dimensional random vectors satisfying EXi = 0
and kXi k2 ≤ M for 1 ≤ i ≤ n. We have for any β > 0 and x > β
P
1/2
P(k ni=1 Xi k2 ≥ x) ≤ P kN k2 ≥ (x − β)/λmax + c1 K 5/2 exp(−c2 K −5/2 β/M ),
P
where λmax is the largest eigenvalue of Cov( ni=1 Xi ), N is a K-dimensional standard normal random
vector and c1 , c2 are positive constants.
Theorem A.4. Assume Conditions 4.5 and 4.6 hold, then there exist constants c1 , c2 , c3 > 0 such that
!
(k)
(k) !2
p X
K
union | log p
X
kˆ
jπ k22 /nk − ω0jπ
cs p log p + |G0π
wk
P ∃π :
≥ c1
≤ c2 exp(−c3 log p),
(k)
n
ω
j=1 k=1
0jπ
where cs is the constant defined in Condition 4.5.
(k)
Proof. We begin by analyzing the asymptotic properties of jπ for all k given a fixed permutation π and
node j. More specifically, consider the following random event
(k) 2
(k) !2
K
X
union
kˆ
jπ k2 /nk − ω0jπ
2 log p (|Paj (G0π )| + cs )
Cjπ :=
wk
≥
c
,
(26)
1
(k)
n
ω
k=1
0jπ
(k)
for some positive constant c1 . Following the proof of Theorem 1 in [3], we define ut
!
(k) 2
(k)
√
[ˆ
]
−ω
jπ t
0jπ
wk
if t ≤ nk ,
(k)
(k)
nk
ω0jπ
ut :=
0
otherwise.
(1)
(K) T
)
as follows
(k)
Denoting by ut = (ut , · · · , ut
the random vector collecting all ut , by definition it follows that
2
(k)
(k) !2
K
n
X
X
kˆ
jπ k22 /nk − ω0jπ
wk
=
ut .
(k)
ω
t=1
k=1
0jπ
2
A straightforward substitution in (26) allows us to rewrite the event Cjπ as
)
( n
union )| + c )
X
log p (|Paj (G0π
s
ut ≥ c1 λn
with λ2n :=
Cjπ :=
.
n
t=1
(27)
2
Notice that in order to apply Lemma A.3 to bound the probability of occurrence of Cjπ , we would need
kut k2 to be smaller than a constant M , which is not true in general. Hence, we split Cjπ into two subevents,
enabling the utilization of Lemma A.3. More precisely, whenever the following random event holds
n
o
(k)
1/2−a
Fa := |ut | ≤ λ−1
K
/n,
∀t,
k
,
n
20
then the `2 norm of ut is smaller than
1−a
kut k2 ≤ Ma := λ−1
/n
n K
where a is a free parameter that will be fixed later in the proof. In detail, we bound the probability of Cjπ
according to
P(Cjπ ) ≤ P(Cjπ |Fa )P(Fa ) + P(¬Fa ),
(28)
where we use Lemma A.3 to bound P(Cjπ |Fa ) and where P(¬Fa ) can be estimated from the chi-squared
tail bound [13].
We first focus on bounding P(Cjπ |Fa ). Let us introduce a new variable ũt obtained by truncating the
(1)
(K)
tail of ut . Formally, recalling that 1{·} denotes the indicator function, we have that ũt := (ũt , · · · , ũt )
where
n
o
h
n
oi
(k)
(k)
(k)
(k)
(k)
1/2−a
1/2−a
ũt := ut 1 |ut | ≤ λ−1
/n − E ut 1 |ut | ≤ λ−1
/n .
n K
n K
Notice that whenever
the
h
n random event Fa holds,
oi then ut and ũt follow the same distribution except for a
(k)
(k)
(k)
(k)
(k)
1/2−a
−1
shift vt := E ut 1 |ut | ≤ λn K
/n . Putting it differently, the distribution of ut − ũt is a
P
(k)
(k)
constant
vt whenever we are on the random event Fa . This implies that k nt=1 ut k2 ≤ n maxt,k |vt | +
Pn
k t=1 ũt k2 .
(k)
Therefore, if we can guarantee that n|vt | = o(1)λn P
for all t and k, then there must exist
P a constant
0 < δ < 1 such that if we are on the random event Fa , k nt=1 ut k2 ≥ c1 λn implies that k nt=1 ũt k2 ≥
(1 − δ)c1 λn . Equivalently, we may write
!
!
n
n
X
X
P(Cjπ |Fa )P(Fa ) ≤ P
ũt ≥ (1 − δ)c1 λn Fa P(Fa ) ≤ P
ũt ≥ (1 − δ)c1 λn , (29)
t=1
t=1
2
2
where the second inequality follows from Bayes’ theorem, and we can bound the last term by applying
(k)
Lemma A.3 since ũt is bounded by definition. We now show that, indeed, n|vt | = o(1)λn for all t and k.
From the cumulative tail bound of a chi-squared random variable with one degree of freedom we have that
√
(k)
(k)
P(ut ≥ l) ≤ exp(−ηnl/ K) for some constant η > 0. Based on this, we can estimate the scale of vt
with respect to p and n as
h
n
oi
h
n
oi
(k)
(k)
(k)
(k)
(k)
1/2−a
−1 1/2−a
−a
|vt | = E ut 1 |ut | ≤ λ−1
K
/n
=
E
u
1
|u
|
>
λ
K
/n
≤ exp −η 0 λ−1
n
n
n K
t
t
(k)
for some 0 < η 0 < η, where the second equality follows from the fact that ut has zero mean. Notice
(k)
(k)
that in the last inequality, ut has been absorbed into the exponential term. As |vt | decays exponentially
(k)
with respect to λ−1
n , we have that n|vt | = o(1)λn . Having justified this, we may now apply Lemma A.3
to the rightmost term in (29) in order to bound P(Cjπ |Fa )P(Fa ). From the definition of ũt it follows that
P
P
(k)
λmax {Cov( nt=1 ũt )} ≤ λmax {Cov( nt=1 ut )}. Furthermore, since the variables ut are independently
distributed for all t, we have that
!
(k) 2
(k)
n
X
[ˆ
]
−
ω
wk
jπ t
0jπ
(k)
≤ c2 /n
var
var
ut
=
(k)
nk
ω
t=1
0jπ
21
P
for some constant c2 > 0. This also implies that λmax {Cov( nt=1 ũt )} ≤ c2 /n. Applying Lemma A.3,
where we select x = (1 − δ)c1 λn , β = δ 0 x for some arbitrary positive constant 0 < δ 0 < 1 and, M =
1−a /n, it follows that
λ−1
n K
!
n
X
2
2
a− 72
2
P
ũt ≥ (1 − δ)c1 λn ≤ exp(−c3 (nλn − K)) + exp
log K − c4 K
nλn ,
(30)
5
t=1
2
for some constants c3 , c4 > 0 that increase if constant c1 is increased. In addition, by choosing a = 7/2,
there must exist a large enough c1 such that c3 , c4 > 1 and therefore
!
n
X
P
(31)
ũt ≥ (1 − δ)c1 λn ≤ exp(−nλ2n ) = exp(−|Paj (G union )| log p − cs log p).
t=1
2
Replacing (31) into (29) gives us the sought exponential bound for the first summand in (28).
We are now left with the task of finding a bound for P(¬Fa ). By relying on the fact that n(1) · · ·
n(K) (cf. Condition 4.6), we get that (maxk nk ) K ≤ c5 n for some constant c5 and therefore
n
o
(k)
1/2−a
P(¬Fa ) ≤ c5 n
max
P |ut | ≥ λ−1
K
/n
.
(32)
n
1≤k≤K,1≤l≤n
Following this, in order to bound the probability that ¬Fa holds we further rely on the tail bound of the
(k)
chi-squared random variable ut to obtain
−a
−a
P(¬Fa ) ≤ c5 n exp −ηλ−1
= c5 exp log n − ηλ−1
.
n K
n K
Recalling the definition of λn from (27), Condition 4.5 implies that
3
λ−1
n
≥ log p(|Paj (G union )| + cs )/α̃ 2
7/2
K
√
and
α̃
λ−1
n
≥ log n.
K 7/2
(33)
By recalling that we have fixed a = 27 , it follows that there exists a constant η 0 such that
√
−a
−a
−1 −a
−1 −a
≤ c5 exp −η 0 λ−1
,
P(¬Fa ) ≤ c5 exp log n − ηλ−1
K
≤
c
exp
α̃λ
K
−
ηλ
K
5
n K
n
n
n
where we have used the second inequality in (33). Furthermore, by leveraging the first inequality in (33) we
obtain that
3
−a
0
union
2
≤
c
exp
−η
log
p
(|Pa
(G
)|
+
c
)/α̃
P(¬Fa ) ≤ c5 exp −η 0 λ−1
K
,
(34)
5
j
s
n
thus obtaining an exponential bound for the second summand in (28).
Having found exponential bounds for both summands in (28), it follows that for c1 > 0 sufficiently large
and α̃ sufficiently small we have that P(Cjπ ) ≤ (1 + c5 ) exp(−|Paj (G union )| log p − cs log p). Following an
argument based on union bounds similar to the one presented in the proof of Theorem A.2, we have that
(k) 2
(k) !2
K
union
X
kˆ
jπ k2 /nk − ω0jπ
log p (|Paj (G
)| + cs )
P
wk
, ∀ j, π
≤ c1
(k)
n
ω0jπ
k=1
p [
p
[
[
≥ 1 − P
Cjπ ≥ 1 − (1 + c5 ) exp(−(cs − 2) log p)
j=1 m=1 π∈Πj (m)
22
for some constant c1 > 0. It is immediately implied from the previous expression that
(k)
(k) !2
p X
K
union
X
kˆ
jπ k22 /nk − ω0 jπ
cs p log p + |G0π | log p
≥ c1
P∃π :
wk
≤ (1 + c5 ) exp(−(cs − 2) log p),
(k)
n
ω0 jπ
j=1 k=1
thus recovering the statement of the theorem (since cs > 2 by Assumption 4.5) after accordingly renaming
the constants on the right hand side.
A.1.3
Random event E3
(k)
(k)
K
Event E3 is defined as the intersection of 2K events that we denote by {E3a }K
k=1 and {E3b }k=1 , where
(k)
(k)
(k)
events E3a and E3b are specific to the kth SEM. More specifically, event E3a ensures that all the estimated
(k)
noise variances ω̂j associated with the kth SEM are finite and bounded away from zero. Formally, we
define the following events for k = 1, . . . , K,
n
o
(k)
(k)
(k)
E3a := min ω̂j , 1/ω̂j
≥ 1/β 2 , for j = 1, . . . , p ,
(35)
(k)
for some β > 0. Event E3b imposes a universal lower bound on the norm achievable by any linear combinations of the data associated with the kth DAG. Mathematically, we consider the ensuing events for
k = 1, . . . , K,
n
o
p
√
(k)
(k)
E3b := kX̂ (k) vk2 / nk ≥ δ3 − λ3
kvk0 kvk2 , ∀v ∈ Rp ,
(36)
(k)
for some δ3 > 0 and λ3
p
(k)
(k)
(k)
(log p)/nk . Based on (35) and (36) we define events E3 := E3a ∩ E3b , and
E3 :=
K
\
(k)
E3 .
(37)
k=1
(k)
Leveraging the fact that Condition 4.4 enforces the maximum in-degree of each G0π to be at most αnk / log p
for some positive constant α, we can generalize Lemmas 7.5 and 7.7 from [43] into the following lemma.
Lemma A.5. [43, Lemmas 7.5 and 7.7] Assume Conditions 4.2, 4.3, 4.4 and 4.5 hold and that
r
p
√
2(t + log p)
3 Λmin /4 −
− 3σ0 α + α̃ ≥ 1/β > 0,
n
for some t > 0. Based on this, define
s
s
p
log p
2(t + log p)
(k)
λ3 := 3σ0
, and δ3 := 3 Λmin /4 −
nk
nk
(k)
(k)
Then P(E3 ) ≥ 1 − 5 exp(−t) and on E3
it holds that
√
(k)
(k)
(k)
(k)
kX̂ (k) (aj − a0j π̂ )k2 / nk ≥ kaj − a0j π̂ k2 /β 2 .
(k)
(38)
Lemma A.5 shows that under certain conditions the events E3 hold with high probability, thus playing
a role analogous to that of Theorem A.2 for event E1 and Theorem A.4 for event E2 .
With the events E1 , E2 , and E3 defined and having shown under which conditions these hold with high
probability, in the next section we leverage these events to prove Theorem 4.9.
23
A.2
Proof of Theorem 4.9
A.2.1
Bounds on new probability space
Through direct manipulation of the likelihood function, in Lemma A.6 we show that the global optimum
(k)
(k) K
(Â(k) , Ω̂(k) )K
k=1 converges to the SEMs (A0π̂ , Ω0π̂ )k=1 , where π̂ is some permutation consistent with the
estimated adjacency matrices Â(1) , · · · , Â(K) . Lemma A.6 is inspired by Lemma 7.1 in [43].
Lemma A.6. Assume we are on E1 ∩E2 ∩E3 and Condition 4.2 holds. Consider anregularizer in (7) satisfying
o
λ2 > λ21 /δ1 + λ22 /δ2 with 0 < δ1 < 1/β 2 and 0 < δ2 < 1/(2β 2 σ02 ). Then, π̂, {(Â(k) , Ω̂(k) )}K
k=1 the
global optimum of (7), satisfies
(k)
X
X
p X
p X
K
K
1
1
(k)
(k) (k)
2
− δ1
wk kX (âj − a0j π̂ )k2 /nk +
− δ2
wk
β2
2β 4 σ04
j=1 k=1
j=1 k=1
union |)
λ2 |G union |
λ2 (p + |G0π̂
λ21 λ22
2
+ λ −
−
+ 1 0π̂ .
|Ĝ| ≤ λ2 |G0union | + 2
δ1
δ2
δ2
δ1
ω̂j
(k) !2
− ω0 j
(k)
ω̂j
(39)
Proof. By definition, the global optimum must satisfy
K
X
K
X
wk `nk (X̂ (k) ; Â(k) , Ω̂(k) ) − λ2 |Ĝ| ≥
k=1
(k)
(k)
wk `nk (X̂ (k) ; A0 , Ω0 ) − λ2 |G0union |.
(40)
k=1
(k)
(k)
Let π̂ denote any permutation consistent with all Â(k) . Since the value of the likelihood `nk (X̂ (k) ; A0 , Ω0 )
(k)
is completely determined by the precision matrices {Θ0 }K
k=1 , it then follows that the likelihood functions
(k)
(k)
(k)
(k)
(k)
(k)
`nk (X̂ ; A0 , Ω0 ) and `nk (X̂ ; A0π̂ , Ω0π̂ ) achieve the same value. We therefore replace the former by
the latter in (40) and expand the definition of the likelihood function in (5) to obtain
p+
p X
K
X
(k)
wk log ω̂j
2
+ λ |Ĝ| ≤
j=1 k=1
p X
K
X
(k)
wk
kˆ
j π̂ k22 /nk
j=1 k=1
(k)
ω0j π̂
+
p X
K
X
(k)
wk log ω0j π̂ + λ2 |G0union |.
j=1 k=1
Basic manipulations transform the above expression into the following inequality
(k) 2
(k)
p X
p X
K
K
X
X
kˆ
k
/n
ω̂j
k
j π̂ 2
wk
− 1 + λ2 |G0union |.
wk log (k) + λ2 |Ĝ| ≤
(k)
ω0j π̂
ω0j π̂
j=1 k=1
j=1 k=1
(k)
Since we are on E3 , we have that 1/ω̂j
(k)
(41)
≤ β 2 [cf. (35)]. By combining this with Condition 4.2 we can
(k)
≤ β 2 σ02 . Then, using the Taylor expansion log(1 + x) ≤ x − x2 /(2(1 + t)2 ), for
(k)
ω̂j
−1 < x ≤ t, we can further replace log
in (41) to obtain
(k)
further bound ω0j π̂ /ω̂j
ω0j π̂
p X
K
X
j=1 k=1
wk
(k)
ω̂j
(k)
− ω0j π̂
(k)
ω̂j
2
(k)
(k) 2
p X
K
X
ω0j π̂
kˆ
k
/n
k
1
j π̂ 2
+
− 1 + λ2 |Ĝ| ≤
wk
− 1 + λ2 |G0union |.
(k)
2β 4 σ04 ω̂ (k)
ω0j π̂
j=1 k=1
j
(42)
24
(k)
Finally, using the fact that X̂j
(k)
ω̂j
(k)
= kX̂j
(k)
(k)
(k)
= ˆj π̂ + X̂ (k) a0j π̂ , we also rewrite ω̂j
(k)
(k)
(k)
as
(k)T
(k)
(k)
(k)
− X̂ (k) âj k22 /nk = kX̂ (k) (âj − a0j π̂ )k22 /nk − 2ˆ
j π̂ X̂ (k) (âj − a0j π̂ )/nk + kˆ
j k22 /nk .
By replacing the above into (42), we get that
p X
K
X
(k)
wk
(k)
kX̂ (k) (âj − a0j π̂ )k22 /nk
(k)
ω̂j
j=1 k=1
≤2
+
p X
K
X
1
2β 4 σ04
p X
K
X
(k)
(k)
− ω0j π̂
(k)
ω̂j
(k)
ˆj π̂ X̂ (k) (âj − a0j π̂ )/nk
(k)
+
ω̂j
j=1 k=1
−
wk
(k)
ω̂j
j=1 k=1
(k)T
wk
(k)
p X
K
X
kˆ
j π̂ k22 /nk −
wk
(k)
ω̂j
j=1 k=1
2
+ λ2 |Ĝ|
p X
K
X
wk
(43)
(k)
kˆ
j π̂ k22 /nk
j=1 k=1
(k)
ω0j π̂
− 1
(k)
ω0j π̂
+ λ2 |G0union |.
In order to further bound the expression in (43), notice that the first summand in the right hand side of the
inequality corresponds to the sum of all empirical correlation coefficients. Leveraging that we are under the
assumption that E1 holds [cf. (18)], we have that
2
p X
K
X
j=1 k=1
(k)T
wk
(k)
(k)
ˆj π̂ X̂ (k) (âj − a0j π̂ )/nk
(k)
ω̂j
≤ δ1
p X
K
X
(k)
(k)
wk kX̂ (k) (âj − a0j π̂ )k22 /nk +
j=1 k=1
λ21
|Ĝ|.
δ1
(44)
In order to bound the second and third terms, we first restate their difference as follows
(k)
(k)
(k)
(k)
(k)
(k)
(k)
p X
p X
K
K
X
X
kˆ
j π̂ k22 /nk − ω0j π̂
kˆ
j π̂ k22 /nk − ω0j π̂ ω0j π̂ − ω̂j
kˆ
j π̂ k22 /nk
−wk
.
wk
wk
− 1 =
(k)
(k)
(k)
(k)
ω̂
ω
ω
ω̂
j=1 k=1
j=1 k=1
j
j
0j π̂
0j π̂
(45)
P
n
2
2
k=1 |vk | , we further bound (45)
j=1 |uj |
Pn
Pn
Next, by using Cauchy-Schwarz inequality, i.e. | i=1 ui vi |2 ≤
as
(k)
(k)
(k)
p X
p X
K
K
X
X
kˆ
j π̂ k22 /nk − ω0j π̂
kˆ
j π̂ k22 /nk
−
wk
− 1
wk
(k)
(k)
ω̂j
ω0j π̂
j=1 k=1
j=1 k=1
2 1/2
2 1/2
(k)
(k)
(k)
(k)
p X
p X
K
K
X
X
ω0j π̂ − ω̂j
kˆ
j π̂ k22 /nk − ω0j π̂
wk
≤
wk
(k)
(k)
ω0j π̂
ω̂j
j=1 k=1
j=1 k=1
(46)
From the fact that event
q E2 holds [cf. (25)], we can upper bound the first of the two factors in the right
hand side of (46) by 2 λ22 (p + |G0union (π̂)|). Further, relying on the inequality 2ab ≤ a2 /δ2 + δ2 b2 for any
δ2 > 0, it follows that
(k)
(k)
(k)
p X
p X
K
K
X
X
kˆ
j π̂ k22 /nk − ω0j π̂
kˆ
j π̂ k22 /nk
−
wk
wk
− 1
(47)
(k)
(k)
ω̂
ω
j=1 k=1
j=1 k=1
j
0j π̂
2
(k)
(k)
p X
K
2
union
X
ω
−
ω̂
λ (p + |G0 (π̂)|)
j
0j π̂
≤ 2
+ δ2
wk
.
(k)
δ2
ω̂
j=1 k=1
25
j
By replacing (44) and (47) into (43), we recover (39), as we wanted to show.
From Lemma A.6 it follows that the global optimum of (7) corresponds to a minimal I-MAP, but no
claim is made about the sparsity level of this I-MAP. In order to show that the solution is indeed sparse, we
union | cannot be much
must rely in Conditions 4.7 and 4.8. Consequently, in Lemma A.7 we show that |G0π̂
union
larger than |Ĝ|. Then, in Thm. A.8 we further show how to cancel out |G0π̂ | with |Ĝ| in (39) to obtain our
main result.
Lemma A.7. Assume Condition 4.7 holds and let λ̃ > 0 be such that
K
X
(k)
union
|.
wk kÂ(k) − A0π̂ k2F ≤ λ̃2 |G0π̂
(48)
k=1
Consider also constants η1 , η2 with 0 ≤ η1 < 1 and 0 < η22 ct < 1−η1 such that
P
i,j 1
n
o
(k)
[A0π̂ ]i,j ≥ λ̃/η2 ≥
(k)
(1 − η1 )|G0π̂ |. Then, it follows that
|Ĝ| ≥
1 − η1 − η22 ct union
|G0π̂ |.
ct
(49)
(k)
Proof. Let N (k) and M(k) be the sets of entries satisfying N (k) := {(i, j) : |[A0π̂ ]i,j | ≥ λ̃/η2 } and
(k)
M(k) := {(i, j) : |[Â(k) ]i,j − [A0π̂ ]i,j | ≥ λ̃/η2 }. From these definitions it follows that
K
X
K
wk |N (k) ∩ M(k) |
k=1
λ̃2 X
wk
≤
η22
k=1
(k)
X
|[Â(k) ]i,j − [A0π̂ ]i,j |2 ≤
K
X
(k)
wk kÂ(k) − A0π̂ k2F .
k=1
(i,j) ∈ N (k) ∩M(k)
Leveraging inequality (48) and Condition 4.7 we further have that
K
X
wk |N
(k)
∩M
(k)
|≤
union
η22 |G0π̂
|
k=1
≤
η22 ct
K
X
(k)
wk |G0π̂ |.
(50)
k=1
C
Notice that for all (i, j)-th entries in the set N (k) ∩ M(k) it must be that |[Â(k) ]i,j | > 0. Hence, N (k) ∩
C
M(k) corresponds to a subset of non-zero entries of Â(k) , which in turn corresponds to a subset of edges
in Ĝ. From this we can infer that
|Ĝ| =
K
X
k=1
wk |Ĝ| ≥
K
X
C
wk |N (k) ∩M(k) | =
k=1
K
X
wk (|N (k) | − |N (k) ∩ M(k) |) ≥ (1 − η1 − η22 ct )
k=1
K
X
(k)
wk |G0π̂ |,
k=1
where the last inequality follows by combining (50) with the definition of η1 in the statement of the lemma.
The proof concludes by replacing Condition 4.7 in the above inequality.
Theorem A.8. Assume Conditions 4.1, 4.7 and 4.8 hold, and suppose that there exist some constants δB
and 0 < δs < 1 as well as λ and λ0 that scale as λ2 λ20 logn p (p/|G0union | ∨ 1) such that
δB
K
X
(k)
union
wk kÂ(k) − A0π̂ k2F + λ2 δs |Ĝ| ≤ λ2 |G0union | + λ20 |G0π̂
|.
k=1
26
(51)
If the constant η0 in Condition 4.8 is sufficiently small, then there exist some constants δs0 , cg , c0g > 0 such
that
δB
K
X
(k)
wk kÂ(k) − A0π̂ k2F + (λ2 δs − λ20 δs0 )|Ĝ| ≤ λ2 |G0union |
(52)
k=1
and
union
| ≥ c0g |G0union |.
|Ĝ| ≥ cg |G0π̂
(53)
(k)
union |. Suppose that for
Proof. Using Conditions 4.1 and 4.7, we have that |G0union | ≤ ct max |G0 | ≤ ct |G0π̂
k
some λ̃ > 0 one has that λ̃2 λ2 λ20 and λ̃2 δB ≥ λ2 ct + λ20 , then it follows from (51) that
δB
K
X
(k)
union
union
|.
| ≤ λ̃2 δB |G0π̂
wk kÂ(k) − A0π̂ k2F ≤ λ2 |G0union | + λ20 |G0π̂
k=1
q
Let η2 be some constant defined as η2 := η0 λ̃/ logn p (p/|G0union | ∨ 1), then we can rewrite Condition 4.8 as
that
o
X n (k)
(k)
1 [A0π̂ ]i,j ≥ λ̃/η2 ≥ (1 − η1 )|G0π̂ |.
i,j
Moreover, for η0 sufficiently small, η2 is also guaranteed that 0 < η22 ct < 1 − η1 . We could therefore
apply Lemma A.7 and get that |Ĝ| ≥
1−η1 −η22 ct
ct
1−η1 −η22 ct union
|G0π̂ |,
ct
which completes the proof of (53) by choosing
c0g
cg =
and = cg · ct . Notice that in order to apply Lemma A.7, it is required that η22 ct < 1 − η1
which is guaranteed by the assumption that η0 is sufficiently small. Leveraging the first inequality in (53),
λ20 ct
union | in (51) by
|Ĝ| in order to obtain
we can replace λ20 |G0π̂
1−η −η 2 c
1
δB
K
X
wk kÂ
(k)
−
(k)
A0π̂ k2F
k=1
2 t
λ20 ct
+ λ δs −
1 − η1 − η22 ct
2
|Ĝ| ≤ λ2 |G0union |.
(54)
Notice that (54) coincides with the sought expression (52) upon substituting δs0 = ct /(1 − η1 − ct η22 ).
A.2.2
Proof of Theorem 4.9
It follows from Theorem A.2, Theorem A.4, and Lemma A.5 that there exist some constants λ1 , λ2 with
(k) 2
p
log
λ21 λ22 logn p as well as some λ3
nk for all k such that with probability 1 − exp(−c log p) for
some constant c > 0, the random event E1 ∩ E2 ∩ E3 occurs.
We may then apply Lemma A.6 to show that there exist some constants δB , δW such that with high
probability, for any λ > 0 satisfying λ2 > λ21 /δ1 + λ22 /δ2 ,
δB β
2
p X
K
X
j=1 k=1
wk kX
(k)
(k)
(âj
−
(k)
a0j π̂ )k22 /nk
+ δW
K
X
(k)
wk kΩ̂(k) − Ω0π̂ k2F + λ2 δs |Ĝ|
(55)
k=1
≤ λ2 |G0union | +
27
union |)
λ22 (p + |G0π̂
λ2 |G union |
+ 1 0π̂ .
δ2
δ1
Pp
(k)
(k)
ω̂j −ω0 j
2
(k)
Note that compared with Lemma A.6, we have replaced j=1
by kΩ̂(k) − Ω0π̂ k2F . This
(k)
ω̂j
P
(k)
(k)
(k)
1
is
follows from combining the facts that pj=1 (ω̂j − ω0 j )2 is equal to kΩ̂(k) − Ω0π̂ k2F and that (k)
ω̂j
1/β 2
bigger than
on the random event E3 . In addition, the constants β, σ0 , δ1 and δ2 in Lemma A.6 have
λ2
λ2
been absorbed into the new constants δB and δW . We also replaced λ2 − δ11 − δ22 by λ2 δs for some
0 < δs < 1.
By applying Lemma A.5 [cf. (38)] we may bound the first summand on the left hand side of (55)
P
(k) − A(k) k2 . Furthermore, replacing λ and λ by some λ that scales as λ2
by δB K
1
2
0
0
k=1 wk kÂ
0π̂ F
log p
union | ∨ 1), we obtain that
(p/|G
0
n
δB
K
X
wk kÂ
(k)
−
(k)
A0π̂ k2F
+ δW
k=1
K
X
(k)
union
wk kΩ̂(k) − Ω0π̂ k2F + λ2 δs |Ĝ| ≤ λ2 |G0union | + λ20 |G0π̂
|.
(56)
k=1
By applying (56), we have that for a constant η0 small enough, (52) and (53) in Theorem A.8 holds by
choosing λ as that λ2 logn p (p/|G0union | ∨ 1) and λ2 δs > λ20 δs0 . Moreover, from (52) we further infer that
δB
K
X
(k)
wk kÂ(k) − A0π̂ k2F + λ2 δs00 |Ĝ| ≤ λ2 |G0union |,
(57)
k=1
where the constant δs00 is chosen such that λ2 δs00 = λ2 δs − λ20 δs0 in (52). From (57) it can thus be inferred
union | ≥ c0 /c |G union |, we recover the
that |Ĝ| ≤ |G0union |/δs00 . Combining this with (53) and the fact that |G0π̂
g g 0
union |. For the relation between |G union | and
first part of (11) in the statement of the theorem, i.e., |Ĝ| |G0π̂
0π̂
union | ≤ |Ĝ|/c ≤ |G union |/(δ 00 · c ) and |G union | ≥ c0 /c |G union |. Finally, to recover
|G0union |, we use that |G0π̂
g
g
g
s
g
0
0
0π̂
(10) we combine (56) with (11).
A.3
Proof of Theorem 4.10
We first introduce a lemma that will be instrumental in proving Theorem 4.10 and that can be obtained by
generalizing Lemmas 7.2 and 7.3 in [43].
(k)
Lemma A.9. [43, Lemmas 7.2 and 7.3] Suppose for some δB , δs , λ0 , λ > 0 one has n
that δB kÂ(k) −A0π̂ ko2F +
P
(k)
(k)
(k)
λ2 δs |Ĝ (k) | ≤ λ2 |G0 | + λ20 |G0π̂ |. Let λ̃2 δB ≥ λ2 + λ20 and assume that i,j 1 [A0π̂ ]i,j ≥ λ̃/η2 ≥
(k)
(1 − η1 )|G0π̂ |. Then
(k)
δB kÂ(k) − A0π̂ k2F + λ2 δs −
λ20
1 − η1 − η22
(k)
|Ĝ (k) | ≤ λ2 |G0 |
and
(k)
(k)
|Ĝ (k) | ≥ (1 − η1 − η22 )|G0π̂ | ≥ (1 − η1 − η22 )|G0 |.
In order to show Theorem 4.10, we begin the proof just like for Theorem
4.9 in Section A.2.2 until we
P
get to expression (56). It then follows from Condition 4.7’ and |Ĝ| ≥ K
w
|Ĝ (k) | that there exists some
k
k=1
28
log p
union | ∨ 1), where C
λ02
max is defined in Condition 4.8, such that for any λ > 0,
0 Cmax n (p/|G0
δB
K
X
wk kÂ
(k)
−
(k)
A0π̂ k2F
+ δW
k=1
K
X
≤ λ2 ct (π0 )
(k)
wk |G0 | +
k=1
λ02
λ2
· ct (π0 ) and
Let
:=
0
any λ > 0,
δs0
K
X
wk kΩ̂
(k)
−
(k)
Ω0π̂ k2F
2
+ λ δs
k=1
K
X
(k)
λ02
wk |G0π̂ |.
0
k=1
K
X
wk |Ĝ (k) |
k=1
:= δs /ct (π0 ), it then follows that there must exist at least one k such that for
(k)
(k)
(k)
(k)
δB kÂ(k) − A0π̂ k2F + δW kΩ̂(k) − Ω0π̂ k2F + λ02 δs0 |Ĝ (k) | ≤ λ02 |G0 | + λ02
0 |G0π̂ |
Since according to Condition 4.7’, ct (π) scales as a constant for permutations consistent with G0union , we have
that δs0 is still a constant and λ0 λ. In this case, it follows from Lemma A.9 and Condition 4.8’ that there
exists some constant 0 < δs < 1 and δs0 > 0 such that by choosing λ0 as that λ0 Cmax logn p (p/|G0union | ∨ 1)
0
and λ02 δs > λ02
0 δs ,
(k)
(k)
(k)
0
δB kÂ(k) − A0π̂ k2F + λ02 δs − λ02
| ≤ λ02 |G0 |
0 δs |Ĝ
(k)
(k)
It also follows from Lemma A.9 that |Ĝ (k) | ≥ cg |G0π̂ | ≥ cg |G0 | for some positive constant cg . Mimicking
the arguments employed in the proof of Theorem 4.9 from (57) until the end of the proof, one can show that
expressions (13) and (14) in the statement of Theorem 4.10 hold true.
A.4
Proof of Corollary 5.1
The following lemma is instrumental in showing the corollary.
Lemma A.10. [17, Lemma 4] Given fixed G, the maximum likelihood estimator in (15) can be written as
p
X
X
X
n
n
−j
(k)
(k)
(k)
k
min
log
kX̂j − X̂Paj (G) · ak22 /nk +
wk log kX̂j k22 /nk
p+
|Paj (G)| n
n
−j
a∈R
k:j6∈Ik
j=1
k:j∈Ik
where n−j is the total number of data from the interventions where node j is not intervened, i.e., n−j =
P
k:j6∈Ik nk .
Recall that in the interventional setting G0union is given by the true graph G0 of the non-intervened model,
(k)
(k)
and that the K models to be inferred (A0 , Ω0 ) correspond to the interventional models (AI0k , ΩI0k ). De(k)
noting by (π̂, Â, Ω̂) the (non-intervened) global optimum of (15), let ω̂j denote the empirical variance of
(k)
(k)
the random variable Xj − X (k) âj if j ∈ Ik and the empirical variance of Xj
Lemma A.10 that the global optimum satisfies that
p
X
X
X
nk (k)
(k)
n−j log
p+
ω̂
+
wk log ω̂j + λ2 |Ĝ|
n
n−j j
j=1
k:j6∈Ik
otherwise. It follows from
k:j∈Ik
≤
p X
K
X
(k)
wk
kˆ
j π̂ k22 /nk
j=1 k=1
29
(k)
ω0j π̂
+
p X
K
X
j=1 k=1
(k)
wk log ω0j π̂ + λ2 |G0union |.
P
P
wk a k ) ≥ K
Then, applying the inequality log( K
k=1 wk log ak for any choices of a1 , . . . , aK > 0
k=1
PK
and w1 , . . . , wK > 0 with k=1 wk = 1, we have that
p+
p X
K
X
j=1 k=1
(k)
wk log ω̂j
2
+ λ |Ĝ| ≤
p X
K
X
(k)
wk
kˆ
j π̂ k22 /nk
j=1 k=1
(k)
ω0j π̂
+
p X
K
X
(k)
wk log ω0j π̂ + λ2 |G0union |.
j=1 k=1
In this case, Corollary 5.1 immediately follows from the proof of Theorem 4.9.
References
[1] P. A. Aguilera, A. Fernández, R. Fernández, R. Rumı́, and A. Salmerón. Bayesian networks in environmental modelling. Environmental Modelling & Software, 26(12):1376–1388, 2011.
[2] M. N. Arbeitman, E. EM. Furlong, F. Imam, E. Johnson, B. H. Null, B. S. Baker, M. A. Krasnow, M. P.
Scott, R. W. Davis, and K. P. White. Gene expression during the life cycle of drosophila melanogaster.
Science, 297(5590):2270–2275, 2002.
[3] T. T. Cai, H. Li, W. Liu, and J. Xie. Joint estimation of multiple high-dimensional precision matrices.
The Annals of Statistics, 38:2118–2144, 2015.
[4] T. T. Cai, W. Liu, and X. Luo. A constrained `1 minimization approach to sparse precision matrix
estimation. Journal of the American Statistical Association, 106(494):594–607, 2011.
[5] D. M. Chickering. Learning Bayesian networks is NP-complete. In Proceedings of the Fifth International Workshop on Artificial Intelligence and Statistics, 1995.
[6] D. M. Chickering. Optimal structure identification with greedy search. Journal of Machine Learning
Research, 3(Nov):507–554, 2002.
[7] P. Danaher, P. Wang, and D. M. Witten. The joint graphical lasso for inverse covariance estimation
across multiple classes. Journal of the Royal Statistical Society: Series B (Statistical Methodology),
76(2):373–397, 2014.
[8] A. Dixit, O. Parnas, B. Li, J. Chen, C. P. Fulco, L. Jerby-Arnon, N. D. Marjanovic, D. Dionne, T. Burks,
R. Raychowdhury, B. Adamson, T. M. Norman, E. S. Lander, J. S. Weissman, N. Friedman, and
A. Regev. Perturb-seq: Dissecting molecular circuits with scalable single-cell RNA profiling of pooled
genetic screens. Cell, 167(7):1853–1866, 2016.
[9] F. Eberhardt, C. Glymour, and R. Scheines. On the number of experiments sufficient and in the worst
case necessary to identify all causal relations among n variables. In Proceedings of the Twenty-First
Conference on Uncertainty in Artificial Intelligence, pages 178–184. AUAI Press, 2005.
[10] J. Friedman, T. Hastie, and R. Tibshirani. Sparse inverse covariance estimation with the graphical
lasso. Biostatistics, 9(3):432–441, 2008.
[11] N. Friedman, M. Linial, I. Nachman, and D. Peter. Using Bayesian networks to analyze expression
data. Journal of Computational Biology, 7:601–620, 2000.
30
[12] D. M. Gau, J. L. Lesnock, B. L. Hood, R. Bhargava, M. Sun, K. Darcy, S. Luthra, U. Chandran, T. P.
Conrads, R. P. Edwards, J. L. Kelley, T. C. Krivak, and P. Roy. BRCA1 deficiency in ovarian cancer is
associated with alteration in expression of several key regulators of cell motility–a proteomics study.
Cell Cycle, 14(12):1884–1892, 2015.
[13] B. George. Probability inequalities for the sum of independent random variables. Journal of the
American Statistical Association, 57(297):33–45, 1962.
[14] C. Glymour, R. Scheines, P. Spirtes, and K. Kelly. Discovering Causal Strucure. Academic Press,
1987.
[15] J. Guo, E. Levina, G. Michailidis, and J. Zhu.
Biometrika, 98(1):1–15, 2011.
Joint estimation of multiple graphical models.
[16] A. Hauser and P. Bühlmann. Characterization and greedy learning of interventional Markov equivalence classes of directed acyclic graphs. Journal of Machine Learning Research, 13(Aug):2409–2464,
2012.
[17] A. Hauser and P. Bühlmann. Jointly interventional and observational data: Estimation of interventional
Markov equivalence classes of directed acyclic graphs. Journal of the Royal Statistical Society: Series
B (Statistical Methodology), 77(1):291–318, 2015.
[18] J. Jönsson, K. Bartuma, M. Dominguez-Valentin, K. Harbst, Z. Ketabi, S. Malander, M. Jönsson,
A. Carneiro, A. Måsbäck, G. Jönsson, and M. Nilbert. Distinct gene expression profiles in ovarian
cancer linked to Lynch syndrome. Familial Cancer, 13:537–545, 2014.
[19] M. Kalisch and P. Bühlmann. Estimating high-dimensional directed acyclic graphs with the PCalgorithm. Journal of Machine Learning Research, 8(Mar):613–636, 2007.
[20] M. Kanehisa, S. Goto, Y. Sato, M. Furumichi, and T. Mao. KEGG for integration and interpretation of
large-scale molecular data sets. Nucleic Acids Research, 40(D1):D109–D114, 2011.
[21] M. Kolar, L. Song, A. Ahmed, and E. P. Xing. Estimating time-varying networks. The Annals of
Applied Statistics, pages 94–123, 2010.
[22] S. L. Lauritzen. Graphical Models, volume 17. Clarendon Press, 1996.
[23] P. Loh and P. Bühlmann. High-dimensional learning of linear causal networks via inverse covariance
estimation. Journal of Machine Learning Research, 15(1):3065–3105, 2014.
[24] E. Z. Macosko, A. Basu, R. Satija, J. Nemesh, K. Shekhar, M. Goldman, I. Tirosh, A. R. Bialas,
N. Kamitaki, E. M. Martersteck, J. J. Trombetta, D. A. Weitz, J. R. Sanes, A. K. Shalek, A. Regev, and
S. A. McCarroll. Highly parallel genome-wide expression profiling of individual cells using nanoliter
droplets. Cell, 161(5):1202–1214, 2015.
[25] N. Meinshausen and P. Bühlmann. High-dimensional graphs and variable selection with the lasso. The
Annals of Statistics, pages 1436–1462, 2006.
[26] N. Meinshausen and P. Bühlmann. Stability selection. Journal of the Royal Statistical Society. Series
B: Statistical Methodology, 72(4):417–473, 2010.
31
[27] K. Mohan, P. London, M. Fazel, D. Witten, and S. Lee. Node-based learning of multiple Gaussian
graphical models. The Journal of Machine Learning Research, 15(1):445–488, 2014.
[28] P. Nandy, A. Hauser, and M. H. Maathuis. High-dimensional consistency in score-based and hybrid
structure learning. The Annals of Statistics (to appear), 2018.
[29] H. Ogata, S. Goto, K. Sato, W. Fujibuchi, H. Bono, and M. Kanehisa. KEGG: Kyoto encyclopedia of
genes and genomes. Nucleic Acids Research, 27(1):29–34, 1999.
[30] J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2000.
[31] J. Pearl and T. S. Vermal. Equivalence and synthesis of causal models. In Proceedings of Sixth
Conference on Uncertainty in Artijicial Intelligence, pages 220–227, 1991.
[32] C. Peterson, F. C. Stingo, and M. Vannucci. Bayesian inference of multiple Gaussian graphical models.
Journal of the American Statistical Association, 110(509):159–174, 2015.
[33] G. Raskutti and C. Uhler. Learning directed acyclic graphs based on sparsest permutations. Stat (to
appear), 2018.
[34] P. Ravikumar, M. J. Wainwright, G. Raskutti, and B. Yu. High-dimensional covariance estimation
by minimizing `1 -penalized log-determinant divergence. Electronic Journal of Statistics, 5:935–980,
2011.
[35] J. M. Robins, M. A. Hernan, and B. Brumback. Marginal structural models and causal inference in
epidemiology. Epidemiology, 11(5):550–560, 2000.
[36] A. D. Santin, F. Zhan, S. Bellone, M. Palmieri, S. Cane, E. Bignotti, S. Anfossi, M. Gokden, D. Dunn,
J. J. Roman, T. J. O’Brien, E. Tian, M. J. Cannon, J. Shaughnessy, and S. Pecorelli. Gene expression profiles in primary ovarian serous papillary tumors and normal ovarian epithelium: identification
of candidate molecular markers for ovarian cancer diagnosis and therapy. International Journal of
Cancer, 112(1):14–25, 2004.
[37] A. K. Shalek, R. Satija, J. Shuga, J. J. Trombetta, D. Gennert, D. Lu, P. Chen, R. S. Gertner, J. T.
Gaublomme, N. Yosef, S. Schwartz, B. Fowler, S. Weaver, J. Wang, X. Wang, R. Ding, R. Raychowdhury, N. Friedman, N. Hacohen, H. Park, A. P. May, and A. Regev. Single cell RNA Seq reveals
dynamic paracrine control of cellular variation. Nature, 510(7505):363, 2014.
[38] L. Song, M. Kolar, and E. P. Xing. Keller: estimating time-varying interactions between genes. Bioinformatics, 25(12):i128–i136, 2009.
[39] P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction and Search. MIT Press, 2000.
[40] E. A. Stronach, G. C. Sellar, C. Blenkiron, G. J. Rabiasz, K. J. Taylor, E. P. Miller, C. E. Massie,
A. Al-Nafussi, J. F. Smyth, D. J. Porteous, and H. Gabra. Identification of clinically relevant genes
on chromosome 11 in a functional model of ovarian cancer tumor suppression. Cancer Research,
63(24):8648–8655, 2003.
[41] R. W. Tothill, A. V. Tinker, J. George, R. Brown, S. B. Fox, S. Lade, D. S. Johnson, M. K. Trivett, D. Etemadmoghadam, B. Locandro, N. Traficante, S. Fereday, J. A. Hung, Y. Chiew, I. Haviv,
32
Australian Ovarian Cancer Study Group, D. Gertig, A. deFazio, and D. D.L. Bowtell. Novel molecular subtypes of serous and endometrioid ovarian cancer linked to clinical outcome. Clinical Cancer
Research, 14(16):5198–5208, 2008.
[42] I. Tsamardinos, L. E. Brown, and C. F. Aliferis. The max-min hill-climbing Bayesian network structure
learning algorithm. Machine learning, 65(1):31–78, 2006.
[43] S. van de Geer and P. Bühlmann. `0 -penalized maximum likelihood for sparse directed acyclic graphs.
The Annals of Statistics, 41(2):536–567, 04 2013.
[44] M. Yuan and Y. Lin. Model selection and estimation in regression with grouped variables. Journal of
the Royal Statistical Society: Series B (Statistical Methodology), 68(1):49–67, 2006.
[45] A. Y. Zaitsev. On the Gaussian approximation of convolutions under multidimensional analogues of
S.N. Bernstein’s inequality conditions. Probability Theory and Related Fields, 74(4):535–566, 1987.
33
| 10math.ST
|
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IN PRESS
1
Identifying Corresponding Patches in SAR and
Optical Images with a Pseudo-Siamese CNN
arXiv:1801.08467v1 [eess.IV] 25 Jan 2018
Lloyd H. Hughes, Student Member, IEEE, Michael Schmitt, Senior Member, IEEE, Lichao Mou,
Yuanyuan Wang, Member, IEEE, and Xiao Xiang Zhu, Senior Member, IEEE
Abstract—This is the pre-print version, to read the final version
please go to IEEE Geoscience and Remote Sensing Letters on
IEEE Xplore.
In this letter, we propose a pseudo-siamese convolutional neural
network (CNN) architecture that enables to solve the task of
identifying corresponding patches in very-high-resolution (VHR)
optical and synthetic aperture radar (SAR) remote sensing
imagery. Using eight convolutional layers each in two parallel
network streams, a fully connected layer for the fusion of the
features learned in each stream, and a loss function based on
binary cross-entropy, we achieve a one-hot indication if two
patches correspond or not.
The network is trained and tested on an automatically generated dataset that is based on a deterministic alignment of
SAR and optical imagery via previously reconstructed and subsequently co-registered 3D point clouds. The satellite images, from
which the patches comprising our dataset are extracted, show
a complex urban scene containing many elevated objects (i.e.
buildings), thus providing one of the most difficult experimental
environments. The achieved results show that the network is
able to predict corresponding patches with high accuracy, thus
indicating great potential for further development towards a
generalized multi-sensor key-point matching procedure.
Index Terms—synthetic aperture radar (SAR), optical imagery,
data fusion, deep learning, convolutional neural networks (CNN),
image matching, deep matching
I. I NTRODUCTION
T
HE identification of corresponding image patches is
used extensively in computer vision and remote sensingrelated image analysis, especially in the framework of stereo
applications or co-registration issues. While many successful
hand-crafted approaches, specifically designed for the matching of optical images, exist [1], to this date the matching of
images acquired by different sensors still remains a widely
unsolved challenge [2]. This particularly holds for a joint
exploitation of SAR and optical imagery, caused by two
completely different sensing modalities: SAR imagery collects information about the physical properties of the scene
The authors are with Signal Processing in Earth Observation,
Technical University of Munich, 80333 Munich, Germany (email:
[email protected],
[email protected],
[email protected],
[email protected], [email protected])
X. Zhu is also with the Remote Sensing Technology Institute, German
Aerospace Center, 82234 Wessling, Germany
Manuscript received XXXX YY, ZZZZ.
This work was supported by the China Scholarship Council, the European
Research Council (ERC) under the European Unions Horizon 2020 research
and innovation programme (grant agreement No. ERC-2016-StG-714087,
Acronym: So2Sat), the Helmholtz Association under the framework of the
Young Investigators Group SiPEO (VH-NG-1018, www.sipeo.bgu.tum.de)
and the German Research Foundation (DFG), grant SCHM 3322/1-1.
and follows a range-based imaging geometry, while optical
imagery reflects the chemical characteristics of the scene and
follows a perspective imaging geometry. Hence, structures
elevated above the ground level, such as buildings or trees,
show strongly different appearances in both SAR and optical
images (cf. Fig. 1), in particular when dealing with very high
resolution data.
Fig. 1. A simple detached multi-story building as SAR amplitude image (left)
and optical photograph (right).
In order to deal with the problem of multi-sensor keypoint matching, several sophisticated approaches have been
proposed, e.g. exploiting phase congruency as a generalization
of gradient information [3]. However, even sophisticated handcrafted descriptors reach their limitations for highly resolving
data showing densely built-up urban scenes, which – in the
SAR case – is often difficult to interpret even for trained
experts.
Therefore, our work aims at learning a multi-sensor correspondence predictor for SAR and optical image patches
of state-of-the-art VHR data. Inspired by promising results
achieved in the context of stereo matching for optical imagery
[4], [5], we also make use of a convolutional neural network
(CNN). The major difference of our work to these purely
optical approaches is that we focus on the aforementioned,
distinctly more complicated multi-sensor setup, and therefore design a specific pseudo-siamese network architecture
with two separate, yet identical convolutional streams for
processing SAR and optical patches in parallel, instead of
a weight-shared siamese network in order to deal with the
heterogeneous nature of the input imagery.
II. T HE N ETWORK A RCHITECTURE
A. Pseudo-Siamese Convolutional Network
Since SAR and optical images lie on different manifolds,
it is not advisable to compare them directly by descriptors
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IN PRESS
IUT\E
2
IUT\E
IUT\E
IUT\E
IUT\E
UVZOIGRVGZIN
IUT\E
IUT\E
IUT\E
Details for Config.
^
^
^
^
^
^
^
^
^
^
^
^
^ ^
IUT\
RGHKR
^
^
^
^
9'8VGZIN
Conv 2_2 + ReLU + BN
stride 2
Max Pooling 1
Max Pooling 1
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_3 + ReLU + BN
Conv 2_3 + ReLU + BN
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_4 + ReLU + BN
Conv 2_4 + ReLU + BN
stride 2
Max Pooling 2
Max Pooling 2
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_5 + ReLU + BN
Conv 2_5 + ReLU + BN
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_6 + ReLU + BN
Conv 2_6 + ReLU + BN
stride 2
Max Pooling 3
Max Pooling 3
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_7 + ReLU + BN
Conv 2_7 + ReLU + BN
conv. 3x3 (l2 norm), stride 1, padding 1
Conv 1_8 + ReLU + BN
Conv 2_8 + ReLU + BN
conv. 3x3 (l2 norm), stride 2, padding 1
Conv 9 + ReLU + BN
conv. 3x3 (l2 norm), stride 2, padding 1
Conv 10 + ReLU + BN
^
IUT\E
Max Pooling 4
FC 1 + ReLU
IUT\E
Conv 2_1 + ReLU + BN
Conv 1_2 + ReLU + BN
stride 2
^
^
^
IUT\E
^
^
^
IUT\E
^
^
^
^
^ ^
IUT\E
Optical Conv. Stream
Conv 1_1 + ReLU + BN
conv. 3x3 (l2 norm), stride 1, padding 1
Concatenation Layer
^
^
^
^
IUT\E
LI
IUTIGZ
IUT\E
IUT\
^
^
^
^
LI
SAR Conv. Stream
conv. 3x3 (l2 norm), stride 1, padding 1
FC 2 + Softmax
Loss
IUT\E
Fig. 2. Proposed pseudo-Siamese network architecture and layer configuration.
designed for matching optical patches. Neither suitable are
conventional siamese CNN architectures (e.g. [5]), which are
designed to share their weights in each layer, as the input
data processed in the two feature extraction streams share
similar properties. In order to cope with the strongly different
geometric and radiometric appearances of SAR and optical
imagery, in [6] we proposed a pseudo-siamese network architecture with two separate, yet identical convolutional streams,
which process the SAR patch and the optical patch in parallel,
and only fuse the resulting information at a later decision
stage. Using this architecture, the network is constrained to
first learn meaningful representations of the input SAR patch
and the optical patch separately, and to combine them on a
higher level. The work presented in this letter is an extension
of [6] by improving the fusion part of the network architecture,
using a different training strategy, and resorting to non-locally
prefiltered SAR patches instead of temporal mean maps.
In addition, we evaluate the network on a deterministically
partitioned dataset instead of a randomly partitioned one, as
random partitioning will always cause positively biased results
due to overlapping regions in patches.
The architecture of the proposed network is shown in Fig. 2.
It is mainly inspired by the philosophy of the well-known
VGG Nets [7]. The SAR and optical image patches are passed
through a stack of convolutional layers, where we make use of
convolutional filters with a very small receptive field of 3 × 3,
rather than using larger ones, such as 5×5 or 7×7. The reason
is that 3×3 convolutional filters are the smallest kernels to capture patterns in different directions, such as center, up/down,
and left/right, but still have an advantage: the use of small
convolutional filters will increase the nonlinearities inside the
network and thus make the network more discriminative
The convolution stride in our network is fixed to 1 pixel;
the spatial padding of convolutional layer input is such that
the spatial resolution is preserved after convolution, i.e. the
padding is 1 pixel for the all 3 × 3 convolutional layers in our
network. Spatial pooling is achieved by carrying out seven
max-pooling layers which follow some of the convolutional
layers. They are used to reduce the dimensionality of the
feature maps. Max-pooling is performed over 2 × 2 pixel
windows with stride 2.
The fusion stage of our proposed network is made up of
two consecutive convolutional layers, followed by two fully
connected layers. The convolutional layers consist of 3 × 3
filters which operate over the concatenated feature maps of the
SAR and optical streams, in order to learn a fusion rule which
minimizes the final loss function. Max-pooling is omitted after
the first convolutional layer in the fusion stage, and a stride
of 2 is used in order to downsample the feature maps while
preserving the spatial information [8]. The use of 3 × 3 filters,
and the absence of max-pooling after the first convolution allows the fusion layer to learn a fusion rule which is somewhat
invariant to spatial mismatches caused by the difference in
imaging modalities. This is due to the fact that the fusion
layer uses 3x3 convolutions to learn relationships between the
features while preserving nearby spatial information. The lack
of max pooling means that these learned spatial relationships
are preserved as not only the maximal response is considered,
while the stride of 2 is used to reduce the feature size. The final
stage of the fusion network consists of two fully connected
layers: the first of which contains 512 channels; while the
second, which performs one-hot binary classification, contains
2 channels.
In a nutshell, the convolutional layers in our network apart
from the fusion layer generally consist of 3 × 3 filters and
follow two rules: 1) The layers with same feature map size
have the same number of filters; and 2) the number of feature
maps increases in the deeper layers, roughly doubling after
each max-pooling layer (except for the last convolutional
stack in each stream). All layers in the network are equipped
with a rectified linear unit (ReLU) as activation function,
except the last fully connected layer, which is activated by
a softmax function. Figure 2 shows the schematic diagram of
the configuration of our network.
B. Loss Function
We
make
use
of
the
binary
cross-entropy
loss
for
training
our
network.
Let
X
=
opt
opt
sar
sar
opt
(xsar
,
x
),
(x
,
x
),
...,
(x
,
x
)
be
a
set
of
SARn
n
1
2
1
2
opt
optical patch pairs, where xsar
∈ RD×D , ∀i = 1, .., n
i , xi
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IN PRESS
opt
and yi is the one-hot label for the pair (xsar
i , xi ) (with
[1, 0] denoting a dissimilar pair, and [0, 1] denoting a similar
pair). We then seek to minimize the binary cross-entropy loss
n
1X
opt
(1)
yi · log f (xsar
E=
i , xi , θ)
n i=1
opt
f (xsar
i , xi , θ)
3
SAR Image Stack
(2-D viewing angles)
TomoSAR
Inversion
Stereo
Matching
TomoSAR Point Cloud
Optical Point Cloud
(3-D UTM)
denotes the output vector of the softmax
where
opt
function when comparing the input pair (xsar
i , xi ) with the
current network parameters θ.
C. Configuration Details
Figure 2 depicts the full configuration of our network. Apart
from the previously discussed architecture we also make use of
batch normalization (BN) after the activation function of each
convolutional layer. This leads to an increase in the training
speed and reduces the effects of internal covariate shift. In
order to reduce over-fitting during training, we made use of
L2 -regularization, with λ = 0.001, for the convolution kernels
of the SAR and optical streams, and dropout with a rate of
0.7 for the first fully connected layer.
III. AUTOMATIC PATCH P OOL G ENERATION
For training and testing purposes, a large pool of corresponding and non-corresponding SAR and optical image
patches is needed. While the classical work on deep matching
for optical imagery can usually rely on easy-to-achieve optical
patch pools (see, for example the Phototourism patch dataset
[9], [4]), annotating corresponding patches in VHR optical
and SAR imagery of complex urban scenes is a highly nontrivial task even for experienced human experts. Thus, one of
the contributions of this letter is the introduction of a fully
automatic procedure for SAR-optical patch pool generation.
A. The “SARptical” Framework
In order to solve the challenge of automatic dataset generation, we resort to the so-called “SARptical” framework of
Wang et al. [10], an object-space-based matching procedure
developed for mapping textures from optical images onto 3D
point clouds derived from SAR tomography. The core of
this algorithm is to match the SAR and the optical images
in 3D space in order to deal with the inevitable differences
caused by different geometrical distortions. Usually, this would
require an accurate digital surface model (DSM) of the area
to link homologue image parts via a known object space. In
contrast, the approach in [10] creates two separate 3D point
clouds – one from SAR tomography and one from optical
stereo matching –, which are then registered in 3D space to
form a joint (“SARptical”) point cloud, which serves as the
necessary representation of the object space. The flowchart of
the approach can be seen in Fig. 3. In order to estimate the 3D
positions of the individual pixels in the images, the algorithm
requires an interferometric stack of SAR images, as well as at
least a pair of optical stereo images. The matching of the two
point clouds in 3D guarantees the matching of the SAR and
the optical images. Finally, we can project the SAR image into
the geometry of the optical image via the “SARptical” point
cloud, and vice versa.
Optical Images
(2-D range, azimuth)
(3-D UTM)
3-D
Matching
Matched Point Clouds
(3-D UTM)
Fig. 3. Flowchart of the patch-pool generation procedure.
Fig. 2
B. Data Preparation
For the work presented in this letter, we made use of a stack
of 109 TerraSAR-X high resolution spotlight images of the city
of Berlin, acquired between 2009 and 2013 with about 1 meter
resolution, and of 9 UltraCAM optical images of the same
area with 20cm ground spacing. After the reconstruction of
the “SARptical” 3D point cloud, 8,840 pixels with high SNR
(> 5dB) were uniformly sampled from the non-locally filtered
master SAR amplitude image and projected into the individual
optical images, yielding a total of 10,108 corresponding optical
pixels. The reason for the difference in pixel numbers is that
each of the 9 optical multi-view stereo images is acquired
from a different viewing angle, making it possible for each
SAR image pixel to have up to 9 corresponding optical image
pixels. The actual number of corresponding optical pixels
is dependent on the visibility of the SAR pixel from the
respective optical point of view.
All SAR patches are centered at their corresponding SAR
image pixels. Their size is fixed at 112 × 112 pixels, with a
pixel spacing of about 1 m. In analogy, the optical patches are
centered at the corresponding optical pixels. After resampling
to adjust the pixel spacing, the SAR patches were rotated, so
that both patches align with each other as a first approximation.
In order to reduce bias when training our network we
randomly selected a single correct optical correspondence for
each SAR image patch during the final dataset preparation.
In addition, we randomly assign one wrong optical correspondence to each SAR patch in order to create negative examples.
Thus, eventually, we end up with 17,680 SAR-optical patch
pairs (see Fig. 1 for an example for the class of correct
matches).
As final pre-processing steps, the optical patches were
converted to gray-scale, and all patches were normalized [11]
to a radiometric range of [0; 1] with subsequent subtraction of
their means.
C. Patch Pool Partitioning
In order to provide a fair experimental design, we partition
the patch pool in the following manner: 9724 (55%) of
the patch-pairs are used as training dataset, 2652 (15%) as
validation set, and 5304 (30%) as test dataset. It has to be noted
that we do not partition the patch pool on a purely randomized
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IN PRESS
IV. E XPERIMENTS & R ESULTS
A. Training Details
The network was trained using the Adam [12] optimization
algorithm as it is computationally efficient and exhibits faster
convergence than standard stochastic gradient descent methods. The optimization hyper-parameters are fixed to β1 = 0.9,
β2 = 0.999 with a learning rate of αt = 0.0009. The learning
rate was found via a grid search method on our training and
validation data, while the β−parameters were kept at their
recommended values. Prior to training the network weight vectors were initialized using the truncated uniform distribution
described in [13], and the bias vectors were initialized with
zero values. Training was conducted with 2 Nvidia TitanX
GPUs using class balanced, mini-batches of 64 SAR-optical
patch pairs (32 corresponding, 32 non-corresponding pairs)
over 30 epochs; training took on average 25 minutes, with a
single forward pass taking around 3ms to complete.
We trained five versions of our proposed network, each at
a different patch size, in order to evaluate the effect of patch
size on classification accuracy. Patch cropping was done onthe-fly with the new patch being cropped from the center
of a larger patch - this was done as the center pixel is the
point of correspondence between the SAR and optical patch.
Furthermore, we seeded our random number generator with a
fixed value of 0, at the start of training for each patch size, in
order to prevent the randomization effects between networks.
the performance of our proposed network in the presence of
large numbers of potential mismatches.
In Fig. 5a we can see a matrix depicting the similarity scores
of the various pair comparisons, where corresponding SAR
and optical patches are given the same index number. It should
be noted that in determining a binary value for correspondence,
a threshold is applied to these similarity scores. Figure 5b
depicts the sorted scores for all non-similar optical patches,
making it easier to see the number and strength of incorrect
matches in the patch pool.
TABLE I
C ONFUSION MATRIX VALUES FOR DIFFERENT PATCH SIZES
TP
TN
64
76
88
100
112
46.6%
86.2%
61.6%
88.0%
66.0%
88.2%
69.8%
86.0%
82.2%
89.8%
Accuracy Curve for Various Patch Sizes
0.9
64x64
76x76
88x88
100x100
112x112
0.85
0.8
Accuracy
basis but rather resort to a deterministic partitioning method
in order to avoid positively biased test results. The full extent
SAR and optical images are first deterministically partitioned
and then each partition is processed to generate positive and
negative samples for training, validation and testing respectively.
4
0.75
0.7
0.65
0.6
0.55
0.5
0
0.1
0.2
0.3
0.4
0.5
0.6
False Positive Rate
0.7
0.8
0.9
1
Fig. 4. Comparison of different patch sizes.
B. Evaluation Results
We evaluate the proposed network with different input patch
sizes, using our testing patch pool (described in Section III)
which has further been cropped around the center pixel to
produce new testing pools with different patch sizes.
The accuracy vs. false positive rate curves corresponding
to different patch sizes can be seen in Fig. 4. Table I reports
the corresponding confusion matrix values for our proposed
network evaluated with each patch size; it is to be noted that
the confusion matrix is reflective of the network at the point
of highest overall performance for each patch size.
a
b
Fig. 5. Results of key-point matching experiment. a) A confusion matrix
showing the matching scores for all SAR and optical key-point patches. b)
depicts the spread of incorrect matches, ordered by the similarity score.
V. D ISCUSSION
C. Key-Point Matching Results
In order to evaluate the proposed network’s performance
in a real-world, key-point matching scenario we selected 100
neighboring TomoSAR key-points in the SAR image and
extracted the corresponding SAR and optical patch pairs. We
selected these key-points from a localized area within our test
set, so as to reproduce the conditions found in a real-world
key-point matching application. We then compared every SAR
and optical patch in the selected patch set in order to determine
Generally, the results summarized in Section IV-B indicate
a promising discriminative power of the proposed network.
However, the following major points must be considered when
interpreting the results.
A. Influence of the Patch Size
As Tab. I and Fig. 4 clearly indicate, the patch size strongly
affects the discriminative power of the network. This result is
likely due to the effects of distortions in SAR images, which
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IN PRESS
True Positives
False Positives
5
False Negatives
True Negatives
Fig. 6. Exemplary patch correspondence results.
are acquired in a range-based imaging geometry. Thus when
patches are cropped to smaller regions we likely crop out
defining features which are used for matching between the
SAR and optical domain. This can be intuitively understood
by referring to Fig. 1 where we can see the effects of layover
and multi-path reflections of the building in the SAR image,
and a near top down view of the same building in the optical
image. Taking away explanatory context will thus render the
matching more difficult. All further discussion will be with
reference to the results we obtained using the largest patch
size, 112 pixels.
B. Comments on the Discriminative Power of the Proposed
Network
In summary, our approach obtains an accuracy exceeding
77% on a separate test dataset when fixing the false positive
rate to 5%, which falls into the same order of magnitude as
what can be achieved using the powerful handcrafted HOPC
descriptor in combination with an L2 -norm cost function [3].
Furthermore, our approach produced a clear diagonal pattern
in Fig. 5a which depicts its ability to accurately determine
the correct correspondence in a key-point matching scenario.
Upon further investigation it was found that the network
achieved 43% top-1 matching accuracy and 74% top-3 accuracy, while 8% of points had no valid matches detected within
the key-point set. This was found to be due to large amounts
of layover and extreme differences in view point between the
SAR and optical patches, see Fig. 6 False Negatives.
C. Possible Reasons for False Predictions
From the randomly chosen prediction examples displayed
in Fig. 6 it can be observed that many of the false positives
and false negatives are erroneously matched under extreme
differences in viewing angle between the SAR and optical
patches. While this may partially be solvable by resorting to
larger patch sizes, providing valuable context, there might be
a need to exclude image parts with all too strong distortions
from further processing.
VI. C ONCLUSION
In this paper, a pseudo-siamese CNN for learning to identify
corresponding patches in very high resolution SAR and optical
images in a fully automatic manner has been presented. A
first evaluation has shown promising potential with respect
to multi-sensor key-point matching procedures. In order to
ensure transferability to other applications not based on keypoints, e.g. dense matching, we will work on the generation of
additional training patches, whose center pixel does not rely on
specific key-points. In addition, we will test the approach on
data coming from a completely different sources. In the end,
we expect our work to help paving the way for generalized
SAR-optical image matching procedures.
ACKNOWLEDGMENT
We thank the NVIDIA Corporation for donating the Titan
X Pascal GPU used in this research.
R EFERENCES
[1] K. Mikolajczyk and C. Schmid, “A performance evaluation of local
descriptors,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 10,
pp. 1615–1630, 2005.
[2] M. Schmitt and X. X. Zhu, “Data fusion and remote sensing: An evergrowing relationship,” IEEE Geosci. Remote Sens. Mag., vol. 4, no. 4,
pp. 6–23, 2016.
[3] Y. Ye, J. Shan, L. Bruzzone, and L. Shen, “Robust registration of
multimodal remote sensing images based on structural similarity,” IEEE
Trans. Geosci. Remote Sens., vol. 55, no. 5, pp. 2941–2958, 2017.
[4] X. Han, T. Leung, Y. Jia, R. Sukthankar, and A. C. Berg, “MatchNet:
Unifying feature and metric learning for patch-based matching,” in Proc.
CVPR, Boston, U.S.A., 2015.
[5] S. Zagoruyko and N. Komodakis, “Learning to compare image patches
via convolutional neural networks,” in Proc. CVPR, Boston, U.S.A.,
2015.
[6] L. Mou, M. Schmitt, Y. Wang, and X. Zhu, “A CNN for the identification
of corresponding patches in SAR and optical imagery of urban scenes,”
in Proc. JURSE, Dubai, U.A.E., 2017.
[7] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” in Proc. ICLR, arXiv:1409.1556, San
Diego, U.S.A., 2014.
[8] A. Handa, M. Bloesch, S. Stent, and A. Davison, “gvnn: Neural network
library for geometric computer vision,” in Proc. ECCV Workshops,
Amsterdam, Netherlands, 2016, pp. 67–82.
[9] S. A. J. Winder and M. Brown, “Learning local image descriptors,” in
Proc. CVPR, Minneapolis, U.S.A., 2007.
[10] Y. Wang, X. Zhu, B. Zeisl, and M. Pollefeys, “Fusing meter-resolution
4-D InSAR point clouds and optical images for semantic urban infrastructure monitoring,” IEEE Trans. Geosci. Remote Sens., vol. 55, no. 1,
pp. 1–13, 2016.
[11] Q. Wang, C. Zou, Y. Yuan, H. Lu, and P. Yan, “Image registration by
normalized mapping,” Neurocomputing, vol. 101, pp. 181–189, 2013.
[12] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
in Proc. ICLR, arXiv:1412.6980v9, San Diego, U.S.A., 2015.
[13] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” in Proc. AISTATS, Sardinia, Italy, 2010,
pp. 249–256.
| 1cs.CV
|
Deep Neural Network Optimized to Resistive Memory with
Nonlinear Current-Voltage Characteristics
Hyungjun Kim∗ , Taesu Kim∗, Jinseok Kim, and Jae-Joon Kim†
Department of Creative IT Engineering, POSTECH, Pohang, South Korea
arXiv:1703.10642v1 [cs.ET] 30 Mar 2017
Abstract
Artificial Neural Network computation relies on intensive vector-matrix multiplications. Recently, the emerging nonvolatile memory (NVM) crossbar array showed a feasibility of implementing such operations with high energy efficiency, thus there are many
works on efficiently utilizing emerging NVM crossbar array as analog vector-matrix multiplier. However, its nonlinear I-V characteristics restrain critical design parameters, such
as the read voltage and weight range, resulting in substantial accuracy loss. In this paper, instead of optimizing hardware parameters to a given neural network, we propose
a methodology of reconstructing a neural network itself optimized to resistive memory
crossbar arrays. To verify the validity of the proposed method, we simulated various
neural network with MNIST and CIFAR-10 dataset using two different specific Resistive Random Access Memory (RRAM) model. Simulation results show that our proposed
neural network produces significantly higher inference accuracies than conventional neural
network when the synapse devices have nonlinear I-V characteristics.
1
Introduction
In recent years, Artificial Neural Network (ANN) has been gaining significant interest by claiming several cutting-edge results in solving various nonlinear problems [1]. The breakthrough
of ANN heavily depends on the expansion of networks in depth, which requires vast amount
of vector-matrix multiplications. With the advent of vector-matrix multiplication acceleration
based on graphics processing units (GPUs), large and deep neural networks have been able
to handle complex tasks using extensive amounts of data [2]. However, despite the fact that
GPUs provide highly parallel computing suitable for ANNs, the high power consumption of
GPUs is an obstacle to be improved. To address the issue, many dedicated accelerators for
vector-matrix multiplications have been proposed [3–6].
Emerging nonvolatile memory (NVM) technologies including Phase-Change Random Access Memory (PCRAM), Resistive Random Access Memory (RRAM), Conductive-Bridge Random Access Memory (CBRAM), and Spin-Transfer-Torque Magnetic Random Access Memory
(STT-MRAM) have been widely studied as next generation memories [7]. While conventional
memories such as Static Random Access Memory (SRAM), and FLASH memory are chargebased, emerging NVM is current-based and represents states with different resistance values.
This current-based nature opens up the opportunity to use emerging NVM for neural network
acceleration. Current-based devices in a crossbar array structure can straightforwardly implement vector-matrix multiplication in neural network computations as shown in Fig. 1. By
∗
†
Hyungjun Kim and Taesu Kim eqaully contributed to this work.
To whom correspondence should be addressed; E-mail: [email protected]
1
Figure 1: An example of mapping vector-matrix multiplication to a RRAM crossbar array
mapping input vector to input voltages and weight matrix to resistive crossbar array, vectormatrix multiplication can be calculated in a single step by sampling the current flowing in
each column [8]. Since this approach can be several orders of magnitude more efficient than
CMOS ASIC approaches in terms of both speed and power [3–6], many studies proposed neural
network accelerators based on emerging NVM crossbar array [9–12].
However, there are several issues with using emerging NVM crossbar array as an analog
multiplier. Sneak path problem is one of the most critical issues [7, 11, 13, 14]. Various works
have tried to solve this problem in different ways [7, 14, 15]. The most common idea is to
use a device with nonlinear I-V characteristics. For example, it has been proposed to serially
connect a selector device such as a transistor or a diode to an emerging NVM cell or to make
the I-V characteristic of an emerging NVM cell as nonlinear as possible [7]. However, although
this approach can overcome the sneak path problem, it degrades the accuracy of current-based
vector-matrix multiplication because nonlinear I-V characteristics hinder precise implementation of linear multiplications required for vector-matrix multiplications. Several works tried to
solve this issue by restricting the range of the reading voltage to use the pseudo-linear subregion of the nonlinear I-V curve [16, 17]. However, limiting reading voltage worsened DAC
resolution issue, making it difficult to compute complex neural networks using emerging NVM
crossbar arrays. Another previous approach was to address the problem by tuning the weights
considering the computational error before mapping [4, 18]. However, it also failed to utilize
full input voltage range. In addition, these approaches did not fully address how the increase
of the nonlinearity of I-V characteristics affects the inference accuracy of neural networks.
Unlike previous approaches which attempted to precisely map pre-determined weights to
the crossbar array to reduce accuracy loss, we propose to rather construct an ANN model itself
which accomodates nonlinear I-V characteristics. This allows nonlinear I-V characteristics to
be taken into account during both of the learning phase and the inference phase of a neural
network, reducing discrepancies between neural network models and emerging NVM-based
hardwares. In this paper, we have selected two nonlinear RRAM devices as proof-of-concept
devices and performed simulations to verify the idea based on the characteristics of the devices.
The main contribution of this paper is as follows:
1) We analyze the correlation between the degree of nonlinearity of I-V characteristics
and the inference accuracy loss in RRAM-based ANN hardware. We show that the degree of
accuracy loss depends on the distribution of activation values.
2
Figure 2: (a) I-V curve of a real RRAM device in different resistance states. (b) Experimental data
and fitted curve of the device under various resistance states [19].
2) We propose a modified perceptron model that is compatible with the nonlinear I-V
characteristics of resistive memory devices and demonstrate how to train neural networks based
on the proposed model. We show that neural networks based on the proposed model can avoid
the loss of inference accuracy which happens while mapping the network to RRAM crossbar
arrays.
2
2.1
Preliminaries
Nonlinear I-V Characteristics of RRAM Devices
Fig. 2 illustrates the I-V characteristics of an actual metal-oxide RRAM device extracted
from [19]. Each line shows the I-V curve of the RRAM device given a specific sequence of
set-voltage pulses. As shown in Fig. 2(a), the I-V relationship of the RRAM device has an
exponential form. [19] suggests that the I-V characteristics can be modeled as a empirical model
with a sinh function
I(V ) = ed/d0 sinh(BV )
(1)
where d is average tunneling gap distance, d0 is a fitting parameter, and B is a constant.
Measurement results and the values from the empirical model match well as shown in Fig.
2(b). Each I-V curve corresponding to a different state can be obtained by appropriately
determining a state variable d.
The degree of nonlinearity of an I-V curve can be represented by half-bias nonlinearity k,
which is defined as
I(Vr,max )
(2)
k=
I(Vr,max /2)
where Vr,max is the maximum value of read voltage that can be used without disturbing the
state of the device. A k value does not guarantee unique I-V characteristics of a particular
device, as two different devices with the same k value can have different I-V curves. However,
it can be still said that k value represents the nonlinearity in some degree. In this paper, we
assume Vr,max = 1 V for simplicity without loss of generality. Then, k can be expressed as a
function of B as follows,
eB − e−B
k = B/2
(3)
e
− e−B/2
3
Figure 3: (a) I-V curves in sinh form with various degrees of nonlinearity. In this case parameter
B determines nonlinearity k. The model device has k = 7.5, leading to B = 4 in Eq. (1). (b)
Error between nonlinear I-V curves of RRAM devices and corresponding linear I-V curves of resistors.
Hollow symbols stand for the linear resistor model and filled symbols stand for the real device model.
When k = 2, the device has a linear I-V characteristic. As k increases, the I-V characteristic
of the corresponding device becomes more nonlinear and exhibits a larger current difference
from the linear I-V characteristic with the same resistance state (Fig. 3(a)). To investigate the
k values of existing RRAM devices, we surveyed several papers and manually extracted I-V
characteristic data of the proposed devices. We could observe that k has a wide distribution
ranging from 2.5 [20] to 70 [21, 22]. Among the devices, we chose a device with k = 7.5 as the
model device for the rest of the paper.
2.2
Weight Mapping
ANN utilizes a vector-matrix multiplication of the corresponding input vector and weight matrix
to obtain a weighted sum for a layer as
s=x·W
(4)
with s as the weighted sum vector, x as the input vector, and W as the weight matrix. To
implement Eq. (4) using an emerging NVM crossbar array, each weight in particular row and
column of the weight matrix must be mapped to a characteristic parameter of a corresponding
device in the crossbar array. In previous approaches [4, 16, 23] which used RRAM crossbar arrays, weights were mapped to the conductance of the devices assuming linear I-V characteristics
as follows,
(Gmax − Gmin )(w − wmax ) Gmax (wmax − wmin )
+
(5)
G(w) =
wmax − wmin
Gmax − Gmin
Eq. (5) takes a naive linear-mapping approach that maps the minimum weight to the
minimum conductance and maps the maximum weight to the maximum conductance. In this
approach, input vectors can be represented by a set of voltages applied to rows of the crossbar
array and the result of the vector-matrix multiplication can be obtained by sampling the current
in each column. Previous works mostly relied on the naive mapping and attempted to mimic
linear I-V characteristics by limiting the reading voltage into a small linear range [16, 24].
4
Figure 4: (a) Accuracy vs. nonlinearity (k) and (b) normalized accuracy loss vs. nonlinearity (k) for
shallow MNIST case (black square), deep MNIST case (red uptriangle), and shallow CIFAR-10 case
(blue diamond). (b) describes the accuracy loss relative to the baseline (linear case) accuracy.
3
Error analysis
In section 3, we analyze the inference error of the neural networks naively mapped to a RRAM
crossbar array consisting of the devices with nonlinear I-V characteristics. To analyze the errors
that occur while using full range of Vread , we simulated the consequences of applying the naive
mapping method to the devices with the I-V characteristics given in [19] (Fig. 2). Based on
the range of device current at 1 V, weights were linearly converted to conductance using Eq.
(5). The empirical I-V model given in Eq. (1) was used for inference simulations. Fig. 3(b)
shows the mapping strategy.
To acquire model weights for analysis, two different Multilayer Perceptron (MLP) networks
were trained with MNIST dataset. A shallow network (784-500-250-10, shallow MNIST case)
with two hidden layers and a deep network (784-2500-2000-1500-1000-500-10, deep MNIST
case) with five hidden layers were each trained with Stochastic Gradient Descent (SGD) using
MATLAB. Also, a shallow MLP (2352-4000-1000-4000-10, shallow CIFAR-10 case) was
trained with the same method to classify CIFAR-10 dataset for additional analysis.
For each neural network model, the inference accuracy was evaluated using the naive mapping discussed above. Several simulations were performed by sweeping the k values to investigate the relationship between the degree of the inference accuracy degradation and the
nonlinearity of the I-V characteristics. For each k, corresponding parameter B in the numerical
I-V model Eq. (1) could be retrieved using Eq. (3). Evaluation results are illustrated in Fig.
4. We could observe that overall inference accuracies decrease as the I-V characteristics of the
devices become more nonlinear. In addition, the inference accuracy of the deep MNIST case
began to drop at lower k compared to the shallow MNIST case. Another observation was that
the inference accuracy of the shallow CIFAR-10 case also started to fall at low k.
To analyze the cause of accuracy degradation, we investigated the relationship between
input value distribution and computation error. Because the current difference between linear
and nonlinear I-V curves is the largest at an input voltage of about 0.5 V (Fig. 3(b)), we
speculated that input values around 0.5 V are prone to errors.
To verify this claim, two distinct distributions were given as input vectors to the second
layer of the deep MNIST case. Input data A (Fig. 5(a)) was generated by truncating a normal
distribution of numbers so that all data were around 0 and 1. Input data B (Fig. 5(d)) was
5
Figure 5: Distribution of input data A (a) and B (d). Desired output and actual device output for
input data A (b,c) and B (e,f).
generated as a normal distribution of numbers with a mean of 0.5 to let all data be around 0.5.
The result showed that when input data A was fed to the layer, an output distribution similar
to that of the ideal vector-matrix multiplication appeared (Fig. 5(b), (c)). While, feeding input
data B to the layer resulted in an output distribution with large error (Fig. 5(e), (f)). Based
on this observation, we investigated two representative cases that can induce mid-range input
and cause inference accuracy degradation.
3.1
Impact of Network Depth
In the shallow MNIST case, the activation values of neurons tend to yield extreme values such
as 0 and Vread (Fig. 6(a), (c), (d)), since the network can easily fall into saturation due to the
lack of training parameters. In contrast, a deeper network with an increased number of training
parameters can reduce the probability of saturation and induce mid-range activation values.
Increased depth of a layer can make the accuracy even worse because computation error due
to mid-range activation values can accumulate on several layers. Fig. 6(e), (f) demonstrates
that the activation values of the third and fourth hidden layer of the deep MNIST case have
relatively more mid-range values compared to shallow networks. This explains why the shallow
MNIST case could maintain relatively high accuracy while the deep MNIST case was vulnerable
to the I-V nonlinearity, resulting in greater accuracy loss.
3.2
Impact of Input Data Distribution
Fig. 6(a), (b) present the distribution of randomly selected training data from the MNIST
dataset and the CIFAR-10 dataset. As the CIFAR-10 dataset consists of natural images with
various RGB data, it has more mid-range values compared to the MNIST dataset. Such an
input data distribution causes degradation in inference accuracy due to imprecise activation
values at the first layers of the network. Besides, hidden layers of the ANNs for the CIFAR10 dataset tend to generate mid-range values as the networks must extract complex features
to classify complex images. As a result, the neuron activation values for CIFAR-10 neural
6
Figure 6: Distributions of input data in (a) MNIST dataset and (b) CIFAR-10. Activation values of
(c) first hidden layer and (d) second hidden layer of shallow MNIST network, (e) third hidden layer
and (f) fourth hidden layer of deep MNIST network, and (g) first hidden layer and (h) second hidden
layer of shallow CIFAR-10 network.
networks have large portion of mid-range values even in the shallow network case (Fig. 6(g),
(h)).
4
Proposed Methodology
Previous approaches attempted to exploit the linear sub-range in the nonlinear I-V curve for
more accurate vector-matrix multiplications. However, its effectiveness was only explored for
particular empirical I-V models. DAC resolution also arose as a critical problem for the cases
with increased nonlinearity. Thus, there is a pressing need to address the I-V nonlinearity
without such limitations. Here, we take a totally reverse approach to solve the problem; we
suggest to reconstruct the neural network model itself to reflect the I-V nonlinearity by replacing
linear vector-matrix multiplications.
Section 4 proposes a method to build an optimized neural network based on a device model.
For the optimized neural network, the basic computation block for vector-matrix multiplications
of a neural network is replaced by the nonlinear I-V model of a given device. Any device with
nonlinear I-V characteristics can be adopted as far as the numerical model of the I-V curve is
differentiable. By reconstructing the neural network considering the nonlinear I-V model of a
given device, we aim at overcoming the causes of accuracy loss discussed in Section 3 without
limiting the functionality of crossbar arrays.
4.1
Network Construction
A perceptron produces its activation value using transfer function and activation function as
y = g(f (w, x))
(6)
with x as a vector of input values, w as a vector of weights, function f as the transfer function
and function g as the activation function. Conventional ANN uses weighted sum as the transfer
function,
X
f (w, x) =
w i xi
(7)
i
7
Figure 7: Conventional and proposed perceptron model
For the activation function, there are several choices such as sigmoid, tanh, and ReLU. In case
of a fully connected layer with k perceptrons for a output vector y and i inputs, input vector
x is fed to each k perceptrons with corresponding weight vectors wk . Since the weight vector
and the input vector are independent, this computation can be simplified to a vector-matrix
multiplication by concatenating the weight vectors into a vector matrix W,
y = g(x · W)
(8)
Different from conventional ANN which uses the weighted sum as the transfer function, we
propose to use the numerical model of the nonlinear I-V characteristics of a given device as the
transfer function of a perceptron. By introducing nonlinearity to the perceptron model itself,
we can reduce the gap between the neural network model and the nonlinear I-V characteristics.
As an example case, let us construct a neural network using the device from [19]. The empirical
model of the device is given as Eq. (1). The equation can be simplified to
I(G, V ) = Gsinh(BV )
(9)
As B is a characteristic constant of each RRAM device, there are two variables G and V which
determine the output current. Based on the observation, we can define a transfer function as
X
f (w, x) =
wi sinh(Bxi )
(10)
i
using conductance Gi as weight wi (Fig. 7). Because the weight vector and the input vector
are independent similar to the transfer function of the conventional ANN, we can also concatenate the weight vectors to simplify the computation of an output layer into a vector-matrix
multiplication as
y = g(sinh(Bx) · W)
(11)
Because conductance value is always positive, two adjacent columns of the crossbar array
are used to express a single column of weights. We decompose a single weight to a pair of
positive and negative sub-weights as
wi = wi+ − wi−
(12)
with wi+ as the positive sub-weight and wi− as the negative sub-weight. With this expression,
expected computation result can be obtained by subtracting the computation results from two
adjacent columns. The transfer function is given as
y = g(sinh(Bx) · W+ − sinh(Bx) · W− )
8
(13)
Besides the modifications, the proposed network use same activation functions, cost functions, optimizers and other components as conventional ANN. For the rest of the paper, we
used modified Rectified Linear Unit (ReLU) as the activation function and logistic regression
with softmax as the cost function. ReLU function was modified to have a upper bound as the
value of maximum read voltage. We used 1 as the upper bound since Vread = 1 V. Under the
condition, logical output value of a layer could be directly fed into the next layer as the input
voltage.
4.2
Training
The proposed network can be trained using gradient descent similar to conventional ANN.
Gradient for kth weight matrix in a n-layer network can be derived using chain rule as follows,
dE
dsn dyn−1 dsn−1
dyk+1 dsk+1
dE
=
·
·
·
· ··· ·
·
dWk
dsn dyn−1 dsn−1 dyn−2
dsk+1 dWk
(14)
where s stands for the result of the transfer function, y stands for the result of the activation
function, and E means the error according to the cost function used. After evaluating the
gradient of each weight matrix, gradient descent can be applied to the proposed network to
update the weights with µ as the learning rate:
Wk∗ = Wk − µ
dE
dWk
(15)
Each term in Eq. (14) can vary depending on the device I-V model used in the neural
network. Let us derive the terms for the example case discussed above.
dE
= (Otrue − Opredict )
dsn
(16)
The derivative term of error is given as the difference between desired output and predicted
output since the example case uses the conventional cross-entrophy loss with softmax function.
(
1, 0 ≤ sn ≤ 1
dyn
=
(17)
dsn
0, else
Eq. (17) shows the derivative term of the activation function used in the example case. At
sn = 0 and sn = 1, we assign 1 to the term for computation although the derivative cannot be
explicitly defined.
dsn
= BWn−1 · cosh(Byn−1 )
(18)
dyn−1
Eq. (18) is derived by taking the derivative of the transfer function with respect to the input
vector.
dsn
= sinh(Byn−1 )
(19)
dWn−1
Eq. (19) can be obtained by taking the derivative of the transfer function with respect to
the weight matrix. For both Eq. (18) and Eq. (19), Eq. (12) is not considered to simplify
the equation. However, simplified sub-weight decomposition can still be used by defining two
sub-weights in adjacent columns as
9
(
w, w ≥ 0
w+ =
0, w < 0
(20)
(
0, w ≥ 0
w− =
|w|, w < 0
(21)
given a weight w. With all the derivative terms above, the proposed network can be trained
with gradient descent algorithm.
5
Evaluation
Several evaluation networks based on the example case in Section 4 were trained and simulated
using MATLAB. A shallow (784-500-250-10) network and a deep (784-2500-2000-1500-1000-50010) network were trained with MNIST dataset. A shallow (2352-4000-1000-4000-10) network
for CIFAR-10 dataset was trained for additional analysis.
Because ed/d0 in Eq. (1) is simplified to G and used as the weight in the proposed network,
trained weights must be mapped back into the range of the term ed/d0 that actual device
exhibits. Manual fitting described in Fig. 2 showed that maximum value for ed/d0 is e−8 and
the minimum value is e−14 . k value was also measured as 7.5 through fitting.
Sub-weights were mapped to the available range via linear transformation. The linear
transformation was as follows,
d/d
0
emapped
= W± ∗
e−8 − e−14
+ e−14
+
−
max(W , W )
(22)
Then, simulated current was sampled from each sub-weight column. After subtracting the
sampled value of each negative sub-weight column from the sampled value of the corresponding
positive sub-weight column, output of the transfer function was retrieved with inverse function
of Eq. (22).
5.1
MNIST Inference Accuracy
Example MLP models were trained with SGD. For the shallow network, we used 30 epochs for
training. Learning rate was set to 5 × 10−6 at first and to 1 × 10−6 after 16 epochs. The deep
network was trained for 65 epochs. Learning rate for the deep network was set to 2 × 10−6
at first and 7 × 10−7 after 15 epochs. To demonstrate the robustness against nonlinear IV characteristics of various devices, model networks with various k values were also tested.
Same SGD was used but learning rate varied for each case. Fig. 8 demonstrates the inference
evaluation results.
The proposed network did not show noticeable degradation in accuracy for various k values
while networks based on naive mapping exhibited drastic accuracy loss as k increased. Accuracies for the case with the example device (k = 7.5) were as shown in Table 1. This result
shows that the proposed network can minimize the error demonstrated in Section 3.1.
Table 1: Evaluation results for device with k = 7.5 (%)
Shallow MNIST Deep MNIST Shallow CIFAR-10
Ideal case
94.8
97.43
54.97
Naive mapping
87.90
9.05
41.94
Proposed network
96.74
96.91
52.09
10
Figure 8: Inference accuracies of networks trained using the proposed method for various k values.
Simulation results from naive mapping are in solid lines and that from proposed mapping are in
symbols. Data for Shallow MNIST, Deep MNIST, and Shallow CIFAR-10 cases are in black, red, blue
color respectively. Proposed networks (symbol) do not show significant dependency on k value while
naively mapped networks (line) do.
5.2
CIFAR-10 Accuracy
Another MLP with the proposed model was trained for CIFAR-10 dataset as a proof-of-concept
model. The training set was randomly cropped into a set of 28x28 images and exposed to
random image distortions. The random distortions included horizontal flips and contrast and
brightness adjustments. Then, the images were divided by 255 to ensure input range between
0 and 1 because the raw data are unsigned 8-bit integers. The MLP was trained for this image
dataset using SGD for 100 epochs. Learning rate was fixed to 1 × 10−6 . MLP networks with
varying k values for CIFAR-10 dataset were also trained.
CIFAR-10 classification results also showed that the accuracy of the proposed network does
not vary much over wide range of the k value. As illustrated in Table 1, the proposed scheme
achieved better inference accuracy than the naive mapping case. Since MLP is not capable of
achieving high inference accuracy for such complex task, inference accuracies of both ideal and
proposed network were limited. Still, the result demonstrates that the proposed network can
also circumvent the error discussed in Section 3.2.
6
Application to more complex I-V model
In addition to the device model derived in [19], there are several other resistive devices and
corresponding empirical I-V models [15, 25]. While each device has a different I-V model, the
proposed method is generally applicable as far as the I-V model is differentiable.
To demonstrate the validity of this claim, we built and tested three MLPs using the device
model obtained by manually fitting the I-V characteristics of the device in [26] as the transfer
function. We intentionally made a complex empirical model to verify our assertion. The
empirical model is as follows,
I(w, V ) = eAw+B (eCV
11
w+D
− 1)
(23)
Figure 9: (a) Experimental data and fitted curve of another device under various resistance states [26].
(b) Accuracy comparison between previous and proposed approaches. This result shows that our
proposed method can also be applied to complex device models.
with A = −53.59, B = −37.058, C = 20, D = 0.2, and 0 < W < 0.15. Note that the
model is more complex than Eq. (1). Fitting result of the model was as shown in Fig. 9(a).
MLPs for MNIST and CIFAR-10 with same structures as the ones described earlier were built
and evaluated. To train the network using gradient descent, same chain rule as Eq. (14) was
derived. Since the cost function and activation function were the same, derivative terms Eq.
(16) and Eq. (17) remained the same. However, for other terms, the derivation process was
very different because the weight term and the input voltage term were correlated. Because of
the correlation, we had to consider the simplified sub-weight decomposition while choosing the
transfer function as
s = I(w+ , V ) − I(w− , V )
(24)
where Eq. (20) and Eq. (21) lead to
P
wn−1,i,j +D
D
eAwn−1,i,j +B (eCxn−1,i
− 1) − eB (eCxn−1,i − 1) , wn−1,i,j ≥ 0
i
sn,j = P
−wn−1,i,j +D
D
− e−Awn−1,i,j +B (eCxn−1,i
− 1) + eB (eCxn−1,i − 1) , wn−1,i,j < 0
(25)
i
for the jth neuron in nth layer. Since the transfer function must be computed element-wise,
derivative terms of the transfer function with respect to the input vector and the weight vector
had to be also expressed element-wise as
P
wn−1,i,j +D
wn−1,i,j +D
+Awn−1,i,j +B
eCxn−1,i
(Cxn−1,i
lnV − A), wn−1,i,j ≥ 0
dsn,j
i
= P
(26)
−wn−1,i,j +D
−wn−1,i,j +D
−Awn−1,i,j +B
dwn−1,i,j
eCxn−1,i
(Cxn−1,i
lnV − A), wn−1,i,j < 0
i
dsn,j
dxn−1,i
X
w
+D
wn−1,i,j +D−1 Cx n−1,i,j
+Awn−1,i,j +B−1
CxD
n−1,i
n−1,i +B ,
x
e
(Cwn−1,i,j + CD) − CDxD−1
n−1,i
n−1,i e
i
wn−1,i,j ≥ 0
= X
−wn−1,i,j +D
−wn−1,i,j +D−1 Cx
−Awn−1,i,j +B−1
CxD
n−1,i +B ,
−xn−1,i
e n−1,i
(−Cwn−1,i,j + CD) + CDxD−1
n−1,i e
i
wn−1,i,j < 0
(27)
12
Since Eq. (27) was not defined for x = 0, we substituted 0 as the term for such cases during the
training phase. Using the terms above, we could obtain the accuracy results as shown in Fig.
9(b). The results showed that the proposed method is applicable to very complex nonlinear
device I-V model.
7
Conclusion & Future Research
In this paper, we aimed for accurate computation of neural networks using emerging NVM
crossbar arrays. We first analyzed the cause of inaccuracy in using naive mapping method.
Simulation results showed that mid-range activation values induce computation error for complex tasks and networks. To overcome the accuracy degradation due to the nonlinear I-V
characteristics of emerging NVM devices, we proposed a method to construct neural networks
optimized to the characteristics. Neural networks based on empirical models of two RRAM
devices were trained and tested to classify MNIST and CIFAR-10 dataset using the proposed
approach. Results showed that proposed networks could achieve inference accuracies comparable to the baseline. Also, proposed networks did not show accuracy degradation for a variety
of nonlinearity values while naive mapping showed significant accuracy loss.
Classification accuracy for the shallow CIFAR-10 case was limited due to the inherent lack
of computing power of the MLP structure. Thus, next step of this research will be applying
the proposed methodology to more complex neural networks such as CNN and RNN. Also, we
did not take into account other non-ideal characteristics of emerging NVM arrays such as I-R
drop for simulations. Another next step of this work will be to simulate and analyze the effects
of such characteristics.
8
Acknowledgement
This work was in part supported by the Ministry of Science, ICT and Future Planning of
South Korea under the ”IT Consilience Creative Program” (IITP-R0346-16-1007) and ”NanoMaterial Technology Development Program” (No. 2016910249) through the National Research
Foundation of Korea (NRF). It is also in part supported by the Industrial Technology Innovation
Program (10067764) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea).
References
[1] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015.
[2] K. Chellapilla, S. Puri, and P. Simard, “High Performance Convolutional Neural Networks for Document
Processing,” Tenth International Workshop on Frontiers in Handwriting Recognition, 2006.
[3] T. Gokmen and Y. Vlasov, “Acceleration of deep neural network training with resistive cross-point devices:
Design considerations,” Frontiers in Neuroscience, vol. 10, no. JUL, pp. 1–13, 2016.
[4] M. Hu, J. P. Strachan, Z. Li, E. M. Grafals, N. Davila, C. Graves, S. Lam, N. Ge, R. S. Williams, and
J. Yang, “Dot-Product Engine for Neuromorphic Computing: Programming 1T1M Crossbar to Accelerate
Matrix-Vector Multiplication,” IEEE Design Automation Conference, p. p. 19, 2016.
[5] A. Shafiee, A. Nag, N. Muralimanohar, and R. Balasubramonian, “ISAAC : A Convolutional Neural
Network Accelerator with In-Situ Analog Arithmetic in Crossbars,” ISCA 2016, 2016.
[6] P. Chi, S. Li, C. Xu, T. Zhang, J. Zhao, Y. Liu, Y. Wang, and Y. Xie, “Prime: A novel processing-inmemory architecture for neural network computation in reram-based main memory,” in Proceedings of the
43rd International Symposium on Computer Architecture, pp. 27–39, IEEE Press, 2016.
[7] S. Yu and P.-Y. Chen, “Emerging memory technologies: recent trends and prospects,” IEEE Solid-State
Circuits Magazine, vol. 8, no. 2, pp. p43–p56, 2016.
[8] M. Hu, H. Li, Y. Chen, Q. Wu, G. S. Rose, and R. W. Linderman, “Memristor crossbar-based neuromorphic
computing system: A case study,” IEEE Transactions on Neural Networks and Learning Systems, vol. 25,
no. 10, pp. 1864–1878, 2014.
13
[9] M. Prezioso, F. Merrikh-Bayat, B. D. Hoskins, G. C. Adam, K. K. Likharev, and D. B. Strukov, “Training
and operation of an integrated neuromorphic network based on metal-oxide memristors.,” Nature, vol. 521,
no. 7550, pp. 61–4, 2015.
[10] L. Xia, T. Tang, W. Huangfu, M. Cheng, X. Yin, B. Li, Y. Wang, and H. Yang, “Switched by input: Power
efficient structure for rram-based convolutional neural network,” in Proceedings of the 53rd Annual Design
Automation Conference, p. 125, ACM, 2016.
[11] G. W. Burr, R. M. Shelby, A. Sebastian, S. Kim, S. Kim, S. Sidler, K. Virwani, M. Ishii, P. Narayanan,
A. Fumarola, et al., “Neuromorphic computing using non-volatile memory,” Advances in Physics: X, vol. 2,
no. 1, pp. 89–124, 2017.
[12] G. W. Burr, R. M. Shelby, S. Sidler, C. Di Nolfo, J. Jang, I. Boybat, R. S. Shenoy, P. Narayanan, K. Virwani,
E. U. Giacometti, B. N. Kurdi, and H. H., “Experimental demonstration and tolerancing of a large-scale
neural network (165 000 synapses) using phase-change memory as the synaptic weight element,” IEEE
Transactions on Electron Devices, vol. 62, no. 11, pp. 3498–3507, 2015.
[13] M. Prezioso, F. Merrikh-Bayat, B. Chakrabarti, and D. Strukov, “Rram-based hardware implementations
of artificial neural networks: progress update and challenges ahead,” in SPIE OPTO, pp. 974918–974918,
International Society for Optics and Photonics, 2016.
[14] M. A. Zidan, H. A. H. Fahmy, M. M. Hussain, and K. N. Salama, “Memristor-based memory: The sneak
paths problem and solutions,” Microelectronics Journal, vol. 44, no. 2, pp. 176–183, 2013.
[15] Y. Deng, P. Huang, B. Chen, X. Yang, B. Gao, J. Wang, L. Zeng, G. Du, J. Kang, and X. Liu, “Rram
crossbar array with cell selection device: A device and circuit interaction study,” IEEE Transactions on
Electron Devices, vol. 60, no. 2, pp. 719–726, 2013.
[16] P. Gu, B. Li, T. Tang, S. Yu, Y. Cao, Y. Wang, and H. Yang, “Technological exploration of rram crossbar
array for matrix-vector multiplication,” in The 20th Asia and South Pacific Design Automation Conference,
pp. 106–111, 2015.
[17] B. Li, P. Gu, Y. Shan, Y. Wang, Y. Chen, and H. Yang, “Rram-based analog approximate computing,”
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 34, no. 12, pp. 1905–
1917, 2015.
[18] M. Hu, J. P. Strachan, Z. Li, R. Stanley, et al., “Dot-product engine as computing memory to accelerate
machine learning algorithms,” in Quality Electronic Design (ISQED), 2016 17th International Symposium
on, pp. 374–379, IEEE, 2016.
[19] S. Yu, B. Gao, Z. Fang, H. Yu, J. Kang, and H. S. P. Wong, “A low energy oxide-based electronic synaptic
device for neuromorphic visual systems with tolerance to device variation,” Advanced Materials, vol. 25,
no. 12, pp. 1774–1779, 2013.
[20] S. H. Misha, N. Tamanna, J. Woo, S. Lee, J. Song, J. Park, S. Lim, J. Park, and H. Hwang, “Effect of
Nitrogen Doping on Variability of TaOx -RRAM for Low-Power 3-Bit MLC Applications,” ECS Solid State
Letters, vol. 4, no. 3, pp. P25–P28, 2015.
[21] J. Zhou, F. Cai, Q. Wang, B. Chen, S. Gaba, and W. D. Lu, “Very low-programming-current RRAM with
self-rectifying characteristics,” IEEE Electron Device Letters, vol. 37, no. 4, pp. 404–407, 2016.
[22] S. Gao, F. Zeng, F. Li, M. Wang, H. Mao, G. Wang, C. Song, and F. Pan, “Forming-free and self-rectifying
resistive switching of the simple Pt/TaOx/n-Si structure for access device-free high-density memory application.,” Nanoscale, vol. 7, no. 14, pp. 6031–8, 2015.
[23] M. Hu, H. Li, Q. Wu, G. S. Rose, and Y. Chen, “Memristor crossbar based hardware realization of BSB
recall function,” in Proceedings of the International Joint Conference on Neural Networks, pp. 498–503,
2012.
[24] P. O. Vontobel, W. Robinett, P. J. Kuekes, D. R. Stewart, J. Straznicky, and R. S. Williams, “Writing to
and reading from a nano-scale crossbar memory based on memristors,” Nanotechnology, vol. 20, no. 42,
p. 425204, 2009.
[25] K. Sonoda, A. Sakai, M. Moniwa, K. Ishikawa, O. Tsuchiya, and Y. Inoue, “A compact model of phasechange memory based on rate equations of crystallization and amorphization,” IEEE Transactions on
Electron Devices, vol. 55, no. 7, pp. 1672–1681, 2008.
[26] S. Park, J. Noh, M.-L. Choo, A. M. Sheri, M. Chang, Y.-B. Kim, C. J. Kim, M. Jeon, B.-G. Lee, B. H. Lee,
and H. Hwang, “Nanoscale RRAM-based synaptic electronics: toward a neuromorphic computing device.,”
Nanotechnology, vol. 24, no. 38, p. 384009, 2013.
14
| 9cs.NE
|
arXiv:1708.02565v2 [math.GR] 20 Mar 2018
DUAL ORE’S THEOREM ON DISTRIBUTIVE
INTERVALS OF FINITE GROUPS
SEBASTIEN PALCOUX
Abstract. This paper gives a self-contained group-theoretic
proof of a dual version of a theorem of Ore on distributive intervals
of finite groups. We deduce a bridge between combinatorics and
representations in finite group theory.
1. Introduction
Øystein Ore proved in 1938 that a finite group is cyclic if and only if
its subgroup lattice is distributive, and he extended one side as follows,
where [H, G] will be an interval in the subgroup lattice of the group G
(idem throughout the paper).
Theorem 1.1 ([4]). Let [H, G] be a distributive interval of finite
groups. Then there is g ∈ G such that hHgi = G.
This paper first recalls our short proof of Theorem 1.1 and then gives a
self-contained group-theoretic proof of the following dual version, where
G(V H ) will be the pointwise stabilizer subgroup of G for the fixed-point
subspace V H (see Definition 3.1).
Theorem 1.2. Let [H, G] be a distributive interval of finite groups.
Then there exists an irreducible complex representation V of G such
that G(V H ) = H.
We deduce a bridge between combinatorics and representations:
Corollary 1.3. The minimal number of irreducible components for
a faithful complex representation of a finite group G is at most the
minimal length ℓ for an ordered chain of subgroups
{e} = H0 < H1 < · · · < Hℓ = G
such that [Hi , Hi+1 ] is distributive (or better, bottom Boolean).
2010 Mathematics Subject Classification. 20D30, 05E15, 20C15, 06C15.
Key words and phrases. finite group; complex representation; subgroup lattice;
distributive; Boolean.
1
2
SEBASTIEN PALCOUX
It is a non-trivial upper bound involving the subgroup lattice only.
These results were first proved by the author as applications to finite
group theory of results on planar algebras [5, Corollaries 6.10, 6.11].
For the convenience of the reader and for being self-contained, this
paper reproduces some preliminaries of [1] and [5].
2. Ore’s theorem on distributive intervals
2.1. Basics in lattice theory. We refer to [7] for the notions of finite
lattice L, meet ∧, join ∨, subgroup lattice L(G), sublattice L′ ⊆ L,
interval [a, b] ⊆ L, minimum 0̂, maximum 1̂, atom, coatom, distributive
lattice, Boolean lattice Bn (of rank n) and complement b∁ (with b ∈ Bn ).
The top interval of a finite lattice L is the interval [t, 1̂], with t the
meet of all the coatoms. The bottom interval of a finite lattice L is the
interval [0̂, b], with b the join of all the atoms. A lattice with a Boolean
top interval will be called top Boolean; idem for bottom Boolean.
Lemma 2.1. A finite distributive lattice is top and bottom Boolean.
Proof. See [7, items a-i p254-255] which uses Birkhoff’s representation
theorem (a finite lattice is distributive if and only if it embeds into
some Bn ).
2.2. The proof.
Øystein Ore proved the following result in [4, Theorem 4, p267].
Theorem 2.2. A finite group G is cyclic if and only if its subgroup
lattice L(G) is distributive.
Theorem 1.1 is an extension by Ore of one side of Theorem 2.2 to any
distributive interval of finite groups [4, Theorem 7, p269].
Definition 2.3. An interval of finite groups [H, G] is said to be Hcyclic if there is g ∈ G such that hH, gi = G. Note that hH, gi = hHgi.
We will give our short alternative proof of Theorem 1.1 by extending
it to any top Boolean interval (see Lemma 2.1) as follows:
Theorem 2.4. A top Boolean interval [H, G] is H-cyclic.
Proof. The proof follows from the claims below.
Claim: Let M be a maximal subgroup of G. Then [M, G] is M-cyclic.
Proof: For g ∈ G with g 6∈ M, we have hM, gi = G by maximality.
Claim: A Boolean interval [H, G] is H-cyclic.
DUAL ORE’S THEOREM ON DISTRIBUTIVE INTERVALS
3
Proof: Let M be a coatom in [H, G], and M ∁ be its complement. By
the previous claim and induction on the rank of the Boolean lattice,
we can assume [H, M] and [H, M ∁ ] both to be H-cyclic, i.e. there are
a, b ∈ G such that hH, ai = M and hH, bi = M ∁ . For g = ab, a = gb−1
and b = a−1 g, so hH, a, gi = hH, g, bi = hH, a, bi = M ∨ M ∁ = G.
Now, hH, gi = hH, gi ∨ H = hH, gi ∨ (M ∧ M ∁ ) but by distributivity hH, gi ∨ (M ∧ M ∁ ) = (hH, gi ∨ Mi) ∧ (hH, gi ∨ M ∁ i). So
hH, gi = hH, a, gi ∧ hH, g, bi = G. The result follows.
Claim: [H, G] is H-cyclic if its top interval [K, G] is K-cyclic.
Proof: Consider g ∈ G with hK, gi = G. For any coatom M ∈ [H, G],
we have K ⊆ M by definition, and so g 6∈ M, then a fortiori hH, gi 6⊆
M. It follows that hH, gi = G.
The converse is false because hS2 , (1234)i = S4 whereas [S2 , S4 ] is not
top Boolean.
3. Dual Ore’s theorem on distributive intervals
3.1. Basics in Galois connections.
Definition 3.1. Let W be a representation of a group G, K a subgroup
of G, and X a subspace of W . We define the fixed-point subspace
W K := {w ∈ W | kw = w , ∀k ∈ K},
and the pointwise stabilizer subgroup
G(X) := {g ∈ G | gx = x , ∀x ∈ X}.
Lemma 3.2. Let G be a finite group, H, K two subgroups, V a representation of G and X, Y two subspaces of V . Then
(1) H ⊆ K ⇒ V K ⊆ V H ,
(2) X ⊆ Y ⇒ G(Y ) ⊆ G(X) ,
(3) V H∨K = V H ∩ V K ,
(4) H ⊆ G(V H ) ,
(5) V G(V H ) = V H ,
(6) [H ⊆ K and V K ( V H ] ⇒ K 6⊆ G(V H ) .
Proof. (1) and (2) are immediate.
(3) First H, K ⊆ H ∨ K, so V H∨K is included in V H and V K , so in
V H ∩ V K . Now take v ∈ V H ∩ V K , then ∀h ∈ H and ∀k ∈ K, hv =
kv = v, but any element g ∈ H ∨K is of the form h1 k1 h2 k2 · · · hr kr with
hi ∈ H and ki ∈ K, it follows that gv = v and so V H ∩ V K ⊆ V H∨K .
(4) Take h ∈ H and v ∈ V H . Then by definition hv = v, so H ⊆ G(V H ) .
(5) From (1) and (4) we deduce that V G(V H ) ⊆ V H . Now take v ∈ V H
4
SEBASTIEN PALCOUX
and g ∈ G(V H ) , by definition gv = v, so V H ⊆ V G(V H ) also.
(6) Suppose that K ⊆ G(V H ) , then V K ⊇ V G(V H ) = V H by (1) and (5).
Hence V K = V H by (1), contradiction with V K ( V H .
3.2. Induced representation.
Definition 3.3. Let G be a finite group and H a subgroup. Consider
the set G/H = {g1 H, . . . , gs H}, with g1 = e. Let V be a complex
representation
of H. The induced representation IndG
H (V ) is a space
L
i gi V (we identify eV with V ) on which G acts as follows:
X
X
g·(
gi vi ) =
gτ (i,g) (hi,g · vi )
i
i
with ggi = gτ (i,g) hi,g , τ (i, g) ∈ {1, . . . , s} and hi,g ∈ H.
Let h·, ·iG be the usual normalized inner product of finite dimensional
complex representations (up to equivalence) of a finite group G.
Lemma 3.4 (Frobenius reciprocity, [3] p62). Let G be a finite group
and H a subgroup. Let V (resp. W ) be a finite dimensional complex
representation of G (resp. of H). Let Ind(W ) be the induction to G
and Res(V ) the restriction to H, then hV, Ind(W )iG = hRes(V ), W iH .
Lemma 3.5. Let [H, G] be an interval of finite groups. Let V1 , . . . , Vr
be the irreducible complex representations of G (up to equivalence).
Then
r
X
|G : H| =
dim(Vi ) dim(ViH ).
i=1
Proof. The following proof is due to Tobias Kildetoft. Let 1G
H be the
trivial representation of H induced to G. On one hand, it has dimension
|G : H|, and on the other hand, this dimension is also
X
X
X
dim(Vi )hVi , 1G
dim(Vi )hVi , 1H iH =
dim(Vi ) dim(ViH ).
H iG =
i
i
i
The first equality follows from Frobenius reciprocity.
Lemma 3.6. Let [H, G] be an interval of finite groups with H 6= G.
Then there is a non-trivial irreducible complex representation V of G
such that V H 6= 0.
P
H
Proof. By Lemma 3.5,
i dim(Vi ) dim(Vi ) = |G : H| ≥ 2 because
H 6= G. The result follows.
DUAL ORE’S THEOREM ON DISTRIBUTIVE INTERVALS
5
Lemma 3.7. Let G be a finite group, K a subgroup and U a finite
dimensional complex representation of K. For any irreducible component V of the induction W = IndG
K (U), there exists Ũ ⊆ V equivalent
to U as a representation of K.
Proof. Direct by Frobenius reciprocity because hV, W iG = hV, UiK .
Lemma 3.8. Let [H, G] be an interval of finite groups, K ∈ [H, G] and
U a finite dimensional complex representation of K such that U H 6= 0.
Let W be the induction IndG
K (U). Let V be an irreducible component
of W . Then
(1) G(W H ) ⊆ K(U H ) ,
(2) V H =
6 0 and K(V H ) ⊆ K(U H ) ,
(3) if K(U H ) = H then G(W H ) = K(V H ) = H.
Proof. (1) U H ⊆ W H , so by Lemma 3.2(2), G(W H ) ⊆ G(U H ) . Now,
U H 6= 0, so by definition of the induction, g · U H ⊆ U if and only
if gK = K, if and only if g ∈ K. Thus G(U H ) ⊆ K(U H ) , and so
G(W H ) ⊆ K(U H ) .
(2) Take Ũ as for Lemma 3.7. Then 0 6= Ũ H ⊆ V H and K(V H ) ⊆
K(Ũ H ) = K(U H ) .
(3) By (1) and Lemma 3.2(4),
H ⊆ G(W H ) ⊆ K(U H ) = H,
so G(W H ) = H. Idem, by (2) and Lemma 3.2(4), K(V H ) = H.
3.3. The proof.
Definition 3.9. The group G is called linearly primitive if it admits an
irreducible complex representation V which is faithful, i.e. G(V ) = {e}.
Definition 3.10. The interval [H, G] is called linearly primitive if
there is an irreducible complex representation V of G such that G(V H ) =
H.
Lemma 3.11. A maximal interval [H, G] is linearly primitive.
Proof. By Lemma 3.6, there is a non-trivial irreducible complex representation V of G with V H 6= 0. By Lemma 3.2(4), H ⊆ G(V H ) . If
G(V H ) = G then V must be trivial (by irreducibility), so by maximality
G(V H ) = H.
Lemma 3.12. The interval [H, G] is linearly primitive if its bottom
interval [H, K] is so.
6
SEBASTIEN PALCOUX
W
Proof. Let [H, K] be the bottom interval of [H, G], i.e. K = i Ki with
K1 , . . . , Kn the atoms of [H, G]. By assumption, there is an irreducible
complex representation U of K such that K(U H ) = H. Let V be an
irreducible component of IndG
K (U). By Lemma 3.8(3), K(V H ) = H.
H
Ki
Now if ∃i such that V = V , then
Ki ⊆ K(V Ki ) = K(V H ) = H,
contradiction with H ( Ki . So ∀i V Ki ( V H . By Lemma 3.2(6), we
deduce that Ki 6⊆ G(V H ) ∀i, so by minimality G(V H ) = H.
A dual version of Theorem 2.4 is the following:
Theorem 3.13. A bottom Boolean interval [H, G] is linearly primitive.
Proof. By Lemma 3.12, we can reduced to Boolean intervals. We make
an induction on the rank of the Boolean lattice. The rank one case is
handled in Lemma 3.11. Assume that it is true at rank < n. We will
write a proof at rank n ≥ 2. Let K be a coatom of [H, G]. Then [H, K]
is Boolean of rank n − 1, so by assumption, it is linearly primitive,
thus there is an irreducible complex representation U of K such that
K(U H ) = H. For any irreducible component V of W = IndG
K (U), we
have
K ∧ G(V H ) = K(V H ) = H
by Lemma 3.8(3). Thus, by the Boolean structure, G(V H ) ≤ K ∁ because
K ∁ = K ∁ ∨H = K ∁ ∨(K∧G(V H ) ) = (K ∁ ∨K)∧(K ∁ ∨G(V H ) ) = K ∁ ∨G(V H ) .
But K is a coatom of [H, G], so K ∁ is an atom and then
G(V H ) ∈ {H, K ∁ }.
Assume that every irreducible component V of W satisfies G(V H ) = K ∁ .
ThereL
are irreducible complex representations V1 , . . . , Vr of G such that
W = i Vi , then by Lemma 3.8(3)
^
H = G(W H ) =
G(ViH ) = K ∁ ,
i
thus H = K ∁ , contradiction. So there is an irreducible component V
of W such that G(V H ) = H, and the result follows.
Theorem 1.2 follows directly from Theorem 3.13 and Lemma 2.1.
DUAL ORE’S THEOREM ON DISTRIBUTIVE INTERVALS
7
4. A bridge between combinatorics and representations
We restate Corollary 1.3 by using the notation H i instead of Hi because
it will be more convenient for the proof.
Corollary 4.1. The minimal number of irreducible components for
a faithful complex representation of a finite group G is at most the
minimal length ℓ for an ordered chain of subgroups
{e} = H 0 < H 1 < · · · < H ℓ = G
such that [H i , H i+1 ] is distributive (or better, bottom Boolean).
Proof. By Theorem 3.13 and Lemma 3.8(3), there are irreducible complex representations V1 , . . . , Vℓ of G such that H i H i−1 = H i−1 . Take
(Vi
)
L
W = ℓi=1 Vi . Then
ker(πW ) =
ℓ
^
ker(πVi ) =
=
i=1
G(Vi ) ≤
H ℓ−1H i−1
(Vi
)
= ··· =
ℓ
^
G(V H i−1 ) = (
i
ℓ−s
^
i=1
ℓ−1
^
G(V H i−1 ) )∧H ℓ−1
i
i=1
i=1
i=1
i=1
ℓ−1
^
ℓ
^
H ℓ−sH i−1 = · · · = H 0 = {e}.
(Vi
)
So W is faithful with ℓ irreducible components. The result follows.
Note that this upper bound involves the subgroup lattice only.
Remark 4.2. The modular maximal-cyclic group M4 (2) and the
abelian group C8 × C2 have the same subgroup lattice, but the first
is linearly primitive whereas the second is not. So the minimal number
in Corollary 4.1 cannot be determined by the subgroup lattice only.
Lemma 4.3. For H core-free, G is linearly primitive if [H, G] is so.
Proof. Let V be an irreducible complex representation of G such that
G(V H ) = H. Now, V H ⊆ V so G(V ) ⊆ G(V H ) , but ker(πV ) = G(V ) ,
it follows that ker(πV ) ⊆ H; but H is a core-free subgroup of G, and
ker(πV ) a normal subgroup of G, so ker(πV ) = {e}.
By Lemma 4.3, we can improve the bound of Corollary 4.1 by taking
for H 0 any core-free subgroup of H 1 (instead of just {e}), and we can
wonder whether it is the right answer in general, in particular:
Question 4.4. Is a finite group G linearly primitive if and only if there
is a core-free subgroup H with [H, G] bottom Boolean?
8
SEBASTIEN PALCOUX
It is true for any finite simple group S, because any proper subgroup
M is core-free, and by choosing it maximal, [M, S] is Boolean of rank
one. Moreover, we have checked by GAP [2] that it is also true for any
finite group G of order less than 512.
Remark 4.5. A normal subgroup N E G is a modular element in
L(G), see [6, p43]. If H is a subgroup of G such that ∀K ∈ (1, H],
K is not modular in L(G), then H is core-free. It follows that we can
also improve the bound of Corollary 4.1 completely combinatorially.
Nevertheless, by Remark 4.2, it cannot be the right answer in general.
5. Acknowledgments
This work is supported by the Institute of Mathematical Sciences,
Chennai. The author thanks an anonymous referee who suggested to
write this group-theoretic proof of dual Ore’s theorem.
References
[1] Mamta Balodi and Sebastien Palcoux, On Boolean intervals of finite groups, J.
Comb. Theory, Ser. A 157 (2018), 49-69, DOI 10.1016/j.jcta.2018.02.004.
[2] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.8.3,
2016. URL: http://www.gap-system.org.
[3] I. Martin Isaacs, Character theory of finite groups, Dover Publications, Inc.,
New York, 1994. MR1280461
[4] Øystein Ore, Structures and group theory. II, Duke Math. J. 4 (1938), no. 2,
247–269, DOI 10.1215/S0012-7094-38-00419-3. MR1546048
[5] Sebastien Palcoux, Ore’s theorem on subfactor planar algebras,
arXiv:1704.00745, 14pp. Under review.
[6] Roland Schmidt, Subgroup lattices of groups, De Gruyter Expositions in Mathematics, vol. 14, 1994.
[7] Richard P. Stanley, Enumerative combinatorics. Volume 1, 2nd ed., Cambridge
Studies in Advanced Mathematics, vol. 49, Cambridge University Press, 2012.
Institute of Mathematical Sciences, Chennai, India
E-mail address: [email protected]
| 4math.GR
|
1
EXPOSE the Line Failures following
a Cyber-Physical Attack on the Power Grid
S. Soltan is with the Department of Electrical Engineering, Princeton
University, NJ, 08544, and G. Zussman is with the Department of Electrical
Engineering, Columbia University, New York, NY, 10027.
E-mails: [email protected], [email protected]
Power Grid
Physical Infrastructure
Supervisory Control and Data
Acquisition (SCADA) system
Recent cyber attack on the Ukrainian grid in December
2015 [1] demonstrated the vulnerability of power grids to
cyber attacks. As indicated in the aftermath report of the
attack [1], once the attackers obtain access to the grid’s
Supervisory Control And Data Acquisition (SCADA) system,
they can delete, modify, and spoof the data as well as remotely
change the grid’s topology by activating the circuit breakers.
The power grid infrastructure is also vulnerable to physical
attacks. Such an attack occured in April 2014 in San Jose,
California, when snipers tried to shut down a substation simply
by shooting at its transformers [2]. Hence, a physical attack
on the power lines and the measurement devices can have a
similar affect to a cyber attack.
To analyze these vulnerabilities, in this paper, we study
cyber-physical attacks that affect both the power grid physical
infrastructure and its SCADA system. Fig. 1 shows the main
components of the power grids. An adversary can attack the
grid by damaging the power lines and measurement devices
with a physical attack, by remotely disconnecting the lines
and erasing the measurements data with a cyber attack, or by
performing a combination of the both.
Independent of the attack strategy, we assume that an adversary attacks an area by: (i) disconnecting some lines within
IED
IED
Data Conc.
IED
Data Conc.
Command
I. I NTRODUCTION
IED
Data
Index Terms—AC Power Flows, State Estimation, Line Failures
Detection, Cyber Attack, Physical Attack.
IED
Command
Abstract—Recent attacks on power grids demonstrated the
vulnerability of the grids to cyber and physical attacks. To analyze this vulnerability, we study cyber-physical attacks that affect
both the power grid physical infrastructure and its underlying
Supervisory Control And Data Acquisition (SCADA) system. We
assume that an adversary attacks an area by: (i) disconnecting
some lines within that area, and (ii) obstructing the information
(e.g., status of the lines and voltage measurements) from within
the area to reach the control center. We leverage the algebraic
properties of the AC power flows to introduce the efficient
EXPOSE Algorithm for detecting line failures and recovering
voltages inside that attacked area after such an attack. The
EXPOSE Algorithm outperforms the state-of-the-art algorithm
for detecting line failures using partial information under the
AC power flow model in terms of scalability and accuracy. The
main advantages of the EXPOSE Algorithm are that its running
time is independent of the size of the grid and number of line
failures, and that it provides accurate information recovery under
some conditions on the attacked area. Moreover, it approximately
recovers the information and provides the confidence of the
solution when these conditions do not hold.
Data
arXiv:1709.07399v1 [cs.SY] 21 Sep 2017
Saleh Soltan, Member, IEEE, and Gil Zussman, Senior Member, IEEE
Control Center
Fig. 1: The main components of the power grid. The Intelligent
Electronic Devices (IED) measure voltage magnitudes and phase
angles, and send these information via Data Concentrators to the
Control Center.
that area (failed lines), and (ii) obstructing the information
(e.g., status of the lines and voltage measurements) from
within the area to reach the control center. We call this area,
the attacked zone. Our objective is to detect the failed lines
and recover the voltages inside the attacked zone using the
information available outside of the attacked zone as well
as the information before the attack. An example of such an
attack on the IEEE 300-bus system is depicted in Fig. 2.
We studied a similar attack scenario in [3] using the linearized DC power flows. In a recent extension [4], the methods
in [3] were modified to statistically recover the information
under the AC power flows. However, due to the inaccuracy of
the DC power flows, the methods in [4] could not guarantee
the correct information recovery under the AC power flows.
In this paper, we directly leverage the properties of the nonlinear AC power flows to detect the line failures and recover
the voltages after an attack with guarantee of performance.
In particular, we prove that if there is a matching between
the nodes (buses) inside and outside of the attacked zone that
2
covers all the nodes inside the attacked zone, the voltages can
be accurately recovered by solving a set of linear equations.
Moreover, given the successful recovery of the voltages, we
prove that if the attacked zone is acyclic (i.e., lines in the
attacked zone do not form any cycles), then the failed lines
can be accurately detected by solving a set of linear equations.
We extend these results and show that given the successful
recovery of the voltages, the failed lines can still be accurately
detected by solving a Linear Program (LP), even if the attacked
zone is not acyclic. We further show that even if there is no
matching between the nodes inside and outside of the attacked
zone that covers all the inside nodes and the attacked zone is
not acyclic, one can still approximately recover the voltages
and detect the line failures using convex optimization.
Based on the results, we then introduce the EXPress line
failure detection using partially ObSErved information (EXPOSE) Algorithm. It outperforms the state-of-the-art algorithm
for detecting line failures using partial information under the
AC power flows in terms of scalability and accuracy. The main
advantages of the EXPOSE Algorithm are that its running
time is independent of the size of the grid and number of
line failures, and that it provides accurate information recovery under some conditions on the attacked zone. Moreover,
it approximately recovers the information and provides the
confidence of the solution when these conditions do not hold.
Most of the related work rely on the DC power flows and
deploy brute force search approaches. These approaches do not
scale well, and therefore, are limited only to detecting single
and double line failures using partial measurements [5], [6],
[7], [8], [9]. To represent these approaches and for comparison
purposes only, we also introduce a naive Brute Force Search
(BFS) Algorithm for detecting line failures after the attack.
Finally, while we analytically prove that the EXPOSE
Algorithm guarantees to accurately recover the voltages and
detect line failures under some conditions, we also numerically
evaluate its performance when those conditions do not hold.
In particular, we evaluate the performance of the EXPOSE
Algorithm as the attacked zone becomes topologically more
complex and compare its running time to the BFS Algorithm
by considering all single, double, and triple line failures in
5 nested attacked zones. Based on the simulation results, we
conclude that despite its accuracy, the BFS Algorithm is not
practical for line failures detection in large networks and that
the EXPOSE Algorithm can provide relatively accurate results
exponentially faster. For example, the EXPOSE algorithm
recovers the voltages with less than 15% error and detects
line failures with less than 1 false negative on average, after
all single, double, and triple line failures in an attacked zone
that satisfies none of the conditions for the accuracy of the
EXPOSE Algorithm.
II. R ELATED W ORK
Vulnerability of power grids to failures and attacks has been
widely studied [10], [11], [12], [13], [14], [15], [16], [17], [18],
[19]. In particular false data injection attacks on power grids
and anomaly detection have been studied using the DC power
flows in [20], [21], [22], [23], [24], [25]. These studies focused
on the observability of the failures and attacks in the grid.
The problem studied in this paper is similar to the problem
of line failures detection using phase angle measurements [5],
[6], [7], [26]. Up to two line failures detection, under the
DC power flow model, was studied in [5], [6]. Since the
provided methods in [5], [6] are greedy-based methods that
need to search the entire failure space, the running time of
these methods grows exponentially as the number of failures
increases. Hence, these methods cannot be generalized to
detect higher order failures. Similar greedy approaches with
likelihood detection functions were studied in [8], [9], [27],
[28], [29] to address the PMU placement problem under the
DC power flow model.
The problem of line failures detection in an internal system
using the information from an external system was also studied
in [7] based (again) on the DC power flow model. The
proposed algorithm works for only one and two line failures,
since it depends on the sparsity of line failures.
In a recent work [26], a linear multinomial regression model
was proposed as a classifier for a single line failure detection
using transient voltage phase angles data. Due to the time
complexity of the learning process for multiple line failures,
this method is impractical for detecting higher order failures.
Moreover, the results provided in [26] are empirical with no
performance guarantees.
Finally, in a recent series of works, the vulnerability of
power grids to undetectable cyber-physical attacks is studied [30], [31], [32] using the DC power flows. These studies
are mainly focused on designing attacks that affect the entire
grid and therefore may remain undetected.
To the best of our knowledge, our methods presented in this
paper and [4] are the only methods for line failures detection
under the AC power flows that can be used to detect any
number of line failures and scale well with size of the grid.
However, the EXPOSE Algorithm provided in this paper is
more accurate than the method provided in [4].
III. M ODEL AND D EFINITIONS
A. AC Power Flow Equations
A power grid with n nodes (buses) and m transmission
lines can be represented by an undirected graph G(N , L),
where N = {1, 2, . . . , n} denotes the set of nodes and
L = {l1 , l2 , . . . , lm } denotes the set of lines or edges. In the
steady-state, the status of each node i is represented by its
voltage Vi = |Vi |eiθi in which |Vi | is the voltage magnitude,
θi is the voltage phase angle, and i denotes the imaginary unit.
The goal of the AC power flow analysis is the computation
of the voltage magnitudes and phase angles at each bus in
steady-state conditions [33]. In the steady-state, the AC power
flow equations can be written in matrix form as follows:
YV = I,
(1)
S = diag(V )I ∗ ,
(2)
where ∗ denotes the complex conjugation, V = [V1 , . . . , Vn ]T
is the vector of node voltages, I = [I1 , I2 , . . . , In ]T is the
vector of injected node currents, S = [S1 , S2 , . . . , Sn ]T is the
vector of injected apparent powers, and Y is the admittance
matrix of the graph.
3
𝐻
IEEE 300-Bus
Fig. 2: The attack model. An adversary attacks a zone by disconnecting some of its lines (red dashed lines) and disallowing the information
from within the zone to reach the control center. G is the power grid graph and H is a subgraph of G that represents the attacked zone.
The elements of the admittance matrix Y which depends
on the topology of the grid as well as the admittance values
of the lines, is defined as follows:
P
yii + i6=k yik , if k = i
Yik = −yik ,
if k ∈ N (i)
0,
if k ∈
/ N (i)
where N (i) denotes the direct neighbors of node i, yik is
the equivalent admittance of the lines from node i to k,
and yii is sum of the shunt admittances at node i. In this
paper, we assume that the shunt admittances are negligible,
and therefore, yii = 0 for all i ∈ N . The admittance matrix
can also be written in term of its real and imaginary parts
as Y = G + iB where G and B are real matrices. Using
this and the definition of the apparent power Si = Pi + iQi
in (1-2) results in the equations for the active power Pi and
the reactive power Qi at each node i as well.
B. Incidence Matrix
Under an arbitrary direction assignment to the edges of G,
the incidence matrix of G is denoted by D ∈ {−1, 0, 1}n×m
and defined as,
if lj is not incident to node i,
0
dij = 1
if lj is coming out of node i,
−1 if lj is going into node i.
For each line lj = (i, k), define ylj := yik . It can be verified
that Y = Ddiag([yl1 , yl2 , . . . , ylm ])DT . As we demonstrate
in Section IV, the incidence matrix is a very useful matrix for
detecting line failures in power grids.
C. Basic Graph Theoretical Terms
Matching: A matching in a graph is a set of pairwise
nonadjacent edges. If M is a matching, the two ends of each
edge of M are said to be matched under M, and each vertex
incident with an edge of M is said to be covered by M.
Cycle: A cycle in a graph is a sequence of its distinct nodes
u1 , u2 , . . . , uk such that for all i < k, {ui , ui+1 } ∈ L, and
also {uk , u1 } ∈ L. A graph with no cycle is called acyclic.
D. Attack Model
We assume that an adversary attacks an area by: (i) disconnecting some lines within that area (failed lines), and
(ii) obstructing all the information (e.g., status of the lines
and voltage measurements) from within the area to reach the
control center. We call this area, the attacked zone.
Fig. 2 shows an example of an attack on the area represented
by H = (NH , LH ). We denote the set of failed lines in the
attacked zone H by F ⊆ LH . Upon failure, the failed lines
are removed from the graph and the flows are redistributed
according to the AC power flows. Our objective is to estimate
the voltages and detect the failed lines inside the attacked zone
using the changes in the voltages outside of the zone.
We use the prime symbol (0 ) to denote the values after an
attack (e.g, Y0 denotes the admittance matrix of the grid and
V 0 denotes node voltages after the attack). Using this notation,
if H̄ denotes the set of nodes outside of the attacked zone, then
given VH̄0 and S 0 , we want to recover VH0 and F. Notice that it
0
is reasonable to assume that we know SH
after the attack, since
for the load nodes inside that attacked zone H, we can assume
that P 0 and Q0 remain similar to their values before the attack
and for the generators we can assume that generator operators
can safely report the generated P 0 and Q0 values. Equivalently,
4
we can assume that P and Q do not change much after the
attack and therefore S 0 = S.
Detecting line failures after such an attack is crucial in
maintaining the stability of the grid, since it may result in
further line overloads and failures, if the proper load shedding
mechanism is not applied. An effective load shedding requires
the exact knowledge of the topology of the grid.
Notation. For any complex number x = Re(x) + i Im(x),
real numbers Re(x) and Im(x) denote its real and imaginary
values, respectively. For a vector X, supp(X) denotes the
set of its nonzero entries. If H1 , H2 are two subgraphs of G,
YH1 |H2 denotes the submatrix of Y with rows from VH1 and
columns from VH2 . Moreover, YH1 denotes the submatrix of
Y with all the rows associated with VH1 . For instance, Y can
be written in any of the following forms,
YH|H YH|H̄
YH
Y=
,Y =
.
YH̄|H YH̄|H̄
YH̄
IV. S TATE E STIMATION
In this section, we provide the analytical building blocks
of the EXPOSE Algorithm which can be used to estimate the
state of the grid following a cyber-physical attack. Notice that
the state estimation problem considered here is different from
the classical state estimation problem in power grids. Here,
besides estimating the voltage magnitudes and phase angles in
the attacked area, the algorithm needs to estimate the topology
of the grid as well.
A. Voltage Recovery
Here, we provide a method to recover the voltages inside
that attacked zone after the attack.
Observation 1: The admittance matrix of the grid does not
0
).
change outside of the attacked zone (i.e., YH̄ = YH̄
Proof: Since the line failures only happen inside H, following the definition of the admittance matrix (see Section III),
after the attack only the entries of YH|H change. Hence, YH̄
remains unchanged.
From Observation 1 and using (1), we have:
0
0
0
∗
0∗
YH̄
V 0 = IH̄
⇒ YH̄ V 0 = IH̄
⇒ YH̄
V 0∗ = IH̄
∗
0∗
⇒diag(VH̄0 )YH̄
V 0∗ = diag(VH̄0 )IH̄
∗
0
⇒diag(VH̄0 )YH̄
V 0∗ = SH̄
∗
0∗
0
∗
0∗
⇒diag(VH̄0 )YH̄|
H̄ VH̄ + diag(VH̄ )YH̄|H VH
Lemma 1: If there is matching between the nodes in H̄ and
H that covers the nodes in H, then the following matrix has
full column rank almost surely,
GH̄|H −BH̄|H
M :=
.
BH̄|H GH̄|H
Proof: Suppose U ⊆ NH̄ are the matched nodes which
are in H̄. Since the matching covers H, thus |U| = |NH |. To
show that M has full column rank, we show that
GU |H −BU |H
6= 0,
det(MU |H ) := det
BU |H GU |H
almost surely. det(MU |H ) can be considered as a polynomial
in terms of the entries of MU |H using Leibniz formula. Now
assume U = {u1 , u2 , . . . , u|NH | } are matched to NH =
Q|N |
{v1 , v2 , . . . , v|NH | } in order. It can be seen that i=1H G2ui vi
Q|NH | 2
and i=1 Bui vi are two terms with nonzero coefficient in
det(MU |H ). Therefore, det(MU |H ) is a nonzero polynomial
in terms of its entries. Now since the set of roots of a nonzero
polynomial is a measure zero set in the real space, thus
det(MU |H ) 6= 0 almost surely.
Corollary 1: If there is matching between the nodes in H̄
and H that covers the nodes in H, then VH0 can be recovered
almost surely.
B. Line Failures Detection
Assume VH0 is successfully recovered using (4). In this
subsection, using VH0 , we provide a method to detect the set
of line failures F.
Lemma 2: There exists a complex vector X ∈ C|LH | such
that
0
YH V 0 = IH
+ DH X,
(5)
and supp(X) = F. Moreover, the vector X is unique if, and
only if, DH has full column rank.
Proof: Without loss of generality assume F =
{l1 , l2 , . . . , lk }. It can be seen that Y0 = Y −
Ddiag([yl1 , yl2 , . . . , ylk , 0, 0, . . . , 0])DT . Hence,
Y0 V 0 = YV 0 − Ddiag([yl1 , yl2 , . . . , ylk , 0, 0, . . . , 0])DT V 0
I 0 = YV 0 − Ddiag([yl1 , yl2 , . . . , ylk , 0, 0, . . . , 0])DT V 0 .
Now if we only focus on the rows associated with the nodes
in H, it can been seen that
0
= SH̄
.
(3)
Notice that in (3) all the variables are known after the attack
∗
0
except VH0∗ . Define EH̄ := −YH̄|
V 0∗ +diag(VH̄0−1 )SH̄
which
H̄ H̄
can be computed from the given variables after the attack.
Then, we can separate the real and imaginary parts of (3)
using block matrices as follows:
GH̄|H −BH̄|H Re(VH0 )
Re(EH̄ )
=
.
(4)
BH̄|H GH̄|H
Im(VH0 )
−Im(EH̄ )
One can see that Re(VH0 ) and Im(VH0 ) can be uniquely
recovered if the matrix on the left hand side of (4) has full
column rank. The following lemma provides the connection
between the rank of that matrix and the topology of the grid.
0
IH
= YH V 0 − DH diag([yl1 , yl2 , . . . , ylk , 0, 0, . . . , 0])DTH V 0 .
Hence, vector X := diag([yl1 , yl2 , . . . , ylk , 0, 0, . . . , 0])DTH V 0
satisfies (5). It can also be seen that only the entries of X that
are associated with the failed lines are nonzero and therefore
supp(X) = F. In order for (5) to have a unique solution, DH
should have full column rank.
Corollary 2: There exist a real vector X ∈ R|LH | such that
∗
0
Re{YH
V 0∗ } = Re{diag(VH0 )−1 SH
} + DH X,
(6)
and supp(X) = F. Moreover, the vector X is unique if, and
only if, DH has full column rank.
Proof: Using (2) and Lemma 2 gives the result.
5
Corollary 2 indicates that the set of line failures can be
detected by solving a matrix equation, if DH has full column
rank. The following lemma provides the connection between
the rank of that matrix and the topology of the attacked zone.
Lemma 3: The solution vector X to (6) is unique if, and
only if, H is acyclic.
Proof: It is easy to verify that (6) has a unique solution
X if, and only if, DH has linearly independent columns. On
the other hand, it is known in graph theory that rank(DH ) =
|NH |−c in which c is the number of connected components of
H [34, Theorem 2.3]. Therefore, DH has linearly independent
columns if and only if LH = |NH | − c which means that each
connected component of DH is acyclic.
Corollary 3: If H is acyclic, then the set of line failures F
can be detected by solving (6) for X.
Corollary 3 states that the set of line failure can accurately be
detected if H is acyclic. The importance of this result is in
demonstrating that the set of line failures can be efficiently
detected by solving a matrix equation, independent of the
number of line failures.
We can use a similar idea as in [3] to extend this approach
to when H is not acyclic. If we assume that the set of line
failures are sparse compare to the total number of lines in H,
we can detect line failures by finding the solution of following
optimization problem instead:
minimize kXk1 s.t.
X∈R|LH |
∗
Re{YH
V 0∗ }
kX − c1k1 =
=
0
= Re{diag(VH0 )−1 SH
} + DH X.
(7)
≥
i=1
minimize
0 ∈C|NH |
X∈R|LH | ,VH
kXk1 s.t.
GH̄|H Re(VH0 ) − BH̄|H Im(VH0 ) = Re(EH̄ )
BH̄|H Re(VH0 ) + GH̄|H Im(VH0 ) = −Im(EH̄ )
∗
0
Re{YH
V 0∗ } = Re{diag(VH0 )−1 SH
} + DH X.
(8)
However, since VH0 is part of the variables, this optimization
problem is not linear and convex anymore. To resolve this
issue, we need to approximate diag(VH0 )−1 with a linear
function in terms of VH0 . For this, we have:
On the other hand, the voltage magnitudes are almost constant
at each node before and after the failure (|VH0 | ≈ |VH |), hence:
diag(VH0 )−1 ≈ diag(|VH |)−2 diag(Re(VH0 ))−i diag(Im(VH0 )) .
We can use the approximation above in optimization (8) in
order to relax its nonconvexity. Notice that since optimization
(8) is for the cases in which the solution to (4) is not unique
and therefore the voltages cannot be recovered uniquely, some
conditions should be placed on the voltages such that the
recovered voltages are near operable conditions. To do so,
we add a convex constraint on the voltage magnitudes of the
nodes in H after the attack as |VH0 | ≤ 1.11H , in which 1H is
an all ones vector of size |NH |. Hence, the following convex
optimization can be used to detect the set of line failures and
recover the voltages when the solution to (4) is not unique:
minimize
0 ∈C|NH |
X∈R|LH | ,VH
kXk1 s.t.
GH̄|H Re(VH0 ) − BH̄|H Im(VH0 ) = Re(EH̄ )
BH̄|H Re(VH0 ) + GH̄|H Im(VH0 ) = −Im(EH̄ )
|VH0 | ≤ 1.11H
+ diag(|VH |)−2 diag(Im(VH0 ))Q0H + DH X.
(|xi − c| + |c|) + (|LH | − 2k)|c|
|xi | + (|LH | − 2k)|c| >
In order to extend our approach to the cases that (4) does
not have a unique solution, we can solve (4) and (7) at the
same time. Therefore, in order to recover the voltages and
detect the line failures at the same time, one needs to solve
the following optimization problem:
∗
0
Re{YH
V 0∗ } ≈ diag(|VH |)−2 diag(Re(VH0 ))PH
|xi − c| + (|LH | − k)|c|
i=1
k
X
i=1
k
X
C. Simultaneous Recovery and Detection
diag(VH0 )−1 = diag(|VH0 |)−2 diag(Re(VH0 ))−i diag(Im(VH0 )) .
Notice that optimization problem (7) can be solved efficiently
using Linear Programming (LP).
Lemma 4: If H is a cycle and less than half of its edges are
failed, then the solution X to the optimization problem (7) is
unique and supp(X) = F.
Proof: The idea of the proof is similar to the idea used
in the proof in [3, Lemma 4]. Here without loss of generality,
we assume that DH is the incidence matrix of H when edges
of the cycle are directed clockwise. Since H is connected, it
is known that rank(DH ) = |NH | − 1 [34, Theorem 2.2].
Therefore, dim(null(DH )) = 1. Suppose 1 ∈ R|LH | is
the all one vector. It can be seen that DH 1 = 0. Since
dim(null(DH )) = 1, 1 is the basis for the null space of D.
Suppose X is a solution to (7) such that supp(X) = F (from
Lemma 2 we know that such a solution exists). To prove that
X is the unique solution for (7), we prove that ∀c ∈ R\{0},
kXk1 < kX −c1k1 . Without loss of generality we can assume
that x1 , x2 , . . . , xk are the nonzero elements of X, in which
k = |F|. From the assumption, we know that k < |LH |/2.
Hence,
k
X
Thus, the solution to (7) is unique and supp(X) = F.
Lemma 4 can be extended to planar graphs similar to the
result for the DC power flows presented in [3, Theorem 2]. The
proof and the argument in [3] should apply with a very slight
change here as well. To avoid repetition, we do not present a
similar Lemma here.
k
X
i=1
|xi | = kXk1 .
(9)
Notice that to enforce AA ≈ BB for arbitrary vectors AA
and BB, we use convex constraint kAA − BBk2 < for a
small value . In Section VII, we evaluate the accuracy of the
results obtained by solving the convex optimization problem
(9) as part of the EXPOSE Algorithm.
6
IEEE 118-Bus
Algorithm 1: EXPress line failure detection using partially
ObSErved information (EXPOSE)
Input: A connected graph G, attacked zone H, V , VH̄0 , and S 0
1:
2:
3:
4:
5:
6:
7:
8:
if M is full rank then
Solve (4) to recover VH0
Use the recovered VH0 in (7) to find the solution vector X
else
Find the solution VH0 and X to the optimization (9)
Compute F = supp(X)
Compute the confidence of the solution cP and cQ
return VH0 , F, cP , and cQ
Algorithm 2: Brute Force Search (BFS)
1: for Any F ⊆ LH do
2:
Compute V † after removing the lines in F † from G
3:
Compute eF † = kRe(VH̄† − VH̄0 )k2 +kIm(VH̄† − VH̄0 )k2
4: Select F = arg minF † ⊆LH eF †
5: return F
D. Confidence of the Solution
Once the set of line failures is detected and the voltages
are recovered, one can compute the confidence of the solution
using (1-2). Assume Y† and V † denote the admittance matrix
of the grid after removing the detected lines and the recovered
voltages after the attack, respectively. If the detection and
recovery are done correctly, then Re{diag(V † )∗ YV † } = P 0
and Im{diag(V † )∗ YV † } = Q0 . However, if the detection and
recovery are not done correctly, this equalities do not hold. We
can use the difference between the two sides of these equalities
as a measure for the correctness of the solution.
We define cP and cQ to denote the confidence of the
solution based on P and Q as follows:
cP := (1−kRe{diag(V † )∗ YV † }−P 0 k2 /kP 0 k2 )+ ∗100, (10)
cQ := (1−kIm{diag(V † )∗ YV † }−Q0 k2 /kQ0 k2 )+ ∗100, (11)
in which (x)+ := max(0, x). If cP , cQ ≈ 100%, then it means
that the solution is reliable. If not, depending on the cP or cQ
values, one can see how reliable the solution actually is.
V. EXPOSE A LGORITHM
In this section, using the results provided in Section IV,
we introduce the EXPress line failure detection using partially
ObSErved information (EXPOSE) Algorithm. The EXPOSE
Algorithm is summarized in Algorithm 1.
Notice that for solving (4), (7), and (9) only the voltages of
the nodes that are at most one hop away from the nodes in H
are required. Hence, not only the running time of the EXPOSE
algorithm is independent of the number of line failures, it is
independent of the size of the entire grid as well. This makes
the EXPOSE Algorithm suitable for detecting line failures in
large networks.
Fig. 3: An attacked zone used in [4], shown by red nodes, in the
IEEE 118-bus system.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
1
2
3
4
5
Failed Line
†
Failed Line
Input: A connected graph G, attacked zone H, V , VH̄0 , and S 0
6
7
8
9
10
11
12
13
14
15
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
Detected Lines
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Detected Lines
(a)
(b)
Fig. 4: Detected line failures after all possible single line failures
using the EXPOSE Algorithm in (a) the zone shown in Fig. 3 in the
IEEE 118-bus system, and (b) the zone shown in Fig. 2 in the IEEE
300-bus system.
Algorithm for detecting line failures after the attack. The BFS
Algorithm considers all possible line failure scenarios and
returns the most likelihood solution. This method is the naive
version of the approaches used in [5], [6], [7], [8], [9] in
similar settings to detect line failures given a partial phase
angle measurements under the DC power flow model.
The idea is to compute the voltages VH̄† for any possible set
of line failures F † ⊆ LH and detect the one that is closest to
VH̄0 as the most likely failure as follows:
F = arg min kRe(VH̄† − VH̄0 )k2 + kIm(VH̄† − VH̄0 )k2 . (12)
F † ⊆LH
The BFS Algrithm is summarized in Algorithm 2. Notice
that the BFS Algorithm is exponentially slower than the
EXPOSE Algorithm, since it requires to solve the AC power
flow solutions 2|LH | times. Moreover, since it requires to solve
the power flow equations for the entire grid, in oppose to the
EXPOSE Algorithm, its running time increases polynomially
with the size of the grid.
The main shortcoming of the BFS Algorithm is its intractability for large networks. One way to speed up the BFS
Algorithm is to stop whenever the eF † (as defined in line 3
of the algorithm) is less than a threshold. This may speed up
the process but does not solve the intractability issue.
VI. B RUTE F ORCE A LGORITHM
In order to compare the performance of the EXPOSE
Algorithm with the previous works that were mostly under the
DC power flows, we introduce the Brute Force Search (BFS)
VII. N UMERICAL R ESULTS
While in Section IV, we analytically proved that the EXPOSE Algorithm guarantees to accurately recover the voltages
7
200
100
0
1
2
3
4
5
6
7
BFS
EXPOSE
Level 1
Level 2
60
40
Level 3
Level 4
Level 5
40
20
0
1
Number of Line Failures
30
Rank Deficit
300
80
BFS
COPSSES
EXPOSE
Running time (sec)
Running time (sec)
400
2
3
4
λH
γH
20
10
5
Attacked Zone Level
(a)
(b)
Fig. 5: Average running times of different line failures detection
algorithms versus (a) the total number of line failures in the attacked
zone shown in Fig. 2, and (b) the size of the attacked zone as shown
in Fig. 6(a) in detecting triple line failures.
and detect line failures under some conditions (i.e., matched
and acyclic attacked zones), in this section, we numerically
confirm those results and also evaluate the EXPOSE Algorithm’s performance when those conditions do not hold (i.e.,
general attacked zones).
A. Matched and acyclic attacked zones
As we mentioned in Section II, to the best of our knowledge,
our work in [4] is the only other method for information
recovery under the AC power flows that can be used to detect
any number of line failures and scales well with the size
of the grid. In [4], we introduced the Convex OPtimization
for Statistical State EStimation (COPSSES) Algorithm and
demonstrated that when the attacked zone is matched and
acyclic (i.e., matrices M and DH have full column rank), it
can detect line failures with few errors. The COPSSES Algorithm uses a relaxation of the methods introduced in [3], which
were based on DC power flow equations, for information
recovery under the AC power flow equations. The advantage
of the COPSSES Algorithm is that similar to the EXPOSE
Algorithm, its running time is independent of the number of
line failures.
In order to demonstrate the superiority of the EXPOSE
Algorithm in this case, here we compare its performance and
running time to the COPSSES Algorithm in addition to the BFS
Algorithm. For comparison purposes, we consider attacks on
the same zones as considered in [4] within the IEEE 118- and
300-bus systems. The zones are depicted in Figs. 3 and 2.
Recall from subsections IV-A and IV-B that when matrices
M and DH have full column rank, as it is the cases here, the
EXPOSE Algorithm can recover the voltages and detect the
line failures accurately. Hence as we expected and can be seen
in Fig. 4, all the single line failures can be exactly detected
using the EXPOSE Algorithm in the selected attacked zones
within the IEEE 118- and 300-bus systems. Notice that the
false positives in failures of lines 6 and 7 as well as 17 and
18 in the IEEE 118-bus system are due to the violation of the
acyclicity of the attacked zone. Lines 6 and 7 (and also 17
and 18) are parallel lines that form a cycle with two nodes.
Moreover, lack of any detections after failures in lines 1
and 15 within that attacked zone in the 300-bus system is due
to the fact that the AC power flows did not have a solution
after those failures. Therefore those cases did not considered
in evaluation of the EXPOSE Algorithm.
0
1
2
3
4
5
Attacked Zone Level
(b)
(a)
Fig. 6: Nested attacked zones in the IEEE 300-bus system. (a) Nodes
corresponding to different levels are shown in different colors, and
(b) the rank deficit in the attacked zones in different levels.
We considered up to 7 line failures in the zone depicted
in Fig. 2. In all the cases, as we expected, the EXPOSE
Algorithm could exactly detect the line failures. The BFS
Algorithm could also detect the line failures exactly in those
scenarios. However, as it was shown in [4], the COPSSES
Algorithm may result in few false positives and negatives in
detecting single line failures, and more false positives and
negatives as the number of line failures increases.
Fig. 5(a) compares the running times of the three Algorithms in detecting line failures versus the number of line
failures. As can be seen since the running times of the
EXPOSE and COPSSES Algorithms are independent of the
number of line failures, they both provide a constant running
time as the number of line failures increases. However, as
can be seen, the running time of the BFS Algorithm increases
exponentially as the total number of line failures increases.
Overall, when the attacked zone is matched and acyclic, the
EXPOSE Algorithm detects line failures as accurately as the
BFS Algorithm, but exponentially faster.
B. General attacked zones
In order to evaluate the performance of the EXPOSE
Algorithm as the attacked zone becomes larger and topologically more complex, in this subsection, we consider 5 nested
attacked zones as depicted in Fig. 6(a). We denote the nodes
that are added to the attacked zone at ith step by the level
i nodes. The level i attacked zone is an attacked zone that
consists of all the nodes in levels 1 to i.
As we proved in Section IV and briefly showed in Subsection VII-A, when matrices M and DH have full column
rank, then the EXPOSE Algorithm can recover the voltages
and detects the line failures accurately. In order to show how
far or close the topological properties of an attacked zone are
to these conditions, we define λH and γH as follows:
λH := 2nH − rank(M),
γH := mH − rank(DH ).
It can be verified that when matrices M and DH have full
column rank, then λH = 0 and γH = 0, respectively. Hence,
λH and γH indicate the rank deficit of matrices M and DH .
Fig. 6(b) shows the λH and γH values for the different
attacked zone levels. As can be seen, both values grow
8
1.5
3
Triple Line Failures
Double Line Failures
Single Line Failures
Number of False Positives
Number of False Negatives
2
1
0.5
0
1
2
3
4
2.5
2
Triple Line Failures
Double Line Failures
Single Line Failures
1.5
1
0.5
0
1
5
Attacked Zone Level
2
3
(a)
30
Triple Line Failures
Double Line Failures
Single Line Failures
20
10
0
1
2
5
(b)
50
Percent Error Phase Ang.
Percent Error Vol. Mag.
40
4
Attacked Zone Level
3
4
Triple Line Failures
Double Line Failures
Single Line Failures
40
30
20
10
0
1
5
Attacked Zone Level
2
3
4
5
Attacked Zone Level
(c)
(d)
100
90
80
80
60
cP
cQ
100
70
60
50
1
40
Triple Line Failures
Double Line Failures
Single Line Failures
2
3
20
4
Attacked Zone Level
(e)
5
0
1
Triple Line Failures
Double Line Failures
Single Line Failures
2
3
4
5
Attacked Zone Level
(f)
Fig. 7: The EXPOSE Algorithm’s performance after all single,
double, and triple line failures versus the size of the attacked zone as
shown in Fig. 6(a). (a) Average number of false negatives, (b) average
number of false positives, (c) average percentage error in recovered
voltage magnitudes, (d) average percentage error in recovered voltage
phase angles, (e) average confidence of the solutions (cP ), and (f)
average confidence of the solutions (cQ ).
significantly in level 4 and 5 attacked zones. This means
that the data outside of the attacked area is very insufficient
to accurately detect the line failures based on the EXPOSE
Algorithm in those levels.
First, in order to show the advantage of the EXPOSE
Algorithm over brute force type algorithms, in Fig. 5(b), we
compare the increase in the running times of the BFS and
the EXPOSE Algorithms in detecting triple line failures as
the number of nodes and lines increases in different levels.
As can be seen in Fig. 5(b), the running time of the BSF
Algorithm exponentially increases with the size of the attacked
zone whereas that of the EXPOSE Algorithm only slightly
increases. This along with Fig. 5(a) clearly indicates that the
BFS Algorithm (and algorithms with similar approaches) do
not scale well with the size of the attacked zone and the
number of line failures.
In order to evaluate the performance of the EXPOSE Algorithm, we consider all single, double, and triple line failures
in the nested zones. The results are presented in Fig. 7.
The average number of false negatives and positives in
detecting line failures in different attacked zone levels for
all single, double, and triple line failures are presented in
Figs. 7(a) and 7(b). It can be seen that as we expected for the
level 1 attacked zone, there are no false negatives or positive.
For the level 2 attacked zone also, although DH does not
have full column rank (see Fig. 6(b)), the EXPOSE Algorithm
can still detect the line failures accurately. However, as the
attacked zone becomes larger in higher levels, and λH and
γH increase, we observe that the EXPOSE Algorithm results
in false positives and negatives. An important observation here
is that the EXPOSE Algorithm results on average in more false
positives than negatives. This is a good characteristic of the
EXPOSE Algorithm, since it means that by having an extra
brute force search step on the detected line failures set, one
can reduce number of false positives significantly.
The average error in recovered voltages in different attacked
zone levels using the EXPOSE Algorithm for all single,
double, and triple line failures are presented in Figs. 7(c) and
7(d). As can be seen, similar to the line failures detection,
the EXPOSE Algorithm recovers the voltages accurately for
the level 1 and 2 attacked zones. Moreover, for the level 3
and 4 attacked zones, the EXPOSE Algorithm recovers the
voltage magnitudes and phase angles with less than 15% and
10% error, respectively. However, for the level 5 attacked zone,
since λH is too high (see Fig. 6(b)), the EXPOSE Algorithm
results in around 30% and 40% error in the recovered voltage
magnitudes and phase angles, respectively.
Finally, as we introduced cP and cQ in subsection IV-D,
these metrics can be used to determine the confidence of
the solutions obtained by the EXPOSE Algorithm. As can
be seen in Figs. 7(e) and 7(f), the cP and cQ values are
directly correlated with the errors in voltages and number of
false negatives. Hence, these values can effectively be used
to compute the confidence of the solution obtained by the
EXPOSE Algorithm. Notice that cQ is more sensitive than
the cP . Therefore, cQ can be used as the upper bound for the
error and cP can be used as the lower bound.
We did not evaluate the performance of the BFS Algorithm
here due to its very high running time (see Fig. 5(b)). However,
we expect that the BFS Algorithm could detect the line failures
and recover the voltages with almost no error. Despite its
accuracy, the BFS Algorithm is not practical for line failures
detection in large networks. As we showed in this section,
the EXPOSE Algorithm can provide relatively accurate results
exponentially faster than the BFS Algorithm.
VIII. C ONCLUSION
We studied cyber-physical attacks on power grids under the
AC power flows. We leveraged the algebraic properties of the
AC power flows to develop the EXPOSE Algorithm for detecting line failures and recovering the voltages after the attack.
We analytically proved that if the attacked zone has certain
topological properties, the EXPOSE Algorithm can accurately
recover the information. We also numerically demonstrated
that in more complex attacked zones, it can still recover the
information approximately well. The main advantages of the
EXPOSE Algorithm are that its running time is independent
9
of the size of the grid and number of line failures, and that it
provides accurate information recovery under some conditions
on the attacked zone. Moreover, it approximately recovers the
information and provides the confidence of the solution when
these conditions do not hold.
The results provided in this paper can be further used in
different context as well. For example, the EXPOSE Algorithm
can be used to detect line failures when measurement devices
are scarce and not ubiquitous. Moreover, the conditions on the
attacked zone such that the EXPOSE Algorithm can accurately
detect the line failures and recover the voltages, can be used
for optimal measurement device placements in the grid.
Despite it strengths, the EXPOSE Algorithm presented in
this paper requires that the power system converges to a stable
state after an attack. However, as the number of line failures
increases, such an assumption may rarely holds. Therefore,
the dynamics of the system after an attack should also be
considered for an effective detection mechanism. Due to their
complexity, study the dynamics of the power system after an
attack is a very challenging task. Hence, exploring this and
other directions is part of our future work.
ACKNOWLEDGEMENT
This work was supported in part by DARPA RADICS under
contract #FA-8750-16-C-0054, funding from the U.S. DOE
OE as part of the DOE Grid Modernization Initiative, and
DTRA grant HDTRA1-13-1-0021.
R EFERENCES
[1] “Analysis of the cyber attack on the Ukrainian power grid,” Mar.
2016, http://www.nerc.com/pa/CI/ESISAC/Documents/E-ISAC SANS
Ukraine DUC 18Mar2016.pdf.
[2] “Assault on California power station raises alarm on potential for
terrorism,” 2014, source: http://goo.gl/RiuhI1.
[3] S. Soltan, M. Yannakakis, and G. Zussman, “Power grid state estimation
following a joint cyber and physical attack,” to appear in IEEE Trans.
Control Netw. Syst., 2017.
[4] S. Soltan and G. Zussman, “Power grid state estimation after a cyberphysical attack under the AC power flow model,” in Proc. IEEE PESGM’17, July 2017.
[5] J. E. Tate and T. J. Overbye, “Line outage detection using phasor angle
measurements,” IEEE Trans. Power Syst., vol. 23, no. 4, pp. 1644–1652,
2008.
[6] ——, “Double line outage detection using phasor angle measurements,”
in Proc. IEEE PES-GM’09, July 2009.
[7] H. Zhu and G. B. Giannakis, “Sparse overcomplete representations for
efficient identification of power line outages,” IEEE Trans. Power Syst.,
vol. 27, no. 4, pp. 2215–2224, 2012.
[8] Y. Zhao, A. Goldsmith, and H. V. Poor, “On PMU location selection
for line outage detection in wide-area transmission networks,” in Proc.
IEEE PES-GM’12, July 2012.
[9] H. Zhu and Y. Shi, “Phasor measurement unit placement for identifying
power line outages in wide-area transmission system monitoring,” in
HICSS’14, 2014, pp. 2483–2492.
[10] A. Ashok, M. Govindarasu, and J. Wang, “Cyber-physical attackresilient wide-area monitoring, protection, and control for the power
grid,” Proc. IEEE, 2017.
[11] A. Pinar, J. Meza, V. Donde, and B. Lesieutre, “Optimization strategies
for the vulnerability analysis of the electric power grid,” SIAM J.
Optimiz., vol. 20, no. 4, pp. 1786–1810, 2010.
[12] T. Kim, S. J. Wright, D. Bienstock, and S. Harnett, “Analyzing vulnerability of power systems with continuous optimization formulations,”
IEEE Trans. Net. Sci. Eng., vol. 3, no. 3, pp. 132–146, 2016.
[13] J. Liu, C. H. Xia, N. B. Shroff, and H. D. Sherali, “Distributed
optimal load shedding for disaster recovery in smart electric power
grids: A second-order approach,” in Proc. ACM SIGMETRICS’14 (poster
description), June 2014.
[14] A. Bernstein, D. Bienstock, D. Hay, M. Uzunoglu, and G. Zussman,
“Sensitivity analysis of the power grid vulnerability to large-scale
cascading failures,” ACM SIGMETRICS Perform. Eval. Rev., vol. 40,
no. 3, pp. 33–37, 2012.
[15] I. Dobson, B. Carreras, V. Lynch, and D. Newman, “Complex systems
analysis of series of blackouts: cascading failure, critical points, and
self-organization,” Chaos, vol. 17, no. 2, p. 026103, 2007.
[16] D. Bienstock, Electrical Transmission System Cascades and Vulnerability: An Operations Research Viewpoint. SIAM, 2016.
[17] S. Soltan, A. Loh, and G. Zussman, “Analyzing and quantifying the
effect of k-line failures in power grids,” to appear in IEEE Trans.
Control Netw. Syst., 2017.
[18] A. Srivastava, T. Morris, T. Ernster, C. Vellaithurai, S. Pan, and U. Adhikari, “Modeling cyber-physical vulnerability of the smart grid with
incomplete information,” IEEE Trans. Smart Grid, vol. 4, no. 1, pp.
235–244, 2013.
[19] F. Pasqualetti, F. Dörfler, and F. Bullo, “Cyber-physical attacks in power
networks: Models, fundamental limitations and monitor design,” in Proc.
IEEE CDC-ECC’11, 2011, pp. 2195–2201.
[20] J. Kim and L. Tong, “On topology attack of a smart grid: undetectable
attacks and countermeasures,” IEEE J. Sel. Areas Commun, vol. 31,
no. 7, pp. 1294–1305, 2013.
[21] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against
state estimation in electric power grids,” ACM Trans. Inf. Syst. Secur.,
vol. 14, no. 1, p. 13, 2011.
[22] G. Dán and H. Sandberg, “Stealth attacks and protection schemes for
state estimators in power systems,” in Proc. IEEE SmartGridComm’10,
2010.
[23] O. Vukovic, K. C. Sou, G. Dán, and H. Sandberg, “Network-layer
protection schemes against stealth attacks on state estimators in power
systems,” in Proc. IEEE SmartGridComm’11, 2011.
[24] S. Li, Y. Yılmaz, and X. Wang, “Quickest detection of false data
injection attack in wide-area smart grids,” IEEE Trans. Smart Grid,
vol. 6, no. 6, pp. 2725–2735, 2015.
[25] J. Kim, L. Tong, and R. J. Thomas, “Subspace methods for data attack on
state estimation: A data driven approach.” IEEE Trans. Signal Process.,
vol. 63, no. 5, pp. 1102–1114, 2015.
[26] M. Garcia, T. Catanach, S. Vander Wiel, R. Bent, and E. Lawrence,
“Line outage localization using phasor measurement data in transient
state,” IEEE Trans. Power Syst., vol. 31, no. 4, pp. 3019–3027, 2016.
[27] N. M. Manousakis, G. N. Korres, and P. S. Georgilakis, “Taxonomy
of PMU placement methodologies,” IEEE Trans. Power Syst., vol. 27,
no. 2, pp. 1070–1077, 2012.
[28] K. Khandeparkar, P. Patre, S. Jain, K. Ramamritham, and R. Gupta,
“Efficient PMU data dissemination in smart grid,” in Proc. ACM eEnergy’14 (poster description), June 2014.
[29] Y. Zhao, J. Chen, A. Goldsmith, and H. V. Poor, “Identification of
outages in power systems with uncertain states and optimal sensor
locations,” IEEE J. Sel. Topics Signal Process., vol. 8, no. 6, pp. 1140–
1153, 2014.
[30] Z. Li, M. Shahidehpour, A. Alabdulwahab, and A. Abusorrah, “Bilevel
model for analyzing coordinated cyber-physical attacks on power systems,” IEEE Trans. Smart Grid, vol. 7, no. 5, pp. 2260–2272, 2016.
[31] R. Deng, P. Zhuang, and H. Liang, “Ccpa: Coordinated cyber-physical
attacks and countermeasures in smart grid,” IEEE Trans. Smart Grid,
2017.
[32] J. Zhang and L. Sankar, “Physical system consequences of unobservable
state-and-topology cyber-physical attacks,” IEEE Trans. Smart Grid,
vol. 7, no. 4, 2016.
[33] J. D. Glover, M. S. Sarma, and T. Overbye, Power System Analysis &
Design, SI Version. Cengage Learning, 2012.
[34] R. Bapat, Graphs and matrices. Springer, 2010.
| 3cs.SY
|
Modeling Temporally Evolving and Spatially Globally
Dependent Data
Emilio Porcu,1 Alfredo Alegria
2
and Reinhard Furrer
3
June 29, 2017
Abstract
arXiv:1706.09233v1 [math.ST] 28 Jun 2017
June 29, 2017
The last decades have seen an unprecedented increase in the availability of data sets that are
inherently global and temporally evolving, from remotely sensed networks to climate model ensembles. This paper provides a view of statistical modeling techniques for space-time processes,
where space is the sphere representing our planet. In particular, we make a distintion between
(a) second order-based, and (b) practical approaches to model temporally evolving global processes. The former are based on the specification of a class of space-time covariance functions,
with space being the two-dimensional sphere. The latter are based on explicit description of
the dynamics of the space-time process, i.e., by specifying its evolution as a function of its past
history with added spatially dependent noise.
We especially focus on approach (a), where the literature has been sparse. We provide new
models of space-time covariance functions for random fields defined on spheres cross time. Practical approaches, (b), are also discussed, with special emphasis on models built directly on the
sphere, without projecting the spherical coordinate on the plane.
We present a case study focused on the analysis of air pollution from the 2015 wildfires in Equatorial Asia, an event which was classified as the year’s worst environmental disaster. The paper
finishes with a list of the main theoretical and applied research problems in the area, where we
expect the statistical community to engage over the next decade.
Keywords: Covariance functions; Great Circle; Massive Data Sets; Spheres.
1
School of Mathematics and Statistics, University of Newcastle, GB.
& Department of Mathematics, University Federico Santa Maria, 2360102 Valparaiso, Chile.
E-mail: [email protected].
Research work of Emilio Porcu was partially supported by grant FONDECYT 1130647 from the Chilean government.
2
Department of Mathematics, University Federico Santa Maria, 2360102 Valparaiso, Chile.
E-mail: [email protected].
3
Department of Mathematics and Department of Computational Science, University of Zurich, 8307 Zurich,
Switzerland.
E-mail: [email protected].
1
1
Introduction
The strong evidence of a changing climate over the last century (IPCC, 2013) has prompted the
scientific community to seek for comprehensive monitoring strategies of the state of the climate
system over the entire planet. The surge of satellite observations, the increase of computational
and storage availability as well as an increase in the horizontal resolution of global climate models,
the deployment of new global and automated network (the ARGO floats and the AERONET, see
Holben et al., 1998) and the recent development of smartphone-based data which potentially allow
every user on the planet to provide scientific data (Citizen Science) has generated an increase in
the size of the data of orders of magnitude. Such an increase in the volume, variate and velocity of
globally indexed data serves as a strong catalyst for the statistical community to develop models
that are inherently global and time evolving.
The research questions underpinning global space-time models span from optimal interpolation
(kriging) for global coverage over both space and time (for variables such as temperature and precipitation, but, also more recently, ozone, carbon dioxide and aerosol optical depth), to interpolation
in the input space of Earth System Models by generating fast approximations (emulators) that can
be used to perform fast and affordable sensitivity analysis. While the aforementioned topics are
of high scientific interest, the development of appropriate statistical methodology for global and
temporal data has been limited, and has advanced in two seemingly very different directions.
The construction of models on the sphere cross the temporal horizon calls for a rigorous development of a theory that would allow for valid processes with the proper distance over the curved
surface of the planet. Under the assumption of Gaussianity, the second order structure can be explicitly specified and the properties of the process can be studied directly from its functional form.
The theory for this approach has been actively developed over the last decade, but thus far has
been limited to the large-scale structure for the covariance function such as isotropy or stationarity
across longitudes. We denote this approach the second order-based approach.
Alternative definitions of the space-time process rely on either on the de-coupling of the spatial
and temporal part through the specification of the dynamics in time, or its representation as a
solution of a stochastic partial differential equation (SPDE). These techniques are designed primarily for inferential purposes, and they are particularly suitable for modern massive data sets. This
increased suitability for inference, however, comes at at the expense of convenient expressions that
allow for an understanding the underlying theoretical properties of the process. We refer to this as
the practical approach.
We start by describing the second order-based approaches, which rely on the first and second
order moment of the underlying random field on the sphere cross time. In particular, the focus
becomes the space-time covariance, where space is the two-dimensional sphere. For stochastic
processes on a sphere, the reader is referred to Jones (1963), Marinucci and Peccati (2011) and
the thorough review in Gneiting (2013). For space-time stochastic processes on the sphere, we
2
refer the reader to the more recent approaches in Porcu et al. (2016b); Berg and Porcu (2017) and
Jeong and Jun (2015). Generalizations to multivariate space-time processes have been considered
in Alegria et al. (2017). The increasing interest in modeling stochastic processes over spheres or
spheres cross time with an explicit covariance function is also reflected in work in areas as diverse
as mathematical analysis (Schoenberg, 1942; Gangolli, 1967; Hannan, 1970; Menegatto, 1994, 1995;
Chen et al., 2003; Menegatto et al., 2006; Beatson et al., 2014; Guella et al., 2017, 2016a; Barbosa
and Menegatto, 2017; Guella et al., 2016b), probability theory (Baldi and Marinucci, 2006; Lang and
Schwab, 2013; Hansen et al., 2015; Clarke et al., 2016), spatial point processes (Møller et al., 2017),
spatial geostatistics (Christakos and Papanicolaou, 2000; Gneiting, 2002; Hitczenko and Stein, 2012;
Huang et al., 2012; Gerber et al., 2017), space-time geostatistics (Christakos, 1991, 2000; Christakos
et al., 2000b; Porcu et al., 2016b; Berg and Porcu, 2017) and mathematical physics (Istas, 2005;
Leonenko and Sakhno, 2012; Malyarenko, 2013).
The natural metric to be used on the sphere is the geodesic or great circle distance (details are
explained in subsequent sections). However, if this metric is used in space-time covariance models
defined on Euclidean spaces cross time, these are generally not valid on the sphere cross time. The
resulting need for new theory of space-time covariance functions has motivated a rich literature
based on positive definite functions based on great circle distance.
Techniques based on covariance functions are certainly accurate both in terms of likelihood
inference and kriging predictions; yet they imply a high computational cost when dealing with
large datasets over the space-time domain. The main computational hurdle is the calculation of
the determinant and of the quadratic form based on the inverse of the covariance matrix. The
so called big N problem in this case requires a compromise between statistical and computational
efficiencies and the reader is referred to, e.g., Bevilacqua et al. (2012, 2010); Stein (2005b) and Furrer
et al. (2006), for statistical approaches based on the covariance functions, that aim to mitigate such
a computational burden. A notable approach for the problem of prediction for very large data sets
can be found in Cressie and Johannesson (2008).
As for practical approaches, when analyzing modern, massive data sets arising from remotely
sensed data, climate models or reanalysis data products, a model that fully specifies the covariance
function would require a prohibitive amount of information to be stored in the covariance matrix,
as well as a prohibitive number of flops and iterations for maximizing the likelihood or exploring
the Markov chain when performing Bayesian inference.
In the context of data in Euclidean spaces, one of the most popular and natural alternatives is
to explicitly describe the dynamics of the process by specifying the evolution as a function of its
past history with added spatially dependent noise. This approach has received strong support from
reference textbooks in spatio-temporal modelling (see Cressie and Wikle, 2011) and recent studies
have extended this methodology to the context of spatio-temporal data. Richardson et al. (2017,
2016) and Tebaldi and Sansó (2009), amongst others, recommend the use of an explicit description
of the dynamics of the process by specifying its evolution as a function of the spatial distribution
3
of the process. Dynamic spatio-temporal models have a long history in Euclidean spaces (Cressie
and Wikle, 2011, with the references therein), but the literature on the sphere is more sparse, with
the exception of Castruccio and Stein (2013) and related work. This paper will review the recent
literature of this approach, with a particular emphasis on scalable methods for large datasets.
Dynamics on large regions of Earths surface have been studied in Kang et al. (2010), who consider
aerosol data from multiangle imaging spectro radiometers along the Americas, the Atlantic Ocean,
and the western part of Europe and North Africa. Other relevant applications are proposed in
Oleson et al. (2013); Nguyen et al. (2014); Banerjee et al. (2008).
In recent years, a powerful modeling approach has emerged based on the consideration of a
space-time global process as a solution to a SPDE defined over the sphere and in time. Earlier work
proposed in Jones and Zhang (1997) was based on a diffusion-injection equation, which is just one
of a multitude of SPDE-based models commonly used to describe physical processes (Christakos,
2000). Later studies have focused on specifying an SPDE over the sphere or, more generally, on
Riemannian geometries, and more recently under the Integrated Nested Laplace Approximation
(INLA) environment (Rue and Tjelmeland, 2002; Lindgren et al., 2011; Lindgren and Rue, 2015;
Cameletti et al., 2012). The key intuition is to use the SPDE approach as a link to approximate a
continuous stochastic process with a Gaussian Markov random field, being a discretized version of
the first. This has remarkable computational advantages (see Lindgren et al., 2011) and also allows
to build flexible models by providing flexible functional expression of the differential operator. Some
notable approaches to model global data under this framework can be found in Bolin and Lindgren
(2011). Although there has been substantial effort to introduce the Markov field architecture
coupled with the INLA routine for spatial data, to our knowledge, only few studies have focused
on the sphere cross time, which will be surveyed below.
The plan of the paper is the following. Section 2 provides the necessary background material
on random fields over spheres or spheres cross time, with their covariance functions. Section
3 details the second order-based approach, with construction principles for characterizing spacetime covariance functions on the sphere cross time domain. New covariance functions are also
introduced with formal proofs given in the Appendix. Section 4 is devoted to practical approaches,
with emphasis on dynamical methods as well as methods based on SPDE and Gaussian Markov
random fields. Section 5 presents a case study focused on the analysis of air pollution from the
2015 wildfires in Equatorial Asia, an event which was classified as the year’s worst environmental
disaster. A massive global space-time data of air quality from NASA’s MERRA2 reanalysis with
more than 12 million data points is provided, and both approaches from (a) and (b) are compared
and their relative merits are discussed. Section 6 completes the paper with discussion, and with
a list of research problems in the area, where we expect the statistical community to engage over
the next decade. Technical proofs are deferred to the Appendix, where we also give some necessary
background material, as well as a list of other new space-time covariance functions that can be used
on spheres cross time.
4
2
Background
2.1
Spatial Fields on Spheres, Coordinates, and Metrics
We consider the unit sphere S2 , defined as S2 = {s ∈ R3 , ∥s∥ = 1}, where ∥ ⋅ ∥ denotes Euclidean
distance. Every point s on the sphere S2 has spherical coordinates s = (φ, ϑ), with φ ∈ [0, π]
and ϑ ∈ [0, 2π) being respectively the polar and the azimuthal angles (equivalently, latitude and
longitude). The extension to the sphere with arbitrary radius is straightforward. For planet Earth,
the radius is approximately 6, 371 km, albeit the Earth is not exactly a sphere.
The natural distance on the sphere is the geodesic or great circle distance, defined as the
mapping dGC ∶ S2 × S2 → [0, π] so that
dGC (s1 , s2 ) = arccos (⟨s1 , s2 ⟩) = arccos ( sin φ1 sin φ2 + cos φ1 cos φ2 cos ϑ),
with si = (φi , ϑi ), i = 1, 2, and ⟨⋅, ⋅⟩ denoting the classical dot product on the sphere, and where
ϑ = ∣ϑ1 − ϑ2 ∣. Thus, the geodesic distance describes an arc between any pair of points located on
the spherical shell. Throughout, we shall equivalently use dGC (s1 , s2 ) or its shortcut dGC to denote
the geodesic distance, whenever no confusion can arise.
An approximation of the true distance between any two points on the sphere is the chordal
distance dCH (s1 , s2 ), given by
dCH (s1 , s2 ) = 2 sin (
dGC (s1 , s2 )
),
2
s1 , s2 ∈ S2 ,
which defines a segment below the arc joining two points on the spherical shell.
We consider zero mean Gaussian fields {X(s), s ∈ S2 } with finite second order moment. Thus,
the finite dimensional distributions are completely specified by the covariance function CS ∶ S2 ×S2 →
R, defined by
CS (s1 , s2 ) = Cov(X(s1 ), X(s2 )),
s1 , s2 ∈ S2 .
2
Covariance functions are positive definite: for any N dimensional collection of points {si }N
i=1 ⊂ S
and constants c1 , . . . , cN ∈ R, we have
N N
∑ ∑ ci CS (si , sj ) cj ≥ 0,
(2.1)
i=1 j=1
see Bingham (1973). We also define the variogram γS of Z on S2 as half the variance of the
increments of X at the given points on the sphere. Namely,
2γS (s1 , s2 ) = Var(X(s2 ) − X(s1 )),
s1 , s2 ∈ S2 .
For a discussion about variograms on spheres, the reader is referred to Huang et al. (2012) and
Gneiting (2013), and the references therein.
The simplest process in the Euclidean framework is the isotropic process, i.e., a process that
does not depend on a particular direction, but just on the distance between points. We say that
CS is geodesically isotropic if CS (s1 , s2 ) = ψS (dGC (s1 , s2 )), for some ψS ∶ [0, π] → R. ψS is called
5
the geodesically isotropic part of CS (Daley and Porcu, 2013). Henceforth, we shall refer to both
CS and ψS as covariance functions, in order to simplify exposition. For a characterization of
geodesic isotropy, the reader is referred to Schoenberg (1942) and the essay in Gneiting (2013).
The definition of a geodesically isotropic variogram is analogous.
While a geodesically isotropic process on S2 is the natural counterpart to an isotropic process in
Euclidean space, it is not necessarily an appropriate process for globally-referenced data. Although
it may be argued that on a sufficiently small scale atmospheric phenomena lack any structured
flow, this does not apply in general for synoptic or mesoscale processes such as prevailing winds,
which follow regular patterns dictated by atmospheric circulation.
As a first approximation for large scale atmospheric phenomena, a process may be assumed to
be nonstationary for different latitudes, but stationary for the same latitude. Indeed, is it expected
that physical quantities such as temperature at surface will display an interannual variability that
is lower in the tropics than at mid-latitude. Therefore, we define the covariance CS to be axially
symmetric if
CS (s1 s2 ) = CS (φ1 , φ2 , ϑ1 − ϑ2 ),
(φi , ϑi ) ∈ [0, π] × [0, 2π), i = 1, 2.
(2.2)
Additionally, an axially symmetric Gaussian field X(s) is called longitudinally reversible if
CS (φ1 , φ2 , ϑ) = CS (φ1 , φ2 , −ϑ),
φi ∈ [0, π], ϑ ∈ [−2π, 2π), i = 1, 2.
(2.3)
An alternative notion of isotropy can be introduced if we assume that the sphere S2 is embedded
in the three dimensional Euclidean space R3 , and that the Gaussian process X is defined on R3 .
The covariance CS can then be defined by restricting X on S2 , which gives rise to the name of
Euclidean isotropy or radial symmetry, since it depends exclusively on the chordal distance (dCH )
between the points. Following Yadrenko (1983) and Yaglom (1987), for any CS being isotropic in
the Euclidean sense in R3 , the function CS (dCH ) is a valid covariance function on S2 . This principle
has been used to create space-time and multivariate covariance functions, and we come back into
this with details in Section 3.5.
2.2
Space-Time Random Fields and Covariance Functions
We now describe zero mean Gaussian fields {Z(s, t), s ∈ S2 , t ∈ R}, evolving temporally over the
sphere S2 . Henceforth, we assume that Z has finite second-order moment. Given the Gaussianity
2
assumption, we focus on the covariance function C ∶ (S2 × R) → R, defined as
C((s1 , t1 ), (s2 , t2 )) = Cov(Z(s1 , t1 ), Z(s2 , t2 )),
(si , ti ) ∈ S2 × R, i = 1, 2.
(2.4)
The definition of positive definiteness follows from Equation (2.1).
The covariance C is called separable (Gneiting et al., 2007), if there exists two mappings CS ∶
(S2 )2 → R and CT ∶ R2 × R such that
C((s1 , t1 ), (s2 , t2 )) = CS (s1 , s2 )CT (t1 , t2 ),
6
(si , ti ) ∈ S2 × R,
i = 1, 2.
(2.5)
Separability can be desirable from a computational standpoint: for a given set of colocated observations (that is, when for every instant t the same spatial sites have available observations), the
related covariance matrix factorizes into the Kronecker product of two smaller covariance matrices.
However, it has been deemed physically unrealistic, as the degree of spatial correlation is the same
at points near or far from the origin in time (Gneiting et al., 2007); a constructive criticism is
offered in Stein (2005a). There is a rich literature on non-separable covariance functions defined on
Euclidean spaces; see the reviews in Gneiting et al. (2007); Mateu et al. (2008) and Banerjee et al.
(2014), with the references therein.
2.3
Temporally Evolving Geodesic Isotropies and Axial Symmetries
A common assumption for space-time covariance functions over spheres cross time is that of geodesic
isotropy coupled with temporal symmetry: there exists a mapping ψ ∶ [0, π] × R → R such that
C((s1 , t1 ), (s2 , t2 )) = ψ(dGC (s1 , s2 ), t1 − t2 ),
(si , ti ) ∈ S2 × R, i = 1, 2.
(2.6)
Following Berg and Porcu (2017), we call P(S2 , R) the class of continuous functions ψ such that
ψ(0, 0) = σ 2 < ∞ and the identity (2.6) holds. The functions ψ are called the geodesically isotropic
and temporally symmetric parts of the space-time covariance functions C. Also, we refer equivalently to C or ψ as covariance functions, in order to simplify the exposition. In Appendix A we
introduce the more general class P(Sn , R) and show many relevant facts about it.
Equation (2.6) can be generalized to the case of temporal nonstationarity and the reader is
referred to Estrade et al. (2016) for a mathematical approach to this problem.
As in the purely spatial case, isotropic models are seldom used in practical applications as they
are deemed overly simplistic. Yet, they can served as building blocks for more sophisticated models,
that can account for local anisotropies and nonstationarities.
We now couple spatial axial symmety with temporal stationarity, so that, for the covariance C
in (2.4), there exists a continuous mapping C ∶ [0, π]2 × [−2π, 2π) × R such that
C((s1 , t1 ), (s2 , t2 )) = C(φ1 , φ2 , ϑ1 − ϑ2 , t1 − t2 ),
(φi , ϑi , ti ) ∈ [0, π] × [0, 2π) × R, i = 1, 2.
Additionally, we call a temporally stationary–spatially axially symmetric random field Z(s, t),
longitudinally reversible, if
C(φ1 , φ2 , ϑ, u) = C(φ1 , φ2 , −ϑ, u),
φi ∈ [0, π], ϑ ∈ [−2π, 2π), u ∈ R, i = 1, 2.
(2.7)
The use of statistical models based on axially symmetric and longitudinal reversible stochastic processes on the sphere is advocated in Stein (2007) for the analysis of total column ozone. We show
throughout the paper that the construction in Equation (2.7) is especially important for implementing dynamical models as described in Section 4.1, where a solely spatial version CS (φ1 , φ2 , ϑ)
7
is used. Note that a longitudinally reversible covariance function C might be separable, in which
case
C(φ1 , φ2 , ϑ, u) = CS (φ1 , φ2 , ϑ) CT (u),
(φi , ϑ, u) ∈ [0, π] × [−2π, 2π) × R, i = 1, 2,
(2.8)
with CS and CT being marginal covariances in their respective spaces.
3
Second Order Approaches
In this section we provide a list of techniques that are used in the literature to implement space-time
covariance functions on the two-dimensional sphere cross time.
3.1
Spectral Representations and Related Stochastic Expansions
Spectral representations in Euclidean spaces have been known since the work of Schoenberg (1938)
and extended to space-time in Cressie and Huang (1999) and then in Gneiting (2002). The analogue
of spectral representations over spheres was then provided by Schoenberg (1942). The case of the
sphere cross time has been unknown until the recent work of Berg and Porcu (2017): finding a
spectral representation for geodesically isotropic space-time covariance function is equivalent to
providing a characterization of the class P(S2 , R). Namely, Berg and Porcu (2017) establish that
a continuous mapping ψ is a member of the class P(S2 , R) if and only if
∞
ψ(dGC , u) = ∑ Ck,T (u)Pk (cos dGC ),
(dGC , u) ∈ [0, π] × R,
(3.1)
k=0
where {Ck,T (⋅)}∞
k=0 is a sequence of temporal covariance functions with the additional requirement
2
that ∑∞
k=0 Ck,T (0) < ∞ in order to guarantee the variance σ = ψ(0, 0) to be finite. Here, Pk (x)
denotes the kth Legendre polynomial, x ∈ [−1, 1] (see Dai and Xu, 2013, for more details). We refer
to (3.1) as spectral representation of the space time covariance ψ(dGC , u).
Berg and Porcu (2017) showed a general representation for the case of the n-dimensional sphere
Sn cross time (see Appendix A for details). This fact is not merely a mathematical artifact, but
also the key for modeling strategies, as shown in Porcu et al. (2016b).
Some comments are in order. Clearly, ψS (dGC ) = ψ(dGC , 0) is the geodesically isotropic part
of a spatial covariance defined over the sphere, a characterization of which can be found in the
notable work by Schoenberg (1942); see also the recent review by Gneiting (2002). Furthermore,
representation (2.6) implies that ψ is separable if and only if the elements Ck,T of the sequence
{Ck,T (⋅)}∞
k=0 in Equation (3.1) are of the form
Ck,T (u) = bk ρT (u),
u ∈ R,
k ∈ {0, 1, . . .},
with {bk }∞
k=0 being a uniquely determined probability sequence, and ρT a temporal covariance
function. We follow Daley and Porcu (2013) and call {bk }∞
k=0 a 2-Schoenberg sequence to emphasize
that the coefficients bk depends on the dimension of the two-dimensional sphere where ψ is defined.
8
Figure 1: Simulated space-time data from the double Karhunen–Loève expansion used by Clarke
et al. (2016), using 17,000 spatial sites on S2 and two temporal instants. The order of trunction is
50 for both expansions.
It can be proved that the covariance functions from the series expansion in (3.1) have a corresponding Gaussian process with an associated series expansion as well. Indeed, direct inspection
together with the addition theorem for spherical harmonics (Marinucci and Peccati, 2011) shows
that (3.1) corresponds to a Gaussian process on S2 ×R, defined by the stochastic expansion of Jones
(1963):
∞
k
Z(s, t) = ∑ ∑ Ak,` (t)Yk,` (s),
(s, t) ∈ S2 × R,
(3.2)
k=0 `=−k
where Yk,` are the deterministic spherical harmonics on S2 (Dai and Xu, 2013), and each of the
k
zero mean Gaussian processes {Ak,` (t)}∞,
k=0,`=−k satisfies
E(Ak,` (t)Ak′ ,`′ (t′ )) = δk=k′ δ`=`′ Ck,T (t − t′ ),
t, t′ ∈ R,
(3.3)
with δ denoting the classical Kronecker delta, and {Ck,T }∞
k=0 a sequence of temporal covariances
summable at zero.
Unfortunately, the representation does not allow for many interesting examples in closed forms.
Hence, using the results in Berg and Porcu (2017), Porcu et al. (2016b) proposed a spectral representation being valid on any n-dimensional sphere cross R, namely
∞
ψ(dGC , u) = ∑ CT (u)k cos(dGC )k ,
(dGC , u) ∈ [0, π] × R,
(3.4)
k=0
where CT ∶ R → (0, 1] is a temporal correlation function. Some examples are reported from Porcu
et al. (2016b) in Table 1. Details are explained in Appendix A.
Clarke et al. (2016) take advantage of the stochastic representation (3.2) to study the regularity
properties of a Gaussian process Z on S2 × R, in terms of dynamical fractal dimensions and Hölder
exponents. This is achieved by taking a double Karhunen–Loève expansion, e.g. by expanding
each term Ak,` (t) in terms of basis functions (in particular, Hermite polynomials). Such a double
Karhunen–Loève representation is also the key for fast simulation. Clarke et al. (2016) propose to
truncate the order of the double expansions and evaluate the error bound in the L2 sense. Figure 1
depicts an example with this method, using 17, 000 points on the sphere and two time instants.
9
Note how in each of the parametric families outlined in Table 1, the spatial margin ψS (dGC ) =
ψ(dGC , 0) is an analytic function. In particular, we have that the spatial margin is either non
differentiable or infinitely differentiable at the origin.
Spectral representations for stochastic processes over spheres, with the additional feature of
axial symmetry, have been proposed in Jones (1963), Narcowich (1995), Hitczenko and Stein (2012)
and Castruccio and Stein (2013). We are not aware of extensions of this representation to spacetime, but the work in Jones (1963) suggests that axial symmetry over the sphere coupled with
temporal symmetry should be obtained by relaxing the assumption (3.3) for the processes Ak,` in
the expansion (3.2).
An alternative approach to model axially symmetric processes over spheres cross time is proposed in Castruccio and Guinness (2017) on the basis of the stochastic representations in Jones
(1963).
Table 1: Parametric families of covariance functions on S2 ×R obtained through the representation in
Equation (3.4). Second column reports the analytic expression, where g is any correlation function
on the real line. An additional condition is required for the Sine Power family (refer to Porcu et al.,
2016b, for details). All of the members ψ in the second column are rescaled so that ψ(0, 0) = 1.
We use the abuse of notation r for the great circle distance dGC .
Family
Negative Binomial
3.2
Analytic expression
Parameters range
1−ε
τ
ψ(r, u) = ( 1−εg(u)
cos(r) )
ε ∈ (0, 1), τ > 0
(1−ε)2
( 1+ε2 −2εg(u) cos(r) )τ
1 g(u) cos(r)−1
(1 + g(u) cos(r))
2e
−α
α/2
ε ∈ (0, 1), τ > 0
(1+g 2 (u))(1−ε)
( 1+g2 (u)−2εg(u) cos(r) )τ
ε ∈ (0, 1), τ > 0
Multiquadric
ψ(r, u) =
Sine Series
ψ(r, u) =
Sine Power
ψ(r, u) = 1 − 2
(1 − g(u) cos(r))
Adapted Multiquadric
ψ(r, u) =
Poisson
ψ(r, u) = exp (λ(cos(r)g(u) − 1))
α ∈ (0, 2]
λ>0
Scale Mixture Representations
Scale mixtures provide a powerful and elegant way to build members of the class P(S2 , R). We
sketch the general principle here and then report some examples. Let F be a positive measure on
the positive real line. Let ψS (dGC ; ξ) be a geodesically isotropic spatial covariance for any ξ ∈ R+ .
Also, let CT (u; ξ) be a covariance for any ξ. Then, arguments in Porcu and Zastavnyi (2011) and
Gneiting et al. (2007) show that the function
ψ(dGC , u) = ∫
ψS (dGC ; ξ)CT (u; ξ)F (dξ),
(dGC , u) ∈ [0, π] × R,
R+
is a nonseparable geodesically isotropic and temporally symmetric space-time covariance function.
The scale mixture approach offers a nice interpretation in terms of the construction of the
associated process. Let Ξ be a random variable with probability distribution F having a finite first
moment. Let X and Y be, respectively, a purely spatial and a purely temporal Gaussian process.
10
Figure 2: Simulated space-time data from the covariance (3.6), over 17,000 spatial sites on S2 and
2 temporal instants.
Also, suppose that the random variable Ξ and the two processes are mutually independent. Let
Z(s, t ∣ Ξ = ξ) = X(sξ)Y (tξ),
(s, t) ∈ S2 × R,
where ξ is a realization from Ξ and let Z(s, t) = EZ(s, t ∣ Ξ), with expectation E taken with respect
to F . Then, the covariance of Z admits a scale mixture representation.
In general, it is not possible adapt the classes of nonseparable covariance functions based on
scale mixtures and defined over R3 × R to the case S2 × R. For instance, Gaussian scale mixtures as proposed in Schlather (2010) cannot be implemented on the sphere because the function
CS (dGC ; ξ) = exp(−(dGC /ξ)α ) is positive definite on the sphere only when α ∈ (0, 1] (see Gneiting,
2013). Some caution needs be taken when considering the scale mixture based covariances defined
on Euclidean spaces by Porcu et al. (2007); Porcu and Mateu (2007) and Fonseca and Steel (2011),
amongst others.
To illustrate some valid examples, further notation is needed. A function f ∶ [0, ∞) → (0, ∞) is
called completely monotonic if it is infinitely differentiable on the positive real line and if its jth
order derivatives f (j) satisfy (−1)j f (j) (t) ≥ 0, t > 0, j = 0, 1, 2, . . .. A scale mixture argument in
Porcu et al. (2016b) shows that the function
ψ(dGC , u) =
σ2
f (dGC γT (u)),
γT (u)3
(dGC , u) ∈ [0, π] × R,
(3.5)
is a nonseparable covariance function on S2 × R provided f is completely monotonic, with f (0) = 1,
and γT a variogram, with γT (0) = 1. Here, σ 2 denotes the variance. We consider a special case of
Equation (3.5), namely
⎛ dGC (1 +
⎜−
exp
ψ(dGC , u) =
∣u∣ 3
bS
⎝
(1 + bT )
σ2
∣u∣
⎞
bT )
⎟,
⎠
(3.6)
which corresponds to f (t) = exp(−t), t ≥ 0, with γT (u) = (1 + ∣u∣), u ∈ R. Here, σ 2 , bS and bT
are positive parameters associated with the variance, spatial scale and temporal scale of the field,
respectively. Figure 2 shows a realization using σ 2 = 1, bS = 0.3 and bT = 0.5.
Another scale mixture approach allows for an adaptation of the Gneiting class (Gneiting, 2002;
Zastavnyi and Porcu, 2011). Again, Porcu et al. (2016b) show how to derive a space-time covariance
11
of the type
ψ(dGC , u) =
σ2
u
f(
),
1/2
g[0,π] (dGC )
g[0,π] (dGC )
(dGC , u) ∈ [0, π] × R,
(3.7)
where f is completely monotonic, with f (0) = 1, and g[0,π] is the restriction to the interval [0, π]
of a function, g ∶ [0, ∞) → R+ , having a completely monotonic derivative (see Porcu and Schilling,
2011, for a description with classes of functions having this property).
We are now going to outline a new result within the scale-mixture based approach. Specifically,
we consider quasi arithmetic means, as defined in Porcu et al. (2010): these allow to obtain spacetime covariances with given margins in space and time. The construction is defined as
ψ(dGC , u) = Qf (ψS (dGC ), CT (u)) ,
(dGC , u) ∈ S2 × R,
(3.8)
where Qf (x, y) = f (1/2f −1 (x) + 1/2f −1 (y)), for a function f that is completely monotonic on the
positive real line, and having a proper inverse, f −1 , which is always well defined because completely
monotonic functions are strictly decreasing. There is a wide list of such functions available and
the reader is referred to Schilling et al. (2012) and to Porcu and Schilling (2011). Quasi arithmetic
means include as special case all the other means (e.g., the geometric, harmonic, and the Gini
means), and the reader is referred to Porcu et al. (2010) for a complete description of this framework.
Using similar arguments as in Porcu et al. (2010), Theorem 6.3 in Appendix C shows the
conditions for which ψ in Equation (3.8) is a geodesically isotropic space-time covariance function.
3.3
Space-Time Compact Supports
Let ψ be a member of the class P(S2 , R). We say that ψ is dynamically compactly supported on
the sphere if there exists a function h ∶ R → (0, c], with c ≤ π, such that for every fixed temporal
lag u0 , the function ψS (dGC ; u0 ) = ψ(dGC , u0 ) is a covariance function on S2 , and is compactly
supported on the interval [0, h(u0 )).
Let (x)+ denote the positive part of a real number x. For µ > 0 and k = 0, 1, 2, . . ., we define
Wendland functions ϕµ,k (Wendland, 1995) by
ϕµ,k (x) = (1 − x)µ+k
+ Pk (x) ,
x ≥ 0,
(3.9)
where Pk is a polynomial of order k. Special cases are depicted in the second column of Table 2.
We provide new results in the remainder of this subsection. Using scale mixture arguments, we
can show that
µ
ψ(dGC , u) =
dGC
σ2
dGC
σ2
α
h(∣u∣)
ϕ
(
)
=
h(∣u∣)α (1 −
) ,
µ,0
α
α
c
h(∣u∣)
c
h(∣u∣) +
(3.10)
(dGC , u) ∈ [0, π] × R, where, for α ≥ 3 and µ ≥ 4, is a covariance function on S2 × R provided
h is positive, decreasing and convex on the positive real line, with h(0) = c, 0 < c ≤ π, and
limt→∞ h(t) = 0. The additional technical restriction on µ and α is explained in Theorem 6.4 in
Appendix C, where a formal proof is given. An example can better clarifly things. A potential
12
Table 2: Wendland correlation ϕµ,k (⋅) and Matérn correlation Mν (⋅) with increasing smoothness
parameters k and ν. SP (k) means that the sample paths of the associated Gaussian field are k
times differentiable.
k
ϕµ,k (r)
ν
0
(1 − r)µ+
0.5
e−r
1
(1 − r)µ+1
+ (1 + r(µ + 1))
1.5
e−r (1 + r)
2
1
2 2
(1 − r)µ+2
+ (1 + r(µ + 2) + r (µ + 4µ + 3) 3 )
1
2
2
(1 − r)µ+3
+ (1 + r(µ + 3) + r (2µ + 12µ + 15) 5
3
+r (µ + 9µ
3
3
2
1
)
+ 23µ + 15) 15
Mν (r)
SP (k)
0
2.5
e−r (1 + r +
3.5
e−r (1 +
r
2
1
r2
3
)
6
+ r2 15
+
2
r3
15 )
3
candidate that can be used as dynamical support in (3.10) is the function h(t) = c(1 + t)−1/α , t ≥ 0,
α ≥ 3 and 0 < c ≤ π. Then, previous construction becomes
µ
dGC
σ2
(1 −
) ,
ψ(dGC , u) =
(1 + ∣u∣)
c(1 + ∣u∣)−1/α +
(dGC , u) ∈ [0, π] × R,
(3.11)
which shows that the spatial margin ψ(dGC , 0) has compact support c. If c = π, then ψ(dGC , 0)
becomes globally supported. Note how the compact support becomes smaller as the temporal lag
increases, which is a very intuitive property.
The following result characterizes a class of dynamically supported Wendland functions on the
sphere cross time, being a generalization of (3.10) to arbitrary k ∈ N.
Theorem 3.1. Let 0 < c ≤ π. Let h ∶ [0, ∞) → (0, c], be positive, decreasing and convex on the
positive real line, with limt→∞ h(t) = 0. Let k be a positive integer and α ≥ 2k + 2. Let ϕµ,k be the
Wendland function defined in Equation (3.9). Then,
ψ(dGC , u) = σ 2 h(∣u∣)α ϕµ,k (
dGC
),
h(∣u∣)
(dGC , u) ∈ [0, π] × R,
(3.12)
is a covariance function on the sphere cross time, provided µ ≥ k + 4.
A more general statement involves compactly supported space-time covariance functions on the
circle, where no parametric forms are imposed on the involved functions.
Theorem 3.2. Let ϕ ∶ R × R → R be a covariance functions that is symmetric in both first and
second argument, such that ϕ(x, u) = 0 whenever ∣x∣ ≥ π, for all u ∈ R. Call ψ the restriction of
ϕ to the interval [0, π] with respect to the first argument. Then, ψ is a geodesically isotropic and
temporally symmetric covariance function on the circle S1 cross time R.
The above result is unfortunately insufficient to build for building compactly supported models
over spheres cross time. Theorem 3.1 offers a special case with dynamically supported Wendland
functions. To obtain a general assertion on the basis of Theorem 3.2, we give some hints in the
research problems at the end of the manuscript.
13
3.4
Covariance Models under the Lagrangian framework
Environmental, atmospheric, and geophysical processes are often influenced by prevailing winds or
ocean currents (Gneiting et al., 2007). In this type of situation, the general idea of a Lagrangian
reference frame applies. Alegria and Porcu (2016a) considered a simple Lagrangian framework
that allows for transport effects over spheres. Namely, they consider random orthogonal (3 × 3)
matrices R with determinant identically equal to one, such that R−1 = R⊺ , with ⊺ denoting the
transpose operator. Standard theory on random orthogonal rotations shows that R = QDQ−1 ,
with Q denoting a matrix containing the eigenvectors of R, and D a diagonal matrix containing
the associated eigenvalues. Also, we have that each eigenvalue λk , k = 1, 2, 3, can be uniquely
written as λk = exp(ıκk ), with ı being the unit imaginary number and κk real, for k = 1, 2, 3. Then,
following Gantmacher (1960), one can define the tth real power of R, as
Rt = Q(diag (exp(ıκk t)) )Q−1 .
Let X be a Gaussian process on S2 with geodesically isotropic covariance CS (s1 , s2 ) = ψS (dGC (s1 , s2 )).
Define
Z(s, t) = X (Rt s) ,
s ∈ S2 , t ∈ R.
(3.13)
Then, the space-time covariance function with transport effect can be expressed as
C(s1 , s2 , u) = E(CS (dGC (Ru s1 , s2 ))),
s1 , s2 ∈ S2 ,
u ∈ R,
(3.14)
where the expectation is taken with respect to the random rotation R. The fact that the resulting
covariance is still geodesically isotropic in the spatial component is nontrivial and is shown formally,
for some specific choice of the random rotation R, in Alegria and Porcu (2016a), at least for the
case of the sphere S2 . Some comments are in order. The resulting field Z in Equation (3.13) is not
Gaussian (it is Gaussian conditionally on R only). Also, obtaining closed forms for the associated
covariance is generally difficult. Alegria and Porcu (2016a) provide some special cases.
For the specification of the random rotation R in the Lagrangian covariance (3.14), various
choices can be physically motivated and justified. The reader is referrerd to Gneiting et al. (2007)
for a thorough discussion. For instance, the random rotation might represent a prevailing wind as
in Gupta and Waymire (1987). It might be a the westerly wind considered by Haslett and Raftery
(1989) or again, it might be updated dynamically according to the current state of the atmosphere.
Of course, the model (3.14) represents only a first step into Lagrangian modeling over spheres. The
concept of transport effect should be put into a broader context, e.g. by following the discussion
at pages 319–320 of Cressie and Wikle (2011), which shows that the Lagrangian model offered here
is a very special case, obtained when the spatial field is moving at a constant velocity. Innovative
approaches to space-time data under transport effects can be found in Wikle et al. (2001) and
Cressie et al. (2010).
14
3.5
Covariance Models based on Chordal Distances
On the basis of the results in Yadrenko (1983) and Yaglom (1987), Jun and Stein (2007, 2008)
exploit the fact that, for any covariance function CS on R3 , the function CS (dCH ) is a covariance
function on S2 . The Matérn function (e.g., Stein, 1999) is defined as
Mν (x) = σ 2
21−ν ν
x Kν (x) ,
Γ(ν)
x ≥ 0,
(3.15)
where ν > 0 governs the mean square differentiability of the associated process Z on S2 , which is k
times mean differentiable if and only if ν > k. Here, Kν is a modified Bessel function.
The Matérn model coupled with chordal distance, that is Mν (dCH ), is valid on S2 for any
positive ν. Unfortunately, arguments in Gneiting (2013) show that Mν (dGC ) is a valid model on
S2 only for 0 < ν ≤ 1/2, making its use impractical. Space-time models based on geodesic distance
inherit this limitation. The same argument of the Matérn covariance is used in Guinness and Fuentes
(2016) and Jeong and Jun (2015) to assert that the chordal distance might be preferable with respect
to the great circle distance. For instance, a Matérn-Gneiting (Gneiting, 2002; Porcu and Zastavnyi,
2011) type covariance function based on chordal distance might be easily implemented. For a
positive valued temporal variogram γT with γT (0) = 1, the function
C(dCH , u) =
σ2
dCH
Mν (
),
3/2
γT (u)
γT (u)
(dCH , u) ∈ [0, 2] × R
(3.16)
is a covariance function on S2 × R for any positive ν. The spatial margin C(dCH , 0) is of Matérn
type and keeps all the desirable features in terms of differentiability at the origin.
The use of chordal distance has been extensively criticized in the literature. For instance,
because the chordal distance underestimates the true distance between the points on the sphere,
Porcu et al. (2016b) argue that this fact has a nonnegligible impact on the estimation of the spatial
scale. Moreover, Gneiting (2013) argues that the chordal distance is counter to spherical geometry
for larger values of the great circle distance, and thus may result in physically unrealistic distortions.
Further, covariance functions based on chordal distance inherit the limitations of isotropic models
in Euclidean spaces in modeling covariances with negative values. For instance, a covariance based
on chordal distance on S2 does not allow for values lower than −0.21σ 2 , with σ 2 being the variance
as before. Instead, properties of Legendre polynomials (see Szegő, 1939) imply that correlations
based on geodesic distance can attain any value between −1 and +1. Another argument in favor
of the great circle distance is that the differentiability of a given covariance function depending
on the great circle distance can be modeled by imposing a given rate of decay of the associated
2-Schoenberg coefficients {bk }∞
k=0 , as shown in Møller et al. (2017) or even modeling the rate of
decay of the associated 2-Schoenberg functions in (3.1), as shown by Clarke et al. (2016).
A good alternative to the model in Equation (3.16) is the space-time Wendland model based on
the great circle distance, as defined through Equation (3.10). Table 2 highlights a striking connection between the two approaches, showing that Wendland functions allow for a parameterization
15
of the differentiability at the origin in the same way that Matérn does. The compact support of
the Wendland functions can imply some problems in terms of loss of accuracy of kriging predictors,
as shown in Stein (1999). However, recent encouraging results (Bevilacqua et al., 2016) show that
such a loss is negligible under infill asymptotics.
3.6
Nonstationary and Anisotropic Space-Time Covariance Functions
The literature regarding the construction of nonstationary space-time covariance functions is sparse.
The works of Jun and Stein (2007, 2008) are notable exceptions. The methods proposed by the
authors are based on the coupling of the chordal distance with certain classes of differential operators. Estrade et al. (2016) have generalized the Berg-Porcu (Berg and Porcu, 2017) representation
of geodesically isotropic-temporally symmetric covariance functions on Sd × R. In particular, two
generalizations are obtained: an extension of the Berg-Porcu class to the case of temporally nonstationary covariances and a new class that allows for local anisotropy. Anisotropy is also considered
in the tour de force by Hitczenko and Stein (2012), on the basis of chordal distances. Anisotropic
components can be induced through the use of Wigner matrices, as explained in Marinucci and
Peccati (2011).
More recently, Alegria and Porcu (2016b) have considered geodesically isotropic space-time covariance functions that allow the separation of the linear from the cyclical component in the temporal lag. The authors show that such approach offers considerable gains in terms of predictive
performance, in particular in the presence of temporal cyclic components or in the presence of
nonstationarities that are normally removed when detrending the data. A recent discussion about
nonstationary approaches can be found in Fuglstad et al. (2015) who work under the SPDE framework.
4
Practical Approaches
4.1
Dynamical Approaches
When analyzing massive data sets arising from, for instance, remotely sensed networks or satellite
constellations, climate models, or reanalysis data product, a model that fully specifies the covariance
function becomes impractical. Indeed, it would require a prohibitive amount of information to
be stored in the covariance matrix, as well as a prohibitive number of flops and iterations for
maximizing the likelihood or exploring the posterior distribution. Hence, a compromise between
inferential feasibility and model flexibility must be achieved. These limitations arise independently
on the space where the random field is defined. For the case of the sphere, the drawbacks are
magnified by the considerably more sparse literature on the construction of valid global space-time
processes, as shown in the previous section, as well as by the extremely large size of global data.
In Euclidean spaces, a very popular approach to drastically reduce the complexity is to separate
the spatial and temporal components and to describe the dynamics of the process by specifying its
evolution as a function of the past. Variability is then achieved by assuming a random spatial in16
novation. For modeling global climate fields this approach has been further simplified by modeling
the temporal dynamics through covariates only (Furrer et al., 2007; Geinitz et al., 2015).
The dynamical approach has received strong support from reference textbooks in space-time modelling (see e.g. Cressie and Wikle, 2011) and recent years have also seen some development of this
methodology in the context of global space-time data (Castruccio and Stein, 2013). In this regard
the EM algorithm has been efficiently implemented by Fassó et al. (2016) and successfully used at
continental level for multivariate spatio temporal data (Finazzi and Fassó, 2014). Thanks to the
Matérn covariance implemented on the sphere, this code works also at the global level.
Let us assume, with no loss of generality, that the process has zero mean and that it is observed at some locations s1 , . . . , sN ∈ S2 , for equally spaced time points t. We denote Z t =
(Z(s1 , t), . . . , Z(sN , t))⊺ the process at time t, and we specify its dynamic through the following
recursive equation
Z t = Et (Z t−1 , Z t−2 , . . . , Z t−p ) + εt ,
(4.1)
where Et incorporates the evolution of past trajectory of the process up to time t − p, and εt =
iid
(ε(s1 , t), . . . , ε(sn , t))⊺ ∼ N (0, Σ) is an innovation vector with purely spatial global covariance
matrix Σ.
The considerable benefit of such an approach is that the spatio-temporal structure of the model
is specified by the dynamical evolution E and the spatial innovation εt . Hence, the temporal
and spatial part of the model are de-coupled, and this allows for a considerably more convenient
inference. Recent work on satellite data has proposed to couple the dynamical approach dimension
reduction techniques, and in particular Fixed Rank Kriging (see Cressie and Johannesson, 2008;
Nguyen et al., 2014) to further reduce the parameter dimensionality, and to achieve a fit for very
large data sets (Fixed Rank Filtering, see Kang et al., 2010; Cressie et al., 2010). While these
approaches have been very effective for interpolating large data sets, they do not explicitly account
for the spherical geometry, but rather project the data on the Euclidean space. Here, we consider
dynamical models for space-time global data with an explicit definition of the covariance function
in S2 .
A particularly appealing class of models with (4.1) is the Vector AutoRegressive model VAR(p),
defined as
p
Z t = ∑ Φj Z t−j + εt ,
(4.2)
j=1
where Φj are N × N matrices that encode the temporal dependence at lag t − j. This model results
in a space-time precision matrix that is block banded, and hence would greatly reduce the storage
burden arising from massive data sets as shown in Section 5.
4.1.1
The Innovation Structure
While models such as (4.2) allow for a substantial computational saving, in typical climate model
applications the number of locations is larger than 10, 000. Thus, even the likelihood of the innovaiid
tion εt ∼ N (0, Σ) is impossible to evaluate for memory issues on a laptop. Therefore, to perform
17
feasible inference on the full data set, additional structure on the model or spatial design of the data
must be leveraged on. Castruccio and Stein (2013) consider the Fourier transform cross longitude
of an axially symmetric process, that is
∞
εt (φ, ϑ) = ∑ eıϑk f(k; φ)̃
εt (k; φ),
k=0
corr(̃
εt (k; φ), ̃
εt (k ′ ; φ′ )) = δk=k′ ρ(k; φ, φ′ ),
φ, φ′ ∈ [0, π], ϑ ∈ [0, 2π),
(4.3)
with ε̃t (k; φ) being the Fourier process for wavenumber k and latitude φ. Here, f(k; φ) is the
spectrum at latitude φ and wavenumber k, and, for any pair of latitudes (φ, φ′ ), the function
ρ(k; φ, φ′ ) defines a spectral correlation (it is also called coherence). Jun and Stein (2008) showed
that there is a considerable computational benefit if the data are on a N = Nφ × Nϑ (latitude ×
longitude) regular grid over the sphere and if the process is axially symmetric. Indeed, they showed
how the resulting covariance matrix is block circulant and, most importantly, block diagonal in
the spectral domain, thus requiring only O(Nφ × Nϑ2 ) entries to store instead of O(Nφ2 × Nϑ2 ), and
O(Nφ ×Nϑ3 ) flops instead of O(Nφ3 ×Nϑ3 ). Castruccio and Stein (2013) showed that such structure can
be used to perform inference to massive data sets from computer model ensembles (in the range of
107 to 109 data points), by first estimating the spectrum f(k; φ) for each of the Nφ latitudinal bands
in parallel, and then conditionally estimating the structure of the spectral correlation ρ(k; φ, φ′ ).
This approach has been extended to analyze three-dimensional temperature profiles in a regular
grid for a data set larger than one billion data points, allowing for an extension of axially symmetric
models in three dimensions (Castruccio and Genton, 2016).
The spectral approach is not just a mere strategy to simplify inference, but also a key to generalize axially symmetric processes to exhibit nonstationarities, by imposing additional structure in
the process transformed in the spectral domain, while still guaranteeing positive definiteness of the
corresponding covariance functions. In particular, it is still possible to retain the computational
convenience of the gridded geometry while assuming nonstationary models cross longitudes. Castruccio and Genton (2014) explored this idea by assuming that ̃
εt (k; φ) in (4.3) are correlated across
frequencies, with a fully nonparametric dependence structure to be estimated using time replicates,
and they showed how the axially symmetric assumption is badly violated for temperature data.
In order to allow for ocean transitions, Castruccio and Guinness (2017) impose the spectrum
f(⋅; φ) in (4.3) to depend on longitude as well, and call it evolutionary spectrum (Priestley, 1965).
Given two spectra, fi (k; φ), i = 1, 2 and a mapping bland ∶ [0, π] × [0, 2π) → [0, 1], an evolutionary
spectrum is attained through the convex combination
f(k; φ, ϑ) = f1 (k; φ)bland (φ, ϑ) + f2 (k; φ) {1 − bland (φ, ϑ)} ,
(4.4)
so that bland plays the role of modulating the relative contribution of the land regime. Castruccio
and Guinness (2017) showed how this approach is able to capture the majority of the nonstationarity
occurring over a single latitudinal band. Recently, Jeong et al. (2017) proposed an extension to this
18
approach to incorporate mountain ranges in the evolutionary spectrum in the context of wind fields.
Even if the data are not on a regular grid, Horrell and Stein (2015) showed that it is still possible
to interpolate satellite data on a gridded structure using interpolated likelihoods to leverage on
spectral methods. They propose to first perform kriging on the original observations, interpolate
them over a grid, and evaluate the likelihood of these pseudo-observations.
The spectral approach to global data allows to achieve a fit for data sets of remarkable size,
and allows to generalize axially symmetric models to capture longitudinal nonstationarities. Such a
great improvement, however, comes at the cost of a loss of interpretability of the notion of distance.
Indeed, the wavenumber k differs in physical length for different latitudes. Thus, interpreting the
dependence structure across latitudes is problematic, especially near the poles, where the physical
distance among points is very small. Additionally, a process specified with a latitudinally varying
spectrum is not, in general, mean square continuous at the poles. Conditions for regularity at these
two singular points have been discussed in Castruccio and Guinness (2017).
4.1.2
The Temporal Structure
For sufficiently aggregated data, (4.2) can be further simplified by assuming Φj =diag(φii;j ), so that
the inference can be performed in parallel for each location. Simple diagnostics have shown that
this structure is adequate for data spanning from multi-decadal to monthly data, while sub-monthly
data would likely require a more sophisticated neighboring-dependent structure. A key feature of
some geophysical processes is their dependence along thousands of miles. These teleconnections
would require a more complex structure for Φk that would link far away locations.
4.2
The SPDE Approach
We start with a brief description of the SPDE approach as in Lindgren et al. (2011) and Bolin and
Lindgren (2011). We skip all the mathematical details and will stick to the main idea. Clearly, we
shall detail our exposition by working on the sphere. The main idea in Lindgren et al. (2011) is
to evade from a direct specification of the Matérn function Mν as defined in Equation (3.15) in
order to be able to work on any manifold, which of course includes the sphere S2 . Taking verbatim
from Lindgren et al. (2011), our main objective is to construct Matérn fields on the sphere, which is
important for the analysis of global spatial and spatiotemporal models. Also, the authors note that
they want to avoid to use the Matérn covariance adapted with chordal distance, in order to avoid
the interpretational disadvantage of using chordal distances to determine the correlation between
points.
The solution is to consider the SPDE having as solution a Gaussian field with Matérn covariance
function: for a merely spatial process X on S2 , the authors study the SPDE defined through
(κ2 − ∆)α X(s) = W(s),
s ∈ S2 ,
(4.5)
where κ > 0, ∆ is the Laplace-Beltrami operator, and W is a Gaussian white noise on the sphere.
Here, the positive exponent α depends on the parameter ν in (3.15) as well as on the dimension of
19
the sphere.
In order to provide a computationally convenient approximation of (4.5), Lindgren et al. (2011) find
a very ingenious computationally efficient Hilbert space approximation. Namely, the weak solution
to (4.5) if found in some approximation space spanned by some basis functions. The computational
efficiency is then attained by imposing local basis functions, that is basis functions being compactly
supported. This all boils into approximating the field X with a Gaussian Markov field, x, with
precision matrix Q. This idea is then generalized in Bolin and Lindgren (2011) through nested
SPDE models.
The SPDE approach proposed in Lindgren et al. (2011) is ingeniously coupled with hierarchical
models by Cameletti et al. (2012) to provide a space-time model. Our exposition is adapted to the
domain S2 × R. The authors propose a model of the type
Z(s, t) = Υ⊺t β + Wt (s) + εt (s),
Wt (s) = aWt−1 (s) + Wt (s),
(s, t) ∈ S2 × R.
Here, Υt is the vector of covariates and β is a parameter vector. The process εt (s) expresses the
measurement error at time t and location s on the sphere. The latent process Wt (s) is autoregressive
over time and Wt (s) is a purely spatial process with Matérn covariance function as in Equation
(3.15). The bridge with the SPDE approach thus comes from the following assumptions: W
is approximated through a Gaussian Markov structure with a given precision matrix. The same
approximation, but with another precision matrix, is then assigned to the process W1 (s). Using the
dynamic approach, Cameletti et al. (2012) shows that the whole latent process W has a Gaussian
Markov structure with a sparse precision matrix that can be calculated explicitly. A direct spacetime formulation of the SPDE approach is also suggested in Lindgren et al. (2011).
5
Data example
The 2015 El Niño Southern Oscillation (ENSO) was registered as the most intense over the last
two decades (Wolter and Timlin, 2011). In September-October 2015, strong ENSO conditions
coupled with the Indian Ocean Dipole suppressed precipitation and resulted in dry highly flammable
landscape in Equatorial Asia. The extent of the haze from fires in the region was the largest
recorded since 1997, and the increased particulate matter concentration over the densely populated
area resulted in tens of millions of people being exposed to very unhealthy to hazardous air quality
(as defined by the Pollutant Standard Index National Environment Agency, 2016) and resulting into
one of the worst environmental disasters on record. The assessment of exposure and mortality from
this event is critical for the implementation of future mitigation strategies, and the estimation of
these numbers with the associated uncertainty has received widespread media attention (Shannon
et al., 2016) and even official of estimates from scientific studies from local governments. While it
is possible to focus on regional data to provide local exposure estimates, such simulations are very
20
Figure 3: Averages for AOD for the months of September (a) and October (b).
hard to perform and have been attempted only by Crippa et al. (2016), while all other studies have
been focused on more affordable and readily available data on a global scale.
Here, we focus on this event using the MERRA2 reanalysis (Molod et al., 2015) data of daily
(aggregated from 3-hourly) Aerosol Optical Depth (AOD), see Figure 3 and movie in the online
supplement. Values around 0.01 mean clear sky conditions, 0.4 very hazy condition and around
1 extremely toxic. We have also removed the trend, where for every point the monthly mean
(location-wise) has been subtracted. We use this data set to compare different statistical models
fitted globally but focused in the region of interest. The practical approaches discussed in Section
4 would allow for a full estimation over the two months (for a total of more than 12 million points),
but that would not be possible with second order approaches. Therefore, we limit our analysis to
the first 6 days of October. While a detailed study would require an additional comparison in terms
of exposure maps from population estimates, we decided to avoid this for the sake of simplicity and
brevity.
5.1
The Second Order Approach
We fit the data set with a second order approach as detailed in Section 3. Despite the subsampling
in time, the data set is still too large for a full analysis. Hence, a further subsampling in space
a grid of 10 degrees in latitude and longitude was performed. Since several points concentred on
the poles, this can produce numerical problems. We thus remove a small portion of observations,
which correspond to latitudes greater than 85 degrees and lower than −85 degrees.
We have considered four models:
Model 1. A modified Gneiting covariance, as detailed by Equation (3.6).
Model 2.
A dynamically compactly supported covariance, as in Equation (3.11).
To explain the choices for Models 3 and 4, we consider the Gneiting (Gneiting, 2002) function
⎛
⎞
∣u∣
⎜
⎟.
K(r, u) =
exp ⎜−
3
1/2 ⎟
r
r
(1 + bS )
⎝ bT (1 + bS ) ⎠
σ2
(5.1)
Model 3. A Gneiting type covariance coupled with the great circle distance, that is ψ(dGC , u) =
K(dGC , u)), with K as in (5.1).
21
Model 4. A Gneiting type covariance K as in Equation (5.1), coupled with the chordal distance,
dCH .
The proposed models have three parameters, (σ 2 , bS , bT )⊺ indicating the variance, spatial and
temporal range respectively. Inference was performed using a pairwise composite likelihood (CL)
approach, which provides approximate but asymptotically unbiased estimates to very large data
sets. We use the CL method with observations whose spatial distance is less than 6,378 kilometers
(equivalent to 1 radians on a unit sphere).
Table 3 shows CL estimates and the Log-CL value at the maximum, and both models have a
similar performance in terms of CL. Figure 4 illustrates the empirical spatial (semi) variogram in
terms of the great circle distance, at different temporal lags, versus the theoretical models. The
models are indeed able to capture this large-scale feature of the data by fitting well the empirical
variograms.
●
●
●
0.002
0.004
●
Model 1
Model 2
Model 3
●
●
0
1000
2000
3000
4000
5000
0.008
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.006
●
0.004
●
●
Model 1
Model 2
Model 3
●
0.002
●
Spatial Semi−variogram
0.006
●
●
0.000
0.008
●
●
|u|=1
0.000
Spatial Semi−variogram
|u|=0
0
Great circle distance (Kilometers)
1000
2000
3000
4000
5000
Great circle distance (Kilometers)
Figure 4: Empirical spatial (semi) variograms versus theoretical covariances according to Models
1, 2 and 3, at different temporal lags.
We now compare the models in terms of their predictive performance. We use the kriging predictor and a drop-one prediction strategy. We consider the following indicators: Mean Squared Prediction Error (MSPE), Log-Score (LSCORE) and Continuous Ranked Probability Score (CRPS).
Table 3 contains the indicators for each model. Small values of these indicators suggest better
predictions.
Models 3 and 4 produce better predictive results with respect to Models 1 and 2. In particular,
Model 3 generates an approximate improvement of 2% with respect to Model 4, in terms of MPSE.
Model 3, based on the great circle distance, outperforms Model 4 in terms of LSCORE and CRPS.
Table 3: Composite likelihood estimates, Log-CL value at the maximum and predictive scores. The
units for the spatial and temporal scales are km and days, respectively.
̂
̂
̂
σ2
bS
bT
Log-CL
MPSE
LSCORE CRPS
Model 1
8.564 × 10−3
816.38
6.320
2223640
5.069 × 10−3
−1.052
0.107
Model 2
8.564 × 10−3
3845.93
6.255
2223574
5.307 × 10−3
−0.935
0.106
Model 3
8.562 × 10−3
1651.90
2.202
2223928
4.442 × 10−3
−1.288
0.117
Model 4
8.563 × 10−3
765.36
5.601
2223828
4.527 × 10−3
−1.281
0.118
22
2
log spectrum
1
0
φ=65.9° N
−1
−2
°
φ=0.5 S
−3
−4
50
100
150
wavenumber
200
250
Figure 5: Fitted parametrical (blue) and non parametrical (i.e., periodogram red) log spectrum for
two different latitudinal bands, one at the equator, one at high northern latitudes. The periodogram
was obtained by averaging across all longitudes and times.
5.2
Dynamical approach
We now fit the same data set with a dynamical model. We choose (4.2) with p = 1, i.e., a VAR(1)
process, with a diagonal autoregressive structure. Providing a location-specific temporal structure is
likely to be too flexible for data subsampled for only 6 days, but further model selection approaches
to reduce model complexity were deemed out of the scope of this work. Since the data are gridded,
we choose a spectral model (4.3), with a latitudinally varying spectrum, and a coherence for the
same wavenumber, but independence otherwise. While the model was subsampled in time for a
comparison with the second-order approach, the inference was scalable and the fit of the entire
data set did not require a significant additional computational time.
Even if the likelihood evaluation can sidestep the big N problem by storing the results in the
spectral domain via a Whittle likelihood (Whittle, 1953), it is not possible to obtain a maximum
likelihood for the entire model. Because the model’s structure, comprising of hundreds of parameters, the optimization over the entire space would be impossible. Therefore, the estimation is
performed step-wise. First, the temporal dependence is estimated. Then, we estimate the longitudinal dependence, then the latitudinal dependence, and hence the parameter estimates are to be
regarded as local approximations.
The spectral model allows for a different spectral shape at different latitudes, and it is flexible
enough to capture very different behaviours, as illustrated in Figure 5. Indeed, AOD residuals near
the equator display a much smoother behaviour than at high latitudes, where the spectrum is more
flat, but the model is able to fit both behaviours naturally.
5.3
Comparison for Equatorial Asia
While both models are fitted globally, the interest lies in their relative performance in the area
of interest, i.e., in Equatorial Asia. Table 4 shows a comparison of the aforementioned models
in terms of the marginal likelihood in this area, defined as all points whose latitudes is between
−11.3 and 15.3 degrees (Nφ = 53) and whose longitude is between 93 to 137 degrees (Nϑ = 71).
The dynamically specified model outperforms almost uniformly Models 1 and 2 in the second order
23
approach: it is faster, richer in complexity yet achieving a overwhelmingly better fit, so it is clearly
more suitable for a more detailed analysis in the area.
While these results are strongly in favor of the dynamically approach, it must be pointed out
that the particular setting of the global space-time data analyzed, i.e., a regular grid in space and
equal observations in time, is such that a spectral model is clearly the best choice. An application
with more complicated geometries such as with satellite data, would have required an adaptation
of spectral methods along the lines of Horrell and Stein (2015) and, likely, a worse fit. While such a
comparison with an irregular design would have been added to the discussion, we decided to avoid
it for the sake of brevity.
Table 4: Comparison between different models in terms of number of parameters (excluding the temporal
ones), computational time (minutes), normalized loglikelihood, and Bayesian Information Criterion for K = 6
days from October 1st to October 6th 2015.
6
# param
Time (minutes)
∆-loglik/(Nφ Nϑ K)
BIC×104
Model 1
3
238
−3.24
−4.50
Model 2
3
236
−3.24
−4.55
Model (4.3)
161
4.2
0
−19.0
Research Problems
This section is devoted to a list of research problems that we consider relevant for future researches.
1. Nonstationary Space-Time Covariances. The models proposed in Section 3 are geodesically isotropic and temporally symmetric. Since real phenomena on the globe are notably
nonstationary, it is necessary to use those models as building blocks for more sophisticated
constructions. In particular, the development of nonstationary models is necessary and a
promising direction of research is to extend the work of Guella et al. (2017, 2016b,a) who
define kernels over products of n-dimensional spheres. Another direction of research might
be to take into account differences in local geometry of the sphere representing planet Earth.
Thus, the natural solution is to work on processes evolving temporally over Riemannian manifolds. In this direction, the works of Menegatto et al. (2006) and the recent work by Barbosa
and Menegatto (2017) might be very useful.
2. Models based on Convolutions. Convolution arguments have been used for Gaussian processes defined over spheres (no time) in order to index the associated fractal dimension.
Hansen et al. (2015) show that Gaussian particles provide a flexible framework for modelling
and simulating three-dimensional star-shaped random sets. The authors show that, if the
kernel is a von Mises-Fisher density, or uniform on a spherical cap, the correlation function
of the associated random field admits a closed form expression. We are not aware of any
convolution argument for space-time covariance functions. The work of Ziegel (2014) might
be very useful in this direction.
24
3. Physical-based Constructions. Constructions based on dynamical models or moving averages have been proposed by Ailliot et al. (2011) and Schlather (2010). We are not aware of
such extensions to the case S2 × R, but certainly it would be worth being studied. Some other
constructions based on physical characteristic of the space-time process might be appealing
for modeling several real processes. In this direction, it would be desirable to study the approaches proposed in Christakos (2000) as well as Kolovos et al. (2004); Christakos (1991);
Christopoulos and Tsantili (2016) and Christakos et al. (2000a).
4. Multivariate Space-Time Models. Often, several variables are observed over the same spatial location and same temporal resolution. Thus, there is substantial need of space-time
multivariate covariance models being geodesically isotropic or axially symmetric in the spatial component. The literature in this subject is very sparse, with the notable exceptions of
Jun and Stein (2008) and Alegria et al. (2017).
5. Matérn Analogues over Spheres. Find the counterpart of the Matérn covariance function
on the sphere. This would allow to have a covariance function that indexes the differentiability at the origin of the associated Gaussian field. Notable attempts have been made by
Guinness and Fuentes (2016), with partial success. Møller et al. (2017) suggest to model the
n-Schoenberg coefficients (see Appendix A for details) imposing a given rate of decay, but
this does not allow for explicit closed forms.
5. Compact Supports with Differentiable Temporal Margins. A potential drawback of the
space-time construction as in Theorem 3.1 is that it only allows for temporal margins being
non differentiable at the origin. An important step ahead would be to improve such a limitation. Following the arguments in the proof of Theorem 3.1, the crux would be in proving
that, for some mapping h ∶ [0, ∞) → R, the function
h(u2 )α+d 1 F2 {1 + k;
µ+2
ν +3
+ k,
+ k; −n2 h(u2 )2 /4},
2
2
u ∈ R,
n ∈ N,
is positive definite for all n ∈ N.
6. Spectral Constructions à la Stein. Stein (2005a) proposes a class of spectral densities in
Rd × R of the type
f(ω, τ ) = ((1 + ∥ω∥)α1 + (1 + ∣τ ∣)α2 )−α3 ,
(ω, τ ) ∈ Rd × R,
αi > 0,
for i = 1, 2, 3. Under the condition α3 < α1 /d + α2 , f is in L1 (Rd × R). The partial Fourier
transforms allow for indexing the differentiability at the origin in a similar way as the Matérn
covariance function. Further, the resulting covariance is smoother away from the origing
than at the origin, which is important as reported in Theorem 1 in Stein (2005a). There is no
analogue of f-based constructions for the class P(S2 , R). One should necessarily start from
Berg-Porcu (Berg and Porcu, 2017) characterization and try to find a related approach that
allows to obtain such a construction on spheres cross time.
25
7. Local Anisotropies and Transport Effects. The geodesically isotropic models should be
extended to allow for local anisotropies as in Paciorek and Schervish (2006). In this direction,
a major step should be made in order to generalize the Lagrangian model in Equation (3.14)
to the nonstationary case.
8. Strictly Positive Definite Functions. A function C ∶ (S2 ×R)2 → R is called strictly positive
definite when inequality
N
∑ ci C((si , ti ), (sj , tj ))cj > 0,
i,j
N
d
holds for any {ck }N
k=1 ⊂ R, unless c1 = ⋅ ⋅ ⋅ = cN = 0, and {sk , tk }k=1 ⊂ S ×R. There is substantial
work on strict positive definiteness and the reader is referred to Chen et al. (2003); Menegatto
(1994, 1995) and Menegatto et al. (2006). A characterization of the subclass of P(Sn , R) (see
Appendix A) with members ψ such that the corresponding covariances C are strictly positive
definite is still elusive.
9. Walks through Dimensions. The literature on walks through dimensions is related to operators that allow, for a given positive definite function on the n-dimensional sphere, to obtain
new classes of positive definite functions on n′ -dimensional spheres, with n ≠ n′ . The application of such operators has consequences on the differentiability at the origin of the involved
functions. Walks on spheres have been proposed by Beatson et al. (2014), Ziegel (2014) and
by Massa et al. (2017), this last extending previous work to the case of complex spheres. It
would be timely to obtain walks through dimensions for the members of the classes P(Sn , R),
for n a positive integer.
10. Optimal Prediction on Spheres cross Time. Much work needs to be done in order to
assess asymptotic optimal kriging prediction over spheres cross time when the covariance
is missspecified. The work of Arafat et al. (2016) opens a bridge between equivalence of
Gaussian measures and asymptotic optimal prediction over spheres. A relevant direction of
research would be to evaluate the screening effect on spheres cross time. After the works
of Stein (1999), there is nothing related to spectral behaviours over spheres that allow for
evaluating the corresponding screening effect.
11. Fast Simulations on Spheres. Lang and Schwab (2013) and Clarke et al. (2016) propose
fast simulations through truncation of Karhunen-Loève expansions. It would be very interesting to propose analogues of circulant embedding methods, where the difficulty is mainly
due to the fact that it is not possible to set a regular grid on the sphere.
12. Chordal vs Great Circle Again. Moreno Bevilacqua (personal communication) points out
that he has tried the following experiment. Simulate any set of points on the sphere cross time.
Take any element ψ from the class P(S2 , R). Replace the great circle distance with the chordal
distance and calculate the corresponding matrix realizations. With all the experiments, he
26
always found strictly positive eigenvalues. Thus, the question is: suppose that ψ ∈ P(S2 , R).
Then, is it true that ψ(dCH ) is positive definite on S2 × S2 × R?
Appendix
Appendix A: Mathematical Background
We need some notation in order to illustrate the following sections. This material follows largely
the exposition in Berg and Porcu (2017).
Let n be a positive integer. We denote Sn the n-dimensional unit sphere of Rn+1 , given as
Sn = {s ∈ Rn+1 ∣ ∥s∥ = 1}, n ≥ 1.
(6.1)
We also consider
∞
S∞ = {(sk )k∈N ∈ RN ∣ ∑ s2k = 1},
k=1
which is the unit sphere in the Hilbert sequence space `2 of square summable real sequences.
Following Berg and Porcu (2017), we thus define the class P(Sn , R) as the class of continuous
functions ψ ∶ [0, π] × R → R such that
C((s1 , t1 ), (s2 , t2 )) = ψ(dGC (s1 , s2 ), t1 − t2 ),
(si , ti ) ∈ Sn × R,
i = 1, 2,
where dGC as already been defined as the great circle distance. Berg and Porcu (2017) define
P(S∞ , R) as ⋂n≥1 P(Sn , R). The inclusion relation
P(S1 , R) ⊃ P(S2 , R) ⊃ ⋯ ⊃ P(S∞ , R)
is strict and the reader is referred to Berg and Porcu (2017) for details.
(λ)
We recall the definition of Gegenbauer polynomials Ck , given by the generating function (see
Dai and Xu, 2013)
∞
(1 − 2xz + r2 )−λ = ∑ Ck (z)rk ,
(λ)
∣r∣ < 1, z ∈ C.
(6.2)
k=0
Here, λ > 0. We have the classical orthogonality relation:
1
∫
−1
(1 − x2 )λ−1/2 Ck (x)Ch (x) dx =
(λ)
(λ)
πΓ(k + 2λ)21−2λ
δh=k ,
Γ2 (λ)(k + λ)k!
(6.3)
with Γ denoting the Gamma function.
It is of fundamental importance that ∣Ck (x)∣ ≤ Ck (1), x ∈ [−1, 1]. The special value λ = (n−1)/2
(λ)
(λ)
is relevant for the sphere Sn because of the relation to spherical harmonics, which is illustrated in
Berg and Porcu (2017) as follows. A spherical harmonic of degree k for Sn is the restriction to Sn
of a real-valued harmonic homogeneous polynomial in Rn+1 of degree k. Together with the zero
function, the spherical harmonics of degree k form a finite dimensional vector space denoted Hk (n).
It is a subspace of the space C(Sn ) of continuous functions on Sn . We have
Nk (n) ∶= dim Hk (n) =
(n)k−1
(2n + n − 1), k ≥ 1,
k!
27
N0 (n) = 1,
(6.4)
(see Dai and Xu, 2013, p.3). Here, (n)k−1 denotes the Pochammer symbol.
The surface measure of the sphere is denoted ωn , and it has total mass
∣∣ωn ∣∣ =
2π (n+1)/2
.
Γ((n + 1)/2)
(6.5)
The spaces Hk (n) are mutual orthogonal subspaces of the Hilbert space L2 (Sn , ωn ), which they
generate. This means that any F ∈ L2 (Sn , ωn ) has an orthogonal expansion
∞
∞
F = ∑ Sk , Sk ∈ Hk (n),
∣∣F ∣∣22 = ∑ ∣∣Sk ∣∣22 ,
k=0
(6.6)
k=0
where the first series converges in L2 (Sn , ωn ), and the second series is Parseval’s equation (Berg
and Porcu, 2017). Here Sk is the orthogonal projection of F onto Hk (n) given as
Sk (ξ) =
Nk (n)
∫ ck (n, ⟨s, η⟩)F (η) dωn (η).
∣∣ωn ∣∣ Sn
(6.7)
See the addition theorem for spherical harmonics, (Schoenberg, 1942). For λ = (n − 1)/2, ck (n, x)
is defined as the normalized Gegenbauer polynomial
ck (n, x) = Ck
((n−1)/2)
(x)/Ck
((n−1)/2)
(1) =
k!
((n−1)/2)
C
(x),
(n − 1)k k
(6.8)
being 1 for x = 1.
Specializing the orthogonality relation (6.3) to λ = (n − 1)/2 and using Equations (6.4), (6.5), Berg
and Porcu (2017) get for n ∈ N
1
∫
(1 − x2 )n/2−1 ck (n, x)ch (n, x) dx =
−1
∣∣ωn ∣∣
δh=k .
∣∣ωn−1 ∣∣Nk (n)
(6.9)
The following result characterizes completely the class P(Sn , R).
Theorem 6.1. (Berg and Porcu, 2017). Let n ∈ N and let ψ ∶ [0, π] × R → C be a continuous
function. Then ψ belongs to the class P(Sn , R) if and only if there exists a sequence {ϕk,n }∞
k=0 of
positive definite functions on R, with ∑ ϕk,n (0) < ∞, such that such that
∞
ψ(dGC , u) = ∑ ϕk,n (u)ck (n, cos dGC ),
(6.10)
k=0
and the above expansion is uniformly convergent for (dGC , u) ∈ [0, π] × R. We have
ϕk,n (u) =
Nk (n)∣∣ωn−1 ∣∣ π
n−1
dx.
∫ ψ(x, u)ck (n, x) sin x
∣∣ωn ∣∣
0
(6.11)
We also report the characterization of the class P(S∞ , R) obtained by the same authors.
Theorem 6.2. (Berg and Porcu, 2017) Let ψ ∶ [0, π] × R → R be a continuous function. Then ψ
belongs to P(S∞ , R) if and only if there exists a sequence {ϕk }∞
k=0 of positive definite functions on
R, with ∑k ϕk (0) < ∞ such that
∞
ψ(dGC , u) = ∑ ϕk (u) cosk dGC ,
(6.12)
k=0
and the above expansion is uniformly convergent for (dGC , u) ∈ [0, π] × G.
Some comments are in order. Theorems 6.1 and 6.2 are fundamental to create the examples
illustrated in Table 1. Also, they are crucial to solve some of the open problems that we reported
in Section 6. Table 5 lists some permissible models on S2 × R.
28
Table 5: Parametric families of geodesically isotropic space-time covariance functions. PMC is the acronym
for Porcu-Mateu-Christakos Porcu et al. (2010). PBG is the acronym for Porcu-Bevilacqua-Genton Porcu
et al. (2016b). We use the abuse of notation r for the great circle distance dGC .
Family
Expression for ψ(r, u)
Parameter Restrictions
Fonseca-Steel
√
δ
1 + (γS (r) + γT (u)) Kλ1 (2 (a + γS (r)) )
√
λ1
Kλ1 (2 aδ)
(1 + γS (r))
a1 , δ, λi > 0, i = 0, 1.
−ρ3
((1 + r )
α ρ1
Clayton PMC
+ (1 + ∣u∣ )
β ρ2
− 1)
α3
exp ( − (rα1 + uα2 )
Gumbel PMC
PBG
α, β, ρ1 , ρ2 ∈ (0, 1]; ρ3 > 0.
)
αi ∈ (0, 1], i = 1, 2, 3.
Equation (3.5)
Families for f
Expression
Parameters
τ
β
Dagum
r
f (r) = 1 − ( 1+r
β)
β, τ ∈ (0, 1]
Matérn
Equation (3.15)
0 < ν ≤ 1/2
Gen. Cauchy
f (r) = (1 + rα )−β/α
α ∈ (0, 1], β > 0
Pow. Exponential
f (r) = exp(−rα )
α ∈ (0, 1]
Adaptive Gneiting
Equation (3.7)
Families for g[0,π]
f as in PBG family
Expression
Parameters
β
τ
r
g[0,π] (r) = 1 + ( 1+r
β)
Dagum
β, τ ∈ (0, 1]
Gen. Cauchy
g[0,π] (r) = (1 + rα )β/α
α ∈ (0, 1], β ≤ α
Power
g[0,π] (r) = c + rα
α ∈ (0, 1], c > 0
PBG2
Equation (3.10)
h positive, decreasing and
convex with limt→∞ h(t) = 0.
Dynamical Wendland
Equation (3.12)
h positive, decreasing and
convex with limt→∞ h(t) = 0.
Families for ϕµ,k
Dynamical Wendland 2
Dynamical Wendland 4
Expression
h(∣u∣)α (1 −
h(∣u∣)α (1 −
6
r
r
) (1 + 6 h(∣u∣)
)
h(∣u∣) +
8
r
r
) (1 + 8 h(∣u∣)
h(∣u∣) +
29
Parameters
+
63
3
α≥4
2
r
( h(∣u∣)
) )
α≥4
Appendix C. Mathematical Proofs
C1. The Quasi Arithmetic Class on Spheres
For ease of exposition, we slightly deviate from the notation of the paper and denote f ○ g the
composition of f to g. Let us recall the expression of quasi arithmetic covariances as in Equation
(3.8):
1
1
(dGC , u) ∈ [0, π] × R.
ψ(dGC , u) = f ( f −1 ○ ψS (dGC ) + f −1 ○ CT (u)) ,
2
2
By Bernstein’s theorem (Feller, 1966, p. 439), a function f ∶ [0, ∞) → R is completely monotonic if
and only if
f (t) = ∫
t ≥ 0,
e−ξt H(dξ),
(6.13)
[0,∞)
where H is a positive and bounded measure. A Bernstein function is a continuous mapping on the
positive real line, having a first derivative being completely monotonic. We are now able to give a
formal assertion for the validity of the quasi arithmetic construction.
Theorem 6.3. Let f ∶ [0, ∞) → R+ be a completely monotonic function. Let f1 ∶ [0, ∞) → R be
a continuous functions such that f −1 ○ f1 is a Bernstein function. Let CT ∶ R → R be a continuous, symmetric covariance function such f −1 ○ CT is a temporal variogram. Call ψS = f1,[0,π] the
restriction of f1 to the interval [0, π]. Then,
ψ(dGC , u) = Qf (ψS (dGC ), CT (u)) ,
(dGC , u) ∈ [0, π] × R,
(6.14)
is a geodesically isotropic space-time covariance function on Sn × R, for all n = 1, 2, 3, . . ..
Proof. Denote g the composition f −1 ○ f1 and call g[0,π] the restriction of g to [0, π], obtained
through g[0,π] = f −1 ○ f1,[0,π] . By assumption, g is a Bernstein function. Thus, arguments in Porcu
and Schilling (2011), with the references therein, show that the function h, defined through
h(t; ξ) = exp(−ξg(t)),
t, ξ ≥ 0,
is a completely monotonic fuction for any positive ξ. Thus,
h[0,π] (dGC ; ξ) = exp (−ξf −1 ○ f1,[0,π] (dGC )) ,
dGC ∈ [0, π],
is the restriction of a completely monotonic function to the interval [0, π]. Invoking Theorem 7
in Gneiting (2013), we obtain that h[0,π] is a geodesically isotropic covariance function on any ndimensional sphere. Additionally, since f −1 ○ CT is a temporal variogram, by Schoenberg’s theorem
(Schoenberg, 1938) we deduce that k(u; ξ) = exp(−ξf −1 ○ CT (u)), u ∈ R, is a covariance function on
the real line for every positive ξ. Thus, the scale mixture covariance
ψ(dGC , u) = ∫
[0,∞)
h[0,π] (dGC ; ξ)k(u; ξ)H(dξ)
= ∫
[0,∞)
exp (−ξg[0,π] (dGC ) − ξf −1 ○ CT (u)) H(dξ)
= Qf (ψS (dGC ), CT (u)),
is a covariance function on Sn × R for all n ∈ N.
30
(dGC , u) ∈ [0, π] × R,
C2. Proofs for Section 3.3
Theorem 6.4. Let h be a positive, decreasing and convex function on the positive real line, with
h(0) = c, 0 < c ≤ π, and limt→∞ h(t) = 0. Let α ≥ 3 and µ ≥ 4. Then, Equation (3.10) defines a
geodesically isotropic and temporally symmetric covariance function on S3 × R.
Proof. The proof is based on scale mixture arguments in concert with the calculations in Porcu
et al. (2016a). In particular, we have that the function in Equation (3.10) is the result of the scale
mixture of the function ψS (dGC ; ξ) = (1 − dGC /ξ)n+ with the function CT (u; ξ) = ξ n (1 −
γ
ξ
h(∣u∣) )+ ,
ξ > 0, t ≥ 0, γ ≥ 1, n ≥ 2. In particular, arguments in Lemmas 3 and 4 in Gneiting (2013) show
that ψS is a covariance function on S3 for every positive ξ. Under the required conditions on the
function h, we have that CT is a covariance function on R for every positive ξ. Thus, the scale
mixture arguments in Porcu et al. (2016a), with µ = n + γ + 1 and α = n + 1 can now be applied,
obtaining the result.
A more sophisticated argument is required to show that the structure in Equation (3.12) is
positive definite on the circle cross time. We start with the proof of Theorem 3.2 because its
arguments will be partially used to prove Theorem 3.1.
Proof of Theorem 3.2. Denote ϕ[0,π] the restriction of ϕ to the interval [0, π] with respect to the
first argument. Let n ∈ N. Consider the sequence of functions bn (⋅), defined through
bn (u) =
=
π
π
2
2
∫ cos(nz)ψ(z, u)dz = ∫ cos(nz)ϕ[0,π] (z, u)dz
π 0
π 0
∞
1
∫ cos(nx)ϕ(x, u)dx.
π −∞
(6.15)
Since ϕ is positive definite on R × R, arguments in Lemma 1 in Gneiting (2002) show that bn (u) is
a covariance function on R for all n ∈ N. Additionally, we have
∞
∞
∞
1
̂N (n) < ∞,
cos(nx)ϕ(x,
0)dx
=
∑ϕ
∫
n=0 π −∞
n=0
∞
∑ bn (0) = ∑
n=0
̂N denotes the Fourier transform of ϕ(x, 0) restricted to natural numbers. Thus, we get
where ϕ
that ∑n bn (0) < ∞ because the Fourier transform of a positive definite function is nonnegative and
integrable. We can thus invoke Theorem 3.3 in Berg and Porcu (2017) to obtain that ψ(dGC , u) =
ϕ[0,π] (dGC , u) is a covariance function on the circle S1 cross time.
Proof of Theorem 3.1. We give a proof of the constructive type. Let k ∈ N. Arguments in Theorem
1 of Porcu et al. (2016a) show that the function
C(x, u) = σ 2 h(∣u∣)α ϕµ,k (
∣x∣
),
h(∣u∣)
(x, u) ∈ R × R,
(6.16)
is positive definite on R × R provided α ≥ 2k + 3 and µ ≥ k + 4. Arguments in Porcu et al. (2016c)
show that
bn (u)
=
−∞
∫
cos(nx)C(x, u)dx
∞
∝ σ 2 h(∣u∣)α+d 1 F2 (1 + k;
ν +3
µ+2
+ k,
+ k; −n2 h(∣u∣)2 /4),
2
2
31
u ∈ R,
n ∈ N.
From the argument in (6.15) in concert with Lemma 1 in Gneiting (2002), we have that bn (u) is
positive definite on the positive real line for each n. Additionally, arguments in Porcu et al. (2016c)
show that, for µ ≥ k + 4, bn (u) is strictly decreasing in n. Application of Proposition 3.6 of Berg
and Porcu (2017) shows that (3.12) is positive definite on S3 × R.
Acknowledgments
We gratefully acknowledge the NASA scientists responsible for MERRA-2 products.
References
Ailliot, P., Baxevani, A., Cuzol, A., Monbet, V., and Raillard, N. (2011). Space-Time Models for Moving
Fields with an Application to Significant Wave Height Fields. Environmetrics, 22(3):354–369.
Alegria, A. and Porcu, E. (2016a). The Dimple Problem related to Space-Time Modeling under the Lagrangian Framework. Technical Report, University Federico Santa Maria.
Alegria, A. and Porcu, E. (2016b). Space-Time Geostatistical Models with both Linear and Seasonal Structures in the Temporal Components. Technical Report, University Federico Santa Maria.
Alegria, A., Porcu, E., Furrer, R., and Mateu, J. (2017). Covariance Functions for Multivariate Gaussian
Fields Evolving Temporally over Planet Earth. Technical Report, University Federico Santa Maria.
Arafat, A., Porcu, E., Bevilacqua, M., and Mateu, J. (2016). Equivalence and Orthogonality of Gaussian
Measures on Spheres. Technical Report, University Federico Santa Maria.
Baldi, P. and Marinucci, D. (2006). Some Characterizations of the Spherical Harmonics Coefficients for
Isotropic Random Fields. Statistics and Probability Letters, 77:490–496.
Banerjee, S., Carlin, B. P., and Gelfand, A. E. (2014). Hierarchical Modeling and Analysis for Spatial Data.
CRC Press, second edition.
Banerjee, S., Gelfand, A. E., Finley, A. O., and Sang, H. (2008). Gaussian Prediction Process Models for
Large Spatial Data Sets. J. R. Statist. Soc. B, 70:825–848.
Barbosa, V. S. and Menegatto, V. A. (2017). Strict Positive Definiteness on Products of Compact Two–Point
Homogeneous Spaces. Integral Transforms Spec. Functions, 28(1):56–73.
Beatson, R. K., zu Castell, W., and Xu, Y. (2014). Pólya Criterion for (strict) Positive Definiteness on the
sphere. IMA Journal of Numerical Analysis, 34:550–568.
Berg, C. and Porcu, E. (2017). From Schoenberg Coefficients to Schoenberg Functions. Constructive Approximation, 45:217–241.
Bevilacqua, M., Faouzi, T., Furrer, R., and Porcu, E. (2016). Estimation and Prediction using Generalized
Wendland Covariance Function under Fixed Domain Asymptotics. Technical Report, University Federico
Santa Maria.
Bevilacqua, M., Gaetan, C., Mateu, J., and Porcu, E. (2012). Estimating Space and Space-Time Covariance
Functions: a Weighted Composite Likelihood Approach. J. Am. Statist. Ass., 107:268–280.
Bevilacqua, M., Mateu, J., Porcu, E., Zhang, H., and Zini, A. (2010). Weighted Composite Likelihood-based
Tests for Space-Time Separability of Covariance Functions. Statistics and Computing, 20(3):283–293.
Bingham, N. H. (1973). Positive Definite Functions on Spheres. Proc. Cambridge Phil. Soc., 73:145–156.
Bolin, D. and Lindgren, F. (2011). Spatial Models Generated by Nested Stochastic Partial Differential
Equations, with an Application to Global Ozone Mapping. Ann. Appl. Statist., 5(1):523–550.
Cameletti, M., Lindgren, F., Simpson, D., and Rue, H. (2012). Spatio-Temporal Modeling of Particulate
Matter Concentration through the SPDE Approach. AStA Advances in Statistical Analysis, pages 1–23.
32
Castruccio, S. and Genton, M. G. (2014). Beyond Axial Symmetry: An Improved Class of Models for Global
Data. Stat, 3(1):48–55.
Castruccio, S. and Genton, M. G. (2016). Compressing an Ensemble with Statistical Models: An Algorithm
for Global 3D Spatio-Temporal Temperature. Technometrics, 58(3):319–328.
Castruccio, S. and Guinness, J. (2017). An Evolutionary Spectrum Approach to Incorporate Large-scale
Geographical Descriptors on Global Processes. J. R. Statist. Soc. C, 66(2):329–344.
Castruccio, S. and Stein, M. L. (2013). Global Space-Time Models for Climate Ensembles. Ann. Appl.
Statist., 7(3):1593–1611.
Chen, D., Menegatto, V. A., and Sun, X. (2003). A Necessary and Sufficient Condition for Strictly Positive
Definite Functions on Spheres. Proc. Amer. Math. Soc., 131:2733–2740.
Christakos, G. (1991). On Certain Classes of Spatiotemporal Random Fields with Application to Space-Time
Data Processing. Systems, Man, and Cybernetics, 21(4):861–875.
Christakos, G. (2000). Modern Spatiotemporal Geostatistics. Oxford University Press.
Christakos, G., Hristopoulos, D. T., and Kolovos, A. (2000a). Stochastic Flowpath Analysis of Multiphase
Flow in Random Porous Media. SIAM J. Appl. Math., 60(5):1520–1542.
Christakos, G., Hristopulos, D. T., and Bogaert, P. (2000b). On the Physical Geometry Hypotheses at the
Basis of Spatiotemporal Analysis of Hydrologic Geostatistics. Advances in Water Resources, 23:799–810.
Christakos, G. and Papanicolaou, V. (2000). Norm-Dependent Covariance Permissibility of Weakly Homogeneous Spatial Random Fields. Stoch. Env. Res. Risk A., 14(6):1–8.
Christopoulos, D. and Tsantili, I. (2016). Space-Time Models based on Random Fields with local Interactions.
International Journal of Modern Physics B, 29:26 pages.
Clarke, J., Alegria, A., and Porcu, E. (2016). Regularity Properties and Simulations of Gaussian Random
Fields on the Sphere cross Time. Technical Report, University Federico Santa Maria.
Cressie, N. and Huang, H. C. (1999). Classes of Nonseparable, Spatio-Temporal Stationary Covariance
Functions. J. Am. Statist. Ass., 94(448):1330–1340.
Cressie, N. and Johannesson, G. (2008). Fixed Rank Kriging for Very Large Spatial Data Sets. Journal of
the Royal Statistical Society, Series B, 70:209–226.
Cressie, N., Shi, T., and Kang, E. (2010). Fixed Rank Filtering for Spatio-Temporal Data. Journal of
Computational and Graphical Statistics, 19(3):724–745.
Cressie, N. and Wikle, C. K. (2011). Statistics for Spatio-Temporal Data. Wiley. Wiley & Sons.
Crippa, P., Castruccio, S., Archer-Nicholls, G. B., Lebron, M., Kuwata, A., Thota, S., Sumin, E., Butt, C.,
Wiedinmyer, W., and Spracklen, D. V. (2016). Population Exposure to Hazardous Air Quality due to the
2015 Fires in Equatorial Asia. Scientific Reports, 6:”Article number: 37074”.
Dai, F. and Xu, Y. (2013). Approximation Theory and Harmonic Analysis on Spheres and Balls. Springer.
Daley, D. J. and Porcu, E. (2013). Dimension Walks and Schoenberg Spectral Measures. Proc. Amer. Math.
Society, 141:1813–1824.
Estrade, A., Fariñas, A., and Porcu, E. (2016). Characterization Theorems for Covariance Functions on the
n-Dimensional Sphere Across Time. Technical Report, University Federico Santa Maria.
Fassó, A., Finasi, F., and Ndongo, F. (2016). European Population Exposure to Airborne Pollutants based on
a Multivariate Spatio-Temporal Model. Journal of Agricultural, Biological, and Environmental Statistics,
21. Online first. Doi:10.1007/s13253-016-0260-7.
Feller, W. (1966). An Introduction to Probability Theory and Its Applications (Vol. II). Wiley.
Finazzi, F. and Fassó, A. (2014). D-STEM: a Software for the Analysis and Mapping of Environmental
Space-Time Variables. Journal of Statistical Software, 62:1–29.
33
Fonseca, T. C. O. and Steel, M. F. J. (2011). A General Class of Nonseparable Space-Time Covariance
Models. Environmetrics, 22(2):224–242.
Fuglstad, G. A., Simpson, D., Lindgren, F., and Rue, H. (2015). Does Non-stationary Spatial Data always
require Non-stationary Random Fields? Spatial Statistics, 14:505–531.
Furrer, R., Genton, M. G., and Nychka, D. (2006). Covariance Tapering for Interpolation of Large Spatial
Datasets. Journal of Computational and Graphical Statistics, 15:502–523.
Furrer, R., Sain, S. R., Nychka, D. W., and Meehl, G. A. (2007). Multivariate Bayesian Analysis of
Atmosphere-Ocean General Circulation Models. Environ. Ecol. Stat., 14(3):249–266.
Gangolli, R. (1967). Positive Definite Kernels on Homogeneous Spaces and certain Stochastic Processes
related to Levy’s Brownian Motion of Several Parameters. Ann Inst H Poincare, 3:121–226.
Gantmacher, F. (1960). The Theory of Matrices. Chelsea Pub Co.
Geinitz, S., Furrer, R., and Sain, S. R. (2015). Bayesian Multilevel Analysis of Variance for Relative
Comparison Across Sources of Global Climate Model Variability. Int. J. Climatol., 35(3):433–443.
Gerber, F., Mösinger, L., and Furrer, R. (2017). Extending R Packages to Support 64-bit Compiled Code:
an Illustration with spam64 and GIMMS NDVI3g Data. Comput. Geosci., 104:107–119.
Gneiting, T. (2002). Nonseparable, Stationary Covariance Functions for Space-Time Data. J. Am. Statist.
Ass., 97:590–600.
Gneiting, T. (2013). Strictly and Non-Strictly Positive Definite Functions on Spheres. Bernoulli, 19(4):1327–
1349.
Gneiting, T., Genton, M. G., and Guttorp, P. (2007). Geostatistical Space-Time Models, Stationarity,
Separability and Full Symmetry. In Finkenstadt, B. et al., editors, Statistical Methods for Spatio-Temporal
Systems, pages 151–175. Chapman & Hall/CRC, Boca Raton.
Guella, J. C., Menegatto, V. A., and Peron, A. P. (2016a). An Extension of a Theorem of Schoenberg to a
Product of Spheres. Banach J. Math. Anal., 10(4):671–685.
Guella, J. C., Menegatto, V. A., and Peron, A. P. (2016b). Strictly Positive Definite Kernels on a Product
of Spheres ii. SIGMA, 12(103).
Guella, J. C., Menegatto, V. A., and Peron, A. P. (2017). Strictly Positive Definite Kernels on a Product of
Circles. Positivity, 21(1):329–342.
Guinness, J. and Fuentes, M. (2016). Isotropic Covariance Functions on Spheres: some Properties and
Modeling Considerations. J. Multivariate Anal., 143:143–152.
Gupta, V. K. and Waymire, E. (1987). On Taylor’s Hypothesis and Dissipation in Rainfall. Journal of
Geophysical Research, 92(3):9657–9660.
Hannan, E. J. (1970). Multiple Time Series. Wiley, New York.
Hansen, L. V., Thorarinsdottir, T. L., Ovcharov, E., and Gneiting, T. (2015). Gaussian Random Particles
with Flexible Hausdorff Dimension. Advances in Appl. Probability, 307–327.
Haslett, J. and Raftery, A. E. (1989). Space-Time Modelling with Long-Memory Dependence: Assessing
Irelands Wind-Power Resource. Applied Statistics, 38:1–50.
Hitczenko, M. and Stein, M. L. (2012). Some Theory for Anisotropic Processes on the Sphere. Statist.
Methodology, 9:211–227.
Holben, B. N. et al. (1998). AERONET - a Federated Instrument Network and Data Archive for Aerosol
Characterization. Rem. Sens. Environment, 66:1–16.
Horrell, M. T. and Stein, M. L. (2015). A Covariance Parameter Estimation Method for Polar-Orbiting
Satellite Data. Statistica Sinica, 25(1):41–59.
34
Huang, C., Zhang, H., and Robeson, S. (2012). A Simplified Representation of the Covariance Structure of
Axially Symmetric Processes on the Sphere. Statist. Probab. Letters, 82:1346–1351.
IPCC (2013). Climate Change 2013: The Physical Science Basis. In Stocker, T. et al., editors, Fifth
Assessment Report of the Intergovernmental Panel on Climate Change. Cambridge, New York.
Istas, J. (2005). Spherical and Hyperbolic Fractional Brownian Motion. Electron. Commun. Probability,
10:254–262.
Jeong, J., Castruccio, S., Crippa, P., and Genton, M. G. (2017). Statistics-Based Compression of Global
Wind Fields. J. R. Statist. Soc. C. in press.
Jeong, J. and Jun, M. (2015). A Class of Matern-like Covariance Functions for Smooth Processes on a
Sphere. Spatial Statistics, 11:1–18.
Jones, R. and Zhang, Y. (1997). Models for Continuous Stationary Space-Time Processes. In Gregoire, T.
et al., editors, Modelling Longitudinal and Spatially Correlated Data, Lecture Notes in Statistics, pages
289–298. Springer, New York.
Jones, R. H. (1963). Stochastic Processes on a Sphere. Ann. Math. Statist., 34:213–218.
Jun, M. and Stein, M. L. (2007). An Approach to Producing Space-Time Covariance Functions on Spheres.
Technometrics, 49:468–479.
Jun, M. and Stein, M. L. (2008). Nonstationary Covariance Models for Global Data. Ann. Appl. Statist.,
2(4):1271–1289.
Kang, E. L., Cressie, N., and Shi, T. (2010). Using Temporal Variability to Improve Spatial Mapping with
Application to Satellite Data. Canadian Journal of Statistics, 38(2):271–289.
Kolovos, A., Christakos, G., Hristopulos, D. T., and Serre, M. L. (2004). Methods for Generating Nonseparable Spatiotemporal Covariance Models with Potential. Environmental Applications, Adv. in Water
Resources, 27(8):815–830.
Lang, A. and Schwab, C. (2013). Isotropic Random Fields on the Sphere: Regularity, Fast Simulation and
Stochastic Partial Differential Equations. Annals of Applied Probabilty, 25:3047–3094.
Leonenko, N. and Sakhno, L. (2012). On Spectral Representation of Tensor Random Fields on the Sphere.
Stoch. Anal. Appl., 31:167–182.
Lindgren, F. and Rue, H. (2015). Bayesian Spatial Modelling with R-INLA. Journal of Statistical Software,
63(19):1–25.
Lindgren, F., Rue, H., and Lindstroem, J. (2011). An Explicit Link between Gaussian Fields and Gaussian
Markov Random Fields: the Stochastic Partial Differential Equation Approach. J. R. Statist. Soc. B,
73:423–498.
Malyarenko, A. (2013). Invariant Random Fields on Spaces with a Group Action. Springer, New York.
Marinucci, D. and Peccati, G. (2011). Random Fields on the Sphere, Representation, Limit Theorems and
Cosmological Applications. Cambridge, New York.
Massa, E., Peron, A., and Porcu, E. (2017). Positive Definite Functions on Complex Spheres, and their
Walks through Dimensions. Technical Report. University Federico Santa Maria.
Mateu, J., Porcu, E., and Gregori, P. (2008). Recent Advances to Model Anisotropic Space-Time Data.
Statistical Methods & Applications, 17:209–223.
Menegatto, V. A. (1994). Strictly Positive Definite Kernels on the Hilbert Sphere. Applied Analysis, 55:91–
101.
Menegatto, V. A. (1995). Strictly Positive Definite Kernels on the Circle. Rocky Mountain J. Math.,
25:1149–1163.
Menegatto, V. A., Oliveira, C. P., and Peron, A. P. (2006). Strictly Positive Definite Kernels on Subsets of
the Complex Plane. Comput. Math. Appl., 51:1233–1250.
35
Møller, J., Nielsen, M., Porcu, E., and Rubak, E. (2017). Determinantal Point Process Models on the Sphere.
Bernoulli. To appear.
Molod, A., Takacs, L., Suarez, M., and Bacmeister, J. (2015). Development of the GEOS-5 Atmospheric
General Circulation Model: Evolution from MERRA to MERRA2. Geoscientific Model Development,
8:1339–1356.
Narcowich, F. J. (1995). Generalized Hermite Interpolation and Positive Definite Kernels on a Riemannian
Manifold. J. Math. Anal. Appl., 190:165–193.
National Environment Agency (2016). Air pollution control. http://www.nea.gov.sg/anti-pollution-radiationprotection/air-pollution-control/.
Nguyen, H., Katzfuss, M., Cressie, N., and Braverman, A. (2014). Spatio-Temporal Data Fusion for Very
Large Remote Sensing Datasets. Technometrics, 56(2):174–185.
Oleson, J. J., Kumar, N., and Smith, B. J. (2013). Spatiotemporal Modeling of Irregularly Spaced Aerosol
Optical Depth Data. Environmental and Ecological Statistics, 20(2):297–314.
Paciorek, C. J. and Schervish, M. J. (2006). Spatial Modelling using a New Class of Nonstationary Covariance
Functions. Environmetrics, 17(5):483–506.
Porcu, E., Bevilacqua, M., and Genton, M. G. (2016a). Space-Time Covariance Functions with Dynamical
Compact Supports. Technical Report, University Federico Santa Maria.
Porcu, E., Bevilacqua, M., and Genton, M. G. (2016b). Spatio-Temporal Covariance and Cross-Covariance
Functions of the Great Circle Distance on a Sphere. J. Am. Statist. Ass., 11:888–898.
Porcu, E. and Mateu, J. (2007). Mixture-Based Modeling for Space-Time Data. Environmetrics, 18:285–302.
Porcu, E., Mateu, J., and Bevilacqua, M. (2007). Covariance Functions which are Stationary or Nonstationary in Space and Stationary in Time. Statistica Neerlandica, 61(3):358–382.
Porcu, E., Mateu, J., and Christakos, G. (2010). Quasi-Arithmetic Means of Covariance Functions with
Potential Applications to Space-Time Data. J. Multivariate Anal., 100(8):1830–1844.
Porcu, E. and Schilling, R. (2011). From Schoenberg to Pick-Nevanlinna: Towards a Complete Picture of
the Variogram Class. Bernoulli, 17(1):441–455.
Porcu, E. and Zastavnyi, V. P. (2011). Characterization Theorems for some Classes of Covariance Functions
associated to Vector Valued Random Field. J. Multivariate Anal., 102:1293–1301.
Porcu, E., Zastavnyi, V. P., and Bevilacqua, M. (2016c). Buhmann Covariance Functions, their Compact
Supports, and their Smoothness. Technical Report, University Federico Santa Maria.
Priestley, M. B. (1965). Evolutionary Spectra and Non-stationary Processes. J. R. Statist. Soc. B, 27:204–
237.
Richardson, R., Kotta, A., and Sansó, B. (2016). Bayesian Non-Parametric Modeling for Integro-Difference
equations. Statistics and Computing, 1–15.
Richardson, R., Kotta, A., and Sansó, B. (2017). Flexible Integro-Difference Equation Modeling for SpatioTemporal Data. Computational Statistics & Data Analysis, 109:182–198.
Rue, H. and Tjelmeland, H. (2002). Fitting Gaussian Markov Random Fields to Gaussian Fields. Scandinavian Journal of Statistics, 29:31–49.
Schilling, R., Song, R., and Vondracek, Z. (2012). Bernstein Functions. Theory and Applications. De
Gruyter.
Schlather, M. (2010). Some Covariance Models based on Normal Scale Mixtures. Bernoulli, 16(3):780–797.
Schoenberg, I. J. (1938). Metric Spaces and Completely Monotone Functions. Annals of Mathematics,
25(39):811–841.
Schoenberg, I. J. (1942). Positive Definite Functions on Spheres. Duke Math. Journal, 9:96–108.
36
Shannon, N., Koplitz, L. J., Mickley, M. E., Marlier, J. J., Buonocore, P. S., Kim, T. L., Sulprizio, M. P.,
DeFries, R., Jacob, D. J., Schwartz, J., Pongsiri, M., and Myers, S. S. (2016). Public Health Impacts of
the Severe Haze in Equatorial Asia in September–October 2015: Demonstration of a New Framework for
Informing Fire Management Strategies to reduce Downwind Smoke Exposure. Environmental Research
Letters, 11. Letter number: 094023.
Stein, M. L. (1999). Statistical Interpolation of Spatial Data: Some Theory for Kriging. Springer, New York.
Stein, M. L. (2005a). Space-Time Covariance Functions. J. Am. Statist. Ass., 100(469):310–321.
Stein, M. L. (2005b). Statistical Methods for Regular Monitoring Data. J. R. Statist. Soc. B, 67:667–687.
Stein, M. L. (2007). Spatial Variation of Total Column Ozone on a Global Scale. Ann. Appl. Statistics,
1:191–210.
Szegő, G. (1939). Orthogonal Polynomials, volume XXIII of COLLOQUIUM PUBLICATIONS. American
Mathematical Society.
Tebaldi, C. and Sansó, B. (2009). Joint Projections of Temperature and Precipitation Change from Multiple
Climate Models: a Hierarchical Bayesian Approach. J. Roy. Statist. Soc. Ser. A, 172:83–106.
Wendland, H. (1995). Piecewise Polynomial, Positive Definite and Compactly Supported Radial Functions
of Minimal Degree. Adv. Comput. Math., 4:389–396.
Whittle, P. (1953). The Analysis of Multiple Stationary Time Series. J. R. Statist. Soc. B, 15(1):125–139.
Wikle, C. K., Milliff, R. F., Nychka, D., and Berliner, L. M. (2001). Spatiotemporal Hierarchical Bayesian
Modeling: Tropical Ocean Surface Winds. J. Am. Statist. Ass., 96(454):382–397.
Wolter, K. and Timlin, M. S. (2011). El Niño/Southern Oscillation Behaviour since 1871 as Diagnosed in
an Extended Multivariate ENSO index (MEI.ext). Int. J. Climatol., 31:1074–1087.
Yadrenko, A. M. (1983). Spectral Theory of Random Fields. Optimization Software, New York.
Yaglom, A. M. (1987). Correlation Theory of Stationary and Related Random Functions, volume vol I.
Springer, New York.
Zastavnyi, V. P. and Porcu, E. (2011). Characterization Theorems for the Gneiting Class of Space Time
Covariances. Bernoulli, 17(1):456–465.
Ziegel, J. (2014). Convolution Roots and Differentiability of Isotropic Positive Definite Functions on Spheres.
Proceedings of the American Mathematical Society, 142:2053–2077.
37
| 10math.ST
|
Toward Rapid Transformation of Ideas into Software
Mehrdad Afshari
Zhendong Su
Department of Computer Science, University of California, Davis
arXiv:1602.01075v1 [cs.SE] 2 Feb 2016
{mafshari, su}@ucdavis.edu
ABSTRACT
we organize around several key principles:
A key mission of computer science is to enable people realize
their creative ideas as naturally and painlessly as possible.
Software engineering is at the center of this mission — software technologies enable reification of ideas into working systems. As computers become ubiquitous, both in availability
and the aspects of human lives they touch, the quantity
and diversity of ideas also rapidly grow. Our programming
systems and technologies need to evolve to make this reification process — transforming ideas to software — as quick
and accessible as possible.
The goal of this paper is twofold. First, it advocates and
highlights the “transforming ideas to software” mission as a
moonshot for software engineering research. This is a longterm direction for the community, and there is no silver bullet that can get us there. To make this mission a reality, as a
community, we need to improve the status quo across many
dimensions. Thus, the second goal is to outline a number
of directions to modernize our contemporary programming
technologies for decades to come, describe work that has
been undertaken along those vectors, and pinpoint critical
challenges.
1.
• Quick experimentation: to provide developers with immediate feedback on their code modifications and allow
them to experiment with incomplete systems;
• Programming knowledge reuse: to allow developers quick
access to the vast amount of accumulated programming knowledge and wisdom;
• Proactive programming assistant: to monitor the developers’ actions and proactively feed them relevant
information about the program; and
• Intelligent, conversational interfaces: to provide alternative interfaces that allow developers to express their
intentions and conduct interactive exchanges with the
system.
The two core questions in programming are “What” and
“How”: (1) “What” specifies the intention, and (2) “How”
concerns the solution. The first three principles center around
the “How” question, while the last principle the “What”.
Next, we discuss the above principles, and pinpoint specific
research problems and challenges.
INTRODUCTION
There has been no shortage of creative technological ideas,
but few have been realized — it is a daunting task to transform an idea into a working prototype. Indeed, software
engineering — the process of expressing and refining ideas
in a programming language — has been regarded one of the
most challenging human endeavors. Programming innovations, such as procedural abstraction and object orientation,
have helped increase programmer productivity. However,
we still build software essentially the same way as we did
decades ago. As a community, we should rethink and redesign methodologies and techniques for programming to
make software development more natural and painless to
help people realize their creative ideas.
We believe that transforming ideas into software (TIIS)
should be identified as a long-term, catalytic mission for
the software engineering community. Decades of research
and development have led to better languages, methodologies, tools, environments, and processes. However, it is fair
to say that most have been incremental improvements and
do not promise significant advances demanded for the mission. Identifying and highlighting the TIIS mission can help
unite the community and clarify important research focuses
to achieve significant innovations.
The TIIS mission requires a multi-faceted approach, which
2. DIRECTIONS AND CHALLENGES
The vision for quick transformation of ideas into software
is broad, and advances in a number of directions are necessary and can move the state of affairs forward. We discuss
several directions that we have identified that can be influential toward our goal. We have done early work along some
of these directions and hope the community as a whole can
help accelerate the progress toward improving programming
and in particular, the pace of concretizing ideas.
2.1 Quick Experimentation
Live programming has gained momentum following Bret
Victor’s presentation [16], in which he highlighted the importance of immediate connection between the idea and observing its effect, not just as a catalyst, but as an enabler,
in an effective creative process. Since then, several live programming environments, e.g. Xcode [2] (via its Playground
feature) and LightTable [7] that have been influenced by this
principle.
Prorogued programming [1] is a programming paradigm
that explicitly deals with the issue of quick experimentation.
It is focused on liberating the programmer from having to
deal with programming concerns that are necessary to get a
1
ment.
In software engineering practice, major effort is expended
to integrate various systems and assemble a program from
building blocks. Given the large amount of code available, it
is conceivable that what a programmer plans to write is already written and available in some shape or form. Effective
code search can help the programmer discover the existing
functionality from existing code bases import it in the code
being written [8].
With a mechanism to locate pieces of functionality through
existing APIs or code snippets mined from the Internet, we
need to be able to run the resulting mashup consisting of
the different pieces and quickly experiment with them. A
programming paradigm like prorogued programming is wellsuited for this task. Proroguing programming concerns not
only helps in piecing together the building blocks of functionality discovered in the existing code bases, but also provides
a way to effectively insert holes in the program, which can
be filled later. Filling these holes can be done through traditional implementation, i.e. writing a body for the unimplemented method, or it can be done through more innovative
means, like acting as a signal in addition to the search query
and helping the search engine know the context in which the
code snippet being searched for is going to be live in. In addition to providing that context, the input/output examples
persisted during runtime invocation of prorogued calls are a
great source of input for an I/O-based code search engine
and act as a final filter for validation of code found by a
simple keyword based code search engine.
Collecting data about the programmer’s actions is helpful
in other ways as well. By looking at the actions the programmer performs within their development environment,
for example, it is possible to predict what they intend to
accomplish and propose shortcuts to achieve what they are
aiming for more efficiently [6, 9], thereby educating the programmer and making them more effective in the future. Obviously, this can help the IDE designer improve the development environment and simplify its user interface as well.
Reusing programming knowledge is also beneficial in activities beyond writing code. For instance, we are able to
leverage debugging knowledge accumulated over the previous debugging sessions to automatically help the programmer fix the new, similar, issues [5]. One way that has been
accomplished is by collecting and matching the program
traces that exhibit buggy behavior and pattern matching
new traces against the ones in the bug database, revealing
information about the nature of the bug and how it was
previously fixed, potentially helping the programmer understand and fix the new issue.
partial, incomplete, program running and meaningfully experiment with it and observe its behavior. It does so by providing the ability to annotate function calls or type instantiations with a special keyword, prorogue. The prorogue
keyword acts as a hint for the compiler to let it know that
the implementation for the particular method being called is
unavailable. At runtime, after a prorogued call is executed,
a lazy future object is returned in lieu of the return value
and the program execution continues. Later, if the value of
that object is consulted during the program execution, the
user will be asked to provide a concrete return value for the
call interactively, while presenting him the actual arguments
in that specific invocation. The user interaction will then be
recorded and persisted for the rest of the program execution and for subsequent runs, so that the program can be
run and experimented with in spite of the unimplemented
method body.
In effect, prorogued programming aids quick experimentation and top-down design by letting the programmer freely
rearrange his workflow as he sees fit, rather than having to
follow an order imposed by the toolchain they are using.
More interestingly, through hybrid computation, prorogued
calls can act as hooks to glue a program written in an imperative textual programming language into more domainspecific programming systems that would capture the human
intent much better and in a more concise fashion for particular purposes. The other end of hybrid computation does not
even have to be an imperative program. It can be a machine
learning model that is trained to provide the desired function
that would be hard to express the host language. Alternatively, it can be an interactive system that computes the
desired output through some user interaction. It is possible
to have a hybrid computation engine that is mostly similar to mainstream textual programming languages, except
it is much softer when it comes to interpreting programmer intent, leaving room for the compiler to make educated
guesses and at the same time be more lenient to programmer
mistakes, at the expense of precision.
2.2 Programming Knowledge Reuse
Software is rarely written from scratch. Rather, programs
are generally composed of smaller pieces. That makes software engineering activity largely a system integration process. Software engineers build more complex abstractions
out of simpler ones and that lets them build increasingly sophisticated systems. While seeing the effects of a program
live helps, the question remains that given that there are vast
amounts of source code available on the Internet, should we
move from writing new code to casting programming as a
search problem?
The programming knowledge publicly available today comes
in various forms, such as questions and answers on Stack
Overflow [14], sometimes including code snippets as well
as answers, or through publicly accessible code repositories
such as the ones hosted on GitHub [4].
Commercial software development endeavors also collect
internal data about their development process, including the
version history of the code base, data about bugs and defects, and free-form knowledge in form of comments written on the code review tool, wikis, and sometimes in other
forms, like tracking the time the programmers spend on various tasks, storing the search queries they perform [12], or
looking at their behavior within the development environ-
2.3 Proactive Programming Assistant
Many programming analysis tools have been developed.
In practice, program analyses are primarily left to compile
time and later. We believe that we should surface as much
relevant information as possible to the programmer as soon
as possible. Programming tools should capture runtime data
and run background static or dynamic analysis while the
code is being written, and guide the programmer throughout the coding process. With the popularity of compiler-asa-library solutions like libclang [15] or Roslyn [10], we are
already seeing this shift accelerating. Our editors are indeed
becoming more proactive in issuing compiler warnings and
providing safe refactoring tools.
2
will require significant efforts spanning many directions. We
have identified, as a first step, several directions centered
around four principles. We hope that the community unite
to move the state-of-the-art forward toward the TIIS vision
along these and other important pertinent directions.
That said, in particular, the potential for capturing dynamic information and surfacing it in useful ways while coding remains largely untapped. Among other things, the captured data can feed into the live programming aspects of the
system, providing the user with a concrete view of the program, instead of a purely abstract one relying solely on static
analysis. We believe what information is useful to the programmer and how to best surface it will be an exciting and
impactful avenue for further research.
Speculative analysis [3] and its follow up work can perhaps
be viewed as a specific instance of this direction, where the
focus is on using speculative analysis in the background to
help developers make certain decisions.
4. REFERENCES
[1] M. Afshari, E. T. Barr, and Z. Su. Liberating the
programmer with prorogued programming. In ACM
International Symposium on New Ideas, New
Paradigms, and Reflections on Programming and
Software, pages 11–26, 2012.
[2] Apple Inc. Xcode. http://apple.com/xcode.
[3] Y. Brun, R. Holmes, M. D. Ernst, and D. Notkin.
Speculative analysis: Exploring future development
states of software. In Workshop on the Future of
Software Engineering Research, pages 59–64, Santa Fe,
NM, USA, November 7–8, 2010.
[4] GitHub Inc. GitHub. http://github.com.
[5] Z. Gu, E. T. Barr, D. Schleck, and Z. Su. Reusing
debugging knowledge via trace-based bug search. In
ACM International Conference on Object Oriented
Programming Systems Languages and Applications,
pages 927–942, 2012.
[6] Z. Gu, D. Schleck, E. T. Barr, and Z. Su. Capturing
and exploiting IDE interactions. In ACM
International Symposium on New Ideas, New
Paradigms, and Reflections on Programming and
Software, pages 83–94, 2014.
[7] Kodowa, Inc. LightTable.
http://lighttable.com.
[8] Microsoft Research. Bing code search.
http://codesnippet.research.microsoft.com.
[9] E. Murphy-Hill, R. Jiresal, and G. C. Murphy.
Improving software developers’ fluency by
recommending development environment commands.
In ACM SIGSOFT 20th International Symposium on
the Foundations of Software Engineering, pages
42:1–42:11, 2012.
[10] .NET Foundation. .NET Compiler Platform (Roslyn).
https://github.com/dotnet/roslyn.
[11] F. Pérez and B. E. Granger. IPython: a system for
interactive scientific computing. Computing in Science
and Engineering, 9(3):21–29, May 2007.
[12] C. Sadowski, K. T. Stolee, and S. Elbaum. How
developers search for code: A case study. In Joint
Meeting of the European Software Engineering
Conference and the Symposium on the Foundations of
Software Engineering (ESEC/FSE), 2015.
[13] T. Schachman. Alternative programming interfaces for
alternative programmers. In ACM International
Symposium on New Ideas, New Paradigms, and
Reflections on Programming and Software, pages 1–10,
2012.
[14] Stack Overflow Inc. Stack Overflow.
http://stackoverflow.com.
[15] The Clang Team. Clang tooling.
http://clang.llvm.org/docs/Tooling.html.
[16] B. Victor. Inventing on principle, 2012.
http://vimeo.com/36579366.
2.4 Human Interface Innovation
Textual code is a precise and expressive medium for communicating intent. Looking back at the past half century
of programming history, it is hard to see it going away anytime soon. However, most of the computing devices shipped
today are phones that do not have a physical keyboard and
mouse. While it is conceivable that most of the professional
programming activity would not be done on such devices,
at least without some external accessories, it is almost certain that end-users would want to use them to accomplish
custom computational goals or control systems by defining
actions that would happen in response to specific events.
Accomplishing this requires innovation both on the human interface front and on the backend engine. It is likely
that many of the functionalities will be exposed via the artificial intelligence-based assistant, and will be expressed as
interactive voice conversations. On the backend, we need to
build more interactive programming systems that can make
educated guesses and synthesize programs with incomplete
specification, and interactively adapt it as the specification
is perfected by gradually asking for and capturing additional
user input.
Moreover, even on more traditional computers, e.g. desktops and laptops, we need fundamental interface innovations
to support alternative programmers [13], i.e. people who are
not professional programmers and write programs that does
computation and produces a result, which is the object of
interest to them, as opposed to the program itself. An important class of people who would benefit from such interface
innovations are people doing analysis on various data sets.
Already, tools like IPython [11] that have more interactive
characteristics and suit domain-specific use-cases well have
gained widespread adoption in that community. We believe
that there is enormous potential to carry out research that
would substantially impact the life of alternative programmers in a positive way in this area.
3.
CONCLUSION
In this paper, we have advocated the TIIS mission for
quick realization of ideas as working systems and the modernization of our techniques and tools to better support
programming in the coming decades. Given the ubiquity
of connected computer systems — mostly in the form of
smartphones — we are just at the beginning of an explosion
of ideas and applications that wait to be realized by professional developers or end users. Consequently, it is even
more important that we do our best as a community to improve our programming practice to adapt it for the future
challenges we will likely face. Achieving the TIIS mission
3
| 6cs.PL
|
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH
ZERO-DIVISORS
arXiv:1312.7403v1 [math.AC] 28 Dec 2013
CHRISTOPHER PARK MOONEY
Abstract. Recently substantial progress has been made on generalized factorization techniques in integral domains, in particular τ -factorization. There has also been advances made
in investigating factorization in commutative rings with zero-divisors. One approach which
has been found to be very successful is that of U-factorization introduced by C.R. Fletcher.
We seek to synthesize work done in these two areas by generalizing τ -factorization to rings
with zero-divisors by using the notion of U-factorization.
2010 AMS Subject Classification: 13A05, 13E99, 13F15
1. Introduction
Much work has been done on generalized factorization techniques in integral domains.
There is an excellent overview in [4], where particular attention is paid to τ -factorization.
Several authors have investigated ways to extend factorization to commutative rings with
zero-divisors. For instance, D.D. Anderson, Valdez-Leon, Aǧargün, Chun [3, 6, 7]. One
particular method was that of U-factorization introduced by C.R. Fletcher in [11] and [12].
This method of factorization has been studied extensively by Michael Axtell and others
in [9, 10, 8]. We synthesize this work done into a single study of what we will call τ -Ufactorization.
In this paper, we will assume R is a commutative ring with 1. Let R∗ = R − {0}, let
U(R) be the set of units of R, and let R# = R∗ − U(R) be the non-zero, non-units of
R. As in [10], we define U-factorization as follows. Let a ∈ R be a non-unit. If a =
λa1 · · · an b1 · · · bm is a factorization with λ ∈ U(R), ai , bi ∈ R# , then we will call a =
λa1 a2 · · · an ⌈b1 b2 · · · bm ⌉ a U-factorization of a if (1) ai (b1 · · · bm ) = (b1 · · · bm ) for all 1 ≤ i ≤
n and (2) bj (b1 · · · bbj · · · bm ) 6= (b1 · · · bbj · · · bm ) for 1 ≤ j ≤ m where bbj means bj is omitted
from the product. Here (b1 · · · bm ) is the principal ideal generated by b1 · · · bm . The bi ’s in
this particular U-factorization above will be referred to as essential divisors. The ai ’s in this
particular U-factorization above will be referred to as inessential divisors. A U-factorization
is said to be trivial if there is only one essential divisor.
Note: we have added a single unit factor in front with the inessential divisors which was
not in M. Axtell’s original paper. This is added for consistency with the τ -factorization
definitions and it is evident that a unit is always inessential. We allow only one unit factor,
so it will not affect any of the finite factorization properties.
Remark. If a = λa1 · · · an ⌈b1 · · · bm ⌉ is a U-factorization, then for any 1 ≤ i0 ≤ m, we have
(a) = (b1 · · · bm ) ( (b1 · · · bc
i0 · · · bm ). This is immediate from the definition of U-factorization.
Date: November 6, 2017.
Key words and phrases. factorization, zero-divisors, commutative rings.
1
2
CHRISTOPHER PARK MOONEY
In [9], M. Axtell defines a non-unit a and b to be associate if (a) = (b) and a non-zero nonunit a said to be irreducible if a = bc implies a is associate to b or c. R is commutative ring
R to be U-atomic if every non-zero non-unit has a U-factorization in which every essential
divisor is irreducible. R is said to be a U-finite factorization ring if every non-zero non-unit
has a finite number of distinct U-factorizations. R is said to be a U-bounded factorization
ring if every non-zero non-unit has a bound on the number of essential divisors in any Ufactorization. R is said to be a U-weak finite factorization ring if every non-zero non-unit
has a finite number of non-associate essential divisors. R is said to be a U-atomic idf-ring if
every non-zero non-unit has a finite number of non-associate irreducible essential divisors. R
is said to be a U-half factorization ring if R is U-atomic and every U-atomic factorization has
the same number of irreducible essential divisors. R is said to be a U-unique factorization
ring if it is a U-HFR and in addition each U-atomic factorization can be arranged so the
essential divisors correspond up to associate. In [10, Theorem 2.1], it is shown this definition
of U-UFR is equivalent to the one given by C.R. Fletcher in [11, 12].
In the second section, we begin with some preliminary definitions and results about τ factorization in integral domains as well as factorization in rings with zero-divisors. In the
third section, we state definitions for τ -U-irreducible elements and τ -U-finite factorization
properties. We also prove some preliminary results using these new definitions. In the
fourth section, we demonstrate the relationship between rings satisfying the various τ -U finite
factorization properties. Furthermore, we compare these properties with the rings satisfying
τ -finite factorization properties studied in [13]. In the final section, we investigate direct
products of rings. We introduce a relation τ× which carries many τ -U-finite factorization
properties of the component rings through the direct product.
2. Preliminary Definitions and Results
As in [6], we let a ∼ b if (a) = (b), a ≈ b if there exists λ ∈ U(R) such that a = λb, and
a∼
= b if (1) a ∼ b and (2) a = b = 0 or if a = rb for some r ∈ R then r ∈ U(R). We say a
and b are associates (resp. strong associates, very strong associates) if a ∼ b (resp. a ≈ b,
a∼
= b). As in [1], a ring R is said to be strongly associate (resp. very strongly associate) ring
if for any a, b ∈ R, a ∼ b implies a ≈ b (resp. a ∼
= b).
Let τ be a relation on R# , that is, τ ⊆ R# × R# . We will always assume further that τ is
symmetric. Let a be a non-unit, ai ∈ R# and λ ∈ U(R), then a = λa1 · · · an is said to be a
τ -factorization if ai τ aj for all i 6= j. If n = 1, then this is said to be a trivial τ -factorization.
Each ai is said to be a τ -factor, or that ai τ -divides a, written ai |τ a.
We say that τ is multiplicative (resp. divisive) if for a, b, c ∈ R# (resp. a, b, b′ ∈ R# ), aτ b
and aτ c imply aτ bc (resp. aτ b and b′ | b imply aτ b′ ). We say τ is associate (resp. strongly
associate, very strongly associate) preserving if for a, b, b′ ∈ R# with b ∼ b′ (resp. b ≈ b′ ,
b ∼
= b′ ) aτ b implies aτ b′ . We define a τ -refinement of a τ -factorization λa1 · · · an to be a
factorization of the form
(λλ1 · · · λn ) · b11 · · · b1m1 · b21 · · · b2m2 · · · bn1 · · · bnmn
where ai = λi bi1 · · · bimi is a τ -factorization for each i. This is slightly different from the
original definition in [4] where no unit factor was allowed, and one can see they are equivalent
when τ is associate preserving. We then say that τ is refinable if every τ -refinement of a
τ -factorization is a τ -factorization. We say τ is combinable if whenever λa1 · · · an is a τ factorization, then so is each λa1 · · · ai−1 (ai ai+1 )ai+2 · · · an .
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
3
We now summarize several of the definitions given in [13]. Let a ∈ R be a non-unit.
Then a is said to be τ -irreducible or τ -atomic if for any τ -factorization a = λa1 · · · an , we
have a ∼ ai for some i. We will say a is τ -strongly irreducible or τ -strongly atomic if for
any τ -factorization a = λa1 · · · an , we have a ≈ ai for some ai . We will say that a is τ -mirreducible or τ -m-atomic if for any τ -factorization a = λa1 · · · an , we have a ∼ ai for all i.
Note: the m is for “maximal” since such an a is maximal among principal ideals generated
by elements which occur as τ -factors of a. We will say that a is τ -very strongly irreducible or
τ -very strongly atomic if a ∼
= a and a has no non-trivial τ -factorizations. See [13] for more
equivalent definitions of these various forms of τ -irreducibility.
From [13, Theorem 3.9], we have the following relations where † represents the implication
requires a strongly associate ring:
+3 τ -strongly irred.
+3 τ -irred.
τ -very strongly irred.
KS
❯❯❯❯
♠♠ 2:
♠
♠
❯❯❯❯❯
♠♠
†
❯❯❯❯❯
♠♠♠
❯ &.
♠♠♠
τ -m-irred.
3. τ -U-irreducible elements
A τ -U-factorization of a non-unit a ∈ R is a U-factorization a = λa1 a2 · · · an ⌈b1 b2 · · · bm ⌉
for which λa1 · · · an b1 · · · bm is also a τ -factorization.
Given a symmetric relation τ on R# , we say R is τ -U-refinable if for every τ -U-factorization
of any non-unit a ∈ U(R), a = λa1 · · · an ⌈b1 · · · bm ⌉, any τ -U-factorization of an essential
divisors, bi = λ′ c1 · · · cn′ ⌈d1 · · · dm′ ⌉ satisfies
a = λλ′ a1 · · · an c1 · · · cn′ ⌈b1 · · · bi−1 d1 · · · dm′ bi+1 · · · bm ⌉
is a τ -U-factorization.
Example 3.1. Let R = Z/20Z, and let τ = R# × R# .
Certainly 0 = ⌈10 · 10⌉ is a τ -U-factorization. But 10 = ⌈2 · 5⌉ is a τ -U-factorization;
however, 0 = ⌈2 · 5 · 2 · 6⌉ is not a U-factorization since 5 becomes inessential after a τ U-refinement. It will sometimes be important to ensure the essential divisors of a τ -Urefinement of a τ -U-factorization’s essential divisors remain essential. We will see that in
a présimplifiable ring, there are no inessential divisors, so for τ -refinable, R will be τ -Urefinable.
As stated in [9], the primary benefit of looking at U-factorizations is the elimination of
troublesome idempotent elements that ruin many of the finite factorization properties. For
instance, even Z6 is not a BFR (a ring in which every non-unit has a bound on the number of
non-unit factors in any factorization) because we have 3 = 32 . Thus, 3 is an idempotent, so
3 = 3n for all n ≥ 1 which yields arbitrarily long factorizations. When we use U-factorization,
we see any of these factorizations can be rearranged to 3 = 3n−1 ⌈3⌉, which has only one
essential divisor.
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}. Let a
be a non-unit. If a = λa1 a2 · · · an ⌈b1 b2 · · · bm ⌉ is a τ -U-factorization, then this factorization
is said to be a τ -U-α-factorization if it is a τ -U-factorization and the essential divisors bi are
τ -α for 1 ≤ i ≤ m.
One must be somewhat more careful with U-factorizations as there is a loss of uniqueness
4
CHRISTOPHER PARK MOONEY
in the factorizations. For instance, if we let R = Z6 × Z8 , then we can factor (3, 4) as
(3, 1) ⌈(3, 3)(1, 4)⌉ or (3, 3) ⌈(3, 1)(1, 4)⌉. On the bright side, we have [8, Proposition 4.1].
Theorem 3.2. Every factorization can be rearranged into a U-factorization.
Corollary 3.3. Let R be a commutative ring with 1 and τ a symmetric relation on R# .
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}. For every
τ -α factorization of a non-unit a ∈ R, a = λa1 · · · an , we can rearrange this factorization
into a τ -U-α-factorization.
Proof. Let a = λa1 · · · an be a τ -α-factorization. By Theorem 3.2 we can rearrange this to
form a U-factorization. This remains a τ -factorization since τ is assumed to be symmetric.
Lastly each ai is τ -α, so the essential divisors are τ -α.
This leads us to another equivalent definition of τ -irreducible.
Theorem 3.4. Let a ∈ R be a non-unit. Then a is τ -irreducible if and only if any τ -Ufactorization of a has only one essential divisor.
Proof. (⇒) Let a be τ -irreducible. Let a = λa1 · · · an ⌈b1 · · · bm ⌉ be a τ -U-factorization.
Suppose m ≥ 2, then a = λa1 · · · an b1 · · · bm is a τ -factorization implies a ∼ ai0 for some
1 ≤ i0 ≤ n or a ∼ bi0 for some 1 ≤ i0 ≤ m. But then either
or
(a) = (a1 · · · an b1 · · · bm ) ( (a1 · · · an bb1 b2 · · · bm ) ⊆ (ai0 ) = (a)
d
c
(a) = (a1 · · · an b1 · · · bm ) = (b1 · · · bm ) ( (bb1 · · · bd
i0 −1 · bi0 · bi0 +1 · · · bm ) ⊆ (bi0 ) = (a)
a contradiction.
(⇐) Suppose a = λa1 · · · an . Then this can be rearranged into a U-factorization, and
hence a τ -U-factorization. By hypothesis, there can only be one essential divisor. Suppose
it is an . We have a = λa1 · · · an−1 ⌈an ⌉ is a τ -U-factorization and a ∼ an as desired.
We now define the finite factorization properties using the τ -U-factorization approach.
Let α ∈ { irreducible, strongly irreducible, m-irreducible, very strongly irreducible } and
let β ∈ {associate, strongly associate, very strongly associate }. R is said to be τ -U-α if
for all non-units a ∈ R, there is a τ -U-α-factorization of a. R is said to satisfy τ -U-ACCP
(ascending chain condition on principal ideals) if every properly ascending chain of principal
ideals (a1 ) ( (a2 ) ( · · · such that ai+1 is an essential divisor in some τ -U-factorization of ai ,
for each i terminates after finitely many principal ideals. R is said to be a τ -U-BFR if for all
non-units a ∈ R, there is a bound on the number of essential divisors in any τ -U-factorization
of a.
R is said to be a τ -U-β-FFR if for all non-units a ∈ R, there are only finitely many
τ -U-factorizations up to rearrangement of the essential divisors and β. R is said to be a τ U-β-WFFR if for all non-units a ∈ R, there are only finitely many essential divisors among
all τ -U-factorizations of a up to β. R is said to be a τ -U-α-β-divisor finite (df ) ring if
for all non-units a ∈ R, there are only finitely many essential τ -α divisors up to β in the
τ -U-factorizations of a.
R is said to be a τ -U-α-HFR if R is τ -U-α and for all non-units a ∈ R, the number of
essential divisors in any τ -U-α-factorization of a is the same. R is said to be a τ -U-α-βUFR if R is a τ -U-α-HFR and the essential divisors of any two τ -U-α-factorizations can be
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
5
rearranged to match up to β.
R is said to be présimplifiable if for every x ∈ R, x = xy implies x = 0 or y ∈ U(R).
This is a condition which has been well studied and is satisfied by any domain or local ring.
We introduce two slight modifications of this. R is said to be τ -présimplifiable if for every
x ∈ R, the only τ -factorizations of x which contain x as a τ -factor are of the form x = λx
for a unit λ. R is said to be τ -U-présimplifiable if for every non-zero non-unit x ∈ R, all
τ -U-factorizations have no non-unit inessential divisors.
Theorem 3.5. Let R be a commutative ring with 1 and let τ be a symmetric relation on
R# . We have the following.
(1) If R is présimplifiable, then R is τ -U-présimplifiable.
(2) If R is τ -U-présimplifiable, then R is R is τ -présimplifiable.
That is présimplifiable ⇒ τ -U-présimplifiable ⇒ τ -présimplifiable. If τ = R# × R# , then all
are equivalent.
Proof. (1) Let R be présimplifiable, and x ∈ R# . Suppose x = λa1 · · · an ⌈b1 · · · bm ⌉ is a
τ -U-factorization. Then (x) = (b1 · · · bm ). R présimplifiable implies that all the associate
relations coincide, so in fact x ∼
= b1 · · · bm implies that λa1 · · · an ∈ U(R) and hence all
inessential divisors are units.
(2) Let R be τ -U-présimplifiable, and x ∈ R such that x = λxa1 · · · an is a τ -factorization.
We claim that x = λa1 · · · an ⌈x⌉ is a τ -U-factorization. For any 1 ≤ i ≤ n, x | ai x and
(ai x)(λa1 · · · abi · · · an ) = x shows ai x | x, proving the claim. This implies λa1 · · · an ∈ U(R)
as desired.
Let τ = R# × R# and suppose R is τ -présimplifiable. Suppose x = xy, for x 6= 0, we show
y ∈ U(R). If x ∈ U(R), then multiplying through by x−1 yields 1 = x−1 x = x−1 xy = y and
y ∈ U(R) as desired. We may now assume x ∈ R# . If y = 0, then x = 0, a contradiction.
If y ∈ U(R) we are already done, so we may assume y ∈ R# . Thus xτ y, and x = xy is a
τ -factorization, so y ∈ U(R) as desired.
4. τ -U-finite factorization relations
We now would like to show the relationship between rings with various τ -U-α-finite factorization properties as well as compare these rings with the τ -α-finite factorization properties
of [13].
Theorem 4.1. Let R be a commutative ring with 1 and let τ be a symmetric relation on
R# . Consider the following statements.
(1) R is a τ -BFR.
(2) R is τ -présimplifiable and for every non-unit a1 ∈ R, there is a fixed bound on the length
of chains of principal ideals (ai ) ascending from a1 such that at each stage ai+1 |τ ai .
(3) R is τ -présimplifiable and a τ -U-BFR.
(4) For every non-unit a ∈ R , there are natural numbers N1 (a) and N2 (a) such that if
a = λa1 · · · an ⌈b1 · · · bm ⌉ is a τ -U-factorization, then n ≤ N1 (a) and m ≤ N2 (a).
Then (4) ⇒ (1) and (2) ⇒ (3). For τ refinable, (1) ⇒ (2) and for R τ -U-présimplifiable,
(3) ⇒ (4). Thus all are equivalent if R is τ -U-présimplifiable and τ is refinable.
Let ⋆ represent τ being refinable, and † represent R being τ -U-présimplifiable, then the
6
CHRISTOPHER PARK MOONEY
following diagram summarizes the theorem.
(1)
⋆
KS
(4) ks
+3
(2)
†
(3)
Proof. (1) ⇒ (2) Let τ be refinable. Suppose there were a non-trivial τ -factorization x =
λxa1 · · · an with n ≥ 1. Since τ is assumed to be refinable we can continue to replace the
τ -factor x with this factorization.
x = λxa1 · · · an = (λλ)xa1 · · · an a1 · · · an = · · · = (λλλ)xa1 · · · an a1 · · · an a1 · · · an = · · ·
yields an unbounded series of τ -factorizations of increasing length.
Let a1 be a non-unit in R. Suppose N is the bound on the length of any τ -factorization
of a1 . We claim that N satisfies the requirement of (2). Let (a1 ) ( (a2 ) ( · · · be an
ascending chain of principal ideals generated by elements which satisfy ai+1 |τ ai for each i.
Say ai = λi ai+1 ai1 · · · aini for each i. Furthermore, we can assume ni ≥ 1 for each i or else
the containment would not be proper. Then we can write
a1 = λ1 a2 a11 · · · a1n1 = λ1 λ2 a3 a21 · · · a2n2 a11 · · · a1n1 = · · · .
Each remains a τ -factorization since τ is refinable and we have added at least one factor
at each step. If the chain were greater than length N we would contradict R being a τ -BFR.
(2) ⇒ (3) Let a ∈ R be a non-unit. Let N be the bound on the length of any properly ascending chain of principle ideals ascending from a such that ai+1 |τ ai . If a =
λa1 · · · an ⌈b1 · · · bm ⌉ is a τ -U-factorization, then we get an ascending chain with b1 · · · bi−1 |τ
b1 · · · bi for each i:
(a) = (b1 · · · bm ) ( (b1 · · · bm−1 ) ( (b1 · · · bm−2 ) ( · · · ( (b1 b2 ) ( (b1 ).
Hence, m ≤ N and we have found a bound on the number of essential divisors in any τ -Ufactorization of a, making R a τ -U-BFR.
(3) ⇒ (4) Let a ∈ R be a non-unit. Let Ne (a) be the bound on the number of essential
divisors in any τ -U-factorization of a. Since R is τ -U-présimplifiable, there are no inessential
τ -U-divisors of a. We can set N1 (a) = 0, and N2 (a) = Ne (a) and see that this satisfies the
requirements of the theorem.
(4) ⇒ (1) Let a ∈ R be a non-unit. Then any τ -factorization
a = λa1 · · · an can be
rearranged into a τ -U-factorization, say a = λas1 · · · asi asi+1 · · · asn . But then n = i + (n −
i) ≤ N1 (a) + N2 (a). Hence the length of any τ -factorization must be less than N1 (a) + N2 (a)
proving R is a τ -BFR as desired.
The way we have defined our finite factorization properties on only the essential divisors
causes a slight problem. Given a τ -U-factorization a = λa1 · · · an ⌈b1 · · · bm ⌉, we only know
that a ∼ b1 · · · bm . This may no longer be a τ -factorization of a, but rather only some
associate of a. This is easily remedied by insisting that our rings are strongly associate.
Lemma 4.2. Let R be a strongly associate ring with τ a symmetric relation on R# , and let
α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}. Let a ∈ R,
a non-unit. If a = λa1 a2 · · · an ⌈b1 b2 · · · bm ⌉ is a τ -U-α-factorization, then there is a unit
µ ∈ U(R) such that a = µb1 · · · bm is a τ -α-factorization.
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
7
Proof. Let a = λa1 a2 · · · an ⌈b1 b2 · · · bm ⌉ be a τ -U-α-factorization. By definition, (a) =
(b1 · · · bm ), and R strongly associate implies that a ≈ b1 · · · bm . Let µ ∈ U(R) be such
that a = µb1 · · · bm . We still have bi τ bj for all i 6= j, and bi is τ -α for every i. Hence
a = µb1 · · · bm is the desired τ -factorization, proving the lemma.
Theorem 4.3. Let R be a commutative ring with 1, and let τ be a symmetric relation on
R# . Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}, and
β ∈ {associate, strongly associate, very strongly associate }. We have the following.
(1) If R is τ -α, then R is τ -U-α.
(2) If R satisfies τ -ACCP, then R satisfies τ -U-ACCP.
(3) If R is a τ -BFR, then R is a τ -U-BFR.
(4) If R is a τ -β-FFR, then R is a τ -U-β-FFR.
(5) Let R be a τ -β-WFFR, then R is a τ -U-β-WFFR.
(6) Let R be a τ -α-β-divisor finite ring, then R is τ -U-α-β-divisor finite ring.
(7) Let R be a strongly associate τ -α-HFR (resp. τ -α-β-UFR), then R is τ -U-α-HFR (resp.
τ -U-α-β-UFR).
Proof. (1) This is immediate from Corollary 3.3.
(2) Suppose there were a infinite properly ascending chain of principal ideals (a1 ) ( (a2 ) (
· · · such that ai+1 is an essential divisor in some τ -U-factorization of ai , for each i. Every
essential τ -U-divisor is certainly a τ -divisor. This would contradict the fact that R satisfies
τ -ACCP.
(3) We suppose that there is a non-unit a ∈ R with τ -U-factorizations having arbitrarily
large numbers of essential τ -U-divisors. Each is certainly a τ -factorization, having at least
as many τ -factors as there are essential τ -divisors, so this would contradict the hypothesis.
(4) Every τ -U-factorization is certainly among the τ -factorizations. If the latter is finite,
then so is the former.
(5) For any given non-unit a ∈ R, every essential τ -U-divisor of a is certainly a τ -factor
of a which has only finitely many up to β. Hence there can be only finitely many essential
τ -U-factors up to β.
(6) Let a ∈ R be a non-unit. Every essential τ -U-α-divisor of a is a τ -α-factor of a.
There are only finitely many τ -α-divisors up to β, so then there can be only finitely many
τ -U-α-divisors of a up to β.
(7) We have already seen that R being τ -α implies R is τ -U-α. Let a ∈ R be a non-unit.
We suppose for a moment there are two τ -α-U-factorizations:
a = λa1 · · · an ⌈b1 · · · bm ⌉ = λ′ a′1 · · · a′n′ ⌈b′1 · · · b′m′ ⌉
such that m 6= m′ (resp. m 6= m′ or there is no rearrangement such that bi and b′i are β
for each i). Lemma 4.2 implies ∃µ, µ′ ∈ U(R) with a = µb1 · · · bm = µ′ b′1 · · · b′m′ are two
τ -α-factorizations of a, so m = m′ (resp. m = m′ and there is a rearrangement so that bi
and b′i are β for each 1 ≤ i ≤ m), a contradiction, proving R is indeed a τ -U-α-HFR (resp.
-β-UFR) as desired.
Theorem 4.4. Let R be a commutative ring with 1 and τ a symmetric relation on R# .
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}, and let
β ∈ {associate, strongly associate, very strongly associate}.
(1) If R is a τ -U-α-β-UFR, then R is a τ -α-U-HFR.
(2) If R is τ -U-refinable and R is a τ -U-α-β-UFR, then R is a τ -U-β-FFR.
8
(3)
(4)
(5)
(6)
(7)
(8)
CHRISTOPHER PARK MOONEY
If
If
If
If
If
If
R
R
R
R
R
R
is
is
is
is
is
is
τ -U-refinable and R is a τ -U-α-HFR, then R is a τ -U-BFR.
a τ -U-β-FFR, then R is a τ -U-BFR.
a τ -U-β-FFR, then R is a τ -U-β-WFFR.
a τ -U-β-WFFR, then R is a τ -U-α-β-divisor finite ring.
τ -U-refinable and R is a τ -U-α-BFR, then R satisfies τ -U-ACCP.
τ -U-refinable and R satisfies τ -U-ACCP, then R is τ -U-α.
Proof. (1) This is immediate from definitions.
(2) Let a ∈ R be a non-unit. Let a = λa1 · · · an ⌈b1 · · · bm ⌉ be the unique τ -α-U-factorization
up to rearrangement and β. Given any other τ -U-factorization, we can τ -U-refine each essential τ -U-divisor into a τ -U-α-factorization of a. There is a rearrangement of the essential
divisors to match up to β with bi for each 1 ≤ i ≤ m. Thus the essential divisors in any τ -Ufactorization come from some combination of products of β of the m τ -U-α essential factors
in our original factorization. Hence there are at most 2m possible distinct τ -U-factorizations
up to β, making this a τ -U-β-FFR as desired.
(3) For a given non-unit a ∈ R, the number of essential divisors in any τ -U-α-factorization
is the same, say N. We claim this is a bound on the number of essential divisors of any
τ -U-factorization. Suppose there were a τ -U-factorization a = λa1 · · · an ⌈b1 · · · bm ⌉ with
m > N. For every i, bi has a τ -U-α-factorization with at least one essential divisor. Since R
is τ -U-refinable, we can τ -U-refine the factorization yielding a τ -U-α-factorization of a with
at least m τ -U-α essential factors. This contradicts the assumption that R is a τ -U-α-HFR.
(4) Let R be a τ -U-β-FFR. Let a ∈ R be a non-unit. There are only finitely many τ -Ufactorizations of a up to rearrangement and β of the essential divisors. We can simply take
the maximum of the number of essential divisors among all of these factorizations. This is
an upper bound for the number of essential divisors in any τ -U-factorization.
(5) Let R be a τ -U-β-FFR, then for any non-unit a ∈ R. Let S be the collection of essential
divisors in the finite number of representative τ -U-factorizations of a up to β. This gives us
a finite collection of elements up to β. Every essential divisor up to β in a τ -U-factorization
of a must be among these, so this collection is finite as desired.
(6) If every non-unit a ∈ R has a finite number of proper essential τ -U divisors, then
certainly there are a finite number of essential τ -α-U-divisors.
(7) Suppose R is a τ -U-BFR, but (a1 ) ( (a2 ) ( · · · is a properly ascending chain of
principal ideals such that ai+1 is an essential factor in some τ -U-factorization of ai , say
ai = λi ai1 · · · aini ⌈ai+1 bi1 · · · bimi ⌉
for each i. Furthermore, mi ≥ 1, for each i otherwise we would have (ai+1 ) = (ai ) contrary
to our assumption that our chain is properly increasing. Our assumption that R is τ -U
refinable allows us to factor a1 as follows:
a1 = λ1 a11 · · · a1n1 ⌈a2 b11 · · · b1m1 ⌉ =
λ1 λ2 a11 · · · a1n1 a21 · · · a2n2 ⌈a3 b21 · · · b2m2 b11 · · · b1m1 ⌉
and so on. At each iteration i we have at least i + 1 essential factors in our τ -U-factorization.
This contradicts the assumption that a1 should have a bound on the number of essential
divisors in any τ -U-factorization.
(8) Let a1 ∈ R be a non-unit. If a1 is τ -U-α we are already done, so there must be a
non-trivial τ -U factorization of a1 , say:
a1 = λ1 a11 · · · a1n1 ⌈a2 b11 · · · b1m1 ⌉ .
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
9
Now if all of the essential divisors are τ -U-α we are done as we have found a τ -U-αfactorization. After rearranging if necessary, we suppose that a2 is not τ -U-α. Therefore a2
has a non-trivial τ -U-factorization, say:
a2 = λ2 a21 · · · a2n1 ⌈a3 b21 · · · b2m2 ⌉ .
Because R is τ -U-refinable, this gives us a τ -U-factorization:
a1 = λ1 λ2 a11 · · · a1n1 a21 · · · a2n2 ⌈a3 b21 · · · b2m2 b11 · · · b1m1 ⌉
which cannot be τ -U-α or else we would be done. We can continue in this fashion and get
an ascending chain of principal ideals
(a1 ) ⊆ (a2 ) ⊆ · · ·
such that ai+1 is an essential τ -U-divisor of ai for each i.
Claim: This chain must be properly ascending. Suppose (ai ) = (ai+1 ) for some i. When
we look at ai = λi ai1 · · · aini ⌈ai+1 bi1 · · · bimi ⌉, we see that (ai ) = (ai+1 bi1 · · · bimi ). But then
we could remove any of the bij for any 1 ≤ j ≤ mi and still have (ai ) = (ai+1 bi1 · · · bc
ij · · · bimi )
contradicting the fact that the factorization was a τ -U-factorization since bij is inessential.
We certainly have (ai ) ⊆ (ai+1 bi1 · · · bc
ij · · · bimi ). To see the other containment holds,
(ai ) = (ai+1 ) ⇒ ai+1 = ai r for some r ∈ R, and we can simply multiply by bi1 · · · bc
ij · · · bimi
on both sides to see that
c
ai+1 bi1 · · · bc
ij · · · bim = ai (rbi1 · · · bij · · · bim )
i
i
showing the other containment. Proving the claim.
This is a contradiction to the fact that R satisfies τ -U-ACCP, proving we must in finitely
many steps arrive at a τ -U-α-factorization of a1 , proving R is indeed τ -U-α as desired.
The following diagram summarizes our results from the Theorems 4.3 and 4.4 where ⋆
represents R being strongly associate, and † represents R is τ -U-refinable:
τ -α-β-UFR
ks
τ ❧-U-α-HFR
◗◗
19
⋆
τ -α-HFR
◗◗◗◗
❧❧❧
❧❧❧
◗◗◗†◗
❧
❧
❧
⋆
◗◗◗◗
❧❧❧
❧
❧
❧
$,
†
+3 τ -U-BFR
+3 τ -U-β-FFR
τ -U-α-β-UFR
KS
]e ❈❈❈
❈❈❈❈
❈❈❈❈
❈❈❈❈
❈❈
+3 τ -U-β-WFFR ❈❈❈❈❈❈❈
τ -β-WFFR
❈❈❈❈❈ τ -BFR
❈❈❈❈
❈❈❈❈
❈❈❈
+3 τ -U-α-β df ring
τ -α-β df ring
τ -β-FFR
†
+3
τ -U-ACCP
KS
τ −ACCP
†
+3
τ -U-α
KS
τ -α
We have left off the relations which were proven in [13, Theorem 4.1], and focused instead
on the rings satisfying the U-finite factorization properties. Examples given in [9, 10, 4, 2]
show that arrows can neither be reversed nor added to the diagram with a few exceptions.
Question 4.5. Does U-atomic imply atomic?
D.D. Anderson and S. Valdez-Leon show in [6, Theorem 3.13] that if R has a finite number
of non-associate irreducibles, then U-atomic and atomic are equivalent. This remains open
10
CHRISTOPHER PARK MOONEY
in general.
Question 4.6. Does U-ACCP imply ACCP?
We can modify M. Axtell’s proof of [9, Theorem 2.9] to add a partial converse to Theorem
4.4 (5) if τ is combinable and associate preserving. The idea is the same, but slight adjustments are required to adapt it to τ -factorizations and to allow uniqueness up to any type of
associate.
Theorem 4.7. Let β ∈ {associate, strongly associate, very strongly associate}. Let R be a
commutative ring with 1 and let τ be a symmetric relation on R# which is both combinable
and associate preserving. R is a τ -U-β-FFR if and only if R is a τ -U-β-WFFR.
Proof. (⇒) was already shown, so we need only prove the converse. (⇐) Suppose R is not
a τ -U-β-FFR. Let a ∈ R be a non-unit which has infinitely many τ -U-factorizations up to
β. Let b1 , b2 , . . . , bm be a complete list of essential τ -U-divisors of a up to β. Let
a = a1 · · · an ⌈c1 · · · ck ⌉ = a′1 · · · a′n′ ⌈d1 · · · dn ⌉
be two τ -U-factorizations of a and assume we have re-ordered the essential divisors in
both factorizations above so that the β of b1 appear first, followed by β of b2 , etc. Let
A = h(c1 ), (c2 ), . . . , (ck )i and B = h(d1 ), (d2 ), . . . , (dn )i be sequences of ideals. We call the
factorizations comparable if A is a subsequence of B or vice versa.
Suppose A is a proper subsequence of B
B = h(d1 ), . . . , (di1 ) = (c1 ), . . . , (di2 ) = (c2 ), . . . , (dik ) = (ck ), . . . , (dn )i
with n > k. Because τ is combinable and symmetric,
l
m
c
c
a = a′1 · · · a′n′ di1 di2 · · · dik (d1 · · · dc
d
·
·
·
d
·
·
·
d
)
i1 i2
ik
n
remains a τ -factorizations and [9, Lemma 1.3] ensures that this remains a U-factorization.
This yields
c
c
(a) = (d1 · · · dc
i1 di2 · · · dik · · · dn )(di1 di2 · · · dik ) = (d1 · · · dn ) = (c1 · · · ck )
= (c1 ) · · · (ck ) = (di1 ) · · · (dik ) = (di1 · · · dik ).
c
c
But then, (d1 · · · dc
i1 di2 · · · dik · · · dn ) cannot be an essential divisor, a contradiction, unless
n = k.
If n = k, then the sequences of ideals are identical, and we seek to prove this means the τ U-factorizations are the same up to β. It is certainly true for β = associate as demonstrated
in [9, Theorem 2.9]. So we have a pairing of the ci and di such that ci ∼ bj ∼ di for one of
the essential τ -U-divisors bj . We know further that ci and bj (resp. di and bj ) are β since R
is by assumption a τ -U-β-WFFR.
It is well established that β is transitive, so we can conclude that this same pairing
demonstrates that ci and di are β, not just associate. Thus the number of distinct τ -Ufactorizations up to β is less than or equal to the number of non-comparable finite sequences
of elements from the set {(b1 ), (b2 ), . . . , (bm )}.
From here we direct the reader to the proof of the second claim in [9, Theorem 2.9] where
it is shown that this set is finite.
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
11
5. Direct Products
For each i, 1 ≤ i ≤ N, let Ri be commutative rings with τi a symmetric relation on Ri# .
We define a relation τ× on R = R1 × · · · × RN which preserves many of the theorems about
direct products from [8] for τ -factorizations. Let (ai ), (bi ) ∈ R# , then (ai )τ× (bi ) if and only
if whenever ai and bi are both non-units in Ri , then ai τi bi .
For convenience we will adopt the following notation: Suppose x ∈ Ri , then x(i) =
(1R1 , · · · , 1Ri−1 , x, 1Ri+1 , · · · 1RN ). so x appears in the ith coordinate, and all other entries
(1) (2)
(n)
are the identity. Thus for any (ai ) ∈ R, we have (ai ) = a1 a2 · · · an is a τ× -factorization.
We will always move any τ× -factors which may become units in this process to the front and
collect them there.
Lemma 5.1. Let R = R1 × · · · × RN for N ∈ N. Then (ai ) ∼ (bi ) (resp. (ai ) ≈ (bi )) if and
only if ai ∼ bi (resp. ai ≈ bi ) for every i. Furthermore, (ai ) ∼
= bi for all i,
= (bi ) implies ai ∼
∼
∼
and for ai , bi all non-zero, ai = bi for all i ⇒ (ai ) = (bi ).
Proof. See [6, Theorem 2.15].
Example 5.2. If ai0 = 0 for even one index 1 ≤ i0 ≤ N, then ai ∼
= bi for all i need not
imply (ai ) ∼
= (bi ).
Consider the ring R = Z × Z, with τi = Z# × Z# for i = 1, 2, the usual factorization. We
6= (0, 1).
have 1 ∼
= 0 since Z is a domain; however (0, 1) = (0, 1)(0, 1) shows (0, 1) ∼
= 1 and 0 ∼
Lemma 5.3. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# for
each i. Let α ∈ { irreducible, strongly irreducible, m-irreducible, very strongly irreducible}.
If (ai ) ∈ R is τ -α, then precisely one coordinate is not a unit.
Proof. Let a = (ai ) ∈ R be a non-unit which is τ× -α. Certainly not all coordinates can
be units, or else a ∈ U(R). Suppose for a moment there were at least two coordinates for
which ai is not a unit in Ri . After reordering, we may assume a1 and a2 are not units. Then
(1)
a = a1 (1R1 , a2 , · · · , aN ) is a τ× -factorization. But a is not even associate to either τ× -factor,
a contradiction.
Theorem 5.4. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# for
each i.
(1) A non-unit (ai ) ∈ R is τ× -atomic (resp. strongly atomic) if and only if ai0 is τi0 -atomic
(resp. strongly atomic) for some 1 ≤ i0 ≤ n and ai ∈ U(Ri ) for all i 6= i0 .
(2) A non-unit (ai ) ∈ R is τ× -m-atomic if and only if ai0 is τi0 -m-atomic for some 1 ≤ i0 ≤ n
and ai ∈ U(Ri ) for all i 6= i0 .
(3) A non-unit (ai ) ∈ R is τ× -very strongly atomic if and only if ai0 is τi0 -very strongly
atomic and non-zero for some 1 ≤ i0 ≤ n and ai ∈ U(Ri ) for all i 6= i0 .
Proof. (1) (⇒) Let a = (ai ) ∈ R be a non-unit which is τ× -atomic (resp. strongly atomic).
By Lemma 5.3, there is only one non-unit coordinate. Suppose after reordering if necessary
that a1 is the non-unit. If a1 were not τ1 -atomic (resp. strongly atomic), then there is a
τ1 -factorization, λ11 a11 a12 · · · a1k for which a1 6∼ a1j (resp. a1 6≈ a1j ) for any 1 ≤ j ≤ k. But
then
(1) (1)
(1)
(ai ) = (λ11 , a2 , . . . , an )a11 a12 · · · a1k
(1)
(1)
is a τ× -factorization. Furthermore, by Lemma 5.1 (ai ) 6∼ a1j (resp. (ai ) 6∼ a1j ) for all
1 ≤ j ≤ k. This would contradict the assumption that a was τ× -atomic (resp. strongly
12
CHRISTOPHER PARK MOONEY
atomic).
(⇐) Let a1 ∈ R1 , a non-unit with a1 being τ1 -atomic (resp. strongly atomic). Let
µi ∈ U(Ri ) for 2 ≤ i ≤ N. We show a = (a1 , µ2 , · · · µN ) is τ× -atomic (resp. strongly
atomic). Suppose a = (λ1 , . . . , λN )(a11 , . . . , a1N ) · · · (ak1 , . . . , akN ) is a τ× -factorization of a.
We first note aij ∈ U(Rj ) for all j ≥ 2. Furthermore, this means ai1 is not a unit in R1
for 1 ≤ i ≤ k, otherwise we would have units as factors in a τ× factorization. This means
a1 = λ1 a11 · · · ak1 is a τ1 factorization of a τ1 -atomic (resp. strongly atomic) element. Thus,
we must have a1 ∼ aj1 (resp. a1 ≈ aj1 ) for some 1 ≤ j ≤ k. Hence by Lemma 5.1, we have
a ∼ (aj1 , . . . , ajN ) (resp. a ≈ (aj1 , . . . , ajN ) for some 1 ≤ j ≤ k and a is τ× atomic (resp.
strongly atomic) as desired.
(2) (⇒) Let a = (ai ) ∈ R be a non-unit which is τ× -m-atomic. By Lemma 5.3, there is
only one non-unit coordinate, say a1 after reordering if necessary. Let a1 = λ11 a11 a12 · · · a1k
be a τ1 factorization for which a1 6∼ a1j0 for at least one 1 ≤ j0 ≤ k. But then
(1) (1)
(1)
(ai ) = (λ11 , a2 , . . . , an )a11 a12 · · · a1k
(1)
is a τ× -factorization of a for which (by Lemma 5.1) a = (ai ) 6∼ a1j . This contradicts the
0
hypothesis that a is τ× -m-atomic.
(⇐) Let a1 ∈ R1 , a non-unit with a1 being τ1 -m-atomic. Let µi ∈ U(Ri ) for 2 ≤ i ≤ N.
We show a = (a1 , µ2 , · · · µN ) is τ× -m-atomic. Suppose
a = (λ1 , . . . , λN )(a11 , . . . , a1N ) · · · (ak1 , . . . , akN )
is a τ× -factorization of a. We first note aij ∈ U(Rj ) for all j ≥ 2. As before, this means
a1 = λ1 a11 · · · ak1 is a τ1 factorization of a τ1 -m-atomic element. Hence a1 ∼ aj1 for each
1 ≤ j ≤ k. By Lemma 5.1 we have a ∼ (aj1 , . . . , ajN ) for all 1 ≤ j ≤ k and thus a is
τ× -m-atomic as desired.
(3) (⇒) Let a = (a1 , . . . aN ) be a non-unit which is τ× -very strongly atomic. By Lemma
5.3, we may assume a1 is the non-unit, and aj is a unit for j ≥ 2. We suppose for a moment
that a1 = 01 . But then (0, a2 , . . . aN ) = (0, 1, . . . 1) · (0, a2 , . . . aN ) shows that a ∼
6= a, a
∼
∼
contradiction. Lemma 5.1 shows that if a = a, then ai = ai for each 1 ≤ i ≤ N. Hence, if a1
were not τ1 -very strongly atomic, then there is a τ1 -factorization, λ11 a11 a12 · · · a1k for which
a1 ∼
6= a1j for any 1 ≤ j ≤ k. But then
(1) (1)
(1)
(ai ) = (λ11 , a2 , . . . , an )a11 a12 · · · a1k
is a τ× -factorization. Furthermore, since every coordinate is non-zero, by Lemma 5.1 (ai ) ∼
6=
(1)
a1j for all 1 ≤ j ≤ k. This would contradict the assumption that a was τ× -very strongly
atomic.
(⇐) Let a1 ∈ R1# be τ1 -very strongly atomic. Let µi ∈ U(Ri ) for 2 ≤ i ≤ N. We
show a = (a1 , µ2 , · · · µN ) is τ× -very strongly atomic. We first check a ∼
= a. By definition
of τ1 -very strongly atomic, a1 ∼
= a1 . Certainly as units, we have µi ∼
= µi for each i ≥ 2.
Lastly, all of these are non-zero, so we may apply Lemma 5.1 to see that a ∼
= a. Suppose
a = (λ1 , . . . , λN )(a11 , . . . , a1N ) · · · (ak1 , . . . , akN ) is a τ× -factorization of a. We first note
aij ∈ U(Rj ) for all j ≥ 2. As before, this means a1 = λ1 a11 · · · ak1 is a τ1 factorization of
a τ1 -very strongly atomic element. Hence a1 ∼
= aj1 for some 1 ≤ j ≤ k. By Lemma 5.1 we
have a ∼
(a
,
.
.
.
,
a
)
and
thus
a
is
τ
-very
strongly
atomic as desired.
= j1
jN
×
Lemma 5.5. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# . Let
α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible }. Then we
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
13
have the following.
(1) If a = λa1 · · · aln ⌈b1 · · · bmm⌉ is a τi -U-α-factorization of some non-unit a ∈ Ri , then
(i)
(i)
(i)
(i)
a(i) = λ(i) a1 · · · an b1 · · · bm is a τ× -U-α-factorization.
(2) Conversely, let ai0 ∈ Ri0 be a non-unit and µi ∈ U(Ri ) for all i 6= i0 . Let
(µ1 , µ2 , . . . , µi0 −1 , ai0 , µi0 +1 . . . , µN ) = (λi )(a1i )(a2i ) · · · (ani ) ⌈(b1i )(b2i ) · · · (bmi )⌉
be a τ× -U-α-factorization. Then
is a τi0 -U-α-factorization.
ai0 = λi0 a1i0 · · · ani0 b1i0 · · · bi0
Proof. (1) Let a = λa1 · · · an ⌈b1 · · · bm ⌉lbe a τi -U-α-factorization
of some non-unit a ∈ Ri . It
m
(i)
(i)
(i)
(i)
(i) (i)
is easy to see that a = λ a1 · · · an b1 · · · bm is a τ× -factorization. Furthermore, bj 6= 0
(i)
for all 1 ≤ j ≤ or else it would not be a τi -factorization. Hence by Theorem 5.4 bj is τ× -α
for each 1 ≤ j ≤ m. Thus it suffices to show that we actually have a U-factorization.
Since a = λa1 · · · an ⌈b1 · · · bm ⌉ is a U-factorization, we know ak (b1 · · · bm ) = (b1 · · · bm ) for
all 1 ≤ k ≤ n. In the other coordinates, we have (1Rj ) = (1Rj ) for all j 6= i. Hence, we apply
(i) (i)
(i)
(i)
(i)
Lemma 5.1 and see that this implies that ak (b1 · · · bm ) = (b1 · · · bm ) for all 1 ≤ k ≤ n.
c
(i) (i)
(i)
(i)
Similarly we have bj (b1 · · · bbj · · · bm ) 6= (b1 · · · bbj · · · bm ) which implies bj (b1 · · · bj · · · bm ) 6=
c
(i)
(i)
(i)
(b1 · · · bj · · · bm ), so this is indeed a U-factorization.
(2) Let
(µ1 , µ2 , . . . , µi0 −1 , ai0 , µi0 +1 . . . , µN ) = (λi )(a1i )(a2i ) · · · (ani ) ⌈(b1i )(b2i ) · · · (bmi )⌉
be a τ× -U-α-factorization. We note that aji ∈ U(Ri ) for all i 6= i0 and all 1 ≤ j ≤ n and
bji ∈ U(Ri ) for all i 6= i0 and all 1 ≤ j ≤ m since they divide the unit µi . Next, every
coordinate in the i0 place must be a non-unit in Ri0 or else this factor would be a unit in R
and therefore could not occur as a factor in a τ× -factorization. This tells us that
ai0 = λi0 a1i0 · · · ani0 b1i0 · · · bi0
is a τi0 -factorization. Furthermore, (bki ) is assumed to be τ× -α for all 1 ≤ k ≤ m, and the
other coordinates are units, so bki0 is τi0 -α for all 1 ≤ k ≤ m by Theorem 5.4. Again, we
need only show that
ai0 = λi0 a1i0 a2i0 · · · ani0 b1i0 b2i0 · · · bmi0
is a U-factorization. Since all the coordinates other than i0 are units, we simply apply Lemma
5.1 and see that we indeed maintain a U-factorization.
Theorem 5.6. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# . Let
α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible}. Then R is
τ× -U-α if and only if Ri is τi -U-α for each 1 ≤ i ≤ N.
Proof. (⇒) Let a ∈ Ri0 be a non-unit. Then a(i0 ) is a non-unit in R and therefore has a τ× U-α-factorization. Furthermore, the only possible non-unit factors in this factorization must
occur in the ith
0 coordinate. Thus as in Lemma 5.5 (2), we have found a τi0 -U-α-factorization
of a by taking the product of the ith
0 entries. This shows Ri0 is τi0 -U-α as desired.
14
CHRISTOPHER PARK MOONEY
(⇐) Let a = (ai ) ∈ R be a non-unit. For each non-unit ai ∈ Ri , there is a τi -U-αfactorization of ai , say
ai = λi ai1 · · · aini bi1 · · · bimi .
(i)
If ai ∈ U(Ri ), then ai ∈ U(R) and we can simply collect these unit factors in the front, so
we need not worry about these factors. This yields a τ× -U-α-factorization
&m
'
n
Y
Y
(i) (i)
(i)
(i)
(i)
a = (ai ) =
λi ai1 · · · ain
bi1 · · · bim .
i
i=1
i
i=0
It is certainly a τ× -factorization. Furthermore, bjk 6= 0j for 1 ≤ j ≤ m and 1 ≤ k ≤ mj , so
(j)
bjk is τ× -α by Theorem 5.4. It is also clear from Lemma 5.5 that this is a U-factorization,
showing every non-unit in R has a τ× -U-α-factorization.
Theorem 5.7. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# .
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible} and let
β ∈ {associate, strongly associate, very strongly associate}. Then R is a τ× -U-α-β-df ring if
and only if Ri is τi -U-α-β-df ring for each 1 ≤ i ≤ N.
Proof. (⇒) Let a ∈ Ri0 be a non-unit. Suppose there were an infinite number of τi0 -U-α
(i0 ) ∞
essential divisors of a, say {bj }∞
j=1 none of which are β. But then {bj }j=1 yields an infinite
set of τ× -U-α-divisors of a(i0 ) by Lemma 5.5. Furthermore, none of them are β by Lemma
5.1.
(⇐) Let (ai ) ∈ R be a non-unit. We look at the collection of τ× -U-α essential divisors of
(ai ). Each must be of the form (λ1 , · · · , bi0 , · · · λN ) with λi ∈ U(Ri ) for each i and with bi0
τi0 -α for some 1 ≤ i0 ≤ N. But then bi0 is a τi0 -α essential divisor of ai0 . For each i between
1 and N, Ri is a τi -U-α-β-df ring, so there can be only finitely many τi -α essential divisors
of ai up to β, say N(ai ). If ai ∈ Ri , then we can simply set N(ai ) = 0 since it is a unit and
has no non-trivial τi -U-factorizations. Hence there can be only
N((ai )) := N(a1 ) + N(a2 ) + · · · + N(aN ) =
N
X
N(ai )
i=1
τ× -α essential divisors of (ai ) up to β. This proves the claim.
Corollary 5.8. Let α and β be as in the theorem. Let R = R1 × · · · × RN for N ∈ N with
τi a symmetric relation on Ri# . Then R is a τ× -U-α τ× -U-α-β-df ring if and only if Ri is
τi -U-α τi -U-α-β-df ring for each 1 ≤ i ≤ N.
Proof. This is immediate from Theorem 5.7 and Theorem 5.6.
Theorem 5.9. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# . Then
R is a τ× -U-BFR if and only if Ri is a τi -U-BFR for every i.
Proof. (⇒) Let a ∈ Ri0 be a non-unit. Then a(i0 ) is a non-unit in R, and hence has a bound
on the number of essential divisors in any τ× -U-factorization, say Ne (a(i0 ) ). We claim this
also bounds the number of essential divisors in any τi0 -U-factorization of a. Suppose for a
moment a = a1 · · · an ⌈b1 · · · bm ⌉ were a τi0 -U-factorization with m > Ne (a(i) ). But then
l
m
(i0 )
(i0 ) (i0 )
(i0 )
(i0 )
a = λ a1 · · · an b1 · · · bm
GENERALIZED U-FACTORIZATION IN COMMUTATIVE RINGS WITH ZERO-DIVISORS
15
is a τ× -U-factorization with more essential divisors than is allowed, a contradiction.
(⇐) Let a = (ai ) ∈ R be a non-unit. Let B(a) = max{Ne (ai )}N
i=1 . Where Ne (ai ) is
the number of essential divisors in any τi -U-factorization of ai , and will say for ai ∈ U(Ri ),
Ne (ai ) = 0. We claim that B(a)N is a bound on the number of essential divisors in any
τ× -U-factorization of a. Let
(ai ) = (λi )(a1i ) · · · (ani ) ⌈(b1i ) · · · (bmi )⌉
be a τ× -U-factorization. We can decompose this factorization so that each factor has at most
one non-unit entry as follows:
(ai ) =
N
Y
i=1
(i) (i)
λi a1i · · ·
N
Y
i=1
a(i)
ni
N
Y
i=1
(i)
b1i · · ·
N
Y
b(i)
mi .
i=1
Some of these factors may indeed be units; however, by allowing a unit factor in the front of
every τ -U-factorization, we simply combine all the units into one at the front, and maintain
a τ× -factorization. We can always rearrange this to be a τ× -U-factorization. Furthermore,
(i)
since aji is inessential, by Lemma 5.1 aji is inessential. Only some of the components of the
essential divisors could become inessential, for instance if one coordinate were a unit. At
(i)
worst when we decompose, bji remains an essential divisor for all 1 ≤ j ≤ m and for all
1 ≤ i ≤ N. But then the product of each of the ith coordinates gives a τi -U-factorization of
ai and thus is bounded by Ne (ai ), so we have m ≤ Ne (ai ) ≤ B(a) and therefore there are no
more than B(a)N essential divisors. Certainly the original factorization is no longer than
the one we constructed through the decomposition, proving the claim and completing the
proof.
Theorem 5.10. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# .
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible }. Then R
is τ× -U-α-HFR if and only if Ri is a τi -U-α-HFR for each i.
Proof. (⇒) Let a ∈ Ri0 be a non-unit. We know by Theorem 5.6 Then a(i0 ) is a non-unit
in R, and has an τ× -U-α-factorization. Suppose there were τi0 -U-α-factorizations of a with
different numbers of essential divisors, say:
a = λa1 · · · an ⌈b1 · · · bm ⌉ = µc1 · · · cn′ ⌈d1 · · · dm′ ⌉
where m 6= m′ . By Lemma 5.5 this yields two τ× -U-α-factorizations:
l
m
l
m
(i0 )
(i0 )
(i0 )
(i0 )
(i0 )
(i0 ) (i0 )
(i0 )
(i0 )
(i0 ) (i0 )
a = λ a1 · · · an b1 · · · bn
= µ c1 · · · cn ′ d 1 · · · d n ′ .
This contradicts the hypothesis that R is a R is τ× -U-α-HFR.
(⇐) Let (ai ) ∈ R be a non-unit. Suppose we had two τ× -U-α factorizations
m
l
′
′
′
′
′
′
(ai ) = (λi )(a1i )(a2i ) · · · (ani ) ⌈(b1i )(b2i ) · · · (bmi )⌉ = (µi )(a1i )(a2i ) · · · (an′i ) (b1i )(b2i ) · · · (bm′i ) .
For each i0 , if ai0 is a non-unit in Ri0 , then since each τ× -α element can only have one
coordinate which is not a unit, we can simply collect all the τ× -divisors which have the i0
coordinate a non-unit. This product forms a τi0 -U-α-factorization of ai0 and therefore the
number of essential τ× -factors with coordinate i0 a non-unit must be the same in the two
factorizations. This is true for each coordinate i0 , hence m = m′ as desired.
16
CHRISTOPHER PARK MOONEY
Theorem 5.11. Let R = R1 × · · · × RN for N ∈ N with τi a symmetric relation on Ri# .
Let α ∈ {irreducible, strongly irreducible, m-irreducible, very strongly irreducible} and let
β ∈ {associate, strongly associate }. Then R is τ× -U-α-β-UFR if and only if Ri is a τi -Uα-β-UFR for each i.
Proof. We simply apply Lemma 5.1 to the proof of Theorem 5.10, to see that the factors can
always be rearranged to match associates of the correct type.
References
[1] D. D. Anderson, M. Axtell, S.J. Forman, and J. Stickles, When are associates unit multiples?, Rocky
Mountain J. Math. 34:3 (2004), 811–828.
[2] D.D. Anderson, D.F. Anderson, and M. Zafrullah, Factorization in integral domains, J. Pure Appl.
Algebra 69 (1990), 1–19.
[3] D.D. Anderson and S. Chun, Irreducible elements in commutative rings with zero-divisors, Rocky Mountain J. Math. 37:3 (2011), 741–744.
[4] D. D. Anderson and Andrea M. Frazier, On a general theory of factorization in integral domains, Rocky
Mountain J. Math. 41:3 (2011), 663–705.
[5] D.D. Anderson and M. Naseer, Beck’s coloring of a commutative ring, J. Algebra 159:2 (1993), 500–514.
[6] D.D. Anderson and S. Valdes-Leon, Factorization in commutative rings with zero divisors, Rocky Mountain J. of Math. 26:2 (1996), 439–480.
[7] A.G. Aǧargün and D.D. Anderson and S. Valdez-Leon, Unique factorization in commutative rings with
zero divisors, Comm. Algebra 27:4 (1999), 1967–1974.
[8] A.G. Aǧargün, D.D. Anderson and S. Valdes-Leon, Factorization in commutative rings with zero divisors,
III, Rocky Mountain J. of Math. 31:1 (2001), 1–21.
[9] M. Axtell, U-factorizations in commutative rings with zero-divisors, Comm. Algebra 30:3 (2002), 1241–
1255.
[10] M. Axtell, S. Forman, N. Roersma and J. Stickles, Properties of U-factorizations, International Journal
of Commutative Rings 2:2 (2003), 83–99.
[11] C.R. Fletcher, Unique factorization rings, Proc. Cambridge Philos. Soc. 65 (1969), 579–583.
[12] C.R. Fletcher, The structure of unique factorization rings, Proc. Cambridge Philos. Soc. 67 (1970),
535–540.
[13] C.P. Mooney, Generalized factorization in commutative rings with zero-divisors, Houston J. of Math.
To appear.
Reinhart Center, Viterbo University, 900 Viterbo Drive, La Crosse, WI 54601
E-mail address: [email protected]
| 0math.AC
|
Clustering Algorithms for the Centralized and Local Models∗
Kobbi Nissim†
Uri Stemmer‡
arXiv:1707.04766v1 [cs.DS] 15 Jul 2017
July 15, 2017
Abstract
We revisit the problem of finding a minimum enclosing ball with differential privacy: Given
a set of n points in the Euclidean space Rd and an integer t ≤ n, the goal is to find a ball of the
smallest radius ropt enclosing at least t input points. The problem is motivated by its various
applications to differential privacy, including the sample and aggregate technique, private data
exploration, and clustering [20, 21, 15].
Without privacy concerns, minimum enclosing ball has a polynomial time approximation
scheme (PTAS), which computes a ball of radius almost ropt (the problem
√ is NP-hard to solve
exactly). In contrast, under differential privacy, until this work, only a O( log n)-approximation
algorithm was known.
We provide new constructions of differentially private algorithms for minimum enclosing ball
achieving constant factor approximation to ropt both in the centralized model (where a trusted
curator collects the sensitive information and analyzes it with differential privacy) and in the
local model (where each respondent randomizes her answers to the data curator to protect her
privacy).
We demonstrate how to use our algorithms as a building block for approximating k-means
in both models.
∗
Research supported by NSF grant No. 1565387.
Dept. of Computer Science, Georgetown University and Center for Research on Computation and Society (CRCS),
Harvard University. [email protected].
‡
Center for Research on Computation and Society (CRCS), Harvard University. [email protected].
†
1
Introduction
We revisit the problem of finding a minimum enclosing ball with differential privacy and provide
efficient algorithms for both the centralized curator model and the local model. The problem is
motivated by its various applications to differential privacy: it can be used as the aggregation
function in the sample and aggregate framework [20, 21] as well as serve as a building block in
private data exploration, and clustering [15]. Given a set of n points in the Euclidean space Rd and
an integer t ≤ n, the goal is to find a ball of smallest radius ropt enclosing at least t input points.
Definition 1.1 (revised in Definition 2.1). An instance of the 1-cluster problem is a collection of n
points in the d-dimensional Euclidean space and a target parameter t ≤ n. Let ropt be the radius of
the smallest ball containing at least t of the points. A (∆, w)-approximate solution for the instance
consists of a ball of radius at most w · ropt containing at least t − ∆ of the points.
Our recent work with Vadhan [21]
√ provided an efficient differentially private (∆,
√ w)-approximation
algorithm for 1-cluster for t > O( d log n/) with ∆ = O(log n/) and w = O( log n) (for clarity
we omit the dependency on some of the√parameters). In words, this algorithm identifies with differential privacy
√ a ball of radius O(ropt · log n) containing at least t − O(log n/) points, provided
that t > O( d log n/). This algorithm works in the centralized curator model, where all data is
collected and processed by a trusted curator.
A new construction based on locality sensitive hashing: Our first contribution is a new
algorithm for the centralized curator model. Our new algorithm LSH-GoodCenter has two main
components. First, a family of locality sensitive hash functions [17] is utilized to identify a small
number of disjoint subsets of the input points, such that (at least) one of these subsets is contained
in an (approximately) minimal enclosing ball. The identification is indirect in that it provides a
predicate that evaluates to one on the “identified” input points and zero on all other input points.
Then, these identified points are averaged with differential privacy to locate a point within their
enclosing ball. We give a construction following these lines for solving 1-cluster in the d-dimensional
Euclidean space. The resulting algorithm provides constant approximation to ropt (i.e., w = O(1)).
Furthermore, the construction lends itself to metrics other than the Euclidean metric, provided
that locality preserving hashing exist and that some “averaging” can be performed in the metric
with differential privacy. Table 1 summarizes this result in comparison with previous algorithms
for the problem.
1.1
Local algorithms
Local algorithms are a sub-class of differentially private algorithms where each individual respondent randomizes her own data to guarantee her own privacy. Hence, respondents in a local algorithm
do not need to trust a centralized curator with their data. The accuracy of local algorithms is,
generally, lower than what is achievable with a trusted curator, however, where data exists in
abundance, local algorithms can provide a practical alternative for computing over sensitive data
with minimal trust assumptions. Indeed, recent industrial implementation of differential privacy
by Google [12] and Apple [24] are reported to utilize local algorithms.
A local algorithm for 1-cluster: Our second contribution is an algorithm for 1-cluster in the
local model. The main part of the construction, Algorithm LDP-GoodCenter, combines locality
1
[20]
(centralized model)
[21]
(centralized model)
This work
(centralized model)
This work
(local model)
Needed cluster size – t
Additive loss in cluster size – ∆
n
o
2
t ≥ max 0.51n, O( d2 log2 |X|)
∆=
1
√
w = O( d/)
· log log(d|X|)
∗
· log1.5 1δ · 2O(log (|X|d))
∗
∆ = 1 · log 1δ · 2O(log (|X|d))
√
0.1
∗
t ≥ n · d · log1.5 1δ · 2O(log (|X|d))
0.1
∗
∆ = n · log 1δ · 2O(log (|X|d))
t ≥ O 1 · n0.67 · d1/3 · log(n|X|)
∆ = O 1 · n0.67 · log(n|X|)
t≥
√
Approximation factor
in radius – w
d
√
w = O( log n)
w = O(1)
w = O(1)
Table 1: Algorithms for 1-cluster
sensitive hashing with the recent optimal heavy hitters local algorithm of Bassily et al. [2] to
identify a subset of input points that fall within an approximately minimal enclosing ball. We
then approximate the average of the identified points using a new (and simple) construction for
approximating averages of points under local differential privacy. Importantly, the respondents
running time and communication complexity is only polylogarithmic in the number of participants,
and the running time for the data curator is slightly super linear.
1.2
Application to clustering
Feldman et al. [15] showed that an algorithm for privately solving the 1-cluster problem can be used
for privately computing approximations to the k-means of the input, by iterating the algorithm for
(roughly) k 0 = k · log(n) times and finding k 0 balls that cover most of the data points (the k 0 centers
can then be post-processed to find k centers that approximate the k-means of the data points). Our
new algorithm for the 1-cluster problem immediately yields new bounds on the achievable error for
privately approximating the k-means.
k-mean clustering in the local model: We show that the construction of [15] can be implemented in the local model. Together with our local algorithm for the 1-cluster problem, this results
in the first local differentially private algorithm that computes a provable approximation of the
k-means clustering problem. As with our local algorithm for the 1-cluster problem, the running
time and the communication complexity of the users is only polylogarithmic in the number of
participants.
2
Preliminaries
Notation. Throughout the paper, we use X to denote a finite totally ordered data domain, and
use X d for the corresponding d-dimensional domain. We identify X d with points in the real d2
dimensional unit cube, quantized with grid step 1/(|X| − 1). Datasets are (ordered) collections of
elements from some data universe U (e.g., U = X or U = X d or U = Rd ). Two datasets S, S 0 ∈ U n
are called neighboring if they differ on at most one entry.
2.1
1-cluster
The reason for our explicit embedding of the finite ordered domain X in R is that certain computations are impossible for differential privacy when the input domain is the unit interval [0, 1].
An example is the Interior Point problem [6] where the task is, given n points in the unit interval
x1 , x2 , . . . , xn , to output a value in the range [min(x1 , . . . , xn ), max(x1 , . . . , xn )]. This impossibility result carries over to the 1-cluster problem (through a reduction from the interior point
problem) [21]. With this fact in mind, we now redefine 1-cluster:
Definition 2.1. A 1-cluster problem (X d , n, t) consists of a d-dimensional domain X d and parameters n ≥ t. We say that algorithm M solves (X d , n, t) with parameters (∆, w, β) if for every input
database S ∈ (X d )n it outputs, with probability at least 1 − β, a center c and a radius r such that
(i) the ball of radius r around c contains at least t − ∆ points from S; and (ii) r ≤ w · ropt , where
ropt is the radius of the smallest ball in X d containing at least t points from S.
2.2
Preliminaries from differential privacy
Consider a database, where each entry contains information pertaining to an individual. An algorithm operating on databases is said to preserve differential privacy if a change of a single record
of the database does not significantly change the output distribution of the algorithm. Intuitively,
this means that individual information is protected: whatever is learned about an individual could
also be learned with her data arbitrarily modified (or without her data).
Definition 2.2 (Differential Privacy [9]). A randomized algorithm M : U n → Y is (, δ) differentially private if for every two neighboring datasets S, S 0 ∈ U n and every T ⊆ Y we have
Pr[M (S) ∈ T ] ≤ e · Pr[M (S 0 ) ∈ T ] + δ, where the probability is over the randomness of M .
2.2.1
The Laplace and Gaussian Mechanisms
The most basic constructions of differentially private algorithms are via the Laplace and Gaussian
mechanisms as follows.
Definition 2.3 (Lp -Sensitivity). A function f mapping databases to Rd has Lp -sensitivity k if
kf (S) − f (S 0 )kp ≤ k for all neighboring S, S 0 .
Theorem 2.4 (Laplace mechanism [9]). A random variable is distributed as Lap(λ) if its probability
1
∗
d
density function is f (y) = 2λ
exp(− |y|
λ ). Let > 0, and assume f : U → R has L1 -sensitivity k.
d
The mechanism that on input S ∈ U ∗ outputs f (S) + Lap( k ) is (, 0)-differentially private.
Theorem 2.5 (GaussianpMechanism [7]). Let , δ ∈ (0, 1), and assume f : U ∗ → Rd has L2 sensitivity k. Let σ ≥ k 2 ln(1.25/δ). The mechanism that on input S ∈ U ∗ outputs f (S) +
d
N (0, σ 2 ) is (, δ)-differentially private.
3
2.2.2
Stability based histogram [8, 23, 3]
Given a database S ∈ U ∗ , consider the task of choosing a “good” solution out of a possible solution
set F where “goodness” is measured by a quality function q : U ∗ × F → N assigning “scores” to
solutions from F (w.r.t. the given database S). One of the constructions in differential privacy –
the exponential mechanism [19] – can privately identify an approximately optimal solution f ∈ F
provided that q has low-sensitivity and that |S| & log |F |.
By focusing on cases where the number of possible solutions with “high” scores is limited, it is
possible to relax the requirement that |S| & log |F |, using what has come to be known as stability
based techniques. In this work we use stability based techniques for the following task: Given a
dataset S ∈ U n and a partition P of U , find a set p ∈ P containing (approximately) maximum
number of elements of S.
n
Theorem 2.6 ([3, 5, 25]). Fix β, , δ, n, and let t ≥ 12
ln( βδ ). There exists an (, δ)-differentially
n
private algorithm that given S ∈ U returns a collection L ⊂ P of sets in the partition P such
that |{x ∈ S : x ∈ p}| ≥ t/4 for every p ∈ L (hence, there are at most 4n/t sets in the collection
L); and, with probability at least (1 − β), the collection L contains every set p ∈ P such that
|{x ∈ S : x ∈ p}| ≥ t.
2.2.3
The sparse vector technique [10]
Consider a sequence of low sensitivity functions f1 , f2 , . . . , fk , which are given (one by one) to a
data curator (holding a database S). Algorithm AboveThreshold by Dwork et al. [10] privately
identifies the first query fi whose value fi (S) is greater than some threshold t:
Theorem 2.7 (Algorithm AboveThreshold [10]). There exists an (, 0)-differentially private algorithm A such that for k rounds, after receiving a sensitivity-1 query fi : U ∗ → R, algorithm A
either outputs > and halts, or outputs ⊥ and waits for the next round. If A was executed with a
database S ∈ U ∗ and a threshold parameter t, then the following holds with probability (1 − β): (i)
If a query fi was answered by > then fi (S) ≥ t − 8 log(2k/β); (ii) If a query fi was answered by ⊥
then fi (S) ≤ t + 8 log(2k/β).
2.2.4
Composition theorems
We construct algorithms that use several differentially private mechanisms as subroutines, and
analyze the overall privacy using the following composition theorems:
Theorem 2.8 (Basic composition [7, 8]). A mechanism that permits k adaptive interactions with
(, δ)-differentially private mechanisms (and does not access the database otherwise) is (k, kδ)differentially private.
Theorem 2.9 (Advanced composition [11]). Let , δ, δ 0 > 0. A mechanism that permits k adaptive
interactions with (, δ)-differentially private mechanisms (and
p does not access the database otherwise) is (0 , kδ + δ 0 )-differentially private, for 0 = 2k2 + 2k ln(1/δ 0 ).
3
Our Algorithms
Let X d be a a finite subset of the Euclidean space Rd . We consider the following problem under
differential privacy:
4
Definition 3.1 (The problem of minimum enclosing ball with t points). Given a set S of n points
in X d and a parameter t ≤ n, find a ball of minimal radius ropt enclosing at least t input points.
In this section we give an informal description of our algorithms for this problem and highlight
some of the ideas behind the constructions. Any informalities made herein will be removed in the
sections that follow. Consider the following relaxed variant of the above problem:
Definition 3.2 (Minimum enclosing ball – promise version). Given a set S of n points in X d , an
integer t ≤ n, and a radius r ∈ R s.t. there exists a ball in X d containing at least t points from S,
the goal is to find a ball of radius r enclosing at least t input points.
That is, in addition to the input set S and the parameter t, in this promise problem we are also
given a target radius r, and our goal is to identify a ball of that radius that contains t of the input
points, under the promise that such a ball exists.
As was observed in [21], it suffices to solve the promise problem. To see this, fix a database
S ∈ (X d )n , and let ropt denote the smallest radius s.t. there is a ball of that radius containing t
points from S. Assume that we have an algorithm A for the promise problem. Now, for every
possible choice for a radius r, consider applying A on the database S with the radius r to obtain a
center cr . As A solves the promise problem, for every choice for r s.t. r ≥ ropt , we will have that
the ball of radius r around the obtained center cr contains at least t points from S. Using standard
tools from differential privacy (e.g., the Laplace mechanism), we can now privately estimate, for
every r, the number of input points contained in the ball of radius r around cr . Afterwards we
can return the center cr∗ and the radius r∗ such that r∗ is the minimal radius for which (our
estimation for) the number of points in the corresponding ball is at least ≈ t. In fact, as we are
only aiming for a solution with an approximated radius r, it suffices to only consider radiuses in
powers of 2. As our domain X d is finite, there are at most log(d|X|) such choices, so this will
introduce error at most polylog(d|X|) in our estimations (meaning that we will identify a ball
containing ≈ t − polylog(d|X|) input points). Furthermore, a binary search for the radius r∗ (out
of all the possible powers of 2) would reduce the error to polyloglog(d|X|). Using the privacypreserving recursion on binary search of [3] it is possible to reduce the incurred error even further,
∗
to 2O(log (d|X|)) . In the rest of this section we focus on solving (or rather approximating) the promise
problem under differential privacy. To that end, let us take an even closer look at the problem, and
define the following restricted variant:
Definition 3.3 (Promise problem, restricted to t = n). Given a set S of t points in X d and a
radius r ∈ R s.t. there exists a ball in X d containing all t input points, the goal is to locate a ball
of radius r enclosing all t input points.
Intuitively, this makes the problem easier since we no longer have to identify which of the n
points are in the 1-cluster (i.e., which t points can be enclosed in a ball of radius r). Indeed, without
privacy concerns, this restricted variant of the problem is easier, and algorithms for the restricted
problem are sometimes used as a subroutine for approximating the general problem.1 Similarly,
Nissim et al. [21] showed that in order to obtain a private approximation for the unrestricted
problem (Definition 3.2) it suffices to privately approximate the restricted problem (Definition 3.3).
1
Computing an exact solution to the problem of a minimum ball enclosing t points is NP-hard, even for the case
where t = n. However, better approximation algorithms exist for the restricted case where t = n. The restricted
version has a FPTAS while the general problem does not (assuming that P 6= N P ). See, e.g., [22].
5
√
However, their reduction introduced a factor of log n in the radius of the computed ball. Our
main technical contribution can be thought of as a new reduction from the problem of privately
approximating a solution for the general problem to privately approximating a solution for the
restricted problem while introducing only a constant blowup in the radius.
We first briefly describe the techniques of [21] for privately approximating the restricted problem:
Assume that we are given a set S of t input points in Rd , and a radius r such that the set S can
be enclosed in a ball of radius r. Our goal is to locate a center c such that a ball of radius O(r)
around c contains all of S. The strategy here is conceptually simple – compute (an estimate for)
the average of the points in S. Indeed, as S is of diameter 2r, a ball of radius 2r around this average
contains all of S. Estimating
√ the average of S can be done privately using the√Gaussian mechanism,
d·r
with error proportional to |S|d times the diameter of S, i.e., proportional to |S|
. Hence, assuming
√
that t = |S| & d, a ball of radius O(r) around the (privately obtained) estimation for the average
contains all of S.2
3.1
Our contributions
Our final task is to transform a private algorithm for the restricted promise problem
into a private
√
algorithm for the unrestricted promise problem, without incurring a blowup of log n in the radius.
As we next explain, this can be achieved using locality sensitive hashing. Informally, a locality
sensitive hash function aims to maximize the probability of a collision for similar items, while
minimizing the probability of collision for dissimilar items. Formally,
Definition 3.4 ([17]). Let M be a metric space, and let r>0, c>1, 0≤q<p≤1. A family H of
functions mapping M into domain U is an (r, cr, p, q) locality sensitive hashing family (LSH) if for
all x, y ∈ M (i) Prh∈R H [h(x) = h(y)] ≥ p if dM (x, y) ≤ r; and (ii) Prh∈R H [h(x) = h(y)] ≤ q if
dM (x, y) ≥ cr.
Now, let S ∈ Rn be an input database and let r ∈ R be s.t. there exists a ball of radius r
that contains at least t points from S. Let P ⊆ S denote the guaranteed set of t input points
contained in a ball of radius r. Assume that we have an (r, cr, p, q) locality sensitive hashing family
H mapping Rd to some domain U , and sample a function h ∈R H. Intuitively, if q (the collision
probability of dissimilar elements) is less than n−2 , then w.h.p. we will have that all dissimilar
elements (i.e., at distance more than cr) are mapped into different hash values. On the other hand,
as P ⊆ S is of diameter r, we expect that ≈ p · |P | = pt of the elements of P will be mapped into
the same hash value. In the following sections we show that this intuition is indeed correct, and
that with noticeable probability the following two events occur (over the choice of h ∈ H):
(E1 ) For every x, y ∈ S s.t. kx − yk ≥ cr it holds that h(x) 6= h(y); and,
(E2 ) There exists a hash value in U , denoted u∗ , such that |{x ∈ P : h(x) = u∗ }| ≥
p
2
· t.
Event (E1 ) states that if two points in S are mapped into the same hash value, then these points
are close. Event (E2 ) states that there is a “heavy” hash value u∗ ∈ U , such that “many” of the
points in P are mapped into u∗ .
2
For technical reasons, we cannot apply the Gaussian mechanism directly. Following [21], before using the Gaussian
mechanism, we first privately compute a box of diameter ≈ r that contains all of S, and use it to bound the diameter
of S.
6
Using standard stability based arguments (see, e.g., Theorem 2.6), we can now privately identify
a list L containing all such “heavy” hash values u ∈ U (in particular, we will have that u∗ ∈ L).
Once we have the list L, we can use every u ∈ L as a “filter” on S that isolates clustered points:
Su = {x ∈ S : h(x) = u}. Observe that, by Event E1 , for every u ∈ L we have that Su can be
enclosed in a ball of radius cr. Hence, for every u ∈ L we can use a private algorithm for the
restricted promise problem on the database Su with the radius cr, to obtain a center yu s.t. a ball
of radius O(cr) around yu contains all of Su .
So, we have generated a set of centers Y , such that there exists yu∗ ∈ Y for which a ball of
radius O(cr) centered at yu∗ contains all of Su∗ . As Su∗ contains some of the points from the
guaranteed cluster P ⊆ S of diameter 2r, we get that the ball of radius O(cr + 2r) = O(cr) around
yu∗ contains all of P , and hence, contains at least t points from S. All that remains is to privately
identify a center y ∈ Y s.t. the ball of radius O(cr) around it contains ≈ t points from S. This can
be done privately, e.g., using the sparse vector technique (see Theorem 2.7).
Choosing the LSH parameters. Recall that for the discussion above we needed that q (the
collision probability of dissimilar elements) is less than n−2 . In addition, in order to obtain a
constant factor approximation in the radius, we want to have c = O(1). Under those requirements,
existing LSH constructions can achieve p = n−b for some small constant b > 0 (for example, see [1]).
This, in turn, means that we need to have t n−b in order for Event E2 above to be meaningful.
We obtain the following theorem.
Theorem 3.5. Let n, t, β, , δ be s.t.
t≥O
n0.1 ·
√
d
!
s
√
∗
1
nd
1
log
log
log
· 9log (2|X| d) .
β
βδ
βδ
There exists an (, δ)-differentially private algorithm that solves the 1-cluster problem (X d , n, t) with
parameters (∆, w) and error probability β, where w = O (1) and
0.1
√
n
1
1
log∗ (2|X| d)
∆=O
log
log
·9
.
βδ
β
In words, there exists an efficient (, δ)-differentially private algorithm that (ignoring logarithmic
0.1
factors) is capable of identifying a ball of radius O(ropt ) containing t − Õ( n ) points, provided
√
that t ≥ Õ(n0.1 · d/). The exponent 0.1 in the factor n0.1 is arbitrary and can be reduced
to any constant by appropriately choosing the LSH parameters. See Section 4 for the complete
construction and analysis.
3.1.1
1-cluster in the local model
In Section 5 we show that many of the ideas discussed above have analogues for the local model,
where each individual holds her private information locally, and only releases the outcomes of
privacy-preserving computations on her data. While our construction for the local model is conceptually similar to the centralized construction, several modifications are required. These include
introducing a new protocol for computing averages of points in Rd (satisfying local differential
privacy), and a modified “filtering” usage of the locality sensitive hash function.
7
We obtain the following theorem, which is similar in spirit to Theorem 3.5, albeit with weaker
guarantees.
Theorem 3.6. Let S = (x1 , . . . , xn ) ∈ (X d )n be a database which is distributed across n players
(each holding one point in X d ). Let t, β, be s.t.
r
1 0.67 1/3
1
t≥O
·n
· d · log(dn|X|) log( ) .
β
There exists an algorithm, satisfying -local differential privacy (LDP), that identifies a center y
and a radius r s.t. with probability at least 1 − β,
1. r = O(ropt ), where ropt is the radius of a smallest ball containing at least t points from S.
2. The number of points from S contained in the ball of radius r around y is at least t − ∆, where
r
1 0.67
1
n
∆=O
·n
· log(d|X|) log( ) log( ) .
β
β
Our algorithm uses three rounds of communication with the users. Constructing a non-interactive
LDP algorithm for the 1-cluster problem (while maintaining constant error in the radius) remains
an intriguing open problem.
4
Locating a small cluster – the centralized setting
We begin with an algorithm for the trusted curator model. As discussed in Section 3, we use a tool
from [21] that enable us to focus on the task of locating a ball of radius r, under the assumption
that r is given to us and that r is such that there exists a ball of radius r enclosing at least t of the
input points.
4.1
4.1.1
Additional preliminaries
Tool from prior work: Algorithm GoodRadius [21]
Let S ⊆ X n be a set of input points. Given a parameter t, consider the task of privately identifying
a radius r s.t. there exists a ball of radius r containing ≈ t points from S, and furthermore, r is not
much bigger than the smallest such possible radius. Using privacy preserving recursion on binary
search [3], Nissim et al. [21] obtained the following construction:
Theorem 4.1 (Algorithm GoodRadius, [21]). Let S ∈ (X d )n be a database containing n points
from X d and let t, β, , δ be parameters. There exists a poly(n, d, log |X|)-time (, δ)-differentially
private algorithm that on input S outputs a radius r ∈ R s.t. with probability at least (1 − β):
∗
1
1. There is a ball in X d of radius r containing at least t−O 1 log( βδ
) · 9log (|X|·d) input points.
2. r ≤ 4 · ropt where ropt is the radius of the smallest ball in X d containing at least t points
from S.
8
4.1.2
Random rotation
We also use the following technical lemma to argue that if a set of points P is contained within a
ball of radius r in Rd , then by randomly rotating the Euclidean√space we get that (w.h.p.) P is
contained within an axis-aligned rectangle with side-length ≈ r/ d.
Lemma 4.2 (e.g., [26]). Let P ∈ (Rd )m be a set of m points in the d dimensional Euclidean space,
and let Z = (z1 , . . . , zd ) be a random orthonormal basis for Rd . Then,
i
h
p
Pr ∀x, y ∈ P : ∀1 ≤ i ≤ d : |hx − y, zi i| ≤ 2 ln(dm/β)/d · kx − yk2 ≥ 1 − β.
Z
4.2
Algorithm GoodCenter
We are now ready to present algorithm LSH-GoodCenter. Given a radius r computed by algorithm
GoodRadius (Theorem 4.1), the algorithm privately locates a ball of radius O(r) containing & t
points.
The privacy properties of LSH-GoodCenter follow directly from composition (see Section 2.2.4).
The analysis appears in Appendix A.
Lemma 4.3. Algorithm LSH-GoodCenter preserves (2, 2δ)-differential privacy.
We now proceed with the utility analysis of algorithm LSH-GoodCenter. We will assume the
existence of a family H of (r, cr, p=n−b , q=n−2−a )-sensitive hash functions mapping Rd to a universe
U , for some constants a > b, r > 0, and c > 1.
Lemma 4.4. Let LSH-GoodCenter be executed with the family H on a database S containing n
points in Rd , with parameters β = n−a /20 and t, , δ s.t.
!
√
r
d
nd
1
t ≥ O nb ·
log
log( ) .
δ
δ
If there exists a ball of radius r in Rd containing at least t points
from S, then with probability at
least n−a /4, the output ŷ in Step 7 is s.t. at least t − O 1 log(n) of the input points are contained
in a ball of radius 2r(c + 1) around ŷ.
Remark 4.5. We can think of a and b as small constants, e.g., a = 0.2 and b√= 0.1. Hence,
ignoring logarithmic factors, the above theorem only requires t to be as big as n0.1 d/. Note that
the algorithm only succeeds with small probability – namely, with probability n−a /4. This can easily
be amplified using repetitions (roughly na = n0.2 repetitions).
Proof. First recall that by the properties of the family H, for every x, y ∈ Rd s.t. kx − yk ≥ cr we
have that Prh∈H [h(x) = h(y)] ≤ q = n−2−a . Using the union bound we get
Pr [h(x) 6= h(y) for all x, y ∈ S s.t. kx − yk ≥ cr] ≥ (1 − n−a /2).
h∈R H
Let P ⊆ S denote the guaranteed set of t input points that are contained in a ball of radius r,
and let x ∈ P be an arbitrary point in P . By linearity of expectation, we have that
E [|{y ∈ P : h(y) 6= h(x)}|] ≤ t(1 − p) = t(1 − n−b ).
h∈H
9
Hence, by Markov’s inequality,
t(1 − n−b )
≤ 1 − n−a .
Pr |{y ∈ P : h(y) 6= h(x)}| ≥
h∈H
1 − n−a
So,
1 − n−b
Pr |{y ∈ P : h(y) = h(x)}| ≥ t 1 −
≥ n−a .
h∈H
1 − n−a
Simplifying, for large enough n (specifically, for na−b ≥ 2) we get
t
−b
Pr |{y ∈ P : h(y) = h(x)}| ≥ · n
≥ n−a .
h∈H
2
So far we have established that with probability at least n−a /2 over the choice of h ∈ H in
Step 1 the following events occur:
(E1 ) For every x, y ∈ S s.t. kx − yk ≥ cr it holds that h(x) 6= h(y); and,
(E2 ) There exists a hash value in U , denoted u∗ , such that |{y ∈ P : h(y) = u∗ }| ≥
t
2
· n−b .
Event (E1 ) states that if two points in S are mapped into the same hash value, then these
points are close. Event (E2 ) states that there is a “heavy” hash value u∗ ∈ U , such that “many” of
the points in P are mapped into u∗ . We proceed with the analysis assuming that these two events
occur.
In Step 2 we use Theorem 2.6 to identify a list L ∈ U ∗ that, w.p. (1 − β), contains all “heavy”
hash values u ∈ U . In particular, with probability at least (1 − β) we have that u∗ ∈ L. We proceed
with the analysis assuming that this is the case.
In Step 3 we generate a random orthonormal basis Z. By Lemma 4.2, with probability at least
(1 − β), for everypx, y ∈ S and for every zi ∈ Z, we have that the projection of (x − y) onto zi is of
length at most 2 ln(dn/β)/d · kx − yk. In particular, for every hash value u ∈ L we have that the
projection of p
Su , {x ∈ S : h(x) = u} onto every axis zi ∈ Z fits within an interval of length at
most p = 2rc ln(dn/β)/d.
In Step 4 we use Theorem 2.6 again to identify, for every u ∈ L and for every axis zi ∈ Z,
an interval of length 3p containing all of u. For our choice of t, all of these applications together
succeed with probability
at least (1 − β). This results, for every u ∈ L, in a box B(u) of side length
√
3p and diameter 3p d containing all of Su .
For every u ∈ L denote the average of the points in Su as yu . In Step 5 we use the Gaussian
mechanism to compute the noisy average ŷu of Su for every u ∈ L. The noise magnitude reduces
with the size of the sets |Su |, and for our choice of t (recall that |Su | ≥ t/4 for every u ∈ L), with
probability at least (1 − β) we have that kyu − ŷu k ≤ cr for every u ∈ L.
Recall that there exists a hash value u∗ ∈ L such that some of the points from the guaranteed
cluster P are mapped (by h) into u∗ . Let v ∈ P be such that h(v) = u∗ . A ball of radius 2r around
v contains all of P , as P is of radius r. In addition, by (E1 ), a ball of radius cr around yu∗ contains
all of Su∗ . As, kyu∗ − ŷu∗ k ≤ cr, we get that a ball of radius 2cr around ŷu∗ contains all of Su∗ ,
and in particular, contains v. Thus, a ball of radius 2cr + 2r around ŷu∗ contains all of P , that
is, contains at least t input points. Hence, by the properties of algorithm AboveThreshold, with
probability at least (1 − β), the loop in Step 7 ends with algorithm AboveThreshold retuning >,
10
identifying a point ŷu such that a ball of radius 2cr + 2r around it contains at least t − 65
log(2n/β)
input points.
All in all, with probability at least n−a /2 − 5β = n−a /4 we have that algorithm GoodCenter
outputs a point ŷ such that a ball of radius 2r(c + 1) around it contains at least t − 65
log(2n/β)
input points.
Theorem 3.5 now follows by combining Theorem 4.1 (algorithm GoodRadius) with Lemmas 4.3
and 4.4 (algorithm LSH-GoodCenter).
5
Locating a small cluster – the distributed setting
We begin by describing private computation in the local model where each individual holds her
private information locally, and only releases the outcomes of privacy-preserving computations on
her data. This is modeled by letting the algorithm access each entry xi in the input database
S = (x1 , . . . , xn ) ∈ X n separately, and only via differentially private local randomizers.
Definition 5.1 (Local Randomizer, LR Oracle [18]). A local randomizer R : X → W is a randomized algorithm that takes a database of size n = 1. Let S = (x1 , . . . , xn ) ∈ X n be a database.
An LR oracle LRS (·, ·) gets an index i ∈ [n] and a local randomizer R, and outputs a random value
w ∈ W chosen according to the distribution R(xi ).
Definition 5.2 (Local differential privacy). An algorithm satisfies (, δ)-local differential privacy
(LDP) if it accesses the database S only via the oracle LRS with the following restriction: for
all possible executions of the algorithm and for all i ∈ [n], if LRS (i, R1 ), . . . , LRS (i, Rk ) are the
algorithm’s invocations of LRS on index i, then the algorithm B(x) = (R1 (x), R2 (x), . . . , Rk (x)) is
(, δ)-differentially private. Local algorithms that prepare all their queries to LRS before receiving
any answers are called noninteractive; otherwise, they are interactive.
5.1
Additional preliminaries
We now present additional preliminaries that enable our construction.
5.1.1
Counting queries and histograms with local differential privacy
The most basic task we can apply under local differential privacy is a counting query. Consider a
database S ∈ X n which is distributed across n users (each holding one element from X). Given
a predicate y : X → {0, 1}, consider the task of estimating the number of users holding a value x
√
such that y(x) = 1. This can be solved privately with error proportional to 1 n.
Notation. For a database S ∈ X n and a domain element x ∈ X, we use fS (x) to denote the
duplicity of x in S, i.e.,
fS (x) = |{xi ∈ S : xi = x}|.
Theorem 5.3 (e.g., [18]). Let ≤ 1. Let S = (x1 , . . . , xn ) ∈ {0, 1}n be a database which is
distributed across n players (each holding one bit). There exists an algorithm satisfying q
-LDP that
returns an estimation a ∈ R s.t. with probability at least 1−β we have that |a−fS (1)| ≤
11
3
n · ln( β8 ).
Algorithm LSH-GoodCenter
Input: Database S ∈ Rd containing n points, radius r, target number of points in cluster t, failure
probability β, privacy parameters , δ, and another parameter c (reflecting multiplicative error in
the radius of the found ball).
Tool used: A family H of (r, c · r, p, q)-locality sensitive hash functions mapping Rd to a universe
U.
1. Sample a hash function h ∈ H mapping Rd to U . For u ∈ U define Su as the multiset
containing all elements of S that are mapped into u by h, i.e., Su , {x ∈ S : h(x) = u}.
2. Use algorithm from Theorem 2.6 with privacy parameters ( 4 , 4δ ) to identify L ⊂ U that, w.p.
q
√
)
(1 − β), L contains every hash value u ∈ U s.t. |Su | ≥ 960 d ln( nd
ln( 8δ ). Moreover, for
βδ
q
√
8
every u ∈ L we have that |Su | ≥ 240 d ln( nd
βδ ) ln( δ ).
% W.h.p. L contains a hash value u s.t. a lot of the input points from the guaranteed cluster of radius r
are mapped into u.
q
3. Let Z = (z1 , . . . , zd ) be a random orthonormal basis of Rd and p = 2rc ln( dn
β )/d.
For each basis vector zi ∈ Z, partition the axis in direction zi into intervals Ii =
{[j · p , (j + 1) · p) : j ∈ Z}.
4. For every hash value u ∈ L, identify (in an axis by axis manner) an axis aligned rectangle
containing all of Su . Specifically, for every u ∈ L:
(a)
For each basis vector
zi ∈ Z, use algorithm from Theorem 2.6 with privacy parameters
δ
√
to choose an interval Ii (u) ∈ Ii containing a large number of points
, 8d
10
d ln(8/δ)
from Su . Let Iˆi (u) denote that chosen interval after extending it by p on each side (that
is Iˆi (u) is of length 3p).
(b) Let B(u) be the box in Rd whose projection on every basis vector zi ∈ Z is Iˆi (u). Define
Su0 = Su ∩ B(u).
% Observe that we defined Su0 = Su ∩ B(u) although we expect that Su ⊆ B(u). This will be useful
in the privacy analysis, as we now have a deterministic bound on the diameter of Su0 .
5. For every hash value u ∈ L: Use the Gaussian mechanism with privacy parameters ( 4 , 4δ ) to
compute the noisy average of the points in Su0 . Denote the outcome as ŷu .
6. Instantiate algorithm AboveThreshold (Theorem 2.7) with database S, privacy parameter
/4, and threshold t − 33
log(2n/β).
7. For every hash value u ∈ L: Query algorithm AboveThreshold for the number of input points
contained in a ball of radius 2r(c + 1) around ŷu . If the answer is > then halt and return ŷu .
Otherwise continue.
12
A slightly more complicated task is when instead of a binary predicate, we have a function
y : X → U mapping domain elements to a (potentially) large set U . We use a tool from the recent
work of Bassily et al. [2] on heavy hitters in the local model.
Theorem 5.4 ([2]). Let ≤ 1. Let U be a data universe, and let S = (u1 , . . . , un ) ∈ U n be a
database which is distributed across n players (each holding one row). There exists an algorithm
√
satisfying -LDP that returns a set L ⊆ (U × R) of size at most |L| ≤ n s.t. with probability 1 − β
we have
q
1. For every (u, a) ∈ L we have that |a − fS (u)| ≤ O 1 n log( βn ) .
2. For every u ∈ U s.t. fS (u) ≥ O
5.1.2
q
1
1
n log( |X|
β ) log( β )
, we have that u appears in the set L.
Concentration bounds
We will use the following variant of the Hoeffding bound for sampling without replacement.
Theorem 5.5 (Hoeffdings inequality, sampling without replacement [16]). Let X = (x1 , . . . , xN )
be a finite population of N points and X1 , . . . , Xn be a random sample drawn without replacement
from X . Let a = mini∈[N ] {xi } and b = maxi∈[N ] {xi }. Then, for all > 0,
Pr
"
#
n
N
1 X
2n2
1X
Xi −
xi ≥ ≤ 2 exp −
.
n
N
(b − a)2
i=1
i=1
We also use the following tail bound on sums of λ-wise independent random variables.
Lemma 5.6 ([4]). Let λ ≥ 6 be an even integer. Suppose X1 , . . . , Xn are λ-wise independent
random variables taking values in [0, 1]. Let X = X1 + · · · + Xn and µ = E[X], and let α > 0.
Then,
λ/2
nλ
Pr[|X − µ| ≥ α] ≤
.
α2
We also use a standard “Chernoff type” concentration bound for sums of i.i.d. samples from
Lap(b):
Lemma 5.7. Suppose X1 , . . . , Xn are independent samples from Lap( 1 ). That is, every Xi has
probability density function f (x) = 2 exp(−|x|). Let X = X1 + · · · + Xn . For every t ≥ 0 it holds
that
·t
Pr[|X| ≥ t] ≤ 6 exp − √
.
2n
Moreover, for every 0 ≤ t < 2n/ we have that
2 2
·t
Pr[|X| ≥ t] ≤ 2 exp −
.
4n
13
5.2
Average of vectors in Rd
For our algorithms we will need a tool for privately computing averages of vectors in Rd . Specifically,
assume that there are n users, where user j is holding a point xj ∈ Rd . Moreover, assume that
we have a bound b ≥ 0 such that we are only interested in points xj ∈ Rd s.t. every coordinate of
xj is between 0 and b (let us them “interesting points”). We would like to obtain an estimation
(satisfying LDP) for the average of these interesting points. This can be achieved by having each
user send a noisy version of its point. An estimation for the average of the interesting points can
then be computed from the average of all the noisy points.
Algorithm R: Local Randomizer for LDP-AVG
Inputs: x ∈ R, privacy parameter , and range parameter b.
1. If x > b or x < 0 then set x = 0.
2. Let z ← x + Lap( b ).
3. Return z.
Algorithm LDP-AVG
Public parameters: Random partition of [n] into d subsets I1 , . . . , Id .
Setting: Each player j ∈ [n] holds a value vj ∈ Rd ∪ {⊥}. Define S = (v1 , . . . , vn ).
For j ∈ [n] let vj0 = vj if every coordinate of vj is in [0, b]. Otherwise set vj0 = ~0 ∈ Rd .
For j ∈ [n] let xj denote the `th coordinate of vj0 , where ` is s.t. j ∈ I` .
Define Se = (x1 , · · · , xn ).
e
Oracle access: LR Oracle to S and S.
1. For j ∈ [n] let zj ← LRSe(j, R) with parameters 2 , b.
P
2. For ` ∈ [d] define y` = |In` | j∈I` zj . Let y = (y1 , . . . , yd ) ∈ Rd .
3. Let t denote the number of users j holding a value vj ∈ Rd s.t. every coordinate of vj is
between 0 and b. Use Theorem 5.3 with parameter /2 to obtain a noisy estimation t̂ for t.
4. Return
1
t̂
· y.
Observation 5.8. Algorithm LDP-AVG satisfies -LDP.
Lemma 5.9. Let b > 0 be a parameter. Fix individual information vj ∈ Rd ∪ {⊥} for every user
j ∈ [n]. Let D ⊆ [n] denote the subset of users holding a value vj 6= ⊥ s.t. every coordinate
of vj
q
is between 0 and b, and denote |D| = t. Assume that n ≥ 8d ln( 8d
β ), and that t ≥
12
n log( 32
β ).
Then, with probability at least 1 − β, algorithm LDP-AVG returns a vector a ∈ Rd such that
s
1X
30bd
32d
a−
vj ≤
n ln(
).
t
t
β
j∈D
2
14
Remark 5.10. For simplicity, in Lemma 5.9 we assumed that “interesting points” are points
x ∈ Rd such that x ∈ B, where B is a box with side-length b, aligned with the origin. This can
easily be generalized to arbitrary axis-aligned boxes B with side-length b (that is, not necessarily
located at the origin) by, e.g., shifting the axes accordingly.
Remark 5.11. Observe that the
√ error in Lemma 5.9 decreases with t (the number of “interesting
points”). Specifically, for t√& nd we get that the error is proportional to the diameter of the box
of interesting points, i.e., db.
Remark 5.12. Note that every user in the above protocol sends (on step 3 of algorithm R) a real
number z. In fact, it suffices to have each user send only O(log n) bits, without effecting the error
guarantees
of
Lemma 5.9. To see this, observe that we can truncate z to an interval of length
b
n
O log( β ) . Indeed, by the properties of the Laplace distribution, w.h.p. this truncation has no
q
effect on the computation. We can now discretize this interval with grid steps O tb nd log( βd ) ,
and have every user send the grid point closest to the
point z, which
can be represented using
q
d
d
b
O(log n) bits. This introduces an error of at most O t n log( β ) per user, and hence error at
q
q
d
most O tb dn log( βd ) per coordinate, and thus, error at most O bd
n
log(
)
in L2 norm.
t
β
Proof of Lemma 5.9. Our goal is to relate y/t̂ (the output obtained on step 4) to
P
will now show that w.h.p. every coordinate of yt̂ − 1t j∈D vj is small.
1
t
P
j∈D vj .
We
First observe that by the Chernoff bound, assuming that n ≥ 8d ln( 8d
β ), with probability at
n
least 1 − β4 , for every ` ∈ [d] we have that 2d
≤ |I` | ≤ 2n
d . We continue the analysis assuming that
this is the case. Also, by Theorem 5.3, with probability at least 1 − β4 we have that t̂ (computed on
q
q
12
).
In
particular,
assuming
that
t
≥
n log( 32
step 3) satisfies |t − t̂| ≤ 6 n log( 32
β
β ) we get that
t
2
≤ t̂ ≤ 2t. We proceed with the analysis assuming that this is the case.
For ` ∈ [d], observe that y` (defined on Step 2 of the algorithm) can be expressed as
y` =
n X
n X
n X
(xj + ηj ) =
vj,` +
ηj ,
|I` |
|I` |
|I` |
j∈I` ∩D
j∈I`
(1)
j∈I`
where vj,` is coordinate ` of vj , and every ηj is sampled from Lap( 2b
). By the Hoeffding bound
(sampling without replacement, Theorem 5.5), with probability at least 1 − β4 we have that
X
n X
vj,` −
vj,` ≤ bn
|I` |
j∈I` ∩D
j∈D
s
15
p
1
ln(8/β) ≤ b dn ln(8/β).
2|I` |
(2)
Combining Equalities (1) and (2), and using the triangle inequality, we get
y` 1 X
−
vj,`
t
t̂
j∈D
≤
p
b dn ln(8/β)
1 1 X
n X
−
+
ηj +
vj,`
t
t̂
t̂|I` | j∈I
t̂ j∈D
`
≤
p
b dn ln(8/β)
n X
|t̂ − t| · b
+
ηj +
t̂
t̂|I` | j∈I
t̂
≤
2b p
4d X
12b
dn ln(8/β) +
ηj +
t
t
t
`
j∈I`
r
32
n log( ).
β
(3)
Finally, by Lemma 5.7 (tail bound for sum of i.i.d. samples from the Laplace
distribution), with
q
P
β
n
4b
8d
probability at least 1 − 4 , for every ` ∈ [d] we have that
j∈I` ηj ≤
d ln( β ). So,
16b
2b p
dn ln(8/β) +
(3) ≤
t
t
s
12b
8d
dn ln( ) +
β
t
s
r
30b
32
32d
).
n log( ) ≤
dn ln(
β
t
β
P
Overall, with probability at least 1−β we get that every coordinate of the vector yt̂ − 1t j∈D vj
q
dn ln( 32d
is at most 30b
t
β ) in absolute value. Hence, with probability at least 1 − β we have
y 1X
−
vj
t
t̂
j∈D
5.3
≤
30bd
t
2
s
n ln(
32d
).
β
Algorithm LDP-GoodCenter
We are now ready to present our local algorithm for the 1-cluster problem. As we mentioned in
the introduction, our algorithm can be used as a building block for privately approximating the
k-means clustering problem, by iterating the algorithm and finding (roughly) k balls that cover
most of the data points. To that end, during the ith iteration we will need the ability to exclude
from the computation the data points that were already covered in the previous iterations. This
is handled by an additional input to the algorithm, a predicate σ, denoting a portion of the data
universe to be excluded from the computation (meaning that if σ(x) = 1 then x is to be excluded).
Our main statement for the 1-cluster problem under LDP is the following.
Theorem 5.13. Fix a predicate σ : X d → {0, 1}. Let S = (x1 , . . . , xn ) ∈ (X d )n be a database
which is distributed across n players (each holding one point in X d ). Let t, β, be s.t.
r
!
n0.51 · d1/2
1
n0.67 · d1/3
1 1/3
t≥O
log(dn|X|) log( ) +
log(dn|X|) log( )
.
β
β
2/3
There exists an -LDP algorithm that identifies a center y and a radius r s.t. with probability at
least 1 − β,
16
1. r = O(ropt ), where ropt is the radius of a smallest ball containing at least t points from
Sσ = {x ∈ S : σ(x) 6= 1}.
2. The number of points from Sσ contained in the ball of radius r around y is at least t − ∆,
where
r
1 n1.01
1
n
· log(d|X|) log( ) log( ) .
∆=O
·
t1/2
β
β
The main part of our construction is algorithm LDP-GoodCenter which, assuming t of the
input points are contained in a ball of radius r (given as an input parameter) identifies a ball
containing ≈ t points. As the radius r is not known in advance, we execute O(log |X|) copies
of LDP-GoodCenter in parallel with exponentially growing values of r (one of which provides a
constant factor approximation to ropt ). Another alternative is to perform a binary search on r,
hence paying in increasing round complexity to log log |X| but saving on composition, and hence
accuracy.
Algorithm LDP-GoodCenter
Input: Radius r, target number of points t, failure probability β, privacy parameter .
Optional input: A predicate σ : X d → {0, 1} (otherwise set σ ≡ 0).
Tool used: Family H of (r, c · r, p, q)-locality sensitive hash functions mapping Rd to a universe U .
Setting: Each player j ∈ [n] holds a value xj ∈ X d . Define S = (x1 , . . . , xn ).
1. Sample a hash function h ∈ H mapping Rd to U .
2. Use Theorem 5.4 with 4 to identify a list L of length at most 32n1+b /t containing all values
t
u ∈ U s.t. |{x ∈ S : h(x) = u}| ≥ 16
· n−b .
q
3. Let Z = (z1 , . . . , zd ) be a random orthonormal basis of Rd , and denote p = 2rc ln( dn
β )/d.
√
Partition every axis into intervals of length p, denoted as I = {I1 , I2 , . . . , I d|X| }.
4. Randomly partition S into subsets S41 , . . . , S4d of size |S4i | = nd . For every basis vector zi ∈ Z,
use Theorem 5.4 with 4 to obtain for every pair (I, u) ∈ I×U an estimation ai (I, u) for
|{x ∈ S4i : σ(x) 6= 1 and h(x) = u and hx, zi i ∈ I}|.
5. For every basis vector zi ∈ Z and for every hash value u ∈ L, denote I(i, u) =
ˆ u) by extending I(i, u) by p to each direction
argmaxI∈I {ai (I, u)}, and define the interval I(i,
ˆ
(that is, I(i, u) is of length 3p).
6. For every hash value u ∈ L, let B(u) denote the box in Rd whose projection on every axis
ˆ u).
zi ∈ Z is I(i,
|L|
n
7. Randomly partition S into subsets S71 , . . . , S7 of size |S7` | = |L|
. That is, for every u ∈ L
u
we have a subset S7 ⊆ S. For every u ∈ L use algorithm LDP-AVG with 4 to obtain an
approximation ŷu for the average of {x ∈ S7u : h(x) = u and x ∈ B(u) and σ(x) 6= 1} .
8. Output {ŷu : u ∈ L}.
17
Observation 5.14. Algorithm LDP-GoodCenter satisfies -LDP.
We now proceed with the utility analysis of algorithm LDP-GoodCenter. We will assume the
existence of a family H of (r, cr, p=n−b , q=n−2−a )-sensitive hash functions mapping Rd to a universe
U , for some constants a > b, r > 0, and c > 1.
Lemma 5.15. Fix a predicate σ : X d → {0, 1}. Let S = (x1 , . . . , xn ) ∈ (X d )n be a database which
is distributed across n players (each holding one row), and let P ⊆ {x ∈ S : σ(x) = 1} be a set of t
input points which can be enclosed in a ball of radius r. Let LDP-GoodCenter be executed with the
family H and with a radius r and with parameters β = n−a /24 and t, s.t.
!
1
2
n 2 +b · d1/2
n 3 +b · d1/3
2/3
t≥O
(log(dn))
.
log(dn|X|) +
2/3
1+b
The algorithm outputs a set Y of 32·nt
vectors in Rd such that with probability at least n−a /4
there exists a vector y ∈ Y such that the ball of radius 3cr around y contains at least 1 point from
P.
Proof. First observe that, w.l.o.g., we can assume that the range U of every function in H is of
size |U | ≤ n3 . If this is not the case, then we can simply apply a (pairwise independent) hash
function with range n3 onto the output of the locally sensitive hash function. Clearly, this will not
decrease the probability of collusion for “close” elements (within distance r), and moreover, this
can increase the probability of collusion for “non-close” elements (at distance at least cr) by at
most n−3 = o(n−2−a ) = o(q).
Now recall that by the properties of the family H, for every x, y ∈ Rd s.t. kx − yk ≥ cr we have
that Prh∈H [h(x) = h(y)] ≤ q = n−2−a . Using the union bound we get
Pr [h(x) 6= h(y) for all x, y ∈ S s.t. kx − yk ≥ cr] ≥ (1 − n−a /2).
h∈R H
Let P ⊆ S denote the guaranteed set of t input points x ∈ S with σ(x) 6= 1 that are contained
in a ball of radius r, and let x ∈ P be an arbitrary point in P . By linearity of expectation, we have
that
E [|{y ∈ P : h(y) 6= h(x)}|] ≤ t(1 − p) = t(1 − n−b ).
h∈H
Hence, by Markov’s inequality,
t(1 − n−b )
Pr |{y ∈ P : h(y) 6= h(x)}| ≥
≤ 1 − n−a .
h∈H
1 − n−a
So,
1 − n−b
Pr |{y ∈ P : h(y) = h(x)}| ≥ t 1 −
≥ n−a .
h∈H
1 − n−a
Simplifying, for large enough n (specifically, for na−b ≥ 2) we get
t
Pr |{y ∈ P : h(y) = h(x)}| ≥ · n−b ≥ n−a .
h∈H
2
So far we have established that with probability at least n−a /2 over the choice of h ∈ H in
Step 1 the following events occur:
18
(E1 ) For every x, y ∈ S s.t. kx − yk ≥ cr it holds that h(x) 6= h(y); and,
(E2 ) There exists a hash value in U , denoted u∗ , such that |{y ∈ P : h(y) = u∗ }| ≥
t
2
· n−b .
Event (E1 ) states that if two points in S are mapped into the same hash value, then these
points are close. Event (E2 ) states that there is a “heavy” hash value u∗ ∈ U , such that “many” of
the points in P are mapped into u∗ . We proceed with the analysis assuming that these two events
occur.
Onstep 2, we identify a list L containing
all such “heavy” hash values u ∈ U . Assuming that
p
1
0.5+b
t ≥ O ·n
· log(n/β) log(1/β) , Theorem 5.4 ensures that with probability at least 1 − β
∗
we have that u ∈ L. We continue with the analysis assuming that this is the case.
On Step 3 we generate a random orthonormal basis Z. By Lemma 4.2, with probability at least
(1 − β), for every x,py ∈ S and for every zi ∈ Z, we have that the projection of (x − y) onto zi is
of length at most 2 ln(dn/β)/d · kx − yk. In particular, for every hash value u ∈ L we have that
the projection ofpSu , {x ∈ S : h(x) = u} onto every axis zi ∈ Z fits within an interval of length
at most p = 2rc ln(dn/β)/d. Recall that we divide (on step 3) every axis zi ∈ Z into intervals of
length p, denoted as I = {I1 , I2 , . . . }. Hence, for every axis zi ∈ Z and for every u ∈ U , we have
that the projection of Su onto zi is contained within 1 or 2 consecutive intervals from I.
On step 4 we partition S into d subsets S4i ⊆ S of size nd . By the Hoeffding bound, assuming
q
that t ≥ 2 · n0.5+b · 2d ln( 2d
β ), with probability at least 1 − β, for every i ∈ [d], we have that
−b
u∗ |
|S4i ∩ Su∗ | ≥ |S2d
≥ t·n4d . Recall that the projection of Su∗ onto every axis zi ∈ Z fits within
(at most) 2 consecutive intervals from I. Hence, for every axis zi ∈ Z, at least 1 interval from
−b
I contains
at least half of the points from
S4i ∩ Su∗ , i.e., at least t·n8d points. Therefore, for
q
1
t ≥ O 1 · n0.5+b · d · log( dn|X|
β ) log( β ) , Theorem 5.4 ensures that with probability at least 1−β,
for every zi ∈ Z we have that I(i, u∗ ) = argmaxi∈I {ai (I, u∗ )} (defined on step 5) contains at least
ˆ u∗ ) obtained by extending I(i, u∗ ) by p to each
one point from Su∗ , and hence, the interval I(i,
direction, contains (the projection of) all of the points from Su∗ (onto the ith axis). As a result,
ˆ u∗ ), contains
the box B(u∗ ), defined on step 7 as the box whose projection onto every axis i is I(i,
all of Su∗ . We continue with the analysis assuming that this is the case.
|L|
On step 7 we partition S into |L| subsets S71 , . . . , S7 . For every u ∈ L we then use S7u to obtain
a an estimation ŷu for the average yu of the points x ∈ S7u s.t. h(x) = u and x ∈ B(u). By the
2
∗
u∗ |
Hoeffding bound, with probability at least 1 − β, we have that |S7u ∩ Su∗ | ≥ |S2|L|
≥ 128·nt 1+2b . If
that is the case, then as B(u∗ ) is of side length p, Lemma 5.9 ensures that with probability at least
2/3
2
+b
1/3 · log( dn )
3
. We continue
1 − β we have that kyu∗ − ŷu∗ k2 ≤ cr, provided that t ≥ 310
·
n
·
d
2/3
β
with the analysis assuming that this is the case.
Observe that yu∗ is the average of (some of) the points in Su∗ , and that every two points in Su∗
are within distance cr from each other. Hence, we get that a ball of radius 2cr around yu∗ contains
all of Su∗ . In particular, as Su∗ contains at least some of the points from P (the guaranteed cluster
radius r with t input points from S), we have that the ball of radius 2cr around yu∗ contains at
least 1 point from P . Therefore, as kyu∗ − ŷu∗ k2 ≤ cr we get that a ball of radius 3cr around ŷu∗
contains at least 1 points from P .
−a
Overall, with probability at least n 2 − 6β we have that the output on step 8 contains at least
one vector ŷ s.t. the ball of radius 3cr around ŷ contains at least one point from P .
19
Algorithm LDP-1Cluster
Input: Radius r, target number of points t, failure probability β, privacy parameters , δ.
Optional input: A predicate σ : X d → {0, 1} (otherwise set σ ≡ 0).
Tool used: Algorithm LDP-GoodCenter with success probability n−a /4, and error factor (in the
radius) at most 3c.
Setting: Each player j ∈ [n] holds a value xj ∈ X d . Define S = (x1 , . . . , xn ).
√
1. Denote J = log( d|X|) and K = 4 · na · ln(1/β). Randomly partition S into R · T subsets
n
S1,1 , . . . , SR,T of size JK
each.
2. For every j ∈ [J] and every k ∈ [K] apply LDP-GoodCenter on the database Sj,k with the
2j
t
radius |X|
, the predicate σ, privacy parameter 2 , and the parameter t0 = 2JK
. Obtain a set
of L =
64·n1+b
t
vectors: Ŷj,k = {ŷj,k,1 , . . . , ŷj,k,L }.
3. Randomly partition S into J · K · L subsets S1,1,1 , . . . , SJ,K,L of size
n
JKL
each.
4. For every j ∈ [J], every k ∈ [K], and every ` ∈ [L], use Theorem 5.3 on the database Sj,k,`
with 2 to obtain an estimation ĉj,k,` for the number of points x ∈ Sj,k,` enclosed in a ball of
2j
radius 5c |X|
around ŷj,k,` , for which σ(x) 6= 1.
5. Let j 0 be the smallest number in [J] for which there exist (k 0 , `0 ) ∈ [K]×[L] s.t.
r
1
32n
224 n1+a
· 1/2 · log(d|X|) log( ) log(
).
JKL · ĉj 0 ,k0 ,`0 ≥ t −
β
β
t
j0
2
Return the vector t̂j 0 ,k0 ,`0 and the radius 5c |X|
.
20
We are now ready to prove Theorem 5.13. The proof is via the construction of algorithm
LDP-1Cluster.
Proof of Theorem 5.13. The privacy properties of algorithm LDP-1Cluster are immediate. We
now proceed with the utility analysis.
Let ropt be the radius of the smallest ball in X d containing at least t points from {x ∈ S :
σ(x) 6= 1}, and let P denote a set of t such points. Let j ∗ denote the smallest integer s.t. ropt ≤
j∗
∗
2j
|X| ,
and denote r = 2|X| .
On step 1 we partition
S into JK subsets S1,1 , . . . , SJ,K . By the Hoeffding bound, assuming
q
that t ≥ 8 · n
1+a
2
·
log(d|X|) log( β1 ) log( 8n
β ), with probability at least 1 − β, for every k ∈ [K] we
have that |Sj ∗ ,k ∩ P | ≥ t0 =
t
2JK .
That is, for every k ∈ [K], the database Sj ∗ ,k contains at least
j∗
points from P , which can be enclosed in a ball of radius 2|X| = r. Hence, by the guarantees of
algorithm LDP-GoodCenter (Theorem 5.15), for every k ∈ [K], with probability at least n−a /4 we
have that Ŷj ∗ ,k contains a vector ŷ s.t. the ball of radius 3cr around ŷ contains at least 1 point
from P , provided that
r
!
1
2
1
n 2 +a+b · d1/2
n 3 +a+b · d1/3
1 1/3
t≥O
log(dn|X|) log( ) +
log(dn|X|) log( )
.
β
β
2/3
t0
S
Therefore, as K = 4na ln( β1 ), we get that with probability at least (1 − β), the set k∈[K] Ŷj ∗ ,k
contains at least one vector, denoted ŷj ∗ ,k∗ ,`∗ , s.t. the ball of radius 3cr around it contains at least
1 point from P . We continue with the analysis assuming that this is the case. Observe that, as the
set P is of diameter (at most) 2r, we get that a ball of radius 5cr around ŷj ∗ ,k∗ ,`∗ contains all of P .
On step 3 we partition S into JKL subsets S1,1,1 , . . . , SJ,K,L . For every (j, k, `) ∈ [J]×[K]×[L],
we then use Theorem 5.3 to obtain an estimation ĉj,k,` for the number of points from Sj,k,` that
2j
are enclosed in the ball of radius 5c |X|
around ŷj,k,` . Let us denote the true number of points from
Sj,k,` in that ball as cj,k,` (Sj,k,` ). Similarly, let cj,k,` (S) denote the number of points from S in that
2j
ball, i.e., the number of points from S that are enclosed in the ball of radius 5c |X|
around ŷj,k,` .
By the Hoeffding bound, with probability at leastq
1 − β, for every (j, k, `) ∈ [J]×[K]×[L] we
have that |cj,k,` (S) − JKL · cj,k,` (Sj,k,` )| ≤ 32 ·
n1+a
t1/2
·
log(d|X|) log( β1 ) log( 32n
β ). In addition, by
the guarantees of Theorem 5.3, with probability at least 1q− β, for every (j, k, `) ∈ [J]×[K]×[L] we
n1+a
have that |JKL · cj,k,` (Sj,k,` ) − JKL · ĉj,k,` | ≤ 192
log(d|X|) log( β1 ) log( 7n
· t1/2 ·
β ), and hence, for
every (j, k, `) ∈ [J]×[K]×[L] we have
224 n1+a
|cj,k,` (S) − JKL · ĉj,k,` | ≤
· 1/2 ·
t
r
1
32n
log(d|X|) log( ) log(
).
β
β
We continue with the analysis assuming that this is the case.
S
We have already established the existence of a vector ŷj ∗ ,k∗ ,`∗ ∈ k∈[K] Ŷj ∗ ,k s.t. the ball of radius
j∗
5c 2|X| around it contains all of P , i.e., contains at least t points from S. That is, cj ∗ ,k∗ ,`∗ (S) ≥ t,
and so
r
224 n1+a
1
32n
JKL · ĉj ∗ ,k∗ ,`∗ ≥ t −
· 1/2 · log(d|X|) log( ) log(
).
(4)
β
β
t
21
Hence, as j 0 , k 0 , `0 is defined on step 5 as the minimal j for which Inequality (4) holds, we get that
j0
2
≤ 2ropt . That is, the radius returned at the end of the execution is at most
j 0 ≤ j ∗ , and hence, |X|
10c · ropt . In addition, as j 0 (from step 5) satisfies Inequality (4), we get that the number of points
from S contained in the ball of radius 10c · ropt around ŷj 0 ,k0 ,`0 is at least
r
448 n1+a
1
32n
t−
· 1/2 · log(d|X|) log( ) log(
).
β
β
t
6
Application to k-means clustering
For a set of centers C ⊆ Rd and a point x ∈ Rd we denote dist(x, C) = minc∈C kx − ck2 . For a
finite set S ⊆ Rd , the sum of squared distances is defined as
X
cost(S, C) =
dist2 (x, C).
x∈S
If C = {c} we denote cost(S, c) = cost(S, {c}) for simplicity. For a weighted set S = {(x1 , α1 ), . . . , (x` , α` )},
the weighted cost is
X
cost(S, C) =
α · dist2 (x, C).
(x,α)∈S
Definition 6.1 (k-means). Let S be a (weighted or unweighted) finite set of points in Rd . A set
C ∗ of k centers (points) in Rd is called k-means of S if it minimizes cost(S, C) over every such set
C.
Definition 6.2 (Approximated k-means). Let S be a (weighted or unweighted) finite set of points
in Rd . A set C of k centers in Rd is a (γ, η)-approximation for the k-means of S if
cost(S, C) ≤ γ · cost(S, C ∗ ) + η,
where C ∗ is a k-means of S. When η = 0 we omit it and say that C is a γ-approximation for the
k-means of S.
6.1
k-means clustering – the centralized setting
As we mentioned in the introduction, Feldman et al. [15] showed that private algorithms for the
1-cluster problem translate to private algorithm for approximating the k-means of the data. Specifically, they showed the following theorem:
Theorem 6.3 ([15]). Fix a domain X d and a number k ≥ 1. Assume the existence of an (, δ)differentially private algorithm that, for every n ≥ nmin , solves the 1-cluster problem (X d , n, t= 3n
8k )
n
0
with parameters (∆, w) and error probability β, where ∆ ≤ 8k . Then, for every δ > 0, there exists
an (0 , 4k log(n)δ+δ 0 )-differentially private algorithm such that the following holds. Given a database
S ∈ (X d )n , with probability at least 1−O(β ·k ·log(n)), the algorithm returns a (γ, η)-approximation
to the k-means of S, where
r
1
2
d
0
γ = O(w · k),
η = O(nmin ) · diam(X ),
= 8k log(n) log( 0 ) + 4k log(n)(e − 1).
δ
22
Using this transformation with the (private) algorithm of [21] for the 1-cluster problem, Feldman
et al. [15] obtained the following result.3
Theorem 6.4 ([15]). Fix β, , δ ≤ 1. There exists an (, δ)-differentially private algorithm that,
given a database S ∈ (X d )n , identifies with probability 1 − β a (γ,η)-approximation for the k-means
of S, where γ = O(k log(n)) and η = Õ 1 · d0.5 · k 1.5 · diam(X d ) .
Applying the transformation of Theorem 6.3 with our new algorithm for the 1-cluster problem,
we obtain the following result.
Theorem 6.5. Fix β, , δ ≤ 1. There exists an (, δ)-differentially private algorithm that, given a
database S ∈ (X d )n , identifies with probability 1 − β a (γ,
η)-approximation for the k-means of S,
1
· d0.51 · k 1.51 · diam(X d ) .
where γ = O(k) and η = Õ 1.01
6.2
k-means clustering – the distributed setting
Towards obtaining our LDP algorithm for k-means, we start by presenting a non-private algorithm
that compute an approximation for the k-means (in fact, this is a family of algorithms, as the
algorithm can make arbitrary decisions throughout the execution). Afterwards we will show that
(a successful run of) our LDP algorithm can be identified with a possible execution of the nonprivate algorithm, showing that our private algorithm indeed computes an approximation for the
k-means. The non-private algorithm we use was first introduced by [13], and was modified by [15]
to obtain the transformation from 1-cluster to k-means in the centralized model.
Theorem 6.6 ([13, 15]). Every execution of algorithm Non-Private-k-Mean on a database S ∈
(X d )n with parameters ν, w, tmin , returns a set C of k centers such that
cost(S, C) ≤ O(w2 k) · cost(S, C ∗ ) + O (νk log(n) + tmin ) · diam(X d ),
where C ∗ are the optimal k-means of the input database S.
The analysis of algorithm Non-Private-k-Mean is almost identical to the analysis of [13, 15].
The only modification is that we allowed some “slackness” in the choices for Gi and αi , which will
be helpful for our LDP variant of the algorithm. The proof of Theorem 6.6 appears in the appendix
for completeness.
p
Lemma 6.7. LDP-k-Mean satisfies (0 , δ)-LDP for 0 = 24k log(n) ln(1/δ) + 12k log(n)(e − 1).
Lemma 6.7 follows directly from Theorem 2.9 (advanced composition). We now present the
utility analysis.
Lemma 6.8. Let S = (x1 , . . . , xn ) ∈ (X d )n be a p
database which is distributed across n players
(each holding one point in X d ), and denote ν = 16
n · log(8/β).
Let algorithm LDP-k-Mean be executed with an algorithm for the 1-cluster problem with parameters w, ∆, tmin , such that tmin ≥ 9ν + 6∆. Except with probability at most O(β · k log n), the
output C is s.t. there exists an execution of algorithm Non-Private-k-Mean on the database S with
parameters ν and w and t0min ≤ 8k
3 · tmin + ν that results in the same output C.
3
) and 2O(log
For simplicity, the Õ notation in theorems 6.4 and 6.5 hides factors of log( kn
βδ
on the additive error η.
23
∗
|X|)
from the statement
Algorithm Non-Private-k-Mean [13, 15]
Inputs: A set of n input points S ∈ (X d )n and an integer k ≥ 1.
Parameter: ν, w, tmin .
Output: A set C of k centers.
1. B ← ∅.
2. While |S| > tmin do (let i denote the index of the current iteration)
(a) ni ← |S|.
(b) Let ri,opt denote the radius of a smallest ball that contains
(c) Let ci ∈
Xd
3ni
8k
points from S.
and Gi ⊆ S be s.t.
i. Gi is contained within a ball of radius ri ≤ w · ri,opt around ci ,
ni
i
ii. 4k
≤ |Gi | ≤ 3n
8k .
(d) S ← S \ Gi .
(e) Arbitrarily choose |Gi | − ν ≤ αi ≤ |Gi |.
(f) B ← B ∪ {(ci , αi )}.
3. C ← an O(1)-approximation to the k-means of B.
4. Return C.
Combining Lemmas 6.7 and 6.8 with our LDP algorithm for the 1-cluster problem, yields the
following theorem.4
Theorem 6.9. Fix β, , δ ≤ 1. There exists an (, δ)-LDP algorithm that, for a (distributed)
database S ∈ (X d )n , identifies with probability 1 − β a (γ, η)-approximation for the k-means of S,
where γ = O(k) and η = Õ 1 · n0.67 · d1/3 · k 0.5 · diam(X d ) .
Proof of Lemma 6.8. We say that an application of step 4a is successful if it identifies a center ci
and a radius ri s.t. the ball of radius ri around ci contains at least (ti − ∆) points from Si , and
furthermore, ri ≤ w · r̂i,opt (where r̂i,opt is the radius of a smallest ball enclosing ti points from Si ).
Similarly, we say that the applications of steps 4b,4c,4e,4f are successful if they result in estimations
satisfying the stated requirements. As every such application succeeds with probability (1 − β),
and as there are at most 4k log n iterations, we have that all of these applications succeed together
with probability at least 1 − O(β · k log n). We proceed with the analysis assuming that this is true,
in which case we say that the entire execution is successful.
The proof proceeds by showing that for every successful execution of algorithm LDP-k-Mean
there exists an execution of algorithm Non-Private-k-Mean that results in the same output C.
Hence, successful executions of algorithm LDP-k-Mean enjoy the utility guarantees of algorithm
Non-Private-k-Mean.
Fix a successful execution of algorithm LDP-k-Mean, denoted as Exec1 (this amounts to fixing
all of the randomness throughout the computation). We analyze the set of points that are excluded
4
For simplicity, the Õ notation in Theorem 6.9 hides factors of log( kn|X|
) from the additive error η.
βδ
24
Algorithm LDP-k-Mean
Inputs: Privacy parameters , δ, failure probability β, and desired number of centers k.
Oracle: LR Oracle access to a database S ∈ (X d )n .
Tool used: An -LDPns algorithm for the 1-cluster problem with parameters w, ∆, tmin .
Output: A set C of k centers.
1. B ← ∅.
2. Define σ : X d → {0, 1} where σ(x) ≡ 0. We represent σ as a subset of X d , i.e., σ = ∅.
p
n1 · log(8/β).
3. Set i = 1. Denote n1 = n and n̂1 = n. Also denote ν = 16
% We assume that tmin ≥ 9ν + 6∆. Otherwise set tmin = 9ν + 6∆.
4. While (n̂i >
8k
3 tmin )
and (i ≤ 4k log n) do
(a) Use the algorithm for the 1-cluster problem with the predicate σ and parameter ti =
to obtain a center ci and a radius ri .
3n̂i
8k
% Let Si denote the data of all users holding an input element x s.t. σ(x) 6= 1. With probability (1 − β)
we have that the ball of radius ri around ci contains at least (ti − ∆) points from Si . Furthermore,
ri ≤ w · r̂i,opt , where r̂i,opt is the radius of a smallest ball enclosing ti points from Si .
(b) Let bi denote the number of users holding an input element x s.t. σ(x) 6= 1 and kx−ci k2 ≤
ri . Let b̂i be a noisy estimation (satisfying -LDP, e.g., using Theorem 5.3) for bi , such
that w.p. (1 − β) we have that bi − ν ≤ b̂i ≤ bi .
(c) Let hi : X d → {0, 1}
hash function s.t. for every x ∈ X d we have that
n be a random
o
Pr[hi (x) = 1] = min 1, 5n̂i , p̂i .
16kb̂i
e i denote the set of
% It suffices to use a λ-wise independent hash function for λ = ln(1/β). Let G
input points x ∈ Si such that kx − ci k2 ≤ ri and hi (x) = 1. Using a standard tail bound for sum
of λ-wise independent random variables (see, e.g., Lemma 5.6), with probability (1 − β) we have that
p
e i | ∈ p̂i · bi ± 8bi ln(1/β).
|G
(d) Let σ ← σ ∪ {x ∈ X d : kx − ci k2 ≤ ri and hi (x) = 1}.
% That is, roughly a p̂i fraction of the points in the found ball are excluded from the rest of the
computation.
(e) B ← B ∪ {(ci , α̂i )}, where α̂i = b̂i · p̂i − ν.
% Let αi denote the number of users that were excluded on the last step. With high probability we will
have that αi − 3ν ≤ α̂i ≤ αi
(f) Set i ← (i + 1), and let ni denote the number of users that are not excluded from the
computation, i.e., ni = |{x ∈ S : σ(x) 6= 1}|. Let n̂i be a noisy estimation (satisfying LDP, e.g., using Theorem 5.3) for ni , such that w.p. (1−β) we have that ni −ν ≤ n̂i ≤ ni .
5. C ← an O(1)-approximation to the k-means of B.
6. Return C.
25
e i (that is, G
e i is the set of input points x ∈ Si
on step 4d of iteration i of Exec1 , denoted as G
e i satisfies
such that kx − ci k2 ≤ ri and hi (x) = 1). We next show that, for every i, the set G
conditions (i),(ii) of step 2c of algorithm Non-Private-k-Mean. For condition (i), recall that as
e i is contained within a ball of radius w · r̂i,opt around ci , where r̂i,opt
step 4a succeeded, the set G
i
is the radius of a smallest ball enclosing ti = 3n̂
8k points from Si . Let ri,opt denote the radius of a
3ni
e i is
smallest ball enclosing 8k points from Si . As n̂i ≤ ni we have that r̂i,opt ≤ ri,opt , and hence, G
contained within a ball of radius w · ri,opt around ci , satisfying condition (i).
Let bi and b̂i be as in step 4b of the ith iteration of Exec1 (that is, bi denotes the number of
users holding an input element x s.t. σ(x) 6= 1 and kx − ci k2 ≤ ri , and b̂i is a noisy estimation
satisfying bi − ν ≤ b̂i ≤ bi ). As the execution is successful, we have that
bi ≥
3n̂i
3ni 3ν
−∆≥
−
− ∆.
8k
8k
8k
(5)
Assuming that tmin ≥ 9ν + 6∆, we have that
ni ≥ n̂i ≥
8k
tmin ≥ 24kν + 16k∆.
3
(6)
Plugging this into Inequality (5) we get that
5ni
+ ν.
(7)
16k
o
n
Recall that on step 4c we define p̂i , min 1, 5n̂i . By Inequality (7) and by the fact that
bi ≥
16kb̂i
ni − ν ≤ n̂i ≤ ni and bi − ν ≤ b̂i ≤ bi , we get that
5ni
ν
5ni
ν
− ≤ p̂i ≤
+ .
16kbi bi
16kbi bi
Finally, as step 4c succeeded, we have that
p
p
e i | ≤ p̂i · bi + 8bi ln(1/β),
p̂i · bi − 8bi ln(1/β) ≤ |G
(8)
(9)
which in combination with Inequalities (8) and (6) yields:
ni
e i | ≤ 3ni .
≤ |G
4k
8k
(10)
e i satisfying conditions (i) and (ii) of algorithm
So, every iteration of Exec1 identifies a set G
Non-Private-k-Mean, and excludes it from future iterations. In addition, by the fact that bi −ν b̂i ≤
e i | − 3ν ≤ α̂i ≤ |G
e i |, as
bi and by Inequality (9), we get that the weight α̂i (from step 4e) satisfies |G
is required on step 2e of algorithm Non-Private-k-Mean. This shows that Exec1 has a matching
execution of algorithm Non-Private-k-Mean such that the set of centers B is identical throughout
both of the executions. We next show that the stopping condition of both executions can also be
unified.
e i | ≥ ni . That
We have already established that in every iteration i of Exec1 we have that |G
4k
ni
is, in every iteration, at least 4k points are excluded from S, and hence, the number of iterations
is at most 4k log n. Thus, the loop of step 4 only halts when n̂i ≤ 8k
3 · tmin (and not because of the
26
condition on the number of iterations). As we next explain, this can be identified with the stopping
condition in Non-Private-k-Mean for a slightly different value t0min .
Consider the iteration i in which LDP-k-Mean halts. In that iteration we had that n̂i ≤ 8k
3 · tmin
8k
and that n̂i ≥ ni − ν. Hence, ni ≤ 3 · tmin + ν. In addition, as ni < ni−1 , there exists a number
t0min such that ni ≤ t0min < ni−1 and t0min ≤ 8k
3 · tmin + ν. The execution Exec1 would remain intact
had we replaced the condition of the while loop with (ni ≥ t0min ).
This shows that Exec1 has a matching execution of algorithm Non-Private-k-Mean, denoted
Exec2 , such that both executions have exactly the same number of iterations, and furthermore, the
set of centers B is identical in Exec1 and in Exec2 on every step of the executions. In particular,
this is the case at the end of the executions, which results in the same output C.
References
[1] A. Andoni and P. Indyk. Near-optimal hashing algorithms for approximate nearest neighbor
in high dimensions. In 47th Annual IEEE Symposium on Foundations of Computer Science
(FOCS 2006), 21-24 October 2006, Berkeley, California, USA, Proceedings, pages 459–468,
2006.
[2] R. Bassily, K. Nissim, U. Stemmer, and A. Thakurta. Practical locally private heavy hitters.
arXiv.org, 2017.
[3] A. Beimel, K. Nissim, and U. Stemmer. Private learning and sanitization: Pure vs. approximate differential privacy. In APPROX-RANDOM, volume 8096 of Lecture Notes in Computer
Science, pages 363–378. Springer, 2013.
[4] M. Bellare and J. Rompel. Randomness-efficient oblivious sampling. In Proceedings of the 35th
Annual Symposium on Foundations of Computer Science, SFCS ’94, pages 276–287, Washington, DC, USA, 1994. IEEE Computer Society.
[5] M. Bun, K. Nissim, and U. Stemmer. Simultaneous private learning of multiple concepts. In
ITCS, pages 369–380. ACM, 2016.
[6] M. Bun, K. Nissim, U. Stemmer, and S. P. Vadhan. Differentially private release and learning
of threshold functions. In FOCS, pages 634–649, 2015.
[7] C. Dwork, K. Kenthapadi, F. McSherry, I. Mironov, and M. Naor. Our data, ourselves: Privacy
via distributed noise generation. In S. Vaudenay, editor, EUROCRYPT, volume 4004 of Lecture
Notes in Computer Science, pages 486–503. Springer, 2006.
[8] C. Dwork and J. Lei. Differential privacy and robust statistics. In M. Mitzenmacher, editor,
STOC, pages 371–380. ACM, 2009.
[9] C. Dwork, F. McSherry, K. Nissim, and A. Smith. Calibrating noise to sensitivity in private
data analysis. In TCC, volume 3876 of Lecture Notes in Computer Science, pages 265–284.
Springer, 2006.
[10] C. Dwork, M. Naor, O. Reingold, G. N. Rothblum, and S. P. Vadhan. On the complexity of differentially private data release: efficient algorithms and hardness results. In M. Mitzenmacher,
editor, STOC, pages 381–390. ACM, 2009.
27
[11] C. Dwork, G. N. Rothblum, and S. P. Vadhan. Boosting and differential privacy. In FOCS,
pages 51–60. IEEE Computer Society, 2010.
[12] U. Erlingsson, V. Pihur, and A. Korolova. Rappor: Randomized aggregatable privacypreserving ordinal response. In Proceedings of the 2014 ACM SIGSAC Conference on Computer
and Communications Security, CCS ’14, pages 1054–1067, New York, NY, USA, 2014. ACM.
[13] D. Feldman, A. Fiat, M. Sharir, and D. Segev. Bi-criteria linear-time approximations for
generalized k-mean/median/center. In Proceedings of the Twenty-third Annual Symposium on
Computational Geometry, SCG ’07, pages 19–26, New York, NY, USA, 2007. ACM.
[14] D. Feldman, M. Schmidt, and C. Sohler. Turning Big data into tiny data: Constant-size
coresets for ¡italic¿k¡/italic¿-means, PCA and projective clustering, pages 1434–1453.
[15] D. Feldman, C. Xiang, R. Zhu, and D. Rus. Coresets for differentially private k-means clustering and applications to privacy in mobile sensor networks. In Proceedings of the 16th
ACM/IEEE International Conference on Information Processing in Sensor Networks, IPSN
’17, pages 3–15, New York, NY, USA, 2017. ACM.
[16] W. Hoeffding. Probability inequalities for sums of bounded random variables. Journal of the
American Statistical Association, 58(301):13–30, 1963.
[17] P. Indyk and R. Motwani. Approximate nearest neighbors: Towards removing the curse
of dimensionality. In Proceedings of the Thirtieth Annual ACM Symposium on Theory of
Computing, STOC ’98, pages 604–613, New York, NY, USA, 1998. ACM.
[18] S. P. Kasiviswanathan, H. K. Lee, K. Nissim, S. Raskhodnikova, and A. Smith. What can we
learn privately? SIAM J. Comput., 40(3):793–826, 2011.
[19] F. McSherry and K. Talwar. Mechanism design via differential privacy. In FOCS, pages 94–103.
IEEE Computer Society, 2007.
[20] K. Nissim, S. Raskhodnikova, and A. Smith. Smooth sensitivity and sampling in private data
analysis. In STOC, pages 75–84. ACM, 2007.
[21] K. Nissim, U. Stemmer, and S. P. Vadhan. Locating a small cluster privately. In Proceedings
of the 35th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems,
PODS 2016, San Francisco, CA, USA, June 26 - July 01, 2016, pages 413–427, 2016.
[22] V. Shenmaier. The problem of a minimal ball enclosing k points. Journal of Applied and
Industrial Mathematics, 7(3):444–448, 2013.
[23] A. Thakurta and A. Smith. Differentially private feature selection via stability arguments, and
the robustness of the lasso. In S. Shalev-Shwartz and I. Steinwart, editors, COLT, volume 30
of JMLR Proceedings, pages 819–850. JMLR.org, 2013.
[24] A. Thakurta, A. Vyrros, U. Vaishampayan, G. Kapoor, J. Freudiger, V. Sridhar, and D. Davidson. Learning new words, Mar. 14 2017. US Patent 9,594,741.
[25] S. Vadhan. The Complexity of Differential Privacy. 2016.
[26] P. Vazirani and P. Rao. Lecture notes in combinatorial algorithms and data structures.
28
A
Missing proofs
Lemma A.1. Algorithm LSH-GoodCenter preserves (2, 2δ)-differential privacy.
Proof. Algorithm LSH-GoodCenter interacts with its input database on steps 2, 4a, 5, 6, and 7.
Step 2 invokes the algorithm from Theorem 2.6 which is ( 4 , 4δ )-differentially private, to obtain
L ⊂ U . For every u ∈ L we make (in Step 4a) d applications of the algorithm from Theorem 2.6
that interact only with Su (and no other points in S). By the advanced composition (Theorem 2.9)
this preserves ( 4 , 4δ )-differential privacy w.r.t. Su . As the sets Su and Su0 are mutually disjoint,
Step 4a preserves ( 2 , 2δ )-differential privacy overall. Similarly, Step 5 interacts with the data for
at most n times (applying Gaussian mechanism). As each application of the Gaussian mechanism
is ( 4 , 4δ )-differentially private, we get that Step 5 preserves ( 2 , 2δ )-differential privacy. Steps 6,
and 7 initialize and use Algorithm AboveThreshold, which is ( 4 , 0)-differentially private. Overall,
GoodCenter is (2, 2δ)-differentially private by application of simple composition (Theorem 2.8).
Lemma A.2. Suppose X1 , . . . , Xn are independent samples from Lap( 1 ). That is, every Xi has
probability density function f (x) = 2 exp(−|x|). Let X = X1 + · · · + Xn . For every t ≥ 0 it holds
that
·t
√
Pr[|X| ≥ t] ≤ 6 exp −
.
2n
Moreover, for every 0 ≤ t < 2n/ we have that
2 2
·t
Pr[|X| ≥ t] ≤ 2 exp −
.
4n
The proof of Lemma A.2 is standard. We include a proof here for completeness as we are unable
to find an appropriate reference.
Proof of Lemma A.2. By symmetry, it suffices to analyze Pr[X ≥ t]. Indeed, for every c > 0 we
have that
Pr[X ≥ t] = Pr[cX ≥ ct] = Pr ecX ≥ ect
n
≤ e−ct · E ecX = e−ct · E ecX1
.
(11)
Furthermore,
cX1
E e
=
=
=
=
Z ∞
cX1
1
Pr X1 ≥ ln z dz
Pr e
≥ z dz =
c
0
0
Z 1
Z ∞
1
1
Pr X1 ≥ ln z dz +
Pr X1 ≥ ln z dz
c
c
0
1
Z 1
Z ∞
· ln z
1
· ln z
1
1 − exp
dz +
exp −
dz
2
c
2
c
1
0
Z
Z
1 1 /c
1 ∞ −/c
1
1−
z dz +
z
dz =
,
2 0
2 1
1 − (c/)2
Z
∞
where the last equality follows by asserting that /c > 1 and solving the integrals. Plugging into
Equality (11) we get that
n
c 2
Pr[X ≥ t] ≤ e−ct · E ecX1
= e−ct · (1 − (c/)2 )−n ≤ e−ct · e2n·( ) ,
29
where the last inequality holds whenever /c ≥ 1.2. The lemma now follows by choosing c =
√
(in which case /c ≥ 2 for every t ≤ 2n/) or by choosing c = / 2n.
B
C∗
Si
Si∗
2 t
4n
Proof of Theorem 6.6 [13, 15]
A set of k centers that minimizes cost(S, C) over the original database S over every such C.
The database S during the ith iteration.
The 3|Si |/4 points x ∈ Si with smallest value dist(x, C ∗ ).
Table 2: Notations for the analysis of algorithm Non-Private-k-Mean
Lemma B.1. In every iteration i we have that cost(Gi , ci ) ≤ w2 · cost(Si∗ , C ∗ )
|S ∗ |
Proof. By the pigeonhole principle there must be a center x ∈ C ∗ that serves m ≥ ki points in
∗ have x as their closest center in C ∗ . At least half
Si∗ , i.e., at least 1/k fraction of the points in Siq
2cost(Si∗ ,C ∗ )
|S ∗ |
of them (i.e., 2ki ) have distance at most r̃ =
to x, as otherwise we would get that
m
cost(Si∗ , C ∗ ) >
|Si∗ |
2k
·
2cost(Si∗ ,C ∗ )
m
≥ cost(Si∗ , C ∗ ).
We conclude that there is a ball of radius r̃ that contains at least m/2 ≥
from Si . By definition of ri,opt , we have that
r
2cost(Si∗ , C ∗ )
ri,opt ≤ r̃ =
.
m
That is,
|Si∗ |
2k
=
3|Si |
8k
points
m 2
r
≤ cost(Si∗ , C ∗ ).
2 i,opt
Hence,
cost(Gi , ci ) ≤ |Gi |·(w·ri,opt )2 ≤
3|Si |
|S ∗
m 2
·(w·ri,opt )2 = i ·(w·ri,opt )2 ≤ w2 · ri,opt
≤ w2 ·cost(Si∗ , C ∗ ),
8k
2k
2
where the first and second inequalities follow from the conditions on Gi on step 2(c) of the algorithm.
Lemma B.2.
|B|
X
i=1
cost(Gi , ci ) ≤ O(w2 · k) · cost(S, C ∗ ).
Proof. We order the points in S by S = (x1 , x2 , . . . , xn ), such that dist(xj1 , C ∗ ) ≤ dist(xj2 , C ∗ ) for
every j1 < j2 , where ties are broken arbitrarily. Let
Ui = {x1 , x2 , . . . , xn−|Si | },
Vi = {xn−|Si |+1 , . . . , xn−|Si |+|Si∗ | }.
30
Clearly, |Ui ∪ Vi | = |Ui | + |Vi | (as the sets Ui and Vi are disjoint). Moreover, as there are at most
(n − |Si |) elements outside of Si , we have that
|(Ui ∪ Vi ) ∩ Si | ≥ |Ui | + |Vi | − (n − |Si |)
= |Vi | = |Si∗ |.
So, the sets (Ui ∪ Vi ) and Si have at least |Si∗ | elements in common. As Si∗ contains the values
x ∈ Si with smallest distance dist(x, C ∗ ), and as (Ui ∪ Vi ) contain all such elements from S, we get
that these common elements contain Si∗ , that is, Si∗ ⊆ (Ui ∪ Vi ).
The set Vi contains the |Si∗ | points x ∈ (Ui ∪ Vi ) with the largest values dist(x, C ∗ ). Hence,
cost(Si∗ , C ∗ ) ≤ cost(Vi , C ∗ ). Combining this with Lemma B.1, and summing over every i we get
|B|
X
i=1
cost(Gi , ci ) ≤
|B|
X
i=1
w2 · cost(Vi , C ∗ ).
We now want to relate the right hand side of the above inequality to the input points S.
Specifically, we want to show that every point x ∈ S is included in at most O(k) sets Vi . To that
end, observe that:
1. The index of the last point in Vi is n − |Si | + |Si∗ | = n −
|Si |
4 .
2. The index of the first point in Vi+12k is n − |Si+12k | + 1 ≥ n − |Si | · (1 −
1 6k
4k )
>n−
|Si |
4 .
Hence, every point appears in O(k) sets in the sequence V1 , V2 , . . . . Hence,
|B|
X
i=1
cost(Gi , ci ) ≤
|B|
X
i=1
w2 · cost(Vi , C ∗ ) ≤ O(w2 · k) · cost(S, C ∗ ).
Lemma B.3.
√
√
cost(S, C) ≤ O(w2 k) · cost(S, C ∗ ) + O tmin · d + νk log(n) d .
Proof. Recall that every iteration defines αi as |Gi | − ν ≤ αi ≤ |Gi |. For every i, let Ĝi ⊆ Gi denote
an arbitrary subset of size αi . Let Ŝe = ∪Ĝi .
We use the weak triangle inequality [14] stating that for every x, y ∈ Rd and a closed set C ⊆ Rd
we have
12kx − yk22 λ
|dist2 (x, C) − dist2 (y, C)| ≤
+ dist2 (x, C).
λ
2
31
In our case, by letting x0 denote the associated ci to x ∈ Ŝe when x is deleted, and using λ = 1/2,
this implies
|cost(Ŝe , C) − cost(B, C)| =
≤
X
x∈Ŝe
dist2 (x, C) −
X
dist2 (x0 , C)
x∈Ŝe
X 12kx − x0 k2 λ
2
2
+ dist (x, C)
λ
2
x∈Ŝe
= 24
|B|
X
i=1
1
cost(Ĝi , ci ) + cost(Ŝe , C).
4
(12)
Hence,
|B|
X
3cost(Ŝe , C)
≤ cost(B, C) + 24
cost(Ĝi , ci ).
4
i=1
By the fact that C is an O(1)-approximation to the k-means of B, we get
cost(B, C) ≤ O(1) · cost(B, C ∗ ).
Similarly to (12),
∗
∗
|cost(Ŝe , C ) − cost(B, C )| ≤ 24
|B|
X
i=1
1
cost(Ĝi , ci ) + cost(Ŝe , C ∗ ).
4
Hence,
|B|
X
5
cost(B, C ) ≤ cost(Ŝe , C ∗ ) + 24
cost(Ĝi , ci ).
4
∗
i=1
So,
cost(Ŝe , C) = O cost(B, C) +
|B|
X
i=1
= O cost(B, C ∗ ) +
= O cost(B, C ∗ ) +
cost(Ĝi , ci )
|B|
X
i=1
|B|
X
i=1
cost(Ĝi , ci )
cost(Gi , ci )
= O cost(Ŝe , C ∗ ) + w2 k · cost(S, C ∗ )
= O(w2 k) · cost(S, C ∗ ).
Notice that cost(S, C) = cost(Ŝe , C) + cost(S \ Ŝe , C), where S \ Ŝe ≤ ν · |B| + tmin . As there
are at most O(k · log(n)) iterations (i.e., |B| = O(k · log(n))), we have that
32
cost(S, C) = cost(Ŝe , C) + cost(S \ Ŝe , C)
≤ O(w2 k) · cost(S, C ∗ ) + O (νk log(n) + tmin ) · diam(X d ).
33
| 8cs.DS
|
Analysis of a chaotic spiking neural model: The
NDS neuron
arXiv:1408.3735v1 [cs.NE] 16 Aug 2014
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
Abstract Further analysis and experimentation is carried out in this paper for a
chaotic dynamic model, viz. the Nonlinear Dynamic State neuron (NDS). The analysis and experimentations are performed to further understand the underlying dynamics of the model and enhance it as well. Chaos provides many interesting properties
that can be exploited to achieve computational tasks. Such properties are sensitivity to initial conditions, space filling, control and synchronization.Chaos might play
an important role in information processing tasks in human brain as suggested by
biologists. If artificial neural networks (ANNs) is equipped with chaos then it will
enrich the dynamic behaviours of such networks. The NDS model has some limitations and can be overcome in different ways. In this paper different approaches are
followed to push the boundaries of the NDS model in order to enhance it. One way
is to study the effects of scaling the parameters of the chaotic equations of the NDS
model and study the resulted dynamics. Another way is to study the method that is
used in discretization of the original Rössler that the NDS model is based on. These
approaches have revealed some facts about the NDS attractor and suggest why such
a model can be stabilized to large number of unstable periodic orbits (UPOs) which
might correspond to memories in phase space.
Mohammad Alhawarat
Department of Computer Science, College of Computer Engineering and Sciences,
Salman Bin Abdulaziz University, Kingdom of Saudi Arabia,
Tel.: +966-1-1588371, Fax: +966-1-1588302, e-mail: [email protected]
Waleed Nazih
Department of Computer Science, College of Computer Engineering and Sciences,
Salman Bin Abdulaziz University, Kingdom of Saudi Arabia,
Tel.: +966-1-1588371, Fax: +966-1-1588302, e-mail: [email protected]
Mohammad Eldesouki
Department of Information Systems, College of Computer Engineering and Sciences,
Salman Bin Abdulaziz University, Kingdom of Saudi Arabia,
Tel.: +966-1-1588371, Fax: +966-1-1588302, e-mail: [email protected]
1
2
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
1 Introduction
Chaos might play an important role in information processing tasks in human brain
as shown in [3, 8, 10, 11, 12, 17, 19, 20]. Some properties that might be useful
for information processing tasks are: sensitivity to initial conditions, space filling,
control, synchronization and a rich dynamics that can be accessed using different
control methods. In theory, if Artificial Neural Networks (ANN) are equipped with
chaos they will enable a large number of rich dynamic behaviors. After applying
control, these dynamics can be accessed using one of the control mechanisms such
as feedback control [14, 15, 16]. Applying such control mechanisms to discrete
chaotic neural models showed that the model would stabilize into one of many UPOs
that are embedded in the chaotic attractor.
Different chaotic neural models have been devised in recent years to explore the
possibilities of exploiting the rich dynamics that such models might provide for
information processing tasks. One of these model is the NDS model [5]. This model
is based on Rössler system [18].
Rössler is a simple chaotic system. It has been studied many times in terms of
control investigation and biological studies [9, 4] to name a few.
The origins of this model dates back to 2003 where the authors in[7] have proposed a chaotic neuron that is based on Rössler system. The idea was to exploit
the rich dynamics of the chaotic attractor to represent internal states and therefore
the chaotic attractor can represent an infinite state machine. Many experiments have
been carried out to show that using periodic input signals would cause the chaotic
attractor to stabilize to an UPO. The control mechanism used was a modified version of Pyragas[16] where the period length is considered a system variable. Small
networks of 2 − 3 neurons have been studied and the network has stabilized to one
UPO according to a periodic length that is implicitly appears in the input pattern.
The model is very interesting due to the fact that it theoretically allows an access
to a large number of UPOs, which correspond to memories in phase space, using
only single NDS neuron. In contrast, the Hopfield neural network can give only
0.15n memory size (where n is the number of neurons in the network).
The NDS model is studied in a series of works [5, 13, 1, 6, 2]. The authors in[13]
have used Lorenz attractor instead of Rössler. They have used transient computation machine to detect human motion from sequences of video frames. In another
paper[6] the authors argued that chaos may equip mammalian brain with the equivalent of kernel trick for solving hard nonlinear problems.
In [2] networks of NDS neurons have been investigated in the context of Spike
Time Dependent Plasticity (STDP) which is a property of cortical neurons. The
author has suggested that NDS neurons may own the realism of biological neural
networks; this has been supported by experiments conducted by the author.
The NDS model has been investigated thoroughly in[1]. In his investigation, the
author has studied the chaotic behavior of the model from both experimental and
analytical perspectives. Explanation of the behavior of the model has become clear
after the experimentations and the mathematical analysis and the study has shown
interesting results.
Analysis of a chaotic spiking neural model: The NDS neuron
3
In this paper some of the limitations that exist in the NDS model will be investigated. This includes tuning the model parameters for the sake of enhancing the
model capacity in terms of the number of successfully stabilized UPOs. Moreover,
the discretization method that used to to convert the continuous Rössler system into
the discrete NDS model will be discussed and compared to other well-known methods of discretization.
The paper is organized as follows: in section 2 the original Rössler model is introduced, in section 3 the NDS model is described, in section 4 a mathematical analysis of the NDS model and related discretization methods are discussed, section 5
is devoted to describe the experimentation setups that are carried out to tune the
parameters of the NDS model, section 6 includes discussions and finally section 7
concludes the paper.
2 Rössler Chaotic attractor
The Rössler system [18] is a simple dynamical system that exhibits chaos and has
only one nonlinear term in its equations. Rössler built the system in 1976; it describes chemical fluctuation and is represented by the following differential equations:
x′ = −y − u
(1)
y′ = x + a ∗ y
(2)
z′ = b + z(x − c)
(3)
Where a and b are usually fixed and c is varied and is called the control parameter.
The familiar parameter settings for the Rössler attractor are a = 0.2, b = 0.2, and
c = 5.7, and the corresponding attractor is shown in figure 1. Note that the Rössler
attractor most of the time lies in the x − y plane and comprises a disk that has a dense
number of orbits. Note also that these orbits are stretching as a result of divergence
and sensitivity to initial conditions. From time to time the Rössler attractor rises in
the z direction and then folds back to the disk which forms a fin-like shape. The
folding and stretching keep the Rössler attractor bounded in phase space.
If a trajectory of a chaotic system evolved starting from an initial point within the
attractor for a long period of time, then that trajectory will fill a bounded part of the
phase space and the attractor of the system will have a fractional dimension. This
bounded space is one of the properties of chaos, and is due to the attracting and the
repelling of the trajectory by the fixed points that govern and organize the system
behaviour. The type of these points determines the shape of the resulting attractor.
This fractal dimension can be recognized in figure 1 where the attractor is not filling
the whole space, instead it is filling part of the space.
4
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
25
20
z
15
10
5
0
10
15
0
10
5
−10
y
0
−20
−5
−10
x
Fig. 1 The Rössler chaotic attractor with parameters a = 0.2, b = 0.2, and c = 5.7.
Unstable periodic orbits
An UPO is one of the dynamic behaviours that a nonlinear system exhibit in phase
space. An UPO is a repeating orbit and is unstable as a result of being attracted
and repelled by fixed points of an attractor. Chaotic attractors usually have a dense
number of UPOs which can be accessed using controlling methods.
For example figure 2 shows two UPOs of period 1 and period 2 in the Rössler
attractor when the c parameter is tuned to 2.5 and 3.5 respectively. Period 2 here
means that the UPO repeat twice in the attractor.
3 Describing the NDS model
In[5], Crook et al. have proposed a chaotic spiking neuron model that is called the
NDS neuron. The NDS neuron is a conceptual discretized model that is based on
Rössler’s chaotic system [18]. The NDS model is a modified version of Rössler’s
equations as described by equations 1- 3 in section 2.
By varying the system parameters such as period length τ , connection time delays
and initial conditions, large number of distinct orbits with varying periodicity may
be stabilised.
The NDS model simulates a novel chaotic spiking neuron and is represented by:
Analysis of a chaotic spiking neural model: The NDS neuron
5
25
Main attractor
Period one
Period two
20
z
15
10
5
0
10
15
0
10
5
−10
0
−20
y
−5
−10
x
Fig. 2 Two UPOs, one of period 1 and the other of period 2 in the Rössler chaotic attractor occurred
when parameters a, b are fixed to 0.2 and parameter c is set to 2.5 and 3.5.
1
0
0.5
γ(t)
−0.5
0
−1
u(t)
u(t)
−0.5
−1
−1.5
−1.5
−2
−2
−2.5
9500
9600
9700
9800
9900 10000
−2.5
−0.4
t
−0.2
0
x(t)
0.2
(a) Time series
(b) phase plot: u(t) vs x(t)
0.4
Fig. 3 The chaotic behaviour of a NDS neuron without input (a) the time series of u(t) and γ (t),
and (b) the phase space of x(t) versus u(t)
x(t + 1) = x(t) + b(−y(t) − u(t))
(4)
y(t + 1) = y(t) + c(x(t) + ay(t))
(5)
6
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
u(t + 1) =
η0
u(t) > θ
u(t) + d(v + u(t)(−x(t)) + ku(t)) + F(t) + In(t) u(t) ≤ θ
(6)
n
F(t) =
∑ w j γ (t − τ j )
(7)
j=1
n
In(t) =
∑ I j (t)
(8)
1 u(t) > θ
0 u(t) ≤ θ
(9)
j=1
γ (t + 1) =
where x(t), y(t) and u(t) describe the internal state of the neuron, γ (t) is the neuron’s binary output, F(t) represent the feedback signals, In(t) is the external input
binary spike train, and the constants and parameters of the model are: a = v = 0.002,
b = c = 0.03, d = 0.8, k = −0.057, θ = −0.01, η0 = −0.7 and τ j is the period
length of the feedback signals.
1
0.5
0
0
u(t)
u(t)
γ(t)
0.5
−0.5
−0.5
−1
−1
−1.5
9500 9600 9700 9800 9900 10000
t
(a) Time series
−1.5
−0.2
−0.1
0
x(t)
0.1
0.2
Phase plot: u(t) vs x(t)
Fig. 4 The stabilizing of period-4 orbit of a NDS neuron with feedback connection.
The NDS model is a discrete version of the Rössler system. The main reason
to have a discrete version of the Rössler system is because spikes should occur in
discrete time. The discretization has been carried out by scaling the system variables x(t), y(t) and u(t) using different scaling constants: b, c, d. The values of these
constants have been tuned experimentally until the dynamics of the Rössler system
are preserved. If the values of these constants are large, then a system trajectory will
miss many dynamic evolutions while moving from one discrete iteration to the next.
Therefore, the time steps of the discrete system need to be very small so that a sys-
Analysis of a chaotic spiking neural model: The NDS neuron
7
tem trajectory will involve most of the dynamic evolutions of the Rössler system.
This will become clear in section 4.
The dynamics of a single NDS neuron without input is shown in Figure 3. However, when the NDS neuron is equipped with a time delayed feedback connection
then the firing pattern of the neuron can become periodic. This is shown in figure 4
where period-4 orbit is stabilised due to the feedback control mechanism F.
4 Discretization method
There are two kinds of discretization methods when converting a continuous system into a discrete one: standard and nonstandard methods. One of the nonstandard
methods is Eulers Forward differentiation method that is usually used in developing
simple simulation systems. To convert a continuous system into the corresponding
discrete one using this method, then a time step T S is used to approximate the next
value of a continuous system variable that it will evolve to. For example, equation1
when converted into a discrete equation using Eulers Forward differentiation it will
become:
x(tk+1 ) − x(tk )
= −y(tk ) − u(tk )
TS
and then solving for x(tk+1 ) gives:
ẋ(tk ) =
x(tk+1 ) = x(tk ) + T S(−y(tk ) − u(tk ))
(10)
(11)
After using the simple notation x(t) instead of x(tk ), the equation becomes:
x(t + 1) = x(t) + T S(−y(t) − u(t))
(12)
if this equation is compared with equation 4, it is obvious that the T S is chosen
to be parameter b = 0.03.
In order for the discretized function to behave similarly to the continuous one
then T S should be chosen to be small. For simulation purposes it is preferable to
choose T S according to the following formula:
TS ≤
0.1
|λ |max
(13)
Where |λ |max is the largest absolute eigenvalues for the Rössler system. According to the mathematical analysis that is carried out in [1], |λ |max = 5.68698. When
substituting this value in equation 14 it becomes:
T S ≤ 0.0176
(14)
If this compared with b it is obvious that the time step that has been chosen
doesn’t follow the simulation preferable setup.
8
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
Moreover, if we look at equation 5 not only the T S is chosen to be c = 0.03 but
also a scaling factor is used to reduce the y down by the factor a = 0.002.
For the third variable in Rössler system, viz. z variable, there were many changes
because the authors of the model according to [5] wanted to invent a spiking model
that is based on a threshold variable. That variable was u which corresponds to z
in the original Rössler system. In addition, the time step T S value that is used in
the discretization process was different from those appear in equations 4-5. While
the T S = b = c = 0.03 used in the aforementioned equations, T S is chosen to be
T S = d = 0.8 in discretization of the variable u. Moreover, the authors have scaled
the constant c from 5.7 down to 0.057 and changed its sign to negative. A final
change was made to the sign of x variable from positive to negative.
All these changes made the new system to behave differently in phase space. To
summarize, the new attractor of the system has different fixed points types. According to the mathematical analysis carried out by [1], the original Rössler system fixed
points, which are two spiral saddle points, have become two spiral repellors due to
the varying scaling factors used and the change of the sign for both k and x.
These results assure that the NDS model, although has a promising results as a
spiking chaotic neuron model, it doesn’t have a strong connection to the properties
of the original Rössler attractor. This is made obvious in [1] when they concluded
that the existence of the UPOs of the NDS attractor is due to the acting forces of the
two spiral repellors and the reset mechanism. Because without the reset mechanism,
the two spiral repellors will enforce any trajectory that starts near by to evolve away
from both of them and approaches infinity.
If the Eulers Forward differentiation discretization method is used to convert the
continuous Rössler system into a discrete model, where the time step is chosen to
be T S = 0.0055, then equations 1- 3 will become:
x(t + 1) = x(t) + T S(−y(t) − u(t))
(15)
y(t + 1) = y(t) + T S(x(t) + ay(t))
(16)
u(t + 1) = u(t) + T S(b + z(t)(x(t) − c))
(17)
Where T S = 0.0055, a = 0.2, b = 0.2, and c = 5.7.
To verify if such settings will preserve the shape and properties of the original
Rössler system, then an experiment is carried out to depict the discretized system’s
attractor as shown in figure 5.
If these equations are converted into the equations of the NDS model then different steps need to be followed:
1.
2.
3.
4.
5.
Specify the value of T S, here T S is chosen to be 0.0055
Change the parameter b in equation 17 to v.
Change the parameter c in equation 17 to k.
Change the constant T S in equation 15 to b and give it the value of T S.
Change the constant T S in equation 16 to c and give it the value of T S.
Analysis of a chaotic spiking neural model: The NDS neuron
9
25
20
u(t)
15
10
5
0
10
15
0
10
5
−10
y(t)
0
−20
−5
−10
x(t)
Fig. 5 The Discrete version of Rössler attractor based on equations 15- 17
6.
7.
8.
9.
10.
Change the constant T S in equation 17 to d and give it the value of T S.
Change the value of a and v and give it the value of 0.2.
Change the value of k to 5.7
Change the variable z to u.
Change the sign of the term (x(t) − k) in equation 17 to become (−x(t) + k)
After applying the previous changes the equations 15- 17 become:
x(t + 1) = x(t) + b(−y(t) − u(t))
(18)
y(t + 1) = y(t) + c(x(t) + ay(t))
(19)
u(t + 1) = u(t) + d(v + u(t)(−x(t) + k))
(20)
Where a = 0.2, b = c = d = 0.0055, and k = 5.7.
Now this new model need to be verified, i.e. an experiment need to be carried out
to depict the attractor such equations will produce. Figure 6 depicts the results of
evolution of equations 18- 20.
Note that the original Rössler attractor has disappeared and this is due to the
change of sign that is made to the term (x(t) − k) in equation 20.
10
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
Fig. 6 The attractor of the model that is based on equations 18- 20
5 Tuning the parameters of the NDS model
In this research, many experiments have been carried out to tune the parameters
of the NDS model. These experiments are carried out by varying one parameter
and fix all other parameters to their default values that are used in the NDS model
definition as stated in section 3. The following parameters have been considered
in these experiments: a, v, b, c, d and k because the other parameters: θ , η0 and τ j
have already been studied thoroughly in [1]. One NDS neuron has been used in the
experiments setup and random initial conditions are chosen for the values of the
variables: x, y and u. After 1000 iterations the feedback control is applied and the
experiment runs for another 9000 iterations.
To decide whether a specific setting is valid; the values of the variables of the
model are recorded and then depicted in phase space. If an attractor exist, then this
setting is considered in the valid ranges for the values of the system parameters.
Table 1 Parameter value’s ranges
Parameter
Range
a, v
b, c
d
k
0.001 − 0.1
0.01 − 0.055
0.8 − 0.9
−(0.055 − 0.58)
Analysis of a chaotic spiking neural model: The NDS neuron
11
To summarize the results; the valid ranges for the model parameters are shown
in table 1.
Table 2 Parameter settings with different selected values from the ranges appear in table1
Parameter
a, v
b, c
d
k
Setup 01
Setup 02
Setup 03
Setup 04
Setup 05
Setup 06
Setup 07
Setup 08
Setup 09
Setup 10
Setup 11
Setup 12
Setup 13
Setup 14
Setup 15
0.001
0.01
0.1
0.002
0.002
0.002
0.002
0.002
0.002
0.002
0.002
0.002
0.01
0.002
0.1
0.03
0.03
0.03
0.001
0.02
0.05
0.03
0.03
0.03
0.03
0.03
0.03
0.05
0.015
0.04
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.85
0.9
0.8
0.8
0.8
0.85
0.8
0.8
−0.057
−0.057
−0.057
−0.057
−0.057
−0.057
−0.057
−0.057
−0.057
−0.055
−0.056
−0.058
−0.055
−0.058
−0.056
1000
Average Number of successfully
stabilised orbits for 1000 runs
900
800
700
600
500
400
300
200
100
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Parameter settings
Fig. 7 Average Stabilized UPOs over 1000 run based on the parameter settings that appear on
table 2.
To judge weather such ranges could enhance the capacity of the attractor in terms
of the number of UPOs that it might encompass, then another experiment setup is
12
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
used. Here the average number of successfully stabilized UPOs is computed over
1000 run for different parameter settings according to table 2 and the results are
depicted in figure 6.
Note that the original parameters values of the NDS model is used in setup 7. This
is done to compare the capacity of the attractor with different parameter settings
compared to the original parameter settings.
The results that are shown in figure 6 suggest that there exist better parameter
settings such as Settings 14 and Settings 05, but these will not make a significant
enhancement in the capacity of the attractor in terms of successfully stabilized UPOs
when compared to the NDS original settings (Setting 07).
Based on the results that have been attained in section 4, if a parameter setting
need to be chosen to represent the Rössler system then first a T S need to be set
according to equation 14, then a mapping between the original Rössler equations
and the NDS equations need to be performed. If T S is chosen to be 0.015, then after
carrying out the mapping between the equations of both systems; the parameter
settings become as listed in table 3
Table 3 Parameter settings for the original Rössler as discussed in section 2
Parameter
a, v
b, c
d
k
Parameter value
0.2
0.015
0.015
5.7
This setting is considered in another experiment setup, and the result is that there
is no single UPO that can be stabilized. One possible reason is that the fixed points
of the system is reserved in this case and not changed as the case in the NDS model
settings. The change in sign in the term (x − c) and the different nested scaling
factors that are applied to the original Rössler equation have led to change in the
types and properties of the fixed points as shown in [1].
This can be proved easily because when multiplying an equation with a constant
and then trying to find its roots, then it should be set to zero. Therefore, the constant
will have no effect on the resulted roots as dividing both sides by constant will
eliminate the constant from the left hand side and will not affect zero in the right
hand side.
6 Discussion
The results of both experimentations and mathematical analysis of the discretization
method suggest that the NDS model has weak connections to the original Rössler
system. This is due to many factors. Firstly, the discretization process that is used in
devising the NDS equations does not follow any known discretization method where
different discretization time steps and scaling factors are used. Secondly, Changing
the sign of the term (x(t) − k) that appears in equation 20 made the system attractor
Analysis of a chaotic spiking neural model: The NDS neuron
13
to disappear and approach infinity instead. These changes affected both the shape
and properties of the original Rössler attractor.
It is important here to stress that even before adding the feedback signal, the input
signals and the reset mechanism to the system equations, the attractor has become
completely different from the original Rössler attractor. This assure that the NDS
model has weak connections to the Rössler model and has different fixed points,
eigenvalues and eigenvectors as demonstrated in [1].
The results also suggest that both the reset mechanism and the feedback signal
are the major ingredients for the NDS model to work and to be stabilised to one of
its available UPOs.
The results of experimental setups made to tune the NDS parameters suggest that
there exist better parameter settings but will not enhance the capacity of the attractor
in terms of successfully stabilized UPOs significantly when compared to the NDS
original settings.
Also, the settings of the Rössler model when used in the NDS model resulted in
no stabilized UPOs because the discretization method that is used to build the NDS
model has led to changes in the shape and properties of the Rössler attractor.
It is important to mention that the main factors that affect the shape and properties
of the original Rössler attractor are both scaling the parameters of the model with
different values and the change in sign that is made to the term (x(t)−k) that appears
in equation 20.
7 Conclusion
One chaotic model, viz., the NDS model has been studied in this paper. NDS is one
of different chaotic models that are devised in recent years to explore the possibilities of exploiting the rich dynamics that such models might provide for carrying out
information processing tasks.
The NDS model might be stabilized to a large number of UPOs. These UPOs
can be stabilised using a feedback control mechanism. The NDS model is a modified version of Rössler chaotic system. The rich dynamics of the Rössler system is
supposed to be inherited by the NDS model. This is suggested by the large number
of UPOs that can be stabilised as shown in figure 3.
However, when the discretization methods are discussed in this paper, it is shown
that the method that is used to discretize the original Rössler equations in devising
the NDS equations is not a known discretization method. This along with the change
in sign in the term u(t)x(t)− ku(t)) has affected the shape and properties of the NDS
attractor when compared with its origin: the Rössler attractor.
Different experimental setups have been prepared and performed to tune the NDS
model parameters. The results of these experimentations have revealed the valid
ranges of the parameters of the model. Also, other experimentations have shown
different capacities for the NDS attractor in terms of the number of stabilized UPOs
with different parameter settings.
14
Mohammad Alhawarat, Waleed Nazih and Mohammad Eldesouki
The results attained in this paper suggest that there are weak relationships between the NDS and the Rössler models. However, the NDS attractor encompasses
large number of UPOs as shown in figure 7. These and the wide range of dynamic
behaviours may be exploited to carry out information processing tasks.
Acknowledgements This work was supported by Salman Bin Abdulaziz University in KSA under
grant 40/H/1432. I’m grateful to this support and would like to thank the deanship of scientific
research at the university, represented by Dr. Abdullah Alqahtani at that time.
References
1. Alhawarat, M.O.I.: Learning and memory in chaotic spiking neural models. Ph.D. thesis,
Oxford Brookes University (2007)
2. Aoun, M.: Stdp within nds neurons. In: Advances in Neural Networks - ISNN 2010, Lecture
Notes in Computer Science, vol. 6063, pp. 33–43. Springer Berlin Heidelberg (2010)
3. Babloyantz, A., Lourenço, C.: Brain chaos and computation. International Journal of Neural
Systems 7, 461–471 (1996)
4. Bershadskii, A., Ikegaya, Y.: Chaotic neuron clock.
Chaos, Solitons & Fractals 44(45), 342 – 347 (2011).
DOI 10.1016/j.chaos.2011.03.001.
URL
http://www.sciencedirect.com/science/article/pii/S096007791100035X
5. Crook, N., Goh, W., Hawarat, M.: The nonlinear dynamic state neuron. In: M. Verleysen (ed.) ESANN, pp. 37–42. 13th European Symposium on Artificial Neural Networks
(ESANN’2005), Bruges, Belgium (2005)
6. Crook, N., Goh, W.J.: Nonlinear transient computation as a potential ”kernel trick” in cortical
processing. Biosystems 94(1-2), 55–59 (2008)
7. Crook, N., olde Scheper, T., Pathirana, V.: Self organised dynamic recognition states for
chaotic neural networks. Information Sciences 150(1-2), 59–75 (2003)
8. Destexhe, A.: Oscillations, complex spatiotemporal behavior, and information transport in
networks of excitatory and inhibitory neurons. Physical Review E 50(2), 1594–1606 (1994)
9. Ding, Y., Jiang, W., Wang, H.: Delayed feedback control and bifurcation analysis of rossler
chaotic system. Nonlinear Dynamics 61, 707–715 (2010). DOI 10.1007/s11071-010-9681-y.
URL http://dx.doi.org/10.1007/s11071-010-9681-y
10. Freeman, W.: How Brains Make up their Minds. Weidenfeld and Nicolson, London (1999)
11. Freeman, W.: A proposed name for aperiodic brain activity: stochastic chaos. Neural Networks
13(1), 11–13 (2000)
12. Freeman, W., Barrie, J.: Chaotic oscillations and the genesis of meaning in cerebral cortex.
Temporal Coding in the Brain pp. 13–37 (1994)
13. Goh, W., Crook, N.: Pattern recognition using chaotic transients. In: 15th European Symposium on Artificial Neural Networks (ESANN-2007), pp. 7–12 (2007)
14. Ott, E., Grebogi, C., Yorke, J.A.: Controlling chaos. Physical Review Letters 64, 1196–1199
(1990)
15. Pasemann, F., Stollenwerk, N.: Attractor switching by neural control of chaotic neurodynamics. Network: Computational Neural Systems 9(4), 549–561 (1998)
16. Pyragas, K.: Continuous control of chaos, by self-controlling feedback. Physics Letters A
170, 421–428 (1992)
17. Rapp, P.E.: Chaos in the neurosciences: Cautionary tales from the frontier. Biologist 40(2),
89–94 (1993)
18. Rössler, O.E.: An equation for continuous chaos. Physics Letters A 57(5), 397–398 (1976)
19. Theiler, J.: On the evidence for low-dimensional chaos in an epileptic electroencephalogram.
Physics Letters A 196, 335–341 (1995)
Analysis of a chaotic spiking neural model: The NDS neuron
15
20. Wu, Y.T., Shyu, K.K., Chen, T.R., Guo, W.Y.: Using three-dimensional fractal dimension to analyze the complexity of fetal cortical surface from magnetic resonance images. Nonlinear Dynamics 58, 745–752 (2009). DOI 10.1007/s11071-009-9515-y. URL
http://dx.doi.org/10.1007/s11071-009-9515-y
| 9cs.NE
|
Active Learning for Visual Question Answering:
arXiv:1711.01732v1 [cs.CV] 6 Nov 2017
An Empirical Study
Xiao Lin
Virginia Tech
[email protected]
Devi Parikh
Georgia Tech
[email protected]
Abstract
We present an empirical study of active learning for Visual Question Answering, where a deep VQA
model selects informative question-image pairs from a pool and queries an oracle for answers to maximally
improve its performance under a limited query budget. Drawing analogies from human learning, we
explore cramming (entropy), curiosity-driven (expected model change), and goal-driven (expected error
reduction) active learning approaches, and propose a new goal-driven active learning scoring function
to pick question-image pairs for deep VQA models under the Bayesian Neural Network framework. We
find that deep VQA models need large amounts of training data before they can start asking informative
questions. But once they do, all three approaches outperform the random selection baseline and achieve
significant query savings. For the scenario where the model is allowed to ask generic questions about
images but is evaluated only on specific questions (e.g., questions whose answer is either yes or no), our
proposed goal-driven scoring function performs the best.
1
Introduction
Visual Question Answering (VQA) [1, 8, 9, 11, 23, 28] is the task of taking in an image and a free-form
natural language question and automatically answering the question. Correctly answering VQA questions
arguably demonstrates machines’ image understanding, language understanding and and perhaps even some
commonsense reasoning abilities. Previous works have demonstrated that deep models which combine image,
question and answer representations, and are trained on large corpora of VQA data are effective at the VQA
task.
Although such deep models are often deemed data-hungry, the flip-side is that their performance scales well
with more training data. In Fig.1 we plot performance versus training set size of two representative deep
VQA models: LSTM+CNN [20] and HieCoAtt [21] trained on random subsets of the VQA v1.0 dataset [1].
We see that for both methods, accuracy improves significantly – by 12% – with every order of magnitude
of more training data. As performance improvements still seem linear, it is reasonable to expect another
12% increase by collecting a VQA dataset 10 times larger. Such trends are invariant to choice of image
feature [15] and is also observed in ImageNet image classification [24]. Note that improvements brought by
additional training data may be orthogonal to improvements in model architecture.
1
Figure 1: Performance of two representative VQA models: LSTM+CNN [20] and HieCoAtt [21] on random
subsets of the VQA v1.0 dataset. Both models improve by 12% with every order of magnitude of more
training data.
LSTM+CNN
VQA Accuracy
(Lu et al. 2015)
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
1000
HieCoAtt
(Lu et al. 2016)
10000
Training Set Size
100000
However, collecting large quantities of annotated data is expensive. Even worse, as a result of long tail
distributions, it will likely result in redundant questions and answers while still having insufficient training
data for rare concepts. This is especially important for learning commonsense knowledge, as it is well known
that humans tend to talk about unusual circumstances more often than commonsense knowledge which can
be boring to talk about [10]. Active learning helps address these issues. In active learning, a model is
first trained on an initial training set. It then iteratively expands its training set by selecting potentially
informative examples according to a query strategy, and seeking annotations on these examples. Previous
works have shown that a carefully designed query strategy effectively reduces annotation effort required in a
variety of tasks for shallow models. For deep models however, active learning literature is scarce and mainly
focuses on classical unimodal tasks such as image and text classification.
In this work we study active learning for deep VQA models. VQA poses several unique challenges and
opportunities for active learning.
First, VQA is a multimodal problem. Deep VQA models may combine Multi-Layer Perceptrons (MLPs),
Convolutional Neural Nets (CNNs), Recurrent Neural Nets (RNNs) and even attention mechanisms to solve
VQA. Such models are much more complex than MLPs or CNNs alone studied in existing active learning
literature and need tailored query strategies.
Second, VQA questions are free-form and open-ended. In fact, VQA can play several roles from answering
any generic question about an image, to answering only specific question types (e.g., questions with “yes/no”
answers, or counting questions), to being a submodule in some other task (e.g., image captioning as in [19]).
Each of these different scenarios may require a different active learning approach.
Finally, VQA can be thought of as a Visual Turing Test [9] for computer vision systems. To answer questions
2
such as “does this person have 20/20 vision” and “will the cat be able to jump onto the shelf”, the computer
not only needs to understand the surface meaning of the image and the question, but it also needs to have
sufficient commonsense knowledge about our world. One could argue that proposing informative questions
about images is also a test of commonsense knowledge and intelligence.
We draw coarse analogies to human learning and explore three types of information-theoretic active learning
query strategies:
Cramming – maximizing information gain in the training domain. The objective of this strategy is
to efficiently memorize knowledge in an unlabeled pool of examples. This strategy selects unlabeled examples
whose label the model is most uncertain about (maximum entropy).
Curiosity-driven learning – maximizing information gain in model space. The objective of this
strategy is to select examples that could potentially change the belief on the model’s parameters (also known
as expected model change). There might exist examples in the pool whose labels have high uncertainty but
the model does not have enough capacity to capture them. In curiosity-driven learning the model will skip
these examples. BALD [7, 13] is one such strategy for deep models, where examples are selected to maximize
the reduction in entropy over model parameter space.
Goal-driven learning – maximizing information gain in the target domain. The objective of this
strategy is to gather knowledge to better achieve a particular goal (also known as expected error reduction).
To give an example from image classification, if the goal is to recognize digits i.e., the target domain is digit
classification, dog images in the unlabeled pool are not relevant even though their labels might be uncertain
or might change model parameters significantly. On the other hand, in addition to digit labels, some other
non-digit labels such as the orientation of the image might be useful to the digit classification task. We
propose a novel goal-driven query strategy that computes mutual information between pool questions and
test questions under the Bayesian Neural Network [2, 6] framework.
We evaluate active learning performance on VQA v1.0 [1] and v2.0 [11] under the pool-based active learning
setting described in Section 3. We show that active learning for deep VQA models requires a large amount of
initial training data before they can achieve better scaling than random selection. In other words, the model
needs to have enough knowledge to ask informative questions. But once it does, all three querying strategies
outperform the random selection baseline, saving 27.3% and 19.0% answer annotation effort for VQA v1.0
and v2.0 respectively. Moreover, when the target task is restricted to answering only “yes/no” questions,
our proposed goal-driven query strategy beats random selection and achieves the best performance out of
the three active query strategies.
2
2.1
Related Work
Active Learning
Active learning query strategies for shallow models [31, 17] often rely on specific model simplifications and
closed-form solutions. Deep neural networks however, are inherently complex non-linear functions. This
poses challenges on uncertainty estimation.
3
In the context of deep active learning for language or image understanding, [34] develops a margin-based
query strategy on Restricted Boltzmann Machines for review sentiment classification. [16] queries highconfidence web images for active fine-grained image classification. [30] proposes a query strategy based on
feature space covering, applied to deep image classification. Closest to our work, [7] studies BALD [13],
an expected model change query strategy computed under the Bayesian Neural Network [2, 6] framework
applied to image classification.
In this work we study active learning for VQA. VQA is a challenging multimodal problem. Today’s state-ofthe-art VQA models are deep neural networks. We take an information-theoretic perspective and study three
active learning objectives: minimizing entropy in training domain (entropy), model space (expected model
change) or target domain (expected error reduction). Drawing coarse analogy from human learning, we
call them cramming, curiosity-driven and goal-driven learning respectively. We apply the Bayesian Neural
Network [2, 6] framework to compute these strategies. In particular, for goal-driven learning which was
deemed impractical beyond binary classification on small datasets [31], we propose a fast and effective query
scoring function that speeds up computation by hundreds of millions of times, and show that it is effective
for VQA which has 1, 000 classes and > 400, 000 examples on contemporary multi-modal deep neural nets.
2.2
Visual Conversations
Building machines that demonstrate curiosity – machines that improve themselves through conversations
with humans – is an important problem in AI.
[26, 25] study generating human-like questions given an image and the context of a conversation about that
image. [33] uses reinforcement learning to learn an agent that plays a “Guess What?” game [5]: finding out
which object in the image the user is looking at by asking questions. [4] studies grounded visual dialog [3]
between two machines in collaborative image retrieval, where one machine as the “answerer” has an image
and answers questions about the image while the other as “questioner” asks questions to retrieve the image
at the end of the conversation. Both machines are learnt to better perform the task using reinforcement
learning.
In this work we study visual “conversations” from an active learning perspective. In each round of the
conversation, a VQA model strategically chooses an informative question about an image and queries an
oracle to get an answer. Each round of “conversation” provides a new VQA training example which improves
the VQA model.
3
Approach
We study a pool-based active learning setting for VQA: A VQA model is first trained on an initial training
set Dtrain . It then iteratively grows Dtrain by greedily selecting batches of high-scoring question-image pairs
(Q, I) from a human-curated pool according to a query scoring function s(Q, I). The selected (Q, I) pairs
are sent to an oracle for one of J ground truth answers A ∈ {a1 , a2 , . . . , aJ }, and (Q, I, A) tuples are added
4
as new examples to Dtrain .
1
We take an information-theoretic perspective and explore cramming, curiosity-driven, and goal-driven query
strategies as described in Section 1. However computing s(Q, I) for those query strategies directly is intractable, as they require taking expectations under the model parameter distribution. So in Section 3.1
we first introduce a Bayesian VQA model which enables variational approximation of the model parameter
distribution. And then Section 3.2 introduces the query scoring functions and their approximations.
3.1
Bayesian LSTM+CNN for VQA
We start with the LSTM+CNN VQA model [20]. The model encodes an image into a feature vector using
the VGG-net [32] CNN, encodes a question into a feature vector by learning a Long Short Term Memory
(LSTM) RNN, and then learns a multi-layer perceptron on top that combines the image feature and the
question feature to predict a probabilistic distribution over top J = 1000 most common answers.
In order to learn a variational approximation of the posterior model distribution, we adopt the Bayesian
Neural Network framework [2, 6] and introduce a Bayesian LSTM+CNN model for VQA. Let ω be the
parameters of the LSTM and the multi-layer perceptron (we use a frozen pre-trained CNN). We learn a
weight-generating model with parameter θ:
ω =θ◦
i ∼ Bernoulli(0.5)
(1)
Let qθ (ω) be the probabilistic distribution of weights generated by this model. Following [2, 6], we learn θ
by minimizing KL divergence KL(qθ (ω)||p(ω|Dtrain )) so qθ (ω) serves as a variational approximation to the
true model parameter posterior p(ω|Dtrain ). Specifically we minimize
KL(qθ (ω)||p(ω|Dtrain )) = Eω∼qθ (ω) [− log P (Dtrain |ω)] +
|
{z
}
Cross entropy loss
KL(qθ (ω)||p(ω))
|
{z
}
(2)
Deviation from weight prior
using batch Stochastic Gradient Descent (SGD) to learn θ. In practice, KL(qθ (ω)||p(ω)) can be naively approxmiated with a parametric hybrid L1 - L2 norm [6]. Experiments show that such an naive approximation
does not have a significant impact on active learning results. So in experiments we set this term to 0. How
to come up with a more informative prior term is an open problem for Bayesian Neural Networks.
Let P (A|Q, I, ω) be the predicted J-dimensional answer distribution of the VQA model for question-image
pair (Q, I) when using ω as model parameters. A Bayesian VQA prediction for (Q, I) using variational
approximation qθ (ω) is therefore given by:
1 VQA
models require a large training set to be effective. To avoid prohibitive data collection cost and focus on evaluating
active learning query strategies, in this work we study pool-based active learning which makes use of existing VQA datasets.
Having the model select or even generate questions for images it would liked answered, as opposed to picking from a pool of
(Q, I) pairs is a direction for future research.
5
P (A = a|Q, I) ≈ Eω∼qθ (ω) P (A = a|Q, I, ω)
3.2
(3)
Query Strategies and Approximations
We experiment with 3 active learning query strategies: cramming, curiosity-driven learning and goal-driven
learning.
Cramming or “uncertainty sampling” [31] minimizes uncertainty (entropy) of answers for questions in the
pool. It selects (Q, I) whose answer A’s distribution has maximum entropy. This is a classical active learning
approach commonly used in practice.
sentropy (Q, I) = H(A)
X
=−
P (A = a|Q, I) log P (A = a|Q, I)
(4)
a
Curiosity-driven learning or “expected model change” minimizes uncertainty (entropy) of model parameter distribution p(ω|Dtrain ). It selects (Q, I) whose answer A would expectedly bring steepest decrease in
model parameter entropy if added to the training set.
scuriosity (Q, I) = H(ω) − H(ω|A)
= I(ω; A)
= H(A) − H(A|ω)
(5)
Intuitively, H(A) − H(A|ω) computes the divergence of answer predictions under different model parameters.
If plausible models are making divergent predictions on a question-image pair (Q, I), the answer to this
(Q, I) would rule out many of those models and thereby reduce confusion.
According to BALD [7], the conditional entropy term H(A|ω) in Eq. 5 can be approximated by:
X
H(A|ω) ≈ −Eω∼qθ (ω)
P (A = a|Q, I, ω) log P (A = a|Q, I, ω)
(6)
a
Goal-driven learning or “expected error reduction” minimizes uncertainty (entropy) on answers A0t to
a given set of unlabeled test question-image pairs (Q0t , It0 ), t = 1, 2, ...T , against which the model will be
evaluated. The goal-driven query strategy selects the pool question-image pair (Q, I) that has the maximum
total mutual information with (Q0t , It0 ), t = 1, 2, ...T . That is, it queries (Q, I) pairs which maximize:
6
Algorithm 1 Active learning for Visual Question Answering
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Initialize Dtrain with N inital training examples. Use the rest of (Q, I) in VQA TRAIN set as pool.Q
Train θ on Dtrain for K epochs using Eq. 2 for initial qθ (ω).
for iter = 1, . . . , L do
Sample ω ∼ qθ (ω) M times.
Using each ω to make predictions P (A|Q, I, ω) on all pool and test question-image pairs.
Compute s(Q, I) for every (Q, I) in pool using Eq. 4, 5 or 9.
Select the top G high-scoring (Q, I) pairs from the pool.2
Lookup answers A for (Q, I) pairs in the VQA training set (proxy for querying a human).
Add (Q, I, A) tuples to Dtrain .
Update θ on new Dtrain for K epochs.
end for
sgoal (Q, I) =
X
H(A0t ) − H(A0t |A)
t
=
X
=
XXX
I(A; A0t )
t
t
a
P (A = a, A0t = a0 |Q, I, Q0t , It0 ) log
a0
P (A = a, A0t = a0 |Q, I, Q0t , It0 )
P (A = a|Q, I)P (A0t = a0 |Q0t , It0 )
(7)
For term P (A = a, A0t = a0 |Q, I, Q0t , It0 ), observe that when the model parameter ω is given, (Q, I) and (Q0t , It0 )
are two different VQA questions so their answers – A and A0t respectively – are predicted independently. In
other words, A and A0t are independent conditioned on ω. Therefore we can take expectation over model
parameter ω to compute this joint probability term:
P (A = a, A0t = a0 |Q, I, Q0t , It0 )
= Eω P (A = a|Q, I, ω)P (A0t = a0 |Q0t , It0 , ω)
≈ Eω∼qθ (ω) P (A = a|Q, I, ω)P (A0t = a0 |Q0t , It0 , ω)
(8)
Let M be the number of samples of ω, J be the number of possible answers, and U be the number of examples
in the pool. Computing I(A; A0t ) for all U examples in the pool following Eq. 8 has a time complexity of
O(U T J 2 M ). For VQA typically the pool and test corpora each contains hundreds of thousands of examples
and there are 1000 possible answers, e.g., U = 400k, T = 100k and J = 1, 000. We typically use M = 50
samples in our experiments. So computing Eq. 8 is still time-consuming and can be prohibitive for even
larger VQA datasets. To speed up computation, we approximate log(·) using first-order Taylor expansion
and discover that the following approximation holds empirically (more details can be found in Appendix A
and B):
7
sgoal (Q, I)
X P (A = a|Q, I, ω)P (A = a|Q, I, ω 0 )
1h
≈ Eω Eω0
2
P (A = a|Q, I)
a
i
X X P (A0 = a|Q0 , I 0 , ω)P (A0 = a|Q0 , I 0 , ω 0 )
t
t t
t
t t
−T
0
0
0
P (At = a|Qt , It )
t
a
(9)
Eq. 9 brings drastic improvements to time complexity. It can be computed as a dot-product between two
vectors of length M 2 . One only involves pool questions (Q, I). The other one only involves test questions
(Q0t , It0 ) and can be precomputed for all pool questions. Precomputing vectors for test questions has a time
complexity of O(T JM 2 ). Computing Eq. 9 using the precomputed test vector has a time complexity of
O(U JM 2 ). So the overall time complexity is linear to dataset size max(U, T ) and the number of possible
answers J.
In previous works, goal-driven learning was deemed impractical beyond binary classification on small datasets [31].
Previous works explore the goal-driven learning objective for shallow classifiers such as Naive Bayes [29],
Support Vector Machines [12] and Gaussian Process [35]. However on VQA, computing such scoring functions would require learning a new set of model parameters for every possible combinations of (Q, I, A) and
then making predictions on all (Q0t , It0 ) using the learnt model. That would require 4×1015 forward-backward
passes (10 billion epochs) for VQA neural nets. Instead our Monte-Carlo approximation of Eq. 9 only involves making predictions on (Q, I) and (Q0t , It0 ), and avoids training new models for each of J = 1, 000
answers when computing sgoal (Q, I). In our approach, the operation with the highest time complexity is a
matrix multiplication operation which in practice is not the bottleneck. The most time-consuming operation
– computing scores for P (A|Q, I, ω) and P (A0t |Q0t , It0 , ω) – costs approximately 3 × 107 forward passes (75
epochs), a speed up of more than 108 times. Our approach is easily parallelizable and works for all Bayesian
Neural Networks.
Our active learning procedure is summarized in Algorithm 1.
4
Experiment
4.1
Experiment Setup
We evaluate cramming (entropy), curiosity-driven and goal-driven active learning strategies against passive
learning on the VQA v1.0 [1] and v2.0 [11] datasets. The VQA v1.0 dataset consists of 614,163 VQA
questions with human answers on 204,721 COCO [18] images. The VQA v2.0 dataset augments the VQA
v1.0 dataset and brings dataset balancing: every question in VQA v2.0 is paired with two similar images
that have different answers to the question. So VQA v2.0 doubles the amount of data and models need to
focus on the image to do well on VQA v2.0.
2 Jointly
selecting a batch of (Q, I) pairs that optimizes the active learning objectives may further improve active learning
performance. Deriving query strategies that can select batches of examples under the Bayesian Neural Network framework is
part of future work.
8
Figure 2: Active learning versus passive learning on (top) VQA v1.0 and (bottom) v2.0. All three active
learning strategies perform better than passive learning. Best viewed in color.
VQA v2.0
0.53
0.48
0.52
0.47
0.51
0.5
Save 27.3%
queries
0.49
0.48
Curiosity-driven
0.46
0.45
Save 19.0%
queries
0.44
100000
150000
0.42
50000
Goal-driven
Entropy
Passive
0.43
0.47
0.46
50000
VQA Accuracy
VQA Accuracy
VQA v1.0
100000
150000
Training Set Size
Training Set Size
We choose a random initial training set of N = 50k (Q, I) pairs from the TRAIN split, use the rest of TRAIN
as pool and report VQA accuracy [1] on the VAL split. We run the active learning loop for L = 50 iterations.
We sample model parameter ω for M = 50 times for query score computation. For passive learning i.e.
querying (Q, I) pairs randomly, we set spassive (Q, I) ∼ uniform(0, 1). In each iteration G = 2, 000 (Q, I, A)
pairs are added to Dtrain , resulting in a training set of 150k examples by the end of iteration 50.
For VQA model, we use the Bayesian LSTM+CNN model described in Section 3.1. In every active learning
iteration we train the model for K = 50 epochs with learning rate 3×10−4 and batch size 8×128 for learning
qθ (ω).
4.2
Active Learning on VQA v1.0 and v2.0
Fig. 2 (left), (right) show the active learning results on VQA v1.0 and v2.0 respectively. On both datasets,
all 3 active learning methods perform similarly and all of them outperform passive learning. On VQA v1.0,
passive learning queries 88k answers before reaching 51% accuracy, where as all active learning methods
need only 64k queries, achieving a saving of 27.3%. It shows that active learning is able to effectively
tell informative VQA questions from redundant questions, even among high-quality questions generated by
humans. Similarly at 46% accuracy, active learning on VQA v2.0 achieves a saving of 19.0%. Savings on
VQA v2.0 is lower, possibly because dataset balancing in VQA v2.0 improves the informativeness of even a
random example.
Table 1 shows that for each pair of active learning methods, what percentage of the query (Q, I) pairs are
selected by both methods on VQA v2.0 (overlap between their training sets). For the VQA task, active
learning methods seem to agree on which (Q, I) pairs are more informative. They have more than 80% of
9
Table 1: On VQA v2.0 for each pair of query strategy, what percentage of (Q, I) pairs are selected by both
methods. Active learning (entropy, curiosity-driven, goal-driven) query strategies select > 80% common
(Q, I) pairs and they are very different from passive learning.
(Q, I)
Overlap (%)
Passive learning
Entropy
Curiosity-driven
Goal-driven
Passive
learning
26.70
26.65
26.64
Entropy
26.70
83.26
82.52
Curiosity
driven
26.65
83.26
85.27
Goal
driven
26.64
82.52
85.27
-
(Q, I) pairs in common, while against passive learning they only share ∼27 % of (Q, I) pairs.
On VQA v2.0, we also experiment with smaller initial training sets N ∈ {20k, 10k, 5k, 2k} to study the
impact of training set size on active learning performance. Fig. 3 shows the results. For all initial training
set sizes, the breakpoint when active learning methods start to outperform passive learning is around 30k
to 50k examples. It shows that active learning methods do require a large training set size to start asking
informative questions. Models with smaller initial training set sizes tend to show less and inconsistent data
savings compared to N = 50k, possibly because such models are less capable of telling informative questions
from redundant ones. In addition, entropy shows fluctuating performance while curiosity-driven learning
performs consistantly better than both entropy and passive learning irrespective of initial training set size.
4.3
Goal-driven Active Learning
To evaluate our goal-driven learning approach, we keep the initial training set and the pool unchanged for
VQA v2.0 – the model is allowed to ask all kinds of questions from the VQA v2.0 TRAIN split – but will
be evaluated on only “yes/no” questions (questions whose answers are “yes” or “no” ) in the VAL split.
This task tests our proposed goal-driven active learning approach’s ability to focus on achieving the goal of
answering “yes/no” questions more accurately.
Fig. 4 (top) shows the performance of active and passive learning approaches on this task.3 Our goal-driven
active learning approach is able to select relevant questions as queries and outperforms passive learning.
Curiosity-driven and entropy approaches perform poorly. They are not aware of the task and tend to be
attracted to harder, open-ended questions, which are not very relevant to the task.
Fig. 4 (bottom) shows a closer examination of the composition of questions queried by the goal-driven
learning approach compared to baseline approaches. The goal-driven learning approach queries mostly
“yes/no” questions, which are presumably more useful for the task. Note that the approach was not told
that the downstream task is to answer “yes/no” questions. The approach figures out which questions will
be informative to ask just based on samples from the downstream task. It shows that the goal-driven
scoring function in Eq. 7, as well as the approximations in Eq. 9 are indeed effective for selecting informative
3 We also found that updating θ from previous iteration in Algorithm 1 step 10 leads to slight overfitting that affects mutual
information approximation. So for this task, θ is initialized from scratch in every iteration.
10
Figure 3: Active learning with N = 20k, 10k, 5k, 2k initial training set size. When dataset size is small,
active learning is unable to outperform passive learning. The breakpoint when active learning methods start
to perform better is around 30k to 50k examples. Best viewed in color.
VQA v2.0 N=20k initial
VQA v2.0 N=10k initial
0.46
0.48
VQA Accuracy
VQA Accuracy
0.46
0.44
0.42
0.4
0.38
20000
40000
80000
0.44
Curiosity-driven
0.42
Training Set Size
0.44
0.44
VQA Accuracy
VQA Accuracy
0.46
0.42
0.4
80000
80000
160000
VQA v2.0 N=2k initial
0.46
40000
40000
Training Set Size
VQA v2.0 N=5k initial
0.38
20000
Passive
0.4
0.38
20000
160000
Entropy
160000
Curiosity-driven
0.42
Entropy
Passive
0.4
0.38
20000
40000
80000
Training Set Size
Training Set Size
11
160000
Figure 4: Top: Goal-driven active learning of VQA for answering only “yes/no” questions. Our goal-driven
active learning approach outperforms passive learning and other active learning approaches. Bottom: Query
compositions of active learning approaches, on VQA v2.0 dataset for the task of answering only “yes/no”
questions. Our goal-driven active learning approach queries mostly “yes/no” questions. Best viewed in color.
VQA v2.0 Test on “Yes/No” Questions
VQA Accuracy
0.72
Goal-driven
0.7
Passive learning
Curiosity-driven
Entropy
0.68
0.66
50000
VQA v2.0 Test on “Yes/No” Questions
Query Composition
1
“Yes/No” Question
150000
Training Set Size
0.8
Goal-driven
Passive learning
0.6
Curiosity-driven
0.4
Entropy
0.2
0
0
10
20
30
Iteration
12
40
50
Figure 5: Goal-driven active learning of VQA for answering only “yes/no” questions, compared to passive
learning that “cheats” and queries only “yes/no” questions. Best viewed in color.
VQA v2.0 Test on “Yes/No” Questions
VQA Accuracy
0.72
"Yes/No"
questions only
0.7
Goal-driven
0.68
0.66
50000
Training Set Size
150000
questions.
As an “upper bound”, it is reasonable to assume4 that “yes/no” questions are more desirable for this task.
Imagine a passive learning method that “cheats”: one that is aware that it will be tested only on “yes/no”
questions, as well as knowing which questions are “yes/no’ questions in the pool, so it restricts itself to query
only “yes/no” questions. How does our goal-driven learning approach compare with such a method that
only learns from “yes/no” questions? Fig. 5 shows the results. Our goal-driven learning is able to compete
with the “cheating” approach. In fact, of all 167,499 “yes/no” questions in the VQA v2.0 TRAIN split,
goal-driven learning finds 38% of them by iteration 25, and 50% of them by iteration 50. That might also
have made finding the remaining “yes/no” questions more difficult which explains the drop of the rate of
“yes/no” question towards later iterations. We expect that a larger pool (i.e. a larger VQA dataset) would
reduce these issues.
5
Discussion
In this work we discussed three active learning strategies – cramming (entropy), curiosity-driven learning
and goal-driven learning – for Visual Question Answering using deep multimodal neural networks. Our
results show that deep VQA models require 30k - 50k training questions for active learning before they are
able to ask informative questions and achieve better scaling than randomly selecting questions for labeling.
4 Note that this is not necessarily the case. Even non-yes/no questions can help a VQA model get better at answering yes/no
questions by learning concepts from non-yes/no questions that can later come handy for yes/no questions. For example “Q:
What is the man doing? A: Surfing” can be as useful as “Q: Is the man surfing? A: Yes”.
13
Once the training set is large enough, several active learning strategies achieve significant savings in answer
annotation cost. Our proposed goal-driven query strategy in particular, shows a significant advantage on
improving performance when the downstream task involves answering a specific type of VQA questions.
Jointly selecting batches of examples as queries [30] and formulating active learning as a decision making
problem [14] (greedily selecting the batch that reduces entropy by the most for the current iteration may
not be the optimal decision) have been shown to improve optimality in active learning query selection.
Combining those approaches with deep neural networks under the Bayesian Neural Network framework are
promising future directions.
The pool-based active learning setup explored in this work selects unlabeled human generated questionimage pairs and asks the oracle for answers. For building VQA datasets however, collecting human-generated
questions paired with each image is also a substantial portion of the overall cost. Hence, starting from a
bank of questions and an unaligned bank of images, and having the model decide which question it would
like to pair with each image to use as a query would result in a further reduction in cost. Note that such
a model would need to not only reason about the informativeness of a question-image pair, but also about
the relevance of a question to the image [27, 22]. Evaluating such an approach would require collecting new
VQA datasets with humans in the loop to give answers – which we show would require 30k - 50k answers
before the model could start selecting informative images and questions. Going one step further, we could
also envision a model that generates new questions rather than selecting from a pool of questions. That
would require a generative model that can perform inference to optimize for the active learning objectives.
We hope that our work serves as a foundation for these future research directions.
Acknowledgements
We thank Michael Cogswell and Qing Sun for discussions about the active learning strategies. This work was
funded in part by an NSF CAREER award, ONR YIP award, Allen Distinguished Investigator award from
the Paul G. Allen Family Foundation, Google Faculty Research Award, and Amazon Academic Research
Award to DP. The views and conclusions contained herein are those of the authors and should not be
interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of
the U.S. Government, or any sponsor.
References
[1] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh. VQA: Visual Question
Answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages
2425–2433. IEEE, 2015.
[2] C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight uncertainty in neural network.
In Proceedings of the 32nd International Conference on Machine Learning (ICML), pages 1613–1622.
PMLR, 2015.
14
[3] A. Das, S. Kottur, K. Gupta, A. Singh, D. Yadav, J. M. Moura, D. Parikh, and D. Batra. Visual Dialog.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE,
2017.
[4] A. Das, S. Kottur, J. M. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with
deep reinforcement learning. arXiv preprint arXiv:1703.06585, 2017.
[5] H. de Vries, F. Strub, S. Chandar, O. Pietquin, H. Larochelle, and A. Courville. GuessWhat?! Visual
object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR). IEEE, 2017.
[6] Y. Gal and Z. Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty
in deep learning. In Proceedings of the 33rd International Conference on Machine Learning (ICML),
pages 1050–1059. PMLR, 2016.
[7] Y. Gal, R. Islam, and Z. Ghahramani. Deep Bayesian active learning with image data. In Proceedings
of the 34th International Conference on Machine Learning (ICML), pages 1183–1192. PMLR, 2017.
[8] H. Gao, J. Mao, J. Zhou, Z. Huang, L. Wang, and W. Xu. Are you talking to a machine? Dataset and
methods for multilingual image question. In Proceedings of the 28th Advances in Neural Information
Processing Systems (NIPS), pages 2296–2304, 2015.
[9] D. Geman, S. Geman, N. Hallonquist, and L. Younes. Visual Turing test for computer vision systems.
volume 112, pages 3618–3623. National Academy of Sciences, 2015.
[10] J. Gordon and B. Van Durme. Reporting bias and knowledge acquisition. In Proceedings of the 2013
Workshop on Automated Knowledge Base Construction (AKBC), pages 25–30. ACM, 2013.
[11] Y. Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh. Making the V in VQA matter: Elevating
the role of image understanding in Visual Question Answering. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017.
[12] Y. Guo and R. Greiner. Optimistic active-learning using mutual information. In Proceedings of the 22nd
International Joint Conference on Artificial Intelligence (IJCAI), volume 7, pages 823–829. AAAI, 2007.
[13] N. Houlsby, F. Huszár, Z. Ghahramani, and M. Lengyel. Bayesian active learning for classification and
preference learning. arXiv preprint arXiv:1112.5745, 2011.
[14] S. Javdani, Y. Chen, A. Karbasi, A. Krause, D. Bagnell, and S. S. Srinivasa. Near optimal Bayesian
active learning for decision making. In Proceedings of the 17th International Conference on Artificial
Intelligence and Statistics (AISTATS), pages 430–438. PMLR, 2014.
[15] K. Kafle and C. Kanan. Visual Question Answering: Datasets, algorithms, and future challenges.
Computer Vision and Image Understanding, 2017.
[16] J. Krause, B. Sapp, A. Howard, H. Zhou, A. Toshev, T. Duerig, J. Philbin, and L. Fei-Fei. The
unreasonable effectiveness of noisy data for fine-grained recognition. In Proceedings of the 14th European
Conference on Computer Vision (ECCV), pages 301–320. Springer, 2016.
[17] A. Krishnakumar. Active learning literature survey. Technical Report, University of California, Santa
Cruz, 2007.
[18] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick.
Microsoft COCO: Common objects in context. In Proceedings of the 13th European Conference on
Computer Vision (ECCV), pages 740–755. Springer, 2014.
15
[19] X. Lin and D. Parikh. Leveraging Visual Question Answering for image-caption ranking. In Proceedings
of the 14th European Conference on Computer Vision (ECCV), pages 261–277. Springer, 2016.
[20] J. Lu, X. Lin, D. Batra, and D. Parikh. Deeper LSTM and normalized CNN Visual Question Answering
model. https://github.com/VT-vision-lab/VQA_LSTM_CNN, 2015.
[21] J. Lu, J. Yang, D. Batra, and D. Parikh. Hierarchical question-image co-attention for visual question
answering. In Proceedings of the 29th Advances in Neural Information Processing Systems (NIPS),
pages 289–297, 2016.
[22] A. Mahendru, V. Prabhu, A. Mohapatra, D. Batra, and S. Lee. The promise of premise: Harnessing
question premises in Visual Question Answering. In Proceedings of the Conference on Empirical Methods
in Natural Language Processing (EMNLP), pages 937–946. ACL, 2017.
[23] M. Malinowski, M. Rohrbach, and M. Fritz. Ask your neurons: A neural-based approach to answering
questions about images. In Proceedings of the IEEE International Conference on Computer Vision
(ICCV), pages 1–9. IEEE, 2015.
[24] D. Mishkin, N. Sergievskiy, and J. Matas. Systematic evaluation of convolution neural network advances
on the ImageNet. Computer Vision and Image Understanding, 161:11–19, 2017.
[25] N. Mostafazadeh, C. Brockett, B. Dolan, M. Galley, J. Gao, G. P. Spithourakis, and L. Vanderwende.
Image-grounded conversations: Multimodal context for natural question and response generation. arXiv
preprint arXiv:1701.08251, 2017.
[26] N. Mostafazadeh, I. Misra, J. Devlin, M. Mitchell, X. He, and L. Vanderwende. Generating natural questions about an image. In Proceedings of the 54th Annual Meeting of the Association for Computational
Linguistics, volume 1, pages 1802–1813. ACL, 2016.
[27] A. Ray, G. Christie, M. Bansal, D. Batra, and D. Parikh. Question relevance in VQA: Identifying nonvisual and false-premise questions. In Proceedings of the Conference on Empirical Methods in Natural
Language Processing (EMNLP), pages 919–924. ACL, 2016.
[28] M. Ren, R. Kiros, and R. Zemel. Exploring models and data for image question answering. In Proceedings
of the 28th Advances in Neural Information Processing Systems (NIPS), pages 2953–2961, 2015.
[29] N. Roy and A. Mccallum. Toward optimal active learning through Monte Carlo estimation of error
reduction. In Proceedings of the 18th International Conference on Machine Learning (ICML), pages
441–448. Morgan Kaufmann Publishers, 2001.
[30] O. Sener and S. Savarese. A geometric approach to active learning for convolutional neural networks.
arXiv preprint arXiv:1708.00489, 2017.
[31] B. Settles. Active learning literature survey. Computer Science Technical Report (TR1648), University
of Wisconsin, Madison, 2010.
[32] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556, 2014.
[33] F. Strub, H. de Vries, J. Mary, B. Piot, A. Courville, and O. Pietquin. End-to-end optimization of
goal-driven and visually grounded dialogue systems. In Proceedings of the 26th International Joint
Conference on Artificial Intelligence (IJCAI), pages 2765–2771. AAAI, 2017.
16
[34] S. Zhou, Q. Chen, and X. Wang. Active deep networks for semi-supervised sentiment classification.
In Proceedings of the 23rd International Conference on Computational Linguistics (Coling): Posters,
pages 1515–1523. ACL, 2010.
[35] X. Zhu, J. Lafferty, and Z. Ghahramani. Combining active learning and semi-supervised learning using
gaussian fields and harmonic functions. In ICML 2003 workshop on The Continuum from Labeled to
Unlabeled Data in Machine Learning and Data Mining, pages 58–65, 2003.
17
A
Fast Approximation of Goal-driven Scoring Function
In Section 3.2, we discuss our proposed goal-driven query strategy that minimizes uncertainty (entropy) on
answers A0t to a given set of unlabeled test question-image pairs (Q0t , It0 ), t = 1, 2, ...T , against which the
model will be evaluated. It queries (Q, I) pairs which maximize:
sgoal (Q, I)
X
=
H(A0t ) − H(A0t |A)
t
=
X
I(A; A0t )
t
=
XXX
t
a
P (A = a, A0t = a0 |Q, I, Q0t , It0 ) log
a0
P (A = a, A0t = a0 |Q, I, Q0t , It0 )
P (A = a|Q, I)P (A0t = a0 |Q0t , It0 )
(10)
Recall that we propose an approximation for term P (A = a, A0t = a0 |Q, I, Q0t , It0 ) as follows:
P (A = a, A0t = a0 |Q, I, Q0t , It0 )
= Eω P (A = a|Q, I, ω)P (A0t = a0 |Q0t , It0 , ω)
≈ Eω∼qθ (ω) P (A = a|Q, I, ω)P (A0t = a0 |Q0t , It0 , ω)
(11)
Let us define four matrices M1 , D1 , M2 (t), D2 (t) as follows:
P (A = aJ |Q, I, ω1 )
P (A = aJ |Q, I, ω2 )
..
.
P (A = aJ |Q, I, ωM )
(12)
i
P (A = aJ |Q, I)
(13)
P (A = a1 |Q, I, ω1 ) P (A = a2 |Q, I, ω1 ) . . .
P (A = a1 |Q, I, ω2 ) P (A = a2 |Q, I, ω2 )
M1 =
..
..
.
.
P (A = a1 |Q, I, ωM ) P (A = a2 |Q, I, ωM ) . . .
D1 = Diag
h
P (A = a1 |Q, I) P (A = a2 |Q, I) . . .
P (A0t = a1 |Q0t , It0 , ω1 ) P (A0t = a2 |Q0t , It0 , ω1 ) . . . P (A0t = aJ |Q0t , It0 , ω1 )
P (A0t = aJ |Q0t , It0 , ω2 )
P (A0t = a1 |Q0t , It0 , ω2 ) P (A0t = a2 |Q0t , It0 , ω2 )
M2 (t) =
..
..
..
.
.
.
P (A0t = a1 |Q0t , It0 , ωM ) P (A0t = a2 |Q0t , It0 , ωM ) . . . P (A0t = aJ |Q0t , It0 , ωM )
D2 (t) = Diag
h
P (A0t = a1 |Q0t , It0 ) P (A0t = a2 |Q0t , It0 ) . . .
18
P (A0t = aJ |Q0t , It0 )
i
(14)
(15)
Here M1 is an M × J matrix, D1 is a J × J matrix, M2 (t) is an M × J matrix and D2 (t) is a J × J matrix.
With M1 , D1 , M2 (t), D2 (t) we could rewrite Eq. 11 in matrix form:
P (A = a1 , A0t = a1 |Q, I, Q0t , It0 ) . . . P (A = a1 , A0t = aJ |Q, I, Q0t , It0 )
..
..
..
.
.
.
0
0
0
0
0
0
P (A = aJ , At = a1 |Q, I, Qt , It ) . . . P (A = aJ , At = aJ |Q, I, Qt , It )
1
≈
M T M2 (t)
M 1
(16)
Let Sum(·) be an operator on a matrix that sums up all elements in that matrix. Combining Eq. 10 and
Eq. 16, our goal-driven scoring function can be approximately computed as follows:
sgoal (Q, I)
XXX
P (A = a, A0t = a0 |Q, I, Q0t , It0 )
=
P (A = a, A0t = a0 |Q, I, Q0t , It0 ) log
P (A = a|Q, I)P (A0t = a0 |Q0t , It0 )
0
a
t
a
≈
X
t
1
1
Sum
M T M2 (t) ◦ log
D1 M1T M2 (t)D2 (t)
M 1
M
X1
1
1 T
Sum −
M1T M2 (t) +
M1 M2 (t) ◦ D1 M1T M2 (t)D2 (t)
2
2
M
M
t
X 1
1
− +
=
Sum M1T M2 (t) ◦ D1 M1T M2 (t)D2 (t)
2
2
2M
t
X 1
T
1
− +
=
Tr M1T M2 (t) D1 M1T M2 (t)D2 (t)
2
2
2M
t
X 1
1
=
− +
Tr M1T M2 (t)D2 (t)M2T (t)M1 D1
2
2
2M
t
X 1
1
=
− +
Tr M1 D1 M1T M2 (t)D2 (t)M2T (t)
2
2
2M
t
X 1
1
=
− +
Sum M1 D1 M1T ◦ M2 (t)D2 (t)M2T (t)
2
2
2M
t
≈
=
X
t
−
X P (A = a|Q, I, ω)P (A = a|Q, I, ω 0 )
1
1
+ Eω Eω0
2
2
P (A = a|Q, I)
a
Rewriting in matrix form.
x log x ≈
1
(−x + x2 ).
2
Sum of P (A, A0t ) reduces to 1.
Sum(A ◦ B) = Tr(AB T ).
Property of trace.
Tr(AB T ) = Sum(A ◦ B).
Rewriting in probability form.
X P (A0 = a|Q0 , I 0 , ω)P (A0 = a|Q0 , I 0 , ω 0 )
t
t t
t
t t
0 = a|Q0 , I 0 )
P
(A
t
t
t
a
X P (A = a|Q, I, ω)P (A = a|Q, I, ω 0 )
1
= Eω Eω0
2
P (A = a|Q, I)
a
Rearranging summation.
X X P (A0 = a|Q0 , I 0 , ω)P (A0 = a|Q0 , I 0 , ω 0 ) X 1
t
t t
t
t t
−
0 = a|Q0 , I 0 )
P
(A
2
t
t
t
a
t
t
(17)
Which is Eq. 9 in Section 3.2.
As stated in Section 3.2, the above equation can be computed as a dot-product between two vectors of
1
M1 D1 M1T expanded into a vector. It only involves pool questions
length M 2 . One vector is matrix M
P
1
T
(Q, I). The other vector is M
t M2 (t)D2 (t)M2 (t) expanded into a vector. It only involves test questions
0
0
(Qt , It ) and it is shared for all pool questions (Q, I), so it can be precomputed for all (Q, I). Precomputing
P
1
T
2
t M2 (t)D2 (t)M2 (t) for test questions has a time complexity of O(T JM ). Note that D2 (t) is a diagonal
M
19
Figure 6: Convergence of Monte Carlo approximation to entropy, curiosity-driven and goa-driven scoring
functions in terms of rank correlation. We compute scores using Eq. 4 (entropy), 5 (curiosity-driven) and
7 (goal-driven) for 200 random examples from the pool using M ∈ {1, 2, 5, 10, 20, 50, 100, 200, 500} samples
from qθ (ω), and compare them with M = 500 in terms of rank correlation (Spearman’s ρ).
Monte Carlo Approximation Quality
1.2
Spearman’s ρ
1
0.8
0.6
Entropy
0.4
Curiosity-driven
0.2
Goal-driven
0
1
10
100
1000
Number of ω samples
matrix, so multiplying D2 (t) with M2T (t) only takes O(JM ) operations. In the same way, computing
1
T
2
M M1 D1 M1 for all (Q, I) has a time complexity of O(U JM ). The time complexity of their dot product
for all (Q, I) is merely O(U M 2 ). So the overall time complexity is O(max(U, T )JM 2 ). The overall time
complexity is linear to both dataset size U and T and the number of possible answers J, so our approach
can easily scale to very large datasets and more VQA answers.
B
Quality of Approximations
Our entropy, curiosity-driven and goal-driven scoring functions use 3 types of approximations
(a) Variational distribution qθ (ω) as approximation to model parameter distribution p(ω|Dtrain ).
(b) Monte Carlo sampling over qθ (ω) for computing expectation over p(ω|Dtrain ).
(c) Fast approximation to mutual information in Eq. 17.
For (a), since the space of model parameters is very large, it is intractable to evaluate how accurate qθ (ω)
substitutes p(ω|Dtrain ) for expectation computation. But nevertheless our goal-driven learning results in
Section 4.3 suggest that Eq. 9 computed using qθ (ω) is indeed useful for selecting relevant examples. It
remains as an open problem that how to quantitatively evaluate the quality of qθ (ω) for the purpose of
uncertainty estimation and expectation computation.
20
Figure 7: Entropy, curiosity-driven and goal-driven scores of 50 examples under different numbers of model
parameter samples.
Entropy Scores
Curiosity-driven Scores
3.5
1.8
3
1.6
1.4
2.5
1.2
2
1
1.5
0.8
0.6
1
0.4
0.5
0.2
0
0
1
10
100
1
1000
Number of ω samples
10
100
Number of ω samples
1000
Goal-driven Scores
50
45
40
35
30
25
20
15
10
5
0
1
10
100
Number of ω samples
1000
For (b), we study the convergence patterns of Monte Carlo sampling. Specifically, given an arbitrary VQA
model5 , we compute scores using Eq. 4 (entropy), 5 (curiosity-driven) and 7 (goal-driven) for 200 random
examples from the pool using M ∈ {1, 2, 5, 10, 20, 50, 100, 200, 500} samples from qθ (ω), and compare them
with M = 500 in terms of rank correlation (Spearman’s ρ). Note that we use different seeds for the
different M values, i.e. ω samples for M = 200 do not overlap with ω samples for M = 500. Fig. 6 shows
the results. Entropy, curiosity-driven and goal-driven scoring functions require increasingly more samples of
model parameters to converge in terms of ranking. To reach ρ = 0.9, entropy, curiosity-driven and goal-driven
scoring functions require 10, 20 and 50 samples from qθ (ω) respectively. Fig. 8 shows how the actual scores
of examples change according to number of samples from qθ (ω) for 50 random examples in the pool. The
entropy and curiosity-driven scores seem to converge with a large number of samples. The goal-driven scores
however, tend to first increase and then decrease with the number of samples and have not yet converged by
5 For
our experiments we use the model from curiosity-driven learning at iteration 50. This choice is made arbitrarily and
does not change conclusions.
21
M = 500 samples, which is a limitation of the Monte Carlo sampling approach. Despite that, the relative
rankings based on which the queries are selected have mostly converged. Upper- and lower-bounds of Eq. 7
that might improve convergence are opportunities for future research.
For (c), we plot goal-driven scores Eq. 7 as the x-axis versus our fast approximations Eq. 9 as the y-axis for
200 random examples from the pool using M = {2, 5, 10, 20, 50, 100, 200, 500} samples from qθ (ω). Because
Eq. 7 does not scale well to large datasets, we use a subset of 200 random (Q0t , It0 ) pairs from the VAL split
as the test domain for both Eq. 7 and Eq. 9. Fig. 8 shows the results. Our fast approximations are mostly
linear to the goal-driven scores. The slope changes according to the number of model parameter samples M .
That is probably because our approximation 21 (−x + x2 ) (see Section A for details) overestimates x log x for
x > 1. The rank correlations between goal-driven scores and their fast approximations remain high, e.g.,
above ρ > 0.96 even for M = 500, which is sufficient for query selection.
22
Figure 8: Our fast approximations using Eq. 9 versus the original goal-driven scores computed using Eq. 7
under M = {2, 5, 10, 20, 50, 100, 200, 500} samples of model parameters. Our approximations have high rank
correlation with scores computed using the original method.
Fast Approximation to Goal-driven Scoring Function
M=5
Our Fast Approximation
Our Fast Approximation
M=2
45
40
35
30
25
20
15
10
5
0
𝜌=0.9945
0
20
40
70
60
50
𝜌=0.9898
40
30
20
10
0
60
0
M = 10
70
60
𝜌=0.9860
50
40
30
20
10
0
0
20
40
20
40
60
𝜌=0.9784
0
20
Our Fast Approximation
Our Fast Approximation
60
𝜌=0.9724
40
30
20
10
0
20
30
40
60
50
𝜌=0.9735
40
30
20
10
0
0
M = 200
𝜌=0.9641
0
5
10
15
10
20
30
Goal-driven Scoring Function
Our Fast Approximation
Our Fast Approximation
Goal-driven Scoring Function
45
40
35
30
25
20
15
10
5
0
60
M = 100
M = 50
10
40
Goal-driven Scoring Function
70
0
80
M = 20
80
70
60
50
40
30
20
10
0
Goal-driven Scoring Function
50
60
Goal-driven Scoring Function
Our Fast Approximation
Our Fast Approximation
Goal-driven Scoring Function
20
M = 500
35
30
𝜌=0.9686
25
20
15
10
5
0
0
Goal-driven Scoring Function
5
10
Goal-driven Scoring Function
23
15
| 1cs.CV
|
All entangled states can generate certified randomness
Xing Chen1, ∗
arXiv:1706.03528v4 [quant-ph] 19 Sep 2017
1
Institute of Physics, Beijing National Laboratory for Condensed Matter Physics,
Chinese Academy of Sciences, Beijing 100190, China
Random number has many applications, it plays an important role in quantum information processing. It’s not difficult to generate true random numbers, the main difficulty is how to certify
the random numbers generated by untrusted devices. In [Nature(London) 464, 1021 (2010)], the
authors provided us a way to generate certified random number by Bell’s theorem. In their scheme,
we can use the nonlocal behavior of entangled states to generate certified randomness. But there are
entangled states, which admit a local hidden variable model, could not be used in their scheme. We
show in our paper that the nonlocal correlations in every entangled state can be used to generate
certified randomness, and we use Werner states as an example to show how to quantify the output
randomness.
I.
INTRODUCTION
A.
Background
Random number has many important applications
nowadays, such as quantum key distribution and the test
of Bell’s theorem. It is believed that we can’t generate true random numbers by classical processes. On the
other hand, the random numbers generated by quantum
physics are truly random because of the superposition
of quantum states. There are three kinds of quantum
random number generators(QRNG)[1], the first kind is
practical QRNG, which is built on trusted devices; the
second kind is self-testing QRNG, which can generate
true random numbers without trusting the measurement
devices; the last kind of QRNG is called semi-self-testing
QRNG, which partially combines the advantages of the
first two QRNGs.
It is not difficult to generate true random numbers by
tools of quantum physics currently. In the first kind of
QRNG, we can actually generate random numbers at a
very satisfied speed[2][3][4][5]. But these true random
numbers are not easy to be certified, because the adversaries may use memory-stick attack[6]. The so called
memory-stick attack is implemented in the following way:
the adversaries may generate a very long true random
numbers and store them in the devices, when the users
use this devices to generate random numbers, the generated random numbers are exactly what the adversaries
stored in the devices.
Nonlocality can be used to generate certified random
numbers [7][8]. In [7][8] the authors connected Bell’s theorem with randomness, the violation of Bell’s inequality
guarantees that the generated random numbers contain
true randomness, and this randomness is measurement
device independent. Especially in [8], a lower bound of
the output randomness was derived by the nonlocality in
entangled states. However, not all entangled states can
∗
[email protected]
violate Bell’s inequality, and those who admit a local hidden variable model can not be used in their randomness
certification system. In order to take advantage of all
entangled states, we need a new certification scheme. Inspired by previous work [7][8], with the method provided
by [9][10][11], we connect randomness with entanglement.
In our paper, we can generate certified private randomness by any entangled state, the output randomness in
our protocol is measurement device independent and it
only needs some fresh randomness as input.
B.
Related work
Similar randomness generation protocol was also mentioned in Chaturvedi and Banik’s paper[12], but there
are some major differences between their paper and our
paper.
In their paper, Chaturvedi and Banik only gave a very
specific entangled state’s output randomness. Also, in
their paper, they claimed that the output randomness
was safe even local operations and classical communication(LOCC) was allowed between two measurement
devices. However, the untrusted measurement devices
could share some extra entangled states which are not
known by Alice and Bob, and we show in our paper that
these extra entangled states have the potential to damage
the output randomness without being detected.
C.
The result of our paper
In our paper, we give a general lower bound for our
randomness generation protocol. The same as[8], the
classical communication between different measurement
devices is forbidden in our protocol. Moreover, the lower
bound derived in our paper takes the secret entangled
states shared between untrusted measurement devices
into consideration.
Generally, the certified randomness in our scheme is
different from the randomness verified by Bell’s theorem. This is because the random numbers in these two
2
protocols are not generated in the same way. In Bell’s
scenario, the randomness is created by measuring entangled states with optimal positive operator valued measure(POVM)(in order to obtain the maximum Bell value
I), and the guessing probability of the measurement results is increasing with the decreasing of I, only when
I > 2, the guessing probability is less than 1, so this protocol needs the violation of Bell’s inequality to guarantee
the randomness in the measurement results. While in our
protocol, the inputs are nonorthogonal states, the maximum guessing probability has an upper bound which is
always less than 1, and the measurement results in our
protocol can be verified by any entangled state, so every
entangled state is a useful resource in our protocol.
Our paper is organised in the following way. In section II we give a brief introduction to the randomness
generation protocol mentioned in [8]. In section III, the
main part of our paper, we show how to construct the
protocol in our paper, and we give an example to show
how to quantify the output randomness. Section IV is a
brief conclusion.
II.
RANDOMNESS CERTIFIED BY BELL’S
THEOREM
Before the illustration of our randomness generation protocol, we first give a brief introduction to the
scheme mentioned in [8], the authors used Clauser-HorneShimony-Holt (CHSH)[13][14]correlation function
I=
X
(−1)xy [P (a = b|xy) − P (a 6= b|xy)]
(1)
x,y
as an example to show how to generate certified randomness. Where x and y represent certain type of measurement, such as the different polarization directions of
polarizers. a and b are the measurement results of Alice and Bob. Choosing certain x and y to measure the
entangled state between Alice and Bob could make the
Bell value I > 2, which means the violation of CHSH
inequality, and the violation of CHSH inequality guarantees the randomness in the measurement results a, b.
The randomness generation structure is shown in FIG.1
In this randomness generation protocol we can use
the random seed x and y to generate certified random
strings a and b if their correlation violate CHSH inequality, or more generally Bell’s inequality.(One thing we
should note here is that the input x and y should both be
true random numbers, if they can be predicted by some
adversaries, the adversaries may pre-create some correlated random numbers and store them into Alice and
Bob’s measurement devices to implement memory-stick
attack.) The private randomness in the output strings a
and b is quantified by min-entropy, and the details could
be find in the supplementary information of [8].
y
x
ρAB
Alice
Bob
a
b
FIG. 1. Randomness certified by Bell nonlocality
III.
RANDOMNESS CERTIFIED BY ANY
ENTANGLED STATE
This section is the main part of our paper. It shows
how to generate certified randomness by any entangled
state, and how to quantify the output randomness.
The randomness generation protocol mentioned in section II takes advantage of the Bell nonlocality in entangled states. However, some entangled states may admit
a local hidden variable model, and they do not contain
Bell nonlocality[15][16]. In order to make every entangled
state useful in certified randomness generation protocol,
we must use other property of entangled states instead
of Bell nonlocality. In [9], the author proved that with
nonorthogonal states as input, the measurement correlation in every entangled state cannot be created by any
separate state with local operations and shared randomness(LOSR). The protocol introduced in our paper uses
this property of entangled state. For simplicity, we use
bipartite scenario to illustrate our protocol.
First, we introduce a Bell-like inequality[10]. In bipartite scenario, Alice and Bob are two separated parties,
they share an entangled state ρAB , then they both input some quantum states {τs } and {ωt } into the measurement devices, the measurement devices then output
measurement results a and b. The correlation between a
and b is represented as P (a, b|τs , ωt ).
0
0
0
0
B
P (a, b|τs , ωt ) = T r[(PaA A ⊗QB
)(τsA ⊗ρAB ⊗ωtB )] (2)
b
where A, B is the Hilbert space of state ρAB , A0 , B 0
is the
Hilbert
space of nonorthogonal states τs , ωt , and
0
0
B
PaA A , QB
are the POVMs chosen by Alice and Bob.
b
By making suitable joint measurement on their respective part0 of ρAB0 and on the input quantum states τs , ωt
B
with PaA A , QB
, Alice and Bob can obtain the correb
lation P (a, b|τs , ωt ), which cannot be explained without
entanglement[9][11]. This allows the existence of the following linear combination of P (a, b|τs , ωt ) :
X
IρAB =
βs,t,a,b P (a, b|τs , ωt ) < 0
(3)
a,b,s,t
where βs,t,a,b are some real coefficients. The coefficients
3
βs,t,a,b could be obtained from the decomposition of entanglement witness. Because of the completeness of {τs }
and {ωt }, the entanglement witness W of ρAB could be
decomposed as
X
W =
βs,t τs> ⊗ ωt>
(4)
τs
ωt
Alice
s,t
with βs,t,1,1 = βs,t , and βs,t,a,b = 0, (a, b) 6= (1, 1), formula 3 becomes:
X
IρAB =
βs,t P (1, 1|τs , ωt ) < 0
(5)
a
ρAB
Bob
b
s,t
It is proved in [11] that this inequality is greater than
or equal to zero for any separated state with any possible POVM. For any entangled state, with the following
POVMs,
0
0
0
+
AA
P1A A = |Φ+
= I − P1A A
dA ihΦdA |, P0
Q1B
0
B
+
B
= |Φ+
dB ihΦdB |, Q0
0
B
= I − QB
1
0
B
FIG. 2. Randomness certified by any entangled state. In
this protocol, the input τs , ωt are nonorthogonal states, and
the measurement devices of Alice and Bob are not allowed to
communicate with each other. As long as the Bell-like value
IρAB is less than zero, Alice and Bob could extract certified
randomness from the measurement results a and b.
(6)
√
PdA −1
where
=
|ii ⊗ |ii/ dA , and |Φ+
dB i =
√ i=0
PdB −1
j=0 |ji ⊗ |ji/ dB . The Bell-like value IρAB will be:
X
IρAB =
βs,t P (1, 1|τs , ωt )
|Φ+
dA i
The input quantum states τs and ωt are nonorthogonal,
they could be any nonorthogonal state which forms a
complete basis. Without losing generality, they could
be:
τs = (I + v~s · ~σ )/2
ωt = (I + v~t · ~σ )/2,
s,t
T r(W ρAB )
dA dB
<0
=
(7)
which is less than zero for entangled states, this
means the measurement results a, b must be unpredictable, otherwise there would be contradiction in quantum theory[11].
With the above Bell-like correlation 5, we can design
a certified randomness generation protocol, where any
entangled state is useful to generate certified randomness. The protocol is shown in FIG.2 . In this protocol, Alice and Bob randomly input some nonorthogonal
states τs , ωt to the measurement devices, and they measure these nonorthogonal states jointly with their respective part of ρAB by the POVM 6. If the linear combination IρAB of the correlation P (1, 1|τs , ωt ) is less than zero,
then the output strings a, b contain randomness which
could not be predicted by the untrusted measurement
devices. The certified randomness in a, b is quantified by
min-entropy H∞ (AB|ST ).
We take Werner states as an example to show how to
get the lower bound of H∞ (AB|ST ). For Werner states,
1−z
I + z|Φ+ ihΦ+ |
(8)
4
√
where |Φ+ i is Bell state (|00i + |11i)/ 2. The entanglement witness is:
ρz =
Wρ z =
1
I − |Φ+ ihΦ+ |
2
(9)
(10)
where ~σ
= √
(σx , σy , σz ), s, t
∈ √
{0, 1, 2, 3},
−1,
−1)/
3, v~2 =
v~0 = (1,√1, 1)/ 3, v~1 = (1,
√
(−1, 1, −1)/ 3, v~3 = (−1, −1, 1)/ 3.
Next, we discuss the guessing probability of the measurement results. It is illustrated in the Appendix that
the average guessing probability of the measurement results a, b can be written as:
pguess =
Iρ0 − Iρz
Iρ
× 1 + z × maxPρ0 (a, b|τs , ωt ) (11)
Iρ0
Iρ0
And the upper bound of this average guessing probability
pguess is
pguess ≤
Iρoptimal − Iρz
Iρz
+
× Pρmax (a, b|τs , ωt )
Iρoptimal
Iρoptimal
For Werner states ρz , Iρz = (1 − 3z)/16, Iρoptimal =
I|Φ+ i = −1/8. Pρmax (a, b|τs , ωt ) is the maximum guessing probability with input 10 and POVM 6, it√is proved
in the Appendix that Pρmax (a, b|τs , ωt ) ≤ 9+16 3 . Then
the upper bound of the average guessing probability is
√
Iρ|Φ+ i − Iρz
Iρz
9+ 3
pguess ≤
+
×
Iρ|Φ+ i
Iρ|Φ+ i
16
√
−1/8 − (1 − 3z)/16 (1 − 3z)/16 9 + 3
=
+
×
−1/8
−1/8
16
√
7 − 3 3z − 1
=1 −
×
16
2
(12)
4
the measurement devices[11], so the generating devices
of the nonorthogonal states should be trusted. Thirdly,
any classical communication between the measurement
devices is forbidden.
b
a
a
FIG. 3. The lower bound of output randomness for Werner
states ρz . The analytical upper bound
of the average guessing
√
probability is pguess ≤ 1 − 7−16 3 × 3z−1
, and this average
2
guessing probability is the optimal one allowed by√ quantum
theory. Then we can generate at least −log2 (1− 7−16 3 × 3z−1
)
2
bits of certified randomness, and this randomness is the curve
a in the figure. Curve b represents the lower bound obtained
from semidefinite programs(SDP).
According to the definition of min-entropy, the lower
bound of the output randomness for Werner states ρz is
√
7 − 3 3z − 1
×
)
(13)
H∞ (AB|ST ) ≥ −log2 (1 −
16
2
This lower bound is shown in FIG.3.
From FIG.3 we can see that, for any entangled Werner
state, the output randomness is greater than zero, which
means any entanglement is useful in our randomness generation protocol.
Comparing the lower bound of our randomness generation protocol with the analytical lower bound of the output randomness in [8], our randomness generation protocol could take advantage of all entangled states by introducing nonorthogonal states as input, and in low degree
entanglement, our protocol could generate more secured
randomness. The compare of the output randomness is
shown in FIG.4
There are some constrains in our randomness generation protocol. Firstly, the corresponding entanglement witness must be decomposable by the nonorthogonal states and the nonorthogonal states cannot be distinguished by the measurement devices. These conditions
require that the input quantum states τs and ωt must
be nonorthogonal states which form a complete basis,
and the labels of these nonorthogonal states must not
be revealed to the measurement devices. Secondly, it is
more reasonable to trust the source devices other than
c
FIG. 4. The compare of output certified randomness of ρz in
our protocol and in Bell protocol[8]. The lower bound
of our
√
protocol is given by H∞ (AB|ST ) ≥ −log2 (1 − 7−16 3 × 3z−1
),
2
and is represented as blue line(curve a) in the figure. The
red line(curve c) is the lower bound of output randomness
certified by Bell’s
p theorem, and its formula is H∞ (AB|XY ) ≥
[1 − log2√(1 + 2 − I 2 /4)], the relationship between z and I
is I = 2 2z.
IV.
CONCLUSION
In this article, we construct a protocol to generate certified randomness by any entangled state. Similar protocol was also presented in [12], but the authors didn’t give
a general lower bound of their protocol, and the lower
bound obtained in[12] is not convincing without considering the secret entanglement shared between untrusted
measurement devices. In our article, we take this situation into consideration and we obtain a general lower
bound for this kind of randomness generation protocol.
Comparing to the protocol in [8], our protocol can
take advantage of any entangled state, and our protocol is based on entanglement theory, the correctness of
our protocol is guaranteed by the validity of quantum
physics. The randomness generation protocol in our paper shows us the deep connection between entanglement
and randomness, and further open questions may arise,
such as what’s the relationship between randomness and
entanglement, are they equal to each other? How do
nonorthogonal states assist entangled states, which admit a local hidden variable model, to generate certified
randomness?
5
[1] X. Ma, X. Yuan, Z. Cao, B. Qi, and Z. Zhang, Nature
Publishing Group , 1 (2016).
[2] F. Xu, B. Qi, X. Ma, H. Xu, H. Zheng, and H.-K. Lo,
Opt. Express 20, 12366 (2012).
[3] Z. L. Yuan, M. Lucamarini, J. F. Dynes, B. Frhlich,
A. Plews, and A. J. Shields, Applied Physics Letters 104,
261112 (2014), http://dx.doi.org/10.1063/1.4886761.
[4] Y.-Q. Nie, L. Huang, Y. Liu, F. Payne, J. Zhang, and
J.-W. Pan, Review of Scientific Instruments 86, 063105
(2015), http://dx.doi.org/10.1063/1.4922417.
[5] D. G. Marangon, G. Vallone, and P. Villoresi, Phys. Rev.
Lett. 118, 060503 (2017).
[6] A. Acı́n and L. Masanes, Nature 540, 213 (2016).
[7] R. Colbeck, arXiv preprint arXiv:0911.3814 (2009).
[8] S. Pironio, A. Acı́n, S. Massar, A. B. de La Giroday,
D. N. Matsukevich, P. Maunz, S. Olmschenk, D. Hayes,
L. Luo, T. A. Manning, et al., Nature 464, 1021 (2010).
[9] F. Buscemi, Phys. Rev. Lett. 108, 200401 (2012).
[10] D. Rosset, C. Branciard, N. Gisin, and Y.-C. Liang, New
Journal of Physics 15, 053025 (2013).
[11] C. Branciard, D. Rosset, Y.-C. Liang, and N. Gisin,
Phys. Rev. Lett. 110, 060405 (2013).
[12] A. Chaturvedi and M. Banik, EPL (Europhysics Letters)
112, 30003 (2015).
[13] J. S. Bell, “On the einstein podolsky rosen paradox,”
(1964).
[14] J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt,
Phys. Rev. Lett. 23, 880 (1969).
[15] R. F. Werner, Phys. Rev. A 40, 4277 (1989).
[16] N. Brunner, D. Cavalcanti, S. Pironio, V. Scarani, and
S. Wehner, Rev. Mod. Phys. 86, 419 (2014).
[17] X. Chen, X. Hu, and D. L. Zhou, Phys. Rev. A 95,
052326 (2017).
[18] A. Horn, Pacific J. Math. 12, 225 (1962).
[19] L. Vandenberghe and S. Boyd, SIAM review 38, 49
(1996).
[20] M. Navascués, S. Pironio, and A. Acı́n, Phys. Rev. Lett.
98, 010401 (2007).
[21] C. C. W. Lim, Phys. Rev. A 93, 020101 (2016).
APPENDIX
In this Appendix, we will show how to get the lower
bound of the output randomness in our randomness generation protocol.
Appendix A: The average guessing probability of
measurement results
The POVMs in our protocol are:
0
0
0
+
AA
P1A A = |Φ+
= I − P1A A
dA ihΦdA |, P0
QB
1
0
B
+
B
= |Φ+
dB ihΦdB |, Q0
0
B
= I − QB
1
0
B
(A1)
and the input nonorthogonal states are
τs = (I + v~s · ~σ )/2
ωt = (I + v~t · ~σ )/2,
(A2)
where ~σ
= √(σx , σy , σz ), s, t
= √ 0, 1, 2, 3,
−1,
−1)/
3, v~2 =
v~0 = (1,√1, 1)/ 3, v~1 = (1,
√
(−1, 1, −1)/ 3, v~3 = (−1, −1, 1)/ 3.
The average guessing probability of the measurement
results is :
Iρ0 − IρAB
Iρ
× 1 + AB × maxPρ0 (a, b|τs , ωt )
Iρ0
Iρ0
Iρoptimal − IρAB
IρAB
≤
+
× maxPρ0 (a, b|τs , ωt )
Iρoptimal
Iρoptimal
Iρ
− IρAB
IρAB
+
× Pρmax (a, b|τs , ωt )
≤ optimal
Iρoptimal
Iρoptimal
pguess =
where Iρ0 is the Bell-like value of entangled state ρ0 , which
shares the same given entanglement witness W of ρAB .
Pρ0 (a, b|τs , ωt ) is the correlation of the measurement results. We have Iρ0 ≤ IρAB < 0 and maxPρ0 (a, b|τs , ωt ) ≥
maxPρAB (a, b|τs , ωt ), otherwise, it’s meaningless for the
measurement devices to forge the measurement results.
Iρoptimal is the minimum Bell-like value for entanglement
witness W . Pρmax (a, b|τs , ωt )) is the maximum measurement correlation for any two-qubit state with POVM
(A1) and with nonorthogonal inputs (A2)
Next, we give an example to show the meaning of
the average guessing probability. Suppose the entangled
state shared between Alice and Bob is
ρz =
1−z
I + z|Φ+ ihΦ+ |
4
(A3)
the corresponding entanglement witness will be
Wρz =
1
I − |Φ+ ihΦ+ |
2
(A4)
6
deceive Alice and Bob. The entanglement witness of |Φ+ i
and ρz0 could both be Wρz , the corresponding Bell-like
value of |Φ+ i is −1/8, it’s much less than that of ρz0 .
Pρz (a, b|τs , ωt ) = zP|Φ+ i (a, b, |τs , ωt ) + (1 − z)PI (a, b|τs , ωt ) As long as the Bell-like value of the measurement results
are equal to Iρz0 , Alice and Bob will not detect any in(3 − 2a)(3 − 2b)
7 − 5a − 5b + 4ab
+ (1 − z)
=z
security in the measurement results. By using this fact,
12
16
the measurement devices can fake about 99% of the mea(3 − 2a)(3 − 2b)
(1 − 2a)(1 − 2b)
=
+z
surement results. These fake measurement results can be
16
48
stored in the measurement devices without being known
and the maximum Pρz (a, b|τs , ωt ) is
by Alice and Bob, and the Bell-like value of this fake part
can be as less as zero. For the rest 1%, the measurement
maxa,b Pρz (a, b|τs , ωt )
devices choose to jointly measure (A2) with the built-in
= Pρz (0, 0|τs , ωt )
state |Φ+ i, the Bell-like value of these measurement results is I|Φ+ i = −1/8. From the view of Alice and Bob,
1
9
(A5)
+z
=
the average Bell-like value of the whole measurement re16
48
sults would be
27 + z
=
48
1
1
Iavaerage = 0 × 99% + − × 1% = −
This equation shows that the guessing probability is be8
800
coming larger with the increasing of z. It seems without
entanglement, the output randomness is better. Actuwhich is the same as Iρz0 . The fake part of the measureally, this is not the case. From formula (A5), the guessment results is known by the measurement devices, the
ing probability is maximum when the shared entangled
guessing probability could be 1. For the measurement
state is Bell state, so the optimal cheating strategy for
results of the built-in state |Φ+ i, the maximum guessing
the measurement devices is to fake the measurement reprobability is 7/12. Then the average guessing probabilsults with built-in state |Φ+ i. Here is the detail, suppose
ity of the whole measurement results for the measurement
z = 0.34, then
devices should be 0.99×1+0.01×7/12 = 0.9958, and the
corresponding lower bound of the output randomness is
I
−log2 (0.9958) = 0.006.
ρz0 = 0.66 × + 0.34 × |Φ+ ihΦ+ |
(A6)
4
One thing we should note here is that, the untrusted
If the measurement devices do the corresponding meameasurement devices could not forge all the measurement
surement honestly, the maximum guessing probability of
results, there must be some truly random measurement
the measurement results will be
results from the shared entanglement between measurement devices to make sure the average Bell-like value is
maxa,b Pρz0 (a, b|τs , ωt )
less than zero. Otherwise, without sharing entanglement,
the measurement devices could make Iρz < 0, which is
= Pρz0 (0, 0|τs , ωt )
(A7)
contradict with quantum physics[11].
27 + 0.34
=
According to the analysis above, the average guessing
48
probability of the measurement results for Werner states
the output randomness is −log2 ( 27+0.34
) = 0.812, it’s
A3 should be:
48
quite large for a low degree entangled state like (A6).
However, most part of the output randomness cannot
I|Φ+ i − Iρz
Iρz
pguess =
×1+
× maxP|Φ+ i (a, b|τs , ωt )
be certified, the untrusted measurement devices could
I|Φ+ i
I|Φ+ i
use the following trick to jeopardize the security in the
measurement results:
The measurement correlation maxP|Φ+ i (a, b|τs , ωt ) may
When Alice and Bob used entangled state (A6) to
not be the optimal one, as there are some entangled
generate certified random number, the expected Bell-like
states’ measurement correlation larger than Bell state.
value Iρz0 is
We use Pρmax (a, b|τs , ωt )) to represent the maximum correlation allowed by quantum physics, then the upper
T r(W ρz0 )
Iρz0 =
bound of the average guessing probability is
4
1 − 3 × 0.34
(A8)
=
I|Φ+ i − Iρz
Iρz
16
pguess =
×1+
× maxP|Φ+ i (a, b|τs , ωt )
1
I|Φ+ i
I|Φ+ i
=−
800
I|Φ+ i − Iρz
Iρz
≤
+
× Pρmax (a, b|τs , ωt ))
I|Φ+ i
I|Φ+ i
Without being detected, the measurement devices
could take advantage of their built-in Bell state |Φ+ i to
(A9)
With the POVM (A1) and with nonorthogonal inputs
(A2),the correlations of measurement results are
7
each basis:
Appendix B: The maximum measurement
correlation Pρmax (a, b|τs , ωt ))
0
In this appendix, we calculate Pρmax (a, b|τs , ωt )). It is
said in [17] that any two-qubit density matrix ρAB is an
element in the linear space expanded by {σi ⊗ σj , i, j ∈
0, 1, 2, 3}(σ0 = I2×2 ). T r(ρAB ) = 1, so the coefficient
of σ0 ⊗ σ0 is 14 . Thus all the two-qubit density matrices has 15 different coefficients, they are c3i +j , i, j ∈
{0, 1, 2, 3}, (i, j) 6= {0, 0}. ρAB could be written as:
0
B
P (a, b|τs , ωt ) = T r[(PaA A ⊗ QB
)(τs ⊗ ρAB ⊗ ωt )]
b
X
0
0
B
=T r[(PaA A ⊗ QB
)(τs ⊗ (
c3i +j σi ⊗ σj ) ⊗ ωt )]
b
i,j=0,1,2,3
X
9
c3i +j P (σi ⊗ σj ))]
= +(
16
i,j=0,1,2,3
(i,j)6=(0,0)
≤
X
9
+
c i maxP (σi ⊗ σj )
16 i,j=0,1,2,3 3 +j
(i,j)6=(0,0)
ρAB =
X
1
σ0 ⊗ σ0 +
c3i +j σi ⊗ σj
4
i,j=0,1,2,3
(B1)
ρAB is a quantum state, which means the minimum
eigenvalue λminAB ≥ 0. From the property of eigenvalue,
the eigenvalue of ρAB is less than the eigenvalue sum of
its each part [18]:
X
c3i +j λ3i +j
i,j=0,1,2,3
The eigenvalues of σ0 ⊗ σ0 are {1, 1, 1, 1}, the eigenvalues
of σi ⊗ σj (i, j) 6= (0, 0) are {−1, −1, 1, 1}. We already
know the coefficient of σ0 ⊗ σ0 is 41 , then the range of
ρAB ’s eigenvalues could be represented as
X
λAB1 ≤ 1/4 −
λAB2 ≤ 1/4 −
X
c3i +j ,
(B2)
c3i +j ,
X
where Ioptimal is the minimum Bell-like value for the
given entanglement witness of entangled state ρAB .
From the perspective of the measurement devices,
searching for the optimal guessing probability pguess is
equal to solve the next optimization problem:
max pguess
i,j=0,1,2,3
(i,j)6=(0,0)
λAB4 ≤ 1/4 +
The similar analysis could be applied to any entangled state ρAB shared between Alice and Bob, thus the
upper bound of the average guessing probability in our
randomness generation protocol is:
√
Iρoptimal − IρAB
IρAB
9+ 3
pguess ≤
+
×
Iρoptimal
Iρoptimal
16
Appendix C: SDP output randomness
i,j=0,1,2,3
(i,j)6=(0,0)
λAB3 ≤ 1/4 +
0
B
where P (σi ⊗σj ) = T r[(PaA A ⊗QB
)(τs ⊗(σi ⊗σj )⊗ωt )].
b
With the given POVM(A1) and with nonorthogonal
states(A2), we can get maxP (σi ⊗ σj ) =
√
3
measurement correlation is
4 , and the maximum
√
Pρmax (a, b|τs , ωt ) = 9+16 3 . Then the upper bound of the
average guessing probability (A9) is
√
Iρ|Φ+ i − Iρz
Iρz
9+ 3
pguess ≤
+
×
Iρ|Φ+ i
Iρ|Φ+ i
16
c3i +j ,
i,j=0,1,2,3
(i,j)6=(0,0)
X
9
maxP (σi ⊗ σj )
+
16
4
0
(i,j)6=(0,0)
λAB ≤
≤
subject to
c3i +j
pguess =
i,j=0,1,2,3
(i,j)6=(0,0)
Iρ0 − IρAB
Iρ
+ AB × maxPρ0 (a, b|τs , ωt )
Iρ0
Iρ0
X
IρAB =
βs,t PρAB (1, 1|τs , ωt )
s,t
Because the minimum element of λAB is greater than or
equal to zero, we have
Iρ0 =
X
βs,t Pρ0 (1, 1|τs , ωt )
s,t
−1/4 ≤
X
c3i +j ≤ 1/4
(B3)
i,j=0,1,2,3
(i,j)6=(0,0)
The correlation P (a, b|τs , ωt ) is the correlation sum of
Iρ0 ≤ IρAB < 0
maxPρ0 (a, b|τs , ωt ) ≥ maxPρAB (a, b|τs , ωt )
This problem can be formulating as semidefinite
programs(SDP)[19][20][21], the lower bound of the output randomness obtained from SPD is shown as curve b
in FIG.3.
| 7cs.IT
|
Hierarchical Memory Management for Mutable State
arXiv:1801.04618v2 [cs.PL] 18 Feb 2018
Extended Technical Appendix
Adrien Guatto
Sam Westrick
Ram Raghunathan
Carnegie Mellon University
[email protected]
Carnegie Mellon University
[email protected]
Carnegie Mellon University
[email protected]
Umut Acar
Matthew Fluet
Carnegie Mellon University
[email protected]
Rochester Institute of Technology
[email protected]
Abstract
strongly typed functional languages is their ability to distinguish between pure and impure code. This aids in writing
correct parallel programs by making it easier to avoid race
conditions, which can become a formidable challenge in
languages whose type systems don’t distinguish between
mutable and immutable data.
In the sequential realm, functional languages compete well
with other garbage-collected languages such as Java and Go,
often running within a factor of 2 or 3 and sometimes even
faster. In many cases, functional languages even compete
well with the C family, where memory is managed by the
programmer [34].
In the parallel realm, however, the gap between functional
and imperative languages is significantly larger. One reason for this is the (poor) support for mutation in parallel
functional languages. A reality of modern hardware is that
imperative algorithms can perform significantly better than
pure functional algorithms by using constant-time random
accesses and updates. Even when a parallel algorithm has a
pure functional interface (immutable inputs and immutable
outputs), it can be more efficient to use mutation internally.
For example, the efficiency of a pure functional parallel
merge-sort can be significantly improved by reverting to
a sequential imperative quick-sort for small inputs. Committing to pure functional algorithms only does not completely
avoid mutation: a language run-time system uses mutation to
implement crucial facilities such as (thread) schedulers and
synchronization primitives, which require communication
between processors via shared memory.
Even though mutation is crucial for efficiency, it remains
poorly supported in parallel functional languages and remains as an active area of research (see Section 6). For example, the Manticore project has developed rich extensions to
the ML language to support parallelism but has focused on
purely functional code where the programmer cannot use
mutation [7, 16]. Other ML dialects such as OCaml and SML#
continue to remain primarily sequential languages, though
there is ongoing work in extending them to support modern
parallelism features. Like Manticore, Haskell has a relatively
rich set of parallelism features and its runtime must support
efficient mutation [27]. Writing efficient parallel programs
It is well known that modern functional programming languages are naturally amenable to parallel programming.
Achieving efficient parallelism using functional languages,
however, remains difficult. Perhaps the most important reason for this is their lack of support for efficient in-place
updates, i.e., mutation, which is important for the implementation of both parallel algorithms and the run-time system
services (e.g., schedulers and synchronization primitives)
used to execute them.
In this paper, we propose techniques for efficient mutation
in parallel functional languages. To this end, we couple the
memory manager with the thread scheduler to make reading
and updating data allocated by nested threads efficient. We
describe the key algorithms behind our technique, implement them in the MLton Standard ML compiler, and present
an empirical evaluation. Our experiments show that the approach performs well, significantly improving efficiency over
existing functional language implementations.
CCS Concepts • Software and its engineering →
Garbage collection; Parallel programming languages;
Functional languages;
Keywords parallel functional language implementation,
garbage collection, hierarchical heaps, mutation, promotion
1
Introduction
With the proliferation of parallel hardware, functional programming languages, such as Haskell and the ML family
(OCaml, Standard ML), have received much attention from
academia and industry. Even non-functional languages today such as C++, Python, and Swift support certain features
of functional languages, including higher-order functions
and, sometimes, rich type systems. An important virtue of
PPoPP ’18, February 24–28, 2018, Vienna, Austria
© 2018 Copyright held by the owner/author(s). Publication rights licensed
to the Association for Computing Machinery.
This is the author’s version of the work. It is posted here for your personal
use. Not for redistribution. The definitive Version of Record was published
in PPoPP ’18: Principles and Practice of Parallel Programming, February 24–28,
2018, Vienna, Austria, https://doi.org/10.1145/3178487.3178494.
1
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
val GRAIN = ...
fun inplaceQSort s = ...
fun msort s =
i f Seq.length s <= GRAIN
then l e t val a = Seq.toArray s
val () = inplaceQSort a
in Seq.fromArray a end
else l e t val (l, r) = Seq.splitMid s
val (l', r') = par (msort l, msort r)
in Seq.merge (l', r') end
in Haskell, however, remains difficult in part because of lazy
evaluation [29].
This state of the art raises the question of whether functional programming can be extended to support mutable data
and parallelism efficiently. At the highest level of abstraction,
this is a challenging problem because its parts—parallelism
and efficient memory management—are individually challenging. The problem is further complicated by the fact that
functional languages allocate and reclaim memory at a much
faster rate than other languages [5–7, 13, 14, 18, 19, 27].
In this paper, we propose techniques for supporting mutable data efficiently within the run-time system of nestedparallel functional languages, focusing on strict languages
in the style of the ML family. Our approach builds upon
that of hierarchical heaps, a memory management technique
developed in prior work [31]. The basic idea is to organize
memory so that it mirrors the structure of the parallel computation. Specifically, each thread is assigned its own heap in a
hierarchy (tree) of heaps which grows and shrinks as threads
are forked and joined. Threads allocate data in their own
heaps, and can read and update objects in ancestor heaps
(including their own). A key invariant is that data in nonancestor heaps remains unreachable to a thread. To enforce
this invariant, we propose a promotion technique for copying
data upwards in the hierarchy as necessary.
Our approach has several important benefits. First, threads
can allocate, read, and update mutable objects in their heap,
without synchronization or copying. This allows local mutable objects to be used efficiently. Second, because heaps are
associated with threads rather than processors, a thread can
be migrated between processors without copying data. These
two properties contrast with the predominant approach to
memory management in parallel functional languages with
local heaps, where both mutation and thread migration require copying [7, 13, 14, 27]. Third, our techniques introduce
no overhead for reads of immutable objects, which are pervasive in functional languages. Finally, any thread can collect
its heap independently and, more generally, any subtree of
heaps in the hierarchy could be collected independently.
The contributions of this paper include techniques and
algorithms for handling mutable data in hierarchichal heaps
(Section 3), an implementation extending the MLton wholeprogram optimizing compiler for Standard ML [28], and an
empirical evaluation considering a number of both pure and
imperative benchmarks (Section 4). Our results show that
these techniques can be implemented efficiently and can
perform well in practice.
2
Figure 1. Code for parallel imperative merge sort.
in
< 3, 2, 1, 9 >
main ( )
s
s
a
msort in[0..2)
3
2
msort in[0..4)
l
< 3, 2 >
r
< 1, 9 >
s
msort in[2..4)
a
1
9
Figure 2. Hierarchical heap example for msort.
Consider the parallel merge sort in Figure 1. The implementation uses an immutable sequence data structure provided by a module Seq, whose details we omit. To sort an
input sequence, the function msort first checks its length. If
the length is less than some constant GRAIN, then the function
uses an imperative in-place sequential quicksort to sort the
input. Otherwise, the input is split into two halves and two
recursive calls are performed in parallel.
Parallelism is exposed by the programmer through the par
construct, which creates new tasks. Initially, there is only one
task (corresponding to the execution of the entire program);
thus all user code implicitly runs within the context of some
task. With par, a task may spawn two new tasks. This establishes a parent/child nesting relationship where a parent
task must wait until both its children complete before continuing its execution. Tasks are managed by a scheduler, which
strives to minimize the completion time of the program by
mapping tasks to processors.
As in Raghunathan et al. [31], to support parallel automatic memory management, each task is assigned its own
heap in which it allocates new data. Heaps are organized
into a hierarchy (tree) with the same parent/child relationships as their associated tasks. When both children of a task
complete, their heaps are logically merged with the parent
heap, allowing the parent to continue with child data stored
locally as though the children had never existed.
Figure 2 illustrates an example where msort is called with
the input sequence ⟨3, 2, 1, 9⟩ and GRAIN = 2. Tasks are drawn
Overview
We present a brief overview of our techniques, using a simple example to illustrate both the programming model and
details of memory management. In the process, we introduce
terminology that will be used throughout the paper.
2
Hierarchical Memory Management for Mutable State
as ellipses connected by straight, black arrows pointing from
parent to child. The grey boxes drawn around each task delimit its heap, and red curved arrows show pointers in memory. At the root of the hierarchy is the initial task main which
allocated the input sequence and then called msort. The middle task corresponds to this initial call of msort, which split
the input into two sequences l and r, allocated locally. The
two leaf tasks are the parallel recursive calls of msort, which
have copied their inputs to local arrays a.
Consider the following definition. We say that the hierarchy is disentangled if, for any pointer from an object x in
heap h x to another object y in heap hy , hy is either equal to
or an ancestor of h x . In other words, in order to be disentangled, the hierarchy must not contain down-pointers that
point from ancestor to descendant, nor may it contain crosspointers that point between unrelated heaps (such as between
two siblings). In a disentangled hierarchy, the lack of pointers into leaf heaps allows us to garbage collect the leaves
independently, without synchronizing with other tasks, and
in parallel with other leaf-heap collections. More generally,
any two disjoint sub-trees of heaps may be collected independently in parallel (although the tasks within a sub-tree
must cooperate). The execution of msort in Figure 2 is disentangled, thus the two leaves of the hierarchy could both
be independently collected in parallel; for example, collecting the array a after computing Seq.fromArray a but before
joining with the parent task. It can be shown that all purely
functional programs naturally exhibit disentanglement [31].
In the presence of arbitrary mutation, disentanglement is
not guaranteed. Consider for example a mutable reference r
which is allocated at a parent task and then passed to two
child tasks. One child could update r to point to a locally
allocated object, creating a down-pointer. The second child
could then read from r to create a cross-pointer.
To enforce disentanglement, we propose a promotion technique. The basic idea behind promotion is to detect when a
down-pointer would be created in the hierarchy, and first
promote (copy) the lower object upwards so that it lies in
the same heap as the mutable object.
There are a number of challenges associated with promotion. In particular, promotion duplicates objects, which
complicates the identity of mutable objects. We solve this
issue by distinguishing one copy as the master copy, to which
all accesses are redirected through forwarding pointers. Additionally, when data being promoted contains pointers to
other objects, all transitively reachable data might need to be
promoted. This introduces concurrency into the system even
when none exists in the user code, since a task might access
an object which is in scope of an in-progress promotion.
In our implementation, we prioritize the efficiency of updates to local objects. This facilitates an important idiom
of practical parallelism where the overhead of parallelism
is amortized by switching to a fast sequential algorithm on
small inputs. As the fastest sequential algorithms on modern
1
2
3
4
5
6
7
8
9
10
11
type task
type data
type objptr
type field
type thunk = unit -> objptr
function forkjoin: thunk * thunk -> objptr * objptr
function alloc: field list -> objptr
function readImmutable: objptr * field -> data
function readMutable: objptr * field -> data
function writeNonptr: objptr * field * data -> unit
function writePtr: objptr * field * objptr -> unit
Figure 3. High-Level Operations.
hardware are often imperative, local updates are thus crucial for efficiency. The msort program exemplifies this idiom,
utilizing a fast imperative quicksort on small inputs. Indeed,
in our results (Section 4.4), we see that msort can be up to
twice as fast as a purely functional alternative.
3
3.1
Hierarchical Heaps for Mutable Data
Programming Model
Our proposal deals with the full ML language, including
mutable data, extended with nested task parallelism. We extended an existing ML compiler to reduce this programming
model to a small set of high-level operations that we implemented in the runtime system. In what follows, we describe
the interface of these high-level operations in abstract terms
and explain the challenges we face when implementing them
in the context of hierarchical heaps.
High-Level Types. Figure 3 describes the types and operations we use to implement nested-parallel ML programs.
While such operations are typically programmed in C or
some other low-level language, here for the sake of simplicity we specify them in ML-like pseudocode.
Tasks compute on data. The exact definition of data
does not matter at this level of description; we can assume
that it consists in machine integers, floating-point numbers, pointers, etc. Amongst general data, we distinguish
the type objptr of object pointers, that is, of pointers to allocated objects, or simply objects. Objects correspond to ML
data types allocated during the course of execution, such as
cons cells. For our purposes, an object consists of a finite list
of fields storing its content. Objects are used for communication between tasks. The body of a task is a thunk, that is, a
function expecting no argument and returning its result as
an object pointer.
High-Level Operations. The types described above are
used through six high-level operations. One of them deals
with task management, the others with objects.
The compiler elaborates the par keyword into calls to
the forkjoin operation. This operation takes two thunks,
3
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
function currentTask: unit -> task
function getField: objptr -> field -> ptr
function fields: objptr -> field list
function ptrFields: objptr -> field list
function nonptrFields: objptr -> field list
function fwdPtr: objptr -> ptr
function hasFwdPtr: objptr -> bool
type heap
function heapOfTask: task -> heap
function newChildHeap: heap -> heap
function joinHeap: heap * heap -> unit
function depth: heap -> int
function freshObj: heap * field list -> objptr
function heapOf: objptr -> heap
function lock: heap -> { READ, WRITE } -> unit
function unlock: heap -> unit
3.2
Low-Level Primitives.
Our algorithms implement the interface of Figure 3 using
hierarchical heaps. To do so, they rely on a number of lowlevel operations (Figure 4) provided by the runtime system.
Task-Related Operations. Calling currentTask retrieves
the currently running task. Since user code is always implicitly running within some task, this never fails.
Memory-Related Operations. The getField operation returns a pointer to the specified field in an object pointer. This
pointer does not point to (the beginning of) an object, but
rather inside it. For this reason, the result type of getField
is the abstract type ptr, rather than objptr.
Calling fields(o) operation returns all the fields of the
memory object. Its variants ptrFields and nonptrFields
return respectively the fields which hold object pointers and
those which do not.
Finally, each object comes equipped with a special field
for storing a forwarding pointer. Forwarding pointers are a
classic ingredient used in copying collectors [22] to perform
bookkeeping during collection. An object may or may not
have a valid forwarding pointer in its dedicated field. This
can be tested using the hasFwdPtr operation. The address
of the field can be obtained by calling fwdPtr.
Figure 4. Low-level Primitives.
creates one task for each, and runs both tasks to completion,
in parallel. This establishes a nesting relationship: we say
that the task calling forkjoin is the parent task, while the
tasks created by the operation are the children tasks. The
parent task is suspended until both children return, at which
point it receives their results and resumes.
Allocation arises from a variety of ML features: constructors of algebraic data types, explicit initialization of references and arrays, closures, etc. The alloc operation allocates
a new object. The caller describes the list of fields of the new
object and receives a pointer to it as a result.
The remaining primitives deal with reading from and writing to objects. Since the type system of ML distinguishes
between mutable and immutable data, we have several reading and writing operations, depending on the type of the data
being read or written. Later in this section, we will exploit
these distinctions for efficiency purposes.
First, we distinguish between reading immutable data with
the readImmutable operation and reading mutable data with
the readMutable operation. Both operations take a pair of
an object pointer and a field descriptor, and read the data
held in the corresponding field of the object. If the field does
not exist, the result is undefined.
Second, when writing to mutable data, we distinguish
between writing non-pointer data (e.g., machine integers)
and writing object pointers. The former takes the object
and field to be written to, as well as the data to write. The
latter takes the same argument, except that the data to write
must be an object pointer. Note that, in the type signature
of writeNonptr, we specify the general data type, which
includes pointers; however, the behavior of the function is
undefined if it is actually passed an object pointer.
Heap-Related Operations. The first group of operations
on heaps (abstract type heap) manages their relationships
with tasks as well as with each other. Calling heapOfTask(t)
returns the heap associated with task t. Like tasks, heaps
are organized into a hierarchy which grows and shrinks
through the newChildHeap and joinHeap primitives. Calling depth(h) returns the depth of a heap h in the hierarchy: the root is at depth zero, its children at depth one, etc.
The operation freshObj allocates a new object in the
specified heap and with the specified fields. Correspondingly,
the heap in which an object was allocated can be retrieved
by the heapOf operation.
Finally, in order to deal with concurrency issues, every
heap comes equipped with a readers-writers lock [20]. Such
locks can be held in reading mode by several threads simultaneously, but by a single thread in writing mode (excluding any other readers or writers). The lock associated
with a heap h can be acquired by calling lock(h,m), with m
being the READ or WRITE mode, and then released by calling unlock(h).
3.3
Implementation of the High-Level Primitives
We can now explain our algorithms as implementations of
the high-level operations in terms of the low-level ones,
starting with an overview of the challenges involved.
Challenges. Our goal is to ensure that disentanglement of the hierarchy holds in the presence of calls
to writePtr(obj,field,ptr). If implemented naïvely
4
Hierarchical Memory Management for Mutable State
1
2
3
4
5
6
7
function forkjoin (f, g) =
heap ← heapOfTask(currentTask ())
heap_f ← newChildHeap(heap)
heap_g ← newChildHeap(heap)
(r_f, r_g) ← run t ∈ {f,g} in heapt and wait
joinHeap(heap, heap_f); joinHeap(heap, heap_g)
return (r_f, r_g)
1
2
3
4
5
6
7
8
Figure 5. Fork/join.
9
10
11
as *getField(obj,field) ← ptr, such writes would
create down-pointers when heapOf(obj) is an ancestor of heapOf(ptr). In order to enforce disentanglement
in this situation, our implementation of writePtr promotes (copies) the object at ptr, as well as all objects reachable from it, into heapOf(obj). The address of the copy can
then be written into *getField(obj,field).
In the presence of repeated writes of ptr to objects held
in heaps of decreasing depth, the object at ptr might be
promoted several times. Ultimately, all copies but one should
be eliminated, and pointers to them updated to point to the
remaining one. Thus, we need a way to link an object with
its copies. We use forwarding pointers to do so, and organize
all the existing copies of an object into a singly-linked list.
All the copies of an object are equivalent as far as immutable fields are concerned, since by definition their content cannot change. In contrast, reads and writes of mutable
fields cannot treat all copies as equivalent, for this could
lead to lost updates (if an updated copy is then eliminated)
or inconsistent reads (if a task updates one copy and then
reads another one). Thus, we have to perform all mutable
accesses on a unique, authoritative copy of an object, that we
call its master copy. Given that all copies are arranged into
a linked-list, we choose to take the last element of that list,
that is the one in the shallowest heap, as the master copy.
A further difficulty is that several tasks might be trying to
copy the same data, and thus update the relevant forwarding pointers, simultaneously. A task might also be updating
forwarding pointers while another one is traversing them
to find the master copy. Our algorithms avoid synchronization issues by acquiring the locks of the heaps they traverse
during mutable accesses, from deepest to shallowest. Additionally, we propose several fast-paths avoiding locking in
common cases.
12
13
14
15
16
17
18
19
20
21
22
23
function alloc (fields) =
return freshObj(heapOfTask(currentTask()), fields)
function readImmutable (obj, field) =
return *getField(obj, field)
function findMaster (obj) =
while true:
while hasFwdPtr(obj): obj ← *fwdPtr(obj)
lock(heapOf(obj), READ)
i f not hasFwdPtr(obj): return obj
unlock(heap(obj))
function readMutable (obj, field) =
res ← *getField(obj, field)
i f not hasFwdPtr(obj): return res
obj ← findMaster(obj)
res ← *getField(obj, field)
unlock(heapOf(obj))
return res
function writeNonptr (obj, field, val) =
*getField(obj, field) ← val
i f not hasFwdPtr(obj): return
obj ← findMaster(obj)
*getField(obj, field) ← val
unlock(heapOf(obj))
Figure 6. Allocation, reads, non-pointer writes.
Allocation. Our implementation of alloc allocates the new
object in the heap of the currently running task (l. 2).
Reading Immutable Data. ML programs read immutable
data when destructuring values such as tuples or lists, e.g.
through projections or pattern matching. Since these are
very common operations, it is important to support them
efficiently. Fortunately, all the potential copies of the same
object hold the same value in their immutable fields, by
definition. Thus, readImmutable(obj,field) does not care
about the forwarding pointer slot of obj and can always
access the contents of field without any indirection (l. 4).
Finding Master Copies. After several promotions occur,
objects may exist in multiple copies linked in a chain by their
forwarding pointers. This chain has to be taken into account
when accessing mutable fields: intuitively, only the last copy,
called the master copy, holds up to date information.
The function findMaster(obj) returns a pointer to the
master copy of obj. Intuitively, it simply has to walk the chain
of forwarding pointers, starting from obj. However, while
doing so it might encounter a forwarding pointer installed
by a promotion that is still ongoing. In that case, we should
wait for the promotion to complete. We do so by acquiring
the lock of the heap to which the copy belongs. Since we
acquire the lock in shared mode, we do not block concurrent
calls to findMaster. In contrast, promotion always locks in
Fork-join. Figure 5 shows a naïve implementation of the
fork/join operation. First, create a heap attached to the heap
of the running task for each child task. Then, run in parallel
each t within its heap heapt , for t ∈ {f, g}, and wait for
them to complete (the exact realization of these operations
depending on the scheduler at hand). Finally, join both child
heap with their parent, and pass the results returned by
each task to the caller. Joining heaps can be done without
physically copying data.
5
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
function writePtr (obj, field, ptr) =
i f heapOf(obj) = heapOfTask(currentTask())
and not hasFwdPtr(obj):
*getField(obj, field) ← ptr
return
obj ← findMaster(obj)
i f depth(heapOf(obj)) ≥ depth(heapOf(ptr)):
*getField(obj, field) ← ptr
unlock(heap(obj))
return
unlock(heap(obj))
writePromote(obj, field, ptr)
function writePromote (obj, field, ptr) =
assert (depth(heapOf(obj)) < depth(heapOf(ptr)))
prev ← ptr
lock(heapOf(prev), WRITE)
while true:
for h from heapOf(prev) excluded
up to heapOf(obj) included: lock(h, WRITE)
i f not hasFwdPtr(obj): break
else:
prev ← obj
obj ← *fwdPtr(obj)
promotedPtr ← promote(heapOf(obj), ptr)
*getField(obj, field) ← promotedPtr
for h from heapOf(obj) included
down to heapOf(ptr) included: unlock(h)
function promote (heap, obj) =
i f depth(heapOf(obj)) ≤ depth(heap): return obj
i f hasFwdPtr(obj):
return promote(heap, *fwdPtr(obj))
newObj ← freshObj(heap, sizeOf(obj))
*fwdPtr(obj) = newObj
for field in nonptrFields(obj):
*getField(newObj, field) ←
*getField(obj, field)
for field in ptrFields(obj):
*getField(newObj, field) ←
promote(heap, *getField(obj, field))
return newObj
start walking the chain again (l. 10). Note that it is the caller’s
responsibility to unlock the heap.
Reading Mutable Data. Mutators can read a mutable field
in an object by calling readMutable(obj,field). It would be
correct to simply acquire the master copy, read the field,
and release the lock (l. 14-17). We add a fast path: read the
mutable field optimistically, then check for the absence of
a forwarding pointer (l. 12-13). This way, accessing mutable fields in objects without copies only takes a couple of
machine instructions.
Writing Non-pointer Data. Writing plain data such as integers or floating-pointer numbers cannot involve promotion,
and thus is always relatively cheap. The implementation
of writeNonptr mimicks that of readMutable. In the fast path,
we optimistically write the value val in field, and then check
whether obj was the master copy (l. 19-20). Otherwise, we
find the master copy, write the value, and unlock (l. 21-23).
Writing Pointer Data. Writing pointer data is the most
difficult case, as it might trigger a promotion. The code
for writePtr(obj,field,ptr), which attempts to write ptr
to field in obj, is given in Figure 7. We may have to promote ptr to the heap of obj if writing it directly would result
in entanglement. The algorithm can be decomposed into
three cases: a fast path, non-promoting writes, and promoting writes. Let us describe each of these in turn.
The fast path of writePtr (l. 2-5) writes ptr into obj only
if the latter has no forwarding pointer and is in the heap
of the currently running task. Since this heap is necessarily
a leaf in the hierarchy, promotion is never needed in this
case. This qualifies as a fast path since testing whether an
object pointer belongs to the currently running task can
be implemented much more efficiently than computing the
depth of an arbitrary heap.
On the slow path, we acquire the master copy of obj and
obtain the depth of its heap. If it is deeper in the hierarchy
than that of ptr, we are not creating entanglement, and can
simply perform the write and unlock (l. 6-10). Otherwise,
we have to promote, and thus unlock and call the dedicated
function writePromote(obj,field,ptr) (l. 11-12).
Figure 7. Pointer writes and promotion.
Promoting Writes. We proceed in three phases. First, we
lock in exclusive mode all the heaps on the path from ptr to
the master copy of obj from the bottom up (l. 15-23). Then,
we promote obj and write the address of the promoted copy
to field (l. 24-25). Finally, we unlock the path from top to
bottom (l. 26-27).
Acquiring the locks serve several purposes. Let us call h 1
the heap that contains obj and h 2 the heap containing the
master copy of obj. By acquiring the locks on the path
from h 1 to h 2 excluded, we take ownership of the forwarding pointer of any object we might need to promote. By
acquiring the lock of h 2 , we ensure that no concurrent call
to findMaster will return before we have finished promoting.
exclusive mode the heaps where it installs new forwarding
pointers, ensuring mutual exclusion.
Our implementation of findMaster uses the classic doublechecked locking pattern to reduce the cost of locking. As long
as we observe forwarding pointers, we move up, without
locking (l. 7). Once we see an object that is a candidate for
being the master copy, we acquire the lock of its heap, and
check whether the object has acquired a forwarding pointer
in the meantime; if not, it is definitely the master copy, and
can be returned (l. 8-9). Otherwise, we unlock the heap and
6
Hierarchical Memory Management for Mutable State
Read
Local
Distant
Promoted
Write
Immutable
Mutable
Non-pointer
Non-promoting
Promoting
✓
✓
✓
✓
✓
✓
✓
∼
✓
✓
✓
✓
∼
✓✓
∼
∼
≈
≈
✓
✓✓
∼
≈
single instruction
few instructions
single-heap locking
path locking + copying
Figure 8. Costs of memory operations.
The recursive function promote returns a pointer to a copy
of obj held in heap or one of its parents. If obj already resides
in heap or above, it can simply be returned (l. 29); if obj has
a forwarding pointer, we follow it (l. 30-31). Taken together,
these two tests ensure that objects with a chain of forwarding
pointers leading above heap are not copied. If both fail, we
have to introduce a new copy. We allocate a copy newObj
of obj in heap (l. 38), set the forwarding pointer slot of obj to
point to newObj (l. 32-33), and copy non-pointer fields (l. 3436). We recursively promote pointer fields, since they might
point strictly below heap (l. 37-39). At this point we can
return newObj (l. 40) since it belongs to heap, as do every
object reachable from it.
Note that while we have expressed promote as a recursive function for simplicity, it can be implemented using a
work list. In particular, we were careful to set the forwarding
pointer of obj before the recursive calls.
3. If the chain ends at an object that has no forwarding
pointer, then this object is in a from-space below h.
In the first two cases, the address of the copy can be reused
directly, whereas in the last case we have to introduce a
new copy. The second case corresponds to the elimination of
duplicates introduced during promotion. Note that since we
do not attempt to access copies outside of the collection zone,
no additional locking is needed. In the third case, we copy
the last element of the forwarding chain into the to-space
and update its forwarding pointer to point to this new copy.
By doing so, we effectively make sure that all pointers to
the object or its promoted copies will point to the new copy
created in the to-space.
4
We implemented our techniques by building upon the parallel MLton compiler developed in prior work [31], from which
we inherit the hierarchical heaps infrastructure, garbage collection policy, and scheduler. We extended this compiler
with support for general mutation by closely following the
algorithms described in Section 3. Further details on the
implementation can be found in Appendix B.
We evaluate our techniques by considering a number of
benchmarks compiled with several compilers for dialects of
parallel ML. Thanks to the shared ML language, our benchmarks remain mostly identical across compilers except for
minor compatibility edits. Our benchmark suite builds on
previous suites from parallel functional languages [7, 16],
and extends them, also to include imperative programs which
use mutable data. Our benchmarks use several standard implementations of data types such as sequences and graphs.
Unless stated otherwise, the elements of the sequences are
64-bit numeric types (integers or floating point) generated
randomly with a hash function.
Cost Summary. Figure 8 summarizes the costs of each memory operation in various situations. Columns correspond to
distinct memory operations, distinguishing between nonpromoting and promoting pointer writes. Rows classify objects being read from or written to: promoted objects are
those with forwarding pointers; local and distant objects
have no forwarding pointers and belong either to the heap
of the task performing the memory operation (local objects)
or to one of its ancestors (distant objects).
3.4
Implementation and Experiments
Promotion-Aware Copy Collection
Promotion introduces redundant copies of objects. However,
it is not difficult to eliminate these copies by piggybacking
on the classic semispace garbage collection algorithm. We
briefly explain our proposal; precise details are available
in Appendix A.
Assume that we are collecting a sub-tree starting at some
heap h; each heap below h (included) thus acquires a tospace. When examining an object in from-space, our collector
traverses its forwarding pointer chain, considering several
possibilities in turn.
Experimental Setup. For the measurements, we use a 72core (4 x 18 core Intel E7-8867 v4) Dell PowerEdge Server
with 1 Terabyte of memory. For the sequential baselines,
we use the whole-program optimizing MLton compiler [28],
which we label mlton. We compare all of our benchmarks
to the work of Blelloch, Spoonhower, and Harper [33], labeled mlton-spoonhower, which extends the MLton compiler
to support nested fork-join parallelism and parallel allocation, but utilizes sequential, stop-the-world collection. For
purely functional benchmarks, we also compare with the
1. If the chain leads into a to-space, it points to a copy
introduced during collection.
2. If the chain leads into a from-space that is strictly
above h in the hierarchy, it leads to a copy introduced
during promotion.
7
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
4.2
Manticore compiler [7, 16], labeled manticore, which provides for parallel functional programming by using syntax
similar to ours. We refer to our hierarchical-heaps compiler
as mlton-parmem.
When taking timing measurements, we exclude initialization times. All reported timings include GC times and are
reported as the median of five runs.
4.1
Imperative benchmarks.
These benchmarks are designed to exercise various different forms of mutation, as summarized in Figure 9. Due to
mutation, they are not implementable in Manticore.
msort, dedup. These benchmarks begin by tabulating a se-
quence of size 107 before sorting it with a technique similar
to that shown in Figure 1. The msort benchmark uses imperative quick-sort below the sequential threshold of 104
elements. The dedup benchmark is similar to msort but removes duplicate keys. Below the sequential threshold, this is
accomplished by imperatively inserting elements into a hash
set before sorting with the in-place quick-sort. For dedup, we
guarantee the sequence has approximately 106 unique keys.
Pure Benchmarks
These benchmarks are purely functional, meaning that their
source code does not use mutation.
fib. This benchmark computes the 42nd Fibonacci number
via the naïve recursive formula F (n) = F (n − 1) + F (n − 2),
with a sequential threshold of n = 25.
tourney. This benchmark tabulates a sequence of 108 contes-
tabulate, map, reduce, filter. These benchmarks each begin
by generating an input sequence of size 108 . The tabulate
tants, and then computes a tournament tree. Each contestant
is represented by an integer which measures their fitness. In
the tournament tree, each contestant c has an associated parent pointer which points to the contestant that eliminated c
from the tournament. This benchmark computes the tournament tree with a simple divide-and-conquer approach, using
mutation at each join point in order to set a parent pointer.
benchmark completes once the input sequence is built. The
map benchmark constructs a second sequence by applying
a simple function to each element. The reduce benchmark
sums the elements of the input sequence. The filter benchmark constructs a second sequence containing only the
elements which satisfy a given predicate. They are each
implemented with straightforward divide-and-conquer approaches, with a sequential threshold of 104 elements.
usp, usp-tree, multi-usp-tree, reachability. These bench-
marks consider variants of parallel breadth-first search (BFS)
on directed, unweighted graphs. BFS visits vertices in rounds.
At round r , BFS visits (in parallel) every vertex which has not
previously been visited and is reachable by r hops from the
source vertex. When a vertex is visited, a piece of mutable
data associated with it is updated.
The BFS variants differ in the types of per-vertex mutable
data. They also differ in the number of times a vertex is
visited. Except in the reachability benchmark, we guarantee
that each vertex is visited exactly once by marking vertices
as visited with an atomic “compare-and-swap” operation.
In the reachability benchmark, we check and update the
visited status of vertices simply by reading and writing to a
shared flag. This creates a data race and potentially causes
some vertices to be visited multiple times (up to at most
P visitations, P = number of processors). In practice, this
variant of BFS often performs better on modern hardware
because (a) atomic operations such as compare-and-swap are
expensive, and (b) observing the data race within a particular
execution is rare.
The specifics of these benchmarks are described below.
msort-pure. This benchmark first tabulates a sequence of
size 107 . It then sorts the sequence with a function similar
to the one shown in Figure 1, except that it uses a purely
functional quick-sort below a sequential threshold of 104
instead of the imperative one.
dmm, smvm. These benchmarks operate on square matrices
of size n × n. Each matrix is represented by a sequence of
rows (or columns). The dmm benchmark multiplies two dense
matrices with the naïve O(n3 ) algorithm, where each of the
n rows is implemented as another sequence of size n. The
smvm benchmark multiplies a sparse matrix by a dense vector,
where each row of the sparse matrix contains only the nonzero entries represented as index-value pairs. In dmm, n = 600.
In smvm, n = 20, 000 and each row has approximately 2, 000
non-zero entries. The sequential threshold is one matrix row.
strassen. This benchmark multiplies two dense square ma-
trices of size n × n using Strassen’s algorithm. The matrices
are represented by quadtrees with leaves of vectors of elements of size 64 × 64. In our experiments, n = 1024 with a
sequential threshold of 64 (that is, the leaves of the quadree
are each processed sequentially).
• reachability identifies which vertices are reachable
from the source.
• usp computes the unweighted single-source shortest
path length of all vertices. Every time a vertex is visited,
the algorithm records the current round number as
the distance to that vertex.
• usp-tree computes all unweighted single-source shortest paths. It is implemented with an array A of ancestor
raytracer. This benchmark is adapted from the raytracer
benchmark written for the Manticore language [7], which
was adapted from an Id program [30]. It renders a 600px ×
600px scene in parallel by tabulating a sequence of pixels
with a sequential granularity of 300 pixels.
8
Hierarchical Memory Management for Mutable State
Benchmark
pure benchmarks
msort
dedup
tourney
reachability
usp
usp-tree
multi-usp-tree
to parallel allocations. Memory inflation gives the memory
consumption as a factor relative to the sequential memory
consumption, Ms .
For mlton-spoonhower, GC 72 includes processor time spent
blocked during a stop-the-world collection. We do not report
GC statistics for Manticore, because it is not able to collect
statistics only for a specific region of code, which we need to
have a meaningful comparison. We are also unable to report
results for Manticore on msort-pure due to a compiler bug.
Our results are summarized in Figures 10, 11, 12, and
13. Figures 10 and 11 list the execution times, overheads,
speedups, and GC percentages of pure and imperative
benchmarks, respectively. Figure 12 shows the speedup of
mlton-parmem on various benchmarks for processor counts
between 1 and 72. Finally, Figure 13 lists the memory consumptions and inflations of all benchmarks.
Representative Operation
immutable reads
local non-pointer writes
local non-pointer writes
local non-promoting writes
distant non-pointer writes
distant non-pointer writes
distant promoting writes
distant promoting writes
Figure 9. Representative operations of all benchmarks.
lists. When a vertex v is visited along an edge (u, v),
the ancestors of v are recorded as A[v] := u :: A[u].
• multi-usp-tree runs 36 copies of usp-tree in parallel.
The input graph is the orkut social network graph [1], which
has approximately 3 million vertices, 117 million edges, and
a diameter of 9. Each benchmark begins by converting the
input graph into a compact adjacency-sequence format suitable for parallel BFS.
4.3
Overheads. Inspecting Figures 10 and 11, we can make several observations and conclusions. First, for both pure and
imperative benchmarks, the overheads of mlton-parmem are
generally comparable to those of mlton-spoonhower, which
serves as a good baseline because it does not support parallel memory management. This shows that our techniques
for maintaining a dynamic memory structure based on hierarchical heaps can be implemented efficiently. Second, we
observe that for pure benchmarks (Figure 10), our overheads
are within a factor 2 of the sequential baseline and are consistently smaller than those of manticore. Third, for imperative
benchmarks (Figure 11), our overheads are higher than those
of the pure benchmarks but still remain within a factor of
approximately 2.6 in comparison to the sequential baseline.
The increase in overheads is due in part to the memory operations which are no longer plain loads and stores (see Figures
8 and 9).
Representative Operations
In Figure 9, using the terminology from Figure 8, we characterize each benchmark by a representative memory operation.
Representative operations summarize which type of operation is most likely to be a dominant cost in execution time.1
These in turn help understand and predict performance. For
example, if a benchmark exhibits mostly “local, non-pointer”
writes, then that benchmark likely has low overhead; in
contrast, if a benchmark is characterized by many “distant,
promoting” writes, then it might have high overhead and
scale poorly.
4.4
Speedups. Inspecting Figure 10, we observe that for pure
benchmarks with 72 cores, mlton-parmem achieves speedups
ranging between 10 and 62. Compared to mlton-spoonhower,
our speedups are significantly higher, which is expected
because mlton-spoonhower suffers from sequential, stop-theworld garbage collections. Our speedups are also significantly better than those of manticore, which is sometimes
as low as 3x. This seems surprising because manticore is
designed for purely functional programs. The reason is that
manticore relies on imperative updates within the run-time
to execute computations in parallel (e.g., to communicate
the result of a remotely executed task to another processor) and employs a promotion technique to preserve local
heap invariants [7]. To verify this, we measured that on the
map benchmark with 72 cores, manticore promoted nearly
340MB of data in total, whereas mlton-parmem performed no
promotions.
Inspecting the imperative benchmarks (Figure 11), we
observe that mlton-parmem achieves good speedups on 72
processors (between 14 for high-overhead benchmarks and
Results
We collect the following statistics for each benchmark.
• Ts is the sequential execution time.
• T1 and T72 are execution times on 1 and 72 processors.
• The overhead is T1 /Ts .
• The speedup is Ts /T72 . In general, the speedup on P
processors is given by Ts /TP .
• GCs is the percent of time spent in GC during a sequential run.
• GC 72 is the percent of processor time spent in GC
during a 72-processor run.
• Ms is the memory consumption of the sequential run.
• I 1 and I 72 are memory inflations on 1 and 72 processors.
The memory consumption statistic is an upper bound on the
amount of physical memory required to store heap-allocated
objects; it is computed by tracking the maximum heap occupancy within one execution, and includes fragmentation due
1 Note
that immutable reads are pervasive in all benchmarks.
9
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
mlton
fib
tabulate
map
reduce
filter
msort-pure
dmm
smvm
strassen
raytracer
mlton-spoonhower
mlton-parmem (our compiler)
manticore
Ts
GCs
T1
Over
head
T72
Speed
up
GC 72
T1
Over
head
T72
Speed
up
T1
Over
head
T72
Speed
up
GC 72
2.67
1.11
1.46
1.13
3.62
7.02
3.76
7.23
2.54
7.41
0.0%
39.4%
29.8%
40.5%
11.9%
12.3%
15.2%
0.0%
1.6%
1.3%
3.72
0.89
1.31
0.9
4.84
8.52
7.02
9.93
2.89
7.0
1.39
0.8
0.9
0.8
1.34
1.21
1.87
1.37
1.14
0.94
0.12
0.42
0.48
0.46
0.54
1.79
0.92
0.2
0.16
0.3
22.25
2.64
3.04
2.46
6.7
3.92
4.09
36.15
15.88
24.7
3.4%
89.5%
82.2%
84.7%
76.6%
75.8%
78.3%
0.0%
40.6%
29.8%
5.19
1.92
4.02
1.93
6.15
–
8.3
12.68
4.36
6.97
1.94
1.73
2.75
1.71
1.7
–
2.21
1.75
1.72
0.94
0.12
0.12
0.49
0.13
0.49
–
0.19
0.32
0.12
0.17
22.25
9.25
2.98
8.69
7.39
–
19.79
22.59
21.17
43.59
3.63
1.62
2.75
1.43
5.75
6.91
5.83
8.69
2.94
6.52
1.36
1.46
1.88
1.27
1.59
0.98
1.55
1.2
1.16
0.88
0.07
0.07
0.14
0.09
0.18
0.36
0.16
0.18
0.12
0.12
38.14
15.86
10.43
12.56
20.11
19.5
23.5
40.17
21.17
61.75
0.0%
15.4%
16.2%
11.7%
12.0%
15.7%
6.8%
0.0%
7.9%
0.5%
Figure 10. Execution times (in seconds), overheads, and speedups of purely functional benchmarks.
mlton
msort
dedup
tourney
reachability
usp
usp-tree
multi-usp-tree
mlton-parmem (our compiler)
mlton-spoonhower
Ts
GCs
T1
Over
head
3.75
3.72
4.64
8.36
8.34
8.63
100.25
3.4%
2.6%
6.3%
0.0%
0.0%
0.0%
3.0%
4.66
4.05
8.17
21.59
23.38
23.79
209.59
1.24
1.09
1.76
2.58
2.8
2.76
2.09
T72
Speed
up
GC 72
T1
Over
head
T72
Speed
up
GC 72
0.36
0.32
0.92
0.52
0.61
0.63
19.07
10.42
11.63
5.04
16.08
13.67
13.7
5.26
58.3%
52.1%
76.2%
0.0%
0.0%
0.0%
34.0%
5.33
4.61
7.86
19.59
21.85
22.3
245.6
1.42
1.24
1.69
2.34
2.62
2.58
2.45
0.18
0.16
0.23
0.46
0.58
7.93
11.18
20.83
23.25
20.17
18.17
14.38
1.09
8.97
7.7%
6.3%
7.1%
0.0%
0.0%
0.0%
8.2%
Figure 11. Execution times (in seconds), overheads, and speedups of imperative benchmarks.
mlton
Figure 12. Speedups of mlton-parmem.
23 for those with low overhead) with a couple exceptions:
the highly concurrent usp-tree and multi-usp-tree benchmarks. Poor performance on usp-tree and multi-usp-tree is
expected, because these benchmarks exhibit close to pessimal
cases for our techniques with frequent concurrent updates
of shared pointer data. For example, every time a vertex is
visited in usp-tree, one cell of a distant array (located at the
root) is updated with a new list, triggering promotion from a
mltonspoonhower
mlton-parmem
(our compiler)
Ms
I1
I 72
I1
I 72
fib
tabulate
map
reduce
filter
msort-pure
dmm
smvm
strassen
raytracer
0.0
3.48
1.6
0.8
2.8
1.43
0.18
1.92
0.22
0.13
+0.0
0.23
1.01
1.01
0.8
1.01
1.78
1.33
0.95
1.15
+0.31
0.34
1.42
2.02
1.41
1.44
3.39
1.47
4.41
3.77
+0.0
0.97
3.07
3.15
1.7
1.49
0.94
2.67
1.86
1.46
+0.4
1.25
5.07
4.95
2.51
4.9
7.39
2.78
17.68
5.38
msort
dedup
tourney
reachability
usp
usp-tree
multi-usp-tree
1.09
0.56
0.8
3.87
3.87
3.97
19.76
0.78
1.18
6.24
1.01
1.01
1.01
1.2
1.5
2.61
8.34
2.29
2.35
2.25
2.85
1.6
1.63
6.91
1.5
1.5
1.55
1.64
4.04
6.48
8.66
2.15
2.19
2.21
1.79
Figure 13. Memory consumption (in GB) and inflations.
leaf heap to the root heap. Since promotions require locking
entire heaps, they can sequentialize otherwise parallel visitations, leading effectively to complete serialization of the
10
Hierarchical Memory Management for Mutable State
entire computation. However, when multiple usp-tree computations are performed in parallel in the multi-usp-tree
benchmark, some promotions remain independent and execute in parallel because the updated array is not always at the
root heap. We indeed see a 9-fold speedup in this benchmark.
In Figure 12, we observe that as the number of processors increases, the speedup of all benchmarks continues to
increase. That is, there are no inversions. For multiple benchmarks the speedup improves nearly linearly, suggesting the
possibility of further scalability to higher core counts.
second, each such collection is sequential. Parallelism is thus
achieved by collecting many leaves independently. Our results suggest that this simple approach can perform well for
highly parallel applications. However, if a collection takes
place at the root heap (when there is no parallelism), such
a collection would be sequential and effectively stop-theworld. More generally, when there is little parallelism, large
collections can take place sequentially. In future work, we
plan to complete the implementation by adding support for
parallel collection of individual heaps, and in general parallel
collection of sub-trees of heaps (not just leaves).
Garbage Collection. Inspecting Figures 10 and 11, we observe that mlton-parmem only loses at most approximately
16% of its time to garbage collection on runs with 72 cores.
As expected, mlton-spoonhower performs poorly due to its
sequential GC. Some benchmarks (smvm, reachability, usp,
usp-tree) spend no time in garbage collection, regardless
of when run sequentially or in parallel. This is due to the
fact that these benchmarks allocate memory in an already
large heap, which was grown to accommodate the input.
(In the case of smvm, the benchmark does not include input
generation; for the graph algorithms, the benchmark does
not include the time it takes to read the graph from disk into
a large string in the heap.)
6
With the proliferation of shared memory parallel computers using modern multicore processors, there has been
significant work on the design and implementation of
high-level programming languages for writing parallel programs [9, 12, 16, 17, 21, 23, 25, 26, 32]. For implicit memory
allocation and reclamation with garbage collection, there are
numerous techniques for incorporating parallelism, concurrency, and real-time features. Jones et al. [22] provides an
excellent survey.
We contrast our work with a number of systems [4, 7, 13–
15, 27, 31, 32] that use processor- or thread-local heaps which
service (most) allocations of the executing computation and
can be collected independently combined with a shared
global heap that must be collected cooperatively.
The Doligez-Leroy-Gonthier (DLG) parallel collector [13,
14] employs this design, with the invariant that there are
no pointers from the shared global heap into any processorlocal heap and no pointers from one processor local-heap
into another processor-local heap. To maintain this invariant, all mutable objects are allocated in the shared global
heap and (transitively reachable) data is promoted (copied)
from a processor-local heap to the shared global heap when
updating a mutable object, which increases the cost of all
mutable allocations and updates. In our approach, mutable
objects are allocated in the thread-local heap and updates to
such never need to promote data, making this common case
significantly less expensive than in DLG. Moreover, in DLG,
scheduling and communication actions, such as migrating
a language-level thread from one processor to another or
returning the result from a child task to a parent task, typically employ mutable objects and require promotions. With
hierarchical heaps, the local heap is associated with the task,
rather than the processor, and returning the result of a child
task to the parent task is accomplished without copying.
Anderson [4] describes TCG, a variant of the DLG collector for a language with implicit parallelism serviced by a
fixed number of worker threads pinned to processors. TCG
allows mutable objects to be allocated in the processor-local
heap and a processor-local collection copies live data to
the global shared heap. When updating a mutable object
Memory Consumption. Inspecting Figure 13, with only a
few exceptions, our compiler on 72 cores consumes at most
7x more memory than the sequential baseline. Note that
in general, any P-processor execution scheduled via workstealing can expect to see inflation of up to a factor P [10,
11]. Thus, using mlton-spoonhower as an alternative baseline
helps determine how much inflation is due simply to parallel
execution, versus how much is due to our techniques. Indeed,
our inflation with respect to mlton-spoonhower is generally
lower, staying consistently within a factor of approximately
4. Our implementation introduces additional inflation in part
through (a) the need for a separate forwarding pointer on
every object, and (b) greater fragmentation of allocation to
distinguish heaps within the hierarchy.
5
Related Work
Discussion and Future Work
The promotion techniques presented in this paper rely on
coarse-grained locks to manage concurrent manipulations
of overlapping data. This approach prevents certain promotions from proceeding in parallel, even when those promotions would otherwise be independent. For example, in the
usp-tree benchmark, every visitation of a vertex triggers a
promotion to the root of hierarchy, causing a serialization of
visitations. However none of these promotions overlap, so
they ought to be able to proceed in parallel. In future work,
we intend to design a more fine-grained promotion strategy
that would permit parallel promotions to the same heap.
With respect to garbage collection, our current implementation has two limitations: first, it can only collect leaf heaps;
11
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
in the shared global heap with a processor-local pointer, a
processor-local garbage collection is triggered, which copies
the to-be-written object (and all other processor-local live
data) to the shared global heap; updating a mutable object
in the processor-local heap proceeds without a collection.
Using collection to over-approximate promotion would not
work well with our hiearchical heaps, because the collection
would need to be triggered for all descendent heaps of the
mutable object being written and could not be performed by
the writing processor independently.
The Manticore garbage collector [7] is another variant of
the DLG design, where the Appel semi-generational collector [5] is used for collection of the processor-local heaps.
Although the high-level language is mutation-free, the implementation uses mutation to realize various parallel constructs and employs promotion to preserve the heap invariants. Recent work [24] has considered extending the Manticore language with mutable state via software transactional
memory, but notes that promotions make a chronologicallyordered read set implemented as a mutable doubly-linked list
inefficient. In contrast, such a data structure would be efficient in our hierarchical heaps, since a transaction’s read set
is necessarily local to the thread executing the transaction.
The current Glasgow Haskell Compiler garbage collector [27] combines elements of the DLG and Domani et al. [15]
collectors. Although Haskell is a pure language, there is significant mutation due to lazy evaluation. The collector allows
mutable objects to be allocated in a dedicated portion of the
processor-local heaps, which use a non-moving collector.
The collector also allows pointers from the global heap to
the processor-local heaps, mediated by proxy objects. When
another processor accesses a proxy, it communicates with the
owning processor to request that the object be promoted to
the global shared heap. Proxy objects fit well with Haskell’s
lazy evalution, where all pointer accesses have a read barrier
to check for unevaluated computations and where proxy
objects can be incorporated into unevaluated computations
without requiring promotion. Standard ML employs strict
evaluation and eager promotion is a better fit.
The MultiMLton project [32] forked from
mlton-spoonhower and shifted the domain to messagepassing concurrency. While one could encode shared-state
fork-join parallelism with their message-passing operations,
the resulting overheads (emulating references with threads;
eager thread creation vs lazy work-stealing) would not lead
to a meaningful performance comparison. Their garbage
collection strategy is tuned to their message-passing
concurrency bias — starting from the DLG design and
invariants, they avoid promotion through procrastination,
blocking the writing thread (and executing another of the
abundant concurrent threads) until a GC can be performed,
which promotes the object and fixes references. They are
concerned that promotion, leaving forwarding pointers that
overwrite object data, would have unacceptable read-barrier
overhead, which also motivates their dynamic cleanliness
analysis; in contrast, we have introduced a dedicated
forwarding-pointer metadata field, which only requires a
read barrier for mutable data.
Raghunathan et al. introduced hierachical heaps [31] to
mirror the hierarchy of tasks in a strict pure functional language with nested parallelism. They prove that the language
enforces disentanglement, formulate a hierarchical garbage
collection technique that allows independent heaps to be collected concurrently, and report the performance of an implementation in MLton. We extend this work to accomodate the
mutable references and arrays of a strict impure functional
language with nested parallelism. The hierarchical-heaps design is partly motivated by a desire to take advantage of the
natural data locality of computation [2], which, with some
care, could be preserved by thread schedulers (e.g., [3, 8]).
7
Conclusion
The high-level nature of functional programming languages
makes them a good fit for parallel programming, but they require sophisticated memory managers which are challenging
to get right in the joint presence of parallelism and mutation.
In this paper, we showed how to provide efficient support for
uses of mutation common in parallel programs by exploiting
the hierarchical structure of functional computations. Our
experiments suggest that these results could be an important step towards making functional programming a serious
contender for performant parallel computing.
Acknowledgments
We thank Rohan Yadav for his assistance in the implementation of the benchmarks used in this paper. This material is
based upon work supported by the National Science Foundation under Grant No. 1408940, Grant No. 1408981, and
Grant No. 1629444. Any opinions, findings, and conclusions
or recommendations expressed in this material are those of
the author and do not necessarily reflect the views of the National Science Foundation. The first author was also partially
supported by the German Research Council (DFG) under
Grant No. ME14271/6-2.
12
Hierarchical Memory Management for Mutable State
References
[19] Marcelo J. R. Gonçalves and Andrew W. Appel. 1995. Cache Performance of Fast-Allocating Programs. In Conference on Functional
Programming and Computer Architecture. ACM Press, La Jolla, CA,
293–305. https://doi.org/10.1145/224164.224219
[20] Maurice Herlihy and Nir Shavit. 2011. The Art of Multiprocessor Programming. Morgan Kaufmann.
[21] Shams Mahmood Imam and Vivek Sarkar. 2014. Habanero-Java library:
a Java 8 framework for multicore programming. In 2014 International
Conference on Principles and Practices of Programming on the Java
Platform Virtual Machines, Languages and Tools, PPPJ ’14. 75–86.
[22] Richard Jones, Antony Hosking, and Eliot Moss. 2012. The Garbage
Collection Handbook: The Art of Automatic Memory Management. Chapman & Hall.
[23] Gabriele Keller, Manuel M.T. Chakravarty, Roman Leshchinskiy, Simon
Peyton Jones, and Ben Lippmeier. 2010. Regular, shape-polymorphic,
parallel arrays in Haskell. In Proceedings of the 15th ACM SIGPLAN
international conference on Functional programming (ICFP ’10). 261–
272.
[24] Matthew Le and Matthew Fluet. 2015. Partial Aborts for Transactions
via First-class Continuations. In Proceedings of the 20th ACM SIGPLAN
International Conference on Functional Programming (ICFP 2015). 230–
242.
[25] Doug Lea. 2000. A Java fork/join framework. In Proceedings of the
ACM 2000 conference on Java Grande (JAVA ’00). 36–43.
[26] Daan Leijen, Wolfram Schulte, and Sebastian Burckhardt. 2009. The
design of a task parallel library. In Proceedings of the 24th ACM SIGPLAN conference on Object Oriented Programming Systems Languages
and Applications (OOPSLA ’09). 227–242.
[27] Simon Marlow and Simon L. Peyton Jones. 2011. Multicore Garbage
Collection with Local Heaps. In 10th International Symposium on Memory Management, Hans Boehm and David Bacon (Eds.). ACM Press,
San Jose, CA, 21–32.
[28] MLton [n. d.]. MLton web site. http://www.mlton.org. ([n. d.]).
[29] Ryan Newton. 2017. (2017). Indiana University. Personal Communication.
[30] Rishiyur S. Nikhil. 1991. Id Language Reference Manual. (1991).
[31] Ram Raghunathan, Stefan K. Muller, Umut A. Acar, and Guy Blelloch.
2016. Hierarchical Memory Management for Parallel Programs. In
ICFP 2016. ACM Press.
[32] K. C. Sivaramakrishnan, Lukasz Ziarek, and Suresh Jagannathan. 2014.
MultiMLton: A multicore-aware runtime for standard ML. Journal of
Functional Programming FirstView (6 2014), 1–62.
[33] Daniel Spoonhower, Guy E. Blelloch, Robert Harper, and Phillip B.
Gibbons. 2010. Space profiling for parallel functional programs. Journal
of Functional Programming 20 (2010), 417–461. Issue Special Issue 5-6.
[34] Philip Wadler. 1998. Why No One Uses Functional Languages. SIGPLAN Notices 33, 8 (1998), 23–27. https://doi.org/10.1145/286385.
286387
[1] [n. d.]. Stanford Large Network Dataset Collection. http://snap.
stanford.edu/. ([n. d.]).
[2] Umut A. Acar, Guy Blelloch, Matthew Fluet, and Stefan K.
Mullerand Ram Raghunathan. 2015. Coupling Memory and Computation for Locality Management. In Summit on Advances in Programming
Languages (SNAPL).
[3] Umut A. Acar, Guy E. Blelloch, and Robert D. Blumofe. 2002. The Data
Locality of Work Stealing. Theory of Computing Systems 35, 3 (2002),
321–347.
[4] Todd A. Anderson. 2010. Optimizations in a Private Nursery-Based
Garbage Collector. In 9th International Symposium on Memory Management, Jan Vitek and Doug Lea (Eds.). ACM Press, Toronto, Canada,
21–30.
[5] Andrew W. Appel. 1989. Simple Generational Garbage Collection and
Fast Allocation. Software: Practice and Experience 19, 2 (1989), 171–183.
[6] Andrew W. Appel and Zhong Shao. 1996. Empirical and analytic
study of stack versus heap cost for languages with closures. Journal
of Functional Programming 6, 1 (Jan. 1996), 47–74. https://doi.org/10.
1017/S095679680000157X
[7] Sven Auhagen, Lars Bergstrom, Matthew Fluet, and John H. Reppy.
2011. Garbage collection for multicore NUMA machines. In Proceedings
of the 2011 ACM SIGPLAN workshop on Memory Systems Performance
and Correctness (MSPC). 51–57.
[8] Guy E. Blelloch, Jeremy T. Fineman, Phillip B. Gibbons, and Harsha Vardhan Simhadri. 2011. Scheduling irregular parallel computations on hierarchical caches. In Proc. ACM Symposium on Parallelism
in Algorithms and Architectures (SPAA). 355–366.
[9] Guy E. Blelloch, Jonathan C. Hardwick, Jay Sipelstein, Marco Zagha,
and Siddhartha Chatterjee. 1994. Implementation of a Portable Nested
Data-Parallel Language. J. Parallel Distrib. Comput. 21, 1 (1994), 4–14.
[10] Robert D. Blumofe and Charles E. Leiserson. 1998. Space-Efficient
Scheduling of Multithreaded Computations. SIAM J. Comput. 27, 1
(1998), 202–229.
[11] Robert D. Blumofe and Charles E. Leiserson. 1999. Scheduling multithreaded computations by work stealing. J. ACM 46 (Sept. 1999),
720–748. Issue 5.
[12] Philippe Charles, Christian Grothoff, Vijay Saraswat, Christopher Donawa, Allan Kielstra, Kemal Ebcioglu, Christoph von Praun, and Vivek
Sarkar. 2005. X10: an object-oriented approach to non-uniform cluster
computing. In Proceedings of the 20th annual ACM SIGPLAN conference
on Object-oriented programming, systems, languages, and applications
(OOPSLA ’05). ACM, 519–538.
[13] Damien Doligez and Georges Gonthier. 1994. Portable, Unobtrusive
Garbage Collection for Multiprocessor Systems. In 21st Annual ACM
Symposium on Principles of Programming Languages. ACM Press, Portland, OR, 70–83. https://doi.org/10.1145/174675.174673
[14] Damien Doligez and Xavier Leroy. 1993. A Concurrent Generational
Garbage Collector for a Multi-Threaded Implementation of ML. In
20th Annual ACM Symposium on Principles of Programming Languages.
ACM Press, Charleston, SC, 113–123.
[15] Tamar Domani, Elliot K. Kolodner, Ethan Lewis, Erez Petrank, and
Dafna Sheinwald. 2002. Thread-Local Heaps for Java. In 3rd International Symposium on Memory Management (ACM SIGPLAN Notices
38(2 supplement)), Hans-J. Boehm and David Detlefs (Eds.). ACM Press,
Berlin, Germany, 76–87. https://doi.org/10.1145/512429.512439
[16] Matthew Fluet, Mike Rainey, John Reppy, and Adam Shaw. 2011. Implicitly threaded parallelism in Manticore. Journal of Functional Programming 20, 5-6 (2011), 1–40.
[17] golang [n. d.]. The Go Programming Language. https://golang.org. ([n.
d.]). Accessed: 2017.
[18] Marcelo J. R. Gonçalves. 1995. Cache Performance of Programs with
Intensive Heap Allocation and Generational Garbage Collection. Ph.D.
Dissertation. Department of Computer Science, Princeton University.
13
A. Guatto, S. Westrick, R. Raghunathan, U. Acar, and M. Fluet
1
2
3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
function toSpaceOf: heap -> heap
function isToSpace: heap -> bool
function switchSemispaces: heap -> unit
in a from-heap strictly above topHeap. The latter case corresponds to the elimination of copies introduced during promotion: copy collection replaces a pointer to an old copy
with a pointer to the a more recent one lying outside of the
collection zone. In addition, since we do not follow forwarding pointers that belong to objects outside of the collection
zone, we do not have to lock heaps during collection.
Like promote, we specify cheneyCopy as a recursive function. Let us call heap the heap where obj resides. If heap is
strictly above topHeap (l. 9-10), or is a to-space (l. 11-12), obj
can be returned. If obj has a forwarding pointer, cheneyCopy
follows it (l. 13-14). Otherwise, as in promote, we create a
new copy, setting up the forwarding pointer of obj to point
to it, recursively call cheneyCopy on its pointer fields, and
return it (l. 17-23).
function collect (topHeap) =
for r in current roots:
*r ← cheneyCopy(heap, *r)
for h below topHeap included:
switchSemispaces(h)
function cheneyCopy (topHeap, obj) =
heap ← heapOf(obj)
i f depth(heap) < depth(topHeap): return obj
i f isToSpace(heap): return obj
i f hasFwdPtr(obj):
return cheneyCopy(topHeap, *fwdPtr(obj))
newObj ← freshObj(toSpaceOf(heap), sizeOf(obj))
*fwdPtr(obj) = newObj
for field in nonptrFields(obj):
*getField(newObj, field) ←
*getField(obj, field)
for field in ptrFields(obj):
*getField(newObj, field) ←
cheneyCopy(topHeap, *getField(obj, field))
return newObj
B
Scheduler. Any implementation of a fork/join programming
model requires a scheduler to coordinate work between
worker threads. Tasks are evaluated within a “user-level
thread” that is scheduled onto a “worker thread”. In addition
to the forkjoin function exposed to the mutator, the scheduler also exposes a schedule function to idle worker threads
to find waiting work.
A naive implementation of a work-stealing scheduler will
create tasks for both thunks passed to it before evaluating
them. However, task creation is expensive and its value is
only realized upon a steal. Steals are far less frequent than
calls to forkjoin, so our implementation ensures that calling
forkjoin is cheap and expensive task creation is deferred
to the steal. In addition, one of the thunks is evaluated immediately in the calling user-level thread, while only the
other thunk is exposed to other worker threads. This reduces the number of user-level threads and thread switches
by allowing a user-level thread to evaluate a path of tasks.
In our implementation, we use a work-stealing scheduler
that has been annotated with heap management operations
at the appropriate places. Worker threads are implemented
as OS-level pthreads and user-level threads are implemented
as the native user-level thread in MLton. As the scheduler
operates outside of the computation, the objects it allocates,
particularly in the schedule function, do not belong to any
heap in the hierarchy. Our implementation has a separate
“global heap” that is used to store scheduler data.
Figure 14. Promotion-aware copy collection.
A
A.1
Implementation
Algorithms
Promotion-Aware Copy Collection
Promotion introduces redundant copies of memory objects.
We now present a way to eliminate these copies by piggybacking on the classic semispace garbage collection algorithm. Our proposal is given in Figure 14, including additional primitives specific to semispace collection.
We now assume that every hierarchical heap accessed
by the mutator is paired with another heap used only during collection. Following standard terminology, we call the
former a from-space and the latter a to-space. Our collection algorithm manipulates semispaces using three primitives: toSpaceOf returns the to-space associated with a given
from-space; isToSpace(heap) returns true iff heap is a tospace; switchSemispaces swaps to-space and from-space.
The function collect(topHeap) is called to collect the subtree starting at topHeap. We assume that every task associated
with a leaf heap below topHeap has been suspended by the
runtime system. Thanks to disentanglement, this is sufficient
to collect the entire subtree independently from other mutators and collectors. It copies every object reachable from
a root to the to-spaces using cheneyCopy (l. 2-3) and then
swaps the semispaces of every heap in the subtree (l. 4-5).
The function cheneyCopy takes a from-heap topHeap and
an object pointer obj, and returns a copy of obj. This copy
is guaranteed to be either in a to-heap below topHeap, or
Superheaps. The goal of making calls to forkjoin cheap
and deferring task creation to steals directly informs our
design and implementation of hierarchical heaps. Our implementation uses a structure called a “superheap”, which
is associated with a user-level thread. Superheaps contain
a linked-list of heaps, each annotated with its depth. This
set of heaps corresponds to the heaps of the path of tasks
evaluated by the superheap’s associated user-level thread.
14
Hierarchical Memory Management for Mutable State
On a call to forkjoin, the current superheap’s depth is
incremented and future allocations take place in the new
heap created for that depth. Once both thunks passed to
forkjoin are completed, the depth is decremented. On depth
decrement, the new heap will be joined to its parent heap
as per the algorithm. As the set of heaps in the superheap is
maintained as a linked list, and heaps are added and joined in
LIFO order, the depth increment and decrement operations
are very cheap.
If no steal occurs, both thunks passed to forkjoin will be
evaluated in the newly created heap. If an idle worker thread
steals one of the thunks, it creates a new user-level thread
and associated superheap for that thunk. This superheap is
then attached as a child to the parent superheap in order for
the runtime to be aware of the complete hierarchy of heaps.
When the stolen thunk is complete, it will reactivate its
parent task which can then merge the superheap and extract
the return value. Merging a superheap is just merging its set
of heaps, which is a simple linked-list operation.
Heaps. We implement a heap as a linked-list of variablesized memory regions called “chunks”. This formulation enables efficient implementations of key heap operations. Increasing heap size and joining heaps (joinHeap) are constanttime linked-list operations that do not require objects to be
copied. Finding the heap of an arbitrary pointer (heapOf) is
implemented by looking up the chunk metadata using address masking, which then contains a pointer to the heap
associated with that chunk.
15
| 6cs.PL
|
arXiv:1706.00409v2 [cs.CV] 28 Jan 2018
Fader Networks:
Manipulating Images by Sliding Attributes
Guillaume Lample1,2 , Neil Zeghidour1,3 , Nicolas Usunier1 ,
Antoine Bordes1 , Ludovic Denoyer2 , Marc’Aurelio Ranzato1
{gl,neilz,usunier,abordes,ranzato}@fb.com
[email protected]
Abstract
This paper introduces a new encoder-decoder architecture that is trained to reconstruct images by disentangling the salient information of the image and the
values of attributes directly in the latent space. As a result, after training, our
model can generate different realistic versions of an input image by varying the
attribute values. By using continuous attribute values, we can choose how much a
specific attribute is perceivable in the generated image. This property could allow
for applications where users can modify an image using sliding knobs, like faders
on a mixing console, to change the facial expression of a portrait, or to update
the color of some objects. Compared to the state-of-the-art which mostly relies
on training adversarial networks in pixel space by altering attribute values at train
time, our approach results in much simpler training schemes and nicely scales to
multiple attributes. We present evidence that our model can significantly change
the perceived value of the attributes while preserving the naturalness of images.
1
Introduction
We are interested in the problem of manipulating natural images by controlling some attributes of
interest. For example, given a photograph of the face of a person described by their gender, age, and
expression, we want to generate a realistic version of this same person looking older or happier, or an
image of a hypothetical twin of the opposite gender. This task and the related problem of unsupervised
domain transfer recently received a lot of interest [17, 24, 9, 26, 21, 23], as a case study for conditional
generative models but also for applications like automatic image edition. The key challenge is that
the transformations are ill-defined and training is unsupervised: the training set contains images
annotated with the attributes of interest, but there is no example of the transformation: In many cases
such as the “gender swapping” example above, there are no pairs of images representing the same
person as a male or as a female. In other cases, collecting examples requires a costly annotation
process, like taking pictures of the same person with and without glasses.
Our approach relies on an encoder-decoder architecture where, given an input image x with its
attributes y, the encoder maps x to a latent representation z, and the decoder is trained to reconstruct
x given (z, y). At inference time, a test image is encoded in the latent space, and the user chooses
the attribute values y that are fed to the decoder. Even with binary attribute values at train time,
each attribute can be considered as a continuous variable during inference to control how much it is
perceived in the final image. We call our architecture Fader Networks, in analogy to the sliders of an
audio mixing console, since the user can choose how much of each attribute they want to incorporate.
1
Facebook AI Research
Sorbonne Universités, UPMC Univ Paris 06, UMR 7606, LIP6
3
LSCP, ENS, EHESS, CNRS, PSL Research University, INRIA
Code available at https://github.com/facebookresearch/FaderNetworks
2
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Figure 1: Interpolation between different attributes (Zoom in for better resolution). Each line shows
reconstructions of the same face with different attribute values, where each attribute is controlled as a
continuous variable. It is then possible to make an old person look older or younger, a man look more
manly or to imagine his female version. Left images are the originals.
The fundamental feature of our approach is to constrain the latent space to be invariant to the attributes
of interest. Concretely, it means that the distribution over images of the latent representations should
be identical for all possible attribute values. This invariance is obtained by using a procedure similar
to domain-adversarial training (see e.g., [20, 6, 14]). In this process, a classifier learns to predict the
attributes y given the latent representation z during training while the encoder-decoder is trained
based on two objectives at the same time. The first objective is the reconstruction error of the decoder,
i.e., the latent representation z must contain enough information to allow for the reconstruction of the
input. The second objective consists in fooling the attribute classifier, i.e., the latent representation
must prevent it from predicting the correct attribute values. In this model, achieving invariance is a
means to filter out, or hide, the properties of the image that are related to the attributes of interest.
A single latent representation thus corresponds to different images that share a common structure
but with different attribute values. The reconstruction objective then forces the decoder to use the
attribute values to choose, from the latent representation, the intended image.
Our motivation is to learn a disentangled latent space in which we have explicit control on some
attributes of interest, without supervision of the intended result of modifying attribute values. With
a similar motivation, several approaches have been tested on the same tasks [17, 24], on related
image-to-image translation problems [9, 26], or for more specific applications like the creation of
parametrized avatars [23]. In addition to a reconstruction loss, the vast majority of these works rely
on adversarial training in pixel space, which compares during training images generated with an
intentional change of attributes from genuine images for the target attribute values. Our approach is
different both because we use adversarial training for the latent space instead of the output, but also
because adversarial training aims at learning invariance to attributes. The assumption underlying our
work is that a high fidelity to the input image is less conflicting with the invariance criterion, than
with a criterion that forces the hallucinated image to match images from the training set.
As a consequence of this principle, our approach results in much simpler training pipelines than those
based on adversarial training in pixel space, and is readily amenable to controlling multiple attributes,
by adding new output variables to the discriminator of the latent space. As shown in Figure 1 on test
images from the CelebA dataset [13], our model can make subtle changes to portraits that end up
sufficient to alter the perceived value of attributes while preserving the natural aspect of the image
and the identity of the person. Our experiments show that our model outperforms previous methods
based on adversarial training on the decoders’ output like [17] in terms of both reconstruction loss
and generation quality as measured by human subjects. We believe this disentanglement approach is
a serious competitor to the widespread adversarial losses on the decoder output for such tasks.
In the remainder of the paper, we discuss in more details the related work in Section 2. We then
present the training procedure in Section 3 before describing the network architecture and the
implementation in Section 4. Experimental results are shown in Section 5.
2
2
Related work
There is substantial literature on attribute-based and/or conditional image generation that can be split
in terms of required supervision, with three different levels. At one extreme are fully supervised
approaches developed to model known transformations, where examples take the form of (input,
transformation, result of the transformation). In that case, the model needs to learn the desired
transformation. This setting was previously explored to learn affine transformations [8], 3D rotations
[25], lighting variations [11] and 2D video game animations [19]. The methods developed in these
works however rely on the supervised setting, and thus cannot be applied in our setup.
At the other extreme of the supervision spectrum lie fully unsupervised methods that aim at learning
deep neural networks that disentangle the factors of variations in the data, without specification of
the attributes. Example methods are InfoGAN [4], or the predictability minimization framework
proposed in [20]. The neural photo editor [3] disentangles factors of variations in natural images for
image edition. This setting is considerably harder than the one we consider, and it may be difficult
with these methods to automatically discover high-level concepts such as gender or age.
Our work lies in between the two previous settings. It is related to information as in [15]. Methods
developed for unsupervised domain transfer [9, 26, 21, 23] can also be applied in our case: given
two different domains of images such as “drawings” and “photograph”, one wants to map an image
from one domain to the other without supervision; in our case, a domain would correspond to an
attribute value. The mappings are trained using adversarial training in pixel space as mentioned in
the introduction, using separate encoders and/or decoders per domain, and thus do not scale well to
multiple attributes. In this line of work but more specifically considering the problem of modifying
attributes, the Invertible conditional GAN [17] first trains a GAN conditioned on the attribute values,
and in a second step learns to map input images to the latent space of the GAN, hence the name of
invertible GANs. It is used as a baseline in our experiments. Antipov et al. [1] use a pre-trained face
recognition system instead of a conditional GAN to learn the latent space, and only focuses on the
age attribute. The attribute-to-image approach [24] is a variational auto-encoder that disentangles
foreground and background to generate images using attribute values only. Conditional generation is
performed by inferring the latent state given the correct attributes and then changing the attributes.
Additionally, our work is related to work on learning invariant latent spaces using adversarial training
in domain adaptation [6], fair classification [5] and robust inference [14]. The training criterion
we use for enforcing invariance is similar to the one used in those works, the difference is that the
end-goal of these works is only to filter out nuisance variables or sensitive information. In our case,
we learn generative models, and invariance is used as a means to force the decoder to use attribute
information in its reconstruction.
Finally, for the application of automatically modifying faces using attributes, the feature interpolation
approach of [22] presents a means to generate alterations of images based on attributes using a
pre-trained network on ImageNet. While their approach is interesting from an application perspective,
their inference is costly and since it relies on pre-trained models, cannot naturally incorporate factors
or attributes that have not been foreseen during the pre-training.
3
Fader Networks
Let X be an image domain and Y the set of possible attributes associated with images in X , where
in the case of people’s faces typical attributes are glasses/no glasses, man/woman, young/old. For
simplicity, we consider here the case where attributes are binary, but our approach could be extended
to categorical attributes. In that setting, Y = {0, 1}n , where n is the number of attributes. We have a
training set D = {(x1 , y 1 ), ..., (xm , y m )}, of m pairs (image, attribute) (xi ∈ X , y i ∈ Y). The end
goal is to learn from D a model that will generate, for any attribute vector y 0 , a version of an input
image x whose attribute values correspond to y 0 .
Encoder-decoder architecture Our model, described in Figure 2, is based on an encoder-decoder
architecture with domain-adversarial training on the latent space. The encoder Eθenc : X → RN is a
convolutional neural network with parameters θenc that maps an input image to its N -dimensional
latent representation Eθenc (x). The decoder Dθdec : (RN , Y) → X is a deconvolutional network with
parameters θdec that produces a new version of the input image given its latent representation Eθenc (x)
3
and any attribute vector y 0 . When the context is clear, we simply use D and E to denote Dθdec and
Eθenc . The precise architectures of the neural networks are described in Section 4. The auto-encoding
loss associated to this architecture is a classical mean squared error (MSE) that measures the quality
of the reconstruction of a training input x given its true attribute vector y:
1 X
2
LAE (θenc , θdec ) =
Dθdec Eθenc (x), y − x 2
m
(x,y)∈D
The exact choice of the reconstruction loss is not fundamental in our approach, and adversarial losses
such as PatchGAN [12] could be used in addition to the MSE at this stage to obtain better textures or
sharper images, as in [9]. Using a mean absolute or mean squared error is still necessary to ensure
that the reconstruction matches the original image.
Ideally, modifying y in D(E(x), y) would generate images with different perceived attributes, but
similar to x in every other aspect. However, without additional constraints, the decoder learns to
ignore the attributes, and modifying y at test time has no effect.
Learning attribute-invariant latent representations To avoid this behavior, our approach is to
learn latent representations that are invariant with respect to the attributes. By invariance, we mean
that given two versions of a same object x and x0 that are the same up to their attribute values, for
instance two images of the same person with and without glasses, the two latent representations
E(x) and E(x0 ) should be the same. When such an invariance is satisfied, the decoder must use the
attribute to reconstruct the original image. Since the training set does not contain different versions
of the same image, this constraint cannot be trivially added in the loss.
We hence propose to incorporate this constraint by doing adversarial training on the latent space.
This idea is inspired by the work on predictability minimization [20] and adversarial training for
domain adaptation [6, 14] where the objective is also to learn an invariant latent representation using
an adversarial formulation of the learning objective. To that end, an additional neural network called
the discriminator is trained to identify the true attributes y of a training pair (x, y) given E(x). The
invariance is obtained by learning the encoder E such that the discriminator is unable to identify the
right attributes. As in GANs [7], this corresponds to a two-player game where the discriminator aims
at maximizing its ability to identify attributes, and E aims at preventing it to be a good discriminator.
The exact structure of our discriminator is described in Section 4.
Discriminator objective The discriminator outputs probabilities of an attribute vector
Pθdis (y|E(x)), where θdis are the discriminator’s parameters. Using the subscript k to refer to
n
P
the k-th attribute, we have log Pθdis (y|E(x)) =
log Pθdis ,k (yk |E(x)). Since the objective of the
k=1
discriminator is to predict the attributes of the input image given its latent representation, its loss
depends on the current state of the encoder and is written as:
Ldis (θdis |θenc ) = −
1
m
X
log Pθdis y Eθenc (x)
(1)
(x,y)∈D
Adversarial objective The objective of the encoder is now to compute a latent representation that
optimizes two objectives. First, the decoder should be able to reconstruct x given E(x) and y, and
at the same time the discriminator should not be able to predict y given E(x). We consider that a
mistake is made when the discriminator predicts 1 − yk for attribute k. Given the discriminator’s
parameters, the complete loss of the encoder-decoder architecture is then:
L(θenc , θdec |θdis ) =
1
m
X
Dθdec Eθenc (x), y − x
2
2
− λE log Pθdis (1 − y|Eθenc (x)) , (2)
(x,y)∈D
where λE > 0 controls the trade-off between the quality of the reconstruction and the invariance
of the latent representations. Large values of λE will restrain the amount of information about x
contained in E(x), and result in blurry images, while low values limit the decoder’s dependency on
the latent code y and will result in poor effects when altering attributes.
4
Figure 2: Main architecture. An (image, attribute) pair (x, y) is given as input. The encoder maps x
to the latent representation z; the discriminator is trained to predict y given z whereas the encoder
is trained to make it impossible for the discriminator to predict y given z only. The decoder should
reconstruct x given (z, y). At test time, the discriminator is discarded and the model can generate
different versions of x when fed with different attribute values.
Learning algorithm Overall, given the current state of the encoder, the optimal discriminator
∗
parameters satisfy θdis
(θenc ) ∈ argminθdis Ldis (θdis |θenc ). If we ignore problems related to multiple
(and local) minima, the overall objective function is
∗
∗
∗
θenc
, θdec
= argmin L(θenc , θdec |θdis
(θenc )) .
θenc ,θdec
∗
In practice, it is unreasonable to solve for θdis
(θenc ) at each update of θenc . Following the practice of
adversarial training for deep networks, we use stochastic gradient updates for all parameters, consid∗
ering the current value of θdis as an approximation for θdis
(θenc ). Given a training example (x, y), let
us denote Ldis θdis θenc , x, y the auto-encoder loss restricted to (x, y) and L θenc , θdec θdis , x, y
(t)
(t)
the corresponding discriminator loss. The update at time t given the current parameters θdis , θenc ,
(t)
and θdec and the training example (x(t) , y (t) ) is:
(t+1)
θdis
(t+1)
(t)
(t)
(t)
= θdis − η∇θdis Ldis θdis θenc
, x(t) , y (t)
(t)
(t)
(t+1)
(t+1)
(t)
(t)
[θenc
, θdec ] = [θenc
, θdec ] − η∇θenc ,θdec L θenc
, θdec θdis
, x(t) , y (t) .
The details of training and models are given in the next section.
4
Implementation
We adapt the architecture of our network from [9]. Let Ck be a Convolution-BatchNorm-ReLU layer
with k filters. Convolutions use kernel of size 4 × 4, with a stride of 2, and a padding of 1, so that
each layer of the encoder divides the size of its input by 2. We use leaky-ReLUs with a slope of 0.2
in the encoder, and simple ReLUs in the decoder.
The encoder consists of the following 7 layers:
C16 − C32 − C64 − C128 − C256 − C512 − C512
Input images have a size of 256 × 256. As a result, the latent representation of an image consists of
512 feature maps of size 2 × 2. In our experiments, using 6 layers gave us similar results, while 8
layers significantly decreased the performance, even when using more feature maps in the latent state.
To provide the decoder with image attributes, we append the latent code to each layer given as input to
the decoder, where the latent code of an image is the concatenation of the one-hot vectors representing
5
Model
Real Image
IcGAN AE
IcGAN Swap
FadNet AE
FadNet Swap
Mouth
92.6
22.7
11.4
88.4
79.0
Naturalness
Smile Glasses
87.0
88.6
21.7
14.8
22.9
9.6
75.2
78.8
31.4
45.3
Mouth
89.0
88.1
10.1
91.8
66.2
Accuracy
Smile
88.3
91.7
9.9
90.1
97.1
Glasses
97.6
86.2
47.5
94.5
76.6
Table 1: Perceptual evaluation of naturalness and swap accuracy for each model. The naturalness
score is the percentage of images that were labeled as “real” by human evaluators to the question “Is
this image a real photograph or a fake generated by a graphics engine?”. The accuracy score is the
classification accuracy by human evaluators on the values of each attribute.
the values of its attributes (binary attributes are represented as [1, 0] and [0, 1]). We append the latent
code as additional constant input channels for all the convolutions of the decoder. Denoting by n the
number of attributes, (hence a code of size 2n), the decoder is symmetric to the encoder, but uses
transposed convolutions for the up-sampling:
C512+2n − C512+2n − C256+2n − C128+2n − C64+2n − C32+2n − C16+2n .
The discriminator is a C512 layer followed by a fully-connected neural network of two layers of size
512 and n repsectively.
Dropout We found it extremely beneficial to add dropout in our discriminator. We set the dropout
rate to 0.3 in all our experiments. Following [9], we also tried to add dropout in the first layers of the
decoder, but in our experiments, this turned out to significantly decrease the performance.
Discriminator cost scheduling Similarly to [2], we use a variable weight for the discriminator loss
coefficient λE . We initially set λE to 0 and the model is trained like a normal auto-encoder. Then,
λE is linearly increased to 0.0001 over the first 500, 000 iterations to slowly encourage the model
to produce invariant representations. This scheduling turned out to be critical in our experiments.
Without it, we observed that the encoder was too affected by the loss coming from the discriminator,
even for low values of λE .
Model selection Model selection was first performed automatically using two criteria. First, we
used the reconstruction error on original images as measured by the MSE. Second, we also want the
model to properly swap the attributes of an image. For this second criterion, we train a classifier
to predict image attributes. At the end of each epoch, we swap the attributes of each image in the
validation set and measure how well the classifier performs on the decoded images. These two
metrics were used to filter out potentially good models. The final model was selected based on human
evaluation on images from the train set reconstructed with swapped attributes.
5
Experiments
5.1
Experiments on the celebA dataset
Experimental setup We first present experiments on the celebA dataset [13], which contains
200, 000 images of celebrity of shape 178 × 218 annotated with 40 attributes. We used the standard
training, validation and test split. All pictures presented in the paper or used for evaluation have been
taken from the test set. For pre-processing, we cropped images to 178 × 178, and resized them to
256 × 256, which is the resolution used in all figures of the paper. Image values were normalized
to [−1, 1]. All models were trained with Adam [10], using a learning rate of 0.002, β1 = 0.5, and a
batch size of 32. We performed data augmentation by flipping horizontally images with a probability
0.5 at each iteration. As model baseline, we used IcGAN [17] with the model provided by the authors
and trained on the same dataset. 4
4
https://github.com/Guim3/IcGAN
6
Figure 3: Swapping the attributes of different faces. Zoom in for better resolution.
Qualitative evaluation Figure 3 shows examples of images generated when swapping different
attributes: the generated images have a high visual quality and clearly handle the attribute value
changes, for example by adding realistic glasses to the different faces. These generated images
confirm that the latent representation learned by Fader Networks is both invariant to the attribute
values, but also captures the information needed to generate any version of a face, for any attribute
value. Indeed, when looking at the shape of the generated glasses, different glasses shapes and colors
have been integrated into the original face depending on the face: our model is not only adding
“generic” glasses to all faces, but generates plausible glasses depending on the input.
Quantitative evaluation protocol We performed a quantitative evaluation of Fader Networks on
Mechanical Turk, using IcGAN as a baseline. We chose the three attributes Mouth (Open/Close),
Smile (With/Without) and Glasses (With/Without) as they were attributes in common between IcGAN
and our model. We evaluated two different aspects of the generated images: the naturalness, that
measures the quality of generated images, and the accuracy, that measures how well swapping an
attribute value is reflected in the generation. Both measures are necessary to assess that we generate
natural images, and that the swap is effective. We compare: R EAL I MAGE , that provides original
images without transformation, FAD N ET AE and I C GAN AE , that reconstruct original images
without attribute alteration, and FAD N ET S WAP and I C GAN S WAP , that generate images with one
swapped attribute, e.g., With Glasses → Without Glasses. Before being submitted to Mechanical Turk,
all images were cropped and resized following the same processing than IcGAN. As a result, output
images were displayed in 64 × 64 resolution, also preventing Workers from basing their judgment on
the sharpness of presented images exclusively.
Technically, we should also assess that the identity of a person is preserved when swapping attributes.
This seemed to be a problem for GAN-based methods, but the reconstruction quality of our model is
very good (RMSE on test of 0.0009, to be compared to 0.028 for IcGAN), and we did not observe
this issue. Therefore, we did not evaluate this aspect.
For naturalness, the first 500 images from the test set such that there are 250 images for each attribute
value were shown to Mechanical Turk Workers, 100 for each of the 5 different models presented
above. For each image, we asked whether the image seems natural or generated. The description
given to the Workers to understand their task showed 4 examples of real images, and 4 examples of
fake images (1 FAD N ET AE , 1 FAD N ET S WAP , 1 I C GAN AE , 1 I C GAN S WAP ).
The accuracy of each model on each attribute was evaluated in a different classification task, resulting
in a total of 15 experiments. For example, the FadNet/Glasses experiment consisted in asking
Workers whether people with glasses being added by FAD N ET S WAP effectively possess glasses,
and vice-versa. This allows us to evaluate how perceptible the swaps are to the human eye. In each
experiment, 100 images were shown (50 images per class, in the order they appear in the test set).
In both quantitative evaluations, each experiment was performed by 10 Workers, resulting in 5, 000
samples per experiment for naturalness, and 1, 000 samples per classification experiment on swapped
attributes. The results on both tasks are shown in Table 1.
7
Figure 4: (Zoom in for better resolution.) Examples of multi-attribute swap (Gender / Opened eyes /
Eye glasses) performed by the same model. Left images are the originals.
Quantitative results In the naturalness experiments, only around 90% of real images were classified as “real” by the Workers, indicating the high level of requirement to generate natural images. Our
model obtained high naturalness accuracies when reconstructing images without swapping attributes:
88.4%, 75.2% and 78.8%, compared to IcGAN reconstructions whose accuracy does not exceed 23%,
whether it be for reconstructed or swapped images. For the swap, FAD N ET S WAP still consistently
outperforms I C GAN S WAP by a large margin. However, the naturalness accuracy varies a lot based
on the swapped attribute: from 79.0% for the opening of the mouth, down to 31.4% for the smile.
Classification experiments show that reconstructions with FAD N ET AE and I C GAN AE have very
high classification scores, and are even on par with real images on both Mouth and Smile. FAD N ET
S WAP obtains an accuracy of 66.2% for the mouth, 76.6% for the glasses and 97.1% for the smile,
indicating that our model can swap these attributes with a very high efficiency. On the other hand,
with accuracies of 10.1%, 47.5% and 9.9% on these same attributes, I C GAN S WAP does not seem
able to generate convincing swaps.
Multi-attributes swapping We present qualitative results for the ability of our model to swap
multiple attributes at once in Figure 4, by jointly modifying the gender, open eyes and glasses
attributes. Even in this more difficult setting, our model can generate convincing images with multiple
swaps.
5.2
Experiments on Flowers dataset
We performed additional experiments on the Oxford-102 dataset, which contains about 9, 000 images
of flowers classified into 102 categories [16]. Since the dataset does not contain other labels than the
flower categories, we built a list of color attributes from the flower captions provided by [18]. Each
flower is provided with 10 different captions. For a given color, we gave a flower the associated color
attribute, if that color appears in at least 5 out of the 10 different captions. Although being naive, this
approach was enough to create accurate labels. We resized images to 64 × 64. Figure 5 represents
reconstructed flowers with different values of the “pink” attribute. We can observe that the color of
the flower changes in the desired direction, while keeping the background cleanly unchanged.
Figure 5: Examples of reconstructed flowers with different values of the pink attribute. First row
images are the originals. Increasing the value of that attribute will turn flower colors into pink, while
decreasing it in images with originally pink flowers will make them turn yellow or orange.
8
6
Conclusion
We presented a new approach to generate variations of images by changing attribute values. The
approach is based on enforcing the invariance of the latent space w.r.t. the attributes. A key advantage
of our method compared to many recent models [26, 9] is that it generates realistic images of high
resolution without needing to apply a GAN to the decoder output. As a result, it could easily be
extended to other domains like speech, or text, where the backpropagation through the decoder can
be really challenging because of the non-differentiable text generation process for instance. However,
methods commonly used in vision to assess the visual quality of the generated images, like PatchGAN,
could totally be applied on top of our model.
Acknowledgments
The authors would like to thank Yedid Hoshen for initial discussions about the core ideas of the paper,
Christian Pursch and Alexander Miller for their help in setting up the experiments and Mechanical
Turk evaluations. The authors are also grateful to David Lopez-Paz and Mouhamadou Moustapha
Cisse for useful feedback and support on this project.
References
[1] Grigory Antipov, Moez Baccouche, and Jean-Luc Dugelay. Face aging with conditional
generative adversarial networks. arXiv preprint arXiv:1702.01983, 2017.
[2] Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy
Bengio. Generating sentences from a continuous space. arXiv preprint arXiv:1511.06349,
2015.
[3] Andrew Brock, Theodore Lim, JM Ritchie, and Nick Weston. Neural photo editing with
introspective adversarial networks. arXiv preprint arXiv:1609.07093, 2016.
[4] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan:
Interpretable representation learning by information maximizing generative adversarial nets. In
Advances in Neural Information Processing Systems, pages 2172–2180, 2016.
[5] Harrison Edwards and Amos Storkey. Censoring representations with an adversary. arXiv
preprint arXiv:1511.05897, 2015.
[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François
Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural
networks. Journal of Machine Learning Research, 17(59):1–35, 2016.
[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural
information processing systems, pages 2672–2680, 2014.
[8] Geoffrey Hinton, Alex Krizhevsky, and Sida Wang. Transforming auto-encoders. Artificial
Neural Networks and Machine Learning–ICANN 2011, pages 44–51, 2011.
[9] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with
conditional adversarial networks. arXiv preprint arXiv:1611.07004, 2016.
[10] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[11] Tejas D Kulkarni, William F Whitney, Pushmeet Kohli, and Josh Tenenbaum. Deep convolutional inverse graphics network. In Advances in Neural Information Processing Systems, pages
2539–2547, 2015.
[12] Chuan Li and Michael Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In European Conference on Computer Vision, pages 702–716.
Springer, 2016.
[13] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the
wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015.
[14] Gilles Louppe, Michael Kagan, and Kyle Cranmer. Learning to pivot with adversarial networks.
arXiv preprint arXiv:1611.01046, 2016.
9
[15] Michael F Mathieu, Junbo Jake Zhao, Junbo Zhao, Aditya Ramesh, Pablo Sprechmann, and
Yann LeCun. Disentangling factors of variation in deep representation using adversarial training.
In Advances in Neural Information Processing Systems, pages 5041–5049, 2016.
[16] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large
number of classes. In Computer Vision, Graphics & Image Processing, 2008. ICVGIP’08.
Sixth Indian Conference on, pages 722–729. IEEE, 2008.
[17] Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M Álvarez. Invertible
conditional gans for image editing. arXiv preprint arXiv:1611.06355, 2016.
[18] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of
fine-grained visual descriptions. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 49–58, 2016.
[19] Scott E Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In
Advances in Neural Information Processing Systems, pages 1252–1260, 2015.
[20] Jürgen Schmidhuber. Learning factorial codes by predictability minimization. Neural
Computation, 4(6):863–879, 1992.
[21] Yaniv Taigman, Adam Polyak, and Lior Wolf. Unsupervised cross-domain image generation.
arXiv preprint arXiv:1611.02200, 2016.
[22] Paul Upchurch, Jacob Gardner, Kavita Bala, Robert Pless, Noah Snavely, and Kilian Weinberger.
Deep feature interpolation for image content changes. arXiv preprint arXiv:1611.05507, 2016.
[23] Lior Wolf, Yaniv Taigman, and Adam Polyak. Unsupervised creation of parameterized avatars.
arXiv preprint arXiv:1704.05693, 2017.
[24] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional
image generation from visual attributes. In European Conference on Computer Vision, pages
776–791. Springer, 2016.
[25] Jimei Yang, Scott E Reed, Ming-Hsuan Yang, and Honglak Lee. Weakly-supervised disentangling with recurrent transformations for 3d view synthesis. In Advances in Neural Information
Processing Systems, pages 1099–1107, 2015.
[26] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image
translation using cycle-consistent adversarial networks. arXiv preprint arXiv:1703.10593, 2017.
10
| 1cs.CV
|
Learning random-walk label propagation
for weakly-supervised semantic segmentation
arXiv:1802.00470v1 [cs.CV] 1 Feb 2018
Paul Vernaza
Manmohan Chandraker
NEC Laboratories America, Media Analytics Department
10080 N Wolfe Road, Cupertino, CA 95014
{pvernaza,manu}@nec-labs.com
Abstract
input
image
Large-scale training for semantic segmentation is challenging due to the expense of obtaining training data for
this task relative to other vision tasks. We propose a novel
training approach to address this difficulty. Given cheaplyobtained sparse image labelings, we propagate the sparse
labels to produce guessed dense labelings. A standard
CNN-based segmentation network is trained to mimic these
labelings. The label-propagation process is defined via
random-walk hitting probabilities, which leads to a differentiable parameterization with uncertainty estimates that
are incorporated into our loss. We show that by learning
the label-propagator jointly with the segmentation predictor, we are able to effectively learn semantic edges given
no direct edge supervision. Experiments also show that
training a segmentation network in this way outperforms
the naive approach. 1
sparse
training
labels
Semantic edge
predictor
Arbitrary
segmentation
net
Random-walk label propagator
propagated
labels
Consistency
loss
training time only
predicted
segmentation
1. Introduction
Figure 1: Overview of proposed training method.
We consider the task of semantic segmentation, which
is to learn a predictor capable of accurately assigning a semantic label to each pixel in an image. As with many other
popular vision problems, convolutional neural nets (CNNs)
have emerged as the leading tool to solve semantic segmentation problems, in part due to their ability to leverage large
datasets effectively. However, datasets for semantic segmentation remain orders of magnitude smaller than for tasks
such as classification and detection, due chiefly to the much
higher annotation expense of this task.
To ease the annotation burden, and in line with previous
work [1, 11], we propose a method for training CNN-based
semantic segmentation networks given sparse annotations,
such as the scribbles depicted in Fig. 1, which also depicts
our proposed training strategy. The idea of our method
is to learn mutually-consistent networks for propagating
the sparse labels to unlabeled points, and predicting the
true labeling given the image alone. Optimizing a mutualconsistency objective obviates the need for dense (or, fullylabeled) supervision. A key innovation of our approach is
proposing to use a specific, probabilistic model of sparse
label propagation that is a differentiable function of semantic boundary predictions. As it is differentiable, minimizing our loss via gradient-based methods results in simultaneous learning of an image-to-semantic-boundary predictor
and an image-to-semantic-segmentation predictor, despite
having no direct observations of semantic edges.
Our method is comparable to recent work by Lin et
al. [11], which proposes alternating between propagating
sparse labels using a CRF defined over superpixels, and
1 This article is a corrected version of an article published in CVPR
2017: https://doi.org/10.1109/CVPR.2017.315
1
Table 1: Notation
X ⊂ Z2
X̂ ⊂ X
L
y:X→L
ŷ : X̂ → L
I
θ, φ
∆n
Qθ,I : X → ∆|L|
Bφ,I : X → R+
Ξ ⊂ Z+ → X
ξ∈Ξ
τ : Ξ → Z+
Pa|b ∈ ∆
δi ∈ {0, 1}|δi | ∩ ∆|δi |
H(p)
H(p, q)
KL(p k q)
set of all pixel locations
set of labeled pixels
set of semantic labels
a semantic labeling function
a sparse labeling
a generic image
predictor parameters
n-dim. probability simplex
label predictor
boundary predictor
set of paths across image
a random walk
path stopping time
distribution of a given b
Kronecker delta vector
Shannon entropy of p
cross-entropy of p, q
KL-divergence of p, q
training a CNN to predict the labels thus inferred. A disadvantage of this approach relative to ours is that [11] employs
a notion of label smoothness that is non-adaptive: specifically, it is assumed that labels are constant within superpixels, and the CRF binary potentials are not learned. This
ultimately places an artificial upper bound on the accuracy
of the training data observed by the CNN—an upper bound
that never improves as more data is collected. By contrast,
by expressing label propagation in terms of a learned semantic boundary predictor, we are able to learn a concept
of label propagation that is entirely data-driven, enabling
our method to scale fully with the data. Furthermore, the
probabilistic nature of our label propagation method allows
us to obtain uncertainty estimates that are directly incorporated into our learning process, mitigating the possibilty of
training on propagated labels that are incorrect.
A crucial technical component of our approach is defining the label propagation process in terms of random-walk
hitting probabilities [6], which enables efficient inference
and gradient-based learning. For this reason, we refer to
our approach as RAWKS, a contraction of RAndom-walk
WeaKly-supervised Segmentation.
2. Method
Given densely labeled training images, typical approaches for deep-learning-based semantic segmentation
minimize the following cross-entropy loss [12, 3]:
X
min
H(δy(x) , Qθ,I (x)),
(1)
θ∈Θ
x∈X
where δy(x) ∈ ∆|L| is the indicator vector of the ground
truth label y(x) and Qθ,I (x) is a label distribution predicted
by a CNN with parameters θ evaluated on image I at location x. See Table 1 for notation. In our case, dense labels
y(x) are not provided: we only have labels ŷ provided at a
subset of points X̂. Our solution is to simultaneously infer
the dense labeling y given the sparse labeling ŷ and use the
inferred labeling to train Q. To achive this, we propose to
train our predictor to minimize the following loss:
X
min
H(Py(x)|ŷ,Bφ,I , Qθ,I (x)).
(2)
θ,φ
x∈X
Here, Q is a predictor of the same form as before, while
Py(x)|ŷ,Bφ,I is a predicted label distribution at x, conditioned on predicted semantic boundaries Bφ,I and the
sparse labels. Bφ,I is assumed to be a nonnegative-valued
CNN-based boundary predictor, in a similar vein as [2, 14].
See Fig. 1 for a graphical overview of our model.
The key to our method is the definition of the propagated
label distributions Py|ŷ,Bφ,I . As in prior work on interactive
segmentation [6], we define these distributions in terms of
random-walk hitting probabilities, which can be computed
analytically, and which allows us to compute the derivatives of the propagated label probabilities with respect to the
predicted boundaries. This enables us to minimize (2) by
pure backpropagation, without tricks such as the alternating
optimization methods employed in prior weakly-supervised
segmentation methods [4, 11].
To elaborate, we first define the set of all 4-connected
paths on the image that end at the first labeled point encountered:
Ξ = {(ξ0 , ξ1 , . . . ,ξτ (ξ) ) | ∀t, ξt ∈ X, kξt+1 − ξt k1 = 1,
ξτ (ξ) ∈ X̂, ∀t0 < τ (ξ) , ξt0 ∈
/ X̂}.
(3)
We now assign each path ξ ∈ Ξ a probability that decays
exponentially as it crosses boundaries:
τ (ξ)−1
X
1 τ (ξ) 2
,
(4)
P (ξ) ∝ exp −
Bφ,I (ξt )
4
t=0
where Bφ,I is assumed to be a nonnegative boundary score
prediction. High- and low-probability paths under this
model are illustrated in Fig. 2b. Given sparse labels ŷ, the
probability that a pixel x has label y 0 is then defined as the
probability that a path starting at x eventually hits a point
labeled y 0 , given the distribution over paths (4):
Py(x)|ŷ,Bφ,I (y 0 ) = P ({ξ ∈ Ξ | ξ0 = x, ŷ(ξτ (ξ) ) = y 0 }).
(5)
2 The factor 4−τ (ξ) is sufficient to ensure convergence of the partition
function as long as Bφ,I ≥ 0. This factor was mistakenly omitted in the
original version of this paper.
(a)
(b)
Figure 2: Illustration of random-walk-based model for
defining label probabilities based on boundaries and sparse
labels.
Boundaries
Image
Sparse
labels
Dense
labels
Figure 3: Independence assumptions as a graphical model
This quantity can be computed efficiently by solving a
sparse linear system, as described in Sec. 2.2. The randomwalk model is illustrated in Fig. 2a. Any pixel in the region
labeled (ii), for example, is very likely to be labeled dog
instead of chair or background, because any path of significant probability starting in (ii) will hit a pixel labeled dog
before it hits a pixel with any other label.
To recap, the overall architecture of our method is summarized in Fig. 1. At training time, an input image is passed
to an arbitrary segmentation predictor Qθ,I and an arbirtary
semantic edge predictor Bφ,I . The semantic edge predictions and sparse training labels ŷ are passed to a module
that computes propagated label probabilities Py|ŷ,B using
the random-walk model described above. The propagated
label probabilities P and the output of the predictor Q are
then passed to a cross-entropy loss H(P, Q), which is minimized in the parameters θ and φ via backpropagation. At
test time, Q is evaluated and used as the prediction. P is not
evaluated at test time, since labels are not available.
over the unobserved dense labels, given the observed sparse
labels. This requires us to define P (y | ŷ, I). We submit
that a natural way to do so is to introduce a new variable B
representing the image’s semantic boundaries. This results
in the proposed graphical model in Fig. 3 to represent the
independence structure of y, ŷ, B, I.
It is then straightforward to show that (2) is equivalent
to marginalizing (1) with respect to a certain distribution
P (y | ŷ, I), after making a few assumptions. First, the
conditional independence structure depicted in Fig. 3 is assumed. y(x) and y(x0 ) are assumed conditionally independent given ŷ, B, ∀x 6= x0 ∈ X, which allows us to specify
P (y | ŷ, B) in terms of marginal distributions and simplifies inference. Finally, B is assumed to be a deterministic
function of I, defined via parameters φ.
We note that the conditional independence assumptions
made in Fig. 3 are significant. In particular, y is assumed
independent of I given ŷ and B. This essentially implies
that there is at least one label for each connected component of the true label image, since knowing the underlying
image usually does give us information as to the label of
unlabeled connected components. In practice, strictly labeling every connected component is not necessary. However,
training data that egregiously violates this assumption will
likely yield poor results.
2.2. Random-walk inference
Key to our method is the efficient computation of the
random-walk hitting probabilities (5). It is well-known that
such probabilities can be computed efficiently via solving
linear systems [6]. We briefly review this result here.
The basic strategy is to compute the partition function
Zxl , which sums the right-hand-side of (4) over all paths
starting at x and ending in a point labeled l. We can then
derive a dynamic programming recursion expressing Zxl in
terms of the same quantity at neighboring points x0 . This recursion defines a set of sparse linear constraints on Z, which
we can then solve using standard sparse solvers.
We first define Ξxl := {ξ ∈ Ξ | ξ0 = x, ŷ(ξτ (ξ) ) = l}.
Zxl is then defined as
τ (ξ)−1
X
X
1 τ (ξ)
. (6)
Zxl :=
exp −
Bφ,I (ξt )
4
t=0
ξ∈Ξxl
The first term in the inner sum can be factored out by introducing a new summation over the four nearest neighbors of
x, denoted x0 ∼ x, easily yielding the recursion
Zxl =
2.1. Probabilistic justification
The proposed loss function (2) arises from a natural
probabilistic extension of (1) to the case where dense labels
are unobserved. Specifically, we consider marginalizing (1)
τ (ξ)−1
1 −Bφ,I (x) X 1 τ (ξ) − Pt=0
Bφ,I (ξt )
e
e
4
4
0
x ∼x,
ξ∈Ξx0 l
=
1 −Bφ,I (x) X
e
Zx0 l .
4
0
x ∼x
(7)
Boundary conditions must also be considered in order to
fully constrain the solution. Paths exiting the image are assumed to have zero probability: hence, Zxl := 0, ∀x ∈
/ X.
Paths starting at a labeled point x ∈ X̂ immediately terminate with probability 1; hence, Zxl := 1, ∀x ∈ X̂. Solving
this system yields a unique solution for Z, from which the
desired probabilities are computed as
Zxy0
.
Py(x)|ŷ,Bφ,I (y 0 ) = P
l∈L Zxl
Image
Boundary scores
Random walk
segmentation
Uncertainty weights
(8)
2.3. Random-walk backpropagation
In order to apply backpropagation, we must ultimately
compute the derivative of the loss with respect to a change
in the boundary score prediction Bφ,I . Here, we focus on
computing the derivative of the partition function Z with respect to the boundary score B, the other steps being trivial.
Since computing Z amounts to solving a linear system, this turns out to be fairly simple. Let us write the
constraints (7) in matrix form Az = b, such that A is
square, zi := Zxi (assigning a unique linear index i to
each xi ∈ X, and temporarily omitting the dependence on
l), and thePith rows of A, b correspond to the constraints
Ci Zxi = x0 ∼xi Zx0 , Zxi = 0, or Zxi = 1 (as appropriate), where Ci := exp(Bφ,I (xi )). Let us consider the effect
of adding a small variation V to A, and then re-solving the
system. It can be shown that
(A + V )−1 b = z − A−1 V z + O(2 ).
(9)
Substituting the first-order dependence on V into a Taylor
expansion of the loss L yields:
dL
−1
−1
, A V z + O(2 )
L((A + V ) b) = L(z) −
dz
−1| dL |
= L(z) − A
z , V + O(2 ).
dz
A first-order variation of Ci corresponds to Vi = −δii ,
which implies that
dL
|−1 dL
= A
zi .
(10)
dCi
dz i
In summary, this implies that computing the loss derivatives
with respect to the boundary score can be implemented efdL
ficiently by solving the sparse adjoint system A| dC
= dL
dz ,
and multiplying the result pointwise by the partition function z, which in turn allows us to efficiently incorporate
sparse label propagation as a function of boundary prediction into an arbitrary deep-learning framework.
2.4. Uncertainty-weighting the loss
An advantage of our method over prior work is that the
random-walk method produces a distribution over dense labelings Py|ŷ,Bφ,I given sparse labels, as opposed to a MAP
Figure 4: Visualization of loss uncertainty weights (blue =
low weight, red = high weight)
estimate. These uncertainty estimtates can be used to downweight the loss in areas where the inferred labels may be
incorrect, as illustrated in Fig. 4. In this example, the
boundary predictor failed to correctly predict parts of object boundaries. In the vicinity of these gaps, the label distribution is uncertain, and the MAP estimate is incorrect.
However, we can mitigate the problem by down-weighting
the loss proportional to the uncertainty estimate.
More concretely, we actually minimize the following
modification of the loss (2):
X
w(x)KL(Py(x)|ŷ,Bφ,I k Qθ,I (x)) + H(Py(x)|ŷ,Bφ,I ),
x∈X
(11)
where we define w(x) := exp(−αH(Py(x)|ŷ,Bφ,I )), for
some fixed parameter α. This loss reduces to (2) for the
case w(x) = 1. Although the KL component of the loss
can be avoided by increasing the prediction entropy, the explicit entropy regularization term prevents trivial solutions
of very large entropy everywhere.
3. Related work
The method most comparable to ours is the work of
Lin et al. [11]. In contrast to [11], our method features
fully-differentiable, gradient-based training (as opposed to
alternating optimization); we learn an inductive rule for
predicting boundaries and propagating labels, as opposed
to using non-adaptive superpixels and a CRF with nonadaptive binary potentials, which enables us to adapt to
large datasets in a data-driven way; and we employ a probabilistic notion of label propagation that enables us to define an uncertainty-weighted loss that mitigates the possibilty of training on propagated labels that are incorrect. Another notable method in the same vein as [11] is the BoxSup
method [4], which also employs alternating optimization,
but uses bounding-box annotations as weak supervision.
Our method was initially inspired by [1], which introduced the idea of training on what we refer to here as sparse
labels as a source of weak supervision. Instead of attempting to directly propagate labels, as we do, that method leverages a notion of objectness to mitigate overfitting. A few
other works have proposed different modes of weak supervision for segmentation. Notably, [15] and [13] both model
weak supervision as imposing linear constraints to be satisfied by the predictor, resulting in models trained by alternating optimization. Our method can be viewed from a
similar perspective, since we also impose our weak supervision via linear constraints (7); however, our constraints
explicitly model the process of spatial label propagation,
whereas the constraints proposed in [15, 13] model only
aggregate statistics over regions, and hence make no provision for learning boundaries as we do in this work. Furthermore, our model is differentiable and can be optimized
via gradient-based methods.
To the extent that it learns semantic edges with a CNN,
our method is similar to previous works such as [14], which
also learns semantic edges using a CNN. However, [14]
achieves this using direct supervision of edges—which we
do not require—and does not jointly train a semantic labeler, as we do. Learning edges with a weaker form of edge
supervision is proposed in [10]; however, this method relies on a combination of heuristic boundary detectors and
bounding boxes for supervision. To reiterate, our method
works without any heuristic source of boundaries as input.
Another vein of prior work relates to some form of joint
reasoning about boundaries and segments. Most prominently, random walks were previously applied to interactive
segmentation in [6]. However, that work did not consider
learning these random walks by learning boundary scores,
as we do, nor did it consider the task of semantic image segmentation in general. More recently, [2] proposed a method
to jointly learn semantic edges and a CNN-based semantic
labeler in an gradient-based learning framework. However,
their method applies only to the strong-supervision case,
and cannot leverage sparse annotations of the kind we employ here.
4. Experiments
We implemented RAWKS in the Caffe [9] framework.
The semantic-boundary-prediction network Bφ,I and the
semantic-segmentation network Qθ,I were both implemented as fully-convolutional CNNs, based on the same
ResNet-101 [8] architecture. The final average-pooling and
fully-connected layers were removed, and features from
the last resulting layer were upsampled and combined with
intermediate-layer features to produce a 4x-downsampled
output for both the semantic boundary and label predictions.
Both networks were initialized from a model trained for
classification on the ImageNet 2012 dataset. We applied no
data augmentation techniques in training any of the methods.
RAWKS was trained on the publicly available scribble
annotations provided by [11]. We used the same training and validation splits as [11], for both the PASCAL
VOC 2012 and PASCAL CONTEXT datasets: for VOC,
the validation set consisted of the VOC 2012 validation set,
while the training set consisted of all other images labeled
in either the VOC 2012 dataset or the PASCAL Semantic Boundary dataset [7] (10582 training images, and 1449
validation images). We trained models for each of these
datasets independently.
We evaluated the performance of both the semanticsegmentation network Q and the label-propagation network
P (propagating the sparse labels given the learned boundaries). To summarize, evaluating the predicted labels Q
on the validation set, RAWKS slightly underperformed the
published results of [11] on VOC 2012, while slightly outperforming [11] on CONTEXT. Our other major observation was that the propagated labels P on the training set
were approximately as accurate as the best possible labeling of a superpixel segmentation of the images.
To elaborate, in Table 2, MIOU refers to the meanintersection-over-union metric, while w/ CRF refers to the
same metric evaluated after post-processing the results with
a fully-connected CRF, as in [11]. RAWKS Qθ,I refers to
the evaluation of the predicted labels Q on the validation
set given the image alone, after jointly training P and Q via
SGD. RAWKS train P,Q then Q also refers to evaluation of
Q, but with a slightly different training protocol: in this case
after jointly training P and Q with SGD, Q was fine-tuned
with P fixed. RAWKS training Py|ŷ,B , 0% abstain refers to
evaluation of P on the training set, given the sparse training labels and learned boundaries Bφ,I . RAWKS training
Py|ŷ,B , 6% abstain consists of the same evaluation, but allowing P to abstain from prediction on 6% of the pixels,
which (for this particular model) corresponds to abstaining
on all pixels with a confidence score w(x) below 0.5 (c.f.
Sec. 2.4). The next section of Table 2 reports baselines:
sparse-loss baseline consists of training the same base network Q, but using loss (1) (i.e., without P ), evaluating it
only at the sparse locations X̂. train on dense ground truth
is the result we obtain training our base network Q on the
dense ground-truth training data. ScribbleSup refers to the
result reported by [11], which we report here verbatim. We
note that [11] used a different base segmentation network
(DeepLab) than we used in our experiments.
The last section of Table 2 reports statistics for baselines meant to represent best-case performance bounds for a
superpixel-based method such as [11]. These were obtained
by segmenting the input images using the method [5] (with
author-suggested parameters), and labeling the resulting superpixels in different ways. SPOPT corresponds to labeling
each superpixel with the majority label from the groundtruth dense segmentation. SPCON differs from SPOPT only
on superpixels containing scribble annotations: to these,
SPCON assigns the majority label from the scribbles contained within. Train on SPOPT/SPCON refer to training
predictors Q using the labelings of SPOPT and SPCON,
respectively. These results are interesting for a number of
reasons. First, the propagated labelings P that we deduce
in the course of training, are nearly as good (VOC) or better
(CONTEXT) than the best possible results obtainable using
superpixels. Second, we see that training with our propagated labelings P is competitive with training on optimal
superpixel labelings. Finally, we emphasize that while the
superpixel baselines cannot improve with training data (as
they are not trained), our label propagation model is naturally refined as we train on larger datasets.
In relative terms, RAWKS performed better on the CONTEXT dataset, as evidenced in Table 3. One potential reason for this is the greater number of classes for this dataset
(60 vs. 21 for VOC), which naturally calls for finer boundaries. Since our method is able to adaptively learn boundaries suited to the task, while [11] uses non-adaptive heuristics to generate superpixels, this may account for the better relative performance of RAWKS in this context. We
also hypothesize that it is easier to learn semanic boundaries when there are a greater number of classes, because
low-level edges and features become a more informative
cue in this case. Surprisingly, our dense ground-truth baseline performed significantly worse than RAWKS; we hypothesize this is due to overfitting, a consequence of the
smaller amount of training data and increased number of
classes in CONTEXT, exacerbated by our use of the verydeep ResNet model. Joint training of the propagator network P in RAWKS seems to have a regularizing effect that
may have prevented overfitting to some extent.
Qualitative validation-set results are shown in Fig. 5 for
the VOC 2012 dataset, while CONTEXT training-set results are shown in Fig. 6. The training-set results of Fig. 6
demonstrate that RAWKS is able to deduce high-quality semantic boundaries, thereby producing propagated labelings
P that are a close approximation to the ground truth dense
labelings (which are not used at training time, to be clear).
In the validation-set results of Fig. 5, the loss weights w(x)
and propagated labels P are shown in addition to the predictions Q—to be clear, these depend on the sparse labels
for these specific examples, which were not used to train
this model. Here we remark that our semantic boundary
predictions also generalize well to the validation set. Although we did not train on these images, we also observe
that had we done so, the loss weights would have behaved
appropriately, down-weighting the loss in regions where the
Method
RAWKS Qθ,I
RAWKS train P, Q, then Q
RAWKS training Py|ŷ,B , 0% abstain
RAWKS training Py|ŷ,B , 6% abstain
Sparse-loss baseline
Train on dense ground truth
ScribbleSup (reported)
Opt. superpixel labels (SPOPT)
Opt. consistent s-p. labels (SPCON)
Train on SPOPT
Train on SPCON
MIOU
57.1
59.5
75.8
81.2
55.8
66.3
.
83.1
76.5
62.8
61.1
w/ CRF
60.0
61.4
.
.
68.8
63.1
.
.
.
.
Table 2: Results on VOC 2012 validation set
Method
RAWKS Qθ,I
RAWKS training Py|ŷ,B , 0% abstain
Sparse-loss baseline
Train on dense ground truth
ScribbleSup (reported)
Opt. consistent s-p. labels (SPCON)
MIOU
36.0
75.5
26.6
31.7
.
70.2
w/ CRF
37.4
.
.
32.4
36.1
.
Table 3: Results on CONTEXT validation set
propagated labels are incorrect. This seems to happen most
often in regions with very fine boundaries (such as the mast
of the boat and the airplane’s wing), where our limited resolution sometimes causes missed boundaries.
In general, we note that subjectively, resolution seemed
to be a limiting factor in the accuracy of our boundary
prediction and label propagation steps. We used quarterresolution outputs (typicaly around 128x96 pixels) for these
steps in order to minimize the computational cost of computing random-walk hitting probabilities. An average
forward-backwards pass of the entire network took about
1.1 s per image, with about 800 ms of that spent solving for
the random-walk hitting probabilities. This layer was implemented using a CPU-based sparse linear system solver,
whereas the rest of the network was run on the GPU (an
NVIDIA GTX 1080). We anticipate that implementing this
layer using GPU operations will allow us to increase the
resolution of these critical steps, which will in turn lead to
increased prediction accuracy.
5. Conclusions
We have presented a novel approach to mitigating the
expense of procuring labeled data in semantic segmentation, through a framework that utilizes only sparse clicks
or scribbles for training. This has a significant impact
on the possibilities for semantic segmentation—for a given
dataset, one may obtain competitive labels at a fraction of
image
predicted
boundaries
loss
weights
propagated
labels P
dense
label GT
predictor
output Q
Figure 5: Validation set results on VOC 2012 dataset
the cost and conversely, for a given budget, one may obtain labeled data at a much larger scale. Our main technical contribution is a random-walk based label propagation mechanism, which is shown to be differentiable and
usable in powerful deep neural network architectures for
semantic segmentation. We achieve this through a novel
predictor-propagator paradigm, which produces uncertainty
estimates for inferred dense labels given sparse labels.
We demonstrate encouraging results on challenging benchmarks. More importantly, we argue that our framework has
inherent advantages over prior works, since our label propagation is not artificially upper-bounded by superpixel baselines, rather, can keep improving with larger-scale training
data. Also, we note that our contribution is equally valid for
any state-of-the-art CNN-based semantic segmentation engines. In future work, we will explore other state-of-the-art
segmentation architectures and incorporate other forms of
weak supervision such as bounding boxes.
image
predicted
boundaries
sparse
labels
propagated
dense
labels P label GT
Figure 6: Training results on CONTEXT dataset
predictor
output Q
References
[1] A. Bearman, O. Russakovsky, V. Ferrari, and L. Fei-Fei.
What’s the point: Semantic segmentation with point supervision. arXiv preprint arXiv:1506.02106, 2015. 1, 5
[2] L.-C. Chen, J. T. Barron, G. Papandreou, K. Murphy, and
A. L. Yuille. Semantic image segmentation with task-specific
edge detection using cnns and a discriminatively trained domain transform. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5
[3] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. In ICLR, 2015. 2
[4] J. Dai, K. He, and J. Sun. Boxsup: Exploiting bounding
boxes to supervise convolutional networks for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1635–1643, 2015. 2, 5
[5] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient graphbased image segmentation. International Journal of Computer Vision, 59(2):167–181, 2004. 5
[6] L. Grady. Random walks for image segmentation. IEEE
transactions on pattern analysis and machine intelligence,
28(11):1768–1783, 2006. 2, 3, 5
[7] B. Hariharan, P. Arbeláez, L. Bourdev, S. Maji, and J. Malik. Semantic contours from inverse detectors. In 2011 International Conference on Computer Vision, pages 991–998.
IEEE, 2011. 5
[8] K. He, X. Zhang, S. Ren, , and J. Sun. Deep residual learning
for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016. 5
[9] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional
architecture for fast feature embedding. arXiv:1408.5093,
2014. 5
[10] A. Khoreva, R. Benenson, M. Omran, M. Hein, and
B. Schiele. Weakly supervised object boundaries. In CVPR,
2016. 5
[11] D. Lin, J. Dai, J. Jia, K. He, , and J. Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In Conference on Computer Vision and
Pattern Recognition (CVPR), 2016. 1, 2, 4, 5, 6
[12] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In CVPR, pages 3431–
3440, 2015. 2
[13] D. Pathak, P. Krahenbuhl, and T. Darrell. Constrained convolutional neural networks for weakly supervised segmentation. In Proceedings of the IEEE International Conference
on Computer Vision, pages 1796–1804, 2015. 5
[14] S. Xie and Z. Tu. Holistically-nested edge detection. In The
IEEE International Conference on Computer Vision (ICCV),
December 2015. 2, 5
[15] J. Xu, A. G. Schwing, and R. Urtasun. Learning to segment under various forms of weak supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 3781–3790, 2015. 5
| 1cs.CV
|
Type theory and its meaning explanations
arXiv:1512.01837v2 [cs.LO] 14 Jul 2016
Jonathan Sterling
Thanks to Bob Harper, Peter Dybjer, Bengt Nordström, Carlo Angiuli and Darin
Morrison for invaluable conversations which helped sharpen my view of the
meaning explanations for computational and intensional type theories. All
mistakes are mine.
Abstract. At the heart of intuitionistic type theory lies an intuitive semantics
called the “meaning explanations”; crucially, when meaning explanations are
taken as definitive for type theory, the core notion is no longer “proof” but
“verification”. We’ll explore how type theories of this sort arise naturally
as enrichments of logical theories with further judgements, and contrast this
with modern proof-theoretic type theories which interpret the judgements and
proofs of logics, not their propositions and verifications.
Contents
Chapter 1. Logical Theories
1. Judgements of a logical theory
5
5
2. Higher-order judgements
3. Propositions and verifications
6
7
4. Judgements for verifications
8
Chapter 2. Computational Type Theories
1. The categorical judgements
11
11
2. The functional sequent judgement
3. The definitions of types
12
14
Chapter 3. Proof Theories and Logical Frameworks
19
1. Proof-theoretic type theory
2. Martin-Löf’s equational logical framework
19
21
3. A critique of MLLF
4. A way forward: verifications & uses
23
25
Bibliography
27
3
CHAPTER 1
Logical Theories
To start, we will consider the notion of a logical theory; in my mind, it begins with a species of judgements that can be proposed, asserted, and (if they are
evident) known.
1. Judgements of a logical theory
The basic forms of judgement for a logical theory will be P prop and P true ;
and what is P ? It is a member of the species of terms, which are made meaningful
in the course of making the judgement P prop (“P is a proposition”) evident for a
proposition P .
The forms of judgement may be construed as containing inputs and outputs; an
input is something which is inspected in the course of knowing a judgement, whereas
an output is something which is synthesized (or created) in the course of knowing
a judgement. The positions of inputs and outputs in a judgement constitute what
is called its mode, and we color-code it in this presentation for clarity.1
To each judgement is assigned a meaning explanation, which explicates the
knowledge-theoretic content of the judgement. For a judgement J , a meaning
explanation should be in the form:
To know J is to know...
The meaning of the judgement P prop is, then, as follows:
To know P prop is to know that P is a proposition, which is to
know what would count as a direct verification of P .
So if a symbol P is taken to denote a proposition, we must know what sort of
thing is to be taken as a direct verification of P , and this is understood as part
of the definition of P . A “direct verification” is understood in constrast with an
“indirect verification”, which is to be thought of as a means or plan for verifying
the proposition; these distinctions will be explained in more detail later on. Now,
the judgement P true (“P is true”) is only meaningful in case we know P prop
(this is called a presupposition). Then the meaning of P true is as follows:
1We will not see any judgements with outputs at first, but it will become necessary as soon as we
consider judgements about computation, where the reduction of a term is synthesized from the
redex. Modes may be used to construe a judgement as having algorithmic content.
5
6
1. LOGICAL THEORIES
To know P true is to have a verification of P .
From the (implicit) presupposition P prop, we already know what counts as a
verification, so the meaning explanation is well-defined. Note that having a means
or plan for verifying P is equivalent to having a (direct) verification; this follows
from the fact that one may put into action a plan for verifying P and achieve such a
verification, and likewise, it is possible to propound a plan of verification by appeal
to an existing verification.
2. Higher-order judgements
The judgements we have described so far are “categorical” in the sense that
they are made without assumption or generality.
2.1. Hypothetical judgement. We will need to define a further form of
judgement, which is called “hypothetical”, and this is the judgement under hypothesis J (J ′ ) , pronounced “J under the assumption J ′ ”.Its meaning explanation
is as follows:
To know the judgement J (J ′ ) is to know the judgement J
assuming you know the judgement J ′ .
Hypothetical judgement may be iterated, and J (J1 , J2 ) will be used as notation for for J (J2 ) (J1 ).
2.2. General judgement. Another kind of higher order judgement is “general judgement”, which is judgement with respect to a variable, |x J , pronounced
“for an arbitrary x, J ”. The meaning explanation for this new judgement is as
follows:
To know the judgement |x J is, to know [E/x]J (i.e. the substitution of E for x in the expression J ) for any arbitrary expression
E, 2
As far as notation is concerned, the bar symbol binds the least tightly of all the
other notations we have considered. Likewise, general judgement may be iterated,
and the notation |x,y J will be used as notation for |x |y J .
2.3. Hypothetico-general judgement. When hypothetical judgement is used
inside general judgement, as in |x A(x) true (B(x) true), we term the whole thing
a “hypothetico-general” judgement. One thing bears clarifying, which is, Why do
we write P true (P true) rather than |P P true (P true)?
The former is really not a single judgement, but rather a scheme for judgements, where P is intended to be replaced with a concrete expression by the person
2Technically, E is qualified as being of the same valence as x, but because we have not developed
a formal theory of expressions in this presentation, I choose to ignore this issue.
3. PROPOSITIONS AND VERIFICATIONS
7
asserting the judgements. On the other hand, the latter is itself a single judgement
which may be asserted all on its own.
3. Propositions and verifications
Now that we have propounded and explained the minimal system of judgements
for a logical theory, let us populate it with propositions. First, we have falsity ⊥,
and we wish to make ⊥ prop evident; to do this, we simply state what counts as a
direct verification of ⊥: there is no direct verification of ⊥.
The next basic proposition is trivial truth ⊤, and to make ⊤ prop evident, we
state that a direct verification of ⊤ is trivial. The definition of ⊤ thus validates the
judgement ⊤ true (i.e. that we have a verification of ⊤; this is immediate).
Next, let us define conjunction; in doing so, we will make evident the hypothetical judgement
P ∧ Q prop (P prop, Q prop)
Equivalently, we can display this as a rule of inference:3
P prop Q prop
P ∧ Q prop
A direct verification of P ∧ Q consists in a verification of P and a verification
of Q; this validates the assertion of the judgement P ∧ Q true (P true, Q true).
Because it is a valid inference, we can write it as an inference rule:
P true Q true
P ∧ Q true
A direct verification of P ∨ Q may be got either from a verification of P or one
of Q. From this definition we know P ∨ Q prop (P prop, Q prop), or
P prop Q prop
P ∨ Q prop
The verification conditions of disjunction give rise to two evident judgements P ∨ Q true (P true)
and P ∨ Q true (Q true), which we can write as inference rules:
P true
P ∨ Q true
Q true
P ∨ Q true
Finally, we must define the circumstances under which P ⊃ Q is a proposition
(i.e. when P ⊃ Q prop is evident). And we intend this to be under the circumstances
3 Evident hypothetical judgements are often written as rules, i.e.
premise
conclusion
rather than conclusion (premise). It must be stressed that only evident/known judgements may
be written in this way.
8
1. LOGICAL THEORIES
that P is a proposition, and also that Q is a proposition assuming that P is true.
In other words, P ⊃ Q prop (P prop, Q prop (P true)), or
P prop Q prop (P true)
P ⊃ Q prop
Now, to validate this judgement will be a bit more complicated than the previous ones. But by unfolding the meaning explanations for hypothetical judgement,
proposition-hood and truth of a proposition, we arrive at the following explanation:
To know P ⊃ Q prop (P prop, Q prop (P true)) is to know what
counts as a direct verification of P ⊃ Q when one knows what
counts as a direct verification of P , and, when one has such a
verification, what counts as a direct verification of Q.4
If the judgement P ⊃ Q prop (P prop, Q prop (P true)) is going to be made
evident, then we must come up with what should count as a direct verification of
P ⊃ Q under the assumptions described above.
And so to have a direct verification of P ⊃ Q is to have a verification of Q
assuming that one has one of P ; this is the meaning of implication, and it validates
the judgement P ⊃ Q true (Q true (P true)), and may be written as an inference
rule as follows:
Q true (P true)
P ⊃ Q true
4. Judgements for verifications
So far, we have given judgements which define what it means to be a proposition, namely P prop, and thence for each proposition, we have by definition a
notion of what should count as a verification of that proposition. And we have a
judgement P true, which in its assertion means that one has (a way to obtain) such
a verification of P , but we have not considered any judgements which actually refer
to the verifications themselves symbolically.
It is a hallmark of Martin-Löf’s program to resolve the contradiction between
syntax and semantics not by choosing symbols over meanings or meanings over
symbols, but by endowing symbols with meaning in the course of knowing the
evident judgements. As such, P is a symbol, but when we assert P prop we
are saying that we know what proposition P denotes.
A similar thing can be done with verifications themselves, by representing them
with symbols in the same way we have done for the propositions. And then, we
can consider a judgement such as “M is a verification of P ”, and in knowing that
4Note that unless P true, it need not be evident that Q prop; in other words, Q only has to be a
proposition if P is true. It would also be acceptable to give a stronger definition to implication,
but this is the one accepted by Martin-Löf.
4. JUDGEMENTS FOR VERIFICATIONS
9
judgement, we know what verification M is meant to denote. In practice, this
judgement has been written in several ways:
Notation
M ∈P
Pronunciation
M is an element of P
M P
M realizes P
P xext M y P is witnessed by M
But they all mean the same thing, so we will choose the notation M ∈ P
and pronounce it “M verifies P ”. Tentatively, the following defective meaning
explanation could be given:
* To know M ∈ P is to know that M is a verification of P .
But now that we have started to assign expressions to verifications, we must be
more careful about differentiating direct verifications (which we will call “canonical”) from indirect verifications (which we will call “non-canonical”). So the domain
of expressions must itself be accorded with a notion of reduction to canonical form,
and this corresponds with putting into action a plan of verification in order to get
a direct (canonical) verification; reduction to canonical form will be represented by
a judgement M ⇓ M ′ , pronounced “M evaluates to M ′ ”.
To know M ⇓ M ′ is to know that M is an expression which
reduces to a canonical form M ′ .
An example of an evident reduction judgement in elementary mathematics
would be 3 + 4 ⇓ 7; note that 3 + 4 ⇓ 1 + 6 is, on the other hand, not evident,
since this judgement describes reduction to canonical form, whereas 1 + 6 is not a
canonical number.
Now, we can correct the previous meaning explanation as follows:
To know M ∈ P is to know an M ′ such that M ⇓ M ′ and M ′ is
a canonical (direct) verification of P .
If it is not yet clear why it would have been a mistake to fail to use the notion of
reduction to canonical form in the above meaning explanation, consider that each
time a proposition is defined, it should be possible to do so without knowing what
other propositions exist in the theory. But if we consider non-canonical forms (as
would be necessary if we omitted the M ⇓ M ′ premise), then we would have to fix
in advance all the possible non-canonical forms in the computation system in the
course of defining each proposition. As such, the open-ended nature of the logic
would be destroyed; in a later chapter, the seriousness of this problem will be made
even more clear.
The meaning explanation for P prop must be accordingly modified to take into
account the computational behavior of expressions:
10
1. LOGICAL THEORIES
To know P prop is to know a P ′ such that P ⇓ P ′ and P ′ is
a canonical proposition, which is to say, that one knows what
counts as a canonical verification for P ′ .
In practice, when it is clear that P is canonical, then we will simply say, “To
know P prop is to know what counts as a canonical verification of P ”. As an
example, then, we will update the evidence of the following assertion:
P ⊃ Q prop (P prop, Q prop (P true))
The meaning of this, expanded into spoken language, is as follows:
To know P ⊃ Q prop (P prop, Q prop (P true)) is to know what
counts as a canonical (direct) verification of P ⊃ Q under the
circumstances that P ⇓ P ′ , such that one knows what counts as a
canonical verification P ′ , and, if one has such a verification, Q ⇓
Q′ such that one knows what counts as a canonical verification
of Q′ .
And the above judgement is evident, since we will say that a canonical verification of P ⊃ Q is an expression λx.E such that we know the hypothetico-general
judgement |x E ∈ Q (x ∈ P ). This validates the assertion λx.E ∈ P ⊃ Q (|x E ∈
Q (x ∈ P )), or, written as an inference rule:
|x E ∈ Q (x ∈ P )
λx.E ∈ P ⊃ Q
By the addition of this judgement, we have graduated from a logical theory to a
type theory, in the sense of Constructive Mathematics and Computer Programming
(Martin-Löf, 1979). In fact, we may dispense with the original P true form of
judgement by defining it in terms of the new M ∈ P judgement as follows:
M ∈P
P true
CHAPTER 2
Computational Type Theories
As alluded to at the end of the previous chapter, we may add a judgement
M ∈ A which deals directly with the objects M which verify the propositions (or
types) P . We will develop a full theory of dependent types in the sense of MartinLöf.1
1. The categorical judgements
Because we will need to consider the introduction of types which do not have a
trivial (intensional) equality relation, we must first amend the meaning explanations
for some of our judgements, and add a few new forms of judgement. First, we will
refer to types rather than propositions in order to emphasize the generality of the
theory; in some presentations, the word set is used instead.
The meaning of hypothetical and general judgement are the same as in the previous chapter, and so we will not reproduce them here. The first form of judgement
is A type , and its meaning explanation is as follows:
To know A type is to know an A′ such that A ⇓ A′ , and you know
what counts as a canonical verification of A′ and when two such
verifications are equal.
The next form of judgement is M ∈ A , which remains the same as before:
To know M ∈ A (presupposing A type) is to know an M ′ such
that M ⇓ M ′ and M ′ is a canonical verification of A.
We’ll need to add new judgements for equality (equality of verifications, and
equality of types respectively). First, equality of verifications is written M = N ∈ A ,
and means the following:
To know M = N ∈ A (presupposing M ∈ A and N ∈ A, and
thence A type) is to know that the values of M and N are equal
as canonical verifications of A.
1Please note that the judgements given here, and their meaning explanations, are not the same as
those used in Constable et al’s “Computational Type Theory” and Nuprl. In this chapter, we use
the term “computational type theory” in a general sense to characterize a family of type theories
which have their origin in Martin-Löf’s 1979 paper Constructive Mathematics and Computer
Programming.
11
12
2. COMPUTATIONAL TYPE THEORIES
The fact that M and N reduce to canonical values which verify A is known
from the presuppositions of the judgement; and what it means for them to be equal
as such is known from the evidence of the presupposition A type which is obtained
from the other presuppositions.
For equality of types A = B type , there are a number of possible meaning
explanations, but we’ll use the one that Martin-Löf used starting in 1979:
To know A = B type (presupposing A type and B type) is to
know |M M ∈ B (M ∈ A) and |M M ∈ A (M ∈ B), and moreover |M,N M = N ∈ B (M = N ∈ A) and |M,N M = N ∈
A (M = N ∈ B).
In other words, two types are equal when they have the same canonical verifications, and moreover, the same equality relation over their canonical verifications.
Note that there are other possible explanations for type equality, including more
intensional ones that appeal to the syntactic structure of type expressions, and
these turn out to be more useful in proof assistants for practical reasons. However,
the extensional equality that we have expounded is the easiest and most obvious
one to formulate, so we will use it here.
2. The functional sequent judgement
Now, because we are allowing the definition of types with arbitrary equivalence
relations, we cannot use plain hypothetico-general judgement in the course of defining our types. For instance, if we were going to try and define the function type
A ⊃ B in the same way as we did in the previous chapter, we would permit “functions” which are not in fact functional, i.e. they do not take equal inputs to equal
outputs. As such, we will need to bake functionality (also called extensionality)
into the definition of functions, and since we will need this in many other places,
we elect to simplify our definitions by baking it into a single judgement which is
meant to be used instead of plain hypothetico-general judgement.
The judgement which expresses simultaneously generality, hypothesis and functionality has been written in multiple ways. Martin-Löf has always written it as
J (Γ), but this is a confusing notation because it appears as though it is merely a
hypothetical judgement (but it is much more, as will be seen). Very frequently, it is
written with a turnstile, Γ ⊢ J , and in the early literature surrounding Constable’s
Computational Type Theory and Nuprl, it was written Γ ≫ J ; we choose this last
option to avoid confusion with a similar judgement form which appears in prooftheoretic, intensional type theories; we’ll call the judgement form a “(functional)
sequent”.
2. THE FUNCTIONAL SEQUENT JUDGEMENT
13
We will define several forms of judgment simultaneously with the sequent judgments: Γ ctx (“Γ is a context”), x # Γ (“x is fresh in Γ”).
To know Γ ctx is to know that Γ ≡ ·, or it is to know a variable
x and expressions ∆, A such that Γ ≡ ∆, x : A and ∆ ≫ A type,
and x # ∆.
To know x # Γ (presupposing Γ ctx ) is to know that Γ ≡ ·, or,
if Γ ≡ ∆, y : A such that x is not y and x # ∆.
In other words, the well-formed contexts are inductively generated by the following grammar:
Γ ctx
· ctx
x#·
Γ ≫ A type
Γ, x : A ctx
x#Γ
x#Γ
x # Γ, y : A
Next, we define the notion of an environment or substitution for a context,
ρ ∈⋆ Γ . It is possible to define this more generally as functional maps from one
context to another, but for our purposes, this is the only case we need.
To know ρ ∈⋆ Γ, presupposing Γ ctx , is to know, if Γ ≡ ·, then
ρ ≡ ·; otherwise, if Γ ≡ ∆, x : A then ρ ≡ ρ′ , M such that ρ′ ∈⋆ ∆
and [ρ′ /∆]M ∈ [ρ′ /∆]A. The substitution [ρ/Γ] is effected by
replacing each variable in Γ with the corresponding term in ρ.
We will also specify when two environments are equal, ρ0 = ρ1 ∈⋆ Γ :
To know ρ0 = ρ1 ∈⋆ Γ, presupposing Γ ctx , ρ0 ∈⋆ Γ and ρ1 ∈⋆ Γ
is to know, if Γ ≡ ·, then ρ0 ≡ · and ρ1 ≡ ·; otherwise, if
Γ ≡ ∆, x : A and thence ρ0 ≡ ρ′0 , M0 and ρ1 ≡ ρ′1 , M1 , then
ρ′0 = ρ′1 ∈⋆ ∆ and [ρ0 /∆]M0 = [ρ1 /∆]M1 ∈ [ρ0 /∆]A.
We will say that Γ ≫ J is only a judgement under the presuppositions that
Γ ctx and that J is a categorical judgement of the form A type, A = B type,
M ∈ A, or M = N ∈ A. Now we may begin giving the meaning explanations for
Γ ≫ J , starting with typehood with respect to a context:
To know Γ ≫ A type (presupposing Γ ctx ) to know |ρ [ρ/Γ]A type (ρ ∈⋆
Γ) and moreover, to know |ρ0 ,ρ1 [ρ0 /Γ]A = [ρ1 /Γ]A type (ρ0 =
ρ1 ∈⋆ Γ).
We can explain type equality sequents in a similar way:
To know Γ ≫ A = B type (presupposing Γ ctx , Γ ≫ A type,
Γ ≫ B type) is to know
|ρ [ρ/Γ]A = [ρ/Γ]B type (ρ ∈⋆ Γ)
14
2. COMPUTATIONAL TYPE THEORIES
and moreover, to know
|ρ0 ,ρ1 [ρ0 /Γ]A = [ρ1 /Γ]B type (ρ0 = ρ1 ∈⋆ Γ)
Next, the meaning of membership sequents is explained:
To know Γ ≫ M ∈ A (presupposing Γ ctx , Γ ≫ A type) is to
know |ρ [ρ/Γ]M ∈ [ρ/Γ]A (ρ ∈⋆ Γ) and moreover, to know
|ρ0 ,ρ1 [ρ0 /Γ]M = [ρ1 /Γ]M ∈ [ρ0 /Γ]A (ρ0 = ρ1 ∈⋆ Γ)
Finally, member equality sequents have an analogous explanation:
To know Γ ≫ M = M ′ ∈ A (presupposing Γ ctx , Γ ≫ A type)
is to know |ρ [ρ/Γ]M = [ρ/Γ]N ∈ [ρ/Γ]A (ρ ∈⋆ Γ) and moreover,
to know
|ρ0 ,ρ1 [ρ0 /Γ]M = [ρ1 /Γ]N ∈ [ρ0 /Γ]A (ρ0 = ρ1 ∈⋆ Γ)
The simultaneous definition of multiple judgements may seem at first concerning, but it can be shown to be non-circular by induction on the length of the context
Γ.
3. The definitions of types
We will now define the types of a simple computational type theory without
universes. In the course of doing so, opportunities will arise for further clarifying
the position of the judgements, meaning explanations and proofs on the one hand,
and the propositions, definitions and verifications on the other hand.
3.1. The unit type. First, we introduce two canonical forms with trivial
reduction rules:
(Canonical)
unit ⇓ unit
•⇓•
Next, we intend to make the judgement unit type evident; and this is done
by defining what counts as a canonical verification of unit and when two such
verifications are equal. To this end, we say that • is a canonical verification of unit,
and that it is equal to itself. I wish to emphasize that this is the entire definition
of the type: we have introduced syntax, and we have defined the canonical forms,
and there is nothing more to be done.
In the presentations of type theory which are currently in vogue, a type is “defined” by writing out a bunch of inference rules, but in type theory, the definitions
that we have given above are prior to the rules, which are justified in respect of the
definitions and the meaning explanations of the judgements. For instance, based on
the meaning of the various forms of sequent judgement, the following rule schemes
3. THE DEFINITIONS OF TYPES
15
are justified:
Γ ≫ unit type
Γ ≫ unit = unit type
Γ ≫ • ∈ unit
Γ ≫ • = • ∈ unit
Each of the assertions above has evidence of a certain kind; since the justification of these rules with respect to the definitions of the logical constants and the
meaning explanations of the judgements is largely self-evident, we omit it in nearly
all cases. It is just important to remember that it is not the rules which define the
types; a type A is defined in the course of causing the judgement A type to become
evident. These rules merely codify standard patterns of use, nothing more, and
they must each be justified.
3.2. The empty type. The empty type is similarly easy to define. First, we
introduce a constant:
(Canonical)
void ⇓ void
To make the judgement void type evident, we will say that there are no canonical
verifications of void, and be done with it. This definition validates some further rules
schemes:
Γ ≫ void type
Γ ≫ void = void type
Γ ≫ M ∈ void
Γ≫J
The last rule simply says that if we have a verification of void, then we may
conclude any judgement whatsoever. Remember that the inference rules are just
notation for an evident hypothetical judgement, e.g. Γ ≫ J (Γ ≫ M ∈ void).
Note that we did not introduce any special constant into the computation system to represent the elimination of a verification of void (in proof-theoretic type
theories, this non-canonical form is usually called abort(R)). This is because, computationally speaking, there is never any chance that we should ever have use for
such a term, since we need only consider the evaluation of closed terms (which
is guaranteed by the meaning explanations), and by its very definition, there can
never be a closed verification of void.
3.3. The cartesian product of a family of types. This will be our first
dependent type, and it will likewise be the first example of a non-trivial addition to
the computation system. First, let us add our canonical and non-canonical forms
16
2. COMPUTATIONAL TYPE THEORIES
and their reduction rules:
(Canonical)
Π(A; x.B) ⇓ Π(A; x.B)
λ(x.E) ⇓ λ(x.E)
M ⇓ λ(x.E) [N/x]E ⇓ M ′
ap(M ; N ) ⇓ M ′
(Non-canonical)
This is the first time in this chapter that we have introduced a term former with
binding structure; it should be noted that the intensional equality of expressions is
up to alpha equivalence, and we will not pay attention to issues of variable renaming
in our presentation.
We will make evident the following judgement scheme:
A type x : A ≫ B type
Π(A; x.B) type
Or, written as a hypothetical judgement:
Π(A; x.B) type (A type, x : A ≫ B type)
This is to say, under the stated assumptions, we know what counts as a
canonical verification of Π(A; x.B) and when two such verifications are equal. We
will say that λ(x.E) is a canonical verification of Π(A; x.B) just when we know
x : A ≫ E ∈ B; moreover, that two verifications λ(x.E) and λ(y.E ′ ) are equal just
when z ∈ A ≫ [z/x]E = [z/y]E ′ ∈ [z/x]B.
By the meaning of the sequent judgement, this is to say that a lambda expression must be functional with respect to its domain (i.e. it must take equals
to equals). We did not need to hypothesize directly two elements of the domain
and their equality because this is part of the meaning explanation for the sequent
judgement already. Likewise, two lambda expressions are equal when equal inputs
yield equal results in both.
The familiar inference rules, which codify the standard mode of use for the
family cartesian product, are justified by this definition:
Γ ≫ A type Γ, x : A ≫ B type
Γ ≫ Π(A; x.B) type
Γ, x : A ≫ E ∈ B
Γ ≫ λ(x.E) ∈ Π(A; x.B)
Γ ≫ M ∈ Π(A; x.B) Γ ≫ N ∈ A
Γ ≫ ap(M ; N ) ∈ [N/x]B
Γ ≫ A = A′ type Γ, z : A ≫ [z/x]B = [z/y]B ′ type
Γ ≫ Π(A; x.B) = Π(A′ ; y.B ′ ) type
Γ, z : A ≫ [z/x]E = [z/y]E ′ ∈ [z/x]B
Γ ≫ λ(x.E) = λ(y.E ′ ) ∈ Π(A; x.B)
Γ ≫ M = M ′ ∈ Π(A; x.B) Γ ≫ N = N ′ ∈ A′
Γ ≫ ap(M ; N ) = ap(M ′ ; N ′ ) ∈ [N/x]B
Note that the type equality rule scheme that we gave above is structural; it is
validated by the meaning explanations, but it is by no means the full totality of
3. THE DEFINITIONS OF TYPES
17
possible equalities between family cartesian product types, which is extensional in
this theory.
It will be instructive to explicitly justify the application rule above with respect
to the meaning explanations, since I have claimed that such rules are posterior to
the definitions we expounded prior to giving these rules.
Proof. It suffices to consider the case that Γ ⇓ ·, because hypotheses may
always be added to the context (this is called weakening). And so, by the meaning
of the sequent judgement at the empty context, the rule amounts to the assertion
ap(M ; N ) ∈ [N/x]B (M ∈ Π(A; x.B), N ∈ A)
By the meaning explanation for hypothetical judgement, and the definition of
the family cartesian product type, we know that M ⇓ λ(x.E) for some E such that
we know x : A ≫ E ∈ B; from the meaning of the sequent judgement and the notion
of an environment for a context, we can conclude |L [L/x]E ∈ [L/x]B (L ∈ A). On
the other hand, from the computation rules, we know that if for some particular
E ′ , [N/x]E ⇓ E ′ , then we know ap(λ(x.E); N ) ⇓ E ′ ; to demonstrate the evidence
of the premise, we may instantiate L at N to know [N/x]E ∈ [N/x]B, whence by
the meaning of membership, we know that there exists some canonical E ′ such that
[N/x] ⇓ E ′ .
CHAPTER 3
Proof Theories and Logical Frameworks
1. Proof-theoretic type theory
1.1. Analytic and synthetic judgement. A synthetic judgement is one for
which the experience of coming to know it necessarily entails some knowledge which
is not implicit in the statement of the judgement; on the other hand, to know an
analytic judgement is to know it purely on the basis of the information contained
inside it. So analytic judgements are decidable, since if they may become evident,
it will be purely on the basis of their own content; whereas synthetic judgements
become evident to someone when they have obtained some particular evidence for
them.
A logical theory has, then, both analytic and synthetic judgements; the judgement P prop is analytic, since its evidence follows from the definition of P , whereas
the assertion of P true entails the knowledge of some extra information, namely a
verification of P . When we have extended the logical theory to a type theory in
the manner of the previous chapter, the judgement M ∈ P is also synthetic, since
M ∈ P is not self-evident in general.
But why is it not enough to assert that M verifies P to know whether M ∈ P ?
It suffices to define a P such that one cannot decide in general whether some term
is a verification of it. Let us define the propositional symbol P, and we intend to
know the judgement P prop, whose meaning is to be expanded as follows:
To know P prop is to know counts as a canonical verification of
P.
We will say, then, that • is a canonical verification of P just when Goldbach’s
conjecture is true. Then it comes immediately that the judgement M ∈ P may not
be known or refuted on its own basis, nor even the judgement • ∈ P, since they
depend on a proposition whose truth is not known:
To know M ∈ P is to know that M ⇓ M ′ to a canonical verification of P .
To know M ∈ P is to know that M ⇓ • such that Goldbach’s conjecture
is true.
1.2. Proof of a judgement vs. verification of a proposition. Because
the judgement M ∈ P is synthetic, we cannot say that it gives rise to a proof
19
20
3. PROOF THEORIES AND LOGICAL FRAMEWORKS
theory for the logic, since the core judgement of a proof theory M : A must be
analytic, in order to avoid the infinite regress of a proof theory requiring a proof
theory requiring a proof theory, and so on.
The notion of verification of a proposition could never be the same as proof
anyway, except in the most trivial circumstances, since a verification is meant to be
an effective operation which realizes the truth of a proposition, and no constraints
whatsoever (termination, totality, etc.) are placed on these operations except those
which come from the meaning of the judgements (see [4], [16], [17]).
So a proof theory is necessarily intensional, and its judgements are to be analytic/decidable. What is it, then, that we have considered so far which corresponds
with a proof M such that M : P in a proof theory? As discussed above, M is not
merely a term such that M ∈ P , since this is not in general enough information
to know whether M is a proof. In fact, M must comprise all the logical inferences
which led to the knowledge that P is true, and so a meaning explanation for the
judgement M : P in a proof theory immediately suggests itself:
To know M : P is to know that M is evidence (demonstration,
proof, derivation) of the judgement P true.
And so the term domain of the proof theory is not the same as the one that
we have considered so far; it must consist in terms which represent traces of the
inferences made in the course of knowing the judgements of a logical theory. There
is a sense in which one can consider the types of a proof theory to interpret the
judgements of the logical theory, and this methodology is called “judgements as
types” (and this implies “derivations as terms”).
What I am calling a “proof-theoretic type theory” is a type theory of the
sort used in the proof assistants Agda, Coq and Idris, whereas the kind of type
theory that I have described in the previous sections, the one based on meaning
explanations, underlies the proof assistant Nuprl.
The proof-theoretic type theories on the one hand are often called “intensional”
and the computational type theories on the other hand are usually “extensional”;
these characterizations are certainly true, though they are not essential ; moreover,
I fear that comparing one of the former with one of the latter is not quite fair, since
there is not any clear analogy to be had. That is to say, the judgement M ∈ P
is a judgement which is added to a logical theory and its meaning is (briefly) “M
evaluates to a canonical verification of P ”, whereas M : P cannot be construed as
a judgement added to a logical theory. Instead, it must be understood as part of
a (proof) theory which is overlayed atop an existing logical theory; it is possible to
understand the theory which contains the judgement M : P to be a metatheory, or
2. MARTIN-LÖF’S EQUATIONAL LOGICAL FRAMEWORK
21
logical framework, for the theory which contains the judgement P true, which can
be construed as the “object language”.
In short, the judgements M ∈ P and M : P are unrelated to each other in
two respects: firstly, that they have different meanings, and secondly that the one
is at the same level as the judgements of a logical theory, whereas the latter is a
judgement in a theory which is defined over a logical theory.
2. Martin-Löf ’s equational logical framework
To make this more concrete, let us expound a proof theoretic type theory called
MLLF, which stands for “Martin-Löf’s (equational) logical framework”;1 in the
course of introducing each type, we will specify which judgement of the underlying
logical theory it is meant to interpret.
We start with four categorical judgements:
Judgement Form Pronunciation
α : type
α is a type
α = β : type
M :α
α and β are equal types
M is of type α
M =N :α
M and N are equal at type α
But we have not defined the meaning of the judgements; let us do so below:
To know α : type is to know what counts as an object of type α,
and when two such objects are equal.
For now, we’ll leave the question of what is an “object” as abstract; in many
cases, types will represent judgements of a logical theory, and the objects will be
the derivations (demonstrations, proofs) of those judgements.
To know α = β : type is to know that any object of type α is
also an object of type β, and two equal objects of type α are
equal as objects of type β (necessarily presupposing α : type and
β : type).
To know M : α is to know that M is an object of type α (necessarily presupposing α : type).
To know M = N : α is to know that M and N are equal objects
of type α (necessarily presupposing M : α and N : α).
In addition to the above judgements, we will need contexts (with their wellformedness judgement Γ ctx ) and an intensional sequent judgement Γ ⊢ J ; their
1For a detailed overview of Martin-Löf’s equational logical framework, see [15].
22
3. PROOF THEORIES AND LOGICAL FRAMEWORKS
meanings here will differ from the sequent judgements of computational type theory, in that they must mean proof-theoretic derivability, rather than semantic consequence.
At this point, we may begin adding types to the logical framework. In practice,
most types which we will introduce in the logical framework will be defined in terms
of a judgement of the logical theory which lies below it. For instance, hypothetical
judgement in the logical theory is represented by a function type in the logical
framework, (x : α)β, whose typehood is meant to be evident under the following
circumstances
α : type x : α ⊢ β : type
(x : α)β : type
Or as a hypothetical judgement, (x : α)β : type (α : type, x : α ⊢ β : type).
Now, to know this judgement is to know that under the circumstances we know
what is an object of type α and when two such objects are equal, and that if we
have such an object x of type α, we know what an object of type β is, and when
two such objects are equal—then we know what an object of type (x : α)β is, and
moreover, for any two objects y, z of type α, that [y/x]β and [z/x]β are equal as
types. To make this evident, then, we will say that under those circumstances an
object of type (x : α)β is an object [x]M such that one knows x : α ⊢ M : β and
|x,y [y/x]M = [z/x]M : [y/x]β (y = z : α); furthermore, two such objects are equal
just when they yield equal outputs for equal inputs.
Then, for each atomic proposition P , we can easily define a type Prf(P ), as
follows. Under the circumstances that P prop in the logical theory, then Prf(P ) :
type in the logical framework, since we will define an object of type Prf(P ) to be a
derivation of P true; beyond reflexivity, further definitional equalities can be added
to reflect the harmony of introduction and elimination rules.
Now, the definitions we have given for the types above are “intuitively” correct,
but they actually fail to satisfy the meaning explanation that we have given for
α : type, because they do not take into account neutral terms. In the following
sections, we will investigate this problem in more detail and propose a solution.
2.1. What is an “object”? It is time to revisit what it means to be an
“object” of a type in the proof-theoretic type theory; we must note how this will
necessarily differ from what it meant to be a “verification” of a proposition in
the previous sections. Namely, a verification of a proposition is either a canonical
verification of that proposition (and what sort of thing this might be is known from
the presupposition P prop), or it is a means of getting such a canonical verification
(i.e. a term which evaluates to a canonical verification).
On the other hand, what we have called an “object” of type P is quite different,
since in addition to the possibility that it is a canonical proof of the judgement
3. A CRITIQUE OF MLLF
23
P true, it may also be neutral (i.e. blocked by a variable); we will call this “normal”
rather than “canonical”. Why does this happen?
In order to keep the judgement M : A analytic (decidable), its meaning explanation can no longer be based on the idea of the computation of closed terms to
canonical form; instead, we will consider the computation of open terms (i.e. terms
with free variables) to normal form. The desire for M : A to be analytic follows
from our intention that it characterize a proof theory: we must be able to recognize
a proof when we see one. But why are closed-term-based meaning explanations
incompatible with this goal? Consider briefly the following judgement:
|n M (n) ∈ P (n ∈ N)
To know this judgement is to know that M (n) computes to a canonical verification of P whenever n is a natural number; when P ’s use of n is not trivial,
this amounts to testing an infinite domain (all of the natural numbers), probably
by means of mathematical induction. The judgement is then clearly synthetic:
to know it is, briefly, to have come up with an (inductive) argument that M (N )
computes to a canonical verification of P at each natural number n.
On the other hand, the judgement n : N ⊢ M (n) : P must have a different
meaning, one which admits its evidence or refutation purely on syntactic/analytic
grounds. In essence, it is to know that M (n) is a proof of P for any arbitrary
object/expression n such that n : N (i.e., the only thing we know about n is that it
is of type N; we do not necessarily know that it is a numeral).
3. A critique of MLLF
The type theory which we constructed in the previous section is to be considered
a proof theory for a logic with the judgements P prop, P true and J (J ′ ). There
are a few reasons to be dissatisfied with this state of affairs, which I shall enumerate
in this section.
3.1. Lack of computational content. Unlike the type theory in the first
chapter, there is no built-in computational content. In a computational type theory
which is defined by the verificationist meaning explanations, the computational
content of terms is understood immediately by means of the M ⇓ M ′ relation; that
is, computation is prior to the main judgements because their meaning explanations
are defined in terms of evaluation to canonical form.
On the other hand, in the type theory above we did not give a primitive reduction relation; instead, we simply permitted the endowement of proofs with definitional equalities which reflect the harmony of introduction and elimination rules.
That is, if we have known the judgement P true by means of an indirect argument (derivation), it must be the case that this derivation corresponds to a direct
24
3. PROOF THEORIES AND LOGICAL FRAMEWORKS
one; we reflect this in the proof theory by defining the indirect derivation to be
definitionally equal to the direct one.
However, this does not amount to computational content being present in terms:
only post facto may the definitional equality be construed as giving rise to computation, through a metamathematical argument which shows that the definitional
equality is confluent and can be used to define a functional normalization relation.
And this is the reason for the peculiarity of the proof-theoretic meaning explanations, namely that they do not include phrases like “evaluates to a canonical...”,
since evaluation may only be understood after taking the meanings of the judgements (α : type, α = β : type, M : α, M = N : α) as giving rise to a closed formal
system which is susceptible to metamathematical argument: to refer to evaluation
in the meaning explanations for the core judgements, then, would be impredicative.
3.2. Modularity of definition. By the same token, the distinction between
canonical (direct) and non-canonical (indirect) proof may not be understood as a
core notion in the theory, but must be understood separately, secondarily. Why is
this a problem? It means that the definition of each type must be made with the
full knowledge of the definitions of every other type; in essence, the open-ended
nature of type theory is obliterated and one is forced into a fixed formal system;
this is in addition to the fact that it causes the epistemic content of α : type for
any type α to be extremely complicated.
To illustrate, let us consider as an example a type theory which has four typeformers: trivial truth ⊤, trivial falsity ⊥, implication (α)β, and conjunction α&β;
we will then introduce the following terms to represent proofs: the trivial element •,
reductio ad absurdum abort(α; E), abstraction [x : α]E, application E(E ′ ), pairing
hE, E ′ i, and projections fst(E), snd(E).
If we will try to make the judgement ⊤ : type evident, the deficiencies of the
formulation will immediately present themselves.
To know ⊤ : type is to know what counts as an object of type
⊤, and when two such objects are equal. An object of type ⊤,
then, is either the expression •, or an expression abort(⊤; E) such
that we know E : ⊥, or an expression E(E ′ ) such that we know
E : (α)⊤ and E ′ : α, or an expression fst(E) such that we know
E : ⊤&β for some β, or an expression snd(E) such that we know
E : α&⊤ for some α; and we additionally have that • is equal to
•, and ...
To save space, we elide the rest of the definition of equality for ⊤; what we have
seen so far already suffices to bring to light a serious problem: the definition of any
type requires knowledge of the entire syntax of the theory. The judgement α : type
4. A WAY FORWARD: VERIFICATIONS & USES
25
may never be made evident in isolation, but must be done with full understanding
of all the other types and their definitions.
Furthermore, to extend an existing theory with a new type, the definitions of
every other type must be rewritten to account for the elimination forms of the new
type.
4. A way forward: verifications & uses
The second critique of MLLF may be partially addressed by fragmenting type
theory into a logic of verifications & uses: instead of a type being defined by its
introduction rules, it must be simultaneously defined by its introduction rules (verifications) and its elimination rules (uses). In practice, this amounts to a standard
technique known as bidirectional type checking.
The semantic priority of the forms of judgement also changes drastically: the
sequent judgement must in this case be explained before the categorical judgements;
moreover, sequents may no longer be explained modularly in terms of general and
hypothetical judgement, since the latter amounts to semantic consequence (admissibility), whereas the meaning of a sequent in a proof theory should be syntactic
consequence (derivability).
Because the target theory lacks computation, it is necessary to rule out redexes
from terms syntactically, but this complicates the definition of substitution; to
address this, Watkins introduced in [18] a technique known as hereditary substitution, which is a family of syntax-directed (algorithmic) judgements which contract
redexes along the way, guaranteeing canonical form in their outputs. Both bidirectional type checking and hereditary substitutions have been used to great effect in
the descendants of the Edinburgh Logical Framework (see [10]).
The first critique, the lack of computational content, is more difficult to address.
Roughly, the right way to do it is to replace the notion of the evaluation of closed
terms to canonical form with the evaluation of open terms to normal form. Peter
Dybjer demonstrates in [5] how this technique may be used to endow the Calculus
of Constructions with a meaning explanation, albeit necessarily of a very different
kind than we have considered here.
Bibliography
[1] Intuitionistic Type Theory. Bibliopolis, 1984.
[2] Stuart F. Allen, Mark Bickford, Robert L. Constable, Richard Eaton, Christoph Kreitz, Lori
Lorigo, and E. Moran. Innovations in computational type theory using nuprl. J. Applied
Logic, 4(4):428–469, 2006.
[3] Robert L. Constable, Stuart F. Allen, Mark Bromley, Rance Cleaveland, J. F. Cremer, R. W.
Harper, Douglas J. Howe, Todd B. Knoblock, N. P. Mendler, Prakash Panangaden, James T.
Sasaki, and Scott F. Smith. Implementing mathematics with the Nuprl proof development
system. Prentice Hall, 1986.
[4] Michael Dummett. Elements of intuitionism, volume 39 of Oxford Logic Guides. The Clarendon Press Oxford University Press, New York, second edition, 2000.
[5] Peter Dybjer. Program testing and the meaning explanations of intuitionistic type theory. In
Peter Dybjer, Sten Lindström, Erik Palmgren, and Göran Sundholm, editors, Epistemology
versus Ontology, volume 27 of Logic, Epistemology, and the Unity of Science, pages 215–241.
Springer, 2012.
[6] Johan Georg Granström. Treatise on Intuitionistic Type Theory. Springer Publishing Company, Incorporated, 2013.
[7] Robert Harper. Constructing type systems over an operational semantics. J. Symb. Comput.,
14(1):71–84, 1992.
[8] Robert Harper. Practical Foundations for Programming Languages. Cambridge University
Press, New York, NY, USA, 2012.
[9] Robert Harper, Furio Honsell, and Gordon Plotkin. A framework for defining logics. J. ACM,
40(1):143–184, January 1993.
[10] Robert Harper and Daniel R. Licata. Mechanizing metatheory in a logical framework. J.
Funct. Program., 17(4-5):613–673, July 2007.
[11] Per Martin-Löf. Constructive mathematics and computer programming. In L. Jonathan Cohen, Jerzy Loś, Helmut Pfeiffer, and Klaus-Peter Podewski, editors, Logic, Methodology and
Philosophy of Science VI, Proceedings of the Sixth International Congress of Logic, Methodology and Philosophy of Science, Hannover 1979, volume 104 of Studies in Logic and the
Foundations of Mathematics, pages 153–175. North-Holland, 1982.
[12] Per Martin-Löf. Truth of a proposition, evidence of a judgement, validity of a proof. Synthese,
73(3):407–420, 1987.
[13] Per Martin-Löf. Analytic and synthetic judgements in type theory. In Paolo Parrini, editor,
Kant and Contemporary Epistemology, volume 54 of The University of Western Ontario
Series in Philosophy of Science, pages 87–99. Springer Netherlands, 1994.
[14] Per Martin-Löf. On the meanings of the logical constants and the justifications of the logical
laws. Nordic Journal of Philosophical Logic, 1(1):11–60, 1996.
[15] B. Nordström, K. Petersson, and J.M. Smith. Programming in Martin Löf ’s Type Theory,
volume 7 of Monographs on Computer Science. Oxford University Press, 1990.
27
28
BIBLIOGRAPHY
[16] Dag Prawitz. Truth and proof in intuitionism. In Peter Dybjer, Sten Lindström, Erik Palmgren, and Göran Sundholm, editors, Epistemology versus Ontology, volume 27 of Logic, Epistemology, and the Unity of Science, pages 45–67. Springer, 2012.
[17] A.S. Troelstra and D. van Dalen. Constructivism in Mathematics. North-Holland, 1988. 2
volumes.
[18] Kevin Watkins, Iliano Cervesato, Frank Pfenning, and David Walker. A Concurrent Logical
Framework: The Propositional Fragment. Types for Proofs and Programs, pages 355–377,
2004.
| 2cs.AI
|
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
1
Distributed Optimal Frequency Control Considering
a Nonlinear Network-Preserving Model
arXiv:1709.01543v3 [cs.SY] 13 Feb 2018
Zhaojian Wang, Feng Liu, John Z. F. Pang, Steven Low, Fellow, IEEE and Shengwei Mei, Fellow, IEEE
Abstract—This paper addresses the distributed optimal frequency control of power systems considering a networkpreserving model with nonlinear power flows and excitation
voltage dynamics. Salient features of the proposed distributed
control strategy are fourfold: i) nonlinearity is considered to
cope with large disturbances; ii) only a part of generators
are controllable; iii) no load measurement is required; iv)
communication connectivity is required only for the controllable
generators. To this end, benefiting from the concept of “virtual
load demand”, we first design the distributed controller for the
controllable generators by leveraging the primal-dual decomposition technique. We then propose a method to estimate the
virtual load demand of each controllable generator based on local
frequencies. We derive incremental passivity conditions for the
uncontrollable generators. Finally, we prove that the closed-loop
system is asymptotically stable and its equilibrium attains the
optimal solution to the associated economic dispatch problem.
Simulations, including small and large-disturbance scenarios,
are carried on the New England system, demonstrating the
effectiveness of our design.
Index Terms—Frequency control, network-preserving model,
distributed control, incremental output passivity
I. I NTRODUCTION
Frequency restoration and economic dispatch (ED) are two
important problems in power system operation. Conventionally, they are implemented hierarchically in a centralized fashion, where the former is addressed in a fast time scale while
the latter in a slow time scale [1], [2]. While this centralized
hierarchy works well for the traditional power system, it
may not be able to keep pace with the fast development of
our power system due to: 1) slow response, 2) insufficient
flexibility, 3) low privacy, 4) intense communication, and 5)
single point of failure issue. In this regard, the idea of breaking
such a hierarchy is proposed in [3], [4].
In [5], an intrinsic connection between the asymptotic
stability of the dynamical frequency control system with the
ED problem is proposed. It leads to a so-called inverseengineering methodology for designing optimal frequency
controllers, where the (partial) primal-dual gradient algorithm
This work was supported by the National Natural Science Foundation of
China (No. 51677100, U1766206, No. 51621065), the US National Science
Foundation through awards EPCN 1619352, CCF 1637598, CNS 1545096,
ARPA-E award de-ar0000699, and Skoltech through Collaboration Agreement
1075-MRA. (Corresponding author: Shengwei Mei)
Z. Wang, F. Liu and S. Mei are with the State Key Laboratory of
Power System and Department of Electrical Engineering, Tsinghua University,
Beijing, China, 100084. e-mail: ([email protected]).
S. H. Low and J. Pang are with the Department of Electrical Engineering, California Institute of Technology, Pasadena, CA, USA, 91105. email:([email protected]).
plays an essential role [1], [6]. When designing optimal frequency controllers, in choice of power flow models, including
the linear model (usually associated with DC power flow, e.g.
[5], [7]–[15]) and the nonlinear model (usually associated with
AC power flow, e.g. [16]–[21]), is crucial. While the closedloop system can be interpreted in a linear model as carrying
out a primal-dual algorithm for solving ED, this interpretation
of frequency control may not hold in a nonlinear model. In
addition to nonlinear power flow, excitation voltage dynamics
are considered in [16]–[18], making the model more realistic.
The aforementioned idea is further developed to enable the
design of distributed optimal frequency controllers. Roughly
speaking, the works of distributed optimal frequency control
can be divided into two categories in terms of different power
system models: network-reduced models e.g. [5], [12]–[16],
[18] and network-preserving models e.g. [7]–[11], [17]. In
network-reduced models, generators and/or loads are aggregated and treated as one bus or control area, which are connected to each other through tie lines. In [5], [13], aggregated
generators in each area are driven by automatic generation
control (AGC) to restore system frequency. [12], [14]–[16],
[18] further consider both the aggregated generators and load
demands in frequency control. In network-preserving models,
generator and load buses are separately handled with different
dynamic models and coupled by power flows, rendering a
set of differential algebraic equations (DAEs). In [7], an
optimal load control (OLC) problem is formulated and a
primary load-side control is derived as a partial primal-dual
gradient algorithm for solving the OLC problem. The design
approach is extended to secondary frequency control (SFC)
that restores nominal frequency in [8]. It is generalized in [9],
where passivity condition guaranteeing stability is proposed
for each local bus. Then, a unified framework combining
load and generator control is advocated in [10]. A similar
model is also utilized in [11], where only limited control
coverage is needed. Similar to [18], the Hamiltonian method
is used to analyze the network-preserving model in [17].
Compared with the network-reduced model, the networkpreserving model describes power systems more precisely
and appear more suitable for analyzing interactions between
different control areas. Therefore, we specifically consider the
network-preserving model in this work.
Almost all of aforementioned works assume that all buses
are controllable and load demands at all buses are accurately measurable, especially for those proposing secondary
frequency control. Moreover, it is usually assumed that the
communication network has the same topology of the power
grid. These assumptions are strong and arguably unrealistic
di d̂i
Pei
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
II. P OWER S YSTEM M ODELS
Pi
g
i
i , i , Eqi'
Pei , Qei
p j 1 q j
i , i ,Vi
E fi
pi 1 qi
Fig. 1. Summary of notations
buses and pure load buses. Denote controllable generator buses
as NCG = {1, 2, . . . , nCG }, uncontrollable generator buses as
NU G = {nCG + 1, nCG + 2, . . . , nCG + nU G }, and pure load
buses as NL = {nCG +nU G +1, . . . , nCG +nU G +nL }. Then
the set of generator buses is NG = NCG ∪ NU G and set of
all the buses is N = NG ∪ NL . It should be noted that load
can be connected to any bus besides pure load buses.
Let E ⊆ N × N be the set of lines, where (i, j) ∈ E if
buses i and j are connected directly. Then the whole system is
modeled as a connected graph
G = (N , E). The admittance of
√
each line is Yij := Gij + −1Bij with Gij = 0 for every line.
Denote the bus voltage by Vi ∠θi , where Vi is the amplitude
and θi is the voltage phase angle. The active and reactive
power Pij , Qij from bus i to bus j is
Pij = Vi Vj Bij sin (θi − θj )
Qij =
Bij Vi2
(1a)
− Vi Vj Bij cos (θi − θj )
(1b)
For convenience, most notations are summarized in Fig.1.
B. Synchronous Generators
For i ∈ NG , we use the standard third-order generator
model (e.g. [18], [22], [23]) (2a)-(2c). Here (2d) is the simplified governor-turbine model, and (2e) is the excitation voltage
control model:
δ̇i = ωi
ω̇i =
(Pig
(2a)
− Di ωi − Pei )/Mi
0
0
(2b)
0
Ėqi = −Eqi /Td0i + Ef i /Td0i
(2c)
Ṗig
(2d)
=
−Pig /Ti
+
ugi
Ėf i = h(Ef i , Eqi )
(2e)
In this model, Mi is the moment of inertia; Di the damping
0
constant; Td0i the d-axis transient time constant; Ti the time
constant of turbine; δi the power angle of generator i; ωi the
generator frequency deviation compared to an steady state
value; Pig the mechanical power input; pj the active load
0
demand; Pei the active power injected to network; Eqi the qaxis transient internal voltage; Eqi the q-axis internal voltage;
Ef i the excitation voltage. Eqi is given by
A. Power Network
Power systems are composed of many generators and loads,
which are integrated in different buses and connected by power
lines, forming a power network. Buses are divided into three
types, controllable generator buses, uncontrollable generator
j , j ,V j
j
Pij , Qij
…
for practice. First, only a part of generators and controllable
load buses can participate in frequency control in practice.
Second, the communication network may not be identical
to the topology of the power grid. Third, it is difficult to
accurately measure the load injection on every bus. In extreme
cases, even the number of load buses is unknown in practice.
These problems highlight some reasons why theoretical work
in this area is hardly applied in practical systems.
In this context, a novel distributed frequency recovery
controller is proposed that only needs to be implemented on
controllable generator buses. To this end, a network-preserving
power system model is adopted. This work is an extension of
our former work [11]. However, in this paper, we design a
totally different controller considering a third-order nonlinear
generator model with excitation voltage dynamics and nonlinear power flow. It is also motivated partly by [17], which
adopts a similar model, although our results are significantly
different from those in [17]. Differing from [11], the controller
avoids load measurement, which greatly facilitates implementation. By using LaSalle’s invariance principle, it is proved that
the closed-loop system converges to an equilibrium point that
solves the economic dispatch problem. The salient features of
the proposed distributed optimal frequency controller are:
1) Model: The network-preserving model of power system
is used, including excitation voltage dynamics and nonlinear power flow. This, unlike work on the linear model,
returns a valid controller even under large disturbances.
2) Controllability: We allow an arbitrary subset of generator buses to be controllable.
3) Controller: The distributed controller achieves the optimal solution to economic dispatch while restoring
the nominal frequency, provided that certain sufficient
conditions on active power dynamics of uncontrollable
generators and excitation voltage dynamics of all generators are satisfied.
4) Communication: Communication is required between
neighboring controllable generators only, and the communication network can be arbitrary as long as it remains connected.
5) Measurement: No load measurement is needed, and the
controller is adaptive to unknown load changes.
The rest of this paper is organized as follows. In Section
II, we introduce the power system model. Section III formulates the optimal economic dispatch problem. The distributed
controller is proposed in Section IV, and we further prove
the optimality and stability of the corresponding equilibrium
point in Section V. The load estimation method is proposed
in Sectin VI. We confirm the performance of controllers via
simulations on a detailed power system in Section VII. Section
VIII concludes the paper.
2
0
Eqi =
xdi − xdi
xdi 0
Vi cos(δi − θi )
0 Eqi −
0
xdi
xdi
(3)
0
where xdi is the d-axis synchronous reactance, and xdi is the
d-axis transient reactance.
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
3
The active and reactive power (denoted by Qei ) injection to
the network are
III. F ORMULATION OF O PTIMAL F REQUENCY C ONTROL
A. Optimal Power-Sharing Problem in Frequency Control
0
Eqi Vi
Pei = 0 sin(δi − θi )
xdi
(4a)
0
Eqi Vi
Vi2
− 0 cos(δi − θi )
0
xdi
xdi
Qei =
(4b)
For controllable generators i ∈ NCG , the capacity limits are
g
P gi ≤ Pig ≤ P i
where
g
P gi , P i
are lower and upper limits of
(5)
The purpose of optimal frequency control is to let all the
controllable generators share power mismatch economically
when restoring frequency. Then we have the following optimization formulation, denoted by SFC.
X
fi (Pig )
(8a)
SFC:
min
g
i∈NCG
Pi ,i∈NCG
X
X
X
(8b)
Pig∗
pi −
Pig =
s.t.
i∈NCG
P gi ≤
Pig .
C. Dynamics of Voltage Phase Angles
To build a network-preserving power system model, relation
between generators and power network should be explicitly
established. In this paper, loads for bus i ∈ N are simply
modeled as constant active and reactive power injections. Then
the following equations are used to dictate the power balance
and voltage phase-angle dynamics at each bus:
θ̇i = ω̃i , i ∈ N
(6a)
X
0 = Pei − D̃i ω̃i − pi −
Pij , i ∈ NG
j∈Ni
X
0 = −D̃i ω̃i − pi −
Pij , i ∈ NL
j∈Ni
X
0 = Qei − qi −
Qij , i ∈ NG
j∈Ni
X
0 = −qi −
Qij , i ∈ NL
j∈Ni
(6b)
(6c)
(6d)
(6e)
where, pi , qi are active and reactive load demands, respectively; ω̃i the frequency deviation at bus i; Ni the set of buses
connected directly to bus i; D̃i the damping constant at bus i;
D̃i ω̃i the change of frequency-sensitive load [7].
In power system, line power flows are mainly related to
power angle difference between two buses rather than the
power angles independently. Then, we define new variables
to denote angle differences as ηii := δi − θi , i ∈ NG and
ηij := θi − θj , i, j ∈ N . The time derivative of ηii and ηij are
η̇ii = ωi − ω̃i , i ∈ NG
(7a)
η̇ij = ω̃i − ω̃j , i, j ∈ N , i 6= j
(7b)
respectively. In the following analysis, we use ηii and ηij as
new state variables instead of δi and θi .
To summarize, (1) − (4), (6b) − (6e), (7a), (7b) constitute the
nonlinear network preserving model of power systems, which
is in a form of differential-algebraic equations (DAE).
D. Communication Network
In this paper, we consider a communication graph among
the buses of controllable generators only. Denote E ⊆
NCG × NCG as the set of communication links. If generator
i and j can communicate directly to each other, we denote
(i, j) ∈ E. Obviously, edges in the communication graph E
can be different from lines in the power network E. For the
communication network, we make the following assumption:
A1: The communication graph E is undirected and connected.
i∈N
g
i∈NU G
Pig ≤ P i ,
i ∈ NCG
(8c)
where Pig∗ is the mechanical power of uncontrollable generator in the steady state. In (8a), fi (Pig ) concerns the controllable
generation Pig , satisfying the following assumption:
A2: The objective fi (Pig ) is second-order continuously dif0
ferentiable, strongly convex and fi (Pig ) is Lipschitz
continuous with Lipschitz constant li > 0. i.e. ∃ αi >
00
0, αi ≤ fi (Pig ) ≤ li .
To ensure the feasibility of the optimization problem, we
make an additional assumption.
A3: The system satisfies
X
X
X
X
g
P gi ≤
pi −
Pig∗ ≤
Pi
(9)
i∈NCG
i∈N
i∈NU G
i∈NCG
Specifically, we say A3 is strictly satisfied if all the inequalities
in (9) strictly hold.
B. Equivalent Optimization Model with Virtual Load Demands
In (8b), load demands are injected to every bus, which
sometimes cannot be measured accurately if at all. As a
consequence, the values of pi may be unknown to both the
controllable generators i, i ∈ NCG and the uncontrollable
generators i, i ∈ NU G . To circumvent such an obstacle in
design, we introduce a set of new variables, p̂i , to re-formulate
SFC as the following equivalent problem:
X
fi (Pig )
(10a)
ESFC:
min
g
i∈NCG
Pi ,i∈NCG
X
X
s.t.
Pig =
p̂i
(10b)
i∈NCG
i∈NCG
g
P gi ≤ Pig ≤ P i ,
i ∈ NCG
(10c)
where p̂i is the virtual load demand supplied byP
generator i in
the
steady
state,
which
is
a
constant,
satisfying
i∈NCG p̂i =
P
P
g∗
i∈N pi −
i∈NU G Pi . Obviously, the number of virtual
loads should be equal to that of the controllable generators.
Note that the power balance constraint (8b) only requires
that all the generators supply all the loads while it is not
necessary to figure out which loads are supplied exactly by
which generators. Hence we treat virtual load demands p̂i as
the effective demands supplied by generator i for dealing with
the issue that onlyPa part of generators
are controllable.
P
P
Simply letting i∈NCG p̂i = i∈N pi − i∈NU G Pig∗ , we
immediately have the following Lemma:
Lemma 1. The problems SFC (8) and ESFC (10) have the
same optimal solutions.
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
4
IV. C ONTROLLER D ESIGN
A. Distributed Frequency Control of Controllable Generators
1) Controller design based on primal-dual gradient algorithm: Invoking the primal-dual gradient algorithm, the
Lagrangian of the ESFC (10) is given by
X
X
X
p̂i
fi (Pig ) + µ
Pig −
L=
i∈NCG
i∈NCG
g
+ γi− (P gi − Pig ) + γi+ (Pig − P i ),
i∈NCG
i ∈ NCG (11)
where µ, γi− , γi+ are Lagrangian multipliers. Based on primaldual update, the controller for i ∈ NCG is designed as
0
ugi = Pig /Ti − kPig ωi + (fi (Pig ) + µ − γi− + γi+ ) (12a)
X
X
p̂i
(12b)
Pig −
µ̇ = kµ
=
kγi [P gi
−
+
Pig ]γ −
i
i
g +
h
γ̇i+ = kγi Pig − P i
(12c)
(12d)
γi+
where kPig , kµ , kγi are positive constants; ugi the control input;
0
fi (Pig ) the marginal cost at Pig . For any xi , ai ∈ R, the
operator is defined as: [xi ]+
ai = xi if ai > 0 or xi > 0; and
[xi ]+
ai = 0 otherwise.
2) Estimating µ by second-order consensus: In (12b), µ is
a global variable, which is a function of mechanical powers
and loads of the entire system. To circumvent the obstacle
in implementation, a second-order consensus based method is
utilized to estimate µ locally by using neighboring information
only. Specifically, for i ∈ NCG , the controller is revised to:
0
ugi = Pig /Ti − kPig ωi + fi (Pig ) + µi − γi− + γi+ (13a)
X
X
g
µ̇i = kµi Pi − p̂i −
(µi − µj ) −
zij
(13b)
j∈Nci
żij = kzi (µi − µj )
γ̇i−
=
kγi [P gi
h
−
j∈Nci
(13c)
+
Pig ]γ −
i
i
g +
γ̇i+ = kγi Pig − P i
where φωi is a positive definite function, and φωi = 0
holds only when Pig = Pig∗ .
The condition C1 on the active power dynamics of uncontrollable generators is easy to verify. As an example, the
commonly-used primary frequency controller
ugi = −ωi + ωi∗ − kωi (Pig − Pig∗ ) + Pig /Ti
(14)
satisfies C1 whenever kωi > 0. In this case, we have Sωi =
g
g∗ 2
g
g∗ 2
k1
2 (Pi − Pi ) with k1 > 0 and φ = k2 (Pi − Pi ) with
0 < k2 ≤ kωi · k1 .
i∈NCG
i∈NCG
γ̇i−
input −ωi and output Pig , i.e., there exists a continuously
differentiable, positive semidefinite function Sωi such that
Ṡωi ≤ (−ωi − (−ωi∗ )) Pig − Pig∗ − φωi (Pig − Pig∗ )
γi+
C. Excitation Voltage Dynamics of All Generators
Similar to the uncontrollable generators, the following
sufficient condition on excitation voltage dynamics of all
generators is needed to guarantee system stability, since we
do not design specific excitation voltage controllers here.
C2: The excitation voltage dynamics are strictly incrementally
output passive in terms of the input −Eqi and output
Ef i , i.e., there exists continuously differentiable, positive
semidefinite function SEi such that
∗
ṠEi ≤ −Eqi − (−Eqi
) Ef i − Ef∗i − φEi (Ef i − Ef∗i )
where φEi is a positive definite function, and φEi = 0
holds only when Ef i = Ef∗i .
C2 is also easy to satisfy. As an example, it can be verified
that the controller given in [23]
∗
h(Ef i , Eqi ) = −Ef i + Ef∗i − kEi (Eqi − Eqi
)
with kEi > 0 satisfies C2. In this case, SEi = k23 (Ef i − Ef∗i )2
with k3 > 0 and φ = k4 (Ef i − Ef∗i )2 with 0 < k4 ≤ kEi · k3 .
(13d)
(13e)
where, kµi , kzi are positive constants; Nci the set of neighbors
of bus i in the communication graph; µi the local estimation of
µ. Here, (13b) and (13c) are used to estimate µ locally, where
only neighboring information is needed. zij is an auxiliary
variable to guarantee the consistency of all µi .
For the Larangian multiplier µ, −µ is often regarded as the
marginal cost of generation. Theoretically, −µi should reach
consensus for all the generators in the steady state.
P Since µ̇i =
0 holds in the steady state, we have Pig − p̂i − j∈Nci zij = 0.
Hence, zij can be regarded as the virtual line power flow of
edge (i, j) in the communication graph.
(15)
V. O PTIMALITY AND S TABILITY
After implementing the controller on the physical power
system, the closed-loop system reads
(1) − (4), (6b) − (6e), (7a), (7b)
(16)
(13a) − (13e)
In this section, we prove the optimality and stability of the
closed-loop system (16).
A. Optimality
v(t) =
Denote the trajectory of closed-loop system as
0
g
−
+
η(t), ω(t), ω̃(t), P (t), µ(t), z(t), γ (t), γ (t), Eq (t), V (t) .
Define the equilibrium set of (16) as
V := {v ∗ |v ∗ is an equilibrium of (16)}
(17)
B. Active Power Dynamics of Uncontrollable Generators
We first present the following Theorem.
To guarantee system stability, a sufficient condition is given
for the active power dynamics of uncontrollable generators.
C1: The active power dynamics of uncontrollable generators
are strictly incrementally output passive in terms of the
Theorem 2. Suppose assumptions A1, A2 and A3 hold. In
equilibrium of (16), following assertions are true.
g
1) The mechanical powers Pig satisfy P gi ≤ Pig∗ ≤ P i ,
∀i ∈ NCG .
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
2) System frequency recovers to the nominal value, i.e.
ωi∗ = 0, ∀i ∈ NCG ∪ NU G , and ω̃i∗ = 0, ∀i ∈ N .
0
3) The marginal generation costs satisfy fi (Pig∗ ) − γi−∗ +
0
g∗
γi+∗ = fj (Pj ) − γj−∗ + γj+∗ , i, j ∈ NCG .
4) Pig∗ is the unique optimal solution of SFC problem (8).
5) µ∗i is unique if A3 is strictly satisfied.
The detailed proof of Theorem 2 is given in Appendix.A of
this paper. It shows that the nominal frequency is recovered
at equilibrium, and marginal generation costs are identical
for all controllable generators, implying the optimality of any
equilibrium.
5
Since the voltage phase deviation between two neighboring
buses is not large in practice, A4 is usually satisfied. Detailed
explanations can be found in Appendix.B of this paper.
The following stability result can be obtained.
Theorem 3. Suppose A1–A4 and C1, C2 hold. For every v ∗ ,
there exists a neighborhood S around v ∗ where all trajectories
v(t) satisfying (16) starting in S converge to the set V. In
addition, each trajectory converges to an equilibrium point.
We can further prove that ∇2 W > 0 and Ẇ ≤ 0. Moreover,
Ẇ = 0 holds only at equilibrium point. Then the theorem
can be proved using the LaSalle’s invariance principle [24,
Theorem 4.4]. Details of the proof are given in Appendix.B.
B. Stability
In this section, the stability of the closed-loop system (16)
is proved. First we define a function as
X
X
µi (Pig − p̂i )
fi (Pig∗ ) +
L̂ :=
i∈NCG
i∈NCG
X
X
1X
−
µi zij −
µi
(µi − µj )
i∈N
i∈NCG
j∈Ni
2
X CG g
X
g
g
g
−
+
+
γi (P i − Pi ) +
(18)
γi (Pi − P i )
i∈NCG
i∈NCG
VI. I MPLEMENTATION BASED ON F REQUENCY
M EASUREMENT
A. Estimation and Optimality
Note that virtual load demands p̂i used in the controller
(13) are difficult to directly measure or estimate
P in practice.
Lemma
1
implies
that
any
p̂
are
valid
as
long
as
i
i∈NCG p̂i =
P
P
g∗
p
−
P
.
Noticing
that
the
power
i∈N i
i∈NU G i
P imbalance
is
very
small
in
normal
operation,
we
have
i∈NCG Pei ≈
P
p̂
.
In
fact,
they
are
identical
in
steady
state. Hence,
i∈N i
we specify p̂i = Pei , which implies Pig − p̂i = Pig − Pei =
Mi ω̇i + Di ωi . That leads to an estimation algorithm of µi
X
X
µ̇i = kµi −
(µi − µj ) −
zij + Mi ω̇i + Di ωi
Denote x1 := (P ), x2 := (µ, z, γi− , γi+ ), x := (x1 , x2 ). Then
L̂(x1 , x2 ) is convex in x1 and concave in x2 .
Before giving the main result, we construct a Lyapunov
candidate function composed of four parts: the quadratic part,
the potential energy part, conditions C1 and C2 related parts,
j∈Ni
j∈Ni
0
as we now explain.
g
+ τi (−µi − fi (Pi ) + γi− − γi+ )
(23)
For i ∈ NG , the quadratic part is given by
where 0 < τi < 4/li . This way, we only need to measure
X 1
1
Wk (ω, x) =
Mi (ωi − ωi∗ )2 + (x − x∗ )T K −1 (x − x∗ ) frequencies ωi at each bus locally, other than the global load
2
2
i∈NG
demands. Since the controller only needs µi of neighboring
(19) buses in the communication graph, it is easy to implement.
Now, we reconstruct the closed-loop system by replacing
where K = diag(kPig , kµi , kzi , kγi ) is a diagonal positive
(13b)
with (23) in (16), which is
definite matrix.
0
Denoting xp = (Eqi , Vi , δi , θi ), the potential energy part is
(1) − (4), (6b) − (6e), (7a), (7b)
(24)
(13a), (13c) − (13e), (23)
∗ T
∗
∗
Wp (xp ) = W̃p (xp ) − (xp − xp ) ∇xp W̃p (xp ) − W̃p (xp ) (20)
We have the following lemma.
where,
Lemma 4. Assertions 1)-5) in Theorem 2 still hold for the
X
X
0
1
W̃p (Eqi , Ei , Vi , δi , θi ) =
Bii Vi2 +
pi θi
equilibrium of (24).
i∈N 2
i∈N
X
1X X
The proof of Lemma 4 is given in Appendix.C.
−
qi ln Vi −
Vi Vj Bij cos (θi − θj )
(21)
2
g
i∈N
i∈N j∈Ni
0
0 2
X Eqi Vi
X
xdi
cos(δi − θi ) +
−
Eqi
0
0
0
xdi
2xdi (xdi − xdi )
i∈NG
i∈NG
P
PConditions 1C1 and C2 related parts are i∈NU G Sωi and
i∈NG T 0 (xdi −x0 ) SEi respectively.
d0i
di
The Lyapunov function is defined as
X
X
SEi
W =Wk + Wp +
Sωi +
(22)
0
0
Td0i (xdi − xdi )
i∈N
i∈N
UG
G
Then, we give the following assumption.
A4: The Hessian of Wp satisfies ∇2v Wp (v) > 0 at desired
equilbirium.
B. Discussion on Stability
Recall (2b), then (23) is derived to
X
X
µ̇i = kµi Pig − p̂i + p̂i − Pei −
(µi − µj ) −
zij
j∈Ni
+ τi (−µi −
0
fi (Pig )
+
γi−
j∈Ni
−
γi+ )
(25)
∗
Denote ρi = p̂i − Pei = Pei
− Pei , which is the difference of
electric power and its value in the steady state. We have the
following assumption
A5: The disturbance can be written as ρi = βi (t)ωi , where
|βi (t)| ≤ β̄i and β̄i is a positive constant. In addition, the
set { t < ∞ | ωi (t) = ωi∗ } has a measure zero.
DFIG
Bus12
Tw XX, XX XXXX
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO.
Bus1
Bus5
Bus7
Bus8
Bus9
6
Bus10 Bus11 Bus3
T3
T1
Gen1
Gen3
Bus6
exists such βi (t).
A5 argues
Whenever ωi 6= ωi∗ , thereT2 alway
T4
Load1
Load2
Bus2
Bus4
∗
that ωi (t) = ωi only happens
at isolated Gen4
points
except
Gen2
equilibrium. Generally, this is reasonable in power system.
Denote the state variables of (24) and its equilibrium set are
ṽ and Ṽ respectively. We have following stability result
Bus1 Bus5 Bus6 Bus7
Bus10 Bus11 Bus2
Theorem 5. Suppose
A1–A5, C1,Bus8C2 Bus9hold
and
(9) is not
T1
T2
∗
binding. For every Gen1
ṽ , there exists a neighborhood Gen2
S around
ṽ ∗ where all trajectories ṽ(t) satisfying (24) starting in S
Bus15 Bus16 Bus17 Bus4
Bus14 Bus13 Bus12
converge to the set ṼBus3whenever
T4
T3
p
Gen4
Gen3
(26)
β̄i < τi Di (4 − τi li ).
G37
Controller
u gj
26
l
g
27
2
18
1
We consider the following scenario: 1) at t = 10s, there
is a step change of 60MW load demands at each of buses 3,
15, 23, 24, 25; 2) at t = 70s, there is another step change of
120MW load at bus 23. Neither the original load demands nor
their changes are known to the generators.
G35
16
15
G39
21
39
4
22
14
5
6
12
7
8
19
13
11
20
23
33
10
9
31
32
G31
34
G32
36
G33
G34
G36
Fig. 2. The New England 39-bus system
Eqi' , Vi , i , i , i , Pi g
Eqj' , V j , j , j , j , Pjg
i , j
Controller
g
i
Controller
u gj
E fi
u
A. Test System
B. Results under Small Disturbances
G38
24
35
uil
To test the proposed controller, the New England 39-bus
system with 10 generators as shown in Fig.2, is utilized. In the
simulation, we apply (23) to estimate the virtual load demands.
All simulations are implemented in the commercial power
system simulation software PSCAD [25], and are carried on
a notebook with 8GB memory and 2.39 GHz CPU.
We control only a subset of these generators, namely G32,
G36, G38, G39, while the remaining are equipped with the
primary frequency control given in (14). In particular, we apply
the controller (13) derived based on a simple model to a much
more realistic and complicated model in PSCAD. The detailed
electromagnetic transient model of three-phase synchronous
machines (sixth-order model) is adopted to simulate generators
with governors and exciters. All the lines and transformers
take both resistance and reactance into account. The loads are
modeled as fixed loads in PSCAD. The communication graph
is undirected and set as G32 ↔ G36 ↔ G38 ↔ G39 ↔ G32.
The objective function is set as fi = 21 ai (Pig )2 + bi Pig ,
which is the generation cost of generator i. Capacity limits of
Pig and parameters ai , bi are given in Table I.
The closed-loop system is shown in Fig.3, where each
generator only needs to measure local frequency, mechanical
power, voltage and phase angle to compute its control command. Note that there is no load measurement and only µi are
communicated between neighboring controllable generators.
38
17
3
Controller
VII.u j S IMULATION Rui ESULTS
29
28
30
Moreover, the convergence of each such trajectory is to a point.
The proof of Theorem 5 is given in Appendix.D. In
fact, the range of β(t) can be very large as long as li is
small enough. For example, set τi = 3/li , then β̄i (t) <
p
3DP
i /li . As we know, if we change the objective function
to k i∈NCG fi (Pig ), k > 0, the optimal solution will not
change. Thus, li can be very small as long as k is small
enough. In this regard, (26) is not conservative. Moreover,
g
l
g
i , Pieven
, Pi l , Pij(9)
, Pki is binding, our
Pjk Section VII that
j , Pj , Pj , Pij ,in
we will illustrate
controller still works.
37
25
G30
E fj
…
Fig. 3. Diagram of the closed-loop system
TABLE I
C APACITY LIMITS OF GENERATORS
Gi
32
36
38
39
g
[P gi , P i ] (MW)
[0, 1000]
[0, 1000]
[0, 850]
[0, 1080]
ai
0.00009
0.00014
0.00010
0.00008
bi
0.032
0.030
0.032
0.032
TABLE II
E QUILIBRIUM POINTS
Stage 1
Stage 2
g
P32
(MW)
927
968
g
P36
(MW)
610
652
g
P38
(MW)
834
850
g
P39
(MW)
1043
1080
1) Equilibrium: In steady states, the nominal frequency is
well recovered. The optimal mechanical powers are given in
Table II, which are identical to the optimal solution of (8)
computed by centralized optimization. Stage 1 is for the period
from 10s to 70s, and Stage 2 from 70s to 130s. The values
in Table II are generations at the end of each stage. In Stage
1, no generation reaches its limit, while in Stage 2 both G38
and G39 reach their upper limits. At the end of each stage, the
marginal generation cost −µi of generator i, converges identically (see Fig. 9), implying the optimality of the results. The
test results confirm the theoretical analyses and demonstrate
that our controller can automatically attain optimal operation
points even in the more realistic and sophisicated model.
2) Dynamic Performance: In this subsection, we analyze
the dynamic performance of the closed-loop system. For
G
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
Frequency (Hz)
60.1
60
59.9
59.7
59.6
10
30
50
70
Time (s)
90
G35
G36
G37
G38
G39
G30
G31
G32
G33
G34
G35
G36
G37
G38
G39
G30
G31
G32
G33
G34
59.8
110
10
130
30
50
70
Time (s)
800
1600
With AGC
Mechanical Power (MW)
Distributed control
60.2
7
110
G32
G36
G38
G39
1200
1000
600
500
G30
G31
G33
G34
G35
G37
400
800
300
600
200
400
90
700
1400
10
30
130
50
70
Time (s)
90
110
130
10
30
50
70
Time (s)
90
110
130
Fig. 6. Dynamics of mechanical powers under AGC
Fig. 4. Dynamics of frequencies
Mechanical Power (MW)
Frequency (Hz)
60.2
60.1
60
59.9
G30
G31
G32
G33
G34
59.8
59.7
59.6
G35
G36
G37
G38
G39
30
50
70
Time (s)
90
110
130
700
1000
600
900
G32
G36
G38
G39
800
700
600
500
G30
G31
G33
G34
G35
G37
400
300
500
400
10
800
1200
1100
200
10
30
50
70
Time (s)
90
110
130
10
30
50
70
Time (s)
90
110
130
Fig. 7. Dynamics of mechanical powers under the proposed control
Fig. 5. Dynamics of frequencies with the controller in [11]
Voltage (p.u.)
Generation (MW)
perfectly in coincidence with the indication given by Theorems
comparison, automatic generation control (AGC) is tested in
2, 3 and 5 and Lemma 4 in this paper.
the same scenario. In the AGC implementation, the signal of
Mechanical power dynamics under the AGC and the proarea control error (ACE) is given by ACE = Kf ω + Pij
posed controller are shown in Fig.6 and Fig.7, respectively.
[26, Chapter 11.6], where, Kf is the frequency response
The left parts show mechanical powers of G32, G36, G38,
coefficient; ω the frequency deviation; Pij the deviation of
G39, while the right parts show mechanical powers of other
tie line power. In the case studies, we can treat the whole
generators adopting conventional controller (14). With both
system as one control area, implying Pij = 0. Hence, the
controls, mechanical powers of the generators adopting (14)
control center computes ACE = Kf ω and allocates it to
remain identical in the steady state. However, there are two
AGC generators, G32, G36, G38 and G39. In this situation,
problems when adopting the AGC: 1) mechanical powers
the
P control command of each generator is ri · ACE, where are not optimal; 2) mechanical power of G39 violates the
i ri = 1. In this paper, we set ri = 0.25 for i = 1, 2, 3, 4.
capacity limit. In contrast, the proposed control can avoid these
The trajectories of frequencies are given in Fig.4, where
problems. In Stage 1 of Fig.7, no generator reaches capacity
the left one stands for the proposed controller and the right
limits. In Stage 2, both G38 and G39 reach their upper limits.
one for the AGC. It is shown in Fig.4 that the frequencies
Then, G38 and G39 stop increasing their outputs while G32
are recovered to the nominal value under both controls. The1200
800
and G36 raise their outputs to balance
the load demands. In
700
frequency drops under two controls are very similar while the1100
addition,
the
steady
states
of
in
both
stages
are optimal, which
recovery time under the proposed control is much less than1000
600
900 are the same as shown in Table II.
that under the conventional AGC.
500
G30
800
We also illustrate in G32
Fig.8 the dynamics of voltage at
G36
G31
G38
Although a number of studies have been devoted to dis- 700
G33
400
G34
buses 3, 15, 23, 24, 25.G39The voltages
converge rapidly, and
G35
tributed frequency control of power systems, most of them 600
300
G37
only
experience
small
drops
when
loads
increase.
This
result
500
assume that all the nodes are controllable except [11]. To
200
validates the effectiveness of the voltage control.
10
30
50
70
90
110
make a fair comparison, the controller proposed in [11] is 400 10 30 50 70 90 110 130
Time generation
(s)
Time
The marginal
cost of generator i, −µ
are shown
i ,(s)
adopted as a rival in our tests. As shown in Eq.(8) of [11], each
controller needs to predict the load it should supply in steady
1.1
state. However, it is hard to acquire an accurate prediction in
1.08
practice, which could lead to steady-state frequency error.
1.06
1.04
In the next case, we compare the two controls in the same
1.02
scenario as that in Section VII.B. The dynamics of frequencies
1
0.98
Bus3
with the controller given in [11] are shown in Fig.5. It is
Bus15
0.96
Bus23
observed that there is a frequency deviation in steady state, as
Bus24
0.94
Bus25
0.92
the prediction is inaccurate. Although the deviation is usually
0.9
10
30
50
70
90
110
130
quite small, it is difficult to completely eliminate. In contrast,
Time (s)
when the proposed method is adopted, there is no frequency
deviation in steady state, as is shown in Fig.4. This result is Fig. 8. Dynamics of bus voltages
130
1.05
1
10
50
70
90
130
JOURNAL
OF LAT30EX CLASS
FILES,
VOL.
XX, 110
NO. XX,
XX XXXX
Time (s)
G32
G36
G38
G39
1.05
10
30
50
70
Time (s)
90
110
130
30
50
70
Time (s)
90
110
59.5
59.4
0 10
20 30 40 50 60 70 80 90
Time (s)
1000
G32
G36
G38
G39
500
0
0 10
20 30 40 50 60 70 80 90
Time (s)
Mechanical Power (MW)
59.6
Mechanical Power (MW)
Frequency (Hz)
59.7
59.3
59.7
130
1500
G30
G31
G33
G34
G35
G36
G37
G38
G39
60
59.9
1
0.8
0.6
0.4
0.2
0
20
40
60
80
Time (s)
0
100
0
20
40
60
80
Time (s)
100
Fig. 11. Dynamics of frequencies and voltages with line trip
60.1
60
59.8
60.1
59.8
10
Fig. 9. Dynamics of −µ and z
59.9
60.2
Bus4
Bus14
1.2
G35
G36
G37
G38
G39
1200
955
1100
950
1000
945
30
50
70
90
110
G32
G36
G38
G39
G32
G36
G38
G39
1.15
1.1
900
800
1.1
G30
G31
G32
G33
G34
Voltage (p.u.)
60.3
Frequency (Hz)
1.15
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
z39-32
z39-38
z32-36
z36-38
z
1.2
1
60.4
103
1.3
1.25
8
1.095
1.09
1.05
1.085
700
1
1.08
600
1.075
0.95
500
400
0
20
40
60
Time (s)
80
100
0.9
1.07
1.065
30
0
20
40
50
70
60
Time (s)
90
80
110
100
Fig. 10. Dynamics of frequencies and mechanical powers with generator trip
Fig. 12. Dynamics of −µ and mechanical powers with line trip
in the left part of Fig.9. They converge in both stages and the
steady-state values in Stage 2 are slightly bigger than that in
Stage 1, as the load changes lead to an increase in the marginal
generation cost. Dynamics of zij , (i, j) ∈ E are illustrated
in the right part of Fig.9, which demonstrate that the steady
state values do not change in the two stages. In addition, the
variation of z in transient is very small as the deviation of µi
is very small.
line is re-closed. At the same time, the voltages of buses 4 and
14 drop to nearly zero when the fault happens. The voltages are
then stabilized to a new steady-state value in around 10s after
the fault line is tripped. They are slightly different from their
initial values because the systems operating point has changed
due to line tripping. When the tripped line is re-closed, the
voltages recover to the initial values quickly.
When line (4, 14) is tripped, the power flow across the
power network varies accordingly. As a consequence, the line
loss also changes, causing variations of mechanical powers,
as shown in Fig.12. In the left part of Fig.12, the inset is
the dynamics of mechanical power of G39 from 30s to 110s.
Similarly, the inset in the right part is dynamics of −µ of
all generators. Mechanical powers and their marginal costs
all increase when the line is tripped. However, the proposed
controller compensates the loss change autonomously.
These simulation results demonstrate that the proposed
distributed optimal frequency controller can cope with large
disturbances such as generator tripping and short-circuit fault.
C. Performance under Large Disturbances
In this subsection, two scenarios of large disturbances are
considered. One is a generator tripping and the other is a shortcircuit fault followed by a line tripping.
1) Generator tripping: At t = 10s, G32 is tripped, followed by occurrence of certain power imbalance. Note that the
communication graph still remains connected. The output of
G32 drops to zero. Frequency and mechanical powers change
accordingly. System dynamics are illustrated in Fig.10. The
left part of Fig.10 shows the frequency dynamics, and the right
shows the mechanical power dynamics of controllable generators. It is observed that the system frequency experiences a big
drop at first, and then recovers to the nominal value quickly as
other controllable generators increase outputs to balance the
power mismatch. These results confirm that our controller can
adapt to generator tripping autonomously even if the tripped
generator is contributing to frequency control.
2) Short-circuit fault: At t = 10s, there occurs a threephase short-circuit fault on line (4, 14). At t = 10.05s, this line
is tripped by breakers. At t = 60s, the fault is removed and line
(4, 14) is re-closed. Frequency dynamics and voltage dynamics
of buses 4 and 14 are given in Fig.11, where the left part shows
frequency dynamics and the right shows voltage dynamics. It
can be seen that the frequency experiences violent oscillations
after the fault happens. And then it is stabilized quickly once
the line is tripped. Small frequency oscillation occurs when the
VIII. C ONCLUSION
In this paper, we have designed a distributed optimal frequency control using a nonlinear network-preserving model,
where only a subset of generator buses is controlled. We have
also simplified the implementation by relaxing the requirements of load measurements and communication topology.
Since nonlinearity of power flow model and dynamics of excitation voltage has been taken into account, our controllers can
cope with large disturbances. We have proved that the closedloop system asymptotically converges to the optimal solution
of the economic dispatch problem. We have also carried out
substantial simulations to verify the good performance of our
controller under both small and large disturbances.
In this work, we have not considered the constraints on
line flows, since the controllable generators are selected arbi-
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
trarily and may not suffice for congestion management. An
interesting problem is how to find out the minimal set of
controllable generators to fulfill the requirement of congestion
management, which is our future work.
R EFERENCES
[1] D. Cai, E. Mallada, and A. Wierman, “Distributed optimization decomposition for joint economic dispatch and frequency regulation,” IEEE
Trans. Power Syst., early access, 2017.
[2] D. J. Shiltz, S. Baros, M. Cvetkovi?, and A. M. Annaswamy, “Integration
of automatic generation control and demand response via a dynamic
regulation market mechanism,” IEEE Trans. Control Syst. Technology,
vol. PP, no. 99, pp. 1–16, 2017.
[3] A. Jokić, M. Lazar, and P. P. van den Bosch, “Real-time control of power
systems using nodal prices,” Int. J. Elect. Power Energy Syst., vol. 31,
no. 9, pp. 522–530, 2009.
[4] F. Dorfler, J. W. Simpson-Porco, and F. Bullo, “Breaking the hierarchy:
Distributed control and economic optimality in microgrids,” IEEE Trans.
Control Netw. Syst., vol. 3, no. 3, pp. 241–253, Sept 2016.
[5] N. Li, C. Zhao, and L. Chen, “Connecting automatic generation control
and economic dispatch from an optimization view,” IEEE Trans. Control
Netw. Syst., vol. 3, no. 3, pp. 254–264, Sept 2016.
[6] D. Feijer and F. Paganini., “Stability of primal-dual gradient dynamics
and applications to network optimization,” Automatica, vol. 46, no. 12,
pp. 1974–1981, Dec 2010.
[7] C. Zhao, U. Topcu, N. Li, and S. H.Low., “Design and stability of loadside primary frequency control in power systems,” IEEE Trans. Autom.
Control, vol. 59, no. 5, pp. 1177–1189, Jan. 2014.
[8] E. Mallada, C. Zhao, and S. H. Low, “Optimal load-side control for
frequency regulation in smart grids,” IEEE Trans. Autom. Control, early
access, 2017.
[9] A. Kasis, E. Devane, C. Spanias, and I. Lestas, “Primary frequency
regulation with load-side participation part i: stability and optimality,”
IEEE Trans. Power Syst. early access, 2016.
[10] C. Zhao, E. Mallada, S. Low, and J. Bialek, “A unified framework
for frequency control and congestion management,” in Power Systems
Computation Conference (PSCC), 2016. IEEE, 2016, pp. 1–7.
[11] J. Z. Pang, L. Guo, and S. H. Low, “Load-side frequency regulation with
limited control coverage,” ACM SIGMETRICS Performance Evaluation
Review, vol. 45, no. 2, pp. 94–96, 2017.
[12] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Decentralized optimal frequency control of interconnected power systems with transient
constraints,” in Proc. 55th IEEE Conf. Decis. Control (CDC), Dec 2016,
pp. 664–671.
[13] X. Zhang, N. Li, and A. Papachristodoulou, “Achieving real-time economic dispatch in power networks via a saddle point design approach,”
in 2015 IEEE PES General Meeting, July 2015, pp. 1–5.
[14] Z. Wang, F. Liu, S. H. Low, C. Zhao, and S. Mei, “Distributed frequency
control with operational constraints, part i: Per-node power balance,”
IEEE Trans. Smart Grid, early access, 2017.
[15] ——, “Distributed frequency control with operational constraints, part ii:
Network power balance,” IEEE Trans. Smart Grid, early access, 2017.
[16] T. Stegink, C. D. Persis, and A. van der Schaft, “A port-hamiltonian
approach to optimal frequency regulation in power grids,” in 54th IEEE
Conference on Decision and Control (CDC), Dec 2015, pp. 3224–3229.
[17] T. W. Stegink, C. De Persis, and A. J. van der Schaft, “Stabilization
of structure-preserving power networks with market dynamics,” arXiv
preprint arXiv:1611.04755, 2016.
[18] T. Stegink, C. D. Persis, and A. van der Schaft, “A unifying energy-based
approach to stability of power grids with market dynamics,” IEEE Trans.
Autom. Control, vol. 62, no. 6, pp. 2612–2622, June 2017.
[19] X. Zhang and A. Papachristodoulou., “A real-time control framework for
smart power networks: Design methodology and stability,” Automatica,
vol. 58, pp. 43–50, Aug 2015.
[20] C. Zhao and S. Low, “Optimal decentralized primary frequency control
in power networks,” in Proc. 53rd IEEE Conf. Decis. Control (CDC),
Dec 2014, pp. 2467–2473.
[21] S. Trip and C. De Persis, “Optimal generation in structure-preserving
power networks with second-order turbine-governor dynamics,” in Control Conference (ECC), 2016 European. IEEE, 2016, pp. 916–921.
[22] P. Kundur, Power System Stability and Control. McGraw-hill New
York, 1994, vol. 7.
[23] Y. Liu, C. Li, and Y. Wang, “Decentralized excitation control of multimachine multi-load power systems using hamiltonian function method,”
Acta Automatica Sinica, vol. 35, no. 7, pp. 919–925, 2009.
9
[24] H. K. Khalil, Nonlinear systems. Prentice hall New Jersey, 1996, vol. 3.
[25] https://hvdc.ca/pscad/.
[26] V. V. Arthur and R. Bergen, Power System Analysis. Upper Saddle
River: NJ: Pretice Hall, 2000.
[27] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
[28] C. D. Persis, N. Monshizadeh, J. Schiffer, and F. Drfler, “A lyapunov
approach to control of microgrids with a network-preserved differentialalgebraic model,” in 2016 IEEE 55th Conference on Decision and
Control (CDC), Dec 2016, pp. 2595–2600.
[29] P. Lin, Y. Jia, and L. Li, “Distributed robust h∞ consensus control in
directed networks of agents with time-delay,” Syst. Control Lett., vol. 57,
no. 8, pp. 643–653, 2008.
A PPENDIX
A. Proof of Theorem 2
Proof. From γ̇i− = γ̇i+ = 0 in (13d), it follows that P gi ≤
g
Pig∗ ≤ P i , which is the first assertion. From żij = 0 in (13c),
we get µ∗i P
= µ∗j = µ0 . Set
P µ̇i = 0, add
P (13b) for i ∈ NCG ,
and recall i∈NCG p̂i = i∈N pi − i∈NU G Pig∗ , we have
X
X
X
Pig∗ −
pi +
Pig∗ = 0 (A.1)
i∈NCG
i∈N
i∈NU G
The right sides of (7a) and (7b) vanish in the equilibrium
points, which implies ωi∗ = ω̃i∗ = ω̃j∗ = ω0 . Set ω̇i = 0 and
add (2b), (6b) and (6c). We have
X
X
X
X
ω0
D̃i =
Pig∗ −
pi +
Pig∗ = 0 (A.2)
i∈N
i∈NCG
i∈N
i∈NU G
This implies that ω0 = 0 due to D̃i > 0, which is the second
assertion.
Combine (13a), (2d) and Ṗig = 0. We have
0
fi (Pig∗ ) − γi−∗ + γi+∗ + ωi∗ + µ∗i = 0
Since ω0 = 0, µ∗i = µ∗j = µ0 , we can obtain the third
assertion.
Next we will prove assertion 4). Since all the constraints
of SFC are linear, A3 implies that Slater’s condition holds
[27, Chapter 5.2.3]. Moreover, the objective function is strictly
convex. We only need to prove that (Pig∗ , µ0 , γi−∗ , γi+∗ ) satisfies the KKT condition of SFC in order to prove the fourth
assertion.
The KKT conditions of SFC problem (8) are
0
fi (Pig∗ ) − γi−∗ + γi+∗ + µ0 = 0
X
X
X
Pig∗ −
pi +
i∈NCG
g
P gi ≤ Pig∗ ≤ P i
γi−∗ ≥ 0, γi+∗ ≥ 0
γi−∗ (P gi − Pig∗ ) =
i∈N
(A.3a)
i∈NU G
Pig∗
=0
(A.3b)
(A.3c)
(A.3d)
0, γi+∗ (Pig∗
−
g
Pi )
=0
(A.3e)
From γ̇i− = γ̇i+ = 0, we have (A.3c), (A.3d) and (A.3e). From
the third assertion, we have (A.3a). From ω̇ = 0 and the second
assertion, we have (A.3b). Therefore, the equilibrium points of
the closed-loop system (16) satisfy the KKT conditions (A.3).
This implies the fourth assertion.
If A3 is strictly satisfied, we know ∃ i ∈ NCG that γi−∗ =
0
+∗
γi = 0. Then, µ∗i = −fi (Pig∗ ) is uniquely determined by
Pig∗ , implying the last assertion. This completes the proof.
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
10
Proof of Theorem 3. Recall (18), and dynamics of (13a) (13e) are rewritten as
(B.1a)
Ṗig = −kPig · ωi + ∂ L̂(x1 , x2 )/∂Pig
µ̇i = kµi · ∂ L̂(x1 , x2 )/∂µi
(B.1c)
g
σ + := {i ∈ NCG | γi+ = 0, Pig − P i < 0}
σ := {i ∈
= 0,
−
Pig
< 0}
(ωi −
X
− Di ωi − Pei ) −
−
(Pig
Pig∗ )(ωi
−
− ωi∗ )
Since L̂ is a convex function of x1 and a concave function of
x2 , we have
≤L̂(x∗1 , x2 ) − L̂(x1 , x2 ) + L̂(x1 , x2 ) − L̂(x1 , x∗2 )
≤0
(B.3b)
(B.7)
where the first inequality follows because L̂ is convex in x1
and concave in x2 and the second inequality follows because
(x∗1 , x∗2 ) is a saddle point. Hence, we have
X
X
Ẇk ≤ −
Di (ωi − ωi∗ )2 −
D̃i (ω̃i − ω̃i∗ )2
i∈NG
(Pig
−
Pig∗ )ωi
X
i∈N
(ωi − ω̃i )(Pei −
∗
Pei
)
i∈NG
+
X
−
(ω̃i − ω̃j )(Pij − Pij∗ )
(i,j)∈E
X
i∈NU G
(Pig
−
Pig∗ )(ωi
− ωi∗ )
(B.8)
The partial of Wp (xp ) is
1
∗
0 (Eqi − Eqi )
xdi − xdi
∇V i W p = 0
∇ E 0 Wp =
· (ωi − ωi∗ )
(B.9a)
qi
∇δi Wp = Pei −
X
X
∗
∇θ i W p =
(Pij − Pij∗ ) −
(Pei − Pei
)
∗
Pei
)
(B.9c)
(B.9d)
i∈NG
(i,j)∈E
The derivative of Wp is
(Pig − Pig∗ )(ωi − ωi∗ )
(B.4)
Ẇp =
∗
X (Eqi − Eqi
)(Ef i − Ef∗i )
0
−
−∗ T
) (P g − P g )
(γ − − γ −∗ )T [P g − P g ]+
γ − ≤ (γ − γ
X
+
0
Td0i (xdi − xdi )
i∈NG
where the inequality is due to
(ωi − ω̃i )(Pei −
∗
Pei
)
+
i∈NG
X
∗ 2
(Eqi − Eqi
)
0
0
Td0i (xdi − xdi )
(ω̃i − ω̃j )(Pij − Pij∗ )
(i,j)∈E
(B.10)
P gi −Pig
Here the inequality holds since γi− = 0 ≤ γi−∗ and
<
0 for i ∈ σ − , i.e., (γi− − γi−∗ ) · (P gi − Pig ) ≥ 0. Similarly for
i ∈ σ + , we have
g
X
−
i∈NG
= (γ − − γ −∗ )T ∇γ − L̂
g
+
+∗ T
(γ + − γ +∗ )T [P g − P ]+
) (P g − P )
γ + ≤ (γ − γ
= (γ + − γ +∗ )T ∇γ + L̂
From (6b) and (6c), it yields
X
X
∗
∗
∗
0=
(ω̃i − ω̃i ) (Pei − Pei ) −
(Pij − Pij ) (B.5)
j∈Ni
(B.9b)
∗
Pei
i∈NU G
i∈NG
(B.6)
(B.3a)
− (x1 − x∗1 )T · ∇x1 L̂ + (x2 − x∗2 )T · ∇x2 L̂
= −(x1 − x∗1 )T · ∇x1 L̂ + (x2 − x∗2 )T · ∇x2 L̂
X
X
−
Di (ωi − ωi∗ )2 −
(ωi − ωi∗ )(Pei −
i∈NG
i∈NG
X
(i,j)∈E
X
(B.2b)
i∈NCG
+
i∈NG
(ω̃i − ω̃j )(Pij − Pij∗ )
−(x1 − x∗1 )T · ∇x1 L̂(x1 , x2 ) + (x2 − x∗2 )T · ∇x2 L̂(x1 , x2 )
i∈NG
−
X
∗
(ω̃i − ωi )(Pei − Pei
)−
(B.2a)
− (x1 − x∗1 )T · ∇x1 L̂ + (x2 − x∗2 )T · ∇x2 L̂
X
∗
=
(ωi − ωi∗ )(Pig − Pig∗ − Di (ωi − ωi∗ ) − (Pei − Pei
))
Pig∗ )
X
− (x1 − x∗1 )T · ∇x1 L̂ + (x2 − x∗2 )T · ∇x2 L̂
i∈NCG
(Pig
i∈N
i∈NG
+
−
i∈NG
X
Add (B.5) to (B.4), then we have
X
X
Di (ωi − ωi∗ )2 −
D̃i (ω̃i − ω̃i∗ )2
Ẇk ≤ −
(B.1e)
i∈NG
≤
j∈Ni
+
The derivative of Wk is
X
Ẇk =
Mi (ωi − ωi∗ )ω̇i + (x − x∗ )T · K −1 ẋ
ωi∗ )(Pig
(Pij − Pij∗ )
=L̂(x∗1 , x2 ) − L̂(x∗1 , x∗2 ) + L̂(x∗1 , x∗2 ) − L̂(x1 , x∗2 )
Then (13d) and (13e) are equivalent to
g
kγi (Pig − P i ), if i ∈
/ σ+ ;
+
γ̇i =
0,
if i ∈ σ + .
kγi (P gi − Pig ), if i ∈
/ σ− ;
−
γ̇i =
0,
if i ∈ σ − .
X
i∈NL
X
(B.1d)
With regard to the closed-loop system, we first define two
sets, σ + and σ − , as follows [6].
P gi
(ω̃i − ω̃i∗ )
i∈NU G
γi
NCG | γi−
X
D̃i (ω̃i − ω̃i∗ )2 −
i∈N
(B.1b)
żij = kzi · ∂ L̂(x1 , x2 )/∂zij
h
i+
γ̇i− = kγi · ∂ L̂(x1 , x2 )/∂γi− −
γi
h
i+
+
+
γ̇i = kγi · ∂ L̂(x1 , x2 )/∂γi +
−
X
−
B. Proof of Theorem 3
The Lyapunov
is defined as W = Wk + Wp +
P
P function
1
Sωi +
S , and its derivative is
0
0
T (x −x ) Ei
i∈NU G
i∈NG
d0i
di
di
X
X
ṠEi
Ṡωi +
Ẇ = Ẇk + Ẇp +
0
0
i∈NU G
i∈NG T
d0i (xdi − xdi )
X
X
≤−
Di (ωi − ωi∗ )2 −
D̃i (ω̃i − ω̃i∗ )2
i∈NG
−
X
i∈NG
∗ 2
(Eqi − Eqi
)
0
0
Td0i (xdi − xdi )
i∈N
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
+
X
(Pig − Pig∗ )(ωi − ωi∗ ) +
Ṡωi
i∈NU G
i∈NU G
+
X
11
∗
X (Eqi − Eqi
)(Ef i − Ef∗i )
0
0
+
X
ṠEi
0
Td0i (xdi − xdi )
0
Td0i (xdi − xdi )
i∈NG
i∈NG
X
X
∗ 2
Di (ωi − ωi ) −
D̃i (ω̃i − ω̃i∗ )2
≤−
i∈NG
−
X
i∈NG
≤0
i∈N
∗ 2
X
X
(Eqi − Eqi
)
φ Ei
φωi −
−
0
0
Td0i (xdi − xdi ) i∈N
i∈N
UG
G
(B.11)
The last two inequalities are due to assumption A4.
To prove the locally asymptotic stability of the closed-loop
system, we also need to prove that W > 0, ∀v ∈ S\v ∗ .
Equivalently, we need ∇2v W > 0, ∀ v ∈ S\v ∗ , i.e. A4.
Consequently, there exists a neighborhood set { v : W (v) ≤
} for all sufficiently small > 0 so that ∇2v W (v) > 0.
Hence, there is a compact set S around v ∗ contained in such
neighborhood, which is forward invariant. Let Z1 := { v :
Ẇ (v) = 0 }. By LaSalle’s invariance principle, the each
of trajectories v(t) starting from S converges to the largest
invariant set Z + contained in S ∩ Z1 . From above analysis, if
Ẇ (v) = 0, v is an equilibrium point of the closed-loop system
(16). Hence, v converges to Z + ∈ V.
Finally, we will prove the convergence of each v(t) starting
from V is to a point by following the proof of Theorem 1 in
[18]. Since v(t) is bounded, its ω-limit set Ω(v) 6= ∅. By contradiction, suppose there exist two different points in Ω(v), i.e.,
v1∗ , v2∗ ∈ Ω(v), v1∗ 6= v2∗ . Since the Hessian of W1 (v), W2 (v)
is positive definite in S, there exist W1 (v), W2 (v) defined by
(22) with respect to v1∗ , v2∗ and scalars c1 > 0, c2 > 0 such
that two sets W1−1 (≤ c1 ) := {v|W1 (v) ≤ c1 }, W2−1 (≤ c2 ) :=
{v|W2 (v) ≤ c2 } are disjoint (i.e. W1−1 (≤ c1 )∩W2−1 (≤ c2 ) =
∅) and compact. In addition, W1−1 (≤ c1 ), W2−1 (≤ c2 ) are
forward invariant. By (B.11), there exists a finite time t1 > 0
such that v(t) ∈ W1−1 (≤ c1 ) for ∀t ≥ t1 . Similarly, there
exists a finite time t2 > 0 such that v(t) ∈ W2−1 (≤ c2 ) for
∀t ≥ t2 . This implies that v(t) ∈ W1−1 (≤ c1 ) ∩ W2−1 (≤ c2 )
for ∀t ≥ max(t1 , t2 ), which is a contradiction. This completes
the proof.
In this part, we discuss the reasonableness of Assumption
A4 by referring to [28, Lemma 3]. Reference [28] investigates
the control of inverter-based microgrids based on a networkpreserving model, while we extend some results to more
complicated synchronous-generator based bulk power systems.
For simplicity, we first present some notations following [28].
Comparing (1) and (4), Pei , Qei have same structures with
Pij , Qij , respectively. We can treat the reactance of generator
0
as a line with admittance as 1/xdi , which connects i ∈ NG
and inner node of the generator. We denote the inner nodes
0
of generators as NG . Then, we can get a augmented power
network, the incidence matrix of which is denoted as Ĉ. The
set of nodes in the augmented power network is denoted as
0
0
0
N = N ∪ NG . Denote V̂ = (Eq , V ), θ̂ = (δ, θ). Let |Ĉ| denote the matrix obtained from Ĉ by replacing all the elements
0
cij with |cij |. Define Γ(V̂ ) :=
Define A as
−|Bij | cos(θi − θj ),
diag(|Bii |),
Aij =
0
−cos(δi − θj )/xdi ,
diag(|Bij |Vi Vj ), i, j ∈ N .
i 6= j, i, j ∈ N ;
i = j, i, j ∈ N ;
0
i ∈ NG , j ∈ NG .
(B.12)
For simplicity, we use the following notation. For an ndimensional vector r := {r1 , r2 , · · · , rn }, the diagonal matrix diag(r1 , r2 , · · · , rn ) is denoted in short by [r]D . And
cos(·), sin(·) are defined component-wise.
From the definition of W in (22), ∇2v W (v) > 0 if and only
if ∇2v Wp (v) > 0, i.e. the matrix
Γ(V̂ )[cos(Ĉ T θ̂)]D
[sin(Ĉ T θ̂)]D Γ(V̂ )|Ĉ|T [V̂ ]−1
D
T
[V̂ ]−1
A + H(V̂ )
D |Ĉ|Γ(V̂ )[sin(Ĉ θ̂)]D
(B.13)
is positive definite, where
" h
H(V̂ ) =
xdi
0
0
2xdi (xdi −xdi )
i
D
0
0
[qi /Vi2 ]D
#
(B.14)
In any steady state of power system (i.e.), the phase-angle
difference between two neighboring nodes is usually small.
In addition, the difference between δi and θi is also small.
This implies that the matrix in (B.13) is diagonal dominant as
well as its positive definiteness. Therefore, Assumption A4 is
usually satisfied and makes sense.
C. Proof of Lemma 4
Proof. From żij = 0, we get µ∗i = µ∗j = µ0 . Set µ̇i = 0 and
add (23) for i ∈ NCG , we have
X
X
Mi ω̇i +
(Di + τi )ωi = 0
(C.1)
i∈NCG
i∈NCG
The right sides of (7a) and (7b) vanish in the equilibrium
points, which implies ωi∗ = ω̃i∗ = ω̃j∗ = ω0 . Set ω̇ = 0 in
(C.1) and we have
X
ω0
(Di + τi ) = 0
(C.2)
i∈NCG
This implies that ω0 = 0 due to Di + τi > 0.
Other proofs are same with that in Theorem 2, which are
omitted.
D. Proof of Theorem 5
Proof of Theorem 5. We still use the Lyapunov function (22)
to analyze the stability of the closed-loop system (24). Denote
y = (ωiT , xT1 , xT2 )T , i ∈ NCG , and define the following
function
−Di ωi
0
g
−
+
− fi (Pi ) + µi − γi + γi
f
˜
µi
f (y) =
, i ∈ NCG
µ
−
µ
i
j
P gi − Pig
g
Pig − P i
(D.1)
JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX XXXX
where fµi = Pig − p̂i −
0
P
(µi − µj ) −
j∈Nci
P
12
zij − τi µi −
j∈Nci
τi fi (Pig ) + βi ωi 1 .
The derivative of Wk is
X
Mi (ωi − ωi∗ )ω̇i + (x − x∗ )T · K −1 ẋ
Ẇk =
i∈NG
=
X
∗
(ωi − ωi∗ )(Pig − Pig∗ − Di (ωi − ωi∗ ) − (Pei − Pei
))
i∈NG
+ (x − x∗ )T · K −1 ẋ
X
∗
))
(ωi − ωi∗ )(Pig − Pig∗ − (Pei − Pei
≤
i∈NG
−
X
Di (ωi − ωi∗ )2 + (y − y ∗ )T f˜(y)
(D.2)
i∈NU G
where the inequality is due to the same reason for that in (B.4).
Divide Ẇk into two parts, and Ẇk1 = (y −y ∗ )T f˜(y), Ẇk2 =
Ẇk − Ẇk1 . Then we will analyze the sign of Ẇk1 .
Ẇk1 = (y − y ∗ )T f˜(y)
Z 1
∂
=
(y − y ∗ )T f˜(z̃(s))(y − y ∗ )ds + (y − y ∗ )T f˜(y ∗ )
∂ z̃
0
T
Z
1 1
∂ ˜
∂ ˜
≤
f (z̃(s)) +
f (z̃(s)) (y − y ∗ )ds
(y − y ∗ )T
2 0
∂ z̃
∂ z̃
Z 1
=
(y − y ∗ )T [H(z̃(s))] (y − y ∗ )ds
(D.3)
0
where z̃(s) = y ∗ + s(y − y ∗ ).RThe second equation is from
1
the fact that f˜(y) − f˜(y ∗ ) = 0 ∂∂z̃ f˜(z̃(s))(y − y ∗ )ds. The
inequality is due to either f˜(y ∗ ) = 0 or f˜(y ∗ ) < 0, yi ≥ 0,
i.e. (y − y ∗ )T f˜(y ∗ ) ≤ 0.
D
0
0
0
0 0
0
∇2P g f (P g )
I
0 −I I
˜
β τ ∇2 g f (P g ) − I τ + Lc C 0 0
∂ f (y)
P
= −
0
−C T
0
0 0
∂y
0
0
I
0
0
0 0
0
−I
0
0
0 0
(D.4)
where D = diag(Di ), β = diag(βi ), τ = diag(τi ), I is the
identity matrix with dimension nCG , C is the incidence matrix
of the communication graph, Lc is the Laplacian matrix of the
communication graph.
Finally, H in (D.3) is
∂ ˜
1 ∂T ˜
f (y) +
f (y)
H=
2 ∂y
∂y
−D
0
− 12 β
0 0 0
0
−∇2P g f (P g ) − τ2 ∇2P g f (P g ) 0 0 0
1
− β − τ ∇2 g f (P g )
−τ − Lc
0 0 0
2
2 P
=
0
0
0
0 0 0
0
0
0
0 0 0
0
0
0
0 0 0
(D.5)
1 Sometimes, we also use β instead of β (t) for simplification. In addition,
i
i
if (9) is not binding, we can omit γi− , γi+ in neighborhoods of equilibrium
points
H ≤ 0, if
−D
0
− 21 β
0
− 12 β
−∇2P g f (P g ) − τ2 ∇2P g f (P g ) < 0
− τ2 ∇2P g f (P g )
−τ
(D.6)
By Schur complement [29], we only need
−1 1
−Di
0
− 2 βi
−τi − − 21 βi − τ2i ci
<0
0
−ci
− τ2i ci
(D.7)
where ci = ∇2P g f (Pig ). Solving (D.7), we can get
i
p
p
− τi Di (4 − τi ci ) < βi < τi Di (4 − τi ci )
(D.8)
By A2, we know ci ≤ li , thus
p
p
τi Di (4 − τi li ) ≤ τi Di (4 − τi ci ),
p
p
− τi Di (4 − τi li ) ≥ − τi Di (4 − τi ci ).
Here we need τi > 0, 4 − τi li > 0, i.e. 0 < τi < 4/li . Finally,
we have
p
p
− τi Di (4 − τi li ) < βi < τi Di (4 − τi li )
(D.9)
p
i.e. β̄i < τi Di (4 − τi li ), implying (26).
Analysis of Ẇk2 and Ẇp as well as the convergence to a
point are same as those in the proof of Theorem 3, which are
omitted here.
| 3cs.SY
|
Wishart Mechanism for Differentially Private Principal Components Analysis
Wuxuan Jiang, Cong Xie and Zhihua Zhang
arXiv:1511.05680v2 [cs.CR] 19 Nov 2015
Department of Computer Science and Engineering
Shanghai Jiao Tong University, Shanghai, China
[email protected], {xcgoner, zhihua}@sjtu.edu.cn
Abstract
We propose a new input perturbation mechanism
for publishing a covariance matrix to achieve (, 0)differential privacy. Our mechanism uses a Wishart distribution to generate matrix noise. In particular, we
apply this mechanism to principal component analysis (PCA). Our mechanism is able to keep the positive semi-definiteness of the published covariance matrix. Thus, our approach gives rise to a general publishing framework for input perturbation of a symmetric positive semidefinite matrix. Moreover, compared
with the classic Laplace mechanism, our method has
better utility guarantee. To the best of our knowledge,
the Wishart mechanism is the best input perturbation
approach for (, 0)-differentially private PCA. We also
compare our work with previous exponential mechanism algorithms in the literature and provide near optimal bound while having more flexibility and less computational intractability.
1
Introduction
Plenty of machine learning tasks deal with sensitive information such as financial and medical data. A common
concern regarding data security arises on account of the
rapid development of data mining techniques. Several data
privacy definitions are proposed in the literature. Among
them differential privacy (DP) has been widely used (Dwork
et al. 2006). Differential privacy controls the fundamental
quantity of information that can be revealed with changing one individual. Beyond a concept in database security,
differential privacy has been used by many researchers to
develop privacy-preserving learning algorithms (Chaudhuri
and Monteleoni 2009; Chaudhuri, Monteleoni, and Sarwate
2011; Bojarski et al. 2014). Indeed, this class of algorithms
is applied to a large number of machine learning models including logistic regression (Chaudhuri and Monteleoni
2009), support vector machine (Chaudhuri, Monteleoni, and
Sarwate 2011), random decision tree (Bojarski et al. 2014),
etc. Accordingly, these methods can protect the raw data
even though the output and algorithm itself are published.
Differential privacy (DP) aims to hide the individual information while keeping basic statistics of the whole dataset.
Copyright c 2016, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
A simple idea to achieve this purpose is to add some special noise to the original model. After that, the attacker, who
has two outputs generated by slightly different inputs, cannot distinguish whether the output change comes from the
artificial noise or input difference. However, the noise might
influence the quality of regular performance of the model.
Thus, we should carefully trade off between privacy and utility.
No matter what the procedure is, a query, a learning algorithm, a game strategy or something else, we are able to
define differential privacy if this procedure takes a dataset as
input and returns the corresponding output. In this paper, we
study the problem of designing differential private principal
component analysis (PCA). PCA reduces the data dimension while keeping the optimal variance. More specifically,
it finds a projection matrix by computing a low rank approximation to the sample covariance matrix of the given data
points.
Privacy-preserving PCA is a well-studied problem in the
literature (Dwork et al. 2014; Hardt and Roth 2012; 2013;
Hardt and Price 2014; Blum et al. 2005; Chaudhuri, Sarwate, and Sinha 2012; Kapralov and Talwar 2013). It outputs a noisy projection matrix for dimension reduction while
preserving the privacy of any single data point. The extant
privacy-preserving PCA algorithms have been devised based
on two major features: the notion of differential privacy
and the stage of randomization. Accordingly, the privacypreserving PCA algorithms can be divided into distinct categories.
The notion of differential privacy has two types: (, 0)-DP
(also called pure DP) and (, δ)-DP (also called approximate
DP). (, δ)-DP is a weaker version of (, 0)-DP as the former
allows the privacy guarantee to be broken with tiny probability (more precisely, δ). In the seminal work on privacypreserving PCA (Dwork et al. 2014; Hardt and Roth 2012;
2013; Hardt and Price 2014; Blum et al. 2005), the authors used the notion of (, δ)-DP. In contrast, there is
only a few work (Chaudhuri, Sarwate, and Sinha 2012;
Kapralov and Talwar 2013), which is based on (, 0)-DP.
In terms of the stage of randomization, there are two
mainstream classes of approaches. The first is randomly
computing the eigenspace (Hardt and Roth 2013; Hardt and
Price 2014; Chaudhuri, Sarwate, and Sinha 2012; Kapralov
and Talwar 2013). The noise is added in the computing pro-
cedure. An alternative way is directly adding noise to the covariance matrix. Then one runs the non-private eigenspace
computing algorithm to produce the output. This class of
approaches is called input perturbation (Blum et al. 2005;
Dwork et al. 2014). The input perturbation algorithms publish a noisy sample covariance matrix before computing the
eigenspace. Thus, any further operation on the noisy covariance matrix does not violate privacy guarantee. So far as
the flexibility is concerned, the input perturbation has better performance because it is not limited only to computing
eigenspace. Besides, the input perturbation approach is efficient because it merely takes extra efforts on generating
the noise. In view of these advantages, our mechanism for
privacy-preserving PCA is also based on input perturbation.
Related Work
Blum et al. (2005) proposed an early input perturbation
framework (named SULQ), and the parameters of noise are
refined by Dwork et al. (2006). Dwork et al. (2014) proved
the state-of-the-art utility bounds for (, δ)-DP. Hardt and
Roth (2012) provided a better bound under the coherence assumption. In (Hardt and Roth 2013; Hardt and Price 2014),
the authors used a noisy power method to produce the principal eigenvector iteratively with removing the previous generated ones. Hardt and Price (2014) provided a special case
for (, 0)-DP as well.
Chaudhuri, Sarwate, and Sinha (2012) proposed the
first useful privacy-preserving PCA algorithm for (, 0)-DP
based on an exponential mechanism (McSherry and Talwar
2007). Kapralov and Talwar (2013) argued that the algorithm in (Chaudhuri, Sarwate, and Sinha 2012) lacks convergence time guarantee and used heuristic tests to check
convergence of the chain, which may affect the privacy guarantee. They also devised a mixed algorithm for low rank matrix approximation. However, their algorithm is quite complicated to implement and takes O(d6 /) running time. Here
d is the dimension of the data point.
Our work is mainly inspired by Dwork et al. (2014). Since
they provided the algorithms for (, δ)-DP, we seek the similar approach for (, 0)-DP with a different noise matrix design. As input perturbation methods, Blum et al. (2005) and
Dwork et al. (2014) both used the Gaussian symmetric noise
matrix for privately publishing a noisy covariance matrix. A
reasonable worry is that the published matrix might be no
longer positive semidefinite, a normal attribute for a covariance matrix.
Contribution and Organization
In this paper we propose a new mechanism for privacypreserving PCA that we call Wishart mechanism. The key
idea is to add a Wishart noise matrix to the original sample covariance matrix. A Wishart matrix is always positive
semidefinite, which in turn makes the perturbed covariance
matrix positive semidefinite. Additionally, Wishart matrix
can be regarded as the scatter matrix of some random Gaussian vectors (Gupta and Nagar 2000). Consequently, our
Wishart mechanism equivalently adds Gaussian noise to the
original data points.
Setting appropriate parameters of Wishart distribution, we
derive the (, 0)-privacy guarantee (Theorem 4). Compared
to the present Laplace mechanism, our Wishart mechanism
adds less noise (Section 4), which implies our mechanism
always has better utility bound. We also provide a general
framework for choosing Laplace or Wishart input perturbation for (, 0)-DP in Section 4.
Not only using the Laplace mechanism as a baseline,
we also conduct theoretical analysis to compare our work
with other privacy-preserving PCA algorithms based on the
(, 0)-DP. With respect to the different criteria, we provide sample complexity bound (Theorem 7) for comparison with Chaudhuri, Sarwate, and Sinha (2012) and derive
the low rank approximation closeness when comparing to
Kapralov and Talwar (2013). Other than the principal eigenvector guarantee in (Chaudhuri, Sarwate, and Sinha 2012),
we have the guarantee for rank-k subspace closeness (Theorem 6). With using a stronger definition of adjacent matrices,
we achieve a k-free utility bound (Theorem 9). Converting
the lower bound construction in (Chaudhuri, Sarwate, and
Sinha 2012; Kapralov and Talwar 2013) into our case, we
can see the Wishart mechanism is near-optimal.
The remainder of the paper is organized as follows. Section 2 gives the notation and definitions used in our paper.
Section 3 lists the baseline and our designed algorithms.
Section 4 provides the thorough analysis on privacy and utility guarantee of our mechanism together with comparison to
several highly-related work. Finally, we conclude the work
in Section 5. Note that we put some proofs and more explanation into the supplementary material.
2
Preliminaries
We first give some notation that will be used in this paper.
Let Im denote the m×m identity matrix. Given an m×n real
matrix Z = [Zij ], let its full singular value decomposition
(SVD) as Z = U ΣV T , where U ∈ Rm×m and V ∈ Rn×n
are orthogonal (i.e., U T U = Im and V T V = In ), and Σ ∈
Rm×n is a diagonal matrix with the ith diagonal entry σi
being the ith largest singular value of Z. Assume that the
rank of Z is ρ (≤ min(m, n)). This implies that Z has ρ
nonzero singular values. Let Uk and Vk be the first k (< ρ)
columns of U and V , respectively, and Σk be the k × k top
sub-block of Σ. Then the m × n matrix Zk = Uk Σk VkT is
the best rank-k approximation to Z.
The Frobenius
s norm of Z is defined as kZkF =
rP
ρ
P
2
Zij
=
σi2 , the spectral norm is defined as
i,j
kZk2 =
kZk∗ =
P
|Zij |.
i=1
kZxk
max kxk 2
2
x6=0
ρ
P
i=1
= σ1 , the nuclear norm is define as
σi , and the `1,1 norm is defined as kZk1,1 =
i,j
Given a set of n raw data points X = [x1 , · · · , xn ] where
xi ∈ Rd , we consider the problem of publishing a noisy empirical sample covariance matrix for doing PCA. Following
previous work on privacy-preserving PCA, we also assume
kxi k2 ≤ 1. The standard PCA computes the sample
Pn covari1
1
T
ance matrix of the raw data A = n XX = n i=1 xi xTi .
Since A is a d × d symmetric positive semidefinite matrix,
its SVD is equivalent to the spectral decomposition. That is,
A = V ΣV T . PCA uses Vk as projection matrix to compute
the low-dimensional representation of raw data: Y , VkT X.
In this work we use Laplace and Wishart distributions,
which are defined as follows.
Definition 1. A random variable z is said to have a Laplace
distribution z ∼ Lap(µ, b), if its probability density function
is
1
|z − µ|
p(z) =
exp(−
).
2b
b
Definition 2 ((Gupta and Nagar 2000)). A d × d random
symmetric positive definite matrix W is said to have a
Wishart distribution W ∼ Wd (m, C), if its probability density function is
p(W ) =
|W |
2
md
2
m−d−1
2
m
|C| 2 Γd ( m
2)
1
exp(− tr(C −1 W )),
2
where m > d − 1 and C is a d × d positive definite matrix.
Now we introduce the formal definition of differential privacy.
Definition 3. A randomized mechanism M takes a dataset
D as input and outputs a structure s ∈ R, where R is the
range of M . For any two adjacent datasets D and D̂ (with
only one distinct entry), M is said to be (, 0)-differential
private if for all S ⊆ R we have
Pr{M (D) ∈ S} ≤ e Pr{M (D̂) ∈ S},
where > 0 is a small parameter controlling the strength of
privacy requirement.
This definition actually sets limitation on the similarity
of output probability distributions for the given similar inputs. Here the adjacent datasets can have several different
interpretations. In the scenario of privacy-preserving PCA,
our definition is as follows. Two datasets X and X̂ are adjacent provided X = [x1 , . . . , xi−1 , xi , xi+1 , . . . , xn ] and
X̂ = [x1 , . . . , xi−1 , x̂i , xi+1 , . . . , xn ] for xi 6= x̂i . It should
be pointed out that our definition of adjacent datasets is
slightly different from (Kapralov and Talwar 2013), which
leads to significant difference on utility bounds. We will give
more specifically discussions in Section 4.
We also give the definition of (, δ)-differential privacy.
This notion requires less privacy protection so that it often
brings better utility guarantee.
Definition 4. A randomized mechanism M takes a dataset
as input and outputs a structure s ∈ R, where R is the range
of M . For any two adjacent datasets D and D̂ (with only
one distinct entry), M is said to be (, δ)-differential private
if for all S ⊆ R we have
Pr[M (D) ∈ S] ≤ e Pr[M (D̂) ∈ S] + δ.
Sensitivity analysis is a general approach to achieving differential privacy. The following definitions show the two
typical kinds of sensitivity.
Definition 5. The `1 sensitivity is defined as
s1 (M ) =
max
M (D) − M (D̂)
.
1
d(D,D̂)=1
The `2 sensitivity is defined as
s2 (M ) =
max
M (D) − M (D̂)
d(D,D̂)=1
.
2
The sensitivity describes the possible largest change as a
result of individual data entry replacement. The `1 sensitivity is used in Laplace Mechanism for (, 0)-differential privacy, while the `2 sensitivity is used in Gaussian Mechanism
for (, δ)-differential privacy. We list the two mechanisms
for comparison.
Theorem 1 (Laplace Mechanism). Let λ > s1 (M )/. Add
Laplace noise Lap(0, λ) to each dimension of M (D). This
mechanism provides (, 0)-differential privacy.
Theorem 2 (Gaussian Mechanism). For c2 > 2 ln(1.25/δ),
let σ > cs2 (M )/. Add Gaussian noise N (0, σ 2 ) to
each dimension of M (D). This mechanism provides (, δ)differential privacy.
The above mechanisms are all perturbation methods. Another widely used method is exponential mechanism (McSherry and Talwar 2007) which is based on sampling techniques.
3
Algorithms
First we look at a general framework of privacy-preserving
PCA. According to the definition of differential privacy, a
privacy-preserving PCA takes the raw data matrix X as input and then calculates the sample covariance matrix A =
1
T
n XX . Finally, it computes the top-k subspace of A as the
output.
The traditional approach adds noise to the computing procedure. For example, Chaudhuri, Sarwate, and Sinha (2012)
and Kapralov and Talwar (2013) used a sampling based
mechanism during computing eigenvectors to obtain approximate results. Our mechanism adds noise in the first
stage, publishing A in a differential private manner. Thus,
our mechanism takes X as input and outputs A. Afterwards
we follow the standard PCA to compute the top-k subspace.
This can be seen as a differential private preprocessing procedure.
Our baseline is the Laplace mechanism (Algorithm 1 and
Theorem 1). To the best of our knowledge, Laplace mechanism is the only input perturbation method for (, 0)-DP
PCA. Since this private procedure ends before computing
the subspace, this shows M (D) = n1 DDT in sensitivity
definition.
Note that to make  be symmetric, we use a symmetric matrix-variate Laplace distribution in Algorithm 1. However, this mechanism cannot guarantee the positive semidefiniteness of Â, a desirable attribute for a covariance matrix. This motivates us to use a Wishart noise alternatively,
giving rise to the Wishart mechanism in Algorithm 2.
Algorithm 1 Laplace input perturbation
d×n
Input: Raw data matrix X ∈ R
; Privacy parameter ;
Number of data n;
2
2d
1: Draw d 2+d i.i.d. samples from Lap(0, n
), then form a
symmetric matrix L. These samples are put in the upper
triangle part. Each entry in lower triangle part is copied
from the opposite position.
2: Compute A = n1 XX T ;
3: Add noise  = A + L;
Output: Â;
Algorithm 2 Wishart input perturbation
Input: Raw data matrix X ∈ Rd×n ; Privacy parameter ;
Number of data n;
1: Draw a sample W from Wd (d + 1, C), where C has d
3
same eigenvalues equal to 2n
;
1
T
2: Compute A = n XX ;
3: Add noise  = A + W ;
Output: Â;
4
Analysis
In this section, we are going to conduct theoretical analysis of Algorithms 1 and 2 under the framework of differential private matrix publishing. The theoretical support has
two parts: privacy and utility guarantee. The former is the
essential requirement for privacy-preserving algorithms and
the latter tells how well the algorithm works against a nonprivate version. Chiefly, we list the valuable theorems and
analysis. All the technical proofs omitted can be found in
the supplementary material.
Privacy guarantee
We first show that both algorithms satisfy privacy guarantee. Suppose there are two adjacent datasets X =
[x1 , . . . , v, . . . , xn ] ∈ Rd×n and X̂ = [x1 , . . . , v̂, . . . , xn ] ∈
Rd×n where v 6= v̂ (i.e., only v and v̂ are distinct). Without
loss of generality, we further assume that each data vector
has the `2 norm at most 1.
Theorem 3. Algorithm 1 provides (, 0)-differential privacy.
This theorem can be quickly proved by some simple
derivations so we put the proof in the supplementary material.
Theorem 4. Algorithm 2 provides (, 0)-differential privacy.
Proof. Assume the outputs for the adjacent inputs X and
X̂ are identical (denoted A + W0 ). Here A = n1 XX T and
 = n1 X̂ X̂ T . We define the difference matrix ∆ , A−  =
1
T
T
n (vv − v̂v̂ ). Actually the privacy guarantee is to bound
the following term:
p(A + W = A + W0 )
p(Â + W = A + W0 )
p(W = W0 )
=
p(W = W0 + ∆)
=
p(W = W0 )
p(W = A + W0 − Â)
As W ∼ Wd (d + 1, C), we have that
exp[− 12 tr(C −1 W0 )]
p(W = W0 )
=
p(W = W0 + ∆)
exp[− 12 tr(C −1 (W0 + ∆))]
1
= exp[ tr(C −1 (W0 + ∆)) − tr(C −1 W0 )]
2
1
= exp[ tr(C −1 ∆)].
2
Then apply Von Neumann’s trace inequality: For matrices
A, B ∈ Rd×d , denote their ith-largest singular value as
Pd
σi (·). Then |tr(AB)| ≤ i=1 σi (A)σi (B). So that
d
1
1X
exp[ tr(C −1 ∆)] ≤ exp[
σi (C −1 )σi (∆)]
2
2 i=1
1
≤ exp[ C −1
2
2
(1)
k∆k∗ ].
Since ∆ = A − Â = n1 (vv T − v̂v̂ T ) has rank at most 2,
and by singular value inequality σi+j−1 (A + B) ≤ σi (A) +
σj (B), we can bound k∆k∗ :
n k∆k∗ ≤ σ1 (vv T ) + σ1 (−v̂v̂ T ) + max{σ1 (vv T )
+ σ2 (−v̂v̂ T ), σ2 (vv T ) + σ1 (−v̂v̂ T )}
= σ1 (vv T ) + σ1 (v̂v̂ T ) + max{σ1 (vv T ), σ1 (v̂v̂ T )}
≤ 3 max σ1 (vv T ) = 3 max vv T
=
2
3 max kvk2
2
≤ 3.
In Algorithm 2, the scale matrix C in Wishart distribu3
tion has d same eigenvalues equal to 2n
, which implies
2n
−1
.
Substituting
these
terms
in
Eq. (1) yields
C
=
3
2
p(A + W = A + W0 )
≤ exp[
p(Â + W = A + W0 )
1 2n 3
≤ exp[ ·
· ] = e .
2 3 n
1
C −1
2
2
k∆k∗ ]
Utility guarantee
Then we give bounds about how far the noisy results are
from optimal. Since the Laplace and Wishart mechanisms
are both input perturbation methods, their analyses are similar.
In order to ensure privacy guarantee, we add a noise matrix to the input data. Such noise may have effects on the
property of the original matrix. For input perturbation methods, the magnitude of the noise matrix directly determines
how large the effects are. For example, if the magnitude of
the noise matrix is even larger than data, the matrix after
perturbation is surely covered by noise. Better utility bound
means less noise added. We choose the spectral norm of the
noise matrix to measure its magnitude. Since we are investigating the privacy-preserving PCA problem, the usefulness
of the subspace of the top-k singular vectors is mainly cared.
The noise matrix in the Laplace mechanism is constructed
2
with d 2+d i.i.d random variables of Lap(2d/n). Using the
tail bound for an ensemble matrix in (Tao 2012), we have
that the spectral norm√of the noise matrix in Algorithm 1
satisfies kLk2 = O(2d d/n) with high probability.
Then we turn to analyze the Wishart mechanism. We use
the tail bound of the Wishart distribution in (Zhu 2012):
Lemma 1 (Tail Bound of Wishart Distribution). Let W ∼
Wd (m, C). Then for θ ≥ 0, with probability at most
d exp(−θ),
p
λ1 (W ) ≥ (m + 2mθ(r + 2) + 2θr)λ1 (C)
where r = tr(C)/ kCk2 .
In our settings that r = d and m = d + 1, we thus have
that with probability at most d exp(−θ),
p
λ1 (W ) ≥ (d + 1 + 2(d + 1)(d + 2)θ + 2θd)λ1 (C).
Let θ = c log d(c > 1). Then d exp(−θ) = d1−c . So we
can say with high probability
p
λ1 (W ) = O([d + 1 + 2(d + 1)(d + 2)θ + 2θd]λ1 (C)).
For convenience, we write
λ1 (W ) = O(d log dλ1 (C)) = O(3d log d/2n).
We can see that the spectral norm of noise matrix generated by the Wishart mechanism is√O(d log d/n) while the
Laplace mechanism requires O(d d/n). This implies that
the Wishart mechanism adds less noise to obtain privacy
guarantee. We list the present four input perturbation approaches for comparison. Compared to the state-of-the-art
results about (, δ) case (Dwork et al. 2014), our noise√magd
d
nitude of O( d log
n ) is obviously worse than their O( n ). It
can be seen as the utility gap between (, δ)-DP and (, 0)DP.
Table 1: Spectral norm of noise matrix in input perturbation.
Approach
Noise magnitude
Privacy
√
Laplace
O(d
(, 0)
√ d/n)
(Blum et al. 2005) O(d d log d/n)
(, δ)
Wishart
O(d √
log d/n)
(, 0)
(Dwork et al. 2014)
O( d/n)
(, δ)
General framework We are talking about the intrinsic difference between the Laplace and Wishart mechanisms. The key element is the difference matrix ∆ of
two adjacent matrices. Laplace mechanism adds a noise
matrix according to the `1 sensitivity, which equals to
max k∆k1,1 . Thus, the spectral norm of noise matrix is
√
O(max k∆k1,1 d/n). When it comes to the Wishart
mechanism, the magnitude of noise is determined by
kCk2 . For purpose of satisfying privacy guarantee, we take
kCk2 = ω(max k∆k∗ /n). Then the spectral norm of noise
matrix is O(max k∆k∗ d log d/n). Consequently, we obtain the following theorem.
Theorem 5. M is a d × d symmetric matrix generated by
some input. For two arbitrary adjacent inputs, the generated matrices are M and M̂ . Let ∆ = M − M̂ . Using the
Wishart mechanism to publish M in differential private manner works better if
max k∆k1,1
max k∆k∗
√
= ω( d log d);
otherwise the Laplace mechanism works better.
Top-k subspace closeness We now conduct comparison
between our mechanism and the algorithm in (Chaudhuri,
Sarwate, and Sinha 2012). Chaudhuri, Sarwate, and Sinha
(2012) proposed an exponential-mechanism-based method,
which outputs the top-k subspace by drawing a sample from
the matrix Bingham-von Mises-Fisher distribution. Wang,
Wu, and Wu (2013) applied this algorithm to private spectral analysis on graph and showed that it outperforms the
Laplace mechanism for output perturbation. Because of
the scoring function defined, it is hard to directly sample
from the original Bingham-von Mises-Fisher distribution.
Instead, Chaudhuri, Sarwate, and Sinha (2012) used Gibbs
sampling techniques to reach an approximate solution. However, there is no guarantee for convergence. They check the
convergence heuristically, which may affect the basic privacy guarantee.
First we provide our result on the top-k subspace closeness:
Theorem 6. Let Vˆk be the top-k subspace of A+W in Algorithm 2. Denote the non-noisy subspace as Vk corresponding
to A. Assume σ1 ≥ σ2 ≥ · · · ≥ σd are singular values of A.
If σk − σk+1 ≥ 2 kW k2 , then with high probability
√
2 k kW k2
T
T
Vk Vk − V̂k V̂k
≤
.
σk − σk+1
F
We apply the well-known Davis-Kahan sin θ theorem (Davis 1963) to obtain this result. This theorem characterizes the usefulness of our noisy top-k subspace. Nevertheless, Chaudhuri, Sarwate, and Sinha (2012) only provided the utility guarantee on the principal eigenvector. So
we can only compare the top-1 subspace closeness, correspondingly.
Before the comparison, we introduce the measure in
(Chaudhuri, Sarwate, and Sinha 2012).
Definition 6. A randomized algorithm A(·) is an (ρ, η)close approximation to the top eigenvector if for all data
sets D of n points, output a vector vˆ1 such that
P(hvˆ1 , v1 i ≥ ρ) ≥ 1 − η.
Under this measure, we derive the sample complexity of
the Wishart mechanism.
Theorem 7. If n ≥
q
3(d+1+ 2(d+1)(d+2) log
d
d
η +2d log η )
2
2(1−ρ )(λ1 −λ2 )
√
and
ρ ≥ 22 , then the Wishart mechanism is a (ρ, η)-close approximation to PCA.
Because a useful algorithm should output
an eigenvector
√
2
making ρ close to 1, our condition of ρ ≥ 2 is quite weak.
Comparing to the sample complexity bound of the algorithm
in (Chaudhuri, Sarwate, and Sinha 2012):
1
Theorem
8.
If
n
≥
log η
d
(1−ρ)(λ1 −λ2 ) ( d
+
1
log (1−ρ24λ
(Chaudhuri,
)(λ1 −λ2 ) ), then the algorithm in
Sarwate, and Sinha 2012) is a (ρ, η)-close approximation
to PCA.
Our result has a factor up to log d with dropping the term
1
1
. Actually, the relationship between d and λ1λ−λ
log λ1λ−λ
2
2
heavily depends on the data. Thus, as a special case of topk subspace closeness, our bound for the top-1 subspace is
comparable to Chaudhuri, Sarwate, and Sinha’s (2012).
Low rank approximation Here we discuss the comparison between the Wishart mechanism and privacy-preserving
rank-k approximation algorithm proposed in (Kapralov and
Talwar 2013; Hardt and Price 2014). PCA can be seen
as a special case of low rank approximation problems.
Kapralov and Talwar (2013) combined the exponential and
Laplace mechanisms to design a low rank approximation algorithm for a symmetric matrix, providing strict guarantee
on convergence. However, the implementation of the algorithm contains too many approximation techniques and it
takes O(d6 /) time complexity while our algorithm takes
O(kd2 ) running time. Hardt and Price (2014) proposed an
efficient meta algorithm, which can be applied to (, δ)differentially private PCA. Additionally, they provided a
(, 0)-differentially private version.
We need to point out that the definition of adjacent matrix
in privacy-preserving low rank approximation is different
from ours (our definition is the same as (Dwork et al. 2014;
Chaudhuri, Sarwate, and Sinha 2012)). In the definition
(Kapralov and Talwar 2013; Hardt and Price 2014), two matrices A and B are called adjacent if kA − Bk2 ≤ 1, while
we restrict the difference to a certain form vv T − v̂v̂ T . In
fact, we make a stronger assumption so that we are dealing
with a case of less sensitivity. This difference impacts the
lower bound provided in (Kapralov and Talwar 2013).
For the consistence of comparison, we remove the term
1
T
n in Algorithm 2, which means we use the XX for PCA
1
instead of n XX T . This is also used by Dwork et al. (2014).
Applying Lemma 1 in (Achlioptas and McSherry 2001),
we can immediately have the following theorem:
Theorem 9. Suppose the original matrix is A = XX T and
Aˆk is the rank-k approximation of output by the Wishart
mechanism. Denote the k-th largest eigenvalue of A as λk .
Then
d log d
A − Âk ≤ λk+1 + O(
).
2
3
Kapralov and Talwar (2013) provided a bound of O( k d )
and Hardt and Price (2014) provided O( k
3
2
d log2 d
)
same scenario. If k 3 is larger than log d, our algorithm will
work better. Moreover, our mechanism has better bounds
than that of Hardt and Price (2014) while both algorithms
are computationally efficient. Kapralov and Talwar (2013)
established a lower bound of O( kd
) according to their definition of adjacent matrix. If replaced with our definition, the
lower bound will become O( d ). The details will be given
in the in supplementary material. So our mechanism is nearoptimal.
for the
5
Concluding Remarks
We have studied the problem of privately publishing a
symmetric matrix and provided an approach for choosing
Laplace or Wishart noise properly. In the scenario of PCA,
our Wishart mechanism adds less noise than the Laplace,
which leads to better utility guarantee. Compared with the
privacy-preserving PCA algorithm in (Chaudhuri, Sarwate,
and Sinha 2012), our mechanism has reliable rank-k utility guarantee while the former (Chaudhuri, Sarwate, and
Sinha 2012) only has rank-1. For rank-1 approximation
we have the comparable performance on sample complexity. Compared with the low rank approximation algorithm
in (Kapralov and Talwar 2013), the bound of our mechanism does not depend on k. Moreover, our method is more
tractable computationally. Compared with the tractable algorithm in (Hardt and Price 2014), our utility bound is better.
Since input perturbation only publishes the matrix for
PCA, any other procedure can take the noisy matrix as input. Thus, our approach has more flexibility. While other
entry-wise input perturbation techniques make the covariance not be positive semidefinite, in our case the noisy covariance matrix still preserves this property.
Acknowledgments
We thank Luo Luo for the meaningful technical discussion. We also thank Yujun Li, Tianfan Fu for support on the
early stage of the work. This work is supported by the National Natural Science Foundation of China (No. 61572017)
and the Natural Science Foundation of Shanghai City (No.
15ZR1424200).
References
Achlioptas, D., and McSherry, F. 2001. Fast computation
of low rank matrix approximations. In Proceedings of the
thirty-third annual ACM symposium on Theory of computing, 611–618. ACM.
Blum, A.; Dwork, C.; McSherry, F.; and Nissim, K. 2005.
Practical privacy: the sulq framework. In Proceedings of the
twenty-fourth ACM SIGMOD-SIGACT-SIGART symposium
on Principles of database systems, 128–138. ACM.
Bojarski, M.; Choromanska, A.; Choromanski, K.; and LeCun, Y. 2014. Differentially-and non-differentially-private
random decision trees. arXiv preprint arXiv:1410.6973.
Chaudhuri, K., and Monteleoni, C. 2009. Privacy-preserving
logistic regression. In Advances in Neural Information Processing Systems, 289–296.
Chaudhuri, K.; Monteleoni, C.; and Sarwate, A. D. 2011.
Differentially private empirical risk minimization. The Journal of Machine Learning Research 12:1069–1109.
Chaudhuri, K.; Sarwate, A.; and Sinha, K. 2012. Nearoptimal differentially private principal components. In Advances in Neural Information Processing Systems, 989–997.
Davis, C. 1963. The rotation of eigenvectors by a perturbation. Journal of Mathematical Analysis and Applications
6(2):159–173.
Dwork, C.; McSherry, F.; Nissim, K.; and Smith, A. 2006.
Calibrating noise to sensitivity in private data analysis. In
Theory of cryptography. Springer. 265–284.
Dwork, C.; Talwar, K.; Thakurta, A.; and Zhang, L. 2014.
Analyze gauss: optimal bounds for privacy-preserving principal component analysis. In Proceedings of the 46th Annual
ACM Symposium on Theory of Computing, 11–20. ACM.
Gupta, A. K., and Nagar, D. K. 2000. Matrix Variate Distributions. Chapman & Hall/CRC.
Hardt, M., and Price, E. 2014. The noisy power method:
A meta algorithm with applications. In Advances in Neural
Information Processing Systems, 2861–2869.
Hardt, M., and Roth, A. 2012. Beating randomized response
on incoherent matrices. In Proceedings of the forty-fourth
annual ACM symposium on Theory of computing, 1255–
1268. ACM.
Hardt, M., and Roth, A. 2013. Beyond worst-case analysis
in private singular vector computation. In Proceedings of the
forty-fifth annual ACM symposium on Theory of computing,
331–340. ACM.
Kapralov, M., and Talwar, K. 2013. On differentially private low rank approximation. In Proceedings of the TwentyFourth Annual ACM-SIAM Symposium on Discrete Algorithms, 1395–1414. SIAM.
McSherry, F., and Talwar, K. 2007. Mechanism design via
differential privacy. In Foundations of Computer Science,
2007. FOCS’07. 48th Annual IEEE Symposium on, 94–103.
IEEE.
Tao, T. 2012. Topics in random matrix theory, volume 132.
American Mathematical Soc.
Wang, Y.; Wu, X.; and Wu, L. 2013. Differential privacy
preserving spectral graph analysis. In Advances in Knowledge Discovery and Data Mining. Springer. 329–340.
Zhu, S. 2012. A short note on the tail bound of wishart
distribution. arXiv preprint arXiv:1212.5860.
Supplementary material
A
Proof of privacy guarantee
The basic settings are the same as section 4.
Proof of Theorem 3
In order to prove Theorem 3, we first give the following
lemma.
Lemma 2. For mechanism M (D) =
tivity s1 satisfies
1
T
n DD ,
the `1 sensi-
d
2d
< s1 (M ) <
.
n
n
Proof. Suppose v = (p1 , · · · , pd )T and v̂ = (q1 , · · · , qd )T .
Then the `1 sensitivity of M (D) can be converted to the
following optimization problem:
1 X
s1 (M ) = max
|pi pj − qi qj |,
n
1≤i,j≤d
subject to
d
X
p2i ≤ 1,
i=1
Setting pi =
√1
d
d
X
qi2 ≤ 1.
i=1
and qi = 0 for i = 1, . . . , d, we can have
a lower bound s1 (M ) ≥ nd . Then applying the triangle inequality, we have the upper bound:
X
X
|pi pj − qi qj | <
|pi pj | + |qi qj |
1≤i,j≤d
1≤i,j≤d
X
=2
|pi pj | ≤
1≤i,j≤d
2d
.
n
Now applying Lemma 2 to Theorem 1 immediately obtains the privacy guarantee for the Laplace mechanism.
B
Proof of utility guarantee
Proof of Theorem 6
Proof. We use the following two lemmas.
Lemma 3 (Davis-Kahan sin θ theorem (Davis 1963)). Let
the k-th eigenvector of A and  be vk and vˆk . Denote Pk =
k
k
P
P
vk vkT and Pˆk =
vˆk vˆk T . If λk (A) > λk+1 (Â), then
i=1
i=1
Pk − Pˆk
A − Â
≤
2
2
λk (A) − λk+1 (Â)
.
Lemma 4 (Weyl’s inequality). If M , H and P are d × d
Hermitian matrices such that M = H + P . Let the k-th
eigenvalues of M , H and P be µk , νk and ρk , respectively.
For i ∈ [n], we have
νi + ρd ≤ µi ≤ νi + ρ1 .
In our case, A and  are both symmetric positive semidefinite (because of the property of Wishart distribution). So the
eigenvalues equal to singular values. Then we use Lemma 4
with A = H and W = P . We obtain
σi (A + W ) ≤ σi (A) + σ1 (W ) = σi (A) + kW k2 .
Applying Lemma3 with A = A and  = A + W leads to
Pk − Pˆk
2
= Vk VkT − V̂k V̂kT
2
≤
2 kW k2
.
σk − σk+1
≤
F
√
k Vk VkT − V̂k V̂kT
2
we finish the proof.
Proof of Theorem 7
We are going to find the condition on sample complexity to
satisfy (ρ, η)-close approximation.
Proof. Set k = 1 in Theorem 6. Then
V1 V1T − V̂1 V̂1T
F
= v1 v1T − v̂1 v̂1T
=
tr[(v1 v1T
−
F
v̂1 v̂1T )(v1 v1T
= 2 − 2(v1T v̂1 )2 ≤
− v̂1 v̂1T )T ]
2 kW k2
.
λ1 − λ2
The condition λ1 − λ2 ≥ 2 kW k2 requires the last term
√
to have a upper bound of 1, which implies |v1T v̂1 | ≥ 22 .
Let η = d exp(−θ), which is θ = log ηd , we have that with
probability 1 − η,
kW k2
≥1−
λ1 − λ2
p
(d + 1 + 2(d + 1)(d + 2)θ + 2θd)λ1 (C)
=1−
λ1 − λ2
q
3(d + 1 + 2(d + 1)(d + 2) log ηd + 2d log ηd )
=1−
.
2n(λ1 − λ2 )
(v1T v̂1 )2
Under theqcondition
n≥
3(d+1+
Which
2(d+1)(d+2) log
2(1−ρ2 )(λ
d
d
η +2d log η )
1 −λ2 )
(v1T v̂1 )2 ≥ 1 − (1 − ρ2 ) =
yields Pr(v1T v̂1 ≥ ρ) ≥ 1 − η.
ρ2
Lemma 5. Define Cδk (Y ) = {S ∈ Gk,d :
Y Y T − SS T 2 ≤ δ}. For each δ > 0 there exists family F = {Y 1 , · · · , Y N } with N = 2Ω(k(d−k) log 1/δ) , where
Y i ∈ Gk,d such that Cδk (Y i ) ∩ Cδk (Y j ) = ∅ for i 6= j.
2
≤ λk+1 + Ω(d/).
Proof. Take a set F = {Y 1 , · · · , Y N } in Lemma 5. Construct a series of matrices Ai = γYi YiT where i ∈ [N ].
Then
h
i
EAi Ai − Aˆik
≤ δγ.
2
Using the property
Vk VkT − V̂k V̂kT
We mainly follow the construction of Kapralov and Talwar
(2013) and make a slight modification to fit into our definition of adjacent matrices.
A − Aˆk
Under the assumption σk −σk+1 ≥ 2 kW k2 , we finally have
2
Lower bound for low rank approximation
Theorem 10. Suppose the original matrix is A = XX T
and Aˆk is the rank-k approximation of output by the any
-differential private mechanism. Denote the k-th largest
eigenvalue of A as λk . Then
kW k2
≤
λk (A) − λk+1 (A + W )
kW k2
≤
λk (A) − λk+1 (A) − kW k2
kW k2
=
.
σk − σk+1 − kW k2
Vk VkT − V̂k V̂kT
C
T
ei = Ŷi Ŷi T , we have
Let Aˆik = Ŷi Σ̂Ŷi . Then letting A
k
i
h
i
i
ek
≤ 2δγ.
EAi A − A
2
Using Markov’s inequality leads to
h
i
eik ≤ 4δγ > 1 .
Pri Ai − A
2
A
2
Here is the main difference between our definition and
Kapralov and Talwar (2013). They consider the distance
from Ai to Aj is at most 2γ since Ai 2 ≤ γ. In our framework, Ai is a dataset consisting of γ data groups, each one is
Y i . Changing Ai to Aj means replacing γk data points with
brand new ones. So we consider the distance is at most 2γk.
The algorithm should put at least half of the probability
k
(Y i ). Meanwhile, to satisfy the privacy guarmass into C4δ
antee
k
(Y i )}
Pr{M (Ai ) ∈ C4δ
≤ e2γk
k (Y i )}
Pr{M (Aj ) ∈ C4δ
k
So Pr{M (Aj ) ∈ C4δ
(Y i )} ≥ 21 e−2γk , we have
1 −2γk Ω(k(d−k) log 1/δ)
e
·2
≤1
2
Which implies γ = Ω(d log(1/δ)/) and completes our
proof.
| 8cs.DS
|
1
Refined Asymptotics for Rate-Distortion using
Gaussian Codebooks for Arbitrary Sources
Lin Zhou, Vincent Y. F. Tan and Mehul Motani
arXiv:1708.04778v2 [cs.IT] 26 Dec 2017
Abstract
The rate-distortion saddle-point problem considered by Lapidoth (1997) consists in finding the minimum rate to compress an
arbitrary ergodic source when one is constrained to use a random Gaussian codebook and minimum (Euclidean) distance encoding
is employed. We extend Lapidoth’s analysis in several directions in this paper. Firstly, we consider refined asymptotics. In particular,
when the source is stationary and memoryless, we establish the second-order, moderate, and large deviation asymptotics of the
problem. Secondly, by “random Gaussian codebook”, Lapidoth refers to a collection of random codewords, each of which is drawn
independently and uniformly from the surface of an n-dimensional sphere. To be more precise, we term this as a spherical Gaussian
codebook. We also consider i.i.d. Gaussian codebooks in which each random codeword is drawn independently from a product
Gaussian distribution. We derive the second-order, moderate, and large deviation asymptotics when i.i.d. Gaussian codebooks are
employed. Interestingly, in contrast to the recent work on the channel coding counterpart by Scarlett, Tan and Durisi (2017), the
dispersions for spherical and i.i.d. Gaussian codebooks are identical. The optimal excess-distortion exponents for both spherical
and i.i.d. Gaussian codebooks are established for all rates. Furthermore, we prove that the i.i.d. Gaussian codebook has a strictly
larger excess-distortion exponent than the spherical counterpart for any rate larger the first order coding rate.
Index Terms
Lossy data compression, Rate-distortion, Gaussian codebook, Mismatched encoding, Minimum distance encoding, Ensemble
tightness, Second-order asymptotics, Dispersion, Moderate deviations, Large deviations
I. I NTRODUCTION
In the traditional lossy data compression problem [1, Section 3.6], one seeks to find the minimum rate of compression
of a source while allowing it to be reconstructed to within a distortion D at the output of the decompressor. Shannon [2]
established the rate-distortion function for stationary and memoryless sources. However, practical considerations on the system
design often necessitate a particular encoding strategy. This then constitutes a mismatch problem in which the codebook is
optimized for a source with one distribution but used to compress a source of a different distribution. For example, one might be
interested to use a Gaussian codebook—a codebook that is optimal for a memoryless Gaussian source—to compress a source
that is arbitrary. For all ergodic sources with second moment σ 2 , Lapidoth [3, Theorem 3] established that the (ensemble)
rate-distortion function is
1
σ2
∗
Rσ2 (D) = log max 1,
.
(1)
2
D
The term “Gaussian codebook” requires some qualifications; Lapidoth [3] uses this term to mean a collection of random
codewords each of which is drawn independently and uniformly from the surface of a sphere in n-dimensions. In this work, we
term this random codebook as a spherical Gaussian codebook and, for the sake of comparison, we also consider i.i.d. Gaussian
codebooks in which each component of each codeword is drawn independently from a (univariate) Gaussian distribution. In the
spirit of recent emphases on refined asymptotics that bring to light the tradeoff between the coding rate, the blocklength, and
the probability of excess-distortion, in this paper, we establish ensemble-tight second-order coding rates, moderate deviations
constants and excess-distortion exponents.
A. Main Contributions and Related Works
Our main contributions are as follows:
(i) We conduct a second-order asymptotic analysis [4]–[6] for the rate-distortion saddle-point problem for stationary and
memoryless sources that satisfy mild technical conditions. Here, the probability of excess-distortion is allowed to be
non-vanishing and the spotlight is shone on the additional rate, above the rate-distortion function, required at finite
blocklengths to compress the source to within the prescribed probability of excess-distortion. This work complements
that of Kostina and Verdú [7] and Ingber and Kochman [8] who established the second-order asymptotics (or dispersion)
for compressing (discrete and Gaussian) memoryless sources when the encoder is unconstrained. Our main result here
is somewhat surprising. We show that the mismatched dispersions (and the first-order coding rates or rate-distortion
functions) for spherical and i.i.d. Gaussian codebooks are identical; this implies that there is no performance loss in
The authors are with the Department of Electrical and Computer Engineering, National University of Singapore (Emails: [email protected], [email protected],
[email protected]).
2
terms of the backoff from the first-order fundamental limit regardless of which type of Gaussian codebook one uses.
This is in stark contrast to the recent work by Scarlett, Tan, and Durisi [9] on the channel coding counterpart of this
problem [10]. It was shown in [9] that the dispersions for both types of codebooks are different. For the lossy source
coding case, the dispersions are common and depend only on the second and fourth moments of the source through a
simple formula. We provide intuition for why this is the case after the statement of Theorem 1. We recover the dispersion
of lossy compression of Gaussian memoryless sources (GMSes) [7], [8] by particularizing the arbitrary source to be a
Gaussian.
(ii) Next, we conduct moderate deviations analysis [11], [12] for the same problem under an additional, albeit mild, assumption
on the source. Here, the rate of the codebook approaches the rate-distortion function at a speed slower than the reciprocal of
the square root of the blocklength. One then seeks the subexponential rate of decay of the probability of excess-distortion.
This analysis complements that of Tan [13] who considered the unconstrained encoding case for (discrete and Gaussian)
memoryless sources. This was generalized to the successive refinement problem by the present authors [14]. We again
show that the moderate deviations constants are identical and that for GMSes can be easily recovered.
(iii) Finally, we consider the large deviations regime [15], [16] in which the rate of the codebook is constrained to be
strictly above the rate-distortion function and one seeks to establish the exponential rate of decay of the probability
of excess-distortion. Our analysis complements that of Ihara and Kubo [17] who used ideas from Marton [18] to find
the excess-distortion exponent for compressing a GMS. We establish the optimal excess-distortion exponents for both
spherical and i.i.d. Gaussian codebooks. We recover the excess-distortion exponent of lossy compression of GMSes [17]
by particularizing the source to be a Gaussian when an i.i.d. Gaussian codebook is used. Furthermore, we show that
the i.i.d. Gaussian codebook has a strictly larger excess-distortion exponent than the spherical counterpart for any rate
R > Rσ∗ 2 (D) (cf. (1)) and illustrate this point using two numerical examples.
B. Organization of the Rest of the Paper
The rest of the paper is organized as follows. We set up the notation, formulate our problem precisely, and present existing
results in Section II. In Section III, we present our main results. These include results concerning second-order, moderate, and
large deviation asymptotics. Sections IV to VI are devoted to the proofs for each of these asymptotic results respectively.
II. T HE R ATE -D ISTORTION S ADDLE -P OINT P ROBLEM
A. Notation
Random variables and their realizations are in upper (e.g., X) and lower case (e.g., x) respectively. All sets are denoted in
calligraphic font (e.g., X ). We use R, R+ , and N to denote the set of real numbers, non-negative real numbers, and natural
numbers respectively. For any two natural numbers a and b we use [a : b] to denote the set of all natural numbers between
a and b (inclusive). We use exp{x} to denote ex . All logarithms are base e. We use Q(·) to denote the standard Gaussian
complementary cumulative distribution function (cdf) and Q−1 (·) its inverse. For any random variable X, we use ΛX (θ) to
denote the cumulant generating function log E[exp{θX}] (where θ ∈ R). We use Λ∗X (t) (where t ∈ R) to denote the FenchelLegendre transform (convex conjugate) of the cumulant generating function, i.e., supθ≥0 {θt−ΛX (θ)}.p
Let X n := (X1 , . . . , Xn )
P 2
n
n
be a random vector of length n and x = (x1 , . . . , xn ) is a particular realization. We use kx k =
i xi to denote the ℓ2
n
n
norm of a vector xn ∈ Rn . Given two sequences
x
and
y
,
the
quadratic
distortion
measure
(squared
Euclidean norm) is
Pn
defined as d(xn , y n ) := n1 kxn − y n k2 = n1 i=1 (xi − yi )2 . For any two sequences {an }n≥1 and {bn }n≥1 , we write an ∼ bn
to mean limn→∞ an /bn = 1.
B. System Model
Consider arbitrary source X with distribution (probability mass function or probability density function) fX satisfying
E[X 2 ] = σ 2 , ζ := E[X 4 ] < ∞, E[X 6 ] < ∞.
(2)
In this paper, we consider memoryless sources and thus X n is an i.i.d. sequence where each component is generated according
to fX . We consider the rate-distortion saddle-point problem [3, Theorem 3] with an admissible distortion level 0 < D < σ 2 .
This is the lossy source coding problem [1, Section 3.6] where one is constrained to use random Gaussian codebooks (spherical
or i.i.d.) and an encoding strategy which chooses the codeword that minimizes the quadratic distortion measure.
Definition 1. An (n, M )-code for the rate-distortion saddle-point problem consists of
n
M
• A set of M codewords {Y (i)}i=1 known by both the encoder and decoder;
n
• An encoder f which maps the source sequence X into the index of the codeword that minimizes the quadratic distortion
n
with respect to the source sequence X , i.e.,
f (X n ) := arg min d X n , Y n (i) .
(3)
i∈[1:M]
3
•
A decoder φ which declares the reproduced sequence as the codeword with index f (X n ), i.e.,
φ(f (X n )) = Y n (f (X n )).
(4)
Throughout the paper, we consider random Gaussian codebooks. To be specific, we consider two types of Gaussian codebooks.
• First, we consider the spherical
Gaussian codebook where each codeword Y n is generated independently and uniformly
p
over a sphere with radius n(σ 2 − D), i.e.,
Y n ∼ fYspn (y n ) =
•
δ(ky n k2 − n(σ 2 − D))
p
,
Sn ( n(σ 2 − D))
(5)
where δ(·) is the Dirac delta function, Sn (r) = nπ n/2 rn−1 /Γ( n+2
2 ) is the surface area of an n-dimensional sphere
with radius r, and Γ(·) is the Gamma function. For GMSes, the spherical codebook is second-order optimal (cf. [7,
Theorem 40]).
Second, we consider the i.i.d. Gaussian codebook where each codeword Y n is generated independently according to the
following product Gaussian distribution with variance σ 2 − D, i.e.,
n
Y
1
yi2
n
iid
n
p
.
(6)
exp −
Y ∼ fY n (y ) =
2(σ 2 − D)
2π(σ 2 − D)
i=1
The i.i.d. Gaussian codebook is also second-order optimal (cf. [7, Theorem 12]) for a GMS.
Here, we mention in passing that the techniques herein may also be applicable to the truncated i.i.d. Gaussian codebook [16,
Eqns. (7.3.12)–(7.3.15)] where the codewords are generated from an i.i.d. Gaussian distribution as in (6) but the distribution
is then truncated to the annulus {y n ∈ Rn : σ 2 − D − nη < n1 ky n k2 ≤ σ 2 − D} where η is a fixed positive number. This
distribution is useful for deriving error exponent and second-order results. See [19], [20] for discussions.
The (ensemble) excess-distortion probability is defined as
Pe,n := Pr{d(X n , φ(f (X n ))) > D}
i
h
M
Xn ,
= EfXn 1 − Pr{d(X n , Y n ) ≤ D}
(7)
(8)
where (8) follows from [7, Theorem 9] and the inner probability (over Y n which is independent of X n ) is calculated either
with respect to the right hand side of (5) if we use a spherical Gaussian codebook or the right hand side of (6) if we use an
i.i.d. Gaussian codebook. Note that the probability in (7) is averaged over the source as well as the random codebook. This
is in contrast to the traditional lossy source coding analysis [7], [18] where the excess-distortion probability is averaged over
the source only. The additional average over the codebook allows us to pose questions concerning ensemble tightness in the
spirit of [9], [21].
C. Existing Results and Definitions
∗
Let Msp
(n, ε, σ 2 , D) be the minimum number of codewords required to compress a length-n source sequence so that the
excess-distortion probability with respect to distortion level D is no larger than ε ∈ (0, 1) when a spherical Gaussian codebook
∗
is used. Similarly, let Miid
(n, ε, σ 2 , D) be the corresponding quantity when an i.i.d. Gaussian codebook is used. Lapidoth [3,
Theorem 3] showed that for any ergodic source with finite second moment σ 2 and any ε ∈ (0, 1),
1
σ2
1
∗
log Msp
(n, ε, σ 2 , D) = log
nats per source symbol.
n→∞ n
2
D
As we will show via a by-product of Theorem 1, for any source satisfying (2) and any ε ∈ (0, 1), we also have
lim
(9)
1
1
σ2
∗
log Miid
(n, ε, σ 2 , D) = log
nats per source symbol.
(10)
n→∞ n
2
D
In this paper, we are interested in second-order, large, and moderate deviations analyses. These analyses provide a refined
understanding of the tradeoff between the rate, the blocklength and the excess-distortion probability. In the study of secondorder asymptotics, a non-vanishing excess-distortion probability is allowed and we aim to find the back-off from the first-order
2
coding rate (the rate-distortion function) 21 log σD .
lim
Definition 2. Fix any ε ∈ [0, 1). A number L is said to be a spherical-achievable second-order coding rate if there exists a
sequence of (n, M )-codes using spherical Gaussian codebooks such that
n
σ2
1
≤ L,
(11)
lim sup √ log M − log
n
2
D
n→∞
lim sup Pe,n ≤ ε.
(12)
n→∞
4
The infimum of all spherical-achievable second-order coding rates is the optimal spherical second-order coding rate and is
denoted as L∗sp (ε). In a similar manner, we define i.i.d.-achievable second-order coding rates and denote their infimum as the
optimal i.i.d. second-order coding rate L∗iid (ε).1
The moderate deviations regime interpolates between the large deviations (cf. Definition 4 to follow) and the second-order
regimes. In this regime, we are interested in a sequence of (n, M )-codes whose rates approach the first-order coding rate
1
σ2
2 log D and whose excess-distortion probabilities vanish simultaneously.
Definition 3. Consider any sequence {ξn }n∈N such that as n → ∞
r
n
ξn → 0 and
ξn → ∞.
log n
(13)
A non-negative number ν is said to be a spherical-achievable moderate deviations constant if there exists a sequence of
(n, M )-codes using spherical Gaussian codebooks such that
n
σ2
1
≤ 1,
(14)
log M − log
lim sup
2
D
n→∞ nξn
1
(15)
lim inf − 2 log Pe,n ≥ ν.
n→∞
nξn
The supremum of all spherical-achievable moderate deviations constants is the optimal spherical moderate deviations constant
∗
and is denoted as νsp
. Similarly, we can define i.i.d.-achievable moderate deviations constants and their supremum is the
∗
optimal i.i.d. moderate deviations constant νiid
.
In the large deviations regime, we characterize the speed of the exponential decay of the excess-distortion probability for
codes with a rate upper bounded by R.
Definition 4. A non-negative number E is said to be a rate-R spherical-achievable excess-distortion exponent if there exists
a sequence of (n, M )-codes using spherical Gaussian codebooks such that
lim sup
n→∞
1
log M ≤ R,
n
(16)
1
log Pe,n ≥ E.
(17)
n
The supremum of all rate-R spherical-achievable excess-distortion exponents is denoted as the optimal spherical excess∗
∗
distortion exponent Esp
(R). Similarly, we can define the optimal i.i.d. excess-distortion exponent Eiid
(R).
lim inf −
n→∞
III. M AIN R ESULTS
A. Second-Order Asymptotics
Our first result pertains to the second-order coding rate. Recall the definitions of σ 2 and ζ in (2). Let the mismatched
dispersion be defined as
V(σ 2 , ζ) :=
ζ − σ4
Var[X 2 ]
=
.
4
4σ
4 (E[X 2]2 )
Theorem 1. Consider an arbitrary memoryless source X satisfying (2). For any ε ∈ [0, 1),
p
L∗sp (ε) = L∗iid (ε) = V(σ 2 , ζ)Q−1 (ε).
(18)
(19)
The proof of Theorem 1 is provided in Section IV. In the proof, we show that for any ε ∈ (0, 1),
∗
∗
log Msp
(n, ε, σ 2 , D) = log Miid
(n, ε, σ 2 , D) =
σ2 p
n
log
+ nV(σ 2 , ζ)Q−1 (ε) + O(log n).
2
D
(20)
A few remarks are in order.
First, in contrast to Scarlett, Tan, and Durisi [9] where spherical and i.i.d. Gaussian codebooks achieve different second-order
coding rates for the channel coding saddle-point problem (where the codebook is Gaussian and the channel is additive and
its noise is non-Gaussian) [10], we observe no performance gap in second-order asymptotics between two kinds of Gaussian
codebooks in the rate-distortion saddle-point counterpart. We provide some intuition why this is so. For the channel coding
problem [9], [10], the performance gap between the dispersions appears because the probability of dominant error event depends
strongly on the choice of the random codebook since the mutual information random variable that appears in the random coding
2
2
∗
1 In Definition 2, we allow ε = 0 (signifying that P
∗
e,n is required to vanish) but in the definitions of Msp (n, ε, σ , D) and Miid (n, ε, σ , D) we enforce
that ε ∈ (0, 1).
5
union (RCU) bound [12] is a strong function of the code distribution. Indeed, it is advantageous to use a spherical code as this
ensures that the codewords have exact power P . For an i.i.d. code, each of the codewords has power P averaged over the n
time slots. Due to this additional randomness (in addition to the channel noise), the dispersion or variability for the i.i.d. case is
worse (higher) than that for the spherical case. This is analogous to the well-known fact that for discrete memoryless channels,
fixed composition codes [15] generally perform better than i.i.d. codes [16] in the error exponents regime [19, Chapter 2] [22,
Section 5]. In contrast, for the rate-distortion saddle-point problem we consider in this paper, there is no “power constraint”
on the codewords. Thus, as can be gleaned in the proof in Section IV, the dominant error event in the present rate-distortion
case concerns the atypicality of the source sequence X n . The Gaussian codebook plays a diminished role in the contribution
to the overall randomness or dispersion of the system.
Second, for a GMS (i.e., X ∼ N (0, σ 2 )), the dispersion is known to be V(σ 2 , ζ) = 12 independent of the distortion level or
the source variance [7], [8]. Hence, when specialized to GMSes, our results are consistent with existing results, namely that
both spherical and i.i.d. Gaussian codebooks achieve the optimal second-order coding rate for the rate-distortion problem [7,
Theorems 13 and 40]. Furthermore, our results strengthen those of Lapidoth [3, Theorem 3] in the sense that, using i.i.d.
Gaussian codebooks, it is true that for an arbitrary memoryless source satisfying (2) and for any ε ∈ (0, 1), the first-order
result in (10) holds.
Third, we remark that as (20) indicates, the larger the mismatched dispersion for a source, the larger the back-off from
the first-order coding rate given a fixed non-vanishing excess-distortion probability ε and a blocklength n. This mismatched
dispersion is a function of the fourth moment of the source. Indeed, V(σ 2 , ζ) is related to the kurtosis of the source.
Finally, we mention in passing that our proof strategy differs significantly from Lapidoth’s [3] who, for the direct and
ensemble converse parts, invoked a theorem due to Wyner [23] concerning packings and coverings of n-spheres. The analysis
used to prove this theorem and the subsequent ones naturally require more refined estimates on various probabilities.
B. Moderate Deviation Asymptotics
Theorem 2. Consider an arbitrary memoryless source X satisfying (2) and ΛX 2 (θ) is finite for some positive number θ. If
V(σ 2 , ζ) is positive,
∗
∗
νsp
= νiid
=
1
.
2V(σ 2 , ζ)
(21)
The proof of Theorem 2 is provided in Section V. Several remarks are in order.
First, we observe that similarly to second-order asymptotics (Theorem 1), the dispersion V(σ 2 , ζ) is a fundamental quantity
that governs the speed of convergence of the excess-distortion probability to zero.
Second, we remark the constraint in (13) is mild. In fact, ξn = n−t for any t ∈ (0, 0.5) satisfies (13). We believe
√ that the
second condition in (13) may be relaxed by the more common moderate deviations condition [11], [12], namely nξn → ∞,
if a more careful analysis is employed.
Third, we remark that the condition on the source where ΛX 2 (θ) is finite for some θ > 0 is also mild. It is used to assert
that the probability that an average of a sum of i.i.d. random variables deviates from its mean by a constant amount decays
exponentially. See [24, Theorem B.4.1] and Lemma 6 to follow. Note that this condition is satisfied if X is distributed according
to many common distributions such as discrete, Gaussian, Maxwell, or Rayleigh. However, this condition is violated if X has
a Cauchy distribution, which does not even have a finite mean.
C. Large Deviation Asymptotics
We present several definitions before stating our main result. Given s ∈ R and any non-negative number z, define
(z + σ 2 − 2D)2
1
,
Rsp (z) := − log 1 −
2
4z(σ 2 − D)
1
sz
sD
Riid (s, z) := log(1 + 2s) +
−
, and
2
(1 + 2s)(σ 2 − D) σ 2 − D
p
σ 2 − 3D + (σ 2 − D)2 + 4zD
∗
.
s (z) := max 0,
4D
(22)
(23)
(24)
We remark that Rsp (z) is the rate of the exponential decay of the non-excess-distortion probability for any source sequence
xn whose power is z = n1 kxn k2 when its reproduction sequence is generated according to fYspn in (5), i.e.,
1
log Pr{d(xn , Y n ) ≤ D} = Rsp (z), where Y n ∼ fYspn .
(25)
n
Similarly, Riid (s∗ (z), z) is the exponent of the non-excess-distortion probability for any source sequence xn with power z
when Y n ∼ fYiidn .
lim −
n→∞
6
The Fenchel-Legendre transform of the cumulant generating function of X 2 (cf. Section II-A) is
Λ∗X 2 (t) := sup{θt − ΛX 2 (θ)},
θ≥0
t ∈ R.
(26)
This is also known as the large deviations rate function [25] of X 2 . For brevity in presentation of the following theorem, let
p
√
(27)
r2 := σ 2 − D + D.
Theorem 3. The following results on the optimal excess-distortion exponents hold.
1
σ2
• If R < 2 log D , then for both spherical and i.i.d. Gaussian codebooks,
∗
∗
Esp
(R) = Eiid
(R) = 0.
(28)
2
•
If R ≥ 21 log σD ,
– For the spherical Gaussian codebook, the optimal excess-distortion exponent satisfies
∗
Esp
(R) = Λ∗X 2 (α),
(29)
where α ∈ [σ 2 , r22 ) is implicitly determined by R through the equation
R = Rsp (α).
(30)
– For the i.i.d. Gaussian codebook, the optimal excess-distortion exponent is
∗
Eiid
(R) = Λ∗X 2 (α),
(31)
where α ≥ σ 2 is implicitly determined by R through the equation
R = Riid (s∗ (α), α).
(32)
The proof of Theorem 3 is provided in Section VI. Several remarks are in order.
First, as can be gleaned in the proof of Theorem 3 (cf. Lemma 7), Rsp (z) and Riid (s∗ (z), z) are both increasing functions
2
of z if z ≥ σ 2 and Rsp (σ 2 ) = Riid (s∗ (σ 2 ), σ 2 ) = 12 log σD . Furthermore, Λ∗X 2 (t) > 0 for t > σ 2 and Λ∗X 2 (t) = 0 otherwise.
2
∗
∗
∗
Combining these two facts, we can conclude that Esp
(R) and Eiid
(R) are both positive for rates R > 21 log σD . That Esp
(R) > 0
2
1
σ
for R > 2 log D recovers the achievability part of [3, Theorem 3] without recourse to Wyner’s theorem [23].
Second, for the i.i.d. Gaussian codebook, if we consider that each codeword is generated according to (6) with σ 2 replaced
α
. Under this scenario, by particularizing the result to a
by α ∈ R+ , then the right hand side of (32) is replaced by 12 log D
GMS, we can recover the achievability result of Ihara and Kubo [17].2
Furthermore, one may wonder the performance comparison of large deviations for the spherical and i.i.d. Gaussian codebooks.
We give a complete answer to this question in the following lemma.
Lemma 4. For any rate R >
1
2
2
log σD , we have
∗
∗
Eiid
(R) > Esp
(R).
(33)
The proof of Lemma 4 is given in Appendix B. To illustrate the result in Lemma 4, we plot optimal excess-distortion
exponents of spherical and i.i.d. Gaussian codebooks for a discrete and a Rayleigh distribution in Figure 1. We verify from
Figure 1 that for both numerical examples, the i.i.d. Gaussian codebook has a strictly larger excess-distortion exponent than
2
the spherical Gaussian codebook for any rate R > 12 log σD .
IV. P ROOF
OF
S ECOND -O RDER A SYMPTOTICS (T HEOREM 1)
A. Preliminaries for the Spherical Codebook
In this subsection, we present some definitions and preliminary results for spherical Gaussian codebooks. For simplicity, let
the variance or power of Y be PY := σ 2 − D. Furthermore, for any ε ∈ (0, 1), let
V := Var[X 2 ] = ζ − σ 4 ,
r
log n
,
an := V
n
r
V −1
bn :=
Q (ε),
n
2 Note
that we use α while Ihara and Kubo [17] use α2 to mean the same quantity.
(34)
(35)
(36)
7
0.7
∗
Esp
(R)
∗
Eiid
(R)
R = 12 log
Bounds on the optimal error exponent
1.6
1.4
Bounds on the optimal error exponent
1.8
σ2
D
1.2
1
0.8
0.6
0.4
∗
Esp
(R)
∗
Eiid
(R)
R = 12 log
0.6
σ2
D
0.5
0.4
0.3
0.2
0.1
0.2
0
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0
0.2
0.4
0.6
0.8
Rate R
Rate R
(a) Discrete
(b) Rayleigh
1
1.2
1.4
∗ (R) and E (R) for a memoryless source distributed according to a discrete and a Rayleigh distribution. The discrete
Fig. 1. Excess-distortion exponents Esp
iid
distribution is ternary with support X = {a, 2a, 3a} where a2 = 0.3σ2 and its probability mass function is PX = [1/2, 1/3, 1/6] so that E[X 2 ] = σ2 .
The Rayleigh distribution has scale parameter σ2 /2 so that E[X 2 ] = σ2 . Note that α is determined through R by (30) for the spherical Gaussian codebook
and by (32) for the i.i.d. Gaussian codebook.
where the second equality in (34) follows from the definition in (2). Note that for any xn , Pr{d(xn , Y n ) ≤ D} depends on
xn only through its norm kxn k. For any xn such that n1 kxn k2 = z > 0, let
Ψ(n, z) := Pr{d(xn , Y n ) ≤ D}
(37)
= Pr{kxn − Y n k2 ≤ nD}
(38)
= Pr{2hxn , Y n i ≥ n(z + PY − D)}
√
n(z + PY − D)
√
,
= Pr Y1 ≥
2 z
(41)
= Pr{kxn k2 + kY n k2 − 2hxn , Y n i ≤ nD}
= Pr{nz + nPY − 2hxn , Y n i ≤ nD}
(39)
(40)
(42)
where Y1 is the √
first element of sequence Y n = (Y1 , . . . , Yn ) and (42) follows because Y n is spherically symmetric so we
n
may take x = ( nz, 0, . . . , 0) (cf. [9]).
n
source sequence
let fZ
Let Z := n1 kX n k2 be the random variable representing the average power of a √
√ X √. Furthermore,
√
2
be the corresponding probability distribution function (pdf) of Z. Recall that r2 = σ − D + D = PY + D (cf. (27))
and let
p
√
(43)
r1 := PY − D.
√
√
Kostina and Verdú [7, Theorem 37] showed that for any z such that z < r1 or z > r2 ,
Ψ(n, z) = 0,
(44)
n−1
2
Γ( n+2
(z + PY − D)2
2 )
1−
Ψ(n, z) ≥ √
=: g(n, z),
n+1
4zPY
πnΓ( 2 )
(45)
and otherwise
where Γ(·) is the Gamma function. Hence, from (8) and (44), we conclude that the excess-distortion probability for the spherical
Gaussian codebook equals to
Z r22
2
2
(1 − Ψ(n, z))M fZ (z) dz.
(46)
Pe,n = Pr{Z < (max{0, r1 }) } + Pr{Z > r2 } +
(max{0,r1 })2
8
B. Achievability Proof for the Spherical Gaussian Codebook
Using the definition of g(·) in (45), we conclude that g(n, z) is a decreasing function of z if z ≥ |PY − D|. Invoking the
definitions of bn in (36), r1 in (43) and r2 in (27), we conclude that r12 ≤ |PY − D| and r22 ≥ σ 2 + bn for n large enough.
Thus, combining (45), (46) and noting that Ψ(n, z) ≤ 1, for sufficiently large n, we can upper the excess-distortion probability
as follows:
Z σ2 +bn
(1 − g(n, z))M fZ (z) dz + Pr{Z > σ 2 + bn }
Pe,n ≤ Pr{Z < |PY − D|} +
(47)
|PY −D|
≤ Pr{Z < |PY − D|} +
Z
σ2 +bn
|PY −D|
exp{−M g(n, z)}fZ (z) dz + Pr{Z > σ 2 + bn },
≤ Pr{Z < |PY − D|} + exp{−M g(n, σ 2 + bn )} + Pr{Z > σ 2 + bn }
(48)
(49)
where (48) follows since (1 − a)M ≤ exp{−M a} for any a ∈ [0, 1); and (49) follows since g(n, z) is decreasing in z for
z ≥ |PY − D|. Let the third central moment of X 2 be defined as
T := E |X 2 − σ 2 |3 .
(50)
Using the definitions of V in (34), T in (50) and the Berry-Esseen theorem, we conclude that
X
n
1
Pr{Z < |PY − D|} = Pr
Xi2 < |σ 2 − 2D|
n i=1
r
n
6T
2
2
+ Q σ − |σ − 2D|
≤ √
V
n V3/2
2
2
2
2n σ − |σ − 2D|
6T
≤ √
+ exp −
V
n V3/2
1
=O √ ,
n
(51)
(52)
(53)
(54)
2
where (53) follows since Q(a) ≤ exp{− a2 } while (54) follows since T (cf. (50)) is finite for sources satisfying (2) and
σ 2 − |σ 2 − 2D| > 0 due to the fact that σ 2 > D. Similarly, using the definition of bn in (36) and the Berry-Esseen theorem,
we have
X
n
1
2
2
2
(55)
X > σ + bn
Pr{Z > σ + bn } = Pr
n i=1 i
6T
≤ε+ √
n V3/2
1
≤ε+O √ .
n
(56)
(57)
Choose M such that
1
log n
log M = − log g(n, σ + bn ) + log
2
1
σ2
bn
log n
=n
log
+ 2 +O
2
D
2σ
n
2
p
σ
n
+ nV(σ 2 , ζ)Q−1 (ε) + O(log n),
= log
2
D
2
(58)
(59)
(60)
√
n+1
where (59) follows from the Taylor expansion of g(n, σ 2 + bn ) (cf. (45)) and noting that Γ( n+2
2 )/Γ( 2 ) = Θ( n); and (60)
follows from the definition of bn (cf. (36)) and V(σ 2 , D) (cf. (18)). Thus, with the choice of M in (58), we conclude that
1
exp{−M g(n, σ 2 + bn )} = √ .
n
(61)
Hence, combining (49), (54), (57) and (61), we have shown that there exists a sequence of (n, M )-codes satisfying (60) and
lim sup Pe,n ≤ ε.
n→∞
(62)
9
C. Ensemble Converse for the Spherical Gaussian Codebook
We now show that the result in (20) is ensemble tight. From Stam’s paper [26, Eq. (4)], the distribution of Y1 is
n−3
2
Γ( n2 )
1
y2
fY1 (y) = √
1−
1{y 2 ≤ nPY }.
n−1
nPY
πnPY Γ( 2 )
(63)
Recall the definitions of an in (35) and bn in (36). Define the sets
P := {r ∈ R : bn < r − σ 2 ≤ an },
(64)
Q := {r ∈ R : r + PY − D ≥ 0}.
√
√
−D)
√Y
≤ nPY , using the definition of Ψ(·) in (42), we obtain that
Then, for any z ∈ P ∩ Q satisfying n(z+P
2 z
√
n(z + PY − D)
√
Ψ(n, z) = Pr Y1 ≥
2 z
n−3
Z √nPY
2
Γ( n2 )
y2
1
√
1
−
= √
dy
n(z+PY −D)
nPY
πnPY Γ( n−1
√
2 )
2 z
n−3
Z √nPY
2
Γ( n2 )
(z + PY − D)2
1
√
≤ √
1
−
dy
n−1
n(z+PY −D)
4zPY
πnPY Γ( 2 )
√
2 z
n−3
2
(z + PY − D)2
1 Γ( n2 )
1−
≤ √
n−1
4zPY
π Γ( 2 )
n
Γ(
)
1
(z + PY − D)2
n−3
2
= √
=: g(n, z),
log
1
−
exp
π Γ( n−1
2
4zPY
2 )
(65)
(66)
(67)
(68)
(69)
(70)
√
−D)
√Y
where (67) follows from the definition in (63) and the condition that z ∈ Q (cf. (65)) which implies n(z+P
≥0>
2 z
√
2
y
− nPY ; (68) follows since (1 − nPY ) is decreasing in y for positive y; and (69) follows by enlarging the integration region
√
n(z+PY −D)
√
2 z
≥ 0). Note that g(n, z) is decreasing in z for z ≥ |PY − D| and g(n, z) ≥ 0 for all z ∈ P. Hence,
√
√
−D)
√Y
for any z ∈ P ∩ Q such that n(z+P
> nPY , we still have g(n, z) ≥ Ψ(n, z).
2 z
Recall that Z = n1 kX n k2 and fZ is the corresponding pdf of Z. Thus, according to (8), for n sufficiently large, we have
(recall that
Pe,n = EX n [(1 − Pr{d(X n , Y n ) ≤ D})M |X n ]
Z ∞
=
(1 − Ψ(n, z))M fZ (z) dz
0
Z ∞
≥
(1 − g(n, z))M 1{z ∈ P ∩ Q}fZ (z) dz
0
Z
(1 − g(n, σ 2 + bn ))M fZ (z) dz
≥
z∈P∩Q
Z
g(n, σ 2 + bn )
exp − M
≥
fZ (z) dz
1 − g(n, σ 2 + bn )
Zz∈P∩Q
1
g(n, σ 2 + bn )
g(n, σ 2 + bn )
exp − M
≥
1 M
≤ √
fZ (z)dz
1 − g(n, σ 2 + bn )
1 − g(n, σ 2 + bn )
n
z∈P∩Q Z
1
1
g(n, σ 2 + bn )
≤ √
≥ 1− √
fZ (z)dz
1 M
2
1
−
g(n,
σ
+
b
)
n
n
n
z∈P∩Q
1 − g(n, σ 2 + bn ) 1
1
√
Pr Z ∈ P ∩ Q, M ≤
= 1− √
n
g(n, σ 2 + bn )
n
1
1
= 1− √
Pr Z ∈ P ∩ Q, log M ≤ log(1 − g(n, σ 2 + bn )) − log g(n, σ 2 + bn ) − log n
2
n
1
1
≥ 1− √
Pr Z ∈ P ∩ Q, log M ≤ − log 2 − log g(n, σ 2 + bn ) − log n} ,
2
n
(71)
(72)
(73)
(74)
(75)
(76)
(77)
(78)
(79)
(80)
where (72) follows from the definition of Ψ(n, z) in (42); (73) follows by restricting z ∈ P ∩ Q and using the definition of
a
g(·) in (70); (74) follows since g(n, z) is decreasing in z for z ∈ P ∩ Q; (75) follows since (1 − a)M ≥ exp{−M 1−a
} for
g(n,z)
1
√
,
exp{−a}
is
decreasing
in
a,
and
exp{−a}
≥
1
−
a
for
a
≥
0;
and
(80)
any a ∈ [0, 1); (77) follows since M 1−g(n,z) ≤ n
1
2
follows since g(n, z) ≤ 2 for n large enough if z > σ .
10
Combining (70), (80) and applying a Taylor expansion of g(n, σ 2 + bn ) similarly to (59), we conclude that for any (n, M )code such that
1
log M ≤ − log 2 − log n − log g(n, σ 2 + bn )
(81)
2
2
σ
bn
log n
1
,
(82)
log
+ 2 +O
=n
2
D
2σ
n
we have
Pe,n
1
≥ 1− √
Pr{Z ∈ P ∩ Q}.
n
(83)
The following lemma is essential to complete the converse proof.
Lemma 5. Consider any source X such that (2) are satisfied and σ 2 < ∞. Then, we have
1
Pr{Z ∈ P ∩ Q} ≥ ε + O √ .
n
(84)
The proof of Lemma 5 is deferred to Appendix A.
Using the definition of V(σ 2 , ζ) in (18), the definition of bn in (36), the bounds in (82), (83), and Lemma 5, we conclude
that for any sequence of (n, M )-codes such that
n
σ2 p
log M ≤ log
(85)
+ nV(σ 2 , ζ)Q−1 (ε) + O(log n),
2
D
we have
lim sup Pe,n ≥ lim inf Pe,n ≥ ε
n→∞
n→∞
(86)
as desired.
D. Preliminaries for the I.I.D. Gaussian Codebook
Now we consider the i.i.d. Gaussian codebook (cf. (6)). Note that Pr{d(xn , Y n ) ≤ D} depends on xn only through its
norm kxn k (cf. [17]). Given any sequence xn such that n1 kxn k2 = z, define
Υ(n, z) := Pr{d(xn , Y n ) ≤ D}.
(87)
From (6), we obtain that
fYiidn (y n )
1
exp
=
2
(2π(σ − D))n/2
ky n k2
.
−
2(σ 2 − D)
(88)
Since fYiidn (y n ) is decreasing in ky n k, we conclude that Υ(n, z) is a decreasing function of z (cf. [17]). Using the definition
of Υ(·) in (87), we have
Υ(n, z) = Pr{kxn − Y n k2 ≤ nD}
X
n
√
(Yi − z)2 ≤ nD
= Pr
(89)
(90)
i=1
= Pr
n
√ 2
1 X
D
(Yi − z) ≥ −
−
.
nPY i=1
PY
(91)
√
where (90) follows since the probability depends on xn only through its power and thus we can choose xn such that x√
z
i =
1
2
for all i ∈ [1 : n] (cf. [9, Eq. (94)]). For the i.i.d. Gaussian codebook, each Yi ∼ N (0, PY ) and hence PY (Yi − z) is
distributed according to a non-central χ2 distribution with one degree of freedom.
Given z and s, let
κ(s, z) :=
(PY (1 + 2s) + 2z)2
PY (1 + 2s)3
(92)
Using the result of [27, Section 2.2.12] concerning the cumulant generating function of a non-central χ2 distribution, the
definition of Riid (·) in (23), the definition of s∗ (·) in (24), and the Bahadur-Ranga Rao (strong large deviations) theorem for
non-lattice random variables [25, Theorem 3.7.4], we obtain
Υ(n, z) ∼
exp{−nRiid(s∗ (z), z)}
p
,
s∗ (z) κ(s∗ (z), z)2πn
n → ∞.
(93)
11
E. Achievability Proof for the I.I.D. Gaussian Codebook
According to (8), the excess-distortion probability under the i.i.d. Gaussian codebook can be upper bounded as follows:
i
h
(94)
Pe,n = E (1 − Pr{d(X n , Y n ) ≤ D})M X n
Z ∞
=
(1 − Υ(n, z))M fZ (z) dz
(95)
0
≤
Z
≤
Z
σ2 −an
0
σ2 +bn
σ2 −an
fZ (z) dz +
Z
σ2 +bn
σ2 −an
(1 − Υ(n, z))M fZ (z) dz +
Z
∞
fZ (z) dz
(96)
σ2 +bn
exp{−M Υ(n, z)}fZ (z) dz + Pr{Z < σ 2 − an } + Pr{Z > σ 2 + bn }
≤ exp{−M Υ(n, σ 2 + bn )} + Pr{Z < σ 2 − an } + Pr{Z > σ 2 + bn },
(97)
(98)
where (96) follows since Υ(n, z) ≥ 0; (97) follows since (1−a)M ≤ exp{−M a}; and (98) follows since Υ(n, z) is decreasing
in z and Pr{σ 2 − an ≤ Z ≤ σ 2 + bn } ≤ 1.
Using the definitions of Riid (·) in (23) and s∗ (·) in (24), we have
Riid (s∗ (σ 2 + bn ), σ 2 + bn )
p
p
PY + PY2 + 4(σ 2 + bn )D z(PY − 2D + PY2 + 4(σ 2 + bn )D)
1
p
= log
+
2
2D
2PY (PY + PY2 + 4(σ 2 + bn )D)
p
PY − 2D + PY2 + 4(σ 2 + bn )D
−
4PY
2
σ
bn
1
+ 2 + O(b2n ),
= log
2
D r
2σ
V(σ 2 , ζ) −1
1
σ2
1
= log
,
+
Q (ε) + O
2
D
n
n
(99)
(100)
(101)
where (100) follows from a Taylor expansion at z = σ 2 and recalling that PY = σ 2 − D; and (101) follows from the definitions
of V(σ 2 , ζ) in (18) and bn in (36).
Choose M such that
1
2
log n .
(102)
log M ≥ − log Υ(n, σ + bn ) + log
2
Then, we have
1
exp{−M Υ(n, σ 2 + bn )} ≤ √ .
n
Furthermore, using the result in (93) and (101), we obtain
σ2 p
n
+ nV(σ 2 , ζ)Q−1 (ε) + O(log n).
log M ≥ log
2
d
Similarly as the proof of Lemma 5, using the Berry-Esseen theorem and the definition of an in (35), we obtain
r
X
n
log n
1
2
2
2
(X − σ ) < V
Pr{Z < σ − an } = Pr
n i=1 i
n
p
6T
≤ Q( log n) + √
n V3/2
1
=O √ .
n
(103)
(104)
(105)
(106)
(107)
Hence, combining (57), (98), (103) and (107), we conclude that there exists a sequence of (n, M )-codes using i.i.d. Gaussian
codebooks with M satisfying (104) and
lim sup Pe,n ≤ ε.
(108)
n→∞
F. Ensemble Converse for the I.I.D. Gaussian Codebook
The ensemble converse proof for the i.i.d. Gaussian codebook is omitted since it is similar to the ensemble converse proof
for the spherical Gaussian codebook in Section IV-C starting from (71) except for the following two points: i) replace g(n, z)
with Υ(n, z); ii) replace P ∩ Q with P.
12
V. P ROOF
OF
M ODERATE D EVIATION A SYMPTOTICS (T HEOREM 2)
A. Preliminaries
The following slight strengthening of the Chernoff bound [24, Theorem B.4.1] is crucial in our proof.
Lemma 6. Given an i.i.d. sequence X n , suppose that the cumulant generating function Λ|X| (θ) is finite for some positive
number θ. Then for any t > E[X],
X
n
1
Xi > t ≤ exp{−nΛ∗X (t)}.
(109)
Pr
n i=1
and for any t < E[X],
Pr
n
1X
Xi < t
n i=1
≤ exp{−nΛ∗X (t)}.
(110)
In both cases, Λ∗X (t) > 0.
In other words, if the threshold t deviates from the mean by a constant amount, the probability in question decays
exponentially fast.
B. Achievability Proof for the Spherical Gaussian Codebook
Let
cn := 2σ 2 ξn .
(111)
The proof of the achievability follows from Section IV-B up till (49) with cn taking the role of bn . Invoking Lemma 6, we
conclude that under the conditions in Theorem 2, we have
X
n
1
2
X < |PY − D|
(112)
Pr{Z < |PY − D|} = Pr
n i=1 i
≤ exp{−nt1 }
2
2
(113)
2
for some t1 > 0 since |PY − D| = |σ − 2D| < σ due to the fact that σ > D. Invoking the moderate deviations theorem [25,
Theorem 3.7.1] and the definition of V in (34), we conclude that
X
n
1
2
2
Pr{Z > σ 2 + cn } = Pr
(114)
(X − σ ) > cn
n i=1 i
nc2
= exp − n + o(nc2n )
(115)
2V
nξn2
+ o(nξn2 ) ,
(116)
= exp −
2V(σ 2 , ζ)
where (116) follows from the definitions of V(σ 2 , ζ) in (18), V in (34) and cn in (111).
Recall the definition of g(·) in (45). Choose M such that
log M = − log g(n, σ 2 + cn ) + log n
1
σ2
=n
log
+ ξn + o(ξn ) ,
2
D
(117)
(118)
where (118) follows from a Taylor expansion similar to (59), the definition of cn in (111), and the conditions on ξn in (13).
With this choice of M , we have
exp{−M g(n, σ 2 + cn )} = exp(−n).
(119)
Combining the results in (49), (113), (116), (118) and (119), we conclude that there exists a sequence of (n, M )-codes such
that
n
σ2
1
≤ 1,
(120)
log M − log
lim sup
2
D
n→∞ nξn
1
1
lim inf − 2 log Pe,n ≥
.
(121)
n→∞
nξn
2V(σ 2 , ζ)
13
C. Ensemble Converse Proof for the Spherical Gaussian Codebook
The converse proof follows from Section IV-C up till (83). However, we use (P ′ , cn ) in place of (P, bn ) where
P ′ := {r ∈ R : ξn < r − σ 2 < 2ξn }.
(122)
Now recall the definition of g(·) in (70). Applying a Taylor expansion similar to (118), invoking Lemma 6, and using the
moderate deviations theorem in [25, Theorem 3.7.1], we conclude that for any (n, M )-code such that
1
log M ≤ − log 2 − log n − log g(n, σ 2 + cn )
2
σ2
1
log
+ ξn + o(ξn ) ,
=n
2
D
(123)
(124)
we have
Pe,n
≥ exp
1 − √1n
4nξn2
nξn2
2
2
+ o(nξn ) − exp −
+ o(nξn ) − exp{−nt2 },
−
2V(σ 2 , ζ)
2V(σ 2 , ζ)
(125)
for some t2 > 0. Note that the first term on the right hand side of (125) dominates as n → ∞. From the results in (124)
and (125), we conclude that for any sequence of (n, M )-codes such that
n
σ2
1
≤ 1,
(126)
log M − log
lim sup
2
D
n→∞ nξn
we have
lim inf −
n→∞
1
1
1
log Pe,n ≤ lim sup − 2 log Pe,n ≤
.
2
nξn
nξn
2V(σ 2 , ζ)
n→∞
(127)
D. Proof for the I.I.D. Gaussian Codebook
The proof for i.i.d. Gaussian codebooks is similar to the proof in Section IV-E and IV-F with the use of Lemma 6 and [25,
Theorem 3.7.1] as in Sections V-B and V-C and is thus omitted.
VI. P ROOF
OF
L ARGE D EVIATION A SYMPTOTICS (T HEOREM 3)
A. Preliminaries
The following properties of the quantities Rsp (z) in (22), Riid (s∗ (z), z) (cf. (23) and (24)) and Λ∗X 2 (t) in (26) are useful
in the proof of Theorem 3.
Lemma 7. The following claims hold.
(i) Concerning Rsp (z),
(a) Rsp (z) is increasing in z if z ≥ |σ 2 − 2D|;
2
(b) Rsp (σ 2 ) = 12 log σD , limz→r22 Rsp (z) = ∞.
(ii) Concerning Riid (s∗ (z), z),
(a) s∗ (z) > 0 if and only if z > max(0, 2D − σ 2 );
2
(b) Riid (s∗ (z), z) = 0 if z = max{0, 2D − σ 2 } and Riid (s∗ (σ 2 ), σ 2 ) = 12 log σD ;
(c) Riid (s∗ (z), z) = sups≥0 Riid (s, z) and thus Riid (s∗ (z), z) is increasing in z for z ≥ max{0, 2D − σ 2 }.
(iii) Concerning Λ∗X 2 (t) (cf. [25, Chapter 3]),
(a) Λ∗X 2 (t) is convex and non-decreasing in t for t ≥ 0;
(b) Λ∗X 2 (t) = 0 if t ≤ σ 2 ;
(c) Λ∗X 2 (t) is increasing in t for t ≥ σ 2 .
The proof of Lemma 7 is omitted since it follows either from simple algebra or from [25, Chapter 3].
Furthermore, we have the following lemma concerning an important property of the function g(n, z) (cf. (45)), which plays
an important role in proving the ensemble tight excess-distortion exponent for spherical Gaussian codebooks.
Lemma 8. For any α ∈ [σ 2 , r22 ), there exists a unique β ∈ (r12 , |σ 2 − 2D|) such that
g(n, β) = g(n, α),
2
α + β ≤ 2σ .
The proof of Lemma 8 is given in Appendix C.
(128)
(129)
14
B. Achievability Proof for the Spherical Gaussian Codebook
Recall that PY = σ 2 − D. Invoking the definitions of r1 in (43) and g(·) in (45), we concluce that r12 ≤ |σ 2 − 2D| < σ 2 ,
g(n, z) is decreasing in z if z ∈ (|σ 2 − 2D|, r22 ) and g(n, z) is increasing in z if z ∈ (r12 , |σ 2 − 2D|)
Using the expression for the excess-distortion probability in (46), given any α such that α ∈ [σ 2 , r22 ), we can upper bound
Pe,n as follows
X
Z r22
X
n
n
1
1
(1 − Ψ(n, z))M fZ (z) dz
(130)
Xi2 < r12 + Pr
Xi2 > r22 +
Pe,n ≤ Pr
n i=1
n i=1
r12
X
X
Z |σ2 −2D|
n
n
1
1
≤ Pr
Xi2 < β + Pr
Xi2 > α +
(1 − g(n, z))M fZ (z) dz
n i=1
n i=1
β
Z α
+
(1 − g(n, z))M fZ (z) dz
(131)
|σ2 −2D|
n
X
n
1X 2
Xi > α + (1 − g(n, α))M
n i=1
i=1
X
X
n
n
1
1
(Xi2 − σ 2 ) > σ 2 − β} + Pr
Xi2 > α + exp{−M g(n, α)}
≤ Pr
n i=1
n i=1
X
n
1
X 2 > α + exp{−M g(n, α)},
≤ 2 Pr
n i=1 i
= Pr
1
n
(Xi2 − σ 2 ) < β − σ 2
+ Pr
(132)
(133)
(134)
where (131) follows from the result in Lemma 8 which states that there exists a unique β ∈ (r22 , |σ 2 − 2D|) such that
g(n, β) = g(n, α) for any α ∈ [σ 2 , r22 ); (132) follows since i) g(n, z) is decreasing for z ≥ |σ 2 − 2D| and α ≥ σ 2 > |σ 2 − 2D|,
ii) g(n, z) is increasing for z ∈ (r12 , |σ 2 − 2D|); and iii) the result in Lemma 8 which states that g(n, β) = g(n, α) (cf. (128))
and β ∈ (r12 , |σ 2 − 2D|); (133) follows since (1 − a)M ≤ exp{−M a} for any a ∈ [0, 1); and (134) follows since 2σ 2 − β ≥ α,
which is implied by (129) in Lemma 8.
Now, given any positive δ ∈ (0, 1), recalling the definition of Rsp (·) in (22), we choose M such that
√
πnΓ( n+1
2 )
log M = (1 + δ)(n − 1)Rsp (α) + log
.
(135)
n+2
Γ( 2 )
Using the definitions of Rsp (·) in (22) and g(n, z) in (45), we obtain that
exp{−M g(n, z)} = exp − exp{(n − 1)δRsp (α)} ,
(136)
which vanishes doubly exponentially fast for α ≥ σ 2 . Invoking Cramér’s Theorem [25, Theorem 2.2.3] and the definition of
Λ∗X 2 (·) in (26), we obtain that
X
n
1
Pr
(137)
Xi2 > α ≤ exp − nΛ∗X 2 (α) .
n i=1
√
√
n+2
Therefore, using (134), (135), (136), (137) and noting that πnΓ( n+1
2 )/Γ( 2 ) = Θ(1/ n), we conclude that for all
α ∈ [σ 2 , r22 ), there exists a sequence of (n, M )-codes such that
lim
n→∞
1
log M = (1 + δ)Rsp (z),
n
(138)
1
log Pe,n ≥ Λ∗X 2 (α).
(139)
n
Now, note that σ 2 > |σ 2 − 2D| due to the fact that σ 2 > D. Recall that Rsp (z) is increasing in z for z ≥ σ 2 (cf. Claim (i)(a)
2
in Lemma 7) and Rsp (σ 2 ) = 21 log σD (cf. Claim (i)(b) in Lemma 7). Hence, by letting δ ↓ 0, (138) and (139) imply that for
2
all R ≥ 12 log σD , there exists a sequence of (n, M )-codes such that
lim inf −
n→∞
1
log M = R,
n
and (139) both hold, where α is determined from R = Rsp (α) (cf. (22)).
2
The proof for R ∈ [0, 12 log σD ) follows trivially by noting that any (n, M )-code satisfies that Pe,n ≤ 1.
lim
n→∞
(140)
15
C. Ensemble Converse Proof for the Spherical Gaussian Codebook
Fix any α such that α ∈ [σ 2 , r22 ) (cf. (27) for the definition of r2 ). Let
P̃ := {r ∈ R : α ≤ r < r22 },
2
Q̃ := {r ∈ R : r − |σ − 2D| ≥ 0}.
Note that r ∈ Q̃ implies that r + (σ 2 − 2D) ≥ 0.
Using the result in (8) and the definition of g(·) in (70), we conclude that for sufficiently large n,
Z r22
(1 − g(n, z))M 1{z ∈ P̃ ∩ Q̃}fZ (z) dz
Pe,n ≥
α
X
n
1
M
2
≥ (1 − g(n, α)) Pr
X ∈ P̃ ∩ Q̃
n i=1 i
X
n
1
g(n, α)
2
X ∈ P̃ ∩ Q̃
Pr
≥ exp − M
1 − g(n, α)
n i=1 i
X
n
1
2
X ∈ P̃ ∩ Q̃ ,
≥ exp − 2M g(n, α) Pr
n i=1 i
(141)
(142)
(143)
(144)
(145)
(146)
where (144) follows since g(n, z) is decreasing in z for z ≥ |σ 2 − 2D| and α ≥ σ 2 > |σ 2 − 2D|; (145) follows since
a
} for any a ∈ [0, 1); and (146) follows since g(n, α) ≤ 21 for n sufficiently large.
(1 − a)M ≥ exp{−M 1−a
For any M such that
1
log n,
2
using (146) and the inequality that exp{−a} ≥ 1 − a, we have that for sufficiently large n,
X
n
1
1
2
Pr
X ∈ P̃ ∩ Q̃ .
Pe,n ≥ 1 − √
n
n i=1 i
log M ≤ − log g(n, α) − log 2 −
(147)
(148)
Note that
n
1X 2
Xi ∈ P̃ ∩ Q̃
n i=1
n
1X 2
= Pr max{α, |σ 2 − 2D|} ≤
Xi < r22
n i=1
X
X
n
n
1
1
Xi2 ≥ α − Pr
Xi2 ≥ r22 ,
= Pr
n i=1
n i=1
Pr
(149)
(150)
where the final equality holds because α2 is chosen to be in [σ 2 , r22 ) so α2 ≥ σ 2 > |σ 2 − 2D|. Invoking Cramér’s theorem [25,
Theorem 2.2.3], we obtain that for sufficiently large n and any positive number δ ∈ (0, 1),
X
n
1
2
X ∈ P̃ ∩ Q̃
Pr
n i=1 i
(151)
≥ exp − n(1 + δ)Λ∗X 2 (α) − exp − n(1 + δ)Λ∗X 2 (r22 )
1
(152)
≥ exp − n(1 + δ)Λ∗X 2 (α) ,
2
where (152) holds since σ 2 ≤ α < r22 and Λ∗X 2 (t) (cf. (26)) is increasing in t for all t ≥ σ 2 (cf. Claim (iii)(c) in Lemma 7).
Invoking the definition of g(·) in (70), the definition of Rsp (·) in (22), and the bounds in (147), (148), (152), we conclude
that for any sequence of (n, M )-codes and any α ∈ [σ 2 , r22 ) if
1
log M ≤ Rsp (α),
n
(153)
1
log Pe,n ≤ (1 + δ)Λ∗X 2 (α).
n
(154)
lim sup
n→∞
then
lim inf −
n→∞
16
Note that σ 2 > |σ 2 − 2D| because σ 2 > D. Recall that Rsp (z) is increasing in z for z ≥ σ 2 (cf. Claim (i)(a) in Lemma 7)
2
and Rsp (σ 2 ) = 12 log σD (cf. Claim (i)(b) in Lemma 7). Hence, the converse proof is complete by letting δ ↓ 0 and noting that
Λ∗X 2 (t) = 0 if t ≤ σ 2 (cf. Claim (iii)(b) in Lemma 7).
D. Achievability Proof for the I.I.D. Gaussian Codebook
Fix α such that α > max{0, 2D − σ 2 }. Invoking the conclusion in (93), for any xn such that
for sufficiently large n and any positive δ,
1
Υ n, kxn k2 ≥ Υ(n, α)
n
≥ exp{−n(1 + δ)Riid (s∗ (α), α)}.
1
n 2
n kx k
Invoking (8), the excess-distortion probability can be upper bounded as follows
i
h
Pe,n = EX n (1 − Pr{d(X n , Y n ) ≤ D})M |X n
Z ∞
=
(1 − Υ(n, z))M fZ (z) dz
0
X
Z α
n
1
Xi2 > α
=
(1 − Υ(n, z))M fZ (z) dz + Pr
n i=1
0
X
Z α
n
M
1
1 − exp{−n(1 + δ)Riid (s∗ (α), α)} fZ (z) dz + Pr
≤
Xi2 > α
n i=1
0
X
n
1
Xi2 > α ,
≤ exp − M exp{−n(1 + δ)Riid (s∗ (α), α)} + Pr
n i=1
≤ α, we have that
(155)
(156)
(157)
(158)
(159)
(160)
(161)
where (158) follows from the definition of Υ(·) in (87); (159) follows since Υ(n, z) ≤ 1; (160) follows from (156); and (161)
follows since (1 − a)M ≤ exp{−M a} for any a ∈ [0, 1).
Recall the definitions of s∗ (·) in (24) and Riid (·) in (23). Choose M such that
log M = n(1 + 2δ)Riid (s∗ (α), α).
(162)
Invoking the definition of Λ∗X 2 (·) in (26), the conclusion in (161) and Cramér’s Theorem [25, Theorem 2.2.3], we conclude
that for sufficiently large n and arbitrary positive δ,
Pe,n ≤ exp − exp nδRiid (s∗ (α), α) + exp − nΛ∗X 2 (α) .
(163)
2
Recall that Riid (s∗ (σ 2 ), σ 2 ) = 12 log σD and Riid (s∗ (z), z) is positive and increasing for z > max(0, 2D − σ 2 ) (cf. Claim
(ii) in Lemma 7. Thus, the first term in (163) vanishes doubly exponentially fast since α > max(0, 2D − σ 2 ). From (162)
and (163), we conclude that for any α > max(0, 2D − σ 2 ), there exists a sequence of (n, M )-codes such that
lim
n→∞
lim inf −
n→∞
1
log M = (1 + 2δ)Riid (s∗ (α), α),
n
1
log Pe,n ≥ Λ∗X 2 (α).
n
(164)
(165)
The proof of the achievability part is done by letting δ ↓ 0 and using the fact that Λ∗X 2 (t) = 0 if t ≤ σ 2 (cf. Claim (iii)(b) in
Lemma 7).
E. Ensemble Converse Proof for the I.I.D. Gaussian Codebook
Fix any α such that α > max{0, 2D − σ 2 }. Using the strong large deviations result in (93), we conclude that for n large
enough and any positive number δ ∈ (0, 1), given any xn such that n1 kxn k2 ≥ α,
1 n 2
Υ n, kx k ≤ Υ(n, α)
(166)
n
≤ exp{−n(1 − δ)Riid (s∗ (α), α)}.
(167)
From (8) and (167), we conclude that for sufficiently large n,
Z ∞
(1 − Υ(n, z))M fZ (z) dz
Pe,n =
0
(168)
17
≥
Z
∞
Zα∞
(1 − Υ(n, z))M fZ (z) dz
(169)
M
1 − exp − n(1 − δ)Riid (s∗ (α), α)
fZ (z) dz
(170)
α
X
n
1
X2 > α ,
(171)
≥ exp − 2M exp − n(1 − δ)Riid (s∗ (α), α) Pr
n i=1 i
a
where (171) follows because i) (1−a)M ≥ exp{−M 1−a
} and ii) for n sufficiently large, exp −n(1−δ)Riid(s∗ (α), α) ≤ 21 .
Using the bound in (171) and Cramér’s theorem [25, Theorem 2.2.3], we conclude that for any α > max(0, 2D − σ 2 ), if M
is chosen such that
≥
log M ≤ n(1 − δ)Riid (s∗ (α), α) − log n − log 2,
then for sufficiently large n, using the inequality exp(−a) ≥ 1 − a for a ∈ [0, 1), we obtain
1
exp − n(1 + δ)Λ∗X 2 (α) .
Pe,n ≥ 1 −
n
(172)
(173)
Hence, given any α > max(0, 2D − σ 2 ), we have shown that for any seqeunce of (n, M )-codes such that
lim sup
n→∞
1
log M ≤ (1 − δ)Riid (s∗ (α), α),
n
(174)
1
log Pe,n ≤ (1 + δ)Λ∗X 2 (α).
n
(175)
we have
lim inf −
n→∞
2
Finally, recall that Riid (s∗ (σ 2 ), σ 2 ) = 12 log σD and Riid (s∗ (z), z) is positive and increasing for z > max(0, 2D − σ 2 )
(cf. Claim (ii) in Lemma 7). Thus, the converse proof is complete by letting δ ↓ 0 and using the fact that Λ∗X 2 (t) = 0 if t ≤ σ 2
(cf. Claim (iii)(b) in Lemma 7).
A PPENDIX
A. Proof of Lemma 5
Note that T (cf. (50)) is finite since E[X 6 ] is finite (cf. (2)). Using the definitions of an in (35), bn in (36), P in (64), and
the Berry-Esseen theorem, we obtain
X
X
n
n
1
1
2
2
2
2
(176)
Pr{Z ∈ P} = Pr
X > σ + bn − Pr
X > σ + an
n i=1 i
n i=1 i
r
r
X
X
n
n
V −1
1
log n
1
2
2
2
2
= Pr
(X − σ ) >
(X − σ ) > V
(177)
Q (ε) − Pr
n i=1 i
n
n i=1 i
n
p
6T
6T
√
−
Q(
(178)
≥ε− √
log
n)
+
n V3/2
n V3/2
1
≥ε+O √ ,
(179)
n
2
where (179) follows since Q(x) ≤ exp{− x2 } and T is finite. Similarly, using the definition of Q in (65) and the Berry-Esseen
theorem, we obtain
X
n
1
2
X + PY − D ≤ 0
(180)
Pr{Z ∈
/ Q} = Pr
n i=1 i
X
n
1
2
2
2
X ≤ σ + (D − PY − σ )
(181)
= Pr
n i=1 i
X
n
1
2
2
= Pr
(182)
X ≤ σ − 2PY
n i=1 i
r
6T
n
PY + √
(183)
≤Q 2
V
n V3/2
6T
2nPY
+√
(184)
≤ exp −
V
n V3/2
18
1
=O √ ,
n
(185)
where (185) follows since the first term in (184) vanishes exponentially fast and is thus dominated by the second term.
Combining (179) and (185), we have
Pr{Z ∈ P ∩ Q} ≥ Pr{Z ∈ P} − Pr{Z ∈
/ Q}
1
≥ε+O √ .
n
(186)
(187)
B. Proof of Lemma 4
For simplicity, in the proof of Lemma 4, for any z ∈ R+ , we let
∗
Riid
(z) := Riid (s∗ (z), z).
(188)
Recalling the results in (29) and (31), we find that the optimal excess-distortion exponents of the spherical and i.i.d Gaussian
codebooks for a given rate R are determined by the corresponding parameters α in Λ∗X 2 (α) (cf. (26)). Using conclusions (i)
∗
and (ii) regarding the increasing properties of Rsp (z) and Riid
(z) in Lemma 7, we obtain that for any given R, there exists
unique αsp (R) and αiid (R) such that,
Rsp (αsp (R)) = R,
∗
Riid
(αiid (R))
= R.
(189)
(190)
From conclusion (iii) in Lemma 7, we know that Λ∗X 2 (t) is an increasing function of t for t ≥ σ 2 . Thus, to prove Lemma 4,
2
it suffices to show that given any rate R > 12 log σD , we have
For any R ≥ Riid (s
∗
(r22 ), r22 ),
αsp (R) < αiid (R).
(191)
αiid (R) ≥ r22 ≥ αsp (R),
(192)
we have
since αsp (R) < r22 for any R ∈ [σ 2 , ∞) (cf. (22)).
2
∗
(r22 )). Recall that both Rsp (z) and Riid (z)
In the following, we will show that (191) also holds for any R ∈ ( 12 log σD , Riid
2
2 2
∗
are increasing in z for z ∈ (σ , r2 ) (cf. Lemma 7). Thus, to prove (191) for R ∈ ( 12 log σD , Riid
(r22 )) is equivalent to show
2 2
that for any z ∈ (σ , r2 ),
∗
Rsp (z) > Riid
(z).
(193)
∗
lim Rsp (z) = ∞ > Riid
(r22 ).
(194)
∗
Rsp (z1 ) = Riid
(r22 ).
(195)
From Lemma 7, we have that
z→∞
Therefore, there exists z1 ∈ (σ 2 , r22 ) such that
∗
Using the increasing nature of Rsp (z) and Riid
(z), we conclude that for any z ∈ [z1 , r22 ), we have
∗
∗
Rsp (z) > Rsp (z1 ) = Riid
(r22 ) > Riid
(z).
(196)
∗
Note that Rsp (z1 ) > Riid
(z1 ). Thus, there exists z2 ∈ (σ 2 , z1 ) such that for any z ∈ [z2 , z1 ), we have
∗
∗
(z).
Rsp (z) > Rsp (z2 ) = Riid
(z1 ) > Riid
(197)
Similarly, we can show that there exists a sequence of values {zi }i≥3 such that zi ∈ (σ 2 , zi−1 ), limi→∞ zi = σ 2 , and for each
z ∈ [zi , zi−1 ), we have
∗
∗
(zi−1 ) > Riid
(z).
Rsp (z) > Rsp (zi ) = Riid
We illustrate the proof idea in Figure 2.
(198)
19
Rate functions
Rsp (z)
∗
Riid
(z)
σ2
z3 z2
z
r22
z1
Fig. 2. Illustration of the proof idea for (193).
C. Proof of Lemma 8
Recall that PY = σ 2 − D. Given any z ∈ R+ , let
h(z) :=
(z + PY − D)2
.
4zPY
(199)
Recall the definition of g(n, z) in (45), we conclude that it suffices to show that for any α ∈ [σ 2 , r22 ), there exists a unique
β ∈ (r12 , |σ 2 − 2D|) such that
h(α) = h(β), α + β ≤ 2σ 2 .
(200)
By simple algebra, we obtain that
z 2 − (σ 2 − 2D)2
∂h(z)
=
,
∂z
4z 2 (σ 2 − D)
h(r12 ) = h(r22 ) = 1.
(201)
(202)
2
2D|, r22 )
2
Note that (201) implies that the function h(z) is increasing in z for z ∈ [|σ −
and decreasing in z for z ∈
(r12 , |σ 2 − 2D|). Hence, using (202), we can conclude that, given any α ∈ [σ 2 , r22 ] ⊆ [|σ − 2D|, r22 ), there exists a unique
β ∈ (r12 , |σ 2 − 2D|) such that h(α) = h(β). Thus, in the following, we just need to show that β ≤ 2σ 2 − α. By noticing that
h(z) ≤ h(α) = h(β) only for all z ∈ [α, β], we conclude that it suffices to show that h(2σ 2 − α) ≤ h(α). To illustrate our
arguments here, we plot h(z) in Figure 3 for the case where σ 2 = 12 and D = 3.
For any α ∈ [σ 2 , r22 ), by calculation, we find that
h(α) − h(2σ 2 − α) =
(α − σ 2 )((α − σ 2 )2 + 4D(D − σ 2 ))
.
2α(σ 2 − D)(α − 2σ 2 )
(203)
Since α ∈ [σ 2 , r22 ), using the definitions of r1 in (43) and r2 in (27), we have r22 < r12 + r22 = 2σ 2 and thus α < 2σ 2 . Recalling
the fact that σ 2 > D, we conclude that
2α(σ 2 − D)(α − 2σ 2 ) < 0.
(204)
20
1
0.95
0.9
h(z)
0.85
0.8
0.75
β
|σ 2 − 2D|
σ2
α
0.7
0.65
0.6
0
5
10
15
20
25
z
Fig. 3. Plot for h(z).
Furthermore, for any α ∈ [σ 2 , r22 ), we have α − σ 2 ≥ 0.
(α − σ 2 )2 + 4D(D − σ 2 ) < (r22 − σ 2 )2 + 4D(D − σ 2 ) = 0.
(205)
Therefore, combining (203), (204) and (205), we conclude that h(2σ 2 − α) ≤ h(α) and thus completing the proof of Lemma
8.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011.
C. E. Shannon, “Coding theorems for a discrete source with a fidelity criterion,” IRE Nat. Conv. Rec, vol. 4, no. 142-163, p. 1, 1959.
A. Lapidoth, “On the role of mismatch in rate distortion theory,” IEEE Trans. Inf. Theory, vol. 43, no. 1, pp. 38–47, 1997.
V. Strassen, “Asymptotische abschätzungen in shannons informationstheorie,” in Trans. Third Prague Conf. Information Theory, 1962, pp. 689–723.
M. Hayashi, “Second-order asymptotics in fixed-length source coding and intrinsic randomness,” IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4619–4637,
2008.
Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56, no. 5, pp. 2307–2359,
2010.
V. Kostina and S. Verdú, “Fixed-length lossy compression in the finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3309–3338,
2012.
A. Ingber and Y. Kochman, “The dispersion of lossy source coding,” in IEEE Data Compression Conference (DCC), 2011, pp. 53–62.
J. Scarlett, V. Y. F. Tan, and G. Durisi, “The dispersion of nearest-neighbor decoding for additive Non-Gaussian channels,” IEEE Trans. Inf. Theory,
vol. 63, no. 1, pp. 81–92, 2017.
A. Lapidoth, “Nearest neighbor decoding for additive non-Gaussian noise channels,” IEEE Trans. Inf. Theory, vol. 42, no. 5, pp. 1520–1529, 1996.
Y. Altuğ and A. B. Wagner, “Moderate deviations in channel coding,” IEEE Trans. Inf. Theory, vol. 60, no. 8, pp. 4417–4426, 2014.
Y. Polyanskiy and S. Verdú, “Channel dispersion and moderate deviations limits for memoryless channels,” in Proc. 48th Annu. Allerton Conf., 2010,
pp. 1334–1339.
V. Y. F. Tan, “Moderate-deviations of lossy source coding for discrete and Gaussian sources,” in IEEE ISIT, 2012, pp. 920–924.
L. Zhou, V. Y. F. Tan, and M. Motani, “Second-order and moderate deviation asymptotics for successive refinement,” IEEE Trans. Inf. Theory, vol. 63,
no. 5, pp. 2896–2921, 2017.
I. Csiszár and J. Körner, Information Theory: Coding Theorems for Discrete Memoryless Systems. Cambridge University Press, 2011.
R. G. Gallager, Information Theory and Reliable Communication. New York: Wiley, 1968.
S. Ihara and M. Kubo, “Error exponent for coding of memoryless Gaussian sources with a fidelity criterion,” IEICE Trans. Fundamentals, vol. 83,
no. 10, pp. 1891–1897, 2000.
K. Marton, “Error exponent for source coding with a fidelity criterion,” IEEE Trans. Inf. Theory, vol. 20, no. 2, pp. 197–199, 1974.
J. Scarlett, “Reliable communication under mismatched decoding,” Ph.D. dissertation, Department of Engineering, University of Cambridge, 2014.
J. Scarlett, A. Martinez, and A. Guillén i Fàbregas, “Mismatched decoding: Error exponents, second-order rates and saddlepoint approximations,” IEEE
Trans. Inf. Theory, vol. 60, no. 5, pp. 2647–2666, May 2014.
R. Gallager, “The random coding bound is tight for the average code (corresp.),” IEEE Trans. Inf. Theory, vol. 19, no. 2, pp. 244–246, 1973.
R. G. Gallager, “Fixed composition arguments and lower bounds to error probability,” MIT, 6.441 Supplementary notes, Jul 1994,
http://www.rle.mit.edu/rgallager/documents/notes5.pdf.
A. D. Wyner, “Random packings and coverings of the unit n-sphere,” Bell Syst. Tech. J., vol. 46, no. 9, pp. 2111–2118, 1967.
N. Bouleau and D. Lépingle, Numerical Methods for Stochastic Processes. Wiley, 1993.
A. Dembo and O. Zeitouni, Large Deviations Techniques and Applications. Springer-Verlag Berlin Heidelberg, 2009, vol. 38.
21
[26] A. Stam, “Limit theorems for uniform distributions on spheres in high-dimensional Euclidean spaces,” J. Appl. Probab., vol. 19, no. 1, pp. 221–228,
1982.
[27] H. Tanizaki, Computational methods in statistics and econometrics. CRC Press, 2004.
| 7cs.IT
|
Reversible DNA codes over a family of non-chain rings
arXiv:1711.02385v1 [math.CO] 7 Nov 2017
Fatmanur Gursoy1 , Elif Segah Oztas2 , Ayten Ozkan1
1
2
Department of Mathematics, Yildiz Technical University, Turkey
Department of Mathematics, Karamanoglu Mehmetbey University, Turkey
([email protected], [email protected], [email protected] )
November 8, 2017
Abstract
In this work we extend results introduced in [15]. Especially, we solve the reversibility problem
for DNA codes over the non chain ring Rk,s = F42k [u1 , . . . , us ]/hu21 − u1 , . . . , u2s − us i. We define
an automorphism θ over Rk,s which helps us both finding the idempotent decomposition of Rk,s
and solving the reversibility problem via skew cyclic codes. Moreover, we introduce a generalized
Gray map that preserves DNA reversibility.
Keywords: Skew Cyclic Codes, DNA Codes, Gray Map, Reversible Codes, DNA k-bases.
2010 Mathematics Subject Classification: 94B15, 94B05, 92D20.
1
Introduction
DNA sequences consist of four bases (nucleotides) which are Adenine (A), Guanine (G), Cytosine (C)
and Thymine (T). Those bases are lined up by the rule Watson-Crick complement (WCC). WCC
ensures that A pairs with T (Ac = T , T c = A) and C pairs with G (C c = G, Gc = C). The attention
into DNA structure and functional properties is attracted by solving a computationally hard (NP)
problem using DNA molecules [3]. With this study, DNA has started to use as an computational tool.
For example; in [4] and [7], authors used DNA to break the Data Encryption System. Also, because
of its natural error correcting structure, DNA has found many applications in coding theory. DNA
single basis( [2, 13, 27, 28]), double bases ([5, 14, 22, 30]), and multiple bases ([6, 16, 23, 24, 25])
over various fields and recently rings have been studied by applying different constraints especially,
Hamming distance constraint. In each case, the reversibility problem has been solved by different
methods.
The reversibility problem while working with sets more than four elements explained by the following example: Let R1,1 be the ring F42 [u1 ]/hu21 −u1 i which has 162 = 44 elements. To make a correspondence between the elements of R1,1 and DNA multiple bases, we need to map each element to a DNA
4-bases. Let (α1 , α2 , α3 ) be a codeword over R1,1 corresponding to ATTCGAACTCCG (a 12-string)
where α1 →ATTC, α2 →GAAC, α3 →TCCG and α1 , α2 , α3 ∈ R1,1 . The reverse of (α1 , α2 , α3 ) is
(α3 , α2 , α1 ), and (α3 , α2 , α1 ) corresponds to TCCGGAACATTC. However, TCCGGAACATTC is not
the reverse of ATTCGAACTCCG. Indeed, the reverse of ATTCGAACTCCG is GCCTCAAGCTTA.
So, basically reversing the entires in the codeword does not lead directly to the reverse of DNA alphabets in general except single alphabet matching.
1
In this study, we focus on extending the problem introduced in [15] and study reversible DNA codes
over Rk,s = F42k [u1 , . . . , us ]/hu21 − u1 , . . . , u2s − us i where ui uj = uj ui . To make a correspondence
between DNA alphabet and the ring Rk,s , we map each ring element to a DNA 2s+1 k-basis (2s+1 kmer). We establish a generalized Gray map that preserves DNA reversibility. Moreover, we obtain
reversible DNA codes by using skew cyclic codes over Rk,s .
The reason of considering the DNA k-bases (k-mers) is that components related to DNA, such as
transcription factor, binding site, genes, proteins etc., are consist of k-bases. DNA researchers focus
on finding the role of the specific k-bases. For example, in [20], the DNA 8-bases (8-mers) are listed
from special areas including binding sites that have important role in opening/closing the genes for
producing proteins. In [20] (Table 6), the frequencies of 8-mers are listed and the top area of the lists
are important to identify binding sites where reversible-complement DNA 8-mers have been shown
to appear intensively. Moreover, finding meaning paths in DNA by using the algebraic codes is an
interesting question for researchers.
2
Preliminaries and definitions
In this section we give basic definition of skew cyclic codes and reversible codes and we introduce
skew cyclic codes over Rk,s . Skew cyclic codes were first introduced by Boucher et al. in [8]. They
generalized cyclic codes by using skew polynomial rings with an automorphism θ over the finite field
with q elements (Fq ). The definition of skew polynomial rings is given below.
Definition 1. [18] Let R be a commutative ring with identity and θ be an automorphism over R. The
set of polynomials R[x; θ] = {a0 + a1 x + ... + an−1 xn−1 |ai ∈ R, n ∈ N} is called the skew polynomial
ring over R where addition is the usual addition of polynomials and the multiplication is defined by
xa = θ(a)x (a ∈ R) and extended to polynomial multiplication naturally.
A skew cyclic code is defined to be a linear code C of length n over Fq satisfying the property
σ(c) = (θ(cn−1 ), θ(c0 ), ..., θ(cn−2 )) ∈ C, for all c = (c0 , c1 , ..., cn−1 ) ∈ C [8]. In skew polynomial
representation of C, we can consider a codeword c = (c0 , c1 , ..., cn−1 ) of C as a polynomial c(x) =
c0 + c1 x + . . . + cn−1 xn−1 ∈ Fq [x; θ]. In this case σ(c) corresponds to the polynomial xc(x) in Fq [x; θ].
If the order of θ, say m, divides n then C corresponds to a left ideal of the quotient ring Fq [x; θ]/(xn −1)
(in [8]), else Fq [x; θ]/(xn − 1) is not a ring anymore but we can consider C as a left Fq [x; θ]−submodule
of Fq [x; θ]/(xn − 1) (in [26]). In both cases C is principally generated by a monic polynomial g(x)
which is a right divisor of xn − 1 in Fq [x; θ]. One of the attractive features of studying skew cyclic codes
is that, factorization of xn − 1 is not unique in Fq [x; θ], so we have much more generating polynomials
in contrast with cyclic codes. Thanks to their rich algebraic structure, many new good codes with
parameters exceeding the parameters of the previously best known linear codes have been obtained
in both [8] and [1]. There are many other studies done by considering skew polynomial rings over
different rings instead of finite fields, such as skew cyclic codes over Galois rings ([9]), finite chain rings
([19]), non-chain rings such as Fq + vFq ([17]) and Fq + uFq + vFq + uvFq ([29]).
In this paper, we study a family of rings Rk,s = F42k [u1 , . . . , us ]/hu21 − u1 , . . . , u2s − us i, where
k, s > 1 and ui uj = uj ui . Rk,s is a commutative non-chain ring with characteristic 2 and cardinality
s
s+1
42k2 = 42 k .
2
We use the following automorphism to define skew cyclic codes over Rk,s . Let
θ :Rk,s −→ Rk,s
k
a −→ a4 ∀a ∈ F42k ,
ui −→ ui + 1 ∀i ∈ {1, . . . , s}.
Then θ is an automorphism over Rk,s of order 2. Naturally Rk,s [x; θ] is a skew polynomial ring and a
skew cyclic code C of length n over Rk,s corresponds to a left Rk,s [x; θ]-submodule of Rk,s [x; θ]/(xn −1).
Whenever g(x) is a right divisor of xn − 1 in Rk,s , the left Rk,s [x; θ]-submodule C = (g(x)) is a skew
cyclic code over Rk,s . But not all skew cyclic codes over Rk,s are principally generated. In this study
we focus on only principally generated ones. The factorization of xn − 1 is not unique in Rk,s [x; θ],
since F42k [x; θ] is a subring of Rk,s [x; θ] and the factorization of xn − 1 is not unique in F42k [x; θ]. So,
there are many different right divisors of xn − 1. This algebraic property provides us an advantage for
finding right divisors with special properties (palindromic and θ−palindromic) in Section 4.
Definition 2. Let C be a code of length n over Fq . If cr = (cn−1 , cn−2 , . . . , c1 , c0 ) ∈ C for all
c = (c0 , c1 , . . . , cn−1 ) ∈ C, then C is called a reversible code.
3
Decomposition of Rk,s
In this section, we introduce a method to find idempotent decomposition of the ring Rk,s . By using
the decomposition of Rk,s we define a Gray map which preserves DNA reversibility.
Our aim is to establish a correspondence between the elements of Rk,s and DNA 2s+1 k-bases in such
a way that the reversibility problem is solved by using skew cyclic codes over Rk,s . First of all we need
to express the elements of Rk,s in a unique way. Let us arrange the elements of the form ur11 ur22 . . . urss
in an order. Let A be the set of such elements, i.e. A = {ur11 ur22 . . . urss |ri ∈ {0, 1}, 1 ≤ i ≤ s}. The
cardinality of A is 2s . Define a map
µ : A −→ Zs
ur11 ur22 . . . urss −→ (r1 , r2 , . . . , rs ).
There are several ways of ordering s-tuples (r1 , r2 , . . . , rs ). We will use the lexicographic order.
Definition 3 ([12]). Let α = (α1 , α2 , . . . , αs ) and β = (β1 , β2 , . . . , βs ) be in Zs>0 . If the leftmost
nonzero entry of the vector difference α − β ∈ Zs is positive then α is greater than β with respect to the
β1 β2
αs
βs
1 α2
lexicographic order, and denote it α >lex β. We write uα
1 u2 . . . us >lex u1 u2 . . . us if α >lex β.
For example; let us put u1 , u2 ∈ R1,3 in order. Since µ(u1 ) = (1, 0, 0), µ(u2 ) = (0, 1, 0) and
(1, 0, 0) − (0, 1, 0) = (1, −1, 0) then u1 >lex u2 .
Here, we rename the elements of Rk,s of the form ur11 ur22 . . . urss where ri ∈ 0, 1 and 1 ≤ i ≤ s, as
follows:
1. First consider the following specific subsets of A;
Aj = {ur11 ur22 . . . urss |exactly j of ri ’s equal 1 the others 0}
where 0 ≤ j ≤ s. So, A0 = {1}, A1 = {u1 , u2 , . . . , us }, A2 = {u1 u2 , u2 u3 . . . , us−1 us }, As =
{u1 u2 . . . us }, etc.
3
2. Arrange the elements of Aj with respect to the lexicographic order, and let Bj be the array
consisting the ordered elements of Aj . For example B0 = [1], B1 = [u1 , . . . , us ], etc.
3. Then concatenate these arrays as B = [B0 , B1 , . . . , Bs ].
4. Rename the ith element of the array B as Ui . For instance, U0 = 1, U1 = u1 , U2s −1 = u1 u2 . . . us
etc.
Define the ordered s-tuples of Rk,s as
T = [µ(U0 ), µ(U1 ), . . . , µ(U2s −1 )] = [T0 , T1 , . . . , T2s −1 ].
(1)
Note that; µ(Ui ) + µ(U2s −1−i ) = Ti + T2s −1−i = (1, 1, . . . , 1) , for 0 ≤ i ≤ 2s − 1.
Example 1. Let us consider the ring R1,3 = F16 [u1 , u2 , u3 ]/hu21 − u1 , u22 − u2 , u23 − u3 i. Then A =
{1, u1, u2 , u3 , u2 u3 , u1 u3 , u1 u2 , u1 u2 u3 } and
B0 = [1], B1 = [u1 , u2 , u3 ],
B2 = [u1 u2 , u1 u3 , u2 u3 ], B3 = [u1 u2 u3 ].
So, B = [1, u1 , u2 , u3 , u1 u2 , u1 u3 , u2 u3 , u1 u2 u3 ] = [U0 , U1 , . . . , U7 ] and the ordered 3 − tuples of R1,3 is;
T = [T0 , . . . , T7 ] = [(0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (1, 0, 1), (0, 1, 1), (1, 1, 1)]. Any element
α ∈ R1,3 can be written uniquely of the form α = a0 U0 + a1 U1 + . . . a7 U7 where αi ∈ F16 . Moreover,
it can be easily seen that Ti + T7−i = (1, 1, 1), for example;
T0 + T7 = µ(U0 ) + µ(U7 ) = µ(1) + µ(u1 u2 u3 ) = (0, 0, 0) + (1, 1, 1) = (1, 1, 1),
T3 + T4 = µ(U3 ) + µ(U4 ) = µ(u3 ) + µ(u1 u2 ) = (0, 0, 1) + (1, 1, 0) = (1, 1, 1).
The following theorem is a result of the Chinese Remainder Theorem;
Theorem 1. Let I = {Ii } be a set of 2s elements of Rk,s . If
(i) Ii2 = Ii ,
(ii) Ii Ij = 0 for i 6= j and 0 ≤ i, j ≤ 2s − 1,
P2s −1
(iii)
i=0 Ii = 1,
then Rk,s = I0 F42k ⊕ I1 F42k ⊕ . . . ⊕ I2s −1 F42k .
Now we construct an idempotent set I satisfying these three conditions above. We associate
some elements of Rk,s with the s-tuple Ti ’s as follows. Let Ti = (r1 , r2 , . . . , rs ). Then define Ii =
θr1 (u1 )θr2 (u2 ) . . . θrs (us ). For example; since T0 = (0, 0, 0, . . . , 0), T1 = (1, 0, 0, . . . , 0) and T2 =
(0, 1, 0, . . . , 0) we have I0 = u1 u2 . . . us , I1 = (u1 + 1)u2 . . . us and I2 = u1 (u2 + 1)u3 . . . us .
Theorem 2. Let I = {Ii = θr1 (u1 )θr2 (u2 ) . . . θrs (us )| Ti = (r1 , r2 , . . . , rs ), 0 ≤ i ≤ 2s − 1}. Then
Rk,s = I0 F42k ⊕ I1 F42k ⊕ . . . ⊕ I2s −1 F42k .
4
Proof. (i) Ii2 = Ii .
Each Ii has the form (uj1 + 1) . . . (ujd + 1)ujd +1 . . . ujs , where {j1 , j2 , . . . , js } = {1, 2, . . . , s}. Since
(uje + 1)2 = (uje + 1) and u2jf = ujf , then Ii2 = Ii .
(ii) Ii Ij = 0 for i 6= j.
Let
Ii = θr1 (u1 )θr2 (u2 ) . . . θrs (us ) → Ti = (r1 , r2 , . . . , rs ),
Ij = θm1 (u1 )θm2 (u2 ) . . . θms (us ) → Tj = (m1 , m2 , . . . , ms ).
Since i 6= j, then Ti and Tj are distinct so, at least one coordinate re 6= me . Without loss of generality,
we may assume that re = 0 and me = 1. Then θre (ue ) = ue and θme (ue ) = ue +1. Since ue (ue +1) = 0,
we have Ii Ij = 0.
P2s −1
(iii) i=0 Ii = 1.
P j −1
The proof is by induction. For s = 0, I0 = 1. Suppose the induction hypothesis 2i=0
Ii = 1
is true for all 0 < j < l. Let us look at the case s = l. Exactly half of the elements of the set
I = {I0 , I1 , . . . , I2l −1 } have ul as a factor and the others have (ul + 1) as another factor.
If we take the subset of I which has idempotents with the factor ul , and delete ul from each
element then we have the idempotent set of Rk,l−1 , say {V0 , V1 , . . . , V2l−1 −1 } thus by induction we
P2l−1 −1
Vi = 1.
have i=0
If we take the subset of I which has idempotents with the factor (ul + 1) , and delete (ul + 1) from
each element then we have the idempotent set of Rk,l−1 , again.
So,
l
2X
−1
i=0
Ii = ul
2l−1
X−1
Vi + (ul + 1)
2l−1
X−1
Vi
i=0
i=0
= ul + (ul + 1) = 1.
Example 2. Let us determine the idempotent set I for the ring R1,3 , given in Example 1. Since
T = [T0 , . . . , T7 ] = [(0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (1, 0, 1), (0, 1, 1), (1, 1, 1)], we have
I0 = u1 u2 u3 ,
I1 = (u1 + 1)u2 u3 ,
I2 = u1 (u2 + 1)u3 ,
I3 = u1 u2 (u3 + 1),
I4 = (u1 + 1)(u2 + 1)u3 ,
I5 = (u1 + 1)u2 (u3 + 1),
I6 = u1 (u2 + 1)(u3 + 1),
I7 = (u1 + 1)(u2 + 1)(u3 + 1).
By using the decomposition of Rk,s in Theorem 2, each element α ∈ Rk,s can uniquely be expressed
as α = α0 I0 + α1 I1 + . . . + α2s −1 I2s −1 , where α0 , α1 , . . . , α2s −1 ∈ F42k . By using this expression we
5
define the following Gray map:
s
ϕ : Rk,s −→ F422k
α −→ (α0 , α1 , . . . , α2s −1 ).
This Gray map is a one-to-one and onto map and can naturally be extended to n-tuples coordinatewise. The following propositon allows us to determine the Gray image of an element of the
form α = b0 U0 + b1 U1 + . . . + b2s −1 U2s −1 . The dot product “·” below denotes the componentwise
multiplication of two vectors, for example, (0, 1, 0) · (1, 0, 0) = (0, 0, 0).
Proposition 1. Let α = b0 U0 + b1 U1 + . . . + b2s −1 U2s −1 ∈ Rk,s where bi ∈ F42k for 0 ≤ i ≤ 2s − 1.
Then ϕ(α) = (α0 , α1 , . . . , α2s −1 ) and
X
bj where 0 ≤ j ≤ 2s − 1 and j satisfies Ti · Tj = (0, 0, . . . , 0).
αi =
j
Proof. Let ϕ(α) = (α0 , α1 , . . . , α2s −1 ). Then α = α0 I0 + α1 I1 + . . . + α2s −1 I2s −1 . Multiplying α by Ii
we have
Ii α = αi Ii = Ii (b0 U0 + b1 U1 + . . . + b2s −1 U2s −1 )
→ αi Ii = b0 U0 Ii + b1 U1 Ii + . . . + b2s −1 U2s −1 Ii
since Ii Ij = 0 for i 6= j. Now let us determine Uj Ii . Let the ordered s-tuple of Ii be Ti = (r1 , r2 , . . . , rs )
and the ordered s-tuple of Uj be Tj = (e1 , e2 , . . . , es ). If rk = ek = 1 for some k ∈ {1, 2, . . . , s} then,
Ii has (uk + 1) as a factor and Uj has uk then Uj Ii = 0. This means bj does not appear in the result
of Ii α. Otherwise, if either rk = ek = 0 or rk 6= ek for all k ∈ {1, 2, . . . , s} then Uj Ii 6= 0 so bj appears
in the result of Ii α. Whenever rkP
= ek = 0 or rk 6= ek for all k ∈ {1, 2, . . . , s}, Ti Tj = (0, 0, . . . , 0).
Hence we can conclude that αi = bj where 0 ≤ j ≤ 2s − 1 and j satisfies Ti · Tj = (0, 0, . . . , 0).
4
Reversible DNA codes over Rk,s
In this section we obtain reversible DNA codes from skew cyclic codes over Rk,s by making use of the
Gray map defined in the previous section.
In [23], Oztas and Siap gave a correspondence between the elements of the field F42k and DNA
k
2k−bases. In order to solve the reversibility problem they mapped each β and β 4 ∈ F42k to the DNA
2k−bases that are reverses of each other. They gave an algorithm to do this mapping. Let us call
s
this map with τ . τ can naturally be extended to a map τ2 from F422k to DNA 2s+1 k- bases as follows;
τ2 (α0 , α1 , . . . , α2s −1 ) = (τ (α0 ), τ (α1 ), . . . , τ (α2s −1 )), where αi ∈ F42k , 0 ≤ i ≤ 2s − 1.
k
In our study β 4 = θ(β) for β ∈ F42k , thus τ (β) and τ (θ(β)) are reverses of each other. Briefly, we
say that β and θ(β) are DNA-reverses of each other.
Let α ∈ Rk,s and ϕ(α) = (α0 , α1 , . . . , α2s −1 ). The corresponding DNA 2s+1 k-bases of α is
τ2 (α0 , α1 , . . . , α2s −1 ) = (τ (α0 ), τ (α1 ), . . . , τ (α2s −1 )) and its reverse is (τ (θ(α2s −1 )), τ (θ(α2s −2 )), . . . , τ (θ(α0 )))
In short, we say (α0 , α1 , . . . , α2s −1 ) and (θ(α2s −1 ), θ(α2s −2 ), . . . , θ(α0 )) are DNA reverses of each other.
Lemma 1. θ(Ij ) = I2s −1−j for all j ∈ {0, 1, . . . , 2s − 1}.
Proof. Let Tj = (r1 , r2 , . . . , rs ). Since Tj +T2s −1−j = (1, 1, . . . , 1), then T2s −1−j = (1−r1 , 1−r2 , . . . , 1−
rs ). Since ri ’s are either 0 or 1 and the order of θ is 2, then θri = θ−ri .
Ij = θr1 (u1 )θr2 (u2 ) . . . θrs (us ) ⇒ θ(Ij ) = θ1−r1 (u1 )θ1−r2 (u2 ) . . . θ1−rs (us ) = I2s −1−j .
6
Following theorem leads us to solve the reversibility problem directly by skew cyclic codes and it
shows that the Gray map ϕ preserves the DNA reversibility.
Theorem 3. Let α ∈ Rk,s . Then the DNA reverse of ϕ(α) is ϕ(θ(α)).
Proof. Let α = α0 I0 + α1 I1 + . . . + α2s −1 I2s −1 . Then ϕ(α) = (α0 , α1 , . . . , α2s −1 ), and
θ(α) = θ(α0 )θ(I0 ) + θ(α1 )θ(I1 ) + . . . + θ(α2s −1 )θ(I2s −1 )
= θ(α0 )I2s −1 + θ(α1 )I2s −2 + . . . + θ(α2s −1 )I0
= θ(α2s −1 )I0 + θ(α2s −2 )I1 + . . . + θ(α0 )I2s −1 .
Thus ϕ(θ(α)) = (θ(α2s −1 ), θ(α2s −2 ), . . . , θ(α0 )) which is exactly the DNA reverse of ϕ(α).
Example 3. In [22], Table-1 gives a correspondence between the elements of F16 and DNA 2-bases.
It has similar properties with the correspondence algorithm given in [23]. Table-1 maps each element
of F16 and its fourth power to the DNA 2-bases that are reverses of each other.
Consider the ring R1,3 , given in Example 1. Let β be a primitive element of F16 and α = β 2 I0 +
βI1 + β 5 I2 + β 3 I3 + I4 + 0I5 + β 7 I6 + I7 ∈ R1,3 . Then, ϕ(α) = (β 2 , β, β 5 , β 3 , 1, 0, β 7 , 1). Using Table-1
given in [22], the corresponding DNA 16-bases is
τ2 (β 2 , β, β 5 , β 3 , 1, 0, β 7 , 1) = (τ (β 2 ), τ (β), τ (β 5 ), τ (β 3 ), τ (1), τ (0), τ (β 7 ), τ (1))
= (GC, AT, CC, AG, T T, AA, GT, T T ).
Also,
θ(α) = I0 + β 13 I1 + 0I2 + I3 + β 12 I4 + β 5 I5 + β 4 I6 + β 8 I7
⇒ ϕ(θ(α)) = (1, β 13 , 0, 1, β 12 , β 5 , β 4 , β 8 )
⇒ τ2 (1, β 13 , 0, 1, β 12 , β 5 , β 4 , β 8 ) = (T T, T G, AA, T T, GA, CC, T A, CG).
Thus, ϕ(α) and ϕ(θ(α)) are DNA reverses of each other.
n
The Gray map ϕ can be extended to n−coordinates. For c = (c0 , . . . , cn−2 , cn−1 ) ∈ Rk,s
we
have ϕ(c) = (ϕ(c0 ), . . . , ϕ(cn−2 ), ϕ(cn−1 )). In this case the DNA reverse of ϕ(c) will be ϕ(θ(c)r ) =
(ϕ(θ(cn−1 )), ϕ(θ(cn−2 )), . . . ϕ(θ(c0 ))) where θ(c) = (θ(c0 ), . . . , θ(cn−2 ), θ(cn−1 )).
n
Definition 4. Let C ⊆ Rk,s
. If the DNA reverse of ϕ(c) exists in ϕ(C), for all c ∈ C, then C or
equivalently ϕ(C) is called a reversible DNA code.
Definition 5 ([16]). Let f (x) = a0 + a1 x + . . . + at xt be a polynomial of degree t over Rk,s . f (x)
is said to be a palindromic polynomial if ai = at−i for all i ∈ {0, 1, . . . , t}. And f (x) is said to be a
θ-palindromic polynomial if ai = θ(at−i ) for all i ∈ {0, 1, . . . , t}.
Let C be a skew cyclic code of length n over Fq with respect to an automorphism θ′ . If the order
of θ′ and n are relatively prime then C is a cyclic code over Fq [26]. Similarly, any skew cyclic code
of odd length over Rk,s with respect to θ is a cyclic code, since the order of θ is 2. For this reason we
restrict the length n to even numbers only.
Theorem 4. Let C = (g(x)) be a skew cyclic code of length n over Rk,s where g(x) is a right divisor
of xn − 1 in Rk,s [x; θ] and deg(g(x)) is odd. If g(x) is a θ-palindromic polynomial then ϕ(C) is a
reversible DNA code.
7
Proof. Let g(x) be a θ-palindromic polynomial and t = n−deg(g(x)). Recall that ϕ(c) and ϕ(θ(c)r ) are
DNA reverses of each other. Note that, we do not distinguish between the vector
representation and the
P
n
i
polynomial representation of a codeword in Rk,s
. For each c ∈ C, c(x) = t−1
i=0 βi x g(x) for some βi ∈
Pt−1
′
′
Rk,s , since C = (g(x)).
the DNA reverse of ϕ(c) is ϕ(c ) where c (x) = i=0 θ(βi )xt−1−i g(x).
P Thent−1−i
′
g(x) ∈ C, then ϕ(c′ ) ∈ ϕ(C). Hence ϕ(C) is a reversible DNA
Since c (x) =
i θ(βi )x
code.
Theorem 5. Let C = (g(x)) be a skew cyclic code of length n over Rk,s where g(x) is a right divisor of
xn − 1 in Rk,s [x; θ] and deg(g(x)) is even. If g(x) is a palindromic polynomial then ϕ(C) is a reversible
DNA code.
Proof. Let g(x) be a palindromic polynomial and t = n − deg(g(x)). For each c ∈ C, c(x) =
Pt−1
xi g(x) for some βi ∈ Rk,s , since C = (g(x)). Then the DNA reverse of ϕ(c) is ϕ(c′ ) where
i=0 βiP
t−1
t−1−i
g(x).
c′ (x) = i=0 θ(βi )x
P
t−1−i
′
g(x) ∈ C, then ϕ(c′ ) ∈ ϕ(C). Hence ϕ(C) is a reversible DNA
Since c (x) =
i θ(βi )x
code.
In Remark 1 of [15], an illustration of proofs has been given for the case k = 1 and s = 2, i.e. for
the ring R1,2 = F16 + u1 F16 + u2 F16 + u1 u2 F16 . This simple method works for all rings Rk,s .
Example 4. Let us consider the skew polynomial ring R1,3 [x; θ]. Let β be a primitive element of F42 ,
then
x6 − 1 = h(x)g(x) =[1 + (β(u2 + u3 ) + β 7 )x + (β(u2 + u3 ) + β 7 )x2 + x3 ]
[1 + (β(u2 + u3 ) + β 7 )x + (β 4 (u2 + u3 ) + β 13 )x2 + x3 ] ∈ R1,3 [x; θ].
Since g(x) = 1 + (β(u2 + u3 ) + β 7 )x + (β 4 (u2 + u3 ) + β 13 )x2 + x3 is a θ-palindromic polynomial with
odd degree then C = (g(x)) is a reversible DNA-code over R1,3 .
Example 5. Let us consider the skew polynomial ring R1,3 [x; θ]. Let β be a primitive element of F42 ,
then
x6 − 1 = h(x)g(x) =[1 + (β 14 + u1 + u2 )x + x2 ]
[1 + (β 14 + u1 + u2 )x + (β 14 + u1 + u2 )x3 + x4 ] ∈ R1,3 [x; θ].
Since g(x) = 1 + (β 14 + u1 + u2 )x + (β 14 + u1 + u2 )x3 + x4 is a palindromic polynomial with even degree
then C = (g(x)) is a reversible DNA-code over R1,3 .
5
Conclusion
In this study, we solve the reversibility problem for DNA codes over non-chain ring Rk,s . We use skew
cyclic codes which provide a direct solution for finding reversible DNA codes. This method that is
based on non-commutativity property of the ring has proven to be more compact and more feasible
than the commutative counterparts (e.g. [5, 22, 23]). Generalization of Rk,s = F42k [u1 , . . . , us ]/hu21 −
u1 , . . . , u2s −us i by changing powers of ui (i = 1, 2, ...s) and obtaining optimal codes are open problems.
There are just a few papers on k-mer based DNA codes including this one. Investigating the reversibility
problem over different rings is an interesting problem. Further, relating these theoretical findings to
real DNA data is even more interesting and challenging problem.
8
References
[1] T. Abualrub, A. Ghrayeb, N. Aydin and I. Siap, “On the construction of skew quasi-cyclic codes”,
IEEE Trans. Inform. Theory, vol. 56, no. 5, pp. 2080–2090, 2010.
[2] T. Abulraub, A. Ghrayeb and X. Nian Zeng, Construction of cyclic codes over GF (4) for DNA
computing, J. Franklin Inst. Vol. 343, No. 4-5, pp. 448–457, 2006.
[3] L. Adleman, Molecular computation of solutions to combinatorial problems, Science, New Series,
Vol. 266, pp. 1021–1024, 1994.
[4] L. Adleman, P. W. K. Rothemund, S. Roweis, E. Winfree, On applying molecular computation
to the Data Encryption Standard, J. Comput. Biol. 6(1) (1999) 53–63.
[5] A. Bayram, E.S. Oztas, I. Siap, Codes over F4 + vF4 and some DNA applications, Designs,Codes
and Cryptography, DOI 10.1007/s10623-015-0100-8, (2015).
[6] N. Bennenni, K. Guenda, S. Mesnager, ”DNA cyclic codes over rings”, Advances in Mathematics
of Communications, vol. 11, pp. 83, 2017
[7] D. Boneh, C. Dunworth, R. Lipton, Breaking DES using molecular computer, Princeton CS Tech–
Report, Number CS–TR–489–95, 1995.
[8] D. Boucher, W. Geiselmann and F. Ulmer, Skew cyclic codes, Appl. Algebra Eng. Comm., Vol.
18, No. 4, pp. 379–389, 2007.
[9] D. Boucher, P. Solé and F. Ulmer, Skew constacyclic codes over Galois rings, Adv. Math. Commun., 2 (2008), 273–292.
[10] D. Boucher and F.Ulmer, A note on the dual codes of module skew codes, Vol. 7089, pp. 230–243,
2011.
[11] D. Boucher and F. Ulmer, Coding with skew polynomial rings, J. Symb. Comput., Vol. 44, No. 12,
pp. 1644–1656, 2009.
[12] D. A. Cox , J. Little , D. O’Shea, Ideals, Varieties, and Algorithms: An Introduction to Computational Algebraic Geometry and Commutative Algebra, 3/e (Undergraduate Texts in Mathematics), Springer-Verlag New York, Inc., Secaucus, NJ, 2007.
[13] K. Guenda and T. A. Gulliver, Construction of cyclic codes over F2 + uF2 for DNA computing,
Appl. Algebra Eng. Commun. Comput, 24 (2013), 445–459.
[14] K. Guenda, T. A. Gulliver and P. Solé, On cyclic DNA codes, in Proc. IEEE Int. Symp. Inform.
Theory, Istanbul, 2013, 121-125.
[15] F. Gursoy, E. S. Oztas, I. Siap, Reversible DNA Codes over F16 + uF16 + vF16 + uvF16 , Adv.
Math. Commun. vol. 11 no. 2, (2017) 307–312.
[16] F. Gursoy, E. S. Oztas, I. Siap, Reversible DNA Codes using Skew Polynomial Rings, Appl.
Algebra Engrg. Comm. Comput., 28(4) (2017) 311–320.
9
[17] F. Gursoy, I. Siap, B. Yildiz, Construction of skew cyclic codes over Fq + vFq , Adv. Math.
Commun., Vol. 8 No. 3, pp. 313–322, 2014.
[18] N. Jacobson, Finite-Dimensional Division Algebras over Fields, Springer, Berlin, 1996.
[19] S. Jitman, S. Ling and P. Udomkavanich, “Skew constacyclic codes over finite chain rings”, Advances in Mathematics of Communications, vol. 6, no. 1, pp. 39–63, 2012.
[20] J. Lichtenberg, A. Yilmaz, J. Welch, K. Kurz, X. Liang, F. Drews, K. Ecker, S. Lee, M. Geisler,
E.Grotewold, L. Welch, The word landscape of the non-coding segments of the Arabidopsis
thaliana genome, BMC Genomics 10(1) (2009) 463–495.
[21] B. R. McDonald, Finite rings with identity, Marcel Dekker Inc., New York, 1974.
[22] E.S. Oztas, I. Siap, Lifted polynomials over F16 and their applications to DNA Codes, Filomat,
Vol. 27, No. 3, pp. 459–466, 2013.
[23] E.S. Oztas, I. Siap, On a generalization of lifted polynomials over finite fields and their applications
to DNA codes, Int. J. Comput. Math. Vol. 92, No. 9, pp. 1976–1988, 2015.
[24] E.S. Oztas, B. Yildiz, I. Siap, A novel approach for constructing reversible codes and applications
to DNA codes over the ring F2 [u]/(u2k − 1), Finite Fields Appl. 46 (2017) 217–234.
[25] E.S. Oztas, B. Yildiz, I. Siap, On DNA codes from a family of chain rings, J. Algebra Comb.
Discrete Struct. Appl., vol. 4, no. 1, 93–102, 2017.
[26] I. Siap, T. Abualrub, N. Aydin and P. Seneviratne, Skew cyclic codes of arbitrary length, Int. J.
Inform. Coding Theory, Vol. 2, No. 1, pp. 10–20, 2011.
[27] I. Siap, T. Abulraub and A. Ghrayeb, Similarity cyclic DNA codes over rings, International
Conference on Bioinformatics and Biomedical Engineering, in Shanghai, PRC, May 16-18th 2008.
[28] I. Siap, T. Abulraub and A. Ghrayeb, Cyclic DNA codes over the ring F2 [u]/(u2 − 1) based on
the deletion distance, J. Franklin Inst., Vol. 346, No. 8, pp. 731–740, 2009.
[29] Ting Yao, Minjia Shi, Patrick Solé, Skew cyclic codes over Fq + uFq + vFq + uvFq , J. Algebra
Comb. Discrete Struct. Appl. 2(3) (2015) 163–168.
[30] B. Yildiz, I. Siap, Cyclic codes over F2 [u]/(u4 − 1) and applications to DNA codes, Comput. Math.
Appl., Vol. 63, No. 7, pp. 1169–1176, 2012.
10
| 7cs.IT
|
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
Optimization of Test Case Generation using Genetic
Algorithm (GA)
Ahmed Mateen
Marriam Nazir
Salman Afsar Awan, PhD
Department of Computer
Science,
University of Agriculture
Faisalabad, Pakistan
Department of Computer
Science,
University of Agriculture
Faisalabad, Pakistan
Department of Computer
Science,
University of Agriculture
Faisalabad, Pakistan
ABSTRACT
Testing provides means pertaining to assuring software
performance. The total aim of software industry is actually to
make a certain start associated with high quality software for
the end user. However, associated with software testing has
quite a few underlying concerns, which are very important
and need to pay attention on these issues. These issues are
effectively generating, prioritization of test cases, etc. These
issues can be overcome by paying attention and focus.
Solitary of the greatest Problems in the software testing
area is usually how to acquire a great proper set associated
with cases to confirm software. Some other strategies and
also methodologies are proposed pertaining to shipping care
of most of these issues. Genetic Algorithm (GA) belongs to
evolutionary algorithms. Evolutionary algorithms have a
significant role in the automatic test generation and many
researchers are focusing on it. In this study explored software
testing related issues by using the GA approach. In addition
to right after applying some analysis, better solution produced,
that is feasible and reliable. The particular research presents
the implementation of GAs because of its generation of
optimized test cases. Along these lines, this paper gives
proficient system for the optimization of test case generation
using genetic algorithm.
Keywords
Optimization, Genetic Algorithm, Test case, Generation,
Design, Testing.
1. INTRODUCTION
Computer software assessment is one of the majorities of
labor strenuous as well as pricey period with the software
program improvement lifetime routine. Computer software
assessment consists of the test circumstance, age group as
well as test suite optimization that includes a strong impact on
the particular usefulness as well as productivity connected
with software program assessment. In the last number of ages,
there was an energetic investigation to automate the task
connected with the test circumstance, age group, however the
tribes are already confined by the dimension and the
complexity connected with the software program. The high
quality software must satisfy the user requirements and
customer demands. Satisfaction of customer has always been
important because it indicates that business people have to
manage and improve their product. Once testing involving
software is actually a good important process connected
with assessing the software to help distinguish it is quality.
It is an important area of the software engineering.
Modern software systems are extremely reliable in addition to
correct.
Represented an optimization tool where their aim is to find a
problem solution to a given problem. Based on inheritance,
natural selection, mutation, and sexual reproduction, they try
to give after many generations the optimal solution in a finite
time [1]. Testing techniques are test case design method. Test
cases are developed using various testing techniques to
achieve more effective testing of application [2]. Genetic
algorithms are generally as outlined by evolutionary ideas
connected with natural menu in addition to genetics. Genetic
algorithms solve the current problems step via step as well as
provide and then generation [3].
Testing is a basic activity of the item headway handles, and
robotic test period adds to reduce cost and time trials. The
ideal test suites are conceived by the strategy for examining
measurements. Testing is the most essential stage in the item
change life cycle. The testing stage is the last channel for all
oversights of rejection and commission. Testing writing
computer programs are altogether more eccentric than
rehearsing a framework to check whether it works. Each
study, audit, survey, walks around, social event code sees, all
is when in doubt a kind of test. The more fruitful that can
make early is static attempting, the less issues involved in the
dynamic periods of testing. IT has shown again and again that
the earlier a defect recognized and cleared, the lower the
additional change cost associated with ousting the mix-up.
Plan for testing begin when each item thing is portrayed. All
evolutionary algorithms similar to Genetic Algorithm acquire
near optimal solution.
The essential objective of testing is to exhibit that the item
thing at any rate, meets a course of action of pre-set up
affirmation criteria under a suggested set of environmental
circumstances. There are two sections of this objective. The
essential fragment is to show that the requirements point of
interest from which the item was laid out is correct. The
second portion is to show that the setup and coding precisely
responds to the necessities. Precision suggests that limit,
execution, and timing necessities match affirmation criteria.
The existing Genetic Algorithm (GA) starts receiving from
building a principal population connected with folks, each
manifested via randomly earned genotype. The present health
of people is normally examined in a lot of problemsdependent technique, and also the GA make an attempt that
can help progress highly suit persons from the first
population. Algorithms were created via Ruben Netherlands
about the sixties and also were produced from the Netherlands
as well as the students as well as acquaintances about the
University about the state of Michigan. Holland’s unique
objective feel, not necessarily every single child style
algorithms every single child remedy were produced
difficulties, but alternatives that can help basically examine
the occurrence about adaptation Every bit as that happens
with mother nature and also every single child generate
methods that the mechanisms concerning pure adaptation
6
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
might be imported with pass [4].
The genetic algorithms tend to base towards the tip of genetics
and evolution. Holland proposed GA as being a heuristic
method as outlined by “Survival of any fittest”. A genetic
algorithm is actually a good evolutionary algorithm, which
my partner and solve optimization problem. Then receive
approximate merchandise in order to optimize ailments
within GA. The genetic algorithm loops greater than a great
iteration method to make the current population evolve [5].
The genetic algorithm is actually a stochastic search technique
that is to base on the idea of the menu of any fittest
chromosome. Throughout genetic algorithm, population of
chromosomes represented as coming from other codes
including binary, real number, permutation etc. genetic
operators (i.e. Selection, crossover, mutation) is usually
applied for the chromosome in order to obtain extra fittest
chromosome. The fitness of the chromosome is defined by a
great proper objective function. As has been a class associated
with the stochastic method, genetic algorithm is another
coming from a random search. Whilst genetic algorithm
executes a good multidimensional search by maintaining a
population involving potential user, random actions
consisting of the combination regarding iterative search
steps as well as simple random search steps probably
acquire a solution for an issued problem [6].
2. PREVIOUS WORK
The testing, including execution of a system by a few
arrangements of test information and contrast the outcomes
and expected results are called programming testing.
Producing of robotic test information is extremely
troublesome errand in item arranged project since legacy,
strategy superseding, polymorphism and formats demonstrate
numerous coupling abnormalities because of element conduct
of articles. Class is the fundamental building square of article
situated programming. Customary testing like basic testing,
practical determination based testing and heuristics testing are
utilized for it. Auxiliary, testing is vital on the grounds that it's
found the bugs in codes by control stream testing, way scope
testing, information stream testing. Utilitarian testing meets
the necessities and detail of programming. Heuristics testing
system test the theoretical classes. Programmed test
information is created in article situated programming from
generally code-based procedure and model-based or plan
based strategy [7].
A good technique is usually proposed to abilities in order to
prioritize check situation scenarios via figuring out the
personal vital way groups utilizing ancestral algorithm. The
individual check situation scenarios usually are derived from
the UML task diagram along with point out chart diagram.
The current screening performance is really optimized for
using people ancestral algorithm for the check facts. The
personal points movement metric is really implemented of the
execute intended for computing the data movement intricacy
relevant to every node from the task diagram together with
point out chart diagram. Whether the software Requirements
alter, people software requirements to assist always modified
and also require retesting of the software [8].
Test information usually are generated with most of these,
a way they may conduct every record at the least once.
Genetic Algorithms applied to path testing no matter
whether the target paths are generally clearly defined, as
well as a good suitable fitness performs concerning this
goal is usually built. GA involves the amount of factors,
which happen to commanded to end up being collection
consequently sizing of any inhabitants likewise pushed to end
up being collected. Human population sizing offers incredible
result for that GA swiftness to uncover the best possible
answer [9].
Automated generation connected with test cases to evaluate
software product is usually much needed. Equally the item
probably decreases
the
night out
as well as costs
associated with the testing process. Whilst
guidebook
testing is actually very night out consuming and also
costly, Software solutions are right now turning towards
the automated testing tools as well as techniques. Inside
the paper, explored how Genetic Algorithm, improves
quality as well as reliability of any software through
bringing in
optimized test cases. Three basic steps
responsible regarding GA are (1) selection of initial
chromosomes through the population, (2) performing
crossover from exchanging the facts between choosing
chromosomes and also (3) performing a mutation operation
towards the selected genes. The actual paper furthermore
provides the automated software testing architecture [10].
Genetic Algorithm (GA), Particle Swarm Optimization (PSO)
in addition to a good hybrid Genetic Particles warm system
algorithm (HGPSTA) has been
designed
for fitness
operate which is according to the dominance relation
between only two nodes. The fitness performs based on the
criteria connected with
facts flow coverage. This is
accepted throughout the research, depending to the
dominance relation between nodes involving details flow
graph. The main goal associated with research, in order to
combine the power involving two algorithms and also
PSO. This proves their power and effectiveness on the
solving the current testing disorders [11].
Examination data creation is basically versions process
regarding determining an incredible collection involving data
that meet versions standards collection created for testing. The
large amount of research is carried out as a result of several
scientists as well as developed several test information
devices similar to arbitrary test data devices, symbolic test
data devices along with powerful test facts devices. The
particular document utilized the set analyze of the test
scenario creation depends on the hereditary formula and in
addition builds test situations [12].
GA inspired from Darwin’s theory information about
evolution. Genetic algorithm functionalities rule involving
menu to evolve a set connected with goods along with
looking for a great optimum solution. Genetic algorithms
simulate the survival of the fittest among men and women
in excess of generations regarding solving a good problem
throughout nature competition among persons results on the
fittest persons dominating the weaker ones. GAs uses
several operations like selection, crossover and mutation [13].
The software testing remains the existing primary course of
action, accustomed to obtain consumers’ confidence with the
software besides categorized the personal check instances
making use of stratified sampling. Anatomical formula (GA)
offers a general-purpose search system, where benefits
guidelines with regards to pure progress. Within the paper,
genetic algorithm considered pertaining to producing check
selection coming from the individually distinct established
linked to check instances. That's why; this is a big challenge.
The kind of utilizing a great genetic algorithm on the inside
software testing creates a new optimized check selection. The
advancement connected with tactics that moreover help types
7
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
automation linked with software testing certain outcome using
substantial financial savings [14].
The use of meta-heuristic worldwide lookup approaches for
software program test info age group is the particular
concentrate connected with research workers in recent times.
Numerous fresh methods as well as cross strategies have also
been planned to handle the challenge more effectively [15].
3. PROPOSED WORK
The real purpose of this work is to provide better optimization
approach, which is introduced in the test case by using
Genetic Algorithm. Optimization approach accommodated
different scale projects to inspect. This research also provides
a survey to determine better quality testing process within the
time.
For testing many techniques used by other researchers to
be able to
accomplish
the effective testing process,
pertaining to guarantee associated with
much better
performance along with quality. But those have some
limitations. In order to address these issues, there is a study
that analyzes how test cases can be optimized and gives best
solution.
Evolutionary tests generally are a very good growing
methodology associated with routinely bringing in high
quality analyzes information. The actual evolutionary
algorithms are now put on inside many correct living
problems. GA is solitary these kinds of evolutionary criteria.
GA offers emerged to become a practical, powerful marketing
process together with seeks procedure. A fantastic GA is
usually a seek criteria. This is prompted with the approach
character evolves variety using an organic food selection of
just about any fittest men and women. Any analyses of unique
computer software test methods probably executed, to confirm
in which GA is successfully used.
Inject four mutants in this program
Now algorithm look like this.
1. Power(a, b)
2. If(a=1)
3. Return 1
4. If(b=1)
5. Return a
6. P=1, i=1
7. While(i<b)
8. {P=P+a
9. i++}
10. return P
According to the optimized algorithm flow as shown in the
figure 1, found the number of mutants by applying this
algorithm. First of all, some mutants added in a program.
Mutants are some errors, which encouraged finding the
optimized test case. Then evaluate the test case performance
and the first step was initialized. After that considered, all
mutants not found. The question was what is the number of
mutants found? For this purpose next step is the evaluation of
mutant numbers found. If found umber is less than the
minimum number of mutants then it means test case failed to
find errors. On the other hand, the condition is no, then
calculated the fitness by finding the exact numbers of mutants.
On the next step if mutant numbers found more than 50%,
then it displays best fit and if found mutants are less than 50%
mutants then it displays mutant number's value. If the total
number of mutants found in the 2nd step, then no more steps
needed to find errors, just end. Thus the test case provided the
optimal solution by finding all the errors.
3.1 Experimental design
In this research first, generated random test cases. Applied
mutates testing to check it. If satisfied, then stop.
Optimized algorithm
1. Inject the mutant in the program.
2. Generate random test cases.
3. Find the mutation score with the formula, mutation score=
(number of mutants found) / (total number of mutants).
4. If the mutant score is satisfactory (Maximum) stop,
otherwise go to step 5.
5. Refine the test case using mutation score. Test case, having
mutation score 20% or less drops them.
6. Apply Genetic Algorithm Operations on remaining test
cases to produce new experiments. Go to step 3.
Algorithm for ab. Where a and b are positive numbers.
1. Power(a, b)
2. If(a= =1)
3. Return 1
4. If(b= =1)
5. Return a
6. P=1, i=1
7. While(i<=b)
8. {P=P*a
9. i++}
10. return P
Fig 1: Optimized algorithm flow
8
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
There has to characterize an approach to decide the nature of
every experiment as an issue of reviewing. It is seen that the
understudies are a solid propensity to rehash botches. On the
off chance that they missed one class, they likely did that for a
large portion of the experiments. At that point saw designs
among a significant number of the understudies, on why
fizzled. Each demonstrating an absence of understanding why
the relating learning is vital for experiment outline:
A. Regulated depiction of experiment execution
B. Make particulars of substantial and invalid inputs
C. Understanding directions/level of points of interest
D. Elaborate experiment creation, and not just utilizing the
most evident experiment or info
E. Comprehend the motivation behind the framework and
current level and connection of testing
F. Characterize a reasonable beginning position for the
experiment
G. Understanding test techniques and how to apply them
H. Suppositions, e.g. concerning accuracy and culmination of
particulars
I. Experiment assessment (ventures to take to make an
unmistakable correlation with expected result ought to be
clear)
J. Tidy up after an experiment, repeatability
These factors are playing very important role in test cases. By
analyzing it finds that elaborated factors having different level
of influence on test cases. All these factors have different
impact ratio and these entire ratio described in fig. 2.
Table 1. Factors and impact ratio
Factors
Regulated depiction
Valid & invalid inputs
Good Detail level
Variation in test cases
Understand system's framework
Clear beginning position
Test design techniques better
Accurate suppositions
Test case assessment
Tidy up after execution
Impact (%)
55%
38%
44%
75%
50%
60%
80%
70%
50%
78%
A. Regulated depiction of experiment
execution
An experiment system is regularly portrayed as an
arrangement of activities, with an orderly depiction of
activities (and possibly additionally halfway reactions). To
make the execution way exceptionally characterized, the
experiment, frequently portrayed in little and extremely
itemized strides, the very same route as composing code.
Impact of regulated depiction on test cases
60
50
40
category A
30
category B
20
category C
10
0
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 2: Impact of regulated depiction
This graph illustrates regulated depiction impact on different
test cases. Here are four sets of test cases, in which maximum
impact of regulated depiction is 55%. Each set contains three
test cases; in the 1st set impact of the regulated depiction of
test cases are respectively 55, 55 and 49. In the 2nd set,
impact of the regulated depiction is respectively 51, 55 and
48. In the 3rd set, impact is respectively 50, 53 and 55. While
in the 4th set impact is respectively 54, 48 and 44.
Approximately, 55 % of the experiments gave deficient data
and subtle element to give steps that are unambiguously
trailed by other human or required personality jumps that are
included while making a system for execution.
B. Substantial and invalid inputs
Portraying data classes in the test/affirmation, determination
enhances the usage of different test diagram methodologies,
and the data decision of the test. This is a proficient approach
to catch the whole info area, furthermore get ready for a
progression of test outline strategies. In the meantime,
acquaint variables with speak to inputs, in this manner
preparing for test computerization. At first hard to
characterize what an information is in this setting, subsequent
to tapping on a predefined menu-thing can at another
reflection level be seen as data.
Impact of input analysis on test case
performance
40
35
30
25
category A
20
category B
15
category C
10
5
0
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 3: Impact of input analysis
This graph describes the impact of input analysis on different
test cases. Here are four sets of test cases, in which maximum
impact of input analysis is 38%. Each set contains three test
9
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
cases; in the 1st set impact of the input analysis of test cases
are respectively 38, 37 and 33. In the 2nd set, impact of the
input analysis is respectively 36, 38 and 37. In the 3 rd set,
impact is respectively 33, 34 and 38. While in the 4th set
impact is respectively 31, 37 and 33. This turned out to be
extremely troublesome for the analyzers, and just 62 % were
even close to the thought proposed with information
examination. The understudies were by and large preferred at
giving legitimate contribution over making an examination on
invalid information. None succeeded to get a handle on the
whole informational space. In this way its impact is 38%.
C. Understanding directions/level of points
of interest
A tester must ready to construe precisely what is implied, and
in a nutty gritty level take after directions and give enough
data, so that the experiment is unambiguous and can be
rehashed by some other tester.
Lack of variation in test cases
80
70
60
50
category A
40
category B
30
category C
20
10
0
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 5: lack of variation
Impact of detail level on test cases
45
44
43
42
41
40
39
38
37
36
35
category A
category B
category C
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 4: Impact of detail level
This study indicates that 29% of the subjects did not read the
direction of conveyance of experiment and 15% did not finish
the whole layout, so it's impact 44%. This graph illustrates the
detail level impact on different test cases. Here are four sets of
test cases, in which maximum impact of detail level is 44%.
Each set contains three test cases; in the 1st set impact of the
detail level on test cases are respectively 41, 38 and 40. In the
2nd set, impact of the detail level is respectively 40, 44 and
42. In the 3rd set, impact is respectively 44, 38 and 40. While
in the 4th set impact is respectively 39, 39 and 43.
D. Not just utilizing the most evident
experiment
This graph demonstrates lack of variation impact on different
test cases. Here are four sets of test cases, in which maximum
impact of lack of variation is 75%. Each set contains three test
cases; in the 1st set impact of the lack of variation on test
cases are respectively 70, 49 and 44. In the 2nd set, impact of
the lack of variation is respectively 75, 75 and 75. In the 3rd
set, impact is respectively 56, 75 and 71. While in the 4th set
impact is respectively 75, 74 and 75. Only a few individuals,
25% make any variety inside the framework, or endeavored
anything inventive with their experiments. The analyzer often
times attempted the limit make a record and the assortment
was greatly obliged (generally names and numbers were
attempted). The second most fundamental test was looking
dated. A couple endeavored to look at a few exchanges which
prompted more important tests with marginally higher scope.
75% neglected to endeavor any variety.
E. Comprehend the motivation behind the
framework and current level and
connection of testing
Framework effect is vital in numerous viewpoints for
comprehension levels and association e.g., detectable quality
of the region is possible, what programming thought is used,
and how that affect the test approach. Understanding the
explanation behind the structure, and current level and
association of the system is related to the reflection levels of
the system. This is probably the most feathery and hard to
handle the thought of a product framework with regards to
testing and is by all accounts an understanding that
individuals secure after a few years working with the
framework.
Amid examination of the few many experiments, it is
expected to see the assortment of experiments made.
Especially, requested that the analyzers are imaginative in
designing legitimate experiments for the framework. Every
product framework has an assortment of info conditions, for
example, the quantity of complete experiments planned
exceptionally awesome, in the testing conditions which
consider all mixes of inputs, so outlined test cases by utilizing
the strategy which can depict a blend of numerous conditions
and create numerous activities correspondingly.
10
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
This is obviously effective identified amid mechanization of
the experiment that data is lacking to distinguish where the
experiment began. The most widely recognized supposition
for this action, which targets test of a little program, is
tolerating that data, for example, "Begin the system" is
sufficient.
Impact of "system's framework
understanding" on test cases
52
50
48
46
category A
44
category B
42
category C
40
38
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 6: Impact of system’s framework understanding
Measuring this appreciation is fairly troublesome. It is
investigated what number of comprehend that all
contributions to the framework were string based, and did not
make test cases that is tolerating the test simply handle digits
and letters. As much as half of the trials failed on this record
(half), which incited a lion's offer of experiments fizzled. This
graph shows the system’s framework impact on different test
cases. Here are four sets of test cases, in which maximum
impact of the system’s framework is 50%. Each set contains
three test cases; in the 1st set impact of the system’s
framework on test cases are respectively 50, 50 and 43. In the
2nd set, impact of the system’s framework is respectively 49,
49 and 50. In the 3rd set, impact is respectively 46, 48 and 42.
While in the 4th set impact is respectively 47, 44 and 50.
G. Understanding test techniques and how
to apply them
This study comprises of the test plan methods and points of
interest, cover and variations, inside and out. The first and
most clear level is to comprehend what the hypothesis is, and
after that the capacity to apply the strategy in the particular
case in the framework, which means, finding an area or
circumstance where can apply the procedure. Most test outline
procedures are identified with information, some are
identified by way of execution, and few are identified with
request of execution. Likewise blends of methods are
conceivable.
Impact of mis-usage of test design
techniques on test cases
82
80
78
76
74
72
70
category A
category B
category C
Test Case Test Case Test Case Test Case
(set 1) (set 2) (set 3) (set 4)
F. Fix beginning position of the experiment
A common place foul up is to simply depict where the starting
position of the test is, i.e., not being specific to the most
proficient method to get to the beginning position or which
moves must be made before; or simply expecting that a
specific area is evident from the experiment connection, or not
saying anything by any means.
Failure rate due to not defining the
beginning position
70
60
50
40
30
20
10
0
category A
category B
category C
Test Case Test Case Test Case Test Case
(set 1) (set 2) (set 3) (set 4)
Fig 7: Impact of defining beginning position
60% of the test cases fizzled on clarifying an unambiguous
beginning position. This graph elaborates failure rate due to
not define the beginning position in different test cases. Here
are four sets of test cases, in which maximum impact of
defining the beginning position is 60%. Each set contains
three test cases; in the 1st set impact of the beginning position
on test cases are respectively 57, 60 and 60. In the 2nd set,
impact of the beginning position is respectively 60, 53 and 60.
In the 3rd set, impact is respectively 55, 59 and 45. While in
the 4th set impact is respectively 59, 60 and 60.
Fig 8: Impact of miss-usage the techniques
The positive experiment system (e.g. giving significant
information), was the most broadly perceived technique
performed, besides the most shocking accomplishment in
making an executable trial. Just 20 % honored that for BVA
every one of three information is executed, regardless of the
fact that was highlighted amid classroom instructing. In this
manner, it's impact is 80%. This graph describes miss usage
technique’s impact on different test cases. Here are four sets
of test cases, in which maximum impact of miss usage
technique is 80%. Each set contains three test cases; in the 1st
set impact of the miss usage technique on test cases are
respectively 80, 75 and 77. In the 2nd set, impact of the miss
usage technique is respectively 78, 79 and 80. In the 3 rd set,
impact is respectively 74, 80 and 80. While in the 4th set,
impact is respectively 77, 78 and 79.
H. Importance of suppositions
This classification identifies with on which suppositions
noticed that an examination has passed or failed. An expert
analyzer is inclined to make judgments on exactness and
zenith of all parts of the framework. On account of judging a
defective necessity, specialists more slanted to expect that the
prerequisite is off base and changed. The beginner accepted
that what is composed is quite often right, and plan the
experiment taking into account this defective supposition.
11
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
Impact of unrealistic supposition on test
cases
71
70
69
68
67
66
65
64
63
62
category A
category B
category C
Test Case Test Case Test Case Test Case
(set 1) (set 2) (set 3) (set 4)
Fig 9: Impact of unrealistic supposition
This graph depicts unrealistic supposition impact on different
test cases. Here are four sets of test cases, in which maximum
impact of unrealistic supposition is 70%. Each set contains
three test cases; in the 1st set impact of the unrealistic
supposition on test cases are respectively 70, 67 and 69. In the
2nd set, impact of the unrealistic supposition is respectively
69, 70 and 70. In the 3rd set, impact is respectively 70, 66 and
69. While in the 4th set, impact is respectively 70, 68 and 65.
More than 70% of analyzers neglected to make a suspicion
that coordinated their normal result. Since all analyzers
neglected to distinguish what's in store taking into account a
comprehension of the framework, they neglected to make
reasonable suspicions in connection to that.
I. Experiment Assessment
The central inspiration driving test execution is to get an
estimation of the item quality, by combining a significant
course of action of analysis appraisal results. To be a useful
trial, it is possible to survey the consequence of the
examination to the portrayed criteria. For structures without
these criteria the possibility of "Suspicion" is valuable.
Proper assessment of test cases
J. Tidy up after an experiment
So also basic to make an accommodating investigation is that
it is possible to repeat the examination, again and again. Tidy
up after an experiment execution incorporates each one of
those activities that are expected to evacuate the impacts of
execution to have the capacity to execute once more. In
industry, there is ordinarily no extra level of documentation
for the test procedure.
This graph demonstrates tidy up’s impact on different test
cases. Here are four sets of test cases, in which maximum
impact of tidy up is 78%. Each set contains three test cases; in
the 1st set impact of the tidy up on test cases are respectively
78, 77 and 78. In the 2nd set, impact of the tidy up is
respectively 71, 70 and 74. In the 3rd set, impact is
respectively 73, 72 and 75. While in the 4th set impact is
respectively 78, 72 and 78. 22% of the test cases attempted to
clean up. So, its impact is 78%.
Impact of "tidy up after execution" on
test cases
80
78
76
74
72
70
68
66
category A
category B
Test
Test
Test
Test
Case Case Case Case
(set 1) (set 2) (set 3) (set 4)
category C
Fig 11: Impact of tidy up after execution
This class is effortlessly overlooked, however an undeniable
classification while doing mechanization. Tidy up contains
numerous activities, and it's especially hard to tidy up in a few
frameworks that dependably store information, and don't
permit expulsion.
52
4. RESULTS
50
The Proposed procedure is a great degree gainful that is used
to evaluate the way of the item. To test the adequacy of the
procedure, diverse exploratory setups are created and the
result is dismembered. Inherited Algorithms are definitely not
hard to apply to a broad assortment of improvement issues,
like the voyaging deals delegate issue, inductive thought
learning, booking, and arrangement issues. Programming,
testing is like a manner to change issue with the objective that
the attempts consumed minimized and the amount of
inadequacies perceived extended. Programming, testing is
seen as most effort exhausting development in the item
change. Regardless of the way that different testing strategies
and plenty fullness criteria have been proposed in the written
work, in any case it has been watched that no
framework/criteria is adequately satisfactory to ensure the
movement of weakness free programming huge to the need of
change test time to minimize the cost of testing.
48
category A
46
44
category B
42
category C
40
Test Case Test Case Test Case Test Case
(set 1)
(set 2)
(set 3)
(set 4)
Fig 10: Proper assessment of test cases
This graph illustrates the assessment’s impact on different test
cases. Here are four sets of test cases, in which maximum
impact of assessment is 50%. Each set contains three test
cases; in the 1st set impact of the assessment of test cases are
respectively 50, 49 and 50. In the 2nd set, impact of the
assessment is respectively 49, 45 and 46. In the 3 rd set, impact
is respectively 50, 50 and 48. While in the 4th set impact is
respectively 44, 47 and 50. Upwards of 40% missed giving
assessment by any means, and around 10% of the experiments
not detail an exact assessment that decide the result. In this
way, its impact is half, 50%.
12
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
Optimized test cases can be generated by considering these
factors. These aspects are very useful and significant to
understand and enhance the performance of test cases.
90
80
70
60
50
40
30
20
10
0
The composed methodology is satisfactory to the association's
configuration measures and gives a smooth stream of data
starting with one stage then onto the next. The primary point
is to evaluate the achievability of utilizing GA to produce
upgraded test information for programming testing. In the
wake of applying the some investigation, the better
arrangement is created, that is attainable and solid.
5 Factors having minimum impact on test
case
Fig 12: Factors impact level on test case
The methodology is additionally time powerful in light of the
fact that the estimations are mechanized. The outcome
contains least mistakes and exceptionally precise as the
information is required. Subsequently, it gives a normal state
of testing. The above figure shows the impact level of
different factors, having influence on test cases. Impact of
good detail level is 44%, understanding system’s framework
impact is 50%, miss using techniques impact is 80%,
suppositions impact is 70%, and variation’s impact is 75%,
beginning position’s impact 60%, input analysis impact is
38%, regulated depiction’s impact is 55%, assessment’s
impact 50% and tidy up impact is 78%. Here are 4 factors that
have high impact level that are described in the following
table:
Table 2. Factors having high impact
4 Factors
Mis-usage of techniques
Tidy up after execution
Lack of Variation
Unrealistic supposition
Impact level
80%
78%
75%
70%
Thus, better enhancement method is presented in the
experiment by utilizing Genetic Algorithm. Streamlining
strategy obliged distinctive scale undertakings to review. In
this study, programming, testing related issues is investigated
by utilizing the Genetic Algorithm (GA) approach.
Analysis for optimized test case
Unrealistic supposition
Lack of Variation
Tidy up after execution
Mis-usage of techniques
65
70
75
80
85
4 Factors having high impact on test case generation
Fig. 13: Analysis for optimized test case
This graph is describing the analysis of four factors that are
most important for test case optimization. These factors play a
significant role and have very high impact on the test case
performance. Misusing of test design techniques has higher
impact, which is 80%. Other three factors comparatively
lower than misusing of test design technique factor.
lack of
detail
24%
understandin
g of system's
framework
27%
Other
56%
invalid
input
20%
lack of
regulated
depiction
29%
Fig 14: Minimum impact level on test case
This graph is an illustration those five factors, that have
minimum impact for optimized test case generation. The
invalid input factor comparatively lower than other four
factors. Lack of regulated depiction factor has 55% impact on
test cases, which is higher by comparing other four factors.
So it is hard to test the item altogether, so this procedure is
extraordinarily useful in selecting the best game plan of
analyses (test suite). The decision relies on upon components
that evaluate the analysis, whether it is incredibly or awful.
5. CONCLUSION
Testing gives a first means relating to guaranteeing
programming execution. The aggregate point of programming
industry is really to make certain begin connected with
amazing programming for the end client. Be that as it may,
one compartment, connected with programming, testing has
many fundamental concerns, which are imperative and need
to focus on these issues. These issues are compelling era,
prioritization of experiments and so forth. These issues are
overcome by focusing and core interest. Problems in the
product testing zone are typically how to gain an
extraordinary appropriate set connected with cases to affirm
programming. Some different systems furthermore techniques
keep on being proposed relating to transportation
consideration of the majority of these issues. The way of the
item layout is measured through different procedures and
methodologies. If any mix-up happened in any part of the
undertaking infers it is essential to change the impacted part
of a framework to evacuate the bug. It is conceivable just by
recognizing blunders and by measuring the nature of
programming. Thusly the way of the item is measured and the
result is motorized by the proposed theory of this
examination. This rationality gives a capable and proper
instrument to evaluate the way of the item. Diverse
estimations are proposed closure by the collection strategy for
different sorts of vernaculars and the distinctive goof
frameworks. Thusly, this investigative work is done
successfully with a beneficial methodology proposed in this
work.
13
International Journal of Computer Applications (0975 – 8887)
Volume 151 – No.7, October 2016
6. REFERENCES
[1] Ali, B. M. Y. and Benmaiza, F., 2012. Generating Test
Case for Object-Oriented Software Using Genetic
Algorithm and Mutation Testing Method. International
Journal of Applied Metaheuristic Computing, 3: 15–23.
[2] Athar, M. and Ahmad, I., 2014. Maximize the Code
Coverage for Test Suit by Genetic Algorithm.
International Journal of Computer Science and
Information Technologies, 5(1): 431–435.
[3] Garg, R., and Mittal, S., 2014. Optimization by Genetic
Algorithm. International Journal of Advanced Research
in Computer Science and Software Engineering, 4(4):
587–589.
[4] Kaur, R. and Dhanda, S. K., 2013. Generation of Test
Data Using Genetic Algorithm. International Journal of
Engineering Research and Applications, 3(5): 573–574.
[5] Khan, R., M. Amjad and Pandey, D., 2014. Automated
Test Case Generation using Nature Inspired Meta
Heuristics- Genetic Algorithm. International Journal of
Application or Innovation in Engineering &
Management, 3(11): 7–9.
[6] Malhotra, R. and Bharadwaj, A., 2012. Test case
priortization using genetic algorithm. International
Journal of Computer Science and Informatics, 3: 63–66.
[7] Mondal, K. and Tahbildar, S. H., 2013. Automated Test
Data Generation Using Fuzzy Logic-Genetic Algorithm
Hybridization System for Class Testing Of Object
Oriented Programming. International Journal of Soft
Computing and Engineering, 3(5): 40–49.
[8] Sabharwal, S., Sibal R., and Sharma, C., 2011. Applying
Genetic Algorithm for Prioritization of Test Case
IJCATM : www.ijcaonline.org
Scenarios Derived from UML Diagrams. International
Journal of Computer Science, 8(3): 433-444.
[9] Saini, P. and Tyagi, S., 2014. Test Data Generation for
Basis Path Testing Using Genetic Algorithm and Clonal
Selection Algorithm. International Journal of Science
and Research, 3(6): 2012–2015.
[10] Sharma, K. A., 2013. Optimized Test Case Generation
Using Genetic Algorithm. International Journal of
Computing and Business Research, 4(3): 4–7.
[11] Singh, A., Garg, N., and Saini, T., 2014. A hybrid
Approach of Genetic Algorithm and Particle Swarm
Technique to Software Test Case Generation.
International Journal of Innovations in Engineering and
Technology, 3(4): 208–214.
[12] Singhal, A., Chandna, S., and Bansal, A., 2012.
Optimization of Test Cases Using Genetic Algorithm.
International Journal of Emerging Technology and
Advanced Engineering, 2(3): 367–369.
[13] Sumalatha, V. M., and Raju, G. S. V. P., 2013. Object
Oriented Test Case Generation Technique using Genetic
Algorithms. International Journal of Computer
Applications, 61(20): 20–26.
[14] Tripathy, P. and Kanhar, S. D., 2013. Optimization of
Software Testing for Discrete Tes tsuite using Genetic
Algorithm and Sampling Technique. International
Journal of Computer Applications, 63(7): 1–5.
[15] Varshney, S. and Mehrotra, M., 2014. Automated
Software Test Data Generation for Data Flow
Dependencies using Genetic Algorithm. International
Journal of Advanced Research in Computer Science and
Software Engineering, 4(2): 472–479.
14
| 9cs.NE
|
The Word Problem for Omega-Terms
over the Trotter-Weil Hierarchy∗
arXiv:1509.05364v2 [cs.FL] 16 May 2017
Manfred Kufleitner† and Jan Philipp Wächter
Institut für Formale Methoden der Informatik
University of Stuttgart, Germany
{kufleitner,jan-philipp.waechter}@fmi.uni-stuttgart.de
May 17, 2017
Abstract. For two given ω-terms α and β, the word problem for ω-terms
over a variety V asks whether α = β in all monoids in V. We show that the
word problem for ω-terms over each level of the Trotter-Weil Hierarchy is decidable. More precisely, for every fixed variety in the Trotter-Weil Hierarchy,
our approach yields an algorithm in nondeterministic logarithmic space (NL).
In addition, we provide deterministic polynomial time algorithms which are
more efficient than straightforward translations of the NL-algorithms. As
an application of our results, we show that separability by the so-called corners of the Trotter-Weil Hierarchy is witnessed by ω-terms (this property is
also known as ω-reducibility). In particular, the separation problem for the
corners of the Trotter-Weil Hierarchy is decidable.
1 Introduction
Algebraic characterizations of classes of regular languages are interesting as they often allow to decide the class’s membership problem. For example, by Schützenberger’s famous
theorem [22], one can decide whether a given regular language is star-free by computing
its syntactic monoid M and checking its aperiodicity. The latter can be achieved by
verifying x|M |! = x|M |! x for all x ∈ M . This equation is also stated as xω = xω x since
this notation is independent of the monoid’s size. More formally, we can see the equation
as a pair of ω-terms: these are finite words built using letters, which are interpreted as
variables, concatenation and an additional formal ω-power. Checking an equation α = β
∗
†
The final publication is available at Springer via http://dx.doi.org/10.1007/s00224-017-9763-z.
The first author was supported by the German Research Foundation (DFG) under grants DI 435/5-2
and KU 2716/1-1.
1
in a finite monoid is easy: one can simply substitute each variable by all elements of the
monoid. For each substitution, this yields a monoid element on the left hand side and
one on the right hand side. The equation holds if and only if they are always equal.
Often, the question whether an equation holds is not only interesting for a single finite
monoid but for a (possibly infinite) class of such monoids. For example, one may ask
whether all monoids in a certain class are aperiodic. This is trivially decidable if the
class is finite. But what if it is infinite? If the class forms a variety (of finite monoids,
sometimes also referred to as a pseudo-variety), i. e. a class of finite monoids closed
under (possibly empty) direct products, submonoids and homomorphic images, then this
problem is called the variety’s word problem for ω-terms. Usually, the study of a variety’s
word problem for ω-terms also gives more insight into the variety’s structure, which is
interesting in its own right. McCammond showed that the word problem for ω-terms of
the variety A of aperiodic finite monoids is decidable [17]. The problem was shown to be
decidable in linear time for J, the class of J -trivial finite monoids, by Almeida [3] and
for R, the class of R-trivial monoids, by Almeida and Zeitoun [4]. For the variety DA,
Moura adapted and expanded those ideas to show decidability in time O((nk)5 ) where
n is the length of the input ω-terms and k is the maximal nesting depth of the ω-power
(which can be linear in n) [18]. Remember that DA is the class of finite monoids whose
regular D-classes form aperiodic semigroups. This variety received a lot of attention due
to its many different characterizations; see e.g. [5, 26]. Most notably is its connection to
two-variable first-order logic [27]. This logic is a natural restriction of first-order logic
over finite words, which in turn is the logic characterization of A.
In this paper, we consider the word problem for ω-terms over the varieties in the
Trotter-Weil Hierarchy. It was introduced by Trotter and Weil [28] with the idea of
using the good understanding of the band varieties (cf. [7]) for studying the lattice of
sub-varieties of DA; bands are semigroups satisfying x2 = x. The levels of this hierarchy
exhaust DA. As it turns out, the Trotter-Weil Hierarchy has tight connections to the
quantifier alternation hierarchy inside two-variable first-order logic [14]. In addition,
many characterizations of DA admit natural restrictions which allow climbing up this
hiearchy (see [12]).
Please note that, in spite of this paper’s title, we will refer to ω-terms as π-terms
for most parts of the paper, the only exception being this introduction. We follow this
notation introduced by Perrin and Pin [19] to avoid notational conflicts. Accordingly, we
use π for the formal power in π-terms and speak of the word problem for π-terms.
Results. In this paper, we present the following results.
• Our main tool for studying a variety V of the Trotter-Weil Hierarchy is a family
of finite index congruences ≡V,n for n ∈ N. These congruences have the property
that a monoid M is in V if and only if there exists n for which M divides a
quotient by ≡V,n . The congruences are not new but they differ in some details
from the ones usually found in the literature, where they are introduced in terms
of rankers [12, 14, 15]. Unfortunately, these differences necessitate new proofs.
• We lift the combinatorics from finite words to ω-terms using the “linear order
approach” introduced by Huschenbett and the first author [10]. They showed that,
2
over varieties of aperiodic monoids, one can use the order N + Z · Q + (−N) for the
formal ω-power. In this paper, we use the simpler order N + (−N). We show that
two ω-terms α and β are equal in some variety V of the Trotter-Weil Hierarchy
if and only if JαKN+(−N) ≡V,n JβKN+(−N) for all n ∈ N. Here, JαKN+(−N) denotes
the generalized word (i. e. the labeled linear order) obtained from replacing every
occurrence of the formal ω-power by the linear order N + (−N). Note that this
order is tailor-made for the Trotter-Weil Hierarchy and does not result from simple
arguments which work in any variety.
• We show that one can effectively check whether JαKN+(−N) ≡V,n JβKN+(−N) for all
n ∈ N.
• We further improve the algorithms and show that, for every variety V of the
Trotter-Weil Hierarchy, the word problem for ω-terms over V is decidable in nondeterministic logarithmic space. The main difficulty is to avoid some blow-up which
(naively) is caused by the nesting depth of the ω-power. For R, which appears
in the hierarchy, this result is incomparable to Almeida and Zeitoun’s linear time
algorithm [4].
• We also introduce polynomial time algorithms, which are more efficient than the
direct translation of these NL algorithms.
• As an application, we prove that the separation problems for the so-called corners
of the Trotter-Weil Hierarchy are decidable by showing ω-reducibility. For J, we
adapt the proof of van Rooijen and Zeitoun [29].
• With little additional effort, we also obtain all of the above results for DA, the
limit of the Trotter-Weil Hierarchy. The decidability of the separation problem
re-proves a result of Place, van Rooijen and Zeitoun [21]. The algorithms for the
word problem for ω-terms are more efficient than those of Moura [18].
Separability of the join levels and the intersection levels is still open. We conjecture
that these problems can be solved with similar but more technical reductions.
2 Preliminaries
Natural Numbers and Finite Words. Let N = {1, 2, . . . }, N0 = {0, 1, . . . } and −N =
{−1, −2, . . . }. For the rest of this paper, we fix a finite alphabet Σ. By Σ∗ , we denote
the set of all finite words over the alphabet Σ, including the empty word ε; Σ+ denotes
Σ∗ \ {ε}.
Order Types. A linearly ordered set (P, ≤P ) consists of a (possibly infinite) set P and
a linear ordering relation ≤P of P , i. e. a reflexive, anti-symmetric, transitive and total
binary relation ≤P ⊆ P × P . To simplify notation we define two special objects −∞ and
+∞. The former is always smaller with regard to ≤P than any element in P while the
3
latter is always larger. We call two linearly ordered sets (P, ≤P ) and (Q, ≤Q ) isomorphic
if there is an order-preserving bijection ϕ : P → Q. Isomorphism between linearly order
sets is an equivalence relation; its classes are called (linear) order types.
The sum of two linearly ordered sets (P, ≤P ) and (Q, ≤Q ) is (P ⊎ Q, ≤P +Q ) where
P ⊎ Q is the disjoint union of P and Q and ≤P +Q orders all elements of P to be smaller
than those of Q while it behaves as ≤P and ≤Q on elements from their respective sets.
Similarly, the product of (P, ≤P ) and (Q, ≤Q ) is (P × Q, ≤P ∗Q ) where (p, q) ≤P ∗Q (p̃, q̃)
holds if and only if either q ≤Q q̃ and q 6= q̃ or q = q̃ and p ≤P p̃ holds. Sum and
product of linearly ordered sets are compatible with taking the order type. This allows
for writing µ + ν and µ ∗ ν for order types µ and ν.
We re-use n ∈ N0 to denote the order type of ({1, 2, . . . , n}, ≤). One should note
that this use of natural numbers to denote order types does not result in contradictions
with sums and products: the usual calculation rules apply. Besides finite linear order
types, we need ω, the order type of (N, ≤), and its dual ω ∗ the order type of (−N, ≤).
Another important order type in the scope of this paper is ω + ω ∗ , whose underlying set
is N ⊎ (−N). Note that, here, natural numbers and the (strictly) negative numbers are
ordered as 1, 2, 3, . . . , . . . , −3, −2, −1; therefore, in this order type, we have for example
−1 ≥ω+ω∗ 1.
Generalized Words. Any finite word w = a1 a2 . . . an of length n ∈ N0 with ai ∈ Σ can
be seen as a function which maps a position i ∈ {1, 2, . . . , n} to the corresponding letter
ai (or, possibly, the empty map). Therefore, it is natural to denote the positions in a
word w by dom(w). By relaxing the requirement of dom(w) to be finite, one obtains the
notion of generalized words: a (generalized) word w over the alphabet Σ of order type µ
is a function w : dom(w) → Σ, where dom(w) is a linearly ordered set in µ. For dom(w),
we usually choose (N, ≤), (−N, ≤) and (N ⊎ (−N), ≤ω+ω∗ ) as representative of ω, ω ∗ and
ω + ω ∗ , respectively. The order type of a finite word of length n is n.
Like finite words, generalized words can be concatenated, i. e. we write u to the left
of v and obtain uv. In that case, the order type of uv is the sum of the order types of
u and v. Beside concatenation, we can also take powers of generalized words. Let w be
a generalized word of order type µ which belongs to (Pµ , ≤µ ) and let ν be an arbitrary
order type belonging to (Pν , ≤ν ). Then, wν is a generalized word of order type µ ∗ ν
which determines the ordering of its letters; w maps (p1 , p2 ) ∈ Pµ × Pν to w(p1 ). If ν = n
for some n ∈ N, then wν = wn is equal to the n-fold concatenation of w.
In this paper, the term word refers to a generalized word. If it is important for a word
to be finite, it is referred to explicitly as a finite word. As a counterpart to the positions
dom(w) in w, we define the set of letters appearing in a word alph(w) as the image of w
seen as a function. For example, for a finite word w = a1 a2 . . . an of length n ∈ N0 (with
a1 , a2 , . . . , an ∈ Σ), we have alph(w) = {a1 , a2 , . . . , an }.
We also introduce notation for factors of words. For a pair (l, r) ∈ ({−∞}⊎ dom(w))×
(dom(w) ⊎ {+∞}), define w(l,r) as the restriction of the word w (seen as a mapping) to
the set of positions (strictly) larger than l and (strictly) smaller than r. Note that
w = w(−∞,+∞) and w(l,r) = ε for any pair (l, r) with no position between l and r.
4
Monoids, Divisors, Congruences and Recognition. In this paper, the term monoid
refers to a finite monoid (except when stated otherwise). It is well known that, for any
monoid M , there is a smallest number n ∈ N such that mn is idempotent (i. e. m2n = mn )
for every element m ∈ M ; this number is called the exponent of M and shall be denoted
by M ! = n.1 A monoid N is a divisor of (another) monoid M , written as N ≺ M , if N
is a homomorphic image of a submonoid of M .
A congruence (relation) in a (not necessarily finite) monoid M is an equivalence relation
C ⊆ M × M such that x1 C x2 and y1 C y2 implies x1 y1 C x2 y2 for all x1 , x2 , y1 , y2 ∈ M .
If M is a (possibly infinite) monoid and C ⊆ M × M is a congruence, then the set of
equivalence classes of C, denoted by M/C, is a well-defined monoid (which might still be
infinite), whose size is called the index of C. For any two congruences C1 and C2 , one can
define their join C1 ∨ C2 as the smallest congruence which includes C1 and C2 ; its index
is at most as large as the index of C1 and the index of C2 .
A (possibly infinite) monoid M recognizes a language (of finite words) L ⊆ Σ∗ if there
is a homomorphism ϕ : Σ∗ → M with L = ϕ−1 (ϕ(L)). A language is regular if and only
if it is recognized by a finite monoid. It is well known that there is a unique smallest
monoid which recognizes a given regular language: the syntactic monoid.
Green’s Relations. Among the most important tools for studying monoids are Green’s
Relations. Let x and y be elements of a monoid M . Define
x R y ⇔ xM = yM,
x L y ⇔ Mx = My
and
x J y ⇔ M xM = M yM
where xM = {xm | m ∈ M } is the right-ideal of x, M x = {mx | m ∈ M } its left-ideal
and M xM = {m1 xm2 | m1 , m2 ∈ M } its (two-sided) ideal.
By simple calculation, one can see that x R y holds if and only if there are z, z ′ ∈ M
such that xz = y and yz ′ = x and, symmetrically, that x L y holds if and only if there
are z, z ′ ∈ M such that zx = y and z ′ y = x.
Varieties, π-Terms, Equations and Word Problem for π-terms. A variety (of finite
monoids) – sometimes also referred to as a pseudo-variety – is a class of monoids which
is closed under submonoids, homomorphic images and – possibly empty – finite direct
products. For example, the class R of R-trivial monoids and the class L of L-trivial
monoids both form a variety, see e.g. [20]. Clearly, if V and W are varieties, then so is
V ∩ W. For example, the class J = R ∩ L is a variety; in fact, it is the variety of all
J -trivial monoids. For two varieties V and W, the smallest variety containing V ∪ W,
the so called join, is denoted by V ∨ W.
Many varieties can be defined in terms of equations (or identities). Because it will
be useful later, we take a more formal approach towards equations by using π-terms 2 .
1
2
Note that all statements remain valid if one assumes that M ! is used to denote |M |!.
As mentioned in the introduction, π-terms are usually referred to as ω-terms. In this paper, however,
5
A π-term is a finite word, built using letters, concatenation and an additional formal
π-power (and appropriate parentheses), whose π-exponents act as a placeholder for a
substitution value. Formally, every letter a ∈ Σ is a π-term (over Σ). As a special case,
also ε is a π-term. For two π-terms α and β, their concatenation αβ is a π-term as well
and, if γ is a π-term, then so is (γ)π , where π is a formal exponent.
To state equations using π-terms, one needs to substitute the formal π-exponents by
actual values resulting in a word. We define JγKµ as the result of substituting the πexponents in γ by an order type µ, i. e. we have JεKµ = ε, JaKµ = a for all a ∈ Σ,
JαβKµ = JαKµ JβKµ and J(γ)π Kµ = (JγKµ )µ . For example3 , we have
Ja ((b)π c)π K3 = a ((bbb)c) ((bbb)c) ((bbb)c) .
An equation α = β consists of two π-terms α and β over the same alphabet Σ, which,
here, can be seen as a set of variables. A homomorphism σ : Σ∗ → M is called an
assignment of variables in this context. An equation α = β holds in a monoid M if for
every assignment of variables σ (JαKM ! ) = σ (JβKM ! ) is satisfied. It holds in a variety V
if it holds in all monoids in V. The word problem for π-terms over a variety V is the
problem to decide whether α = β holds in V for the input π-terms α and β.
Mal’cev Products. Besides intersection and join, we need one more constructions for
varieties: the Mal’cev product, which is often defined using relational morphisms. In this
paper, we use a different, yet equivalent, approach based on the congruences ∼K and ∼D ,
see [11] or [8, Corollary 4.3]. For their definition, let x and y be elements of a monoid M
and define
x ∼K y ⇔ ∀e ∈ E(S) : ex R e or ey R e ⇒ ex = ey
and x ∼D y ⇔ ∀e ∈ E(S) : xe L e or ye L e ⇒ xe = ye,
where E(S) denotes the set of idempotents in S.
Obviously, ∼K and ∼D are of finite index in any (finite) monoid M . Thus, we have
that M/∼K and M/∼D are (finite) monoids and can define Mal’cev products of varieties.
Let V be a variety. The varieties K m V and D m V are defined by
M ∈ K m V ⇔ M/∼K ∈ V and
M ∈ D m V ⇔ M/∼D ∈ V,
where M is a monoid. Note that, indeed, K m V and D m V are varieties for any variety
V and that, furthermore, we have V ⊆ K m V and V ⊆ D m V.
3 The Trotter-Weil Hierarchy
The main object of study in this paper, is the so-called Trotter-Weil Hierarchy. We will
approach it primarily using certain combinatorial congruence. We will show that this
we use ω to denote the order type of the natural numbers. Therefore, we follow the approach of
Perrin and Pin [19] and use π instead of ω.
3
For a more elaborate example (involving ω + ω ∗ ) see Example 1 on page 35.
6
R3
L3
R3 ∩ L3
R2 ∨ L2
R = R2
L2 = L
m∈N
R1 = L1 = J,
..
.
[
The Trotter-Weil Hierarchy. As the name implies, this hierarchy was first studied by Trotter and
Weil [28]. We will define it using Mal’cev products.
Though this approach is different to the original
one used by Trotter and Weil, both are equivalent
[15]. We define:
Rm ∨ Lm = DA
approach is equivalent to the more common algebraic approach, which we will use as the
definition of the hierarchy.
R2 ∩ L2 = J = R1 = L1
Rm+1 = K m Lm and
Figure 1: Trotter-Weil Hierarchy
Lm+1 = D m Rm .
These varieties form the so-called corners of the hierarchy. Additionally, it contains the
join levels Rm ∨ Lm and the intersection levels Rm ∩ Lm . The term “hierarchy” is
justified by the following inclusions: we have Rm ∩ Lm ⊆ Rm , Lm ⊆ Rm ∨ Lm and
Rm ∨ Lm ⊆ Rm+1 ∩ Lm+1 ; the latter can be seen by induction.
Among the corners of the Trotter-Weil Hierarchy are some well known varieties: we
have R1 = L1 = J, R2 = R and L2 = L (for the last two, see [20]; the others are
straightforward).
By taking the union of all varieties in the hierarchy, one gets the variety DA [13],
which is usually defined as the class of monoids whose regular D-classes form aperiodic
semigroups4 :
[
[
[
Fact 1.
DA =
Rm ∨ Lm =
Rm =
Lm
m∈N
m∈N
m∈N
These considerations yield the graphic representation given in Figure 1.
Connections to Two-Variable Logic. The variety DA is closely connected to twovariable first-order logic. By FO2 [<], denote the set of all first-order sentences over finite
words which may only use the < predicate (and equality) and no more than two variables.
A language L ⊆ Σ∗ of finite words is definable by a sentence ϕ ∈ FO2 [<] if and only if its
syntactic monoid is in DA [27], which it is if and only if it is in one of the Trotter-Weil
Hierarchy’s varieties.
The intersection levels corresponds to the quantifier alternation hierarchy within FO2 [<
] [14]: a first-order sentence using at most two variables belongs to FO2m [<] if, on any
path in its syntax tree, there is no quantifier after the first negation and there are at
most m blocks of quantifiers. A language is definable by a sentence in FO2m [<] if and
only if its syntactic monoid is in Rm+1 ∩ Lm+1 .
4
In finite monoids, D-classes coincide with J -classes; a D-class is called regular if it contains an idempotent. A semigroup is called aperiodic (or group-free) if it has no divisor which is a nontrivial
group.
7
Equational Characterization. Besides its definition using Mal’cev products and its connections to logic, the Trotter-Weil Hierarchy can also be characterized in terms of equations. For our proofs, we only need the direction of this characterization stated in the
lemma below. The other direction does hold as well; we will see later on in Lemma 5
that it is an easy consequence of the hierarchy’s combinatorial characterization which we
state below.
Lemma 1. Define the π-terms
U1 = (sx1 )π s(y1 t)π
and
V1 = (sx1 )π t(y1 t)π
over the alphabet Σ1 = {s, t, x1 }. For m ∈ N, let xm+1 and ym+1 be new characters not
in the alphabet Σm and define the π-terms
Um+1 = (Um xm+1 )π Um (ym+1 Um )π
and
Vm+1 = (Um xm+1 )π Vm (ym+1 Um )π
over the alphabet Σm+1 = Σm ⊎ {xm+1 , ym+1 }.
Then we have
M ∈ R1 = L1 = J ⇐ U1 = V1 holds in M ,
M ∈ Rm+1 ⇐ (Um xm+1 )π Um = (Um xm+1 )π Vm holds in M ,
M ∈ Lm+1 ⇐ Um (ym+1 Um )π = Vm (ym+1 Um )π holds in M and
M ∈ Rm+1 ∩ Lm+1 ⇐ Um = Vm holds in M
for all m ∈ N.
Proof. The first implication is a well-known characterization of J [20]. We show the next
two implications by induction over m (see also [12]).
First, assume m = 1 and consider a monoid M in which (U1 x2 )π U1 = (U1 x2 )π V1 holds.
We need to show M ∈ R2 , which is equivalent to showing M/∼K ∈ L1 = J. For this,
we show u1 = σ (JU1 KM ! ) ∼K σ (JV1 KM ! ) = v1 for an arbitrary assignment of variables
σ : Σ∗1 → M . Afterwards, we are done by the first implication since the exponent of
M/∼K is a divisor of M !. Let e be an arbitrary idempotent of M and assume eu1 R e
(the other case from the definition of ∼K is symmetrical). Thus, there is an element
x2 ∈ M such that eu1 x2 = e holds. We can extend σ by mapping the letter x2 ∈ Σ2 to
the just defined monoid element x2 . We then have
eu1 = e(u1 x2 )u1 = . . . = e(u1 x2 )M ! u1
= e(u1 x2 )M ! v1 = · · · = e(u1 x2 )v1 = ev1 .
The equality in the middle holds because (U1 x2 )π U1 = (U1 x2 )π V1 holds in M by assumption. This concludes the m = 1 case because the third implication is symmetrical.
Now, assume m > 1 and consider a monoid M in which (Um xm+1 )π Um = (Um xm+1 )π Vm
holds. We need to show M ∈ Rm+1 and we do this by showing M/∼K ∈ Lm . By induction, we only have to show that Um−1 (ym Um−1 )π = Vm−1 (ym Um−1 )π holds in M/∼K . As
8
before, let σ : (Σm−1 ⊎ {ym })∗ → M be an arbitrary assignment of variables. For convenience, let um−1 = σ (JUm−1 KM ! ) and vm−1 = σ (JVm−1 KM ! ) and identify ym with σ(ym ).
Let e be an arbitrary idempotent in M such that eum−1 (ym um−1 )M ! R e holds. Clearly,
this implies eum−1 R e and there are elements xm , xm+1 ∈ M such that e = eum−1 xm
and e = eum−1 (ym um−1 )M ! xm+1 holds. Extend σ to map the letters xm and xm+1 to
the respective monoid elements. Then, we have
eum−1 (ym um−1 )M ! = e(um−1 xm )um−1 (ym um−1 )M !
= · · · = e(um−1 xm )M ! um−1 (ym um−1 )M ! = eum
with um = σ (JUm KM ! ). This yields
e = eum−1 (ym um−1 )M ! xm+1 = e (um xm+1 ) = · · · = e (um xm+1 )M ! .
{z
}
|
eum
Using the fact that (Um xm+1 )π Um = (Um xm+1 )π Vm holds in M (by assumption), we get
eum = e (um xm+1 )M ! um = e (um xm+1 )M ! vm = evm
where vm = σ (JVm KM ! ). In combination, we have
eum−1 (ym um−1 )M ! = eum = evm = e(um−1 xm )M ! vm−1 (ym um−1 )M !
where the last equality holds due to the definition of Vm . Finally, we get
eum−1 (ym um−1 )M ! = e(um−1 xm )M ! vm−1 (ym um−1 )M !
= · · · = e(um−1 xm )vm−1 (ym um−1 )M !
= evm−1 (ym um−1 )M !
Thus, we have shown um−1 (ym um−1 )M ! ∼K vm−1 (ym um−1 )M ! (the other case from the
definition of ∼K is symmetrical) and are done. The implication for Lm+1 in the case
m > 1 follows by symmetry again.
Finally, for the intersection levels, suppose that Um = Vm holds in a monoid M . By
the identities for the corners, we directly have M ∈ Rm+1 ∩ Lm+1 .
Besides the equational characterization of the individual varieties in the hierarchy, one
can also characterize their union DA in terms of an equation:
Fact 2. Let M be a monoid. Then, we have
M ∈ DA ⇔ (xyz)π y(xyz)π = (xyz)π holds in M .
A proof of this fact can be found in [26].
9
4 Relations for the Trotter-Weil Hierarchy
Although we define the Trotter-Weil Hierarchy algebraically using Mal’cev products, we
will primarily use a different characterization which is based on certain combinatorial
congruences. Before we can finally introduce these congruences, however, we need to
give some definitions for factorizations of words at the first or last a-position (i. e. an
a-labeled position).
Factorizations and accessible words. For a word w, a position p ∈ dom(w) ⊎ {−∞}
and a letter a ∈ alph(w), let Xa (w; p) denote the first a-position (strictly) larger than
p (or the first a-position in w if p = −∞). It is undefined if there is no such position
(i. e. there is no a to the right of p in w) or if the position is not well-defined. Define
Ya (w; p) symmetrically as the first a-position from the right which is (strictly) smaller
than p. Notice that, with generalized words, the first a-position to the right of a position
p is not necessarily well-defined even if there is an a-position larger than p. For example,
∗
aω = . . . aa does not have a first a-position. We call words for which this situation does
not occur accessible; i. e. a word w is accessible if, for every position p in w (including
the special cases p ∈ {±∞}), Xa (w; p) is defined if and only if there is an a-position in
w(p,+∞) and Ya (w; p) is defined if and only if there is an a-position in w(−∞,p). Note that
all finite words are accessible and that so are all words of the form JγKω+ω∗ for a π-term
γ.
Let w be an accessible word, define
w · XaL = w(−∞,Xa (w;−∞)) ,
w · XaR = w(Xa (w;−∞),+∞),
w · YaL = w(−∞,Ya (w;+∞)) and
w · YaR = w(Ya (w;+∞),+∞)
for all a ∈ alph(w). Additionally, define Ca,b as a special form of applying XaL first
and then YbR which is only defined if Xa (w; −∞) is strictly larger than Yb (w; +∞).
For an example of XaL and XaR acting on a word see Figure 2. Note that we have
w = (w · XaL )a(w · XaR ) = (w · YaL )a(w · YaR ) = (w · YbL )b(w · Ca,b )a(w · XaR ) (whenever
these factors are defined). For example, we have cbbcdcaca · Ca,b = cdc; see the upper
part of Figure 4 for a graphical representation of this example. If we apply a sequence of
factorizations, we omit the · between them, e. g. we write w·XaL YbR = w·XaL ·YbR = w·Ca,b .
l
w=
Xa (w; l)
r
abbabbbbbabbbababbab
w(l,r) · XaL
w(l,r) · XaR
Figure 2: Application of XaL and XaR to an example word.
Relations for the Trotter-Weil Hierarchy. With these definitions in place, we define
Y
WI
5
for m, n ∈ N the relations ≡X
m,n , ≡m,n and ≡m,n of accessible words. The idea is that
5
The presented relations could also be defined by (condensed) rankers (as it is done in [14] and [15]).
Rankers were introduced by Weis and Immerman [30] (thus, the WI exponent in ≡WI
m,n ) who reused
10
these relations hold on two words u and v if both words allow for the same sequence of
factorizations at the first or last occurrence of a letter. For this, we use the following
recursive definition.
Definition 1. Let m, n ∈ N and let u and v be accessible words. Define recursively:
Z
Z
1. u ≡Z
0,0 v, u ≡m,0 v and u ≡0,n v for Z ∈ {X, Y, WI} always hold.
Y
2. u ≡X
m,n v ⇔ alph(u) = alph(v), u ≡m−1,n−1 v and
∀a ∈ alph (u) : u · XaL ≡Ym−1,n−1 v · XaL and
R
u · XaR ≡X
m,n−1 v · Xa
u ≡Ym,n v ⇔ alph(u) = alph(v), u ≡X
m−1,n−1 v and
∀a ∈ alph (u) : u · YaL ≡Ym,n−1 v · YaL and
R
u · YaR ≡X
m−1,n−1 v · Ya
u ≡WI
m,n v ⇔ alph(u) = alph(v),
L
∀a ∈ alph (u) : u · XaL ≡WI
m−1,n−1 v · Xa and
R
u · XaR ≡WI
m,n−1 v · Xa ,
L
∀a ∈ alph (u) : u · YaL ≡WI
m,n−1 v · Ya and
R
u · YaR ≡WI
m−1,n−1 v · Ya and
∀a, b ∈ alph (u) : u·Ca,b and v·Ca,b are either both undefined
or both defined and u · Ca,b ≡WI
m−1,n−1 v ·
Ca,b holds.
X
Y
Additionally, define u ≡XY
m,n v ⇔ u ≡m,n v and u ≡m,n v for all m, n ∈ N0 .
X
≡m,n−1
Y
≡m−1,n−1
As this definition is vital for the understanding of the rest
X
of this paper, we try to give an intuitive understanding of
Y
X
how the relations work. In the parameter m, we rememu= b b ca a b c a
ber the remaining number of direction changes (which are
R
L
caused by an Xa or Ya factorizations) in a factorization
sequence and the parameter n is the number of remaining
factorization moves (independent of their direction). Thus,
if m or n is zero, then the relations shall be satisfied on all
accessible word pairs. For m and n larger than zero, our first
v= bcbabcaa
assertion is that both words have the same alphabet; othY
X
erwise, one of them would admit a factorization at a letter
X
while the other would not, as the letter is not in its alphaX
bet. For the letters of the common alphabet, we want to be Figure 3: ≡m,n illustrated.
able to perform further factorization (until we reach n = 0). The X or Y exponent of
the turtle programs by Schwentick, Thérien and Vollmer [23]. Another concept related to condensed
rankers is the unambiguous interval temporal logic by Lodaya, Pandya and Shah [16].
11
WI
≡m−1,n−1
the relation indicates whether we start at the beginnings or at the ends of the two words;
the WI exponent is a special case, which we will discuss below. So, if we want u ≡X
m,n v
to hold and the alphabets of u and v coincide, then we can continue factorizing at the
first a in u and in v. In the next step, we continue either in the two left parts or in
the two right parts. If we continue in the right parts (i. e. we do an XaR factorization),
R
then we require u · XaR ≡X
m,n−1 v · Xa . The X in the exponent indicates that the last
factorization position (the first a in u and v, respectively) was at the beginning of the
words. We have made one additional factorization – thus, we only have n − 1 remaining
factorizations in this part – but we did not change the direction because we were at the
beginnings of the words before and still are for the new word pair – thus, we still have
m such changes of direction. If, instead of taking the right parts after the first a, we had
taken the left parts, then we would still have made a single factorization; thus we have
n − 1 remaining factorizations. However, we would also have a direction change: the
factorization position (i. e. the first a) is to the right of the new words u · XaL and v · XaL .
Therefore, we have m − 1 remaining direction changes and we change the X exponent
into a Y exponent. This is summed up by requiring u · XaL ≡Ym−1,n−1 v · XaL . Both
situations are illustrated in Figure 3. Additionally, we also have the choice to switch
from the beginning to the end at the cost of one direction change and one factorization.
This is reflected in the fact that we require u ≡Ym−1,n−1 v for u ≡X
m,n v.
Y
The relation ≡m,n works symmetrically to ≡X
m,n . HowY
X
WI
ever, ≡m,n is a bit different. Here, we still lose one diu= c b b c d ca c a
rection change after an XaL or YaR factorization while the
number keeps the same for XaR and YaL factorizations but
we also have the special Ca,b factorizations, which start simultaneously at the beginning and the end. Consider the
situation given in Figure 4. We go to the first a and the
last b; the former is to the right of the latter in both words
v= bcbdccaac
(or the ≡WI
m,n relation does not hold). For the pair of parts
X
Y
in the middle, we require u · Ca,b ≡WI
m−1,n−1 v · Ca,b , i. e.
we count this as a single factorization move which loses a
Figure 4: ≡WI
single direction change. Finally, there is ≡XY
m,n illustrated.
m,n where we
can start factorizing at the beginning or at the end.
By simple inductions, one can see that the relations are congruences of finite index
Z
Z
over Σ∗ . Also note that u ≡Z
m,n v implies u ≡m,k v and, if m > 0, also u ≡m−1,k v for all
k ≤ n and Z ∈ {X, Y, XY, WI}.
The importance of the just defined relations for this paper yields from their connection
to the Trotter-Weil Hierarchy (and, thus, to DA), which we state in the next theorem.
Theorem 1. Let M be a monoid and m ∈ N. Then
• M ≺ Σ∗ /≡X
m,n for some n ∈ N0 ⇔ M ∈ Rm ,
• M ≺ Σ∗ /≡Ym,n for some n ∈ N0 ⇔ M ∈ Lm ,
• M ≺ Σ∗ /≡XY
m,n for some n ∈ N0 ⇔ M ∈ Rm ∨ Lm and
• M ≺ Σ∗ /≡WI
m,n for some n ∈ N0 ⇔ M ∈ Rm ∩ Lm hold.
12
The combinatorial nature of the relations will turn out to be useful in the remainder of
this paper since it allow us, for example, to obtain efficient algorithms for the problems we
consider. Unfortunately, on the other hand, it makes proving Theorem 1 quite technical.
In fact, we will dedicate the next section to this proof.
5 A Proof for Theorem 1
We will prove both directions of Theorem 1 individually. Before we can do this, however,
we need to introduce some more concepts.
R- and L-factorizations. Let ϕ : Σ∗ → M be a (monoid) homomorphism into a monoid
M . The R-factorization of a finite word w is the (unique) factorization w = w0 a1 w1 a2 w2
. . . ak wk with w0 , w1 . . . , wk ∈ Σ∗ and a1 , a2 , . . . , ak ∈ Σ such that, on the one hand,
ϕ(ε) R ϕ(w0 ) and
ϕ(w0 a1 w1 a2 w2 . . . ai ) R ϕ(w0 a1 w1 a2 w2 . . . ai wi )
hold for i = 1, 2, . . . , k and, on the other hand,
ϕ(w0 a1 w1 a2 w2 . . . ai wi ) 6R ϕ(w0 a1 w1 a2 w2 . . . ai wi ai+1 )
holds for i = 0, 1, . . . , k − 1. Symmetrically, the L-factorization of w is the factorization
w = w0 a1 w1 a2 w2 . . . ak wk with w0 , w1 . . . , wk ∈ Σ∗ and a1 , a2 , . . . , ak ∈ Σ such that, on
the one hand,
ϕ(wk ) L ϕ(ε) and
ϕ(wi−1 ai wi ai+1 wi+1 . . . ak wk ) L ϕ(ai wi ai+1 wi+1 . . . ak wk )
hold for i = 1, 2, . . . , k and, on the other hand,
ϕ(ai wi ai+1 wi+1 ai+2 wi+2 . . . ak wk ) 6L ϕ(wi ai+1 wi+1 ai+2 wi+2 . . . ak wk )
holds for i = 1, 2, . . . , k.
DA and R-classes. In DA, getting into a new R-class is strictly coupled to an element’s
alphabet, as the following lemma shows6 , where a can be seen as one of the monoids
generators (i. e. a letter in its alphabet).
Lemma 2. Let M ∈ DA be a monoid and let s, t ∈ M such that s R t. Then
s R sa ⇒ t R ta
holds for all a ∈ M .
6
The curious reader might be interested in the fact that the lemma’s assertion also holds for monoids
in DS, the variety of monoids whose regular D-classes form (arbitrary, but finite) semigroups. More
on DS can, for example, be found in [2].
13
Proof. Since we have t R s R sa, there are x, y ∈ M with s = tx and t = say. We then
have
t = txay = t(xay)2 = · · · = t(xay)M ! ,
which yields
ta(xay)M ! = t(xay)M ! a(xay)M ! = t(xay)M ! = t.
using the equation from Fact 2. Thus, we have ta R t.
One of the main applications of the previous lemma is the following. If we have
a monoid M ∈ DA, a homomorphism ϕ : Σ∗ → M and the R-factorization w =
w0 a1 w1 a2 w2 . . . ak wk of a finite word w ∈ Σ∗ , then we know that ai 6∈ alph(wi−1 ) for
i = 1, 2, . . . , k. If we had ai ∈ alph(wi−1 ), we could factorize wi−1 = uai v and would
have
ϕ(w0 a1 w1 a2 w2 . . . ai−1 u) R ϕ(w0 a1 w1 a2 w2 . . . ai−1 uai )
and, by the previous lemma, also
ϕ(w0 a1 w1 a2 w2 . . . ai−1 uai v) R ϕ(w0 a1 w1 a2 w2 . . . ai−1 uai vai ),
which results in a contradiction to the definition of R-factorizations. Of course, we can
apply a left-right dual of the lemma to get an analogue statement for L-factorizations.
Another application of Lemma 2 is stated in the following lemma.
Lemma 3. Let M ∈ DA be a monoid, s ∈ M and let ϕ : Σ∗ → M be a homomorphism.
Then, for all finite words u, v ∈ Σ∗ with alph(u) = alph(v), we have
s R sϕ(u) ⇒ s R sϕ(v).
Proof. The case u = v = ε is trivial. Therefore, assume s R sϕ(u) but s 6R sϕ(v ′ a)
for some prefix v ′ a of v with a ∈ Σ. Without loss of generality, let v ′ a be the shortest
such prefix; thus, we have s R sϕ(v ′ ). Since we have alph(u) = alph(v), the is a
prefix u′ a of u. Because of s R sϕ(u), we have s R sϕ(u′ ) R sϕ(u′ a). Thus, we have
sϕ(v ′ ) R s R sϕ(u′ ) R sϕ(u′ )ϕ(a). Now, Lemma 2 yields sϕ(v ′ ) R sϕ(v ′ )ϕ(a), which is
a contradiction to our assumption.
Now, we are prepared to prove the characterization of the Trotter-Weil Hierarchy stated
in Theorem 1. This is done in the following two theorems (see also [15] for the corners
and [14] for the intersection levels). We use the notations XΣD = {XaL , XaR | a ∈ Σ},
YΣD = {YaL , YaR | a ∈ Σ} and some natural variations of it.
Theorem 2. Let M be a finite monoid, ϕ : Σ∗ → M a homomorphism and m ∈ N.
Then:
• M ∈ Rm ⇒ ∃n ∈ N ∀u, v ∈ Σ∗ : u ≡X
m,n v ⇒ ϕ(u) = ϕ(v)
• M ∈ Lm ⇒ ∃n ∈ N ∀u, v ∈ Σ∗ : u ≡Ym,n v ⇒ ϕ(u) = ϕ(v)
• M ∈ Rm ∨ Lm ⇒ ∃n ∈ N ∀u, v ∈ Σ∗ : u ≡XY
m,n v ⇒ ϕ(u) = ϕ(v)
• M ∈ Rm+1 ∩ Lm+1 ⇒ ∃n ∈ N ∀u, v ∈ Σ∗ : u ≡WI
m,n v ⇒ ϕ(u) = ϕ(v)
14
Proof. We fix a homomorphism ϕ : Σ∗ → M and proceed by induction over m. For
m = 1, we have R1 = L1 = R1 ∨ L1 = R2 ∩ L2 = J. Thus, the assertion follows
from a result of Simon [24]. However, we also include a full proof for completeness. Let
M ∈ J and n = |M |, which is the number of J -classes in M (and equal to the number
of R-classes and the number of L-classes). Assume that u ≡X
1,n v for two finite words
u, v ∈ Σ∗ and let u = u0 a1 u1 a2 u2 . . . ak uk be the R-factorization of u. We have k + 1 ≤ n
and, because M is R-trivial, ϕ(u0 ) = ϕ(u1 ) = · · · = ϕ(uk ) = 1 or, more precisely, that
no letter in u0 u1 . . . uk can be mapped by ϕ to a value different from 1. Notice that this
implies ϕ(u) = ϕ(a1 a2 . . . ak ). Furthermore, we have ϕ(a1 ), ϕ(a2 ), . . . , ϕ(ak ) 6= 1 due to
the definition of an R-factorization.
R
X
By definition of ≡X
m,n , we have a1 ∈ alph(v) and u · Xa1 = a2 u2 a3 u3 . . . ak uk ≡1,n−1 v ·
XaR1 . Therefore, we can find a2 in v·XaR1 and have u·XaR1 ·XaR2 = a3 u3 a4 u4 . . . ak uk ≡X
m,n−2
v·XaR1 ·XaR2 . Iterating this approach yields that a1 a2 . . . ak is a subword7 of v. Now, let v =
v0 b1 v1 b2 v2 . . . bl vl be the R-factorization of v. By symmetry, we get ϕ(b1 ), ϕ(b2 ), . . . , ϕ(bl ) 6=
1 and that b1 b2 . . . bl is a subword of u. However, for no j ∈ {1, 2, . . . , l}, the letter bj
can occur in u0 u1 . . . uk since all letters in that word must be mapped to 1. Thus,
b1 b2 . . . bl must in fact be a subword of a1 a2 . . . ak . Again by symmetry, a1 a2 . . . ak
must be a subword of b1 b2 . . . bl and, thus, the two words must be equal. This implies ϕ(u) = ϕ(a1 a2 . . . ak ) = ϕ(b1 b2 . . . bl ) = ϕ(v) where the last equality follows from
ϕ(v0 ) = ϕ(v1 ) = · · · = ϕ(vl ) = 1, which holds due to the R-triviality of M .
The argumentation for u ≡Y1,n v is symmetric using the L-factorization, the case for
WI
u ≡XY
1,n v follows trivially and the case for u ≡1,n v uses the same argumentation.
Now, let M ∈ Rm for an m > 1. This implies M/∼K ∈ Lm−1 and there is an
n′ ∈ N such that u′ ≡Ym−1,n′ v ′ ⇒ ϕ(u′ ) ∼K ϕ(v ′ ) holds for all u′ , v ′ ∈ Σ∗ . Let
r be the number of R-classes in M and let n = n′ + r. Consider the R-factorization
u = u0 a1 u1 a2 u2 . . . ak uk of a finite word u ∈ Σ∗ ; note that k + 1 ≤ r must hold. We have
ui = u · XaR1 XaR2 . . . XaRi−1 XaLi for i = 0, 1, . . . , k − 1 and
uk = u · XaR1 XaR2 . . . XaRk .
For a second finite word v ∈ Σ∗ with u ≡X
m,n v, we know that alph(u) = alph(v). Thus,
L
R
we can apply Xa1 and Xa1 to v and obtain
v0 = v · XaL1
and
v ′ = v · XaR1 .
Y
′
X
By definition of ≡X
m,n , we have v0 ≡m−1,n−1 u0 and v ≡m,n−1 u1 a2 u2 a3 u3 . . . ak uk .
′
Because of k ≤ r < n, we can apply the same argument on v and, by iteration, get
vi = v · XaR1 XaR2 . . . XaRi−1 XaLi for i = 0, 1, . . . , k − 1 and
vk = v · XaR1 XaR2 . . . XaRk
with ui ≡Ym−1,n−i−1 vi for i = 0, 1, . . . , k − 1 and uk ≡X
m,n−k vk . Because of i ≤ k ≤ r − 1,
′
′
we have n − i − 1 = n + r − i − 1 ≥ n + r − (r − 1) − 1 = n′ and ui ≡Ym−1,n′ vi for
7
A finite word c1 c2 . . . cs with ci ∈ Σ is a subword of a (not necessarily finite) word w if we can write
w = w0 c1 w1 c2 w2 . . . cs ws for some words w0 , w1 , . . . , ws .
15
i = 0, 1, . . . , k − 1. For uk and vk , we have uk ≡Ym−1,n−k−1 vk by the definition of the
congruences and, therefore, uk ≡Ym−1,n′ vk because of n−k −1 ≥ n−i−1 ≥ n′ . Summing
this up, we have ui ≡Ym−1,n′ vi and, thus, ϕ(ui ) ∼K ϕ(vi ) for all i = 0, 1, . . . , k.
Since we have defined ui by the R-factorization of u, there is an si ∈ M for any
i ∈ {0, 1, . . . , k} such that ϕ(u0 a1 u1 a2 u2 . . . ai ui )si = ϕ(u0 a1 u1 a2 u2 . . . ai ) holds. For
these, we have
(ϕ(ui )si )M ! ϕ(ui ) R (ϕ(ui )si )M !
because of (ϕ(ui )si )M ! ϕ(ui )si (ϕ(ui )si )M !−1 = (ϕ(ui )si )M ! , which yields
(ϕ(ui )si )M ! ϕ(ui ) = (ϕ(ui )si )M ! ϕ(vi )
by ϕ(ui ) ∼K ϕ(vi ). Thus, we have
ϕ(u0 a1 u1 a2 u2 . . . ak uk ) = ϕ(u0 a1 u1 a2 u2 . . . ak uk ) (sk ϕ(uk ))M !
= ϕ(u0 a1 u1 a2 u2 . . . ak ) (ϕ(uk )sk )M ! ϕ(uk )
= ϕ(u0 a1 u1 a2 u2 . . . ak ) (ϕ(uk )sk )M ! ϕ(vk )
= ϕ(u0 a1 u1 a2 u2 . . . ak )ϕ(vk )
= ϕ(u0 a1 u1 a2 u2 . . . ak−1 uk−1 )ϕ(ak vk )
= ϕ(u0 a1 u1 a2 u2 . . . ak−1 ) (ϕ(uk−1 )sk−1 )M ! ϕ(uk−1 )ϕ(ak vk )
= ϕ(u0 a1 u1 a2 u2 . . . ak−1 ) (ϕ(uk−1 )sk−1 )M ! ϕ(vk−1 )ϕ(ak vk )
= ϕ(u0 a1 u1 a2 u2 . . . ak−2 uk−2 )ϕ(ak−1 vk−1 ak vk )
= ...
= ϕ(v0 a1 v1 a2 v2 . . . ak vk ),
which concludes the proof for Rm .
The proof for Lm is symmetrical. For Rm ∨ Lm , we observe that a monoid is in the
join V ∨ W of two varieties V and W if and only if it is a divisor (i. e. the homomorphic
image of a submonoid) of a direct product M1 × M2 such that M1 ∈ V and M2 ∈ W
[6, Exercise 1.1]. Therefore, if we have a monoid M ∈ Rm ∨ Lm , there are monoids
M1 ∈ Rm and M2 ∈ Lm such that M is a divisor of M1 × M2 ; i. e. there is a submonoid
N of M1 × M2 and a surjective monoid homomorphism ψ : N ։ M . For every a ∈ Σ,
we can find elements ma,1 ∈ M1 and ma,2 ∈ M2 with (ma,1 , ma,2 ) ∈ N such that
ϕ(a) = ψ(ma,1 , ma,2 ). Indeed, we can define the maps ϕ1 : Σ → M1 and ϕ2 : Σ → M2 by
setting ϕ1 (a) := ma,1 and ϕ2 (a) := ma,2 . These maps can be lifted into homomorphisms
ϕ1 : Σ∗ → M1 and ϕ2 : Σ∗ → M2 . By induction, there are n1 and n2 such that u ≡X
m,n1 v
Y
implies ϕ1 (u) = ϕ2 (v) and u ≡m,n2 v implies ϕ2 (u) = ϕ2 (v) for any two finite words
u, v ∈ Σ∗ . By setting n = max{n1 , n2 }, we have
u ≡XY
m,n v ⇒ ϕ1 (u) = ϕ1 (v) and ϕ2 (u) = ϕ2 (v)
16
for all u, v ∈ Σ∗ . For all u, v ∈ Σ∗ with u ≡XY
m,n v, this yields
ϕ(a1 a2 . . . ak ) = ϕ(a1 )ϕ(a2 ) . . . ϕ(ak )
= ψ(ma1 ,1 , ma1 ,2 )ψ(ma2 ,1 , ma2 ,2 ) . . . ψ(mak ,1 , mak ,2 )
= ψ((ma1 ,1 , ma1 ,2 )(ma2 ,1 , ma2 ,2 ) . . . (mak ,1 , mak ,2 ))
= ψ(ma1 ,1 ma2 ,1 . . . mak ,1 , ma1 ,2 ma2 ,2 . . . mak ,2 )
= ψ (ϕ1 (u), ϕ2 (u))
= ψ (ϕ1 (v), ϕ2 (v))
= ϕ(b1 b2 . . . bl )
where u = a1 a2 . . . ak , v = b1 b2 . . . bl and a1 , a2 , . . . , ak , b1 , b2 , . . . , bl ∈ Σ.
Finally, let M ∈ Rm+1 ∩ Lm+1 with m > 1. Denote by 2Σ the monoid of subsets
of Σ whose binary operation is the union of sets. It is easy to see that 2Σ is J -trivial.
Therefore, we have M × 2Σ ∈ Rm+1 ∩ Lm+1 . Next, we lift ϕ : Σ∗ → M into a
homomorphism ϕ̂ : Σ∗ → M × 2Σ by taking the word’s alphabet as the entry in the
second component. If we show u ≡WI
m,n v ⇒ ϕ̂(u) = ϕ̂(v) for a suitable n ∈ N, we have,
v
⇒
ϕ(u)
=
ϕ(v).
The advantage of this approach is that we have
in particular, u ≡WI
m,n
ϕ̂(u) = ϕ̂(v) ⇒ alph(u) = alph(v) for all u, v ∈ Σ∗ by the construction of ϕ̂. Instead of
continuing to write ϕ̂, we simply substitute M by M × 2Σ and ϕ by ϕ̂.
We have M/∼K ∈ Lm and M/∼D ∈ Rm . By ≈, denote the join of ∼K and ∼D .
Since it is a homomorphic image of both, M/∼K and M/∼D , the monoid M/≈ is in
Rm ∩ Lm and we can apply induction, which yields an n′ ∈ N such that u ≡WI
m−1,n′ v
∗
implies ϕ(u) ≈ ϕ(v) for all finite words u, v ∈ Σ . Let c be the sum of the number of Rclasses and the number of L-classes in M and set n = n′ + c. Suppose we have u ≡WI
m,n v
for two finite words u, v ∈ Σ∗ . Consider the R-factorization u = u′0 a1 u′1 a2 u′2 . . . ar u′r of
u and the L-factorization v = v0′ b1 v1′ b2 v2′ . . . bl vl′ of v. Clearly, we have r + 1 + l + 1 ≤ c.
w
w
w
Define the positions pw
0 = −∞, pr+1 = +∞ and pi = Xai (w; pi−1 ) for i = 1, 2, . . . , r
and w = u, v. By Lemma 2, we know that pui denotes the position of ai in the Rw = +∞, q w = −∞ and
factorization for i = 1, 2, . . . , r. Symmetrically, we can define ql+1
0
w
w
qj = Yaj (w; qj+1 ) for j = l, l − 1, . . . , 1 and w = u, v. Again, we know that qjv is the
position of bj in the L-factorization of v for j = 1, 2, . . . , l. Additionally, we have
w
w
w
pw
0 < p1 < · · · < pr < pr+1 and
w
q0w < q1w < · · · < qlw < ql+1
for w = u and w = v by their definition. We are going to show that we have pui ▽ qju ⇔
pvi ▽ qjv for ▽ ∈ {<, =, >} and all i = 1, 2, . . . , r and j = 1, 2, . . . , l. Together, these
results yield that the sequence which is obtained by ordering the pi and qj positions in u
is equal to the corresponding sequence in v. To prove this assertion, assume that we have
qju ≤ pui but qjv > pvi for an i ∈ {1, 2, . . . , r} and a j ∈ {1, 2, . . . , l} (all other cases are
symmetric or analogous). Without loss of generality, we may assume that pui−1 < qju ≤ pui
holds since, otherwise, we can substitute i by a smaller i for which the former holds. Note
that this substitution does not violate the condition qjv > pvi as pvi gets strictly smaller if
17
pui−1
qju
pui
u
qj+1
u=
ai−1
bj
ai
bj+1
v=
ai−1
ai
bj
bj+1
pvi−1
pvi
qjv
v
qj+1
Figure 5: Contradiction: pi is to the right of qj in u but to its left in v.
u
i decreases. Equally without loss of generality, we may assume qju ≤ pui < qj+1
by a dual
argumentation. The situation is presented in Figure 5. We have
R R
R
L L
L
u ) = u ·X
u(pui−1 ,qj+1
a1 Xa2 . . . Xai−1 Ybl Ybl−1 . . . Ybj+1 and
L
R R
R
L L
v
v(pvi−1 ,qj+1
) = v · Xa1 Xa2 . . . Xai−1 Ybl Ybl−1 . . . Ybj+1
WI
WI
v
v
u ) ≡
v
v
u
u
and u(pui−1 ,qj+1
m,n−(i−1)−(l−(j+1)+1) v(pi−1 ,qj+1 ) , which yields u(pi−1 ,qj+1 ) ≡m,2 v(pi−1 ,qj+1 )
because of
n − (i − 1) − (l − (j + 1) + 1) = n′ + c − i + 1 − l + j + 1 − 1
= n′ + c − i − l + j + 1
≥ n′ + c − (r + l) + 1
≥ n′ + c − (c − 2) + 1 = n′ + 3
> 2.
L
v
u )·Y
If qju = pui , we have a contradiction since u(pui−1 ,qj+1
)·
bj contains no ai while v(pvi−1 ,qj+1
u ) while we cannot apply it to
YbLj does. For qju < pui , we can apply Cai ,bj to u(pui−1 ,qj+1
v
v(pvi−1 ,qj+1
) by its definition. Both situations constitute a contradiction.
We have proved that if we order the set {pui , qju | i = 1, 2, . . . , r, j = 1, 2, . . . , l} =
{P1u , P2u , . . . , Ptu } (with t ∈ N0 ) such that
P1u < P2u < · · · < Ptu
holds, then we can set
Psv
=
(
pvi
qjv
Psu = pui for some i ∈ {1, 2, . . . , r}
Psu = qju for some j ∈ {1, 2, . . . , l}
for s = 1, 2, . . . , t and get
P1v < P2v < · · · < Ptv .
18
These positions yield factorizations u = u0 c1 u1 c2 u2 . . . ct ut and v = v0 c1 v1 c2 v2 . . . ct vt
such that cs ∈ {ai , bj | i = 1, 2, . . . , r, j = 1, 2, . . . , l} and Psw denotes the position of cs
in w ∈ {u, v} for s = 1, 2, . . . , t. To apply induction, we are going to show us ≡WI
m−1,n′ vs
for all s = 1, 2, . . . , t next.
To simplify notation, we say “Ps is an R-position” for any s ∈ {1, 2, . . . , t} if Psu = pui
for some i ∈ {1, 2, . . . , r} (or, equivalently, if Psv = pvi for some i) and we say “Ps is an
L-position” if Psu = qju for some j ∈ {1, 2, . . . , l} (or, equivalently again, if Psv = qjv for
some j). Note that this definition is not exclusive, i. e. there can be a position which is
both, an R-position and an L-position.
Next, we consider the corner cases of u0 /v0 and ut /vt . If P1 is an R-position, we have
c1 = a1 and
u0 = u · XaL1
as well as
v0 = v · XaL1 ,
WI
which yields u0 ≡WI
m−1,n′ v0 by definition of ≡m,n and because of c > 0. If P1 is an
L-position, we have c1 = b1 and
u0 = u · YbLl YbLl−1 . . . YbL1 as well as
v0 = v · YbLl YbLl−1 . . . YbL1 .
Because l < c, u0 ≡WI
m−1,n′ v0 holds also in this case. For ut and vt , we can apply a
symmetric argumentation.
Finally, we distinguish four cases for a fixed s ∈ {1, 2, . . . , t − 1}. If Ps and Ps+1 are
both R-positions, then we have cs = ai and cs+1 = ai+1 for some i ∈ {1, 2, . . . , r} and
also
us = u · XaR1 XaR2 . . . XaRi XaLi+1 as well as
vs = v · XaR1 XaR2 . . . XaRi XaLi+1 .
WI
By definition of ≡WI
m,n , because of i + 1 ≤ c, we thus have us ≡m−1,n′ vs . A symmetric
argument applies if both, Ps and Ps+1 , are L-positions. If Ps is an R-position but
Ps+1 is an L-position, then cs = ai for some i ∈ {1, 2, . . . , r} and cs+1 = bj for some
j ∈ {1, 2, . . . , l}, which yields
us = u · XaR1 XaR2 . . . XaRi YbLl YbLl−1 . . . YbLj as well as
vs = v · XaR1 XaR2 . . . XaRi YbLl YbLl−1 . . . YbLj .
WI
Therefore, we have us ≡WI
m−1,n′ vs because of the definition of ≡m,n and n−i−(l−j +1) =
n′ + c − (i + 1 + l) + j ≥ n′ + c − c + 0 = n′ . The fourth case is the most interesting: if
Ps is an L but not an R-position while Ps+1 is an R but not an L-position, then cs = bj
for some j ∈ {1, 2, . . . , l} and cs+1 = ai for some i ∈ {1, 2, . . . , r}. Additionally, we have
w
w
w
w
pw
i−1 < Ps = qj < Ps+1 = pi < qj+1 for w = u and for w = v. We define
ũ = u · XaR1 XaR2 . . . XaRi−1 YbLl YbLl−1 . . . YbLj+1 as well as
ṽ = v · XaR1 XaR2 . . . XaRi−1 YbLl YbLl−1 . . . YbLj+1
19
(we consider the X-blocks as empty – meaning that we do not factorize – if i = 1 and the
Y -blocks as empty if j = l). We have ũ ≡WI
m,n−(i−1)−(l−j) ṽ. Because of n−(i−1)−(l−j) =
n′ + c − (i + l) + j + 1 ≥ n′ + c − (r + l) + 1 ≥ n′ + 1, us = ũ · Cai ,bj , vs = ṽ · Cai ,bj and
WI
the definition of ≡WI
m,n , we have us ≡m−1,n′ vs .
We have shown us ≡WI
m−1,n′ vs for all s = 1, 2, . . . , t and, by induction, therefore, know
that ϕ(us ) ≈ ϕ(vs ), i. e. for a fixed s ∈ {1, 2, . . . , t}, there are w1 , w2 , . . . , wk ∈ Σ∗ such
that
ϕ(us ) = ϕ(w1 ) ∼K ϕ(w2 ) ∼D · · · ∼K ϕ(wk−1 ) ∼D ϕ(wk ) = ϕ(vs )
holds.
Remember that we substituted M by M × 2Σ so that we can assume ϕ(u) = ϕ(v) ⇒
alph(u) = alph(v) for all u, v ∈ Σ∗ . We can extend this implication: if we have ϕ(u) ∼K
ϕ(v) for two u, v ∈ Σ∗ , then, by definition of ∼K , we also have ϕ(u)M ! ϕ(u) = ϕ(u)M ! ϕ(v)
because of ϕ(u)M ! ϕ(u) R ϕ(u)M ! . Therefore, we have alph(u) = alph(u) ∪ alph(v) by
the implication stated above. By symmetry, we, thus, have alph(u) = alph(v). Since we
can apply a similar argumentation for ∼D , we have ϕ(u) ∼K ϕ(v) or ϕ(u) ∼D ϕ(v) ⇒
alph(u) = alph(v) for all u, v ∈ Σ∗ . This yields alph(us ) = alph(w1 ) = alph(w2 ) = · · · =
alph(wk ) = alph(vs ).
Since the factorizations u = u0 c1 u1 c2 u2 . . . ct ut and v = v0 c1 v1 c2 v2 . . . ct vt are subfactorizations from the R-factorization of u and the L-factorization of v, there are xs , ys ∈ M
with
ϕ(u0 c1 u1 c2 u2 . . . cs ) = ϕ(u0 c1 u1 c2 u2 . . . cs us )xs and
ϕ(cs+1 vs+1 cs+2 vs+2 . . . ct vt ) = ys ϕ(vs cs+1 vs+1 cs+2 vs+2 . . . ct vt ).
Because of alph(us ) = alph(wi ) for all i ∈ {1, 2, . . . , k} and by Lemma 3,
(ϕ(us )xs )M ! R (ϕ(us )xs )M ! ϕ(us ) implies (ϕ(us )xs )M ! R (ϕ(us )xs )M ! ϕ(wi ).
Similarly, we have
(ys ϕ(vs ))M ! L ϕ(wi ) (ys ϕ(vs ))M !
for all i ∈ {1, 2, . . . , k}. For ϕ(wi ) ∼K ϕ(wi+1 ), this implies
(ϕ(us )xs )M ! ϕ(wi ) = (ϕ(us )xs )M ! ϕ(wi+1 )
and
ϕ(wi ) (ys ϕ(vs ))M ! = ϕ(wi+1 ) (ys ϕ(vs ))M !
for ϕ(wi ) ∼D ϕ(wi+1 ). In either case, we have
(ϕ(us )xs )M ! ϕ(wi ) (ys ϕ(vs ))M ! = (ϕ(us )xs )M ! ϕ(wi+1 ) (ys ϕ(vs ))M ! ,
which yields for any i ∈ {1, 2, . . . , k − 1}:
ϕ(u0 c1 u1 c2 u2 . . . cs wi cs+1 vs+1 cs+2 vs+2 . . . ct vt )
= ϕ(u0 c1 u1 c2 u2 . . . cs ) (ϕ(us )xs )M ! ϕ(wi ) (ys ϕ(vs ))M ! ϕ(cs+1 vs+1 cs+2 vs+2 . . . ct vt )
= ϕ(u0 c1 u1 c2 u2 . . . cs ) (ϕ(us )xs )M ! ϕ(wi+1 ) (ys ϕ(vs ))M ! ϕ(cs+1 vs+1 cs+2 vs+2 . . . ct vt )
= ϕ(u0 c1 u1 c2 u2 . . . cs wi+1 cs+1 vs+1 cs+2 vs+2 . . . ct vt )
20
So, we can substitute wi by wi+1 and, therefore, also us by vs , i. e. we have
ϕ(u0 c1 u1 c2 u2 . . . cs us cs+1 vs+1 cs+2 vs+2 . . . ct ut )
= ϕ(u0 c1 u1 c2 u2 . . . cs vs cs+1 vs+1 cs+2 vs+2 . . . ct vt ).
Consecutively applying the former equation for s = t, then for s = t − 1 and so on yields
ϕ(u) = ϕ(u0 c1 u1 c2 u2 . . . ct−1 ut−1 ct ut )
= ϕ(u0 c1 u1 c2 u2 . . . ct−1 ut−1 ct vt )
= ϕ(u0 c1 u1 c2 u2 . . . ct−1 vt−1 ct vt )
..
.
= ϕ(v0 c1 v1 c2 v2 . . . ct−1 vt−1 ct vt )
= ϕ(v),
which concludes the proof.
It remains to show the other direction of Theorem 1 (i. e. the converse of Theorem 2).
We will do this by using the equations from Lemma 1. Therefore, we begin with the
following lemma.
Lemma 4. Let Σm , Um and Vm be as in Lemma 1 and let k ∈ N and n ∈ N0 . Then,
for every assignment of variables σ : Σ∗m → Σ∗ and all Z ∈ {X, Y, XY, WI}, we have
σ (JUm Knk ) ≡Z
m,n σ (JV1 Knk )
π
σ (J(Um xm+1 )π Um Knk ) ≡X
m+1,n σ (J(Um xm+1 ) Vm Knk )
σ (JUm (Um xm+1 )π Knk ) ≡Ym+1,n σ (JVm (Um xm+1 )π Knk )
for all m ∈ N.
Proof. We only show the first assertion by induction over m. The other assertions follow
by similar arguments. For m = 1, we may assume n > 0 since, otherwise, there is nothing
to show. We have
U1 = (sx1 )π s(y1 t)π
and
V1 = (sx1 )π t(y1 t)π .
Let u1 = σ (JU1 Knk ) and v1 = σ (JV1 Knk ). First, assume Z = X. We are only interested
in at most n consecutive simultaneous XΣR factorizations of u1 and v1 because, as soon
as we apply at least one XΣL factorization, we know that ≡X
0,n holds. As long as we
R
apply only factorizations Xa with a ∈ alph(σ(sx1 )), the factorization position stays in
the (sx1 )π part of u1 and v1 . Since the number of remaining factorizations decreases,
R
the right parts will eventually be in relation under ≡X
m,0 . If there is at least one Xa
factorization in the sequence where a is in alph(σ(y1 t)) \ alph(σ(sx1 )), the right-hand
side of u1 belongs to the (y1 t)π part and the right-hand side of v1 belongs to the t(y1 t)π
part; but in both words, there are still at least n − 1 instances of σ(y1 t), which implies
21
part containing the difference
um+1 =
(um xm+1 ) (um xm+1 ) . . . (um xm+1 ) um (ym+1 um ) (ym+1 um ) . . . (ym+1 um )
vm+1 =
(um xm+1 ) (um xm+1 ) . . . (um xm+1 ) vm (ym+1 um ) (ym+1 um ) . . . (ym+1 um )
Repetition of the same word
for nk ≥ n times.
Repetition of the same word
for nk ≥ n times.
Figure 6: Schematic representation of um+1 and vm+1 . Note that xm+1 and ym+1 are
identified with σ(xm+1 ) and σ(ym+1 ), respectively.
that the right-hand sides are equal under ≡X
m,n−1 . For Z = Y , the argumentation is
symmetric, which also handles the Z = XY case. The additional Ca,b of Z = WI needs
no special handling since it decreases the first index of ≡WI
m,n to m − 1 = 0 anyway.
To conclude the induction, we show
um+1 = σ (JUm+1 Knk ) ≡Z
m+1,n σ (JVm+1 Knk ) = vm+1
next. For a schematic representation of um+1 and vm+1 , the reader is advised to refer
to Figure 6. For convenience, we identify xm+1 with σ(xm+1 ) and ym+1 with σ(ym+1 ).
We will only show the case Z = WI since this is the most difficult one and the other
ones are similar. In fact, we will proof the following claim by an inner induction on l for
0 ≤ l ≤ n:
l
l
(um xm+1 )l um (ym+1 um )l ≡WI
m+1,l (um xm+1 ) vm (ym+1 um )
The actual assertion then follows for the case l = n. For l = 0, there is nothing to
show. So, let l > 0. First, consider an XaL or XaR factorizations. Only two cases can
emerge: the factorization happens in the (um xm+1 )l part or the factorization happens in
the (ym+1 um )l part (in both words simultaneously). The factorization cannot happen in
the central um or vm part because we have alph(vm ) = alph(um ) ⊆ alph(um xm+1 ).
First, consider the case in which the factorization happens at the beginning. Clearly,
if this is the case, then the factorization must occur in the first (um xm+1 ) factor of both
words at the same position:
Xa
(um xm+1 ) (um xm+1 )l−1 um (ym+1 um )l−1 (ym+1 um )
=
WI
≡m+1,l−1
=
=
(um xm+1 ) (um xm+1 )l−1 vm (ym+1 um )l−1 (ym+1 um )
Xa
22
l
l
L
So, we have (um xm+1 )l um (ym+1 um )l ·XaL ≡WI
m,l−1 (um xm+1 ) vm (ym+1 um ) ·Xa because
the two words are equal. For an XaR factorization, we have to apply induction on l and
get
l−1
vm (ym+1 um )l−1
(um xm+1 )l−1 um (ym+1 um )l−1 ≡WI
m+1,l−1 (um xm+1 )
l
l
in the middle. Thus, we also get (um xm+1 )l um (ym+1 um )l ·XaR ≡WI
m+1,l−1 (um xm+1 ) vm (ym+1 um ) ·
XaR (because the other word parts are equal and because ≡WI
m+1,l−1 is a congruence).
If the factorization happens at the end, the situation is similar. Clearly, the factorization can only happen at the same position in the first (ym+1 um ) factor of each word,
respectively (in fact, it can only happen within the first ym+1 ):
Xa
(um xm+1 ) (um xm+1 )l−1 um (ym+1 um ) (ym+1 um )l−1
=
=
WI
≡m,l−1
=
(um xm+1 ) (um xm+1 )l−1 vm (ym+1 um ) (ym+1 um )l−1
Xa
WI
We have um ≡WI
m,n vm by induction on m and (because l ≤ n) also um ≡m,l−1 vm .
Because the other parts of the words are equal, respectively, and because ≡WI
m,l−1 is a
l v (y
l
L
(u
x
)
u
congruence, we get (um xm+1 )l um (ym+1 um )l · XaL ≡WI
m m+1
m m+1 m ) · Xa .
m,l−1
l
l
R
We have (um xm+1 )l um (ym+1 um )l · XaR ≡WI
m+1,l−1 (um xm+1 ) vm (ym+1 um ) · Xa directly
because the word parts on the right coincide.
Because YaL and YaR factorizations are symmetrical, the only remaining case is a Ca,b
factorization. It is not possible that Ca,b is defined on one of the words but not on
the other because in each respective part the same letters appear. If Ca,b is defined on
(um xm+1 )l um (ym+1 um )l and on (um xm+1 )l vm (ym+1 um )l , then the only possible situation
is that the XaL factorization happens at the same position of the first (ym+1 um ) part in
each word and the YbR factorization happens at the same position of the last (um xm+1 )
factor in each word:
23
Xa
Yb
(um xm+1 )l−1 (um xm+1 ) um (ym+1 um ) (ym+1 um )l−1
=
=
WI
≡m,l−1
(um xm+1 )l−1 (um xm+1 ) vm (ym+1 um ) (ym+1 um )l−1
Yb
Xa
As in the previous case, we have um ≡WI
m,l−1 vm and, because the other parts of the words
l
l
l
coincide respectively, also (um xm+1 ) um (ym+1 um )l ·Ca,b ≡WI
m,l−1 (um xm+1 ) vm (ym+1 um ) ·
Ca,b .
Theorem 3. Let m, n ∈ N. Then:
• Σ∗ /≡X
m,n ∈ Rm
• Σ∗ /≡XY
m,n ∈ Rm ∨ Lm
• Σ∗ /≡Ym,n ∈ Lm
• Σ∗ /≡WI
m,n ∈ Rm+1 ∩ Lm+1
Proof. To prove the theorem, one needs to show that the equations from Lemma 1 hold
in the respective monoid. To do this, it is worthwhile to make an observation: choose
m, n ∈ N and Z ∈ {X, Y, XY, WI} arbitrarily and let M = Σ∗ /≡Z
m,n . The observation is
Z
that an equation α = β holds in M if and only if σ (JαKn·M ! ) ≡m,n σ (JβKn·M ! ) holds for
all assignments σ : Γ → Σ∗ where Γ is the alphabet of α and β (i. e. the set of variables
appearing in α and β).
WI
So, to show Σ∗ /≡WI
m,n ∈ Rm+1 ∩ Lm+1 , it suffices to show that σ (JUm Kn·M ! ) ≡m,n
σ (JVm Kn·M ! ) holds for all assignments of variables σ : Σ∗m → Σ∗ . However, this follows from the first assertion of Lemma 4. The same is true for R1 and L1 . To show
Σ∗ /≡X
m+1,n ∈ Rm+1 for m ∈ N0 , we use the second assertion of Lemma 4 and, for
∗
Y
Σ /≡m+1,n ∈ Lm+1 , we use the third one.
To prove that M/≡XY
m,n is in Rm ∨Lm , one can recycle an observation from the proof of
Theorem 2: a monoid is in the join V ∨ W of two varieties V and W if and only if it is a
divisor of a direct product M1 × M2 such that M1 ∈ V and M2 ∈ W. Indeed, for any two
congruences C1 and C2 over a monoid N , N/ (C1 ∩ C2 ) is a divisor of N/C1 × N/C2 (as can
X
be shown easily). Therefore, M/≡XY
m,n is a divisor of the direct product of M/≡m,n ∈ Rm
and M/≡Ym,n ∈ Lm .
To conclude this section, we finally state and prove the converse of Lemma 1. This
gives us a full characterization of the Trotter-Weil Hierarchy in terms of equations (see
24
also [12]). Note that we include this proof only for the sake of completeness. Strictly
speaking, it is not necessary for the remainder of this paper.
Lemma 5. Define the π-terms
U1 = (sx1 )π s(y1 t)π
and
V1 = (sx1 )π t(y1 t)π
over the alphabet Σ1 = {s, t, x1 }. For m ∈ N, let xm+1 and ym+1 be new characters not
in the alphabet Σm and define the π-terms
Um+1 = (Um xm+1 )π Um (ym+1 Um )π
and
Vm+1 = (Um xm+1 )π Vm (ym+1 Um )π
over the alphabet Σm+1 = Σm ⊎ {xm+1 , ym+1 }.
Then we have
M ∈ R1 = L1 = J ⇒ U1 = V1 holds in M ,
M ∈ Rm+1 ⇒ (Um xm+1 )π Um = (Um xm+1 )π Vm holds in M ,
M ∈ Lm+1 ⇒ Um (ym+1 Um )π = Vm (ym+1 Um )π holds in M and
M ∈ Rm+1 ∩ Lm+1 ⇒ Um = Vm holds in M
for all m ∈ N.
Proof. As stated in the proof of Lemma 1, the first implication is due to a well-known
characterization in terms of equations of J [20]. The other implications can be proved
by combining Lemma 4 and Theorem 2. For example, suppose M ∈ Rm+1 and let
σ : (Σm ⊎ {xm+1 })∗ → M be an arbitrary assignment of variables. We need to show
σ (J(Um xm+1 )π Um KM ! ) = σ (J(Um xm+1 )π Vm KM ! ). By Theorem 2, there is an n ∈ N such
∗
that u ≡X
m+1,n v implies σ(u) = σ(v) for all u, v ∈ (Σm ⊎ {xm+1 }) . Thus, for this n, we
have
σ (J(Um xm+1 )π Um KM ! ) = σ (J(Um xm+1 )π Um Kn·M ! )
= σ (J(Um xm+1 )π Vm Kn·M ! ) = σ (J(Um xm+1 )π Vm KM ! )
π
since we have J(Um xm+1 )π Um Kn·M ! ≡X
m+1,n J(Um xm+1 ) Vm Kn·M ! by Lemma 4.
6 Relations and Equations
Infinite Version of the Relations. So far, we have mainly considered finite words. In
this section, which connects the relational approach outlined above with equations, we
finally need to consider generalized words. Thus, we also define an infinite version of the
above relations. Here, we allow an arbitrary number of factorizations: for m ∈ N0 and
Z ∈ {X, Y, XY, WI}, define
Z
u ≡Z
m v ⇔ ∀n ∈ N : u ≡m,n v.
Notice that this definition is not useful for finite words as every one of them is in its own
class.
25
Connecting the Relations and Equations. In order to solve the word problem for
π-terms over the varieties in the Trotter-Weil Hierarchy, one can use the following connection between the relations defined above and equations in these varieties, which is
straightforward if one makes the transition from finite to infinite words. Besides its use
for the word problem for π-terms, this connection is also interesting in its own right as
it can be used to prove or disprove equations in any of the varieties. As the class of
monoids in which an equation α = β holds is a variety, one can see the assertion for the
join levels as an implication of the ones for the corners.
Theorem 4. Let α and β be two π-terms. For every m ∈ N, we have:
JαKω+ω∗ ≡X
m JβKω+ω ∗ ⇔ α = β holds in Rm
JαKω+ω∗ ≡Ym JβKω+ω∗ ⇔ α = β holds in Lm
JαKω+ω∗ ≡XY
m JβKω+ω ∗ ⇔ α = β holds in Rm ∨ Lm
JαKω+ω∗ ≡WI
m JβKω+ω ∗ ⇔ α = β holds in Rm+1 ∩ Lm+1
Using Fact 1, which states that DA is equal to the union of all varieties in the TrotterWeil Hierarchy, we immediately get the following corollary.
⇔ α = β holds in DA
Corollary 1. ∀m ∈ N : JαKω+ω∗ ≡XY
m JβKω+ω ∗
To prove Theorem 4, we need two technical lemmas. The first one basically says that
a sufficiently large finite power is as good as an ω + ω ∗ power.
Lemma 6. Let m ∈ N0 , Z ∈ {X, Y, XY, WI} and let u and v be accessible words. Then:
ω+ω
k+1
u ≡Z
≡Z
m,n v ⇒ ∀ 0 ≤ k ≤ n : u
m,k v
∗
Proof. The case m = 0 is trivial. Therefore, let m > 0 and continue by induction over
k. Again, the case k = 0 is trivial. To complete the induction, it remains to
show that
ω+ω ∗ holds for k < n. Obviously, alph(uk+2 ) = alph v ω+ω ∗ is satisfied
uuk+1 ≡Z
v
m,k+1
Y
by assumption. Now assume Z = X. The assumption u ≡X
m,n v implies u ≡m−1,n−1 v.
∗
By induction on m, this yields uk+1 ≡Ym−1,k v ω+ω and u ≡Ym−1,k v. Because ≡Ym−1,k is
∗
∗
a congruence, this shows uuk+1 ≡Ym−1,k vv ω+ω = v ω+ω . Let a ∈ alph(u) = alph(v). It
∗
ω+ω ∗ · X R . For
remains to show uuk+1 · XaL ≡Ym−1,k v ω+ω · XaL and uuk+1 · XaR ≡X
a
m,k v
L , which in turn implies
L ≡Y
v
·
X
v
implies
u
·
X
the former, note that u ≡X
a
a
m,n
m−1,n−1
∗
∗
uuk+1 · XaL = u · XaL ≡Ym−1,k v · XaL = vv ω+ω · XaL = v ω+ω · XaL . For the latter, note that
R
R
R
X
R
X
u ≡X
m,n v implies u · Xa ≡m,n−1 v · Xa and, thus, u · Xa ≡m,k v · Xa . By induction on
∗
ω+ω . Together these yield uuk+1 · X R = (u · X R )uk+1 ≡X
k, we also have uk+1 ≡X
a
a
m,k
m,k v
∗
∗
(v · XaR )v ω+ω = v ω+ω · XaR .
The case for Z = Y is symmetric and the case for Z = XY follows directly. Finally,
for Z = WI the argumentation is analogous because, for k > 0, neither uuk+1 · Ca,b nor
∗
v ω+ω · Ca,b is defined for any pair (a, b) of letters.
26
The second technical lemma states that, for finitely many factorization steps, only
finitely many positions of ω + ω ∗ are relevant.
Lemma 7. Let m, n ∈ N0 , Z ∈ {X, Y, XY, WI} and let γ be a π-term. Then
JγKk ≡Z
m,n JγKω+ω ∗
holds for all k ∈ N0 with k > n.
Proof. The cases for m = 0 or n = 0 are trivial. Thus, assume m > 0 and n > 0. If
γ = ε or γ = a for an a ∈ Σ, then JγKk = γ = JγKω+ω∗ . If γ = αβ for two π-terms α
Z
Z
and β, then by induction JαKk ≡Z
m,n JαKω+ω ∗ and JβKk ≡m,n JβKω+ω ∗ hold. As ≡m,n is a
congruence, this implies JγKk ≡Z
m,n JγKω+ω ∗ .
ω+ω ∗
π
Finally, let γ = (α) for a π-term α. It remains to show that JαKkk ≡Z
m,n JαKω+ω ∗ .
Clearly, the alphabetic condition is satisfied and, by induction, JαKk ≡Z
m,n JαKω+ω ∗ holds.
k
For Z = X, let a ∈ alph(α). By induction on m and n, we have JαKk = JγKk ≡Ym−1,n−1
∗
L
k
L
L
Y
JγKω+ω∗ = JαKω+ω
ω+ω ∗ . We also have JαKk · Xa = JαKk · Xa ≡m−1,n−1 JαKω+ω ∗ · Xa =
∗
∗
ω+ω
R
X
R
k
X
L
JαKω+ω
ω+ω ∗ ·Xa by definition of JαKk ≡m,n JαKω+ω ∗ . To show JαKk ·Xa ≡m,n−1 JαKω+ω ∗ ·Xa ,
X
R
we use JαKk · XaR ≡X
m,n−1 JαKω+ω ∗ · Xa , which holds because of JαKk ≡m,n JαKω+ω ∗ , and
ω+ω ∗
k−1
X
X
show JαKk ≡m,n−1 JαKω+ω . This is sufficient because ≡m,n−1 is a congruence and we
∗
∗
ω+ω ∗
have uuω+ω = uω+ω for all words u. For showing JαKkk−1 ≡X
m,n−1 JαKω+ω , we write
k − 1 = k′ + n for a k′ ∈ N0 and
′
JαKkk−1 = JαKkk JαKnk and
∗
′
∗
ω+ω
k
JαKω+ω
ω+ω ∗ = JαKω+ω ∗ JαKω+ω ∗ .
∗
ω+ω
Setting k = n − 1 in Lemma 6 yields JαKnk ≡X
m,n−1 JαKω+ω ∗ , which concludes the proof
for Z = X since ≡X
m,n−1 is a congruence.
The case for Z = Y is symmetric, which also shows the case for Z = XY . For
∗
Z = WI, we note that JαKω+ω
ω+ω ∗ · Ca,b is defined for no pair a, b of letters. On the other
hand, JαKkk · Ca,b can only be defined for k = 1, in which case we are done because we
have n = 0.
Now, we are finally prepared to prove Theorem 4.
Proof for Theorem 4. The proof is structurally identical for all stated varieties. Therefore, we limit our discussion to Rm .
First, let JαKω+ω∗ ≡X
m JβKω+ω ∗ . Choose a monoid M ∈ Rm and an assignment of
∗
variables σ : Σ → M . Because M is in Rm , there is an n ∈ N such that u ≡X
m,n v
implies σ(u) = σ(v) for any two words u, v ∈ Σ∗ (see also Theorem 2). Now, choose
c ∈ N with M ! · c > n. Then by assumption and Lemma 7, we have
X
X
∗
Σ∗ ∋ JαKM !·c ≡X
m,n JαKω+ω ∗ ≡m,n JβKω+ω ∗ ≡m,n JβKM !·c ∈ Σ
and, therefore, σ(JαKM ! ) = σ(JαKM !·c ) = σ(JβKM !·c ) = σ(JβKM ! ), which is equivalent to
α = β holding in M .
27
Now, let JαKω+ω∗ 6≡X
m JβKω+ω ∗ , which implies that there is an n ∈ N such that
X
JαKω+ω∗ 6≡m,n JβKω+ω∗ . Define M := Σ∗ /≡X
m,n , which is in Rm (by Theorem 3), and
choose c ∈ N such that M ! · c > n. Then, by assumption and Lemma 7, we have
X
X
∗
Σ∗ ∋ JαKM !·c ≡X
m,n JαKω+ω ∗ 6≡m,n JβKω+ω ∗ ≡m,n JβKM !·c ∈ Σ .
As assignment of variables σ : Σ∗ → M choose the canonical projection. This yields
σ(JαKM ! ) = σ(JαKM !·c ) 6= σ(JβKM !·c ) = σ(JβKM ! ), which means that α = β does not hold
in M .
7 Decidability
In the previous section, we saw that checking whether α = β holds in a variety of the
Z
Trotter-Weil Hierarchy boils down to checking JαKω+ω∗ ≡Z
m JβKω+ω ∗ (where ≡m depends
on the variety in question). In this section, we give an introduction on how to do this.
The presented approach works uniformly for all varieties in the Trotter-Weil Hierarchy (in
particular, it also works for the intersection levels, which tend to be more complicated)
and is designed to yield efficient algorithms.
The definition of the relations which need to be tested is inherently recursive. One
would factorize JαKω+ω∗ and JβKω+ω∗ on the first a and/or last b (for a, b ∈ Σ) and test
the factors recursively. Therefore, the computation is based on working with factors of
words of the form JγKω+ω∗ where γ is a π-term.
What happens if we consecutively factorize at a
first/last a is best understood if one considers the
∗
ω+ω ∗ = w,
structure of J(α)π Kω+ω∗ = JαKω+ω
ω+ω ∗ = u
... u u u
u u u ...
which is schematically represented in Figure 7.
1 2 3
−3 −2 −1
Suppose u only contains a single a and we start
with the whole word w(−∞,+∞) . If we factorize
ω-part
ω ∗ -part
on the first a taking the part to the right, then
we end up with the factor w(Xa (w;−∞),+∞) with Figure 7: Representation of uω+ω∗
Xa (w; −∞) = (p, 1) where p is the single a-position in u. If we do this again, we obtain
w((p,2),+∞) . If we now factorize on the next a but take the part to the left, then we get
w((p,2),(p,3)) . Notice that the difference between 2 and 3 is 1 and that there is no way of
getting a (finite) difference larger than one by factorizing on the respective first a. On
the other hand, we can reach any number in N as long as the right position is not in the
ω-part.
Notice that there is also no way of reaching (p, −2) as left border without having
(q, −1) or (q, −2) as right border for a position q ∈ dom(u). These observations (and
their symmetrical duals) lead to the notion of normalizable pairs of positions.
Definition 2. Let γ be a π-term and let w = JγKω+ω∗ . A pair (l, r) of positions in w
such that l is strictly smaller than r is called normalizable (with respect to γ) based on
the following rules:
• Any pair is normalizable with respect to γ = ε or γ = a for an a ∈ Σ.
28
• (−∞, +∞) is normalizable with respect to any π-term.
• If γ = αβ for π-terms α and β, l ∈ dom(JαKω+ω∗ ) ⊎ {−∞} and r ∈ dom(JβKω+ω∗ ) ⊎
{+∞}, then (l, r) is normalizable with respect to γ if (l, +∞) is with respect to α
and (−∞, r) is with respect to β.
• If γ = αβ for π-terms α and β and l ∈ dom(JαKω+ω∗ ) ⊎ {−∞} as well as r ∈
dom(JαKω+ω∗ ) (or l ∈ dom(JβKω+ω∗ ) as well as r ∈ dom(JβKω+ω∗ ) ⊎ {+∞}), then
(l, r) is normalizable with respect to γ if it is with respect to α (or β, respectively).
• If γ = (α)π for a π-term α, l = (l′ , n) for l′ ∈ dom(JαKω+ω∗ ) and n ∈ N ⊎ −N and
r = +∞, then (l, r) is normalizable with respect to γ if (l′ , +∞) is with respect to
α and n is in N ⊎ {−1}.
• If γ = (α)π for a π-term α, l = −∞, and r = (r ′ , m) for r ′ ∈ dom(JαKω+ω∗ ) and
m ∈ N ⊎ −N, then (l, r) is normalizable with respect to γ if (−∞, r ′ ) is with respect
to α and m is in {1} ⊎ −N.
• If γ = (α)π for a π-term α, l = (l′ , n) for l′ ∈ dom(JαKω+ω∗ ) and n ∈ N ⊎ −N and
r = (r ′ , m) for r ′ ∈ dom(JαKω+ω∗ ) and m ∈ N ⊎ −N, then (l, r) is normalizable with
respect to γ if
– n ∈ N, m ∈ −N and (l′ , +∞) and (−∞, r ′ ) are normalizable with respect to
α,
– n, m ∈ N or n, m ∈ −N and in both cases m = n and (l′ , r ′ ) is normalizable
with respect to α, or
– n, m ∈ N or n, m ∈ −N and in both cases m = n+1 and (l′ , +∞) and (−∞, r ′ )
are normalizable with respect to α.
This definition looks cumbersome at first. All it does, however, is formalizing our previous observations. This allows us to give a formal inductive proof that these observations
hold for all positions reachable by iterated first/last a factorization. First, we need to extend our notation. For this, we consider (abstract) positions p ∈ {−∞, +∞} ⊎ dom(w) in
a word w to be implicitly linked to w. For a pair of positions (l, r) ∈ ({−∞} ⊎ dom(w))×
(dom(w) ⊎ {+∞}) in an accessible word w, we write
(l, r) · XaL = (l, Xa (w; l)),
(l, r) · XaR = (Xa (w; l), r),
(l, r) · YaL = (l, Ya (w; r)) and
(l, r) · YaR = (Ya (w; r), r)
for all a ∈ alph w(l,r) . Note that we now have w(l,r) · ZaD = w(l,r)·ZaD for all Z ∈ {X, Y }
and all D ∈ {L, R}. In fact, we could have used this as the definition previously. We also
write (l, r) · Ca,b for (l, r) · XaL · YbR if Xa (w; l) is strictly larger than Yb (w; r). Therefore,
we also have w(l,r) · Ca,b = w(l,r)·Ca,b . As with words, we omit the · if we apply a sequence
of steps longer than one, e. g. we simply write (l, r)·XaL YbR = (l, r)·XaL ·YbR = (l, r)·Ca,b .
29
Lemma 8. Let γ be a π-term and let w = JγKω+ω∗ . Additionally, let (l, r) be a normalizable (with respect to γ) pair of positions in w. Then the pairs
(l, r) · XaL , (l, r) · XaR , (l, r) · YaL and (l, r) · YaR
are normalizable with respect to γ for any a ∈ alph(w(l,r) ).
Therefore, (−∞, +∞)·F1 F2 . . . Fn is normalizable with respect to γ for any F1 , F2 , . . . ,
Fn ∈ {XaL , XaR , YaL , YaR , Ca,b | a, b ∈ Σ} (if it is defined).
Proof. As the cases for YaL and YaR are symmetrical, we only show those for XaL and XaR .
Let p = Xa (w; l) for an a ∈ alph(w(l,r) ). Clearly, we have l <µ p <µ r, where µ is the
order type of w, and we need to show that (l, p) and (p, r) are normalizable. For this, we
proceed by induction on the structure of γ. The base case γ = ε or γ ∈ Σ is trivial.
Case γ = αβ. Define u = JαKω+ω∗ and v = JβKω+ω∗ . For l ∈ dom(u) ⊎ {−∞} and
r ∈ dom(u) we have p ∈ dom(u) as well. Additionally, (l, r) needs to be normalizable
with respect to α by the definition of normalizability and we can apply induction. The
same argument, but on β, works for l ∈ dom(v) and r ∈ dom(v) ⊎ {+∞}. For l ∈
dom(u) ⊎ {−∞} and r ∈ dom(v) ⊎ {+∞} we know that (l, +∞) is normalizable with
respect to α and (−∞, r) is with respect to β by the definition of normalizablity. If
p ∈ dom(u), then (p, +∞) = (l, +∞) · XaR and (l, p) = (l, +∞) · XaL . Induction yields
normalizability with respect to α for both and, by the definition of normalizability, we
have that (p, r) and (l, p) are normalizable with respect to γ. For p ∈ dom(v), we can
apply a similar argument, as then (−∞, p) = (−∞, +∞) · XaL and (p, r) = (−∞, r) · XaR
are normalizable with respect to β.
Case γ = (α)π . Define u = JαKω+ω∗ and let p = (p′ , k). If l = (l′ , n) for an n ∈
N ⊎ −N and r = +∞, then, by the definition of normalizability, we have that (l′ , +∞)
is normalizable with respect to α and n ∈ N ⊎ {−1}. There are two cases: for k =
n ∈ N ⊎ {−1} we know that p′ = Xa (u; l′ ) and, by induction, that (l′ , p′ ), (p′ , +∞) are
normalizable with respect to α. This yields the normalizability with respect to γ of (l, p)
and (p, +∞). For k = n + 1 we know that n 6= −1 and, therefore, that n, k ∈ N. We
also have p′ = Xa (u; −∞) and, thus, that (−∞, p′ ) and (p′ , +∞) are normalizable with
respect to α by induction. By definition, (p, +∞) and (l, p) are normalizable with respect
to γ then. Note that k cannot have any other value than n or n + 1 since otherwise it
could not be the smallest a-position to the right of l.
If l = −∞ and r = (r ′ , m), then k = 1, and p′ = Xa (u; −∞), which yields (−∞, p′ ) =
(−∞, +∞) · XaL and (p′ , +∞) = (−∞, +∞) · XaR . By induction, both of these pairs
are normalizable with respect to α and, by definition of the normalizability, (−∞, p)
is normalizable with respect to γ. Furthermore in this case, we know that (−∞, r ′ ) is
normalizable with respect to α and that m is in {1} ⊎ −N. For m ∈ −N, this shows the
normalizability with respect to γ of (p, r). For m = 1, we have (p′ , r ′ ) = (−∞, r ′ )·XaR and,
by induction, its normalizability with respect to α. This yields that (p, r) is normalizable
with respect to γ.
If l = (l′ , n) and r = (r ′ , m) for n ∈ N and m ∈ −N, we know that (l′ , +∞) and
(−∞, r ′ ) are normalizable with respect to α. For k = n ∈ N, we also know that p′ =
30
Xa (u; l′ ) and, therefore, that (l′ , p′ ) = (l′ , +∞) · XaL and (p′ , +∞) = (l′ , +∞) · XaR are
normalizable with respect to α by induction. Then, by definition, (l, p) and (p, r) are
normalizable with respect to γ. For k = n + 1 ∈ N we have that p′ = Xa (u; −∞)
and, therefore, the normalizability with respect to α of (−∞, p′ ) = (−∞, +∞) · XaL and
(p′ , +∞) = (−∞, +∞) · XaR by induction. This yields the normalizability with respect
to γ of (l, p) and (p, r).
Finally, if l = (l′ , n) and r = (r ′ , m) for n, m ∈ N or n, m ∈ −N, we know that
0 ≤ m−n ≤ 1. Because p must be in between l and r, n = m also implies n = m = k and
that p′ is in between l′ and r ′ as well as p′ = Xa (u; l′ ). In that case, we have that (l′ , r ′ )
and, by induction, also (l′ , p′ ) = (l′ , r ′ ) · XaL and (p′ , r ′ ) = (l′ , r ′ ) · XaR are normalizable
with respect to α. This yields the normalizability with respect to γ of (l, p) and (p, r).
For m = n + 1, we know that (l′ , +∞) and (−∞, r ′ ) are normalizable with respect to
α. Moreover, there are only two cases: k = n and k = m. In the former case, we have
p′ = Xa (u; l′ ) and the normalizability with respect to α of (l′ , p′ ) = (l′ , +∞) · XaL and
(p′ , +∞) = (l′ , +∞)·XaR by induction, which yields the normalizability of (l, p) and (p, r)
with respect to γ. In the latter case, we have p′ = Xa (u; −∞) and the normalizability
with respect to α of (−∞, p′ ) = (−∞, +∞) · XaL and (p′ , r ′ ) = (−∞, r ′ ) · XaR , which
yields the normalizability with respect to γ of (l, p) and (p, r).
The choice of words indicates that normalizability of a pair (l, r) can be used to define
a normalization. Before we give a formal – unfortunately, quite technical – definition of
this, we describe its idea informally. Let us refer back to the schematic representation of
J(α)π Kω+ω∗ = w as given in Figure 7. Basically, there are three different cases for relative
positions of the left border l and the right border r which describe the factor w(l,r) :
1. l is in the ω-part and r is in the ω ∗ -part,
2. l and r are either both in the ω-part or both in the ω ∗ -part and have the same
value there, or
3. l and r are either both in the ω-part or both in the ω ∗ -part but r has a value
exactly larger by one than l.
This is ensured by the normalizability of (l, r). Now, in the first case, we can safely move
l to value 1 (the first position) and r to value −1 (the last position) without changing
the described factor. In the second and third case, we can move l and r to any value –
as long as we retain the difference between the values – without changing the described
factor. Here, we move them to the left-most values (which are 1, 1 or 1, 2). Afterwards,
we go on recursively.
Unfortunately, things get a bit more complicated because l might be −∞ and r might
be +∞. In these cases, we normalize to the left-most or right-most value without changing the factor.
For concatenations of π-terms, we have a similar situation: either l and r belong both
to the left or to the right factor, in which case we can continue by normalization with
respect to that, or l belongs to the left factor and r belongs to the right one. In this
31
case, we have to continue the normalization with (l, +∞) and (−∞, r) in the respective
concatenation parts, as this ensures that the described factor remains unchanged.
Formalizing these ideas results in the following inductive definition.
Definition 3. Let γ be a π-term, w = JγKω+ω∗ and (l, r) a normalizable pair of positions
γ
in w. The normalized pair (l, r) = (¯l, r̄) with respect to γ is defined recursively:
• For γ = ε or γ = a ∈ Σ define ¯l = l and r̄ = r.
• If γ = αβ for π-terms α and β, l ∈ dom(JαKω+ω∗ ) ⊎ {−∞} and r ∈ dom(JβKω+ω∗ ) ⊎
α
{+∞}, then define ¯l as the first component of (l, +∞) and r̄ as the second comβ
ponent of (−∞, r) .
• If γ = αβ for π-terms α and β and l ∈ dom(JαKω+ω∗ ) ⊎ {−∞} as well as r ∈
dom(JαKω+ω∗ ) (or l ∈ dom(JβKω+ω∗ ) as well as r ∈ dom(JβKω+ω∗ ) ⊎ {+∞}), then
α
β
define (¯l, r̄) = (l, r) (or (¯l, r̄) = (l, r) , respectively).
• If γ = (α)π for a π-term α, then:
– if l = −∞, define ¯
l = −∞,
– if r = +∞, define r̄ = +∞,
– if l = (l′ , n) and r = +∞, define ¯l = (l¯′ , n̄) with l¯′ given by the first component
α
of (l′ , +∞) and n̄ given by
(
1
if n ∈ N
n̄ =
−1 if n = −1,
– if l = −∞ and r = (r ′ , m), define r̄
α
component of (−∞, r ′ ) and m̄ given
(
1
m̄ =
−1
= (r¯′ , m̄) with r¯′ given by the second
by
if n = 1
if n ∈ −N,
– if l = (l′ , n) and r = (r ′ , m) with n ∈ N and m ∈ −N, define ¯l = (l¯′ , 1) with
α
l¯′ being by the first component of (l′ , +∞) and define r̄ = (r¯′ , −1) with r̄ ′
α
given by the second component of (−∞, r ′ ) ,
– if l = (l′ , n) and r = (r ′ , m) with n = m, define ¯l = (l¯′ , n̄) and r̄ = (r¯′ , m̄)
α
with (l¯′ , r¯′ ) = (l′ , r ′ ) and n̄ = m̄ = 1, and
– if l = (l′ , n) and r = (r ′ , m) with m = n + 1, define ¯l = (l¯′ , n̄) and r̄ = (r¯′ , m̄)
α
with l¯′ given by the first component of (l′ , +∞) , r¯′ given by the second
α
component of (−∞, r ′ ) , n̄ = 1 and m̄ = n̄ + 1 = 2.
One should note that if we normalize a normalizable pair (l, r), then the resulting pair
is normalizable itself. Indeed, if we normalize an already normalized pair again, we do
not change any values. Formally, this can be proved by an induction on the structure of
the π-term. As an example for such an induction, we prove the following lemma which
states that normalizing a pair of positions does not change the described factor.
32
Lemma 9. Let γ be a π-term and let (l, r) be a normalizable pair of positions in w =
JγKω+ω∗ . Then
w(l,r) = w(l,r)γ
holds.
γ
Proof. Define (l, r) = (¯l, r̄) and proceed by induction on the structure of γ. The base
cases for γ = ε and γ ∈ Σ are trivial.
If γ = αβ for π-terms α and β, then define u = JαKω+ω∗ and v = JβKω+ω∗ . If
l ∈ dom(u) ⊎ {−∞} and r ∈ dom(v) ⊎ {+∞}, then
w(l,r) = u(l,+∞) v(−∞,r) = u(l,+∞)α v
β
(−∞,r)
= w(l,r)γ .
If l ∈ dom(u) ⊎ {−∞} and r ∈ dom(u), then
w(l,r) = u(l,r) = u(l,r)α = w(l,r)γ .
The case l ∈ dom(v) and r ∈ dom(v) ⊎ {+∞} is symmetrical.
If γ = (α)π for a π-term α, then define u = JαKω+ω∗ . The case l = −∞ and r = +∞
α
is trivial. If l = (l′ , n) for an n ∈ N ⊎ −N and r = +∞, define l¯′ by (l′ , +∞) = (l¯′ , +∞).
For n ∈ N we then have
∗
ω+ω ∗
w(l,r) = w((l′ ,n),+∞) = uω+ω
=
u
′
′
((l ,n),+∞)
∗
((l ,1),+∞)
∗
because of uω+ω = uuω+ω and further
∗
∗
w(l,r) = u(l′ ,+∞)uω+ω = u(l′ ,+∞)α uω+ω = u(l¯′ ,+∞) uω+ω
∗
= uω+ω
= w((l′ ,n),+∞)γ = w(l,r)γ
¯′
∗
((l ,1),+∞)
and for n = −1 – the only remaining case – we have
w(l,r) = w((l′ ,−1),+∞) = u(l′ ,+∞) = u(l′ ,+∞)α = u(l¯′ ,+∞) = w((l¯′ ,−1),+∞) = w(l,r)γ .
The case for l = −∞ and r = (r ′ , m) is symmetrical.
Therefore, we can assume l = (l′ , n) and r = (r ′ , m). The case n ∈ N and m ∈ −N is
proved by a calculation similar to the one given above. For n = m we have
w(l,r) = w((l′ ,n),(r′ ,n)) = u(l′ ,r′ ) = u(l′ ,r′ )α = w(l,r)γ
and for m = n + 1 we have
w(l,r) = u(l′ ,+∞) u(−∞,r′ ) = u(l′ ,+∞)α u(−∞,r′ )α = w(l,r)γ .
33
Another observation is crucial for the proof of the decidability: after normalizing a pair
(l, r) the values belonging to the ω+ω ∗ parts for the two positions are all in {1, 2, −2, −1}.
But: there are only finitely many such positions in any word w = JγKω+ω∗ for a π-term
γ. Because the normalization preserves the described factor, this means that there are
only finitely many factors which can result from a sequence of first/last a factorizations.
Plugging all these ideas and observations together yields a proof for the next theorem.
Theorem 5. For two π-terms α and β over the same alphabet Σ, it is decidable for all
m ∈ N0 and all Z ∈ {X, Y, XY, WI} whether
JαKω+ω∗ ≡Z
m JβKω+ω ∗
holds. Furthermore, it is decidable whether
∀m ∈ N : JαKω+ω∗ ≡XY
m JβKω+ω ∗
holds.
Proof. We decide whether u = JαKω+ω∗ ≡Z
m JβKω+ω ∗ = v holds by trying to find a
factorization sequence F1 F2 . . . Fn with F1 , F2 , . . . , Fn ∈ {XaL , XaR , YaL , YaR , Ca,b | a, b ∈
Σ} that can be applied to u but not to v (or vice versa), i. e. we rather try to decide
u 6≡Z
m v instead.
Which sequences need to be tested depends on the actual relation. For example, for
R
L
≡X
m , we can start with arbitrary many Xa factorizations but, as soon as we apply an Xa ,
L
R
Ya or Ya factorization, we have changed the direction and we have to decrease m by one.
If we did this because of a YaD factorization (with D ∈ {R, L}), we also need to switch
to a ≡Ym−1 mode, in which case we can continue with arbitrary many YaL factorizations
while any other factorization decreases the remaining number of direction changes and
WI
might also switch back to a ≡X
m−2 mode. If we want to test ≡m , we also need to allow
Ca,b factorizations and we need to count the number of direction changes appropriately.
For testing ≡XY
m for all m ∈ N, the situation is simpler: here, we do not need to keep
track of the remaining direction changes as there always is an arbitrary number of them
left. Clearly, we can construct a deterministic finite automaton for any of the relations
which accepts exactly those sequences F1 F2 . . . Fn which need to be tested.
To test whether a factorization sequence can be applied on u, we construct an additional deterministic finite automaton. The states of this automaton are the normalized
pairs of positions in u (of which there only finitely many, as discussed above). The initial
state is (−∞, +∞) and all states are final. This does not result in a trivial automaton because it will not be complete in general. We have an F -labeled transition with
α
F ∈ {XaD , YaD , Ca,b | a, b ∈ Σ, D ∈ {L, R}} from (l, r) to (l, r) · F if (l, r) · F is defined.
Note that normalization does not change the implicitly stored factor of u by Lemma 9.
This automaton, by construction, accepts exactly those factorization sequences which
can be applied to u.
Using the same construction, we can also get such an automaton for v. We intersect
both automata with the one which accepts the relevant factorization sequences. For the
34
resulting pair of automata, we check the symmetric difference of the accepted languages
for emptiness. It is empty if any only if u and v are in relation.
Together with Theorem 4, this gives the following decidability result. Note that previous partial results exist: Almeida proved decidability for J [1, 3], Almeida and Zeitoun
proved it for R [4] and Moura for DA [18].
Corollary 2. The word problems for π-terms over Rm , Lm , Rm ∨ Lm and Rm ∩ Lm are
decidable for any m ∈ N. Moreover, the word problem for π-terms over DA is decidable.
8 Nondeterministic Logarithmic Space
In the presented algorithm, we construct for a π-term γ over the alphabet Σ a deterministic finite automaton which accepts exactly those factorization sequences which can be
applied to JγKω+ω∗ = w. For this, we had to store normalized pairs (l, r) of positions
γ
in w and we had to compute (l, r) · F for a factorization F ∈ {XaD , YaD , Ca,b | a, b ∈
Σ, D ∈ {L, R}}. In this section, we will show that both can be done by a deterministic
Turing machine in logarithmic space. Afterwards, we will show that this yields membership of the word problems for π-terms to the class of problems which can be solved by
nondeterministic Turing machines within logarithmic space.
Observations and Ideas. We will start by having a close look on how to store a position
p (and, therefore, a pair of positions) in the word w = JγKω+ω∗ on a Turing machine. To
do this, we first store the position in γ which corresponds to p in w; this, basically, is a
simple pointer. Additionally, we need to store to which value in N ⊎ −N the position p
belongs for all relevant π-exponents; this, we can do by storing a pointer to the π-position
in γ together with the corresponding value. If p is part of a normalized pair of positions,
the values can only be in {1, 2, −2, −1}, which is a finite set and, thus, needs only finite
information.
Example 1. Have a look at the π-term γ = a (b(c)π )π a(b)π c. The word JγKω+ω∗ can be
represented by the following tree:
a
a
b
c ...
...
... c
...
c
b ...
b
c c ...
35
... c
... b
The highlighted position can also be represented this way:
a(b(c)
π
)
π
a(b)
π
c
2 −1
Note that the last π-position does not have a value since it is not relevant for the position.
Also note that storing the values on a second tape directly under the π-position would
require linear space (for a position belonging to a normalized pair). We store them in a
position/value list which will turn out to be more efficient after some modifications.
Unfortunately, this approach still requires at least linear space. But for normalized
pairs, we can optimize it further if we look at the definition of the normalization. At
some point the two positions l and r in the normalized pair (l, r) branch since l is a
smaller position than r. This can either happen because l = −∞ or r = +∞, because
there are sub-π-terms α and β and l belongs to JαKω+ω∗ while r belongs to JβKω+ω∗ , or
because of a sub-π-term of the form (α)π where l has a different value compared to r. In
the former two cases, we call the pair c-branching 8 and in the latter case π-branching.
Whichever is the case, we know that the values of l and r belonging to hierarchically
higher π-positions are always equal and, by definition of the normalization, are equal to
1. Thus, we do not need to store these values explicitly; instead, we are going to store
the branching position.
If the branching position is a π-position (i. e. the pair is π-branching), then, for a
normalized pair, the values of l and r at this position can, by definition, only be 1 for
l and 2 for r or 1 for l and −1 for r. This information can be stored alongside the
branching position in constant space.
To store the values hierarchically below the branching position, we need to have an even
closer look at normalized position pairs. Before we do this, however, it is convenient to
define four position forms: a position is in +-form if all its values for relevant π-positions
are from N and it is in −-form if they are from −N. Positions in ∓-form may only have
values from N for π-positions which are hierarchically lower than the first π-position with
a value from N and positions in ±-form are defined symmetrically. More formally, we
define:
Definition 4. A position p ∈ dom(JγKω+ω∗ ) for a π-term γ is in
• +-form (with respect to γ) if
– γ = ε or γ ∈ Σ,
– γ = αβ for π-terms α and β and p is in +-form with respect to its respective
sub-π-term, or
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ N and p′ is in +-form with
respect to α,
8
The “c” is for concatenation.
36
• −-form (with respect to γ) if
– γ = ε or γ ∈ Σ,
– γ = αβ for π-terms α and β and p is in −-form with respect to its respective
sub-π-term, or
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ −N and p′ is in −-form with
respect to α,
• ∓-form (with respect to γ) if
– γ = ε or γ ∈ Σ,
– γ = αβ for π-terms α and β and p is in ∓-form with respect to its respective
sub-π-term,
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ N and p′ has +-form, or
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ −N and p′ has ∓-form, or
• ±-form (with respect to γ) if
– γ = ε or γ ∈ Σ,
– γ = αβ for π-terms α and β and p is in ±-form with respect to its respective
sub-π-term,
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ −N and p′ has −-form, or
– γ = (α)π for a π-term α, p = (p′ , n) with n ∈ N and p′ has ±-form.
Example 2. The position
a(b(c)
π
)
π
a(b)
π
c
2 −1
from the previous example is in ∓-form but not in any of the other three forms.
The similar position represented by
a(b(c)
π
)
2
π
a(b)
π
c
3
is in +-form, in ∓-form and in ±-form but it is not in −-form.
What use are these definitions for our goal of storing positions efficiently? If we know
that a position is in +-form or −-form, then we do not need to store whether a value is
from N or from −N. Similarly, if a position is in ∓-form or in ±-form, then we only need
to store one potential π-position at which the values switch form −N to N or vice versa.
While this does not seem to be a huge gain since we still need to store the actual value,
it will turn out to be crucial later on.
37
Next, we need to make some further observations. Consider a π-term γ and define
w = JγKω+ω∗ . If we start with a position p ∈ dom(w) in ∓-form and we go to the next
a on the right (i. e. we compute Xa (w; p)), then the follow-up position p′ is in ∓-form
as well because strictly after the branching of p and p′ all values for relevant π-position
have to be 1 (since otherwise there would already have been an a before). By symmetry,
if we start in ±-form and go to the previous a on the left, the resulting position will also
be in ±-form.
Example 3. Look again at the π-term γ = a (b(c)π )π a(b)π c. Suppose we are in the
position p in w = JγKω+ω∗ and advance to p′ = Xb (w; p):
a
b
a
...
...
c
b ...
b
... b
p′
c ...
... c
c c ...
p
... c
Clearly, p′ is in ∓-form.
If we advanced to p′′ = Xc (w; p) instead, the situation would be as follows:
a
b
c ...
a
...
... c
...
c
b ...
b
c c c ...
p p′′
... b
... c
Again, p′′ clearly has ∓-form.
More formally, we can prove the following lemma:
Lemma 10. Let γ be a π-term and w = JγKω+ω∗ . Then
• Xa (w; −∞) is in +-form for any a ∈ alph(w),
• if l ∈ dom(w) is in +-form, then so is Xa (w; l) for all a ∈ alph(w(l,+∞) ),
38
• if l ∈ dom(w) is in ∓-form, then so is Xa (w; l) for all a ∈ alph(w(l,+∞) ),
• Ya (w; +∞) is in −-form for any a ∈ alph(w),
• if r ∈ dom(w) is in −-form, then so is Ya (w; r) for all a ∈ alph(w(−∞,r) ), and
• if r ∈ dom(w) is in ∓-form, then so is Ya (w; r) for all a ∈ alph(w(−∞,r) ).
Proof. By symmetry, we only have to prove the first three assertions.
We start by proving the first one. The cases γ = ε and γ ∈ Σ are trivial. For the other
cases, let a ∈ alph(w) and define p = Xa (w; −∞). If γ = αβ for two π-terms α and β, we
define u = JαKω+ω∗ and v = JβKω+ω∗ . Now, we have p = Xa (u; −∞) or p = Xa (v; −∞).
In either case, we can apply induction, which yields that p is in +-form with respect
to its sub-π-term, and we are done. If γ = (α)π for a π-term α, define u = JαKω+ω∗ .
Clearly, p has to be equal to (p′ , 1) for p′ = Xa (u; −∞). By induction, we have that p′ is
in +-form and, since 1 ∈ N, we are done.
The second assertion can be proved similarly. Let l ∈ dom(w) be in +-form. If γ = ε
or γ ∈ Σ, we do not have any remaining positions in dom(w) for p = Xa (w; l). In
the other cases, let a ∈ alph(w(l,+∞) ). If γ = αβ for two π-terms α and β, we define
u = JαKω+ω∗ and v = JβKω+ω∗ . We have to distinguish: if l and p both are in dom(u)
(or, symmetrically, in dom(v)), then we have p = Xa (u; l) and we can apply induction.
Therefore, p is in +-form with respect to α, which yields that p is also in +-form with
respect to γ. If l ∈ dom(u) and p ∈ dom(v), then we know that p = Xa (v; −∞), which
is in +-form with respect to β by the first assertion and, therefore, also in +-form with
respect to γ.
If γ = (α)π for a π-term α, define u = JαKω+ω∗ . We can write l = (l′ , m) and p = (p′ , k)
for some m, k ∈ N ⊎ −N and l′ , p′ ∈ dom(u). By definition, l′ is in +-form because l is so.
For k, there are only two possible cases: k = m and k = m + 1. In the former case, we
know that p′ = Xa (u; l′ ) and can apply induction to get that p′ is in +-form with respect
to α. Since l is in +-form, we also know that k = m ∈ N. Together, this yields that p is
in +-form with respect to γ. In the latter case k = m + 1, we know that p′ = Xa (u; −∞)
is in +-form by the first assertion. Since l is in +-form, we have m ∈ N and, therefore,
also k = m + 1 ∈ N. Thus, p is in +-form with respect to γ.
Now, we prove the third assertion. Let l be in ∓-form with respect to γ. Again, for
γ = ε or γ ∈ Σ, there is nothing to show. The case for γ = αβ for two π-terms α
and β can be proved analogously to the corresponding case in the proof for the second
assertion.
If γ = (α)π for a π-term α, define u = JαKω+ω∗ and let p = Xa (w; l) for an a ∈
alph(w(l,+∞) ). We can write l = (l′ , m) and p = (p′ , k) for some m, k ∈ N ⊎ −N and
l′ , p′ ∈ dom(u). Again, there are only two possible cases: k = m and k = m + 1. In the
former case we have p′ = Xa (u; l′ ). If k = m ∈ N, we know that l′ is in +-form since
l has to be in ∓-form. By the second assertion, this yields that p′ is in +-form as well.
Therefore, we have that p is in ∓-form. If k = m ∈ −N, we can simply apply induction
and get that p′ is in ∓-form with respect to α. This yields that p is in ∓-form with
respect to γ. In the latter case k = m + 1, we know that p′ = Xa (u; −∞) is in +-form
39
by the first assertion and that m 6= −1. This yields ∓-form for p if k = m + 1 ∈ N or
k = m + 1 ∈ −N.
This observation is also important for pairs which arise by consecutive factorization
at the
first/last a. If we start in (−∞, +∞) and apply a sequence X of elements from
XΣR = XaR a ∈ Σ , then l′ in the resulting pair (l′ , +∞) = (−∞, +∞) · X will be in
∓-form. Equally, for a sequence over YΣL = YaL a ∈ Σ , the right position will be in
±-form. Can we assume that the left position is always in ∓-form and the right one is
always in ±-form? Unfortunately, the answer to this question is “no”. Suppose we start
in the pair (l, r) where l is in ∓-form and r is in ±-form. If we apply XaL for some a ∈ Σ,
then, obviously, we end up in a pair (l, r ′ ) whose right position is in ∓-form. But: the
right position r ′ branches form l at some point and strictly below that point values for
relevant π-positions are equal to 1. So, we could say that for this lower part r ′ , indeed, is
in ±-form. In fact, we will prove that, strictly below the branching, l is always in ∓-form
and r is always in ±-form for any pair (l, r) = (−∞, +∞) · F where F is a sequence
of elements from FΣ = {XaD , YaD , Ca,b | a, b ∈ Σ, D ∈ {L, R}}. The situation at the
branching point itself depends on whether we have a c-branching or a π-branching pair.
For a π-branching pair, we cannot make an assumption on the value of l and r at the
actual branching π-position. We accommodate for this by having two definitions.
Definition 5. Let (l, r) be a pair of positions in JγKω+ω∗ for a π-term γ such that l is
strictly smaller than r.
The pair is called well-c-shaped (with respect to γ) if
• γ = ε or γ ∈ Σ,
• γ = αβ for π-terms α and β, l ∈ dom(JαKω+ω∗ ) ⊎ {−∞}, r ∈ dom(JαKω+ω∗ ) and
(l, r) is well-c-shaped with respect to α,
• γ = αβ for π-terms α and β, l ∈ dom(JβKω+ω∗ ), r ∈ dom(JβKω+ω∗ ) ⊎ {+∞} and
(l, r) is well-c-shaped with respect to β,
• γ = αβ for π-terms α and β, l ∈ dom(JαKω+ω∗ )⊎{−∞}, r ∈ dom(JβKω+ω∗ )⊎{+∞},
and
– l = −∞ or l is in ∓-form with respect to α, and
– r = +∞ or r is in ±-form with respect to β,
or
• γ = (α)π for a π-term α and
– l = −∞ and r = +∞,
– l = −∞ and r is in ±-form,
– l is in ∓-form and r = +∞, or
– l = (l′ , n) and r = (r ′ , n) for an n ∈ N ⊎ −N and (l′ , r ′ ) is well-c-shaped with
respect to α,
40
It is called well-π-shaped (with respect to γ) if γ 6= ε and γ 6∈ Σ as well as
• γ = αβ for π-terms α and β, l, r ∈ dom(JαKω+ω∗ ) and (l, r) is well-π-shaped with
respect to α,
• γ = αβ for π-terms α and β, l, r ∈ dom(JβKω+ω∗ ) and (l, r) is well-π-shaped with
respect to β, or
• γ = (α)π for a π-term α, l = (l′ , n), r = (r ′ , m) for n, m ∈ N ⊎ −N and l′ , r ′ ∈
dom(JαKω+ω∗ ) and
– n = m and (l′ , r ′ ) is well-π-shaped with respect to α, or
– n 6= m, l′ is in ∓-form and r ′ is in ±-form.
Finally, it is called well-shaped (with respect to γ) if it is well-c-shaped or well-π-shaped
(with respect to γ).
The definition of well-c-shapedness is related to c-branching pairs and the definition of
well-π-shapedness is related to π-branching pairs. Note that any pair (l, r) of positions
in JγKω+ω∗ for a π-term is well-c-shaped if l = −∞ or r = +∞. This results in some
asymmetry in the definition.
We proceed by showing that any pair (l, r) which arises by consecutive factorization
at the first/last a is well-c-shaped or well-π-shaped.
Lemma 11. Let (l, r) be a pair of positions in JγKω+ω∗ for a π-term γ. If (l, r) is
well-shaped, then so is (l, r) · F for any F ∈ FΣ (if it is defined).
Therefore, (−∞, +∞) · F1 F2 . . . Fn is well-shaped with respect to γ for any F1 , F2 , . . . ,
Fn ∈ FΣ (if it is defined).
Proof. The second part follows from the first since (−∞, +∞) is well-c-shaped by definition.
Let (l, r) be well-shaped and let p = Xa (w; l) be defined for an a ∈ Σ. Due to
symmetry, it remains to show that (l, p) and (p, r) are well-shaped. Note that this also
includes the Ca,b factorization.
First, consider (l, p). If l = −∞, then p is in +-form by Lemma 10 and, therefore,
also in ±-form. This yields that (−∞, p) is well-c-shaped. Thus, we may safely assume
that l 6= −∞ and proceed by induction on the structure of γ. With this assumption, the
cases γ = ε and γ ∈ Σ cannot occur.
If γ = αβ for two π-terms α and β, define u = JαKω+ω∗ and v = JβKω+ω∗ . If l and r are
both in dom(u), then we know, by definition of well-shapedness, that (l, r) is well-shaped
with respect to α. We also know that p = Xa (u; l) since p must be between l and r. By
induction, we have that (l, p) is well-shaped with respect to α, which yields that (l, p)
is also well-shaped with respect to γ. If l ∈ dom(v) and r ∈ dom(v) ⊎ {+∞}, we can
apply a similar argument. For l ∈ dom(u) and r ∈ dom(v) ⊎ {+∞}, we know that (l, r)
is well-c-shaped and that l is in ∓-form with respect to α. This yields that (l, +∞) is
well-c-shaped with respect to α. By induction, we then have that (l, p) = (l, Xa (u; l)) is
well-shaped with respect to α if p ∈ dom(u). Well-shapedness with respect to γ follows
41
in both cases, i. e. (l, p) is well-c-shaped or (l, p) is well-π-shaped, by definition of wellshapedness. If p ∈ dom(v), then we have p = Xa (v; −∞), which is in +-form (and, thus,
in ±-form) by Lemma 10. By definition, we have the well-c-shapedness of (l, p) with
respect to γ.
If γ = (α)π for a π-term α, then define u = JαKω+ω∗ . We can write l = (l′ , m) and
p = (p′ , k) for some m, k ∈ N⊎−N and l′ , p′ ∈ dom(u). For k = m, we have p′ = Xa (u; l′ ).
This yields well-shapedness with respect to α of (l′ , p′ ) by induction. By definition, we
also have well-shapedness with respect to γ. For k = m + 1 (the only other possible
case), we know p′ = Xa (u; −∞) which is in +-form with respect to α by Lemma 10 and,
therefore, also in ±-form. If r = +∞, then (l, r) had to be well-c-shaped and l has to be
in ∓-form. Then, (l, p) is well-π-shaped. If r = (r ′ , n) for a n ∈ N ⊎ −N and r ′ ∈ dom(u),
then n must be greater than m (with respect to ω + ω ∗ ), because it must be greater than
or equal to k. Therefore, (l, r) has to be well-π-shaped and l has to be in ∓-form. Again,
(l, p) is well-π-shaped then, which concludes the proof that (l, p) is always well-shaped.
Next, consider (p, r). If r = +∞, then (l, r) has to be well-c-shaped. This implies that
l = −∞ or l is in ∓-form. In either case, we have that p = Xa (w; l) is in ∓-form by
Lemma 10. Thus, (p, +∞) is well-c-shaped. Again, we can safely assume that r 6= +∞
and proceed by induction on the structure of γ where the cases γ = ε and γ ∈ Σ do not
occur.
If γ = αβ for two π-terms α and β, then define u = JαKω+ω∗ and v = JβKω+ω∗ . The
case l ∈ dom(u) ⊎ {−∞} and r ∈ dom(u) and the case l, r ∈ dom(v) are similar to
the argumentation for (l, p). For l ∈ dom(u) ⊎ {−∞} and r ∈ dom(v), (l, r) has to
be well-c-shaped. Therefore, r has to be in ±-form with respect to β and (−∞, r) is
well-c-shaped with respect to β. Induction yields the well-shapedness with respect to β,
and, thus, also with respect to γ, of (p, r) if p ∈ dom(v). If p ∈ dom(u), we observe that
well-c-shapedness of (l, r) yields l = −∞ or l in ∓-form. In either case, p is in ∓-form
by Lemma 10. Thus, (p, r) is well-c-shaped with respect to γ.
If γ = (α)π for a π-term α, define u = JαKω+ω∗ . We can write p = (p′ , k) and r = (r ′ , n)
for k, n ∈ N ⊎ −N. For k = n, distinguish: if l = −∞, then (l, r) is well-c-shaped and r
has to be in ±-form with respect to γ. Therefore, r ′ has to be in ±-form with respect
to α. If l = (l′ , m) for l′ ∈ dom(u) and m ∈ N ⊎ −N with m <ω+ω∗ k = n, then (l, r)
is well-π-shaped and r ′ , again, is in ±-form. In both cases, this yields that (−∞, r ′ )
is well-c-shaped with respect to α. Induction yields that (p′ , r ′ ) = (Xa (u; −∞), r ′ ) is
well-shaped with respect to α. Since k = n, this implies that (p, r) is well-shaped. If
l = (l′ , m) but m = k = n, then (l′ , r ′ ) has to be well-shaped with respect to α and so
has to be (p′ , r ′ ) by induction, which again means that (p, r) is well-shaped with respect
to γ. If k <ω+ω∗ n and l = −∞, then (l, r) is well-c-shaped, r is in ±-form and p′ is in
+-form by Lemma 10. Therefore, (p, r) is well-π-shaped. If l = (l′ , m) for l′ ∈ dom(u)
and m ∈ N ⊎ −N and m ≤ω+ω∗ k <ω+ω∗ n, then l′ and, thus, also p′ by Lemma 10 has
to be in ∓-form while r ′ has to be in ±-form. This concludes the proof because, then,
(p, r) is well-π-shaped.
By the previous lemma, we know that there is maximally one switching between values
form −N to values from N in l below the branching position and that the same – but
42
in reverse – is true for r when we consider a pair (l, r) which arises from (−∞, +∞) by
applying a sequence of factorizations from FΣ . However, the algorithm described in the
proof of Theorem 5 performs a normalization after each factorization step. So far, we
have ignored this normalization but the next lemma states that normalization preserves
well-shapedness.
Lemma 12. Let (l, r) be a well-shaped, normalizable pair of positions in JγKω+ω∗ for a
γ
π-term γ over Σ. Then, (l, r) is well-shaped.
γ
γ
Proof. Let (l, r) = (¯l, r̄). In the special case l = −∞ and r = +∞, we have (l, r) = (l, r)
and we are done. For the other cases, we proceed by induction over γ.
γ
For γ = ε and γ ∈ Σ, we have that (l, r) is well-c-shaped. In the case γ = αβ for two
π-terms α and β, we distinguish: if l ∈ {−∞} ⊎ dom(JαKω+ω∗ ) and r ∈ dom(JβKω+ω∗ ) ⊎
{+∞}, we have that (l, r) must be well-c-shaped, that ¯l is given by the first component
α
β
of (l, +∞) and that r̄ is given by the second component of (−∞, r) . For l = −∞,
we also have ¯l = −∞ and, for r = +∞, we also have r̄ = +∞. If l 6= −∞, then l
must be in ∓-form. Note that normalization preserves the form of l when normalizing
(l, +∞), so ¯l is in ∓-from as well. Symmetrically, r̄ must be in ±-form if r 6= +∞. Thus,
γ
(l, r) is well-c-shaped. For l ∈ {−∞} ⊎ dom(JαKω+ω∗ ) and also r ∈ dom(JαKω+ω∗ ), we
γ
α
have (l, r) = (l, r) , which is well-shaped by induction. The same argument proves the
symmetric case l ∈ dom(JβKω+ω∗ ) and r ∈ dom(JβKω+ω∗ ) ⊎ {+∞}.
The remaining case is γ = (α)π for a π-term α. If we have l = −∞ or r = +∞, then
we know that l is in ∓-form or that r is in ±-form, respectively, because (l, r) has to
be well-c-shaped. As before, normalizing (−∞, r) or (l, +∞) preserves this form, which
makes the result well-c-shaped as well. Therefore, we have l = (l′ , n) and r = (r ′ , m) for
n, m ∈ N ⊎ −N and positions l′ , r ′ ∈ dom(JαKω+ω∗ ). For n = m, we have ¯l = (l¯′ , 1) and
α
α
r̄ = (r¯′ , 1) with (l¯′ , r¯′ ) = (l′ , r ′ ) . By induction, we have that (l′ , r ′ ) is well-c-shaped
or well-π-shaped with respect to α. In either case, (¯l, r̄) is also well-c-shaped or well-πshaped with respect to γ. For n 6= m, the pair (l, r) has to be well-π-shaped, l′ must be
in ∓-form and r ′ must be in ±-form. Additionally, we have ¯l = (l¯′ , n̄) and r̄ = (r¯′ , m̄)
for n̄, m̄ ∈ N ⊎ −N and positions l¯′ , r¯′ ∈ dom(JαKω+ω∗ ). Note that we also have n̄ 6= m̄
(in both of the cases which can occur when normalizing). Furthermore, normalization
of (l′ , +∞) (with respect to α) preserves the ∓-form of l′ and normalization of (−∞, r ′ )
γ
preserves the ±-form of r ′ , i. e. we have l¯′ in ∓-form and r̄ ′ in ±-form, which makes (l, r)
well-π-shaped.
Combining the previous two lemmas shows the following. Suppose we start with the
position pair (−∞, +∞) and apply a single factorization F1 ∈ FΣ , then we get (l1′ , r1′ ) =
(−∞, +∞) · F1 , which is well-shaped. If we then normalize (l1′ , r1′ ), the resulting pair
γ
(l1 , r1 ) = (l1′ , r1′ ) is also well-shaped. We can continue with another factorization and
normalize again; the result (l1 , r2 ) will still be well-shaped. Therefore, all the pairs of
positions (l, r) which are states in the deterministic finite automaton constructed in the
proof of Theorem 5 are well-shaped and normalized.
One last small observation is necessary before we discuss the details of storing those
pairs: by definition of the normalization, l and r may only have values from {1, −1} for
43
Name
ShapeType
BranchPosition
BranchValues
lSwitchPosition
rSwitchPosition
lEndPosition
rEndPosition
Values
{“well-c-shaped” , “well-π-shaped” }
a π-position or +∞
{(1, 2), (1, −1)}
a π-position or ⊥
a π-position or ⊥
a Σ-position or −∞
a Σ-position or +∞
Size
O(1)
O(log n)
O(1)
O(log n)
O(log n)
O(log n)
O(log n)
Table 1: The variables required to compute the values at π-positions and their sizes.
π-positions which are hierarchically lower than the branching position. Therefore, it is
sufficient to store the potential π-position at which the values switch (from −1 to 1 for
l or form 1 to −1 for r).
Storing a Well-Shaped, Normalized Pair. Summing up all of our observations results in
the situation which is schematically represented in Figure 8. Therefore, we can construct
values: (1, 1)
Branching, possible values:
values: −1
(1, 2) and (1, −1)
values: 1
Switching, value: 1
Switching, value: −1
values: −1
values: 1
r
l
Figure 8: Schematic representation of the possible values of a normalized pair (l, r)
the values of l and r for a normalized, well-shaped pair (l, r) of positions in JγKω+ω∗ at
all relevant π-positions in the π-term γ from the variables specified in Table 1.
We are going to explain the variables in more detail. Obviously, ShapeType gets the
value “well-c-shaped” if (l, r) is well-c-shaped and the value “well-π-shaped” if (l, r) is
well-π-shaped. For a well-c-shaped pair, we store the hierarchically lowest π-position for
which l and r share the same value in BranchPosition. If l and r differ already in the
hierarchically highest π-position, we store +∞ in BranchPosition. For well-π-shaped
pairs, we store the π-position at which l and r branch in BranchPosition and their
respective values there in BranchValues. The value of BranchValues is not relevant in
44
other cases. If there is a π-position at which the values of l switch from −1 to 1 (that
position is bound to be hierarchically lower than the branching position), then we store
the position of the first occurrence of 1 in lSwitchPosition. rSwitchPosition stores
the (potential) corresponding position for r. Finally, in lEndPosition and rEndPosition
we store the position in the π-term γ which corresponds to l or r, respectively.
One may verify that all these values can be stored in the size specified in the table,
where n is the length of the π-term γ seen as a finite word in Σ ⊎ { (, ), π }.
The variables are sufficient to compute the values of l and r at all relevant π-positions
deterministically within logarithmic space bounds. We only discuss how to do this for r,
as this version can easily be adapted for l. Assume we want to compute the value of r at a
π-position given in piPos. Notice that piPos is relevant for r if and only if rEndPosition
lies in between the opening and closing parenthesis belonging to piPos. Because finding
the matching left or right parenthesis is a simple matter of counting the opening and
closing parentheses this can be done deterministically in logarithmic space. Thus, we
may safely assume that our routine is only ever called for π-positions which are relevant
for r. To compute the value of r at piPos, we only have to know to which part of r it
belongs with respect to the schematic representation in Figure 8. The first part where
all values of r are equal to 1 consists of all π-positions which are hierarchically higher
than BranchPosition (assuming that it is not equal to ⊥), including BranchPosition if
ShapeType is “well-c-shape” or excluding if BranchPosition is “well-π-shaped”. Because
we already know that both, BranchPosition and piPos are relevant for r, checking
whether one is hierarchically higher than the other can be done by comparing their
positions in γ, which is possible in logarithmic space. The hierarchically higher one is to
the right of the lower one. If piPos belongs to the first part, we can return 1 immediately.
Next, we check whether BranchPosition is equal to piPos and ShapeType is “well-πshaped”. If this is the case, we can return the value 2 or −1 depending on the value of
BranchValues. The next part is from BranchPosition (excluding) to rSwitchPosition
(excluding). Again, we can check whether piPos belongs to this part by comparing the
position in the π-term and return 1 immediately. If piPos was not in any part so far, we
know that it belongs to the last part from rSwitchPosition (including) onwards and can
return −1. If BranchPosition is +∞ or rSwitchPosition is ⊥, then the corresponding
parts in Figure 8 simply do not exist and we can omit the respective checks.
Compute a Follow-Up Pair. Now that we know how to store the well-shaped, normalized pair (l, r) efficiently, we need to find a way to compute the normalized follow-up
pair if we apply an element from ZΣD = {XaD , YaD | a ∈ Σ, D ∈ {L, R}} (Ca,b needs to be
handled a bit differently; see below). By symmetry, we can restrict our considerations to
elements from XΣD , which means that we have to find the first a-position for an a ∈ Σ
on the right of the position l in JγKω+ω∗ for the π-term γ. For this, we use the additional
variables from Table 2.
We start by setting pBranchPosition to ⊥ and OpenParentheses to 0. CurrentPPosition gets assigned the next position to the right of the value of lEndPosition. The
algorithm now iteratively moves CurrentPPosition to the right one position at a time. If
45
Name
CurrentPPosition
pBranchPosition
OpenParentheses
Values
a position in γ
a π-position or +∞
a value to count open parentheses
Size
O(log n)
O(log n)
O(log n)
Table 2: The variables required to compute the follow-up pair and their sizes.
CurrentPPosition reaches an a-position, we are done. If CurrentPPosition reaches the
end of γ, then we know that there is no further a and we can stop. If CurrentPPosition
reaches an opening parenthesis, then we increment OpenParentheses. If CurrentPPosition reaches a closing parenthesis and pBranchPosition is ⊥, we distinguish: if OpenParentheses is > 0, then we decrement it and move on; if it is = 0, then we know that
the corresponding π-position is relevant for l. We check if the value of l at that position
is equal to −1. If that is the case, we can simply move on. If the value is from N or
< −1, then we set pBranchPosition to the π-position which corresponds to the closing
parenthesis and move CurrentPPosition to the matching opening parentheses using
OpenParentheses to count the number of opening and closing parentheses. From that
position we continue with the normal algorithm. If we then reach a closing parenthesis
and pBranchPosition is equal to CurrentPPosition, then we set pBranchPosition to ⊥
and OpenParenthesis to 0 and continue normally. Any other closing parenthesis which
is reached while pBranchPosition is not equal to ⊥ simply gets ignored.
Example 4. The only interesting part of the algorithm is that when a closing parenthesis
is reached. Therefore, we explain this part in more detail. Suppose we are in the following
situation9 and want to move to the next a:
a (1 (2 a )2
π2
b (3 c b (4 b )4
π4
)3
π3
2
l:
)1
π1
a (5 b )5
π5
c
−1
The algorithm would move CurrentPPosition to the right until it reaches (4 where OpenParentheses gets increased to 1. When )4 gets reached, OpenParentheses is set back
to 0, which triggers a different behavior when reaching )3 in the next step: pBranchPosition gets set to π3 and CurrentPPosition gets moved to (3 since l has the value
2 (and not −1) at π3 . From there, CurrentPPosition moves back to the right until it
reaches )3 because )4 simply gets ignored and because there is no a between (3 and )3 .
At )3 , OpenPrantheses and pBranchPosition are reset to 0 and ⊥. At the next closing
parenthesis )1 , the value of l gets checked again. But this time it is equal to −1 and we
continue to the right where we find the sought-after next a.
If l had had the value −2 at π1 , then the algorithm would have moved to (1 and the
next a would have been the one between (2 and )2 . In that case, pBranchPosition would
have been set to π1 when the final a is reached.
9
The numbers here do not have a special meaning other than to tell the parenthesis pairs and their
exponents apart from one another.
46
When the algorithm stops (and CurrentPPosition is not moved beyond the end of
γ) then CurrentPPosition points to the position in γ which corresponds to the next
a-position p in JγKω+ω∗ . The values at the π-positions can be reconstructed from the
stored variables. Again, this can be done by dividing the π-positions into parts for which
the value is well-known. If a π-position is relevant for p but not for l, then the value of
p at that position has to be 1.
Example 5. In the last example π2 , π4 and π5 are not relevant for l. If CurrentPPosition
would end up within the pair of parentheses belonging to any of these positions, its value
there would be 1.
If pBranchPosition has a value other than ⊥, then p has the value of l plus 1 there.
Note that the value of l cannot be −1 in that case by the definition of the algorithm. For
π-positions which are relevant for l and for p and which are hierarchically higher than
the position in pBranchPosition, we know that the values of l and p are equal. At those
positions which are hierarchically lower than pBranchPosition and which are relevant
for l and for p, the value of p has to be 1 just like in the first case. If pBranchPosition
is ⊥, then the values of p at positions relevant for l and for p are equal to the values of l.
Because – as said – it can be checked in logarithmic space whether a π-position is
relevant for l, p or r, this yields a deterministic algorithm with logarithmic space bounds
for calculating the values for p at all its relevant π-positions. It remains to test whether
p is still strictly smaller than r and to compute the new values of the variables in Table 1
after an XaD -step and normalization. This test can be done by walking through all πpositions in γ from the right to the left and checking whether they are relevant for p and
for r. If that is the case, we check their values. If the value of p is strictly smaller than
that of r, we are done. If they are equal everywhere, then CurrentPPosition needs to
be to the left of rEndPosition in γ.
For calculating the new values for the variables, a similar approach can be applied.
Consider the (slightly more difficult) case of XaR where p gets the new l. We walk through
the π-positions in γ from right to left and check their relevance for p and for r. If they
are relevant for one or the other, we can compute the values. As long as the values are
equal (which means that the position is relevant for both, p and r) we know that the
π-positions belong to the part in Figure 8 where l and r share the value 1. If there is a
π-position where the values are different, we can store that position in BranchPosition
and update BranchValues according to the value pair which normalization would yield.
In that case, we can set ShapeType to “well-π-shaped”. After the BranchPosition, we
have to check p and r individually for a change from values in N to values in −N or vice
versa and store the corresponding position. If the values at all positions have been equal
so far and we reach a position which is relevant for r but not for p (the other way round is
not possible because p has to be smaller than r), we can set ShapeType to “well-c-shaped”
and BranchPosition to the last π-position where the values were equal. ShapeType gets
also set to “well-c-shaped” if we reach rEndPos before there is a difference in the values.
Special Handling of Ca,b . By Theorem 3, we have to test whether JαKω+ω∗ ≡WI
m JβKω+ω ∗
holds in order to test whether α = β holds in Rm+1 ∩ Lm+1 . The definition of ≡WI
m ,
47
however, does not only rely on factorizations at the first or last a. It uses an additional
special factorization which behaves like first factorizing on the first a and then factorizing
on the last b but which is only possible if the first a is to the right of the last b. As defined
above, this kind of factorization is represented by Ca,b .
Clearly, (l, r) · Ca,b can only be defined if (l, r) · Xa Yb is defined. But the other direction
does not hold: it is possible that (l, r) · Xa Yb is defined while (l, r) · Ca,b is undefined.
We have described how we can compute the follow-up values of the variables in Table 1
for a factorization at the first a and, by symmetry, also for factorizations at the last b
when we start with a factor which is given by a normalized pair (l, r). We can combine
these two calculations into a single one for Ca,b . For this, we assume that we have two
instances of the variables in Table 2, one instance for the factorization at the first a and
one instance for the factorization at the last b. For both instances, we apply the normal
algorithm for the corresponding factorization. Afterwards, we check whether the position
of the first a is to the left of r and whether the position of the last b is to the right of
l just like we did before. For Ca,b , we simply add a third check which can be done in
a similar manner to the other checks: we check whether the position of the first a is to
the right of the position of the last b. If this check fails, then we know that (l, r) · Ca,b is
undefined, otherwise it indeed is defined.
Decidability in Nondeterministic Logarithmic Space. Let us recapitulate the proof for
decidability for the word problems for π-terms. First, we saw that the actual problem
we needed to decide was whether JαKω+ω∗ and JβKω+ω are equivalent with respect to the
relation belonging to the variety in question. To decide this, we constructed a deterministic finite automata Aα and Aβ for each of the input π-terms α and β. The automaton
Aα accepted exactly those factorization sequences F ∈ FΣ∗ for which JαKω+ω∗ · F (or
JβKω+ω∗ · F in the case of Aβ ) is defined. Additionally, we also constructed a deterministic finite automaton B which accepted those factorization sequences that need to be
tested for the respective relation. Finally, we created for Aα and for Aβ an automaton
accepting the intersection with the sequences accepted by B. For these automata, we
checked the symmetric difference for emptiness.
Checking the symmetric difference can be done in nondeterministic logarithmic space
by a naïve (iterative) guess and check algorithm. The automaton B is fixed for every
variety in the Trotter-Weil Hierarchy and for DA (or can be constructed deterministically
in space logarithmic in m if one considers this as an input). Creating an automaton for
the intersection can be done in deterministic logarithmic space. The only interesting
operation is the construction of Aα and Aβ . With the ideas from this section, we can
describe how to do this in deterministic logarithmic space.
Input: a π-term α
Output: Aα (which accepts a subset of FΣ∗ )
Algorithm:
for all possible values V of the variables in Table 1 do
for all F ∈ FΣ do
48
Compute the values V ′ of the variables for the follow-up pair after
application of F and normalization.
If there is such a follow-up pair, then output the transition
F
V
V′
end for
end for
Mark the state which represents (−∞, +∞) as initial state.
Mark any state as final.
In difference to our previous construction, the states in this automaton are not the
normalized pairs of positions anymore. Instead, we are representing them by their corresponding values for the variables in Table 1. Of course, not all possible values of the
variables represent a valid normalized pair, but, since the corresponding states cannot
be reached from the initial state, these errors do not affect the result.
These considerations allow us to state the following theorem.
Theorem 6. Each of the word problems for π-terms over Rm , Lm , Rm ∨ Lm , Rm ∩ Lm
and DA can be solved by a nondeterministic Turing machine in logarithmic space (for
every m ∈ N).
9 Deterministic Polynomial Time
While NL is quite efficient from a complexity class perspective, directly translating the
algorithm to polynomial time does not result in a better running time than the algorithm
for DA given by Moura [18]. However, with some additional tweaks, the algorithm’s
efficiency can be improved.
Encoding and Calculating Positions. Because we are not limited in space, we can
encode a position p ∈ dom(w) with w = JγKω+ω∗ for a π-term γ simply by storing the
values at the relevant π-positions and the corresponding position in γ. Clearly, we can
obtain the values of p at a given π-position in constant time. Additionally, we can store
a (possibly normalized) pair (l, r) by storing l and r. If we want to normalized such a
pair, we walk through all relevant π-positions and update the values of l and r there,
which requires at most linear time. Similarly, we can test whether a position is strictly
smaller (or larger) than an other position in linear time.
Suppose we have stored a position p ∈ dom(w) and want to compute Xa (w; p) for an
a ∈ Σ. We can re-use the algorithm which we used previously to solve the problem in
logarithmic space (see page 45). This algorithm moves a pointer, which belongs to a
position in γ, to the right in every step. The only time it is moved to the left is when it
hits a closing parenthesis and p has a value at the corresponding π-position which is not
−1. In that case, we move to the matching opening parenthesis and continue to move
to the right from there until we hit the closing parenthesis again. Note that there is no
49
“back-setting to the left” in that process. Therefore, we can compute the position of the
next a in at most quadratic time.
Computing the Automata and Equivalence Test. As we did before, we construct
a (not necessarily complete) deterministic finite automaton for each input π-term γ.
The automaton accepts exactly those sequences of factorizations which can be applied to
JγKω+ω∗ and which need to be tested for the variety in question. We will only demonstrate
the details of the construction for the variety Rm+1 ∩ Lm+1 and the input term α; the
construction for the other varieties is similar. We need the variables core and fringe,
which contain subsets of (dom(u) × dom(u)) × {1, 2, . . . , m} where u = JγKω+ω∗ such that
all pairs (l, r) ∈ dom(u) × dom(u) are normalized. The algorithm works as follows:
core ← ∅
fringe ← {((−∞, +∞), m)}
while fringe 6= ∅ do
Remove ((l, r), k) from fringe
if k > 0 then
for all a ∈ Σ do
Compute (l′ , r ′ ) = (l, r) · XaL if defined
⊲ time limited by O(n2 )
′
′
if (l , r ) is defined then
α
⊲ time limited by O(n)
(¯l′ , r̄ ′ ) ← (l′ , r ′ )
Save transition
(l, r)
k
XaL
(¯l′ , r̄ ′ )
k−1
Add ((¯l′ , r̄ ′ ), k − 1) to fringe unless it is in core
end if
Compute (l′ , r ′ ) = (l, r) · XaR if it is defined
if (l′ , r ′ ) is defined then
α
(¯l′ , r̄ ′ ) ← (l′ , r ′ )
Save transition
(l, r)
k
XaR
(¯l′ , r̄ ′ )
k
Add ((¯l′ , r̄ ′ ), k) to fringe unless it is in core
end if
Handle YaL and YaR analogously
for all b ∈ Σ do
⊲ Special case which is only required for ≡WI
m
′
′
Compute (l , r ) = (l, r) · Ca,b if it is defined
if (l′ , r ′ ) is defined then
α
(¯l′ , r̄ ′ ) ← (l′ , r ′ )
50
Save transition
(l, r)
k
Ca,b
(¯l′ , r̄ ′ )
k−1
Add ((¯l′ , r̄ ′ ), k − 1) to fringe unless it is in core
end if
end for
end for
Add ((l, r), k) to core
end if
end while
Set ((−∞, +∞), m) as initial state
Mark all states as final states
Clearly, the resulting automaton accepts the desired factorization sequences. We know
that any normalized pair can be encoded by the variables in Table 1. Therefore, the
number of such pairs is limited by O(n5 ) when n is the length of γ seen as a finite word.
This directly yields that the constructed automaton has at most O(n5 m) states. Because
the outer loop handles any ((l, r), k) at most once the algorithm’s running time is limited
by O(n7 m).
After the construction of the two automata for α and β, we need to test them for
equivalence. They are equivalent if and only if JαKω+ω∗ ≡WI
m JβKω+ω ∗ holds. The test for
equivalence can be done by the algorithm of Hopcroft and Karp [9] in almost linear time
in the size of the automata. Since the number of states is bounded by O(n5 m), the total
running time of our complete algorithm is dominated by O(n7 m2 ).
The automata for the varieties Rm , Lm and Rm ∨ Lm need to store whether the last
factor was obtained by an element from XΣD or by an element from YΣD . This information,
however, is only of constant size and, therefore, does not change the asymptotic running
time of the overall algorithm. For DA we can omit the counting for m in k by Fact 1,
which even reduces the number of states.
This finally shows the following theorem.
Theorem 7. The word problems for π-terms over Rm , Lm , Rm ∨ Lm and Rm ∩ Lm
can be solved by a deterministic algorithm with running time in O(n7 m2 ) where n is the
length of the input π-terms. Moreover, the word problem for π-terms over DA can be
solved by a deterministic algorithms in time O(n7 ).
10 Separability
Two languages L1 , L2 ⊆ Σ∗ are separable by a variety V if there is a language S ⊆ Σ∗
with L1 ⊆ S and L2 ∩ S = ∅ such that S can be recognized by a monoid M ∈ V. The
separation problem of a variety V is the problem to decide whether two regular input
languages of finite words are separable by V.
51
We are going to show the decidability of the separations problems of Rm for all m ∈ N
as well as for DA using the techniques presented in this paper10 . Note that, by symmetry,
this also shows decidability for Lm .
The general idea is as follows. If the input languages are separable, then we can find
a separating language S which is recognized by a monoid in the variety in question. We
can do this by recursively enumerating all monoids and all languages in a suitable representation. For the other direction, we show that, if the input languages are inseparable,
then there are π-terms α and β which witness their inseparability. Since we can also
recursively enumerate these π-terms, we have decidability.
To construct suitable π-terms we need an additional combinatorial property of the
≡X
m,n relations (which, in a slightly different form, can also be found in [15]).
Lemma 13. Let n, m ∈ N with m ≥ 2 and let u ≡X
m,n v for two accessible words u and
L
L
X
v. Then, u · Xa ≡m,n−1 v · Xa holds for all a ∈ alph(u) = alph(v).
Proof. We prove the lemma by induction over n. For n = 1, the assertion is satisfied by
X
L
definition. Therefore, assume we have u ≡X
m,n+1 v and we want to show u0 := u·Xa ≡m,n
Y
X
L
v · Xa =: v0 . We already have u0 ≡m−1,n v0 by definition of ≡m,n+1 . This especially
implies alph(u0 ) = alph(v0 ) since we have m ≥ 2 and n ≥ 1, as well as u0 ≡Ym−1,n−1 v0 .
Additionally, we have
u0 · XbL = u · XbL ≡Ym−1,n v · XbL = v0 · XbL
for all b ∈ alph(u0 ) = alph(v0 ), which implies u0 · XbL ≡Ym−1,n−1 v0 · XbL . All that remains
R
to be shown is that u0 ·XbR ≡X
m,n−1 v0 ·Xb holds for all b ∈ alph(u0 ) = alph(v0 ). Applying
R L
R
R L
X
induction on u · XbR ≡X
m,n v · Xb (for the same a) yields u · Xb Xa ≡m,n−1 v · Xb Xa .
R
L
R
R
L
R
L
R
R
Since we have u0 · Xb = u · Xa Xb = u · Xb Xa and v0 · Xb = v · Xa Xb = v · Xb XaL ,
we are done.
Using this property, one can prove the following lemma about the π-term construction.
Lemma 14. Let M be a monoid, ϕ : Σ∗ → M a homomorphism and m ∈ N0 . Let
(un , vn )n∈N0 be an infinite sequence of word pairs (un , vn )n∈N0 with
• u n , vn ∈ Σ ∗ ,
• ϕ(un ) = mu and
• un ≡ X
m,n vn ,
• ϕ(vn ) = mv
for fixed monoid elements mu , mv ∈ M and all n ∈ N0 . Then, the sequence yields π-terms
α and β (over Σ) such that ϕ (JαKM ! ) = mu , ϕ (JβKM ! ) = mv and JαKω+ω∗ ≡X
m JβKω+ω ∗
hold.
10
Decidability of the separation problem of DA is already known [21]. The proof, however, uses a fixed
point algorithm, which is different from our approach.
52
Before we give a proof of the general case, we give a separate one for the case m = 1. It is
basically an adaption of the ideas from the proof showing decidability of the separation
problem for the variety J of J -trivial monoids given by van Rooijen and Zeitoun [29] to
our setting.
Proof (m = 1). This proof is based on Simon’s Factorization Forest Theorem [25]. For
a finite word w ∈ Σ+ , a factorization tree is a rooted, finite, unranked, labeled ordered
tree such that
• the tree’s root is labeled with w,
• the leaves are labeled with letters (from Σ) and
• any internal node has at least two children and, if its children are labeled with
w1 , w2 , . . . , wk ∈ Σ+ , then the node is labeled with w1 w2 . . . wk .
For every homomorphism ψ : Σ∗ → N into a monoid N , Simon’s Factorization Forest
Theorem yields a factorization tree for every finite word w ∈ Σ+ such that ψ maps the
labels of a node’s children to the same idempotent in N if the node has at least three
children. Furthermore, the tree’s height11 is finite and limited by some constant that
solely depends on |N | (and, in particular, not on w).
Before we begin with the actual proof, we note that, if we remove pairs from the
sequence (un , vn )n∈N0 and still have an infinite sequence, then the resulting sequence still
satisfies all conditions stated above, in particular un ≡X
1,n vn .
∗
Σ
We extend ϕ into a homomorphism Σ → M ×2 which maps a word w to its alphabet
alph(w) for the second component.12 Then, we observe that there has to be an infinite
subsequence such that all first components as well as all second components have the
same alphabet. Indeed, these two alphabets have to coincide by the definition of ≡X
1,n !
We remove all other words from the sequence. If the remaining words un and vn are all
empty (i. e. they have alphabet ∅), we can choose α = β = ε as well. Otherwise, we apply
Simon’s Factorization Forest Theorem to the remaining words un and vn , which yields
a sequence of factorization tree pairs (Tu,n , Tv,n ). We first construct α from (Tu,n )n∈N0
such that we have ϕ(JαKM ! ) = mu and the following conditions:
• If w ∈ Σ∗ is a subword13 of un for an n ∈ N0 , then w is a subword of JαKω+ω∗ .
• If w ∈ Σ∗ is a subword of JαKω+ω∗ , then it is a subword of all un with n ≥ n0 for
an n0 ∈ N0 .
Afterwards, we proceed with (Tv,n )n∈N0 to construct β in the same manner.
We may assume that all trees Tu,n have the same height H as the height is bounded
by a constant and we can remove all words un from the underlying sequence which yield
a tree not of height H. If H is zero, all trees consist of a single leaf and all words un
11
A single node has height 0.
2 is the monoid of all subsets of Σ with taking union as the monoid’s operation.
13
Recall that a finite word u = a1 a2 . . . an with ai ∈ Σ is a subword of a (not necessarily finite) word v
if we can write v = v0 a1 v1 a2 v2 . . . an vn for some words v0 , v1 , . . . , vn .
12 Σ
53
consist of a single letter. Among these, one letter a ∈ Σ has to appear infinitely often;
we remove all other words from the sequence and choose α = a. Clearly, all conditions
for α are satisfied.
For H > 0, we consider the situation at the root of each Tu,n . Let un,1 , un,2 , . . . , un,Kn
be the labels of the root’s children in Tu,n . If the sequence (Kn )n∈N0 is bounded, there
is an infinite subsequence such that Kn is equal to a specific K ≥ 2 for all indexes n of
the subsequence; we remove all words not belonging to this subsequence. In the result,
there is an infinite subsequence such that, for each sequence (un,k )n∈N0 with 1 ≤ k ≤ K,
all un,k get mapped to the same monoid element by ϕ; we remove all other words. As
each child of the root yields a subtree, taking these subtrees gives K infinite sequences
of factorization trees of height H − 1. Applying induction on H yields α1 , α2 , . . . , αK .
We define α := α1 α2 . . . αK . Because α1 , α2 , . . . , αK satisfy the conditions stated above
for their respective subtree sequence, so does α for (Tu,n )n∈N0 .
If the sequence (Kn )n∈N0 is unbounded, we can, without loss of generality, assume
Kn ≥ 3 for all n ∈ N0 and that it is strictly increasing – again taking the appropriate
infinite subsequence. Also, we can assume that all un,1 , un,2 , . . . , un,Kn get mapped to the
same idempotent e ∈ M × 2Σ . Choose w ∈ ϕ−1 (e) arbitrarily and define α := (w)π . Note
that we now have alph(un,1 ) = alph(un,2 ) = · · · = alph(un,Kn ) = alph(un ) = alph(w) for
all n ∈ N0 . Therefore, α satisfies the conditions above.
All which remains to be shown is that we now have JαKω+ω∗ ≡X
1 JβKω+ω ∗ . The imw
with
n
∈
N
holds
if
and
only if w1 and w2
portant observation here is that w1 ≡X
2
0
1,n
have the same subwords of length ≤ n. This means we have to show that JαKω+ω∗ and
JβKω+ω∗ have the same subwords (of arbitrary length). To show the subword equality,
assume w is a subword of JαKω+ω∗ (without loss of generality). By the conditions above,
w is a subword of all un with n ≥ n0 for an n0 ∈ N0 . Let ñ = max{n0 , |w|}. Since we
X
have uñ ≡X
1,ñ vñ and by applying our observation regarding subwords and ≡1,ñ , w is a
subword of vñ and, thus, a subword of JβKω+ω∗ .
Next, we give a proof for the general case m ≥ 1.
Proof of Lemma 14. The assertion is trivial for m = 0. The case m = 1 has already
been covered. For m > 1, we proceed by induction over |Σ|. For Σ = ∅, we set α =
β = ε = un = vn . For |Σ| > 0, remember the observation from the previous proof: if we
take an infinite subsequence (u′n , vn′ )n∈N0 of (un , vn )n∈N0 , this sequence will still satisfy
′
all conditions of the lemma. In particular, we will still have u′n ≡X
m,n vn for all n ∈ N0 .
Now, we factorize un = wn,0 an,0 wn,1 an,1 . . . wn,Kn an,Kn wn,Kn +1 for all n ∈ N0 such that
alph(wn,k ) = alph(un ) \ {an,k } for all k ∈ {0, 1, . . . , Kn } and alph(wn,Kn +1 ) ( alph(un ).
If the sequence (Kn )n∈N0 is bounded, let K be one of the numbers which appear infinitely
often in it and restrict all further considerations to the corresponding subsequence of
words. If (Kn )n∈N0 is unbounded, let K = |M |2 + 1 and remove all word pairs (un , vn )
for which Kn is smaller than K from the sequence. For all k ∈ {0, 1, . . . , K}, a single
letter ak ∈ Σ has to appear infinitely often in the sequence (an,k )n∈N0 because Σ is of finite
size. We restrict our consideration to the corresponding subsequence. Then, we define
xn,k = un ·XaR0 XaR1 . . . XaRk−1 XaLk and yn,k = vn ·XaR0 XaR1 . . . XaRk−1 XaLk for k ∈ {0, 1, . . . , K}
54
as well as xn,K+1 = un · XaR0 XaR1 . . . XaRK and yn,K+1 = vn · XaR0 XaR1 . . . XaRK . We, thus,
have un = xn,0 a0 xn,1 a1 . . . xn,K aK xn,K+1 and vn = yn,0 a0 yn,1 a1 . . . yn,K aK yn,K+1 for all
n ∈ N0 . Because K is constant, we can safely assume that ϕ maps all elements of the
sequence (xn,k )n∈N0 (for every k ∈ {0, 1, . . . , K + 1}) to the same element sk ∈ M : one
element has to appear infinitely often and we take the corresponding subsequence. In
the same way, we can ensure that ϕ maps all element of (yn,k )n∈N0 to the same element
tk ∈ M (again, for all k ∈ {0, 1, . . . , K + 1}). By removing the first K + 2 pairs of words,
X
we can also ensure un ≡X
m,n+K+2 vn for all n ∈ N0 . This implies xn,k ≡m,n+K+2−k−1 yn,k
for all n ∈ N0 and all k ∈ {0, 1, . . . , K} by Lemma 13. Directly by the definition of ≡X
m,n ,
X
X
we already have xn,K+1 ≡m,n+K+2−K−1 yn,K+1 and, therefore, xn,k ≡m,n yn,k for all k ∈
{0, 1, . . . , K+1}. We can apply induction to (xn,k , yn,k )n∈N0 for k ∈ {0, 1, . . . , K} since we
have ak 6∈ alph(xn,k ) by construction. This yields π-terms α0 , α1 , . . . , αK , β0 , β1 , . . . , βK .
If (Kn )n∈N0 was bounded, then alph(xn,K+1 ) = alph(yn,K+1) ( alph(un ) = alph(vn )
holds and we can apply induction as well, which yields π-terms αK+1 and βK+1 . Setting
α = α0 a0 α1 a1 . . . αK aK αK+1 and β = β0 a0 β1 a1 . . . βK aK βK+1 satisfies JαKω+ω∗ ≡X
m
2 +1
is
a
congruence.
If
(K
)
was
unbounded,
we
set
K
=
|M
|
JβKω+ω∗ since ≡X
n
n∈N
m
0
and, by the pigeon hole principle, there are i, j ∈ {0, 1, . . . , K} with i < j and
s0 ϕ(a0 )s1 ϕ(a1 ) . . . si ϕ(ai ) = s0 ϕ(a0 )s1 ϕ(a1 ) . . . sj ϕ(aj ) and
t0 ϕ(a0 )t1 ϕ(a1 ) . . . ti ϕ(ai ) = t0 ϕ(a0 )t1 ϕ(a1 ) . . . tj ϕ(aj ).
We define
α = α0 a0 α1 a1 . . . αi ai (αi+1 ai+1 αi+2 ai+2 . . . αj aj )π αK+1 and
β = β0 a0 β1 a1 . . . βi ai (βi+1 ai+1 βi+2 ai+2 . . . βj aj )π βK+1
where αK+1 and βK+1 are obtained by using induction on m (and symmetry) for the sequences (xn,K+1 )n∈N0 and (yn,K+1)n∈N0 , i. e. we have ϕ(JαK+1 KM ! ) = sK+1 , ϕ(JβK+1 KM ! ) =
tK+1 and JαK+1 Kω+ω∗ ≡Ym−1 JβK+1 Kω+ω∗ . Therefore, we have ϕ (JαKM ! ) = mu and
ϕ (JβKM ! ) = mv by construction. We also have JαKω+ω∗ ≡X
m JβKω+ω ∗ : for the part left
π
up to and including the (·) , we have equivalence by induction and because ≡X
m is a
congruence; the right part, we cannot reach by arbitrarily many XaD factorizations since
all letters appear infinitely often in the (·)π part and, if we reach it by using at least one
YaD factorization, we are done since we have JαKω+ω∗ ≡Ym−1 JβKω+ω∗ .
We can now plug everything together and prove the following theorem.
Theorem 8. For each m ∈ N, the separation problem for Rm is decidable and so is the
one for Lm .
Proof. We only consider Rm as the case for Lm is symmetric. If the input languages
are separable, we can find a separating language by enumerating all candidates. If the
languages are inseparable, we have to apply the previous lemma. As regular languages,
the input languages L1 ⊆ Σ∗ and L2 ⊆ Σ∗ can be recognized by monoids M1 and M2
via the homomorphisms ϕ1 and ϕ2 and the homomorphism can be computed. Therefore,
55
they are also recognized by M := M1 × M2 via the homomorphism ϕ which maps a finite
word to a pair whose first component is determined by ϕ1 and whose second component
is determined by ϕ2 . Let n ∈ N0 be arbitrary. Since we have Σ∗ /≡X
m,n ∈ Rm and
since L1 and L2 cannot be separated by Rm , there have to be finite words un , vn ∈ Σ∗
with un ∈ L1 , vn ∈ L2 and un ≡X
m,n vn ; otherwise, we could construct a separating
language. The homomorphism ϕ has to map infinitely many elements of the sequence
(un , vn )n∈N0 to the same element in M since M is finite. If we remove all other elements,
we still have an infinite sequence (un , vn )n∈N0 with un ≡X
m,n vn for all n ∈ N0 which
also satisfies all conditions of Lemma 14. Therefore, there are π-terms α and β with
JαKω+ω∗ ≡X
m JβKω+ω ∗ , ϕ (JαKM ! ) ∈ ϕ(L1 ) and ϕ (JβKM ! ) ∈ ϕ(L2 ). Since we can test
whether JαKω+ω∗ ≡X
m JβKω+ω ∗ holds for any two π-terms α and β by Theorem 5, we can
also recursively enumerate all possible π-term pairs and check whether the conditions
above are met. We know that we can find such a pair if L1 and L2 are inseparable. On
the other hand, suppose L1 and L2 can be separated by S ⊆ Σ∗ which is recognized
by the monoid N ∈ Rm via a homomorphism ψ : Σ∗ → N and we have found a
pair α and β with JαKω+ω∗ ≡X
m JβKω+ω ∗ , ϕ (JαKM ! ) ∈ ϕ(L1 ) and ϕ (JβKM ! ) ∈ ϕ(L2 ).
Then, we have ϕ (JαKN !·M !) = ϕ (JαKM ! ) ∈ ϕ(L1 ) and, thus, JαKN !·M ! ∈ L1 as well as
JβKN !·M ! ∈ L2 (by a similar argument). Also, α = β holds in Rm by Theorem 4, which
implies s := ψ (JαKN !·M ! ) = ψ (JαKN ! ) = ψ (JβKN ! ) = ψ (JβKN !·M ! ). If we have s ∈ ψ(S),
then we have JβKN !·M ! ∈ S ∩ L2 ; otherwise, we have JαKN !·M ! ∈ L1 but JαKN !·M ! 6∈ S and,
thus, a contradiction in either case.
Since two languages are separable by Rm for m = |Σ| + 1 if they are separable by DA
[30], we also get decidability of the separation problem of DA, which has already been
shown by Place, van Rooijen and Zeitoun [21].
Corollary 3. The separation problem for DA is decidable.
References
[1] Jorge Almeida. Implicit operations on finite J -trivial semigroups and a conjecture
of I. Simon. Journal of Pure and Applied Algebra, 69(3):205 – 218, 1991.
[2] Jorge Almeida. Finite Semigroups and Universal Algebra. World Scientific, 1994.
[3] Jorge Almeida. Finite semigroups: an introduction to a unified theory of pseudovarieties. In G.M̃. dos Gomes Moreira da Cunha, Pedro Ventura Alves da Silva, and
Jean-Éric Pin, editors, Semigroups, Algorithms, Automata and Languages, pages
3–64. World Scientific, 2002.
[4] Jorge Almeida and Marc Zeitoun. An automata-theoretic approach to the word
problem for ω-terms over R. Theoretical Computer Science, 370(1):131–169, 2007.
[5] Volker Diekert, Paul Gastin, and Manfred Kufleitner. A survey on small fragments
of first-order logic over finite words. Int. J. Found. Comput. Sci., 19:513–548, 2008.
56
[6] Samuel Eilenberg. Automata, Languages, and Machines, volume B. Academic press,
1976.
[7] J. A. Gerhard and Mario Petrich. Varieties of bands revisited. Proceedings of the
London Mathematical Society, 58(3):323–350, 1989.
[8] Tom E. Hall and Pascal Weil. On radical congruence systems. Semigroup Forum,
59(1):56–73, 1999.
[9] John E. Hopcroft and Richard M. Karp. A linear algorithm for testing equivalence
of finite automata. Technical report, Cornell University, 1971.
[10] Martin Huschenbett and Manfred Kufleitner. Ehrenfeucht-Fraïssé games on omegaterms. In Ernst W. Mayr and Natacha Portier, editors, STACS 2014, Proceedings,
volume 25 of LIPIcs, pages 374–385, Dagstuhl, Germany, 2014. Dagstuhl Publishing.
[11] Kenneth Krohn, John Lewis Rhodes, and Bret Tilson. Homomorphisms and semilocal theory. In Michael A. Arbib, editor, Algebraic Theory of Machines, Languages,
and Semigroups, chapter 8, pages 191–231. Academic Press, New York and London,
1968.
[12] Manfred Kufleitner and Alexander Lauser. The join levels of the Trotter-Weil Hierarchy are decidable. In Branislav Rovan, Vladimiro Sassone, and Peter Widmayer,
editors, MFCS 2012, Proceedings, volume 7464 of LNCS, pages 603–614. Springer,
2012.
[13] Manfred Kufleitner and Pascal Weil. On the lattice of sub-pseudovarieties of DA.
Semigroup Forum, 81:243–254, 2010.
[14] Manfred Kufleitner and Pascal Weil. The FO2 alternation hierarchy is decidable. In
Patrick Cégielski and Arnaud Durand, editors, CSL 2012, Proceedings, volume 16
of LIPIcs, pages 426–439, Dagstuhl, Germany, 2012. Dagstuhl Publishing.
[15] Manfred Kufleitner and Pascal Weil. On logical hierarchies within FO2 -definable
languages. Logical Methods in Computer Science, 8(3):1–30, 2012.
[16] Kamal Lodaya, Paritosh K. Pandya, and Simoni S. Shah. Marking the chops: an
unambiguous temporal logic. In IFIP TCS 2008, Proceedings, IFIP, pages 461–476.
Springer, 2008.
[17] Jonathan P. McCammond. Normal forms for free aperiodic semigroups. Int. J.
Algebra Comput., 11(5):581–625, 2001.
[18] Ana Moura. The word problem for ω-terms over DA. Theoretical Computer Science,
412(46):6556–6569, 2011.
[19] Dominique Perrin and Jean-Éric Pin. Infinite words, volume 141 of Pure and Applied
Mathematics. Elsevier, Amsterdam, 2004.
57
[20] Jean-Éric Pin. Varieties of Formal Languages. North Oxford Academic Publishers
Ltd, 1986.
[21] Thomas Place, Lorijn van Rooijen, and Marc Zeitoun. Separating regular languages
by piecewise testable and unambiguous languages. In MFCS 2013, Proceedings,
pages 729–740. Springer, 2013.
[22] Marcel Paul Schützenberger. On finite monoids having only trivial subgroups. Information and Control, 8:190–194, 1965.
[23] Thomas Schwentick, Denis Thérien, and Heribert Vollmer. Partially-ordered twoway automata: A new characterization of DA. In Werner Kuich, Grzegorz Rozenberg, and Arto Salomaa, editors, DLT 2001, Proceedings, volume 2295 of LNCS,
pages 239–250. Springer, 2002.
[24] Imre Simon. Piecewise testable events. In Autom. Theor. Form. Lang., 2nd GI
Conf., volume 33 of LNCS, pages 214–222. Springer, 1975.
[25] Imre Simon. Factorization forests of finite height. Theoretical Computer Science,
72(1):65–94, 1990.
[26] Pascal Tesson and Denis Thérien. Diamonds are forever: The variety DA. In
Gracinda Maria dos Gomes Moreira da Cunha, Pedro Ventura Alves da Silva, and
Jean-Éric Pin, editors, Semigroups, Algorithms, Automata and Languages, pages
475–500. World Scientific, 2002.
[27] Denis Thérien and Thomas Wilke. Over words, two variables are as powerful as one
quantifier alternation. In Proceedings of the Thirtieth Annual ACM Symposium on
Theory of Computing, pages 234–240. ACM, 1998.
[28] Peter Trotter and Pascal Weil. The lattice of pseudovarieties of idempotent semigroups and a non-regular analogue. Algebra Universalis, 37(4):491–526, 1997.
[29] Lorijn van Rooijen and Marc Zeitoun. The separation problem for regular languages
by piecewise testable languages. CoRR, abs/1303.2143, 2013.
[30] Philipp Weis and Neil Immerman. Structure Theorem and Strict Alternation Hierarchy for FO2 on Words. Logical Methods in Computer Science, 5(3):1–23, 2009.
58
| 4math.GR
|
Scene recognition with CNNs: objects, scales and dataset bias
Luis Herranz, Shuqiang Jiang, Xiangyang Li
Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS)
Institute of Computer Technology, CAS, Beijing, 100190, China
arXiv:1801.06867v1 [cs.CV] 21 Jan 2018
{luis.herranz,shuqiang.jiang,xiangyang.li}@vipl.ict.ac.cn
Abstract
some recent works[4, 1, 21, 16] have shown that CNN features extracted locally in patches can be also aggregated into
effective scene representations. Often, these approaches
combine multiple scales, that are pooled using VLAD[4]
or Fisher vector (FV)[21] encoding. Dixit et al[1] suggested applying the pooling directly on the semantic representation, arguing that semantic representations are more
invariant. Recently, Wu et al[16] proposed an architecture
in which dense sampling of patches is replaced by region
proposals and discrimintive patch mining. In general, these
works use ImageNet-CNN to extract the local activations
instead of Places-CNN, since local patches are closer to objects than to scenes. However, a largely overlooked aspect
in this multi-scale scenario is the role of the scale and its
relation with the feature extractor (i.e. CNN). One limitation of current multi-scale approaches is the naive use of
CNNs by simply considering CNNs as general purpose feature extractors[10, 2]. Using the same fixed CNN model
for all the scales inevitably leads to dataset bias[13], since
the properties of the data vary at different scales, while the
feature extractor remains fixed.
Since scenes are composed in part of objects, accurate
recognition of scenes requires knowledge about both scenes
and objects. In this paper we address two related problems:
1) scale induced dataset bias in multi-scale convolutional
neural network (CNN) architectures, and 2) how to combine effectively scene-centric and object-centric knowledge
(i.e. Places and ImageNet) in CNNs. An earlier attempt,
Hybrid-CNN[23], showed that incorporating ImageNet did
not help much. Here we propose an alternative method taking the scale into account, resulting in significant recognition gains. By analyzing the response of ImageNet-CNNs
and Places-CNNs at different scales we find that both operate in different scale ranges, so using the same network
for all the scales induces dataset bias resulting in limited
performance. Thus, adapting the feature extractor to each
particular scale (i.e. scale-specific CNNs) is crucial to improve recognition, since the objects in the scenes have their
specific range of scales. Experimental results show that
the recognition accuracy highly depends on the scale, and
that simple yet carefully chosen multi-scale combinations
of ImageNet-CNNs and Places-CNNs, can push the stateof-the-art recognition accuracy in SUN397 up to 66.26%
(and even 70.17% with deeper architectures, comparable to
human performance).
Since objects are main components of scenes, knowledge about objects may be helpful in scene recognition.
Although, Places-CNN itself develops suitable object models at intermediate layers[22], the information in ImageNet
might be valuable. However, in a previous attempt, a network trained with the combined dataset ImageNet+Places
(Hybrid-CNN[23]) show that including ImageNet, far from
helpful was harmful. We will see how this problem is also
connected to scale-related dataset bias.
1. Introduction
State-of-the-art in visual recognition is based on the
successful combination of deep representations and massive datasets. Deep convolutional neural networks (CNNs)
trained on ImageNet (i.e. ImageNet-CNNs) achieve impressive performance in object recognition, while CNNs trained
on Places (Places-CNNs) do in scene recognition[23, 2].
However, CNNs also have limitations, such as the lack of
invariance to significant scaling. This problem is particularly important in scene recognition, due to a wider range
of scales and a larger amount of objects per image.
As an alternative to Places-CNN holistic representation,
In this paper we will study these two problems (i.e.
dataset bias in patch-based CNNs under different scaling
conditions, and how to effectively combine Places and ImageNet) and will see that they are related. Torralba and
Efros[13] studied the dataset bias as a cross-dataset generalization problem, in which the same classes may have
slightly different feature distributions in different datasets.
In our particular case, this bias in the feature distribution
is induced by scaling the image. If the scaling operation
is considerable, the characteristics of the data may change
1
• We show that using a single CNN as a generic feature extractor from patches is quite limited, due to the
dataset bias induced by scale changes. We show how
networks trained on different datasets are suitable for
different scale ranges. In particular, ImageNet-CNN
and Places-CNN have very different optimal ranges,
due to their object-centric and scene-centric natures.
• We evaluate two strategies to alleviate the dataset
bias by using scale-specific networks:
hybrid
Places/ImageNet architectures and fine tuning. By
combining after reducing the dataset bias, our method
is also a more effective way to combine Places and
ImageNet. Extensive experiments with different scale
combinations and hybrid variations (optionally fine
tuned) lead to some variations achieving state-of-theart performance in scene recognition.
2. Objects and scenes
2.1. Objects in object datasets and scene datasets
The knowledge learned by CNNs lies in the data seen
during training, and will be of limited use if tested in a different type of data. Thus, CNNs trained with ImageNet are
limited when used for scene recognition due to this training/test bias, while Places-CNNs perform better in this task.
While this is essentially true, objects and scenes are closely
related, so knowledge about objects can be still helpful to
recognize scenes, if used properly.
Understanding the characteristics of the datasets involved is essential to better explain the causes of dataset
bias. In our case, we want to analyze the properties of objects found in scene and object datasets. We focus on two
aspects related with the objects: scales and density.
To evaluate the dataset bias we use SUN397[18, 17]
as target dataset. Since Places contains scene data, with
205 scene categories overlapping with SUN397, and significantly more data, we can expect a low dataset bias. Thus
we focus on ImageNet (in particular ILSVRC2012), which
contains mostly object data. Fortunately, both ImageNet
and SUN have a fraction of images with region annotations and labels, so we can collect some relevant statistics and compare their distributions (we used the LabelMe
toolbox[14]). Since we will use this information to interpret
the variations in recognition accuracy in next experiements,
4.5
90
4
80
Fraction of images (%)
Fraction of objects (%)
completely, switching from scene data to object data. Understanding and quantifying this bias can help us to design better multi-scale architectures, and even better ways to
combine object and scene knowledge. In particular, we propose multi-scale architectures with scale-specific networks
as a principled way to address scale-related dataset bias and
combine scene and object knowledge (i.e. Places and ImageNet). In the next sections:
3.5
3
ILSVRC2012
SUN397
2.5
2
1.5
1
0.5
0
0
0.2
0.4
0.6
0.8
70
60
40
30
20
10
0
0
1
Normalized size
ILSVRC2012
SUN397
50
10
20
30
40
50
Objects in the scene
(a)
(b)
(c)
Figure 1. Characteristics of objects in ILSVRC2012 (object data)
and SUN397 (scene data): (a) distribution of objects sizes (normalized), (b) number of objects per scene, and (c) examples of
objects by increasing normalized size.
we focus on the 397 categories of the SUN397 benchmark
(rather than the 908 categories of the full SUN database).
Scale. Fig. 1a shows the distribution of object sizes, and
Fig. 1c some examples of objects of different normalized
sizes. We normalized the size of the object relative to the
equivalent training crop. While objects in ImageNet are
mostly large, often covering the whole image, objects in
SUN397 are much smaller, corresponding to the real distribution in scenes. Thus Fig. 1a shows an obvious mismatch
between both datasets.
Density. Fig. 1b shows the distribution of object annotations per scene image. We can observe that images in
ImageNet usually contain just one big object, while images
in SUN397 typically contain many small objects.
2.2. Dataset bias in object recognition
In order to study the behaviour of ImageNet-CNNs and
Places-CNNs in object recognition, we need object data
extracted from scenes datasets. We selected 100 images
per category from the 75 most frequent object categories
in SUN397, so we can have enough images to train SVM
classifiers. We took some precautions to avoid selecting too
small objects.
In contrast to most object and scene datasets, in this case
we have the segmentation of the object within the scene,
so we can use it to create variations over the same objects.
Thus we defined two scales:
• Original scale: the scale of the object in original scene.
• Canonical scale: the object is centered and rescaled to
fill the crop (keeping the aspect ratio). So in this case
its normalized size is 1.
Then we created four variations (see Fig. 2): original
masked, original with background, canonical masked and
canonical with background. In particular, to study the response to different scaling, the canonical variant is scaled
in the range 10%-100%. Note how scaling the variant with
background shifts progressively the content of the crop from
object to scene.
2.3. Scale sensitivity and object density
We trained a SVM classifier with 50 images per class,
and tested on the remaining 50 images. The input feature
was the output of the fc7 activation. The results are shown
in Fig. 3. We use two variants: objects masked and objects
with background (see Fig. 2). Regarding objects masked,
where the background is removed, we can see that in general the performance is optimal when the object is near full
size, above 70-80%. This is actually the most interesting region, with ImageNet-CNN performing slightly better than Places-CNN. This is interesting, since Places-CNN
was trained with scenes containing more similar objects to
the ones in the test set, while ImageNet-CNN was trained
with the less related categories found in ILSVRC2012 (e.g.
dogs, cats). However, as we saw in Fig. 1a, objects in
ILSVRC2012 cover a large portion of the image in contrast to smaller objects in SUN397, suggesting that a more
similar scale in the training data may be more important
than more similar object categories. As the object becomes
smaller, the performance of both models degrades similarly,
again showing a limited robustness to scale changes.
Focusing now on the objects with background variant,
the performance is worse than when the object is isolated
from the background. This behaviour suggests that the
background may introduce some noise in the feature and
lead to poorer performance. In the range close to full object
size, both ImageNet-CNN and Places-CNN have similar
performance. However, as the object becomes smaller, and
the content is more similar to scenes, Places-CNN has much
better performance than ImageNet-CNN, arguably due to
the fact it has learn contextual relations between objects and
global scene properties. In any case, scales with low accuracy are probably too noisy and not suitable for our purpose.
3. Multi-scale architecture with scale-specific
networks
3.1. Overview
For scene recognition we introduce our multi-scale architecture, which combines several networks that operate
in parallel over patches extracted from increasingly larger
versions of the input image. We use a standard multiscale architecture combining several AlexNet CNNs (CaffeNet[5] in practice) where 227x227 patches are extracted
from each full image. For faster processing, instead of ex-
tracting patches independently we use a fully convolutional
network. In contrast to recent works[4, 3, 21, 1], we adopt
simple max pooling to aggregate patch features into image
features.
The previous analysis and experimental results on object recognition evidence the limitations of using either
ImageNet-CNNs or Places-CNNs to deal with such a broad
range of scales, and will be confirmed in the next sections
by the experiments on scene data. For these reasons, we
propose a hybrid architecture introducing two simple, yet
crucial modifications in the architecture (discussed previously in Section 2.2).
• Instead of using naively the same CNN model for all
the scales, we select the most suitable one for each
(ImageNet-CNN, Places-CNN or fine tuned).
• Optionally we fine tune each CNN model to further
adapt it to the range of each scale. This requires resizing the image to target size and extracting patches for
training.
3.2. Differences with previous works
Our architecture is similar to others proposed in previous
multi-scale approaches[4, 21, 1], with the subtle difference
of using scale-specific networks in a principled way to alleviate the dataset bias induced by scaling. The main emphasis in these works is on the way multi-scale features are
combined, implemented as either VLAD or FV encoding,
while leaving the CNN model fixed. While adding a BOW
encoding layer can help to alleviate somewhat the dataset
bias, the main problem is still the rigid CNN model. In contrast, our method addresses better the dataset bias related
with scale and achieves significantly better performance, by
simply adapting the CNN model to the target scale, even
without relying to sophisticated pooling methods.
We can also regard our approach as a way to combine the
training data available in Places and ImageNet. This was explored previously by Zhou et al[23], who trained a HybridCNN using the AlexNet architecture and the combined
Places+ImageNet dataset. However, Hybrid-CNN performs
just slightly better than Places-CNN on MIT Indoor 67 and
worse on SUN397. We believe that the main reason was that
this way of combining data from ImageNet and Places ignores the fact that objects found in both datasets in two different scale ranges (as shown in Fig. 1). In contrast, our architecture combines the knowledge in a scale-adaptive way
via either ImageNet-CNN or Places-CNN. Wu et al[16] use
Hybrid-CNN on patches at different scales. Again, the main
limitation is that the CNN model is fixed, not adapting to the
scale-dependent distributions of patches.
(a)
(b)
Figure 2. The two variants used in the object recognition experiments: object masked (top row) and object with background (bottom row)
with two examples of (a) armchair and (b) streetlight. Left crops show the object in the original scale in the scene. Right crops show the
object scaled progressively from the canonical size (100%) down to 10%. All the images are centered in the object of interest.
55
scene benchmark (at least considered as such before Places)
containing 397 categories, including indoor, man-made and
natural categories. This dataset is very challenging, not only
because of the large number of categories, but also because
the more limited amount of training data (50 images per
category) and a much larger variability in objects and layout properties. It is widely accepted as the reference benchmark for scene recognition. We consider seven scales in our
experiments, obtained by scaling images between 227x227
and 1827x1827 pixels.
Accuracy (%)
50
45
40
35
30
objects masked (imnet)
objects masked (places)
objects w/ bg (imnet)
objects w/ bg (places)
25
20
0
0.2
0.4
0.6
0.8
4.2. Single scale
1
4.2.1
Accuracy
Normalized scale
Figure 3. Object recognition accuracy on SUN397 (75 categories).
4. Experiments on scene recognition
In this section we perform experiments directly over
scene data, to evaluate the relation beween scale, training
dataset and dataset bias by analyzing the scene recognition
performance. Then we combine and evaluate multi-scale
architectures.
4.1. Datasets
We evaluate the proposed architectures with three widely
used scene benchmarks. 15 scenes[6] is a small yet popular
dataset with 15 natural and indoor categories. Models are
trained with 100 images per category. MIT Indoor 67[9]
contains 67 categories of indoor images, with 80 images
per category available for training. Indoor scenes tend to
be rich in objects, which in general makes the task more
challenging, but also more amenable to architectures using
ImageNet-CNNs on patches. SUN397[18, 19] is a larger
Average accuracy is a reasonable metric to evaluate a deep
representation in the context of a classification task. For
the different scales, we extracted fc7 activations locally in
pacthes as features, and then trained SVMs. In addition to
the seven scales evaluated, we included 256x256 pixels as
a baseline, since off-the-shelf ImageNet-CNN and PlacesCNN are trained on this scale. The results for the three
datasets are shown in Fig. 5, with similar patterns. PlacesCNN achieves the best performance when is applied globally at scene level (227x227 or 256x256), while rapidly degrades for more local scales. ImageNet-CNN exhibits a
very different behaviour, with a more modest performance
at global scales, and achieving optimal performance on
patches at intermediate scales, and outperforming PlacesCNN at most local scales. These curves somewhat represent the operational curve of CNNs and the scale. In particular, the performance of ImageNet-CNN can be increases
notably just by using an appropriate scale.
An interesting observation is that there is one point
(around 643 or scale 0.35) that splits the range into two
parts, one dominated by ImageNet-CNN and another one
Figure 4. Multi-scale architecture combining scale-specific networks (spliced architecture). ImageNet-CNNs and Places-CNNs are combined according to the scale of the input patches. This can effectively alleviate the dataset bias by adapting test data to the underlying
training data. Intra-scale features are obtained using max pooling within each scale, and then concatenated into a single multi-scale feature.
dominated by Places-CNN, which we can loosely identify
as object range and scene range. We will use this observation later in Section 4.4 to design spliced architectures.
ever, the effectiveness is limited, since intermediate layers
remain biased to the (pre-)training data.
4.2.3
4.2.2
Effect of fine tuning
A popular way to deal with dataset bias in CNNs is fine
tuning, which basically continues training on a pretrained
model with the target dataset data. Similarly in our case, we
expect that fine tuning can modify somehow the weights
and thus adapt to the objects or at least the scales in the
target dataset. However, in practice that is often not possible because of the limited data, overfitting and difficulty
of setting the training process itself. In our case, fine tuning on scales where patches are very local is very difficult
due since the patch often contains objects or parts while labels indicates scenes categories. In addition, the number of
patches is huge, so only a tiny fraction of them can be used
in practice, rendering fine tuning not very effective.
We evaluated fine tuning on MIT Indoor 67. For scales
with few patches, and thus limited training data, we only
fine tune the fully connected layers. For larger images we
can collect more patches, up to 500K patches (randomly selected). Fig. 5b shows the results. Interestingly, there is
a moderate gain in those range of scales where the original
CNNs perform poorly, i.e. global scales for ImageNet-CNN
and local scales for Places-CNN, while marginal or no gain
in ranges where they have already good performance. Thus,
fine tuning has certain “equalizing” effect over the accuracy
vs scale curve. but limited overall improvement. In particular the gain is such that now Places-CNN (fine tuned) has
the best performance in the whole range of scales.
Fine tuning has impact mostly on the top layers, obtaining a similar effect to adding a BOW pooling layer. How-
Discriminability and redundancy
Accuracy provides a good indirect measure of the utility of
the feature for a given target task (e.g. scene recognition)
via a classifier (e.g. SVM). Here we also consider two information theoretic metrics measuring directly the discriminability and redundancy of the deep feature[8]. We define
the discriminability of a feature x = (x1,··· , x4096 ) with respect to a set of classes C = {1, · · · , M }
D (x, C) =
1 XX
I (xi ; c)
|C| |S|
x ∈x
c∈C
i
where I (xi ; c) is the filter xi and the class c. In order to
evaluate how redundant is the feature (compared with other
filters), we use the redundancy of a feature x, defined as
R (x) =
1
X X
2
|S|
I (xi ; xj )
xj ∈x xi ∈x
In the next experiment we compute D (x, C) and R (x)
of the fc7 activation for ImageNet-CNN and Places-CNN
in MIT Indoor 67. While we can find similarities with
the accuracy curve, a direct comparison is not easy, since
more discriminability not always means higher accuracy.
If we observe the discriminability of ImageNet-CNN (see
Fig. 6a), the curve follows a similar pattern to the accuracy,
with a peak around the scales where the accuracy was best,
and bad discriminability at global scales. Places-CNN extracts the most discriminative features at more global scales.
Comparing ImageNet-CNN and Places-CNN, the former
Accuracy (%)
64
1
3
6
3
45
25
32
0.6
90
85
imnet
places
80
75
0
0.2
0.4
0.8
0.6
1
Patch size (normalized)
18
1
9
89
3
3
1
64
32
45
25
6
imnet
places
imnet-ft
places-ft
55
50
0
0.2
0.4
0.8
0.6
1
Patch size (normalized)
(b)
27 283 99
8
1
18
Accuracy (%)
60
3
3
1
64
6
32
45
25
7
22
7
22
0.3
0.2
0
0.5
3
27 83 9
18 12 89 64
0.14
60
6
25
0.4
7
65
3
32
imnet
places
22
70
1
45
1
Patch size (normalized)
(a)
Mutual information
Accuracy (%)
75
3
28
3
64
0.5
0.1
(a)
27
27 83 9
18 12 89
7
22
Mutual information
27 83 99
18 12
8
95
1
45
0.12
3
32
6
25
7
22
imnet
places
0.1
0.08
0.06
0.04
0.02
0
0
0.5
1
Patch size (b)
(normalized)
55
Figure 6. Discriminability (a) and redundancy (b) of fc7 feature in
MIT Indoor 67.
50
obtains more discriminative features yet also more redundant. Too local scales (e.g. 1827x1827) increase significantly the redundancy of the feature and the noise
Note that this way of combining ImageNet and Places
data is much more effective than Hybrid-CNN[23] (see Table 1). Our dual architecture does not mix object and scene
knowledge (obtained from ImageNet and Places, respectively) and adapts the learned models to scales with similar properties. Dixit et al[1] combine Places-CNN with
a four-scales architecture built on top of ImageNet-CNN.
Similarly to our framework, Places-CNN operates at scene
scale while ImageNet-CNN operates at object scales. Note,
however, that we obtain comparable performance on MIT
Indoor 67 and significantly better on SUN397, using just
two networks instead of five.
4.3. Two scales
4.4. Multiple scales
In the next experiment we evaluated pairwise combinations of CNNs used at different scales. This dual architecture consists simply of two CNNs processing images at different scales. We then concatenate the two resulting fc7
activations into a 8192-dim feature and then train the SVM.
The results in Fig. 7 show that the dual architectures with
best performance are hybrid combinations of Places-CNN
extracting features at global scales (typically 227x227) with
ImageNet-CNN extracting features from patches at more
local scales. The result is a considerable boost in the performance, achieving a remarkable accuracy of 64.10% on
SUN397 using only two AlexNet CNNs.
Finally, we consider the combination of all the scales to
see whether more complex architectures could be helpful in
terms of accuracy. In this experiment we evaluate the concatenation of all the fc7 features of each of the seven scalespecific networks. In this case we use PCA to reduce the
dimension of each features vector so the combined dimension is approximately 4096. We achieve 74.33% (all scales
using ImageNet-CNN) and 78.21% (all scales using PlacesCNN) accuracy for MIT Indoor 67, and 58.71% and 63.81%
for SUN397, respectively. Note that both are better than the
corresponding dual architecture, yet below the corresponding dual hybrids (78.28% and 64.10%). This suggests than
45
40
35
imnet
places
0
0.2
0.4
0.6
0.8
1
Patch size (normalized)
(c)
Figure 5. Scene recognition accuracy for different scales: (a) 15
scenes, (b) MIT Indoor 67, and (c) SUN397.
places (79.10%-93.80%)
hybrid (82.81%-93.80%)
imnet (41.00%-56.62%)
places (35.65%-62.60%)
hybrid (45.09%-64.10%)
227
323
323
323
323
323
323
451
451
451
451
451
451
643
899
1283
643
899
1283
1827
643
899
1283
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
places
643
899
1283
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
imnet
899
1283
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
643
imnet
1827
hybrid (60.22%-78.28%)
imnet-ft (57.01%-72.46%)
hybrid-ft (60.60%-78.73%)
places-ft (56.04%-79.10%)
227
323
323
323
323
323
323
451
451
451
451
451
899
1283
1827
imnet
places
3 7
7 3 1 3 9
22 32 45 64 89 128 182
643
899
1283
1827
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
899
1283
1283
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
899
643
643
899
1283
1827
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
3 7
7 3 1 3 9
22 32 45 64 89 128 182
imnet-ft
imnet
451
places-ft
899
643
places-ft
227
imnet-ft
227
places
227
places
imnet
places (50.37%-76.87%)
1283
imnet
(b)
227
643
3 7
7 3 1 3 9
22 32 45 64 89 128 182
places
227
643
899
3 7
7 3 1 3 9
22 32 45 64 89 128 182
(a)
imnet (57.46%-71.87%)
643
1283
1827
3 7
7 3 1 3 9
22 32 45 64 89 128 182
imnet
places
227
places
227
imnet
227
places
227
places
imnet
imnet (79.60%-91.16%)
227
3 7
7 3 1 3 9
22 32 45 64 89 128 182
places-ft
imnet-ft
(c)
Figure 7. Accuracy in dual architectures combining two networks (only ImageNet-CNNs, only ImageNet-CNNs and hybrid combinations):
(a) 15 scenes, (b) SUN397 and, (c) MIT Indoor 67 (and fine tuned versions). Diagonals of only ImageNet-CNNs and only Places-CNNs
variations show single scale accuracy.
4.5. Deeper networks and other works
The experiments presented so far are based on the
AlexNet architecture.
Deeper architectures such as
GoogLeNet[12] and VGG-net[11] have demonstrated superior performance by exploiting deeper models. We repeated some of the experiments using the 16 layer VGG
architecture, obtaining state-of-the-art results in the three
datasets. The experiments with VGG in dual architectures
66
Accuracy (%)
including more scales while keeping the same CNN model
is marginally helpful and increases significantly the extraction cost and the noise in the representation.
So the key is to find an appropriate combination of
Places-CNNs and ImageNet-CNNs. While in dual architectures evaluating all the combinations is very costly, with
seven networks the combinations is impractical. Since the
optimal ranges of both are complementary, we can design
the full hybrid architecture as global scales using PlacesCNN and local scales using ImageNet-CNN, just as shown
in Fig. 4. We can consider only one free parameter which
is the splicing point. The results for SUN397 are shown
in Fig. 8. As expected, we can reach slightly better performance (80.97% and 65.38%) than in dual architectures.
The performance of hybrid spliced is also significantly better than a 7 network architecture with a fixed CNN model.
Finally we also evaluate a double full architecture, in
which both full ImageNet-CNN and full Places-CNN are
combined in a complex 14 CNNs architecture by concatenating the previous features. This combination does not
help in MIT Indoor 67, and slightly in SUN397, reaching
an accuracy of 66.26%.
64
places-imnet
imnet-places
places
imnet
62
60
58
23
227-3
51
323-4
3
4
451-6
99
643-8
283
899-1
7
-182
1283
same
Splicing point
Figure 8. Accuracy on SUN397 of full and hybrid spliced architectures (7 AlexNet networks). The combination same indicates
that the 7 networks share the same CNN model (i.e. trained with
the same dataset).
are consistent with those with AlexNet, but with a more
moderate gain. However, experiments combining more
networks were surprisingly disappointing, performing even
worse than single network baselines. VGG applied on small
patches tends to be very noisy with poor performance. We
tried an intermediate hybrid architecture, including a total
of three scales, achieving slightly better performance than
with dual architectures.
Overall, for the small 15 scenes dataset, it seems that
the performance is somewhat saturated, with a best performance of 95.18% (94.51% with AlexNet). The best performance in MIT Indoor 67 is 86.04% (compared with 80.97%
with AlexNet) and in SUN397 is 70.17% (compared with
66.26% with AlexNet). This performance is better than human recognition by “good workers” (68.5%), and close to
human expert performance (70.6%), as reported in [18].
Table 1. Accuracy for different multi-scale variations and architectures.
Architecture
Baseline1
Best single2
Dual
Dual hybrid
Three4
Full
Full hybrid (spliced)
Double full hybrid
Hybrid-CNN[23]
MOP-CNN[4]
MPP[21]
MPP+DSFL[21]
SFV[1]
SFV+Places[1]
MetaObject-CNN[16]
DAG-CNN[20]
DSP[3, 15]
Human (good)[17]
Human (expert)[17]
1
2
3
4
Pretraining dataset
#scales
IN
PL
IN
PL
IN
PL
IN/PL
IN/PL
IN
PL
IN/PL
IN/PL
IN+PL
IN
IN
IN
IN
IN/PL
Hybrid (IN+PL)[23]
IN
IN
1
1
1
1
2
2
1+1
1+2
7
7
7
2x7
1
3
7
7+DSFL
4
4+1
1 (variable)
1
1
15 scenes
Alex VGG
87.60 90.69
91.16 92.90
88.54 91.86
91.65 93.73
91.16 93.84
93.80 95.18
93.80 95.18
93.37 95.14
91.66 92.86
93.77 94.51
93.90 94.08
94.51 94.84
53.86
92.9
91.78
-
MIT Indoor 67 (w/ FT)
Alex
VGG
61.49
72.31
74.18
80.45
66.64 (68.21) 76.42
72.76 (73.35) 80.90
71.87 (72.46) 79.04
76.87 (79.40) 83.43
78.28 (78.81) 85.59
78.28
86.04
74.33 (75.97) 70.22
78.21 (79.70) 77.81
80.97 (80.75) 80.22
80.97 (79.85)
80.7
70.80
68.88
75.67
80.78
72.86
79.0
78.90
77.5
78.28
-
SUN 397
Alex VGG3
47.93 55.19
58.87 66.50
52.42 59.71
58.88 66.23
56.62 61.07
62.60 68.49
64.10 69.20
63.03 70.17
58.71 55.18
63.81 58.80
65.38 63.19
66.26 62.01
53.86
51.98
54.4
61.72
58.11
56.2
59.78
68.5%
70.6%
256x256 central crop (conventional settings for single crop).
Excluding 256x256.
Six scales (1827x1827 was not included).
Only evaluated the combination Places-CNN 227x227, Places-CNN 451x451, ImageNet-CNN 899x899.
5. Conclusions
In contrast to previous works, in this paper we analyzed
multi-scale CNN architectures focusing on the local CNN
model, rather than on the pooling method. In particular, we
showed that scaling images induces a bias between training
and test data, which has a significant impact on the recognition performance. We also showed how ImageNet-CNN
and Places-CNN in this context are implicitly tuned for different scale ranges (object scales and scene scales). Based
on these findings, we suggest that addressing this bias is
critical to improve scene recognition, and propose including scale-specific networks in the multi-scale architecture.
The proposed method is also a more principled way to combine scene-centric knowledge (Places) and object-centric
knowledge (ImageNet) than previous attempts (e.g. HybridCNN).
In fact, recent scene recognition approaches fall into two
apparently opposite directions: global holistic recognition
(Places-CNN) versus local object recognition and pooling
(multi-scale CNNs). In this paper we describe them as two
particular cases in a more general view of how multi-scale
features can be combined for scene recognition. They are
not incompatible, and actually when combined properly to
reduce the dataset bias the results can be excellent, even
reaching human recognition performance simply with just
two or three networks carefully chosen. Our hybrid parallel
architecture also suggests some similarities with perceptual
and cognitive models, where object recognition and global
scene features follow two distinct yet complementary
neural pathways which are later integrated to accurately
recognize the visual scene[7].
Acknowledgements. This work was supported in part
by the National Basic Research Program of China (973
Program): 2012CB316400, the National Natural Science
Foundation of China: 61532018, 61322212, 51550110505
and 61550110505, the National Hi-Tech Development
Program (863 Program) of China: 2014AA015202, and
the CAS President’s International Fellowship Initiative:
2011Y1GB05. This work is also funded by Lenovo Outstanding Young Scientists Program (LOYS).
References
[1] M. Dixit, S. Chen, D. Gao, N. Rasiwasia, and N. Vasconcelos. Scene classification with semantic fisher vectors.
In Computer Vision and Pattern Recognition (CVPR), June
2015. 1, 3, 6, 8
[2] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell. DeCAF: A deep convolutional activation feature for generic visual recognition. In International
Conference on Machine Learning (ICML), 2014. 1
[3] B. Gao, X. Wei, J. Wu, and W. Lin. Deep spatial pyramid:
The devil is once again in the details. arXiv, 1504.05277,
2015. 3, 8
[4] Y. Gong, L. Wang, R. Guo, and S. Lazebnik. Multi-scale
orderless pooling of deep convolutional activation features.
In European Conference on Computer Vision (ECCV), pages
392–407, 2014. 1, 3, 8
[5] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional
architecture for fast feature embedding. In ACM International Conference on Multimedia, pages 675–678. ACM,
2014. 3
[6] S. Lazebnik, C. Schmid, and J. Ponce. Beyond bags of
features: Spatial pyramid matching for recognizing natural
scene categories. In Computer Vision and Pattern Recognition (CVPR), 2006. 4
[7] A. Oliva. Scene perception. In J. S. Werner and L. M.
Chalupa, editors, The New Visual Neurosciences, pages 725–
732. The MIT Press, 2014. 8
[8] H. Peng, F. Long, and C. Ding. Feature selection based
on mutual information criteria of max-dependency, maxrelevance, and min-redundancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(8):1226–1238,
Aug 2005. 5
[9] A. Quattoni and A. Torralba. Recognizing indoor scenes.
In Computer Vision and Pattern Recognition (CVPR), pages
413–420, 2009. 4
[10] A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson. CNN features off-the-shelf: An astounding baseline for
recognition. In Computer Vision and Pattern Recognition
(CVPR) Workshops, pages 512–519, June 2014. 1
[11] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. In International
Conference on Learning Representations (ICLR), 2015. 7
[12] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Computer Vision and
Pattern Recognition (CVPR), pages 1–9, June 2015. 7
[13] A. Torralba and A. A. Efros. Unbiased look at dataset bias.
In Computer Vision and Pattern Recognition (CVPR), pages
1521–1528, June 2011. 1
[14] A. Torralba, B. C. Russell, and J. Yuen. Labelme: Online
image annotation and applications. Proceedings of the IEEE,
98(8):1467–1484, Aug. 2010. 2
[15] X. S. Wei, B. B. Gao, and J. Wu. Deep spatial pyramid ensemble for cultural event recognition. In International Conference on Computer Vision (ICCV) Workshops, pages 280–
286, Dec. 2015. 8
[16] R. Wu, B. Wang, W. Wang, and Y. Yu. Harvesting discriminative meta objects with deep CNN features for scene classification. In International Conference on Computer Vision
(ICCV), 2015. 1, 3, 8
[17] J. Xiao, K. Ehinger, J. Hays, A. Torralba, and A. Oliva. Sun
database: Exploring a large collection of scene categories.
International Journal of Computer Vision, pages 1–20, 2014.
2, 8
[18] J. Xiao, J. Hayes, K. Ehringer, A. Olivia, and A. Torralba.
SUN database: Largescale scene recognition from abbey to
zoo. In Computer Vision and Pattern Recognition (CVPR),
2010. 2, 4, 7
[19] Y. Xiao, J. Wu, and J. Yuan. mCENTRIST: A multichannel feature generation mechanism for scene categorization. IEEE Transactions on Image Processing, 23(2):823–
836, Feb 2014. 4
[20] S. Yang and D. Ramanan. Multi-scale recognition with
DAG-CNNs. In International Conference on Computer Vision (ICCV), pages 1215–1223, Dec. 2015. 8
[21] D. Yoo, S. Park, J.-Y. Lee, and I. So Kweon. Multi-scale
pyramid pooling for deep convolutional representation. In
Computer Vision and Pattern Recognition (CVPR) Workshops, June 2015. 1, 3, 8
[22] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba.
Object detectors emerge in deep scene cnns. In International
Conference on Learning Representations (ICLR), 2015. 1
[23] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva.
Learning deep features for scene recognition using places
database. In Neural Information Processing Systems (NIPS),
pages 487–495, 2014. 1, 3, 6, 8
| 1cs.CV
|
AttnGAN: Fine-Grained Text to Image Generation
with Attentional Generative Adversarial Networks
arXiv:1711.10485v1 [cs.CV] 28 Nov 2017
Tao Xu∗1 , Pengchuan Zhang2 , Qiuyuan Huang2 ,
Han Zhang3 , Zhe Gan4 , Xiaolei Huang1 , Xiaodong He2
1
Lehigh University
2
Microsoft Research
3
Rutgers University
4
Duke University
{tax313, xih206}@lehigh.edu, {penzhan, qihua, xiaohe}@microsoft.com
[email protected], [email protected]
Abstract
this bird is red with white and has a very short beak
In this paper, we propose an Attentional Generative Adversarial Network (AttnGAN) that allows attention-driven,
multi-stage refinement for fine-grained text-to-image generation. With a novel attentional generative network, the AttnGAN can synthesize fine-grained details at different subregions of the image by paying attentions to the relevant
words in the natural language description. In addition, a
deep attentional multimodal similarity model is proposed to
compute a fine-grained image-text matching loss for training the generator. The proposed AttnGAN significantly outperforms the previous state of the art, boosting the best reported inception score by 14.14% on the CUB dataset and
170.25% on the more challenging COCO dataset. A detailed analysis is also performed by visualizing the attention layers of the AttnGAN. It for the first time shows that
the layered attentional GAN is able to automatically select
the condition at the word level for generating different parts
of the image.
3:red
11:beak
9:very
8:a
3:red
5:white
1:bird
10:short
0:this
Figure 1. Example results of the proposed AttnGAN. The first row
gives the low-to-high resolution images generated by G0 , G1 and
G2 of the AttnGAN; the second and third row shows the top-5
most attended words by F1attn and F2attn of the AttnGAN, respectively. Here, images of G0 and G1 are bilinearly upsampled
to have the same size as that of G2 for better visualization.
1. Introduction
Automatically generating images according to natural
language descriptions is a fundamental problem in many
applications, such as art generation and computer-aided design. It also drives research progress in multimodal learning
and inference across vision and language, which is one of
the most active research areas in recent years [20, 18, 31,
19, 4, 29, 5, 1, 30]
Most recently proposed text-to-image synthesis methods
are based on Generative Adversarial Networks (GANs) [6].
A commonly used approach is to encode the whole text description into a global sentence vector as the condition for
GAN-based image generation [20, 18, 31, 32]. Although
∗ work
10:short
impressive results have been presented, conditioning GAN
only on the global sentence vector lacks important finegrained information at the word level, and prevents the generation of high quality images. This problem becomes even
more severe when generating complex scenes such as those
in the COCO dataset [14].
To address this issue, we propose an Attentional Generative Adversarial Network (AttnGAN) that allows attentiondriven, multi-stage refinement for fine-grained text-toimage generation. The overall architecture of the AttnGAN
is illustrated in Figure 2. The model consists of two novel
components. The first component is an attentional gener-
was performed when was an intern with Microsoft Research
1
ative network, in which an attention mechanism is developed for the generator to draw different sub-regions of the
image by focusing on words that are most relevant to the
sub-region being drawn (see Figure 1). More specifically,
besides encoding the natural language description into a
global sentence vector, each word in the sentence is also
encoded into a word vector. The generative network utilizes the global sentence vector to generate a low-resolution
image in the first stage. In the following stages, it uses
the image vector in each sub-region to query word vectors
by using an attention layer to form a word-context vector.
It then combines the regional image vector and the corresponding word-context vector to form a multimodal context
vector, based on which the model generates new image features in the surrounding sub-regions. This effectively yields
a higher resolution picture with more details at each stage.
The other component in the AttnGAN is a Deep Attentional
Multimodal Similarity Model (DAMSM). With an attention
mechanism, the DAMSM is able to compute the similarity
between the generated image and the sentence using both
the global sentence level information and the fine-grained
word level information. Thus, the DAMSM provides an additional fine-grained image-text matching loss for training
the generator.
The contribution of our method is threefold. (i) An
Attentional Generative Adversarial Network is proposed
for synthesizing images from text descriptions. Specifically, two novel components are proposed in the AttnGAN, including the attentional generative network and
the DAMSM. (ii) Comprehensive study is carried out to empirically evaluate the proposed AttnGAN. Experimental results show that the AttnGAN significantly outperforms previous state-of-the-art GAN models. (iii) A detailed analysis is performed through visualizing the attention layers of
the AttnGAN. For the first time, it is demonstrated that the
layered conditional GAN is able to automatically attend to
relevant words to form the condition for image generation.
2. Related Work
Generating high resolution images from text descriptions, though very challenging, is important for many practical applications such as art generation and computeraided design. Recently, great progress has been achieved
in this direction with the emergence of deep generative
models [12, 26, 6]. Mansimov et al. [15] built the alignDRAW model, extending the Deep Recurrent Attention
Writer (DRAW) [7] to iteratively draw image patches while
attending to the relevant words in the caption. Nguyen
et al. [16] proposed an approximate Langevin approach
to generate images from captions. Reed et al. [21] used
conditional PixelCNN [26] to synthesize images from text
with a multi-scale model structure. Compared with other
deep generative models, Generative Adversarial Networks
(GANs) [6] have shown great performance for generating
sharper samples [17, 3, 23, 13, 10]. Reed et al. [20] first
showed that the conditional GAN was capable of synthesizing plausible images from text descriptions. Their followup work [18] also demonstrated that GAN was able to generate better samples by incorporating additional conditions
(e.g., object locations). Zhang et al. [31, 32] stacked several GANs for text-to-image synthesis and used different
GANs to generate images of different sizes. However, all
of their GANs are conditioned on the global sentence vector, missing fine-grained word level information for image
generation.
The attention mechanism has recently become an integral part of sequence transduction models. It has been successfully used in modeling multi-level dependencies in image captioning [29], image question answering [30] and
machine translation [2]. Vaswani et al. [27] also demonstrated that machine translation models could achieve stateof-the-art results by solely using an attention model. In
spite of these progress, the attention mechanism has not
been explored in GANs for text-to-image synthesis yet. It is
worth mentioning that the alignDRAW [15] also used LAPGAN [3] to scale the image to a higher resolution. However, the GAN in their framework was only utilized as a
post-processing step without attention. To our knowledge,
the proposed AttnGAN for the first time develops an attention mechanism that enables GANs to generate fine-grained
high quality images via multi-level (e.g., word level and
sentence level) conditioning.
3. Attentional Generative Adversarial Network
As shown in Figure 2, the proposed Attentional Generative Adversarial Network (AttnGAN) has two novel components: the attentional generative network and the deep
attentional multimodal similarity model. We will elaborate
each of them in the rest of this section.
3.1. Attentional Generative Network
Current GAN-based models for text-to-image generation [20, 18, 31, 32] typically encode the whole-sentence
text description into a single vector as the condition for image generation, but lack fine-grained word level information. In this section, we propose a novel attention model
that enables the generative network to draw different subregions of the image conditioned on words that are most
relevant to those sub-regions.
As shown in Figure 2, the proposed attentional generative network has m generators (G0 , G1 , ..., Gm−1 ), which
take the hidden states (h0 , h1 , ..., hm−1 ) as input and generate images of small-to-large scales (x̂0 , x̂1 , ..., x̂m−1 ).
Residual
FC with reshape
Upsampling
Conv3x3
Attentional Generative Network
word
features
Local image
features
Attention models
sentence
feature
F1attn
F0
z~N(0,I)
Text
Encoder
Joining
Deep Attentional Multimodal Similarity Model (DAMSM)
F
F2
Image
Encoder
h2
h1
h0
ca
F2attn
F1
G2
c
256x256x3
G0
this bird is red with
white and has a
very short beak
G1
64x64x3
D0
128x128x3
D1
D2
Figure 2. The architecture of the proposed AttnGAN. Each attention model automatically retrieves the conditions (i.e., the most relevant
word vectors) for generating different sub-regions of the image; the DAMSM provides the fine-grained image-text matching loss for the
generative network.
Specifically,
h0 = F0 (z, F ca (e));
hi = Fi (hi−1 , Fiattn (e, hi−1 )) for i = 1, 2, ..., m − 1;
x̂i = Gi (hi ).
(1)
Here, z is a noise vector usually sampled from a standard
normal distribution. e is a global sentence vector, and e is
the matrix of word vectors. F ca represents the Conditioning
Augmentation [31] that converts the sentence vector e to the
conditioning vector. Fiattn is the proposed attention model
at the ith stage of the AttnGAN. F ca , Fiattn , Fi , and Gi are
modeled as neural networks.
The attention model F attn (e, h) has two inputs: the
word features e ∈ RD×T and the image features from the
previous hidden layer h ∈ RD̂×N . The word features are
first converted into the common semantic space of the image features by adding a new perceptron layer, i.e., e0 = U e,
where U ∈ RD̂×D . Then, a word-context vector is computed for each sub-region of the image based on its hidden
features h (query). Each column of h is a feature vector of
a sub-region of the image. For the j th sub-region, its wordcontext vector is a dynamic representation of word vectors
relevant to hj , which is calculated by
cj =
T
−1
X
i=0
βj,i e0i ,
exp(s0j,i )
where βj,i = PT −1
k=0
exp(s0j,k )
,
(2)
s0j,i = hTj e0i , and βj,i indicates the weight the model attends
to the ith word when generating the j th sub-region of the
image. We then donate the word-context matrix for image
feature set h by F attn (e, h) = (c0 , c1 , ..., cN −1 ) ∈ RD̂×N .
Finally, image features and the corresponding word-context
features are combined to generate images at the next stage.
To generate realistic images with multiple levels (i.e.,
sentence level and word level) of conditions, the final objective function of the attentional generative network is defined
as
L = LG + λLDAM SM , where LG =
m−1
X
LGi .
(3)
i=0
Here, λ is a hyperparameter to balance the two terms of
Eq. (3). The first term is the GAN loss that jointly approximates conditional and unconditional distributions [32]. At
the ith stage of the AttnGAN, the generator Gi has a corresponding discriminator Di . The adversarial loss for Gi is
defined as
1
1
LGi = − Ex̂i ∼pG [log(Di (x̂i )] − Ex̂i ∼pG [log(Di (x̂i , e)],
i
i
2
2
|
{z
} |
{z
}
unconditional loss
conditional loss
(4)
where the unconditional loss determines whether the image
is real or fake while the conditional loss determines whether
the image and the sentence match or not.
Alternately to the training of Gi , each discriminator Di
is trained to classify the input into the class of real or fake
by minimizing the cross-entropy loss defined by
1
1
LDi = − Exi ∼pdata [log Di (xi )] − Ex̂i ∼pG [log(1 − Di (x̂i )] +
i
i
2
2
|
{z
}
unconditional loss
1
1
− Exi ∼pdata [log Di (xi , e)] − Ex̂i ∼pG [log(1 − Di (x̂i , e)],
i
i
2
2
|
{z
}
conditional loss
(5)
where xi is from the true image distribution pdatai at the
ith scale, and x̂i is from the model distribution pGi at the
same scale. Discriminators of the AttnGAN are structurally
disjoint, so they can be trained in parallel and each of them
focuses on a single image scale.
The second term of Eq. (3), LDAM SM , is a word level
fine-grained image-text matching loss computed by the
DAMSM, which will be elaborated in Subsection 3.2.
3.2. Deep Attentional Multimodal Similarity Model
The DAMSM learns two neural networks that map subregions of the image and words of the sentence to a common
semantic space, thus measures the image-text similarity at
the word level to compute a fine-grained loss for image generation.
The text encoder is a bi-directional Long Short-Term
Memory (LSTM) [24] that extracts semantic vectors from
the text description. In the bi-directional LSTM, each word
corresponds to two hidden states, one for each direction.
Thus, we concatenate its two hidden states to represent the
semantic meaning of a word. The feature matrix of all
words is indicated by e ∈ RD×T . Its ith column ei is the
feature vector for the ith word. D is the dimension of the
word vector and T is the number of words. Meanwhile, the
last hidden states of the bi-directional LSTM are concatenated to be the global sentence vector, denoted by e ∈ RD .
The image encoder is a Convolutional Neural Network
(CNN) that maps images to semantic vectors. The intermediate layers of the CNN learn local features of different
sub-regions of the image, while the later layers learn global
features of the image. More specifically, our image encoder is built upon the Inception-v3 model [25] pretrained
on ImageNet [22]. We first rescale the input image to be
299×299 pixels. And then, we extract the local feature matrix f ∈ R768×289 (reshaped from 768×17×17) from the
“mixed 6e” layer of Inception-v3. Each column of f is the
feature vector of a sub-region of the image. 768 is the dimension of the local feature vector, and 289 is the number
of sub-regions in the image. Meanwhile, the global feature
vector f ∈ R2048 is extracted from the last average pooling
layer of Inception-v3. Finally, we convert the image features to a common semantic space of text features by adding
a perceptron layer:
v = Wf ,
D×289
v = W f,
th
(6)
where v ∈ R
and its i column vi is the visual feature vector for the ith sub-region of the image; and v ∈ RD
is the global vector for the whole image. D is the dimension
of the multimodal (i.e., image and text modalities) feature
space. For efficiency, all parameters in layers built from the
Inception-v3 model are fixed, and the parameters in newly
added layers are jointly learned with the rest of the network.
The attention-driven image-text matching score is
designed to measure the matching of an image-sentence pair
based on an attention model between the image and the text.
We first calculate the similarity matrix for all possible
pairs of words in the sentence and sub-regions in the image
by
s = eT v,
(7)
where s ∈ RT ×289 and si,j is the dot-product similarity
between the ith word of the sentence and the j th sub-region
of the image. We find that it is beneficial to normalize the
similarity matrix as follows
exp(si,j )
.
si,j = PT −1
k=0 exp(sk,j )
(8)
Then, we build an attention model to compute a regioncontext vector for each word (query). The region-context
vector ci is a dynamic representation of the image’s subregions related to the ith word of the sentence. It is computed as the weighted sum over all regional visual vectors,
i.e.,
ci =
288
X
exp(γ1 si,j )
.
αj vj , where αj = P288
k=0 exp(γ1 si,k )
j=0
(9)
Here, γ1 is a factor that determines how much attention is
paid to features of its relevant sub-regions when computing
the region-context vector for a word.
Finally, we define the relevance between the ith word
and the image using the cosine similarity between ci and ei ,
i.e., R(ci , ei ) = (cTi ei )/(||ci ||||ei ||). Inspired by the minimum classification error formulation in speech recognition
(see, e.g., [11, 8]), the attention-driven image-text matching score between the entire image (Q) and the whole text
description (D) is defined as
R(Q, D) = log
−1
TX
exp(γ2 R(ci , ei ))
γ1
2
,
(10)
i=1
where γ2 is a factor that determines how much to magnify the importance of the most relevant word-to-regioncontext pair. When γ2 → ∞, R(Q, D) approximates to
−1
maxTi=1
R(ci , ei ).
The DAMSM loss is designed to learn the attention
model in a semi-supervised manner, in which the only supervision is the matching between entire images and whole
sentences (a sequence of words). Similar to [4, 9], for a
batch of image-sentence pairs {(Qi , Di )}M
i=1 , the posterior
probability of sentence Di being matching with image Qi
is computed as
exp(γ3 R(Qi , Di ))
,
P (Di |Qi ) = PM
j=1 exp(γ3 R(Qi , Dj ))
(11)
where γ3 is a smoothing factor determined by experiments.
In this batch of sentences, only Di matches the image Qi ,
and treat all other M − 1 sentences as mismatching descriptions. Following [4, 9], we define the loss function as
the negative log posterior probability that the images are
matched with their corresponding text descriptions (ground
truth), i.e.,
M
X
Lw
=
−
log P (Di |Qi ),
(12)
1
i=1
where ‘w’ stands for “word”.
Symmetrically, we also minimize
Lw
2
=−
M
X
log P (Qi |Di ),
(13)
i=1
where P (Qi |Di ) =
exp(γ3 R(Qi ,Di ))
PM
j=1 exp(γ3 R(Qj ,Di ))
is the posterior
probability that sentence Di is matched with its corresponding image
Qi . If we redefine Eq. (10) by R(Q, D) =
v T e / ||v||||e|| and substitute it to Eq. (11), (12) and
(13), we can obtain loss functions Ls1 and Ls2 (where ‘s’
stands for “sentence”) using the sentence vector e and the
global image vector v.
Finally, the DAMSM loss is defined as
w
s
s
LDAM SM = Lw
1 + L2 + L1 + L2 .
(14)
Based on experiments on a held-out validation set, we set
the hyperparameters in this section as: γ1 = 5, γ2 = 5,
γ3 = 10 and M = 50. Our DAMSM is pretrained 1 by
minimizing LDAM SM using real image-text pairs. Since
the size of images for pretraining DAMSM is not limited
by the size of images that can be generated, real images of
size 299×299 are utilized. In addition, the pretrained textencoder in the DAMSM provides visually-discriminative
word vectors learned from image-text paired data for the
attentional generative network. In comparison, conventional word vectors pretrained on pure text data are often
not visually-discriminative, e.g., word vectors of different
colors, such as red, blue, yellow, etc., are often clustered
together in the vector space, due to the lack of grounding
them to the actual visual signals.
In sum, we propose two novel attention models, the attentional generative network and the DAMSM, which play
different roles in the AttnGAN. (i) The attention mechanism
in the generative network (see Eq. 2) enables the AttnGAN
to automatically select word level condition for generating
different sub-regions of the image. (ii) With an attention
mechanism (see Eq. 9), the DAMSM is able to compute
the fine-grained text-image matching loss LDAM SM . It is
worth mentioning that, LDAM SM is applied only on the
output of the last generator Gm−1 , because the eventual
goal of the AttnGAN is to generate large images by the last
1 We also finetuned the DAMSM with the whole network, however the
performance was not improved.
Dataset
#samples
caption/image
CUB [28]
train
test
8,855 2,933
10
10
COCO [14]
train test
80k
40k
5
5
Table 1. Statistics of datasets.
generator. We tried to apply LDAM SM on images of all
resolutions generated by (G0 , G1 , ..., Gm−1 ). However, the
performance was not improved but the computational cost
was increased.
4. Experiments
Extensive experimentation is carried out to evaluate the
proposed AttnGAN. We first study the important components of the AttnGAN, including the attentional generative network and the DAMSM. Then, we compare our AttnGAN with previous state-of-the-art GAN models for textto-image synthesis [31, 32, 20, 18, 16].
Datasets.
Same as previous text-to-image methods [31, 32, 20, 18], our method is evaluated on CUB [28]
and COCO [14] datasets. We preprocess the CUB dataset
according to the method in [31]. Table 1 lists the statistics
of datasets.
Evaluation. Following Zhang et al. [31], we use the
inception score [23] as the quantitative evaluation measure.
Since the inception score cannot reflect whether the generated image is well conditioned on the given text description,
we propose to use R-precision, a common evaluation metric for ranking retrieval results, as a complementary evaluation metric for the text-to-image synthesis task. If there
are R relevant documents for a query, we examine the top
R ranked retrieval results of a system, and find that r are
relevant, and then by definition, the R-precision is r/R.
More specifically, we conduct a retrieval experiment, i.e.,
we use generated images to query their corresponding text
descriptions. First, the image and text encoders learned in
our pretrained DAMSM are utilized to extract global feature
vectors of the generated images and the given text descriptions. And then, we compute cosine similarities between the
global image vectors and the global text vectors. Finally,
we rank candidate text descriptions for each image in descending similarity and find the top r relevant descriptions
for computing the R-precision. To compute the inception
score and the R-precision, each model generates 30,000 images from randomly selected unseen text descriptions. The
candidate text descriptions for each query image consist of
one ground truth (i.e., R = 1) and 99 randomly selected
mismatching descriptions.
Besides quantitative evaluation, we also qualitatively
examine the samples generated by our models. Specifically, we visualize the intermediate results with attention
learned by the attention models F attn . As defined in
Eq. (2), weights βj,i indicates which words the model at-
4.4
AttnGAN1,no DAMSM
AttnGAN1,𝞴=1
AttnGAN1,𝞴=10
70
4.2
60
4
3.8
AttnGAN1,no DAMSM
AttnGAN1,𝞴=0.1
AttnGAN1,𝞴=1
AttnGAN1,𝞴=5
AttnGAN1,𝞴=10
AttnGAN2,𝞴=5
3.6
3.4
50
150
250
350
Epoch
450
R-precision (%)
Inception score
AttnGAN1,𝞴=0.1
AttnGAN1,𝞴=5
AttnGAN2,𝞴=5
50
40
30
20
10
0
50
550
30
150
250
350
Epoch
450
550
100
90
80
70
20
R-precision(%)
Inception score
25
15
10
60
50
40
30
5
0
10
AttnGAN1,𝞴=0.1
AttnGAN1,𝞴=1
AttnGAN1,𝞴=10
AttnGAN1,𝞴=50
AttnGAN1,𝞴=100
AttnGAN2,𝞴=50
30
50
70 90
Epoch
20
110 130 150
10
0
10
AttnGAN1,𝞴=0.1
AttnGAN1,𝞴=1
AttnGAN1,𝞴=10
AttnGAN1,𝞴=50
AttnGAN1,𝞴=100
AttnGAN2,𝞴=50
30
50
70 90
Epoch
110 130 150
Figure 3. Inception scores and R-precision rates by our AttnGAN
and its variants at different epochs on CUB (top) and COCO (bottom) test sets. For the text-to-image synthesis task, R = 1.
Method
AttnGAN1, no DAMSM
AttnGAN1, λ = 0.1
AttnGAN1, λ = 1
AttnGAN1, λ = 5
AttnGAN1, λ = 10
AttnGAN2, λ = 5
AttnGAN2, λ = 50
(COCO)
inception score
3.98 ± .04
4.19 ± .06
4.35 ± .05
4.35 ± .04
4.29 ± .05
4.36 ± .03
R-precision(%)
10.37± 5.88
16.55± 4.83
34.96± 4.02
58.65± 5.41
63.87± 4.85
67.82 ± 4.43
25.89 ± .47
85.47 ± 3.69
Table 2. The best inception score and the corresponding Rprecision rate of each AttnGAN model on CUB (top six rows) and
COCO (the last row) test sets. More results in Figure 3.
tends
PT −1to when generating a sub-region of the image, and
i=0 βj,i = 1. We suppress the less-relevant words for an
image’s sub-region via
(
β̂j,i =
βj,i , if βj,i > 1/T,
0,
otherwise.
(15)
For better visualization, we fix the word and compute its attention weights with N different sub-regions of an image,
β̂0,i√, β̂1,i , √
..., β̂N −1,i . We reshape the N attention weights
to N × N pixels, which are then upsampled with Gaussian filters to have the same size as the generated images.
Limited by the length of the paper, we only visualize the
top-5
PN −1most attended words (i.e., words with top-5 highest
j=0 β̂j,i values) for each attention model.
4.1. Component analysis
In this section, we first quantitatively evaluate the AttnGAN and its variants. The results are shown in Table 2
and Figure 3. Our “AttnGAN1” architecture has one attention model and two generators, while the “AttnGAN2” architecture has two attention models stacked with three generators (see Figure 2). In addition, as illustrated in Figure 4,
Figure 5, Figure 6, and Figure 7, we qualitatively examine
the images generated by our AttnGAN.
The DAMSM loss. To test the proposed LDAM SM ,
we adjust the value of λ (see Eq. (3)). As shown in Figure 3, a larger λ leads to a significantly higher R-precision
rate on both CUB and COCO datasets. On the CUB dataset,
when the value of λ is increased from 0.1 to 5, the inception score of the AttnGAN1 is improved from 4.19 to 4.35
and the corresponding R-precision rate is increased from
16.55% to 58.65% (see Table 2). On the COCO dataset,
by increasing the value of λ from 0.1 to 50, the AttnGAN1
achieves both high inception score and R-precision rate (see
Figure 3). This comparison demonstrates that properly increasing the weight of LDAM SM helps to generate higher
quality images that are better conditioned on given text descriptions. The reason is that the proposed fine-grained
image-text matching loss LDAM SM provides additional supervision (i.e., word level matching information) for training the generator. Moreover, in our experiments, we do not
observe any collapsed nonsensical mode in the visualization
of AttnGAN-generated images. It indicates that, with extra
supervision, the fine-grained image-text matching loss also
helps to stabilize the training process of the AttnGAN. In
addition, if we replace the proposed DAMSM sub-network
with the text encoder used in [19], on the CUB dataset, the
inception score and R-precision drops to 3.98 and 10.37%,
respectively (i.e., the “AttnGAN1, no DAMSM” entry in table 2), which further demonstrates the effectiveness of the
proposed LDAM SM .
The attentional generative network. As shown in Table 2 and Figure 3, stacking two attention models in the
generative networks not only generates images of a higher
resolution (from 128×128 to 256×256 resolution), but also
yields higher inception scores on both CUB and COCO
datasets. In order to guarantee the image quality, we find
the best value of λ for each dataset by increasing the value
of λ until the overall inception score is starting to drop on
a held-out validation set. “AttnGAN1” models are built for
searching the best λ, based on which a “AttnGAN2” model
is built to generate higher resolution images. Due to GPU
memory constraints, we did not try the AttnGAN with three
attention models. As the result, our final model for CUB
and COCO is “AttnGAN2, λ=5” and “AttnGAN2, λ=50”,
respectively. The final λ of the COCO dataset turns out to
be much larger than that of the CUB dataset, indicating that
the proposed LDAM SM is especially important for generat-
the bird has a yellow crown and a black eyering that is round
this bird has a green crown black primaries and a white belly
1:bird
4:yellow
0:the
12:round
11:is
1:bird
0:this
2:has
11:belly
10:white
1:bird
4:yellow
0:the
8:black
12:round
6:black
4:green
10:white
0:this
1:bird
a photo of a homemade swirly pasta with broccoli carrots and onions
a fruit stand display with bananas and kiwi
0:a
7:with
5:swirly
8:broccoli
10:and
0:a
6:and
1:fruit
7:kiwi
5:bananas
8:broccoli
6:pasta
0:a
9:carrot
5:swirly
0:a
5:bananas
1:fruit
7:kiwi
6:and
Figure 4. Intermediate results of our AttnGAN on CUB (top) and COCO (bottom) test sets. In each block, the first row gives 64×64 images
by G0 , 128×128 images by G1 and 256×256 images by G2 of the AttnGAN; the second and third row shows the top-5 most attended
words by F1attn and F2attn of the AttnGAN, respectively. Refer to the supplementary material for more examples.
Dataset
CUB
COCO
GAN-INT-CLS [20]
2.88 ± .04
7.88 ± .07
GAWWN [18]
3.62 ± .07
/
StackGAN [31]
3.70 ± .04
8.45 ± .03
StackGAN-v2 [32]
3.82 ± .06
/
PPGN [16]
/
9.58 ± .21
Our AttnGAN
4.36 ± .03
25.89 ± .47
Table 3. Inception scores by state-of-the-art GAN models [20, 18, 31, 32, 16] and our AttnGAN on CUB and COCO test sets.
ing complex scenarios like those in the COCO dataset.
To better understand what has been learned by the AttnGAN, we visualize its intermediate results with attention.
As shown in Figure 4, the first stage of the AttnGAN (G0 )
just sketches the primitive shape and colors of objects and
generates low resolution images. Since only the global sentence vectors are utilized in this stage, the generated images
lack details described by exact words, e.g., the beak and
eyes of a bird. Based on word vectors, the following stages
(G1 and G2 ) learn to rectify defects in results of the previous stage and add more details to generate higher-resolution
images. Some sub-regions/pixels of G1 or G2 images can
be inferred directly from images generated by the previous
stage. For those sub-regions, the attention is equally allo-
this bird has wings that are black and has a white belly
this bird has wings that are red and has a yellow belly
this bird has wings that are blue and has a red belly
Figure 5. Example results of our AttnGAN model trained on CUB
while changing some most attended words in the text descriptions.
a fluffy black
cat floating on
top of a lake
a red double
decker bus
is floating on
top of a lake
a stop sign
is floating on
top of a lake
a stop sign
is flying in
the blue sky
Figure 6. 256×256 images generated from descriptions of novel
scenarios using the AttnGAN model trained on COCO. (Intermediate results are given in the supplementary material.)
observation is that our second attention model F2attn is able
to attend to some new words that were omitted by the first
attention model F1attn (see Figure 4). It demonstrates that,
to provide richer information for generating higher resolution images at latter stages of the AttnGAN, the corresponding attention models learn to recover objects and attributes
omitted at previous stages.
Generalization ability. Our experimental results above
have quantitatively and qualitatively shown the generalization ability of the AttnGAN by generating images from
unseen text descriptions. Here we further test how sensitive the outputs are to changes in the input sentences by
changing some most attended words in the text descriptions.
Some examples are shown in Figure 5. It illustrates that the
generated images are modified according to the changes in
the input sentences, showing that the model can catch subtle semantic differences in the text description. Moreover,
as shown in Figure 6, our AttnGAN can generate images to
reflect the semantic meaning of descriptions of novel scenarios that are not likely to happen in the real world, e.g.,
a stop sign is floating on top of a lake. On the other hand,
we also observe that the AttnGAN sometimes generates images which are sharp and detailed, but are not likely realistic. As examples shown in Figure 7, the AttnGAN creates
birds with multiple heads, eyes or tails, which only exist in
fairy tales. This indicates that our current method is still
not perfect in capturing global coherent structures, which
leaves room to improve. To sum up, observations shown
in Figure 5, Figure 6 and Figure 7 further demonstrate the
generalization ability of the AttnGAN.
4.2. Comparison with previous methods
Figure 7. Novel images by our AttnGAN on the CUB test set.
cated to all words and shown to be black in the attention
map (see Figure 4). For other sub-regions, which usually
have semantic meaning expressed in the text description
such as the attributes of objects, the attention is allocated to
their most relevant words (bright regions in Figure 4). Thus,
those regions are inferred from both word-context features
and previous image features of those regions. As shown in
Figure 4, on the CUB dataset, the words the, this, bird are
usually attended by the F attn models for locating the object; the words describing object attributes, such as colors
and parts of birds, are also attended for correcting defects
and drawing details. On the COCO dataset, we have similar
observations. Since there are usually more than one object in each COCO image, it is more visible that the words
describing different objects are attended by different subregions of the image, e.g., bananas, kiwi in the bottom-right
block of Figure 4. Those observations demonstrate that the
AttnGAN learns to understand the detailed semantic meaning expressed in the text description of an image. Another
We compare our AttnGAN with previous state-of-theart GAN models for text-to-image generation on CUB and
COCO test sets. As shown in Table 3, on the CUB dataset,
our AttnGAN achieves 4.36 inception score, which significantly outperforms the previous best inception score of
3.82. More impressively, our AttnGAN boosts the best
reported inception score on the COCO dataset from 9.58
to 25.89, a 170.25% improvement relatively. The COCO
dataset is known to be much more challenging than the
CUB dataset because it consists of images with more complex scenarios. Existing methods struggle in generating realistic high-resolution images on this dataset. Examples
in Figure 4 and Figure 6 illustrate that our AttnGAN succeeds in generating 256×256 images for various scenarios
on the COCO dataset, although those generated images of
the COCO dataset are not as photo-realistic as that of the
CUB dataset. The experimental results show that, compared to previous state-of-the-art approaches, the AttnGAN
is more effective for generating complex scenes due to its
novel attention mechanism that catches fine-grained word
level and sub-region level information in text-to-image generation.
5. Conclusions
In this paper, we propose an Attentional Generative Adversarial Network, named AttnGAN, for fine-grained textto-image synthesis. First, we build a novel attentional generative network for the AttnGAN to generate high quality image through a multi-stage process. Second, we propose a deep attentional multimodal similarity model to compute the fine-grained image-text matching loss for training the generator of the AttnGAN. Our AttnGAN significantly outperforms previous state-of-the-art GAN models,
boosting the best reported inception score by 14.14% on the
CUB dataset and 170.25% on the more challenging COCO
dataset. Extensive experimental results clearly demonstrate
the effectiveness of the proposed attention mechanism in
the AttnGAN, which is especially critical for text-to-image
generation for complex scenes.
References
[1] A. Agrawal, J. Lu, S. Antol, M. Mitchell, C. L. Zitnick, D. Parikh,
and D. Batra. VQA: visual question answering. IJCV, 123(1):4–31,
2017. 1
[2] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by
jointly learning to align and translate. arXiv:1409.0473, 2014. 2
[3] E. L. Denton, S. Chintala, A. Szlam, and R. Fergus. Deep generative
image models using a laplacian pyramid of adversarial networks. In
NIPS, 2015. 2
[4] H. Fang, S. Gupta, F. N. Iandola, R. K. Srivastava, L. Deng, P. Dollár,
J. Gao, X. He, M. Mitchell, J. C. Platt, C. L. Zitnick, and G. Zweig.
From captions to visual concepts and back. In CVPR, 2015. 1, 4, 5
[5] Z. Gan, C. Gan, X. He, Y. Pu, K. Tran, J. Gao, L. Carin, and L. Deng.
Semantic compositional networks for visual captioning. In CVPR,
2017. 1
[6] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. C. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014. 1, 2
[7] K. Gregor, I. Danihelka, A. Graves, D. J. Rezende, and D. Wierstra.
DRAW: A recurrent neural network for image generation. In ICML,
2015. 2
[8] X. He, L. Deng, and W. Chou. Discriminative learning in sequential
pattern recognition. IEEE Signal Processing Magazine, 25(5):14–36,
2008. 4
[9] P.-S. Huang, X. He, J. Gao, L. Deng, A. Acero, and L. Heck. Learning deep structured semantic models for web search using clickthrough data. In CIKM, 2013. 4, 5
[10] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2
[11] B.-H. Juang, W. Chou, and C.-H. Lee. Minimum classification error
rate methods for speech recognition. IEEE Transactions on Speech
and Audio Processing, 5(3):257–265, 1997. 4
[12] D. P. Kingma and M. Welling. Auto-encoding variational bayes. In
ICLR, 2014. 2
[13] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Aitken, A. Tejani,
J. Totz, Z. Wang, and W. Shi. Photo-realistic single image superresolution using a generative adversarial network. In CVPR, 2017.
2
[14] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollr, and C. L. Zitnick. Microsoft coco: Common objects in
context. In ECCV, 2014. 1, 5
[15] E. Mansimov, E. Parisotto, L. J. Ba, and R. Salakhutdinov. Generating images from captions with attention. In ICLR, 2016. 2
[16] A. Nguyen, J. Yosinski, Y. Bengio, A. Dosovitskiy, and J. Clune.
Plug & play generative networks: Conditional iterative generation of
images in latent space. In CVPR, 2017. 2, 5, 7
[17] A. Radford, L. Metz, and S. Chintala. Unsupervised representation
learning with deep convolutional generative adversarial networks. In
ICLR, 2016. 2
[18] S. Reed, Z. Akata, S. Mohan, S. Tenka, B. Schiele, and H. Lee.
Learning what and where to draw. In NIPS, 2016. 1, 2, 5, 7
[19] S. Reed, Z. Akata, B. Schiele, and H. Lee. Learning deep representations of fine-grained visual descriptions. In CVPR, 2016. 1, 6
[20] S. Reed, Z. Akata, X. Yan, L. Logeswaran, B. Schiele, and H. Lee.
Generative adversarial text-to-image synthesis. In ICML, 2016. 1, 2,
5, 7
[21] S. E. Reed, A. van den Oord, N. Kalchbrenner, S. G. Colmenarejo,
Z. Wang, Y. Chen, D. Belov, and N. de Freitas. Parallel multiscale
autoregressive density estimation. In ICML, 2017. 2
[22] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and
L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge.
IJCV, 115(3):211–252, 2015. 4
[23] T. Salimans, I. J. Goodfellow, W. Zaremba, V. Cheung, A. Radford,
and X. Chen. Improved techniques for training gans. In NIPS, 2016.
2, 5
[24] M. Schuster and K. K. Paliwal. Bidirectional recurrent neural networks. IEEE Trans. Signal Processing, 45(11):2673–2681, 1997. 4
[25] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architecture for computer vision. In CVPR,
2016. 4
[26] A. van den Oord, N. Kalchbrenner, O. Vinyals, L. Espeholt,
A. Graves, and K. Kavukcuoglu. Conditional image generation with
pixelcnn decoders. In NIPS, 2016. 2
[27] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N.
Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need.
arXiv:1706.03762, 2017. 2
[28] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The
Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR2011-001, California Institute of Technology, 2011. 5
[29] K. Xu, J. Ba, R. Kiros, K. Cho, A. C. Courville, R. Salakhutdinov,
R. S. Zemel, and Y. Bengio. Show, attend and tell: Neural image
caption generation with visual attention. In ICML, 2015. 1, 2
[30] Z. Yang, X. He, J. Gao, L. Deng, and A. J. Smola. Stacked attention
networks for image question answering. In CVPR, 2016. 1, 2
[31] H. Zhang, T. Xu, H. Li, S. Zhang, X. Wang, X. Huang, and
D. Metaxas. Stackgan: Text to photo-realistic image synthesis with
stacked generative adversarial networks. In ICCV, 2017. 1, 2, 3, 5, 7
[32] H. Zhang, T. Xu, H. Li, S. Zhang, X. Wang, X. Huang, and D. N.
Metaxas. Stackgan++: Realistic image synthesis with stacked generative adversarial networks. arXiv: 1710.10916, 2017. 1, 2, 3, 5,
7
Supplementary Material
Due to the size limit, more examples are available in the
appendix, which can be download from this link.
| 1cs.CV
|
Generating events with style
Matthieu Boutier1 and Gabriel Kerneis1,2
arXiv:1210.4263v2 [cs.PL] 17 Oct 2012
1
PPS, Université Paris Diderot, {first.last}@pps.univ-paris-diderot.fr
2
University of Cambridge
Abstract. Threads and events are two common abstractions for writing
concurrent programs. Because threads are often more convenient, but
events more efficient, it is natural to want to translate the former into
the latter. However, whereas there are many different event-driven styles,
existing translators often apply ad-hoc rules which do not reflect this
diversity.
We analyse various control-flow and data-flow encodings in real-world
event-driven code, and we observe that it is possible to generate any
of these styles automatically from threaded code, by applying certain
carefully chosen classical program transformations. In particular, we
implement two of these transformations, lambda lifting and environments,
in CPC, an extension of the C language for writing concurrent systems.
Finally, we find out that, although rarely used in real-world programs
because it is tedious to perform manually, lambda lifting yields better
performance than environments in most of our benchmarks.
Keywords: Concurrency, program transformations, event-driven style
1
Introduction
Most computer programs are concurrent programs, which need to perform several
tasks at the same time. For example, a network server needs to serve multiple
clients at a time; a GUI needs to handle multiple keyboard and mouse inputs;
and a network program with a graphical interface (e.g. a web browser) needs to
do both simultaneously.
Translating threads into events There are many different techniques to
implement concurrent programs. A very common abstraction is provided by
threads, or lightweight processes. In a threaded program, concurrent tasks are
executed by a number of independent threads which communicate through a
shared memory heap. An alternative to threads is event-driven programming.
An event-driven program interacts with its environment by reacting to a set of
stimuli called events. At any given point in time, to every event is associated a
piece of code known as the handler for this event. A global scheduler, known as
the event loop, repeatedly waits for an event to occur and invokes the associated
handler. Performing a complex task requires to coordinate several event handlers
by exchanging appropriate events.
2
Unlike threads, event handlers do not have an associated stack; event-driven
programs are therefore more lightweight and often faster than their threaded
counterparts. However, because it splits the flow of control into multiple tiny
event handlers, event-driven programming is generally deemed more difficult and
error-prone. Additionally, event-driven programming alone is often not powerful
enough, in particular when accessing blocking APIs or using multiple processor
cores; it is then necessary to write hybrid code, that uses both preemptivelyscheduled threads and cooperatively-scheduled event handlers, which is even more
difficult.
Since event-driven programming is more difficult but more efficient than
threaded programming, it is natural to want to at least partially automate
it. Continuation-Passing C (CPC [10]) is an extension of the C programming
language for writing concurrent systems. The CPC programmer manipulates
very lightweight threads, choosing whether they should be cooperatively or
preemptively scheduled at any given point. The CPC program is then processed
by the CPC translator, which produces highly efficient sequentialised event-loop
code, and uses native threads to execute the preemptive parts. The translation
from threads into events is performed by a series of classical source-to-source
program transformations: splitting of the control flow into mutually recursive
inner functions, lambda lifting of these functions created by the splitting pass,
and CPS conversion of the resulting code. This approach retains the best of
both worlds: the relative convenience of programming with threads, and the low
memory usage of event-loop code.
The many styles of events Not all event-driven programs look the same:
several styles and implementations exist, depending on the programmer’s taste.
Since event-driven programming consists in manually handling the control flow
and data flow of each task, a tedious and error-prone activity, the programmer
often choses a style based on some trade-off between (his intuition of) efficiency
and code-readability, and then sticks with it in the whole program. Even if
the representation of control or data turns out to be suboptimal, changing it
would generally require a complete refactoring of the program, not likely to be
undertaken for an uncertain performance gain. In large event-driven programs,
written by several people or over a long timespan, it is even possible to find a
mix of several styles making the code even harder to decipher.
For example, the transformations performed by the CPC translator yield
event-driven code where control flow is encoded as long, intricate chains of
callbacks, and where local state is stored in tiny data structures, repeatedly
copied from one event-handler to the next. We can afford these techniques
because we generate the code automatically. Hand-written programs often use
less tedious approaches, such as state machines to encode control flow and coarse
long-lived data structures to store local state; these are easier to understand
and debug but might be less efficient. Since the transformations performed by
the CPC translator are completely automated, it offers an ideal opportunity
3
to generate several event-driven variants of the same threaded program, and
compare their efficiency.
Contributions We first review existing translators from threads to events
(Section 2), and analyse several examples of event-driven styles found in realworld programs (Section 3). We identify two typical kinds of control-flow and
data-flow encodings: callbacks or state machines for the control flow, and coarsegrained or minimal data structures for the data flow.
We then propose a set of automatic program transformations to produce each
of these four variants (Section 4). Control flow is translated by splitting and CPS
conversion to produce callbacks; adding a pass of defunctionalisation yields state
machines. Data flow is translated either by lambda lifting, to produce minimal,
short-lived data structures, or using shared environments for coarse-grained ones.
Finally, we implement eCPC, a variant of CPC using shared environments
instead of lambda lifting to handle the data flow in the generated programs
(Section 5). We find out that, although rarely used in real-world event-driven
programs because it is tedious to perform manually, lambda lifting yields faster
code than environments in most of our benchmarks. To the best of our knowledge,
CPC is currently the only threads-to-events translator using lambda lifting.
2
Related work
The translation of threads into events has been rediscovered many times [5,11,12].
In this section, we review existing solutions, and observe that each of them generates only one particular kind of event-driven style. As we shall see in Section 4, we
believe that these implementations are in fact a few classical transformation techniques, studied extensively in the context of functional languages, and adapted to
imperative languages, sometimes unknowingly, by programmers trying to solve
the issue of writing events in a threaded style.
The first example known to us is Weave, an unpublished tool used at IBM
in the late 1990’s to write firmware and drivers for SSA-SCSI RAID storage
adapters [11]. It translates annotated Woven-C code, written in threaded style,
into C code hooked into the underlying event-driven kernel.
Adya et al. [1] provide a detailed analysis of control flow in threads and events
programs, and implement adaptors between event-driven and threaded code to
write hybrid programs mixing both styles.
Duff introduces a technique, known as Duff ’s device [4], to express general loop
unrolling directly in C, using the switch statement. Much later, this technique
has been employed multiple times to express state machines and event-driven
programs in a threaded style: protothreads [5], FairThreads’ automata [2]. These
libraries help keep a clearer flow of control but they provide no automatic handling
of data flow: the programmer is expected to save local variables manually in his
own data structures, just like in event-driven style.
Tame [12] is a C++ language extension and library which exposes events
to the programmer but does not impose event-driven style: it generates state
4
machines to avoid the stack ripping issue and retain a thread-like feeling. Similarly
to Weave, the programmer needs to annotate local variables that must be saved
across context switches.
TaskJava [6] implements the same idea as Tame, in Java, but preserves local
variables automatically, storing them in a state record. Kilim [17] is a messagepassing framework for Java providing actor-based, lightweight threads. It is also
implemented by a partial CPS conversion performed on annotated functions, but
contrary to TaskJava, it works at the JVM bytecode level.
MapJAX [13] is a conservative extension of Javascript for writing asynchronous
RPC, compiled to plain Javascript using some kind of ad-hoc splitting and CPS
conversion. Interestingly enough, the authors note that, in spite of Javascript’s
support for nested functions, they need to perform “function denesting” for
performance reasons; they store free variables in environments (“closure objects”)
rather than using lambda lifting.
AC [7] is a set of language constructs for composable asynchronous I/O
in C and C++. Harris et al. introduce do..finish and async operators to
write asynchronous requests in a synchronous style, and give an operational
semantics. The language constructs are somewhat similar to those of Tame but
the implementation is very different, using LLVM code blocks or macros based
on GCC’s nested functions rather than source-to-source transformations.
3
Control flow and data flow in event-driven code
Because event-driven programs do not use the native call stack to store return
addresses and local variables, they must encode the control flow and data flow
in data structures, the bookkeeping of which is the programmer’s responsibility.
This yields a diversity of styles among event-driven programs, depending on the
programmer’s taste, creativity, and his perception of efficiency. In this section,
we analyse how control flow and data flow are encoded in several examples of
real-world event-driven programs, and compare them to equivalent threaded-style
programs.
3.1
Control flow
Two main techniques are used to represent the control flow in event-driven
programming: callbacks and state machines.
Callbacks Most of the time, control flow is implemented with callbacks. Instead
of performing a blocking function call, the programmer calls a non-blocking
equivalent that cooperates with the event loop, providing a function pointer to be
called back once the non-blocking call is done. This callback function is actually
the continuation of the blocking operation.
Developing large programs raises the issue of composing event handlers.
Whereas threaded code has return addresses stored on the stack and a standard
calling sequence to coordinate the caller and the callee, event-driven code needs to
5
define its own strategy to layer callbacks, storing the callback to the next layer in
some data structure associated with the event handler. The “continuation stack”
of callbacks is often split in various places of the code, each callback encoding its
chunk of the stack in an ad-hoc manner.
Consider for instance the accept loop of an HTTP server that accepts clients
and starts two tasks for each of them: a client handler, and a timeout to disconnect
idle clients. With cooperative threads, this would be implemented as a mere
infinite loop with a cooperation point. The following code is an example of such
an accept loop written with CPC.
cps int cpc_accept ( int fd ) {
cpc_io_wait ( fd , CPC_IO_IN );
return accept ( fd , NULL , NULL );
}
cps int accept_loop ( int fd ) {
int client_fd ;
while (1) {
client_fd = cpc_accept ( fd );
cpc_spawn httpTimeout ( client_fd , clientTimeout );
cpc_spawn httpClientHandler ( client_fd );
}
}
The programmer calls cpc spawn accept loop(fd) to create a new thread that
runs the accept loop; the function accept loop then waits for incoming connections with the cooperating primitive cpc io wait, and creates two new threads
for each client (httpTimeout and httpClientHandler), which kill each other
upon completion. Note that cooperative functions are annotated with the cps
keyword; such cps functions are to be converted into event-driven style by the
CPC translator.
Figure 1 shows the (very simplified) code of the accepting loop in Polipo,
a caching web-proxy written by Chroboczek.3 This code is equivalent to the
threaded version above, and uses several levels of callbacks.
In Polipo, the accept loop is started by a call to schedule accept(fd, httpAccept, NULL). This function stores the pointer to the (second-level) callback
httpAccept in the handler field of the request data structure (line 10), and
registers a (first-level) callback to do scheduled accept, through registerFdEvent. Each time the file descriptor fd becomes ready (not shown), the event
loop calls the (first-level) callback do scheduled accept, which performs the
actual accept system call (line 23) and finally invokes the (second-level) callback
httpAccept stored in request->handler (line 24).
This callback schedules two new event handlers, httpTimeout and httpClientHandler. The former is a timeout handler, registered by scheduleTimeEvent (line 35); the latter reacts I/O events to read requests from the client, and
is registered by do stream buf (line 41). Note that those helper functions that
register callbacks with the event loop use other intermediary callbacks themselves,
just like schedule accept uses do schedule accept.
3
http://www.pps.univ-paris-diderot.fr/~jch/software/polipo/.
6
2
4
6
FdEventHandlerPtr
schedule_accept ( int fd ,
int (* handler )( int , FdEventHandlerPtr , AcceptRequestPtr ) ,
void * data ) {
FdEventHandlerPtr event ;
AcceptRequestRec request ;
int done ;
8
request . fd = fd ;
request . handler = handler ;
request . data = data ;
event = registerFdEvent ( fd , POLLOUT | POLLIN ,
do_scheduled_accept ,
sizeof ( request ) , & request );
return event ;
10
12
14
16
}
18
int
do_scheduled_accept ( int status , FdEventHandlerPtr event ) {
AcceptRequestPtr request = ( AcceptRequestPtr )& event - > data ;
int rc , done ;
20
22
rc = accept ( request - > fd , NULL , NULL );
done = request - > handler ( rc , event , request );
return done ;
24
26
}
28
int
httpAccept ( int fd , FdEventHandlerPtr event ,
AcceptRequestPtr request ) {
HTTPConnectionPtr connection ;
TimeEventHandlerPtr timeout ;
30
32
connection = httpMakeConnection ();
timeout = scheduleTimeEvent ( clientTimeout ,
httpTimeoutHandler ,
sizeof ( connection ) , & connection );
connection - > fd = fd ;
connection - > timeout = timeout ;
connection - > flags = CONN_READER ;
do_stream_buf ( IO_READ | IO_NOTNOW ,
connection - > fd , 0 , & connection - > reqbuf ,
CHUNK_SIZE , httpClientHandler , connection );
return 0;
34
36
38
40
42
44
}
Fig. 1. Accept loop callbacks in Polipo (simplified)
In the original Polipo code, things are even more complex since schedule
accept is called from httpAcceptAgain, yet another callback that is registered
by httpAccept itself in some error cases. The control flow becomes very hard to
follow, in particular when errors are triggered: each callback must be prepared
to cope with error codes, or to follow-up the unexpected value to the next layer.
In some parts of the code, this style looks a lot like an error monad manually
interleaved with a continuation monad. Without a strict discipline and well-
7
defined conventions about composition, the flexibility of callbacks easily traps
the programmer in a control-flow and storage-allocation maze.
State machines When the multiplication of callbacks becomes unbearable, the
event-loop programmer might refactor his code to use a state machine. Instead
of splitting a computation into as many callbacks as it has atomic steps, the
programmer registers a single callback that will be called over and over until the
computation is done. This callback implements a state machine: it stores the
current state of the computation into an ad-hoc data structure, just like threaded
code would store the program counter, and uses it upon resuming to jump to the
appropriate location.
Figure 2 shows how the initial handshake of a BitTorrent connection is handled
in Transmission,4 a popular and efficient BitTorrent client written in (mostly)
event-driven style. Until the handshake is over, all data arriving from a peer is
handed over by the event loop to the canRead callback. This function implements
a state machine, whose state is stored in the state field of a handshake data
structure. This field is initialised to AWAITING HANDSHAKE when the connection is
established (not shown) and updated by the functions responsible for each step
of the handshake.
The first part of the handshake is dispatched by canRead to the readHandshake function (line 7). It receives the buffer inbuf containing the bytes received
so far; if not enough data has yet been received to carry on the handshake, it
returns READ LATER to canRead (line 26), which forwards it to the event loop to
be called back when more data is available (line 16). Otherwise, it checks the
BitTorrent header (line 28), parses the first part of the handshake, registers a
callback to send a reply handshake (not shown), and finally updates the state
(line 33) and returns READ NOW to indicate that the rest of the handshake should
be processed immediately (line 34).
Note what happens when the BitTorrent header is wrong (line 28): the function
tr handshakeDone is called with false as its second parameter, indicating that
some error occurred. This function (not shown) is responsible for invoking the
callback handshake->doneCB and then deallocating the handshake structure.
This is another example of the multiple layers of callbacks mentioned above.
If the first part of the handshake completes without error, canRead then
dispatches the buffer to readPeerId which completes the handshake (line 10). Just
like readHandshake, it returns READ LATER if the second part of the handshake
has not arrived yet (line 41) and finally calls tr handshakeDone with true to
indicate that the handshake has been successfully completed (line 45).
In the original code, ten additional states are used to deal with the various steps
of negotiating encryption keys. The last of these steps finally rolls back the state to
AWAITING HANDSHAKE and the keys are used by the function tr peerIoReadBytes
to decrypt the rest of the exchange transparently. The state machine approach
makes the code slightly more readable than using pure callbacks.
4
http://www.transmissionbt.com/.
8
1
3
static ReadState
canRead ( struct evbuffer * inbuf , tr_handshake * handshake ) {
ReadState ret = READ_NOW ;
while ( ret == READ_NOW ) {
switch ( handshake - > state ) {
case AWAITING_HANDSHAKE :
ret = readHandshake ( handshake , inbuf );
break ;
case AWAITING_PEER_ID :
ret = readPeerId
( handshake , inbuf );
break ;
/* ... cases dealing with encryption omitted */
}
}
return ret ;
5
7
9
11
13
15
17
}
19
static int
readHandshake ( tr_handshake * handshake ,
struct evbuffer * inbuf ) {
uint8_t pstr [20] , reserved [ HANDSHAKE_FLAGS_LEN ] ,
hash [ SHA_DIGEST_LENGTH ];
21
23
if ( evbuffer_get_length ( inbuf ) < I NC OMI NG _H AN DSH AK E_ LEN )
return READ_LATER ;
tr_peerIoReadBytes ( handshake - > io , inbuf , pstr , 20);
if ( memcmp ( pstr , " \023 BitTorrent protocol " , 20))
return tr_handshakeDone ( handshake , false );
tr_peerIoReadBytes ( handshake - > io , inbuf , reserved , ...);
tr_peerIoReadBytes ( handshake - > io , inbuf , hash , ...);
/* ... parsing of handshake and sending reply omitted */
handshake - > state = AWAITING_PEER_ID ;
return READ_NOW ;
25
27
29
31
33
35
}
37
static int
readPeerId ( tr_handshake * handshake , struct evbuffer * inbuf ) {
uint8_t peer_id [ PEER_ID_LEN ];
39
if ( evbuffer_get_length ( inbuf ) < PEER_ID_LEN )
return READ_LATER ;
tr_peerIoReadBytes ( handshake - > io , inbuf , peer_id , ...);
/* ... parsing of peer id omitted */
return tr_handshakeDone ( handshake , true );
41
43
45
}
Fig. 2. Handshake state-machine in Transmission (simplified)
3.2
Data flow
Since each callback function performs only a small part of the whole computation,
the event-loop programmer needs to store temporary data required to carry
on the computation in heap-allocated data structures, whereas stack-allocated
variables would sometimes seem more natural in threaded style. The content of
9
these data structures depends heavily on the program being developed but we
can characterise some common patterns.
Event loops generally provide some means to specify a void* pointer when
registering an event handler. When the expected event triggers, the pointer is
passed as a parameter to the callback function, along with information about the
event itself. This allows the programmer to store partial results in a structure of
his choice, and recover it through the pointer without bothering to maintain the
association between event handlers and data himself.
Coarse-grained, long lived data structures These data structures are usually large
and coarse-grained. Each of them correponds to some meaningful object in the
context of the program, and is passed from callback to callback through a pointer.
For instance, the connection structure used in Polipo (Figure 1) is allocated
by httpMakeConnection when a connection starts (line 34) and passed to the
callbacks httpTimeoutHandler and httpClientHandler through the registering
functions scheduleTimeEvent (line 35) and do stream buf (line 41). It lives as
long as the HTTP connection it describes and contains no less than 22 fields: fd,
timeout, buf, pipelined, etc. The tr handshake structure passed to canRead
in Transmission is similarly large, with 18 fields.
Some of these fields need to live for the whole connection (eg. fd which stores
the file descriptor of the socket) but others are used only transiently (eg. buf
which is filled only when sending a reply), or even not at all in some cases (eg. the
structure HTTPConnectionPtr is used for both client and server connections, but
the pipelined field is never used in the client case). Even if it wastes memory in
some cases, it would be too much of a hassle for the programmer to track every
possible data flow in the program and create ad-hoc data structures for each of
them.
Minimal, short-lived data structures In some simple cases, however, the event-loop
programmer is able to allocate very small and short-lived data structures. These
minimal data structures are allocated directly within an event handler and are
deallocated when the associated callback returns. They might even be allocated
on the stack by the programmer and copied inside the event-loop internals by
the helper function registering the event handler. The overhead is therefore kept
as low as possible.
For instance, the function schedule accept passes a tiny, stack-allocated
structure request to the helper function registerFdEvent (Fig. 1, line 12).
This structure is of type AcceptRequestRec (not shown), which contains only
three fields: an integer fd and two pointers handler and data. It is copied by
registerFdEvent in the event-loop data structure associated with the event,
and freed automatically after the callback do scheduled accept has returned;
it is as short-lived and (almost) as compact as possible.
As it turns out, creating truly minimal structures is hard: AcceptRequestRec
could in fact be optimised to get rid off the fields data, which is always NULL in
practice in Polipo, and fd, which is also present in the encapsulating event data
structure. Finding every such redundancy in the data flow of a large event-driven
10
program would be a daunting task, hence the spurious and redundant fields used
to lighten the programmer’s burden.
4
Generating various event-driven styles
In this section, we first demonstrate the effect of CPC transformation passes on
a small example; we show that code produced by the CPC translator is very
close to event-driven code using callbacks for control flow, and minimal data
structures for data flow (Section 4.1). We then show how two other classical
translation passes produce different event-driven styles: defunctionalising inner
function yields state machines (Section 4.2), and encapsulating local variables in
shared environments yields larger, long-lived data structures with full context
(Section 4.3).
4.1
The CPC compilation technique
Consider the following function, which counts seconds down from an initial value x
to zero.
cps void countdown ( int x ) {
while ( x > 0) {
printf ( " % d \ n " , x - -);
cpc_sleep (1);
}
printf ( " time is over !\ n " );
}
This function is annotated with the cps keyword to indicate that it yields to the
CPC scheduler. This is necessary because it calls the CPC primitive cpc sleep,
which also yields to the scheduler.
The CPC translator is structured in a series of proven source-to-source
transformations [10], which turn a threaded-style CPC program into an equivalent
event-driven C program. Boxing first encapsulates a small number of variables in
environments. Splitting then splits the flow of control of each cps function into a
set of inner functions. Lambda lifting removes free local variables introduced by
the splitting step; it copies them from one inner function to the next, yielding
closed inner functions. Finally, the program is in a form simple enough to perform
a one-pass partial CPS conversion. The resulting continuations are used at
runtime to schedule threads.
In the rest of this section, we show how splitting, lambda lifting and CPS
conversion transform the function countdown. The boxing pass has no effect on
this example because it only applies to extruded variables, the address of which
is retained by the “address of” operator (&).
Splitting The first transformation performed by the CPC translator is splitting.
Splitting has been first described by van Wijngaarden for Algol 60 [20], and later
adapted by Thielecke to C, albeit in a restrictive way [19]. It translates control
structures into mutually recursive functions.
11
Splitting is done in two steps. The first step consists in replacing every controlflow structure, such as for and while loops, by its equivalent in terms of if and
goto.
cps void countdown ( int x ) {
loop :
if ( x <= 0) goto timeout ;
printf ( " % d \ n " , x - -);
cpc_sleep (1);
goto loop ;
timeout :
printf ( " time is over !\ n " );
}
The second step uses the fact that goto are equivalent to tail calls [18]. It
translates every labelled block into an inner function, and every jump to that
label into a tail call (followed by a return) to that function.
1
3
5
7
9
cps void countdown ( int x ) {
cps void loop () {
if ( x <= 0) { timeout (); return ; }
printf ( " % d \ n " , x - -);
cpc_sleep (1); loop (); return ;
}
cps void timeout () { printf ( " time is over !\ n " ); return ; }
loop (); return ;
}
Fig. 3. CPC code after splitting
Splitting yields a program where each cps function is split in several mutually
recursive, atomic functions, very similar to event handlers. Additionally, the tail
positions of these inner functions are always either:
– a return statement (for instance, on line 7 in the previous example),
– a tail call to another cps function (line 3),
– a call to an external cps function followed by a call to an inner cps function
(line 5).
We recognise the typical patterns of an event-driven program that we studied in
Section 3: respectively returning a value to the upper layer (Fig. 1 (4)), calling a
function to carry on the current computation (Fig. 2 (1)), or calling a function
with a callback to resume the computation once it has returned (Fig. 1 (2)).
Another effect of splitting is the introduction of free variables, which are
bound to the original encapsulating function rather than the new inner ones.
For instance, the variable x is free in the function loop above. Because inner
functions and free variables are not allowed in C, we perform a pass of lambda
lifting to eliminate them.
12
Lambda lifting The CPC translator then makes the data flow explicit with a
lambda-lifting pass. Lambda lifting, also called closure conversion, is a standard
technique to remove free variables introduced by Johnsson [8]. It is also performed
in two steps: parameter lifting and block floating.
Parameter lifting binds every free variable to the inner function where it
appears (for instance x to loop on line 2 below). The variable is also added as a
parameter at every call point of the function (lines 5 and 8).
1
3
5
7
9
cps void countdown ( int x ) {
cps void loop ( int x ) {
if ( x <= 0) { timeout (); return ; }
printf ( " % d \ n " , x - -);
cpc_sleep (1); loop ( x ); return ;
}
cps void timeout () { printf ( " time is over !\ n " ); return ; }
loop ( x ); return ;
}
Note that because C is a call-by-value language, lifted parameters are duplicated
rather than shared and this step is not correct in general. It is however sound in
the case of CPC because lifted functions are called in tail position: they never
return, which guarantees that at most one copy of each parameter is reachable
at any given time [10]. Block floating is then a trivial extraction of closed, inner
functions at top-level.
Lambda lifting yields a program where the data is copied from function to
function, each copy living as long as the associated handler. If some piece of data
is no longer needed during the computation, it will not be copied in the subsequent
handlers; for instance, the variable x is not passed to the function timeout. Hence,
lambda lifting produces short-lived, almost minimal data structures.
CPS conversion Finally, the control flow is made explicit with a CPS conversion
[14, 16]. The continuations store callbacks and their parameters in a regular
stack-like structure cont with two primitive operations: push to add a function
on the continuation, and invoke to call the first function of the continuation.
cps void loop ( int x , cont * k ) {
if ( x <= 0) { timeout ( k ); return ; }
printf ( " % d \ n " , x - -);
cpc_sleep (1 , push ( loop , x , k )); return ;
}
cps void timeout ( cont * k ) {
printf ( " time is over !\ n " );
invoke ( k ); return ;
}
cps void countdown ( int x , cont * k ) { loop (x , k ); return ; }
CPS conversion turns out to be an efficient and systematic implementation of
the layered callback scheme described in Section 3.1. Note that, just like lambda
lifting, CPS conversion is not correct in general in an imperative call-by-value
language, because of duplicated variables on the continuation. It is however
13
correct in the case of CPC, for reasons similar to the correctness of lambda
lifting [10].
4.2
Defunctionalising inner functions
Defunctionalisation is a compilation technique introduced by Reynolds to translate
higher-order programs into first-order ones [15]. It maps every first-class function
to a first-order structure that contains both an index representing the function,
and the values of its free variables. These data structures are usually a constructor,
whose parameters store the free variables. Function calls are then performed by
a dedicated function that dispatches on the constructor, restores the content of
the free variables and executes the code of the relevant function.
The dispatch function introduced by defunctionalisation is very close to a state
automaton. It is therefore not surprising that defunctionalising inner functions
in CPC yields an event-driven style similar to state machines (Section 3.1).
Defunctionalisation of CPC programs Usually, defunctionalisation contains an
implicit lambda-lifting pass, to make free variables explicit and store them in
constructors. For example, a function fn x => x + y would be replaced by an
instance of LAMBDA of int, with the free variable y copied in the constructor
LAMBDA. The dispatch function would then have a case: dispatch (LAMBDA y,
x) = x + y.
In this discussion, we wish to decouple this data-flow transformation from
the translation of the control flow into a state machine. Therefore, we define the
dispatch function as an inner function which merges the content of the other inner
functions but still contains free variables. This is possible because the splitting
pass does not create any closure: it introduces inner functions with free variables,
but these are always called directly, not stored as first-class values whose free
variables must be captured.
Consider again our countdown example after the splitting pass (Fig. 3). Once
defunctionalised, it contains a single inner function dispatch that dispatches on
an enumeration representing the former inner function loop and timeout.
2
4
6
8
10
12
14
enum state { LOOP , TIMEOUT };
cps void countdown ( int x ) {
cps void dispatch ( enum state s ) {
switch ( s ) {
case LOOP :
if ( x <= 0) { dispatch ( TIMEOUT ); return ; }
printf ( " % d \ n " , x - -);
cpc_sleep (1); dispatch ( LOOP ); return ;
case TIMEOUT :
printf ( " time is over !\ n " ); return ;
}
}
dispatch ( LOOP ); return ;
}
14
As an optimisation, the recursive call to dispatch on line 6 can be replaced by a
goto statement. However, we cannot replace the call that follows the cps function
cpc sleep(1) on line 8, since we will need to provide dispatch as a callback to
cpc sleep during CPS conversion, to avoid blocking.
We must then eliminate free variables and inner functions, with a lambdalifting pass. It is still correct because defunctionnalisation does not break the
required invariants on tail calls. We finally reach code that is similar in style to
the state-machine shown in Fig. 2.
cps void dispatch ( enum state s , int x ) {
switch ( s ) {
case LOOP :
if ( x <= 0) goto timeout_label ;
printf ( " % d \ n " , x - -);
cpc_sleep (1); dispatch ( LOOP , x ); return ;
case TIMEOUT : timeout_label :
printf ( " time is over !\ n " ); return ;
}
}
cps void countdown ( int x ) { dispatch ( LOOP , x ); return ; }
In this example, we have also replaced the first occurrence of dispatch with
goto timeout label, as discussed above, which avoids the final function call
when the counter reaches zero.
If we ignore the switch – which serves mainly as an entry point to the
dispatch function, à la Duff’s device [4] – we recognise the intermediate code
generated during the first step of splitting, as having an explicit control flow
using gotos but without inner functions. In retrospect, the second step of splitting, which translates gotos to inner functions, can be considered as a form a
refunctionalisation, the left-inverse of defunctionalisation [3].
Benefits The translation presented here is in fact a partial defunctionalisation:
each cps function in the original program gets its own dispatch function, and only
inner functions are defunctionalised. A global defunctionalisation would imply a
whole program analysis, would break modular compilation, and would probably
not be very efficient because C compilers are optimised to compile hand-written,
reasonably-sized functions rather than a giant state automaton with hundreds
of states. On the other hand, since it is only partial, this translation does not
eliminate the need for a subsequent CPS conversion step to translate calls to
external cps functions into operations on continuations.
Despite adding a new translation step while keeping the final CPS conversion,
this approach has several advantages over the CPS conversion of many smaller,
mutually recursive functions performed by the current CPC translator. First, we
do not pay the cost of a CPS call for inner functions. This might bring significant
speed-ups in the case of tight loops or complex control flows. Moreover, it leaves
with much more optimisation opportunities for the C compiler, for instance to
store certain variables in registers, and reduces the number of operations on the
continuations. It also makes debugging easier, avoiding numerous hops through
ancillary cps functions.
15
4.3
Shared environments
The two main compilation techniques to handle free variables are lambda lifting,
illustrated in Section 4.1 and discussed extensively in a previous article [10], and
environments. An environment is a data structure used to capture every free
variable of a first-class function when it is defined; when the function is later
applied, it accesses its variables through the environment. Environments add
a layer of indirection, but contrary to lambda lifting they do not require free
variables to be copied on every function call.
In most functional languages, each environment represents the free variables
of a single function; a pair of a function pointer and its environment is called
a closure. However, nothing prevents in principle an environment from being
shared between several functions, provided they have the same free variables. We
use this technique to allocate a single environment shared by inner functions,
containing all local variables and function parameters.
An example of shared environments Consider once again our countdown example
after splitting (Fig. 3). We introduce an environment to contain the local variables
of countdown (here, there is only x).
1
3
5
7
9
11
13
15
struct env_countdown { int x };
cps void countdown ( int x ) {
struct env_countdown * e =
malloc ( sizeof ( struct env_countdown ));
e->x = x;
cps void loop ( struct env_countdown * e ) {
if (e - > x <= 0) { timeout (); return ; }
printf ( " % d \ n " , e - >x - -);
cpc_sleep (1); loop ( e ); return ;
}
cps void timeout ( struct env_countdown * e ) {
printf ( " time is over !\ n " );
free ( e ); return ;
}
loop ( e ); return ;
}
The environment is allocated (line 4) and initialised (line 5) when the function
countdown is entered. The inner functions access x through the environment,
either to read (line 7) or to write it (line 8). A pointer to the environment is
passed from function to function (line 9); hence all inner functions share the
same environment. Finally, the environment is deallocated just before the last
inner function exits (line 13).
The resulting code is similar in style to hand-written event-driven code, with
a single, heap-allocated data structure sharing the local state between a set of
callbacks. Note that inner functions have no remaining free variable and can
therefore be lambda-lifted trivially.
Benefits Encapsulating local variables in environments avoids having to copy
them back and forth between the continuation and the native call stack. However,
16
it does not necessarily mean that the generated programs are faster; in fact,
lambda-lifted programs are often more efficient (Section 5). Another advantage of
environments is that they make programs easier to debug, because the local state
is always fully available, whereas in a lambda-lifted program “useless” variables
are discarded as soon as possible, even though they might be useful to understand
what went wrong before the program crashed.
5
Evaluation
In this section, we describe the implementation of eCPC, a CPC variant using
shared environments instead of lambda lifting to encapsulate the local state
of cps functions. We then compare the efficiency of programs generated with
eCPC and CPC, and show that the latter is more efficient in most cases. This
demonstrates the benefits of generating events automatically: most real-world
event-driven programs are based on environments, because they are much easier
to use, although systematic lambda lifting would probably yield faster code.
5.1
Implementation
The implementation of eCPC is designed to reuse as much of the existing CPC
infrastructure as possible. The eCPC translator introduces two new passes:
preparation and generation of environments. The former replaces the boxing pass;
the latter replaces lambda lifting.
Environment preparation Environments must be introduced before the splitting
pass for two reasons. First, it is easier to identify the exit points of cps functions,
where the environments must be deallocated, before they are split into multiple,
mutually recursive, inner functions. Furthermore, these environment deallocations
occur in tail position, and have therefore an impact on the splitting pass itself.
Although deallocation points are introduced before splitting, neither allocation nor initialisation or indirect memory accesses are performed at this stage.
Environments introduced during this preparatory pass are empty shells, of type
void*, that merely serve to mark the deallocation points. This is necessary
because not all temporary variables have been introduced at this stage (the
splitting pass will generate more of them). Deciding which variables will be stored
in environments is delayed to a later pass.
This preparatory pass also needs to modify how return values are handled. In
the original CPC, return values are written directly in the continuation when
the returning function invokes its continuation. This is made possible by the
convention that the return value of a cps function is the last parameter of its
continuation, hence at a fixed position on the continuation stack. Such is not
the case in eCPC, where function parameters are kept in the environment rather
than copied on the continuation.
17
In eCPC, the caller function passes a pointer to the callee, indicating the
address where the callee must write its return value.5 The preparatory pass
transforms every cps function returning a type T (different from void) into a
function returning void with an additional parameter of type T ∗; call and return
points are modified accordingly. The implementation of CPC primitives in the
CPC runtime is also modified to reflect this change.
Environment generation After the splitting pass, the eCPC translator allocates
and initialises environments, and replaces variables by their counterpart in the
environment.
First, it collects local variables (except the environment pointer itself) and
function parameters and generates the layout of the associated environment.
Then, it allocates the environment and initialises the fields corresponding to the
function parameters. Because this initialisation is done at the very beginning of
the translated function, it does not affect the tails, thus preserving the correctness
of CPS conversion. Finally, every use of variables is replaced by its counterpart in
the environment, local variables are discarded, and inner functions are modified
to receive the environment as a parameter instead.
The CPS conversion is kept unchanged: the issue of return values is dealt
with completely in the preparatory pass and every cps function returns void at
this stage.
5.2
Benchmark results
We previously designed a set of benchmarks to compare CPC to other thread
libraries, and have shown that CPC is as fast as the fastest thread libraries
available to us while providing at least an order of magnitude more threads [10].
We reuse these benchmarks here to compare the speed of CPC and eCPC; our
experimental setup is unchanged, and detailed in our previous work.
Primitive operations We first measure the time of individual CPC primitives.
Table 1 shows the relative speed of eCPC compared with CPC for each of our
micro-benchmarks: teCPC /tCPC . A value greater than 1 indicates that eCPC is
slower than CPC. The slowest primitive operation in CPC is a cps function call
(cps-call ), mostly because of the multiple layers of indirection introduced by
continuations. This overhead is even larger in the case of eCPC: performing a
cps function call is 2 to 3 times slower than with CPC.
This difference of cost for cps function calls probably has an impact on the
other benchmarks, making them more difficult to interpret. Context switches
(switch) are around 50 % slower on every architecture, which is surprisingly high
since they involve almost no manipulation of environments. Thread creation
(spawn) varies a lot across architectures: more than 3 times slower on the Pentium M, but only 59 % slower on a MIPS embedded processor. Finally, condition
5
Note that a similar device would be necessary to implement defunctionalisation,
because the dispatch function is a generic callback which might receive many
different types of return values.
18
Table 1. Ratio of speeds of eCPC to CPC
Architecture
Core 2 Duo (x86-64)
Pentium M (x86)
MIPS-32 4KEc
cps-call switch condvar spawn
2.45
2.35
2.92
1.67
1.75
1.43
1.13
1.08
0.91
2.18
3.12
1.59
variables (condvar ) are even more surprising: not much slower on x86 and x86-64,
and even 9 % faster on MIPS. It is unclear which combination of factors leads
eCPC to outperform CPC on this particular benchmark only: we believe that
the larger number of registers helps to limit the number of memory accesses, but
we were not able to quantify this effect precisely.
These benchmarks of CPC primitives show that the allocation of environments
slows down eCPC in most cases, and confirms our intuition that avoiding boxing
as much as possible in favour of lambda lifting is very important in CPC.
Tic-tac-toe generator Unfortunately, benchmarking individual CPC primitives
gives little information on the performance of a whole program, because their cost
might be negligible compared to other operations. To get a better understanding
of the performance behaviour of eCPC, we wrote a trivial but highly concurrent
program with intensive memory operations: a tic-tac-toe generator that explores
the space of grids. It creates three threads at each step, each one receiving a copy
of the current grid, hence 39 = 19 683 threads and as many copies of the grid.
We implemented two variants of the code, to test different schemes of memory
usage. The former is a manual scheme that allocates copies of the grids with
malloc before creating the associated threads, and frees each of them in one of
the “leaf” threads, once the grid is completed. The latter is an automatic scheme
that declares the grids as local variables and synchronises their deallocation with
barriers; the grids are then automatically encapsulated, either by the boxing pass
(for CPC) or in the environment (for eCPC).
Our experiment consists in launching an increasing number of generator tasks
simultaneously, each one generating the 19 683 grids and threads mentioned
above. We run up to 100 tasks simultaneously, ie. almost 2 000 000 CPC threads
in total, and the slowest benchmark takes around 3 seconds to complete on an
Intel Centrino 1,87 Ghz, downclocked to 800 MHz.
Finally, we compute the mean time per tic-tac-toe task. This ratio turns out
to be independent of the number of simultaneous tasks: both CPC and eCPC
scale linearly in this benchmark. We measured that eCPC is 20 % slower than
CPC in the case of manual allocation (13.2 vs. 11.0 ms per task), and 18 % slower
in the automatic case (31.3 vs. 26.5 ms per task). This benchmark confirms that
environments add a significant overhead in programs performing a lot a memory
accesses, although it is not as important as in benchmarks of CPC primitives.
Web servers To evaluate the impact of environments on more realistic programs,
we reuse our web server benchmark [9]. We measure the mean response time of a
19
small web server under the load of an increasing number of simultaneous clients.
The server is deliberately kept minimal, and uses one CPC thread per client. The
results are shown in Fig. 4. In this benchmark, the web server compiled with
eCPC is 12 % slower than the server compiled with CPC. Even on programs that
spend most of their time performing network I/O, the overhead of environments
remains measurable.
Mean time per request (ms)
250
eCPC (slope: 0.205)
CPC (slope: 0.183)
200
150
100
50
0
0
200
400
600
800
1000
Number of simultaneous clients
Fig. 4. Web server benchmark
6
Conclusions
Through the analyse of real-world programs, we have identified several typical
styles of control-flow and data-flow encodings in event-driven programs: callbacks
or state machines for the control flow, and coarse-grained or minimal data
structures for the data flow. We have then shown how these various styles can be
generated from a common threaded description, by a set of automatic program
transformations. Finally, we have implemented eCPC, a variant of the CPC
translator using shared environments instead of lambda lifting. We have found
out that, although rarely used in real-world programs because it is tedious to
perform manually, lambda lifting yields better performance than environments
in most of our benchmarks.
An interesting extension of our work would be to try and reverse our program
transformations, in order to reconstruct threaded code from event-driven programs. This could help analysing and debugging event-driven code, or migrating
legacy, hard-to-maintain event-driven programs like Polipo towards CPC or other
cooperative threads implementations.
Acknowledgments The authors would like to thank Andy Key, who has inspired
this work, Juliusz Chroboczek for his continuous support, Peter Sewell, Chloé
Azencott and Alan Mycroft for their valuable comments.
20
References
1. Adya, A., Howell, J., Theimer, M., Bolosky, W.J., Douceur, J.R.: Cooperative
task management without manual stack management. In: Ellis, C.S. (ed.) USENIX
Annual Technical Conference, General Track. pp. 289–302. USENIX (2002)
2. Boussinot, F.: FairThreads: mixing cooperative and preemptive threads in C.
Concurrency and Computation: Practice and Experience 18(5), 445–469 (2006)
3. Danvy, O., Nielsen, L.R.: Defunctionalization at work. In: PPDP. pp. 162–174.
ACM (2001)
4. Duff, T.: Duff’s device (Nov 1983), electronic mail to R. Gomes, D. M. Ritchie and
R. Pike
5. Dunkels, A., Schmidt, O., Voigt, T., Ali, M.: Protothreads: simplifying eventdriven programming of memory-constrained embedded systems. In: Campbell, A.T.,
Bonnet, P., Heidemann, J.S. (eds.) SenSys. pp. 29–42. ACM (2006)
6. Fischer, J., Majumdar, R., Millstein, T.D.: Tasks: language support for event-driven
programming. In: Ramalingam, G., Visser, E. (eds.) PEPM. pp. 134–143. ACM
(2007)
7. Harris, T., Abadi, M., Isaacs, R., McIlroy, R.: AC: composable asynchronous IO for
native languages. In: Lopes, C.V., Fisher, K. (eds.) OOPSLA. pp. 903–920. ACM
(2011)
8. Johnsson, T.: Lambda lifting: Transforming programs to recursive equations. In:
FPCA. pp. 190–203 (1985)
9. Kerneis, G., Chroboczek, J.: Are events fast? Tech. rep., PPS, Université Paris 7
(Jan 2009)
10. Kerneis, G., Chroboczek, J.: Continuation-passing C, compiling threads to events
through continuations. Higher-Order and Symbolic Computation 24, 239–279 (2011)
11. Key, A.: Weave: translated threaded source (with annotations) to fibers with context
passing (ca 1995–2000), personal communication (2010)
12. Krohn, M.N., Kohler, E., Kaashoek, M.F.: Events can make sense. In: USENIX
Annual Technical Conference. pp. 87–100 (2007)
13. Myers, D.S., Carlisle, J.N., Cowling, J.A., Liskov, B.: Mapjax: Data structure
abstractions for asynchronous web applications. In: USENIX Annual Technical
Conference. pp. 101–114 (2007)
14. Plotkin, G.D.: Call-by-name, call-by-value and the lambda-calculus. Theoretical
Computer Science 1(2), 125–159 (1975)
15. Reynolds, J.C.: Definitional interpreters for higher-order programming languages.
Proceedings of the ACM National Conference 2(4), 717–740 (Aug 1972)
16. Reynolds, J.C.: The discoveries of continuations. Lisp and Symbolic Computation
6(3-4), 233–248 (1993)
17. Srinivasan, S., Mycroft, A.: Kilim: Isolation-typed actors for java. In: Vitek, J.
(ed.) ECOOP. Lecture Notes in Computer Science, vol. 5142, pp. 104–128. Springer
(2008)
18. Steele Jr., G.L., Sussman, G.J.: Lambda, the ultimate imperative (Mar 1976)
19. Thielecke, H.: Continuations, functions and jumps. SIGACT News 30(2), 33–42
(1999)
20. van Wijngaarden, A.: Recursive definition of syntax and semantics. In: Formal
Language Description Languages for Computer Programming. pp. 13–24. NorthHolland Publishing Company, Amsterdam, Netherlands (1966)
| 6cs.PL
|
1
Distributed Gauss-Newton Method for AC State
Estimation Using Belief Propagation
arXiv:1702.05781v2 [cs.IT] 1 Oct 2017
Mirsad Cosovic, Student Member, IEEE, Dejan Vukobratovic, Member, IEEE
Abstract—We present a detailed study on application of
factor graphs and the belief propagation (BP) algorithm to the
power system state estimation (SE) problem. We start from the
BP solution for the linear DC model, for which we provide a
detailed convergence analysis. Using insights from the DC
model, we use two different approaches to derive the BP
algorithm for the non-linear AC model. The first method
directly applies BP methodology, however, providing only
approximate BP solution for the AC model. In the second
approach, we make a key further step by providing the
solution in which the BP is applied sequentially over the AC
model, akin to what is done by the Gauss-Newton method. The
resulting BP-based Gauss-Newton algorithm has the
interpretation of a fully distributed Gauss-Newton method with
the same accuracy as the centralized SE, preserving a number
of advantages of the BP framework. The paper provides
extensive numerical study of the proposed algorithms, provides
details on their convergence properties, and gives a number of
useful insights for their implementation.
Index Terms—State Estimation, Electric Power System, Factor
Graphs, Belief Propagation, Distributed Gauss-Newton Method
I. I NTRODUCTION
Motivation: Electric power systems consist of generation,
transmission and consumption spread over wide geographical
areas and operated from the control centers by the system
operators. Maintaining normal operation conditions is of the
central importance for the power system operators [1, Ch. 1].
Control centers are traditionally operated in centralized and
independent fashion. However, increase in the system size
and complexity, as well as external socio-economic factors,
lead to deregulation of power systems, resulting in
decentralized structure with distributed control centers.
Cooperation in control and monitoring across distributed
control centers is critical for efficient system operation.
Consequently, existing centralized algorithms have to be
redefined based on a new requirements for distributed
operation, scalability and computational efficiency [2].
The system monitoring is an essential part of the control
centers, providing control and optimization functionality
This paper was presented in part at the IEEE International Conference
on Smart Grid Communications (SmartGridComm), Sydney, Australia,
November 2016; in part at the IEEE International Workshop on Signal
Processing Advances in Wireless Communications (SPAWC), Edinburgh,
United Kingdom, August 2016.
M. Cosovic is with Schneider Electric DMS NS, Novi Sad, Serbia
(e-mail: [email protected]). D. Vukobratovic is
with Department of Power, Electronic and Communications Engineering,
University of Novi Sad, Novi Sad, Serbia (e-mail: [email protected]). Demo
source code available online at https://github.com/mcosovic.
This paper has received funding from the EU 7th Framework Programme
for research, technological development and demonstration under grant
agreement no. 607774.
whose efficiency relies on accurate state estimation (SE).
The core of the SE is the SE algorithm that provides an
estimate of the system state (i.e., the set of all complex bus
voltages) based on the network topology and available
measurements. Besides the SE algorithm, the SE includes
several additional routines such as network topology
processors, observability analysis and bad data analysis.
The centralized SE assumes that the measurements
collected across the system are available at the control
center, where the centralized SE algorithm provides the
system state estimate. Precisely, the centralized SE algorithm
typically uses the Gauss-Newton method to solve the
non-linear weighted least-squares (WLS) problem [3], [4]. In
contrast, decentralized SE distributes communication and
computational effort across multiple control centers to
provide the system state estimate. There are two main
approaches to distributed SE: i) algorithms which require a
global control center to exchange data with local control
centers, and ii) algorithms with local control centers only
[5]. Distributed SE algorithms target the same state estimate
accuracy as achievable using the centralized SE algorithms.
Literature review: The mainstream distributed SE
algorithms exploit matrix decomposition techniques applied
over the Gauss-Newton method. These algorithms usually
achieve the same accuracy as the centralized SE algorithm
and work with either global control center [6]–[8] or without
it [9]–[12]. Recently, the SE algorithms based on distributed
optimization [13], and in particular, the alternating direction
method of multipliers became very popular [14]. Examples
include algorithms based on convex relaxation [15], [16],
and without convex relaxation [17]. In [18], the gossip-based
Gauss-Newton algorithm is proposed which provides flexible
communication model, but suffers from slight performance
degradation compared to the centralized SE. The work in
[19] presents a fully distributed SE algorithm for wide-area
monitoring which provably converges to the centralized SE.
We refer the reader to [20] for a detailed survey of the
distributed multi-area SE.
Belief-Propagation Approach: In this paper, we solve the
SE problem using probabilistic graphical models [21], a
powerful tool for modeling the dependencies among the
systems of random variables. We represent the SE problem
using a popular class of probabilistic graphical models called
factor graphs and solve it using the belief propagation (BP)
algorithm. Applying the BP algorithm on probabilistic
graphical models without loops, one obtains exact marginal
distributions or a mode of the joint distribution of the system
of random variables [21], [22]. The BP algorithm can be
2
also applied to graphical models with loops (loopy BP) [23],
although in that case, the solution may not converge to the
correct marginals/modes of the joint distribution. BP is a
fully distributed algorithm suitable for accommodation of
distributed power sources and time-varying loads. Moreover,
placing the SE into the probabilistic graphical modelling
framework enables not only efficient inference, but also, a
rich collection of tools for learning parameters or structure
of the graphical model from observed data [24].
The work in [25] provides the first demonstration of BP
applied to the SE problem. Although this work is elaborate
in terms of using, e.g., environmental correlation via
historical data, it applies BP to a simple linearized DC
model. The AC model is recently addressed in [26], where
tree-reweighted BP is applied using preprocessed weights
obtained by randomly sampling the space of spanning trees.
The work in [27] investigates Gaussian BP convergence
performance for the DC model. Although the above results
provide initial insights on using BP for distributed SE, it is
fair to say that a systematic analysis of applying the BP
algorithm on the SE problem, and in particular for the AC
model, is still missing. This paper intends to fill this gap.
Contributions: In this paper, we present a novel
distributed BP-based Gauss-Newton algorithm. In the process
of deriving the proposed algorithm, we provide a
step-by-step guide for application of BP algorithm to the SE
problem, giving this paper strong tutorial flavor. Our
methodology is to start with the simplest linear DC model
and use insights obtained therein to derive the native BP
solution for the non-linear AC model. Unfortunately, as
closed-form expressions for certain classes of BP messages
cannot be obtained, we present suitable approximations that
lead us to propose the AC-BP algorithm as an (approximate)
BP solution for the AC model. Then, as a main contribution,
we make a key further step where we change the perspective
of our BP approach and, instead of applying the BP directly
onto the non-linear AC model, we present the solution where
the BP is applied sequentially over the AC model, akin to
what is done by the Gauss-Newton method. The resulting
Gauss-Newton BP (GN-BP) represents a BP counterpart of
the Gauss-Newton method achieving the same accuracy,
however, preserving a number of advantages brought in by
the BP framework. For example, the proposed GN-BP
algorithm is of considerably lower computational complexity,
can be easily designed to provide asynchronous operation,
could be integrated as part of real-time systems [28], and is
flexible and easy to distribute and parallelize. Finally, this
paper provides a novel and detailed convergence analysis of
the DC-BP algorithm and points to extension of this analysis
for the proposed GN-BP algorithm. We close this paper by
extensive numerical study where we compare the proposed
BP-based algorithms with the centralized SE, and provide a
number of useful insights for their implementation.
Paper Organization: In Section II, we present basic
concepts of factor graphs and BP. Section III reviews the
centralized DC and AC SE. Section IV presents closed form
BP expressions and detailed convergence analysis for the DC
model. In Section V, the AC model is solved via BP-based
Gauss-Newton method. Section VI presents extended
numerical convergence and performance results. Concluding
remarks are given in Section VII.
II. FACTOR G RAPHS AND BP A LGORITHM
Factor graphs and BP algorithm are widely used tools for
probabilistic inference [21], [22]. In the standard setup, the
goal of the BP algorithm is to efficiently evaluate the marginals
of a system of random variables x = (x1 , . . . , xn ) described
via the joint probability density function g(x). Assuming that
the function g(x) can be factorized proportionally (∝) to a
product of local functions:
g(x) ∝
k
Y
ψi (Xi ),
(1)
i=1
where Xi ⊆ {x1 , . . . , xn }, the marginalization problem can
be efficiently solved using BP algorithm.
The first step is forming a factor graph, which is a
bipartite graph that describes the structure of the
factorization (1). The factor graph structure comprises the set
of factor nodes F = {f1 , . . . , fk }, where each factor node fi
represents local function ψi (Xi ), and the set of variable
nodes X = {x1 , . . . , xn }. The factor node fi connects to the
variable node xs if and only if xs ∈ Xi [29].
The BP algorithm on factor graphs proceeds by passing
two types of messages along the edges of the factor graph: i)
a variable node to a factor node, and ii) a factor node to a
variable node messages. Both variable and factor nodes in a
factor graph process the incoming messages and calculate
outgoing messages, where an output message on any edge
depends on incoming messages from all other edges. BP
messages represent ”beliefs” about variable nodes, thus a
message that arrives or departs a certain variable node is a
function (distribution) of the random variable corresponding
to the variable node.
Message from a variable node to a factor node: Consider
a part of a factor graph shown in Fig. 1 with a group of factor
nodes Fs = {fi , fw , ..., fW } ⊆ F that are neighbours of the
variable node xs ∈ X . The message µxs →fi (xs ) from the
µ f w → xs ( x s )
fw
.
.
.
fW
x s µ xs → f i ( x s )
fi
µ f W → xs ( x s )
Fig. 1. Message µxs →fi (xs ) from variable node xs to factor node fi
variable node xs to the factor node fi is equal to the product
of all incoming factor node to variable node messages arriving
at all the other incident edges:
Y
µxs →fi (xs ) =
µfa →xs (xs ),
(2)
fa ∈Fs \fi
3
where Fs \ fi represents the set of factor nodes incident to the
variable node xs , excluding the factor node fi . Note that each
message is a function of the variable xs .
Message from a factor node to a variable node:
Consider a part of a factor graph shown in Fig. 2 that
consists of a group of variable nodes Xi = {xs , xl , ..., xL } ⊆
X that are neighbours of the factor node fi ∈ F. The
µ xl → f i ( x l )
xl
.
.
.
f i µ f i → xs ( x s )
xs
µ xL → f i ( x L )
xL
Fig. 2. Message µfi →xs (xs ) from factor node fi to variable node xs
message µfi →xs (xs ) from the factor node fi to the variable
node xs is defined as a product of all incoming variable
node to factor node messages arriving at other incident
edges, multiplied by the function ψi (Xi ) associated to the
factor node fi , and marginalized over all of the variables
associated with the incoming messages:
Z
Z
···
µfi →xs (xs ) =
xL
xl
Y
ψi (Xi )
µxb →fi (xb ) · dxb , (3)
xb ∈Xi \xs
where Xi \xs is the set of variable nodes incident to the factor
node fi , excluding the variable node xs .
Marginal inference: The marginal of the variable node xs ,
illustrated in Fig. 3, is obtained as the product of all incoming
messages into the variable node xs :
Y
p(xs ) =
µfc →xs (xs ),
(4)
fc ∈Fs
where Fs is the set of factor nodes incident to the variable
node xs .
µ f w → xs ( x s )
fw
.
.
.
fW
x s µ f i → xs ( x s )
fi
III. SE IN E LECTRIC P OWER S YSTEMS
The SE algorithm estimates the values of the state
variables based on the knowledge of network topology and
parameters, and measured values obtained from measurement
devices spread across the power system. The knowledge of
the network topology and parameters is provided by the
network topology processor in the form of the bus/branch
model, where branches of the grid are usually described
using the two-port π-model [1, Ch. 1,2]. The bus/branch
model can be represented using a graph G = (V, E), where
the set of nodes V = {1, . . . , N } represents the set of buses,
while the set of edges E ⊆ V × V represents the set of
branches of the power network.
As an input, the SE requires a set of measurements M of
different electrical quantities spread across the power
network. Using the bus/branch model and available
measurements, the observability analysis defines the
measurement model [1, Ch. 4] which can be described as the
system of equations [4]:
z = h(x) + u,
where x = (x1 , . . . , xn ) is the vector of the state variables,
h(x) = (h1 (x), . . . , hk (x)) is the vector of measurement
functions, z = (z1 , . . . , zk ) is the vector of measurement
values, and u = (u1 , . . . , uk ) is the vector of uncorrelated
measurement errors. The SE problem is commonly an
overdetermined system of equations (k > n) [31, Sec. 2.1].
Each measurement Mi ∈ M is associated with measured
value zi , measurement error ui , and measurement function
hi (x). Under the assumption that measurement errors ui
follow a zero-mean Gaussian distribution, the probability
density function associated with the i-th measurement equals:
(
)
[zi − hi (x)]2
1
exp
,
(6)
N (zi |x, vi ) = √
2vi
2πvi
where vi is the measurement variance defined by the
measurement error ui , and the measurement function hi (x)
connects the vector of state variables x to the value of the
i-th measurement.
The SE in electric power systems deals with the problem of
determining state variables x according to the noisy observed
data z and a prior knowledge:
µ f W → xs ( x s )
Fig. 3. Marginal inference of the variable node xs
In the following, we focus on the Gaussian BP (GBP)
over the set of real-valued random variables where all the
local functions ψi (Xi ) and all the inputs to the BP algorithm
represent Gaussian distributions. Due to the fact that variable
node and factor node processing preserves ”Gaussianity” of
the messages, the resulting GBP algorithm operates by
exchanging the messages that represent Gaussian functions.
Therefore, each message exchanged in GBP is completely
represented using only two values: the mean and the
variance [30].
(5)
p(x|z) =
p(z|x)p(x)
.
p(z)
(7)
Assuming that the prior probability distribution p(x) is
uniform, and given that p(z) does not depend on x, the
maximum a posteriori (MAP) solution of (7) reduces to the
maximum likelihood solution, as given below [32]:
x̂ = arg max p(x|z) = arg max p(z|x) = arg max L(z|x). (8)
x
x
x
One can find the solution (8) via maximization of the
likelihood function L(z|x), which is defined via likelihoods
of k independent measurements:
x̂ = arg max L(z|x) = arg max
x
x
k
Y
i=1
N (zi |x, vi ).
(9)
4
It can be shown that the solution of the MAP problem can
be obtained by solving the following optimization problem,
known as the WLS problem [33, Sec. 9.3]:
x̂ = arg min
x
k
X
[zi − hi (x)]2
i=1
vi
.
(10)
The state estimate x̂ representing the solution of the
optimization problem (10) is known as the WLS estimator,
and is equivalent to the MAP estimator (9).
A. AC State Estimation
only bus voltage angles x ≡ θ as state variables.
Consequently, the number of state variables is n = N .
The set of DC model measurements M involves only active
power flow {MPij }, (i, j) ∈ E, active power injection {MPi },
i ∈ V, and bus voltage angle {Mθi }, i ∈ V, measurements and
the model is dealing with linear measurement functions h(x)
given Appendix A.
The DC state estimate x̂ ≡ θ̂, which is a solution to the
WLS problem (10), is obtained through the non-iterative
procedure by solving the system of linear equations:
HT WH x̂ = HT Wz,
(12)
The AC SE model is defined using the measurement
functions h(x) that precisely follow the physical laws that
connect the measured variables and the state variables. As a
result, the system (5) in general represents the system of
non-linear equations.
In a usual scenario, the AC SE model takes bus voltage
magnitudes and bus voltage angles, transformer magnitudes
of turns ratio and transformer angles of turns ratio as state
variables x. Without loss of generality, in the rest of the
paper, for the AC SE we observe bus voltage angles θ =
(θ1 , . . . , θN ) and bus voltage magnitudes V = (V1 , . . . , VN )
as state variables x ≡ (θ, V). Consequently, the number of
state variables is n = 2N .
The typical set of measurements M in electric power
systems includes: active and reactive power flow and current
magnitude, {MPij , MQij , MIij }, (i, j) ∈ E, respectively;
active and reactive power injection, {MPi , MQi }, i ∈ V; and
bus voltage angle and magnitude, {Mθi , MVi }, i ∈ V. For
completeness, we provide the set of non-linear measurement
functions h(x) of the AC SE model given in Appendix A,
equations (62) - (65).
Based on the available set of measurements, the WLS
estimator x̂ ≡ (θ̂, V̂), i.e., the solution of the WLS problem
(10), can be found using the Gauss-Newton method:
h
i
J(x(ν) )T WJ(x(ν) ) ∆x(ν) = J(x(ν) )T Wr(x(ν) ) (11a)
where H ∈ RkxN is the Jacobian matrix of measurement
functions given in Appendix A, equations (66) - (68).
x(ν+1) = x(ν) + ∆x(ν) ,
B. DC State Estimation
For the DC model, the set of variable nodes is defined by
the state variables θ, i.e., X = {θ1 , . . . , θN }. The set of
factor nodes F = {f1 , . . . , fk } is defined by the set of
measurements M, with measurement functions given in
Appendix A, equations (66) - (68). Measurements define
likelihood functions N (zi |x, vi ) that are in turn equal to
local functions ψi (Xi ) associated to factor nodes. More
precisely, a factor node fi connects to a variable node
xs ∈ X if and only if the state variable xs is an argument of
the corresponding measurement function hi (x).
The DC model is obtained by linearisation of the AC
model. In the typical operating conditions, the difference of
bus voltage angles between adjacent buses (i, j) ∈ E is very
small θi − θj ≈ 0, which implies cos θij ≈ 1 and
sin θij ≈ θij . Further, all bus voltage magnitudes are Vi ≈ 1,
i ∈ V, and all shunt elements and branch resistances can be
neglected. This implies that the DC model ignores the
reactive powers and transmission losses and takes into
account only the active powers. Therefore, the DC SE takes
Example 1 (Constructing factor graph). In this toy example,
using a simple 3-bus model presented in Fig. 4a, we
demonstrate the conversion from a bus/branch model with a
given measurement configuration into the corresponding
factor graph for the DC model. The variable nodes represent
state variables, i.e., X = {θ1 , θ2 , θ3 }. Factor nodes are
defined by corresponding measurements, where in our
example, measurements MP12 and MP3 are mapped into
factor nodes F = {fP12 , fP3 }.
4
(11b)
where ν = {0, 1, 2, . . . } is the iteration index, ∆x(ν) ∈ Rn is
the vector of increments of the state variables, J(x(ν) ) ∈ Rkxn
is the Jacobian matrix of measurement functions h(x(ν) ) at
x = x(ν) (see Appendix A, for details), W ∈ Rkxk is a
diagonal matrix containing inverses of measurement variances,
and r(x(ν) ) = z−h(x(ν) ) is the vector of residuals [3, Ch. 10].
IV. BP- BASED A LGORITHM FOR THE DC SE
In the SE problem defined above, each measurement
function hi (x) depends on a limited (typically small) subset
of state variables x. Hence, the likelihood function L(z|x)
can be factorized into factors (9) affecting small subsets of
state variables. This motivates solving the SE problem
scalably and efficiently using probabilistic graphical models.
The solution involves defining the factor graph
corresponding to (9), and deriving expressions for BP
messages exchanged over the factor graph, as detailed next.
For completeness of exposition, we first present the
solution of the DC SE problem using the BP algorithm; we
refer to the corresponding method as the DC-BP.
Furthermore, and as a novel contribution, we provide an
in-depth convergence analysis of the DC-BP algorithm, and
present methods to improve its convergence. The material in
this section sets the stage for the main contribution of this
paper - the BP-based Gauss-Newton method for AC model
presented in the following section.
A. The Factor Graph Construction
5
bus 1
θ1
2
1 MP12
fP12
The DC model contains only linear measurement functions
which we represent in a general form as:
X
hi (xs , xl , . . . , xL ) = Cxs xs +
Cxb xb ,
(18)
bus 2
θ2
xb ∈Xi \xs
f P3
3
MP 3
θ3
bus 3
(a)
(b)
Fig. 4.
Transformation of the bus/branch model and measurement
configuration (subfigure a) into the corresponding factor graph for the DC
model (subfigure b).
B. Derivation of BP Messages
where Xi \xs is the set of variable nodes incident to the factor
node fi , excluding the variable node xs . From the expression
(3), and using (16)-(18), it can be shown that the message
µfi →xs (xs ) from the factor node fi to the variable node xs is
represented by the Gaussian function (15), with mean zfi →xs
and variance vfi →xs obtained as:
!
X
1
zi −
Cxb zxb →fi
(19a)
zfi →xs =
Cxs
xb ∈Xi \xs
!
X
1
2
vfi →xs = 2 vi +
Cxb vxb →fi .
(19b)
Cxs
xb ∈Xi \xs
Message from a variable node to a factor node: Let us
assume that the incoming messages µfw →xs (xs ), . . . ,
µfW →xs (xs ) into the variable node xs are Gaussian and
represented by their mean-variance pairs (zfw →xs , vfw →xs ),
. . . , (zfW →xs , vfW →xs ). According to (2), it can be shown
that the message µxs →fi (xs ) from the variable node xs to
the factor node fi is proportional (∝) to:
To summarize, after the factor node fi receives the messages
from all of the neighbouring variable nodes from the set Xi \
xs , it evaluates the message µfi →xs (xs ) according to (19a)
and (19b), and sends it to the variable node xs .
Marginal inference: According to (4), it can be shown that
the marginal of the state variable xs is represented by:
µxs →fi (xs ) ∝ N (zxs →fi |xs , vxs →fi ),
p(xs ) ∝ N (x̂s |xs , vxs ),
with mean zxs →fi and variance vxs →fi obtained as:
!
X zfa →xs
vxs →fi
zxs →fi =
vfa →xs
(13)
(14a)
fa ∈Fs \fi
1
vxs →fi
X
=
fa ∈Fs \fi
1
.
vfa →xs
(14b)
(15)
(16)
µxL →fi (xL ) ∝ N (zxL →fi |xL , vxL →fi ).
The Gaussian function associated with the factor node fi is
given by (Section III, equation (6)):
N (zi |xs , xl , . . . , xL , vi )
(
∝ exp
X
1
1
=
.
vx s
vfc →xs
(21b)
fc ∈Fs
Let us assume that the messages into factor nodes are
Gaussian, denoted by:
µxl →fi (xl ) ∝ N (zxl →fi |xl , vxl →fi )
..
.
(21a)
fc ∈Fs
After the variable node xs receives the messages from all of
the neighbouring factor nodes from the set Fs \ fi , it evaluates
the message µxs →fi (xs ) according to (14) and sends it to the
factor node fi .
Message from a factor node to a variable node: Due to
linearity of measurement functions hi (·), closed form
expressions for these messages is easy to obtain and follow a
Gaussian form:
µfi →xs (xs ) ∝ N (zfi →xs |xs , vfi →xs ).
with the mean value x̂s and variance vxs :
!
X zfc →xs
vxs
x̂s =
vfc →xs
(20)
)
[zi − hi (xs , xl , . . . , xL )]2
. (17)
2vi
Finally, the mean-value x̂s is adopted as the estimated value
of the state variable xs .
C. Iterative DC-BP Algorithm
The SE scenario is in general an instance of loopy BP
since the corresponding factor graph usually contains cycles.
Loopy BP is an iterative algorithm, with an iteration index
τ = {0, 1, 2, . . . }, and requires a message-passing schedule.
The scheduling where messages from variable to factor
nodes, and messages from factor nodes to variable nodes, are
updated in parallel in respective half-iterations, is known as
synchronous scheduling. Synchronous scheduling updates all
messages in a given iteration using the output of the
previous iteration as an input [34].
To present the algorithm precisely, we need to introduce
different types of factor nodes. The indirect factor nodes
Find ⊂ F correspond to measurements that measure state
variables indirectly. In the DC scenario, this includes active
power flow and power injection measurements. The direct
factor nodes Fdir ⊂ F correspond to the measurements that
measure state variables directly. For our choice of state
variables for the DC scenario, an example includes
measurements of angles.
6
Besides direct and indirect factor nodes, we define three
additional types of singly-connected factor nodes. The slack
factor node corresponds to the slack or reference bus where
the voltage angle has a given value. The initialization factor
node is required to initialize the algorithm at certain variable
nodes. Finally, the virtual factor node is a is singly-connected
factor node used if the variable node is not directly measured.
Both initialization and virtual factor nodes take the value of
”flat start” with variance σx2i → ∞ or a priori given mean
value and variance of state variables. We refer to direct factor
nodes and three additional types of singly-connected factor
nodes1 as local factor nodes Floc ⊂ F.
Algorithm 1 The DC-BP
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
procedure I NITIALIZATION τ = 0
for each fs ∈ Floc do
(0)
send µfs →xs to incident xs ∈ X
end for
for each xs ∈ X do
(0)
(0)
send µxs →f = µfs →xs , to incident fi ∈ Find
i
end for
end procedure
procedure I TERATION LOOP τ = 1, 2, . . .
while stopping criterion is not met do
for each fi ∈ Find do
(τ )
Compute µf →xs using (19a)*, (19b)*
i
end for
for each xs ∈ X do
(τ )
Compute µxs →f using (14)
i
end for
end while
end procedure
procedure O UTPUT
for each xs ∈ X do
Compute x̂s , vxs using (21)
end for
end procedure
*Incoming messages are obtained in previous iteration τ − 1
Example 2 (Different types of factor nodes). In this example,
we consider the bus/branch model with four measurements
illustrated in Fig. 5a that we use to describe different types
of factor nodes. The corresponding factor graph is given in
Fig. 5b, with indirect factor nodes (red squares), direct factor
nodes (orange squares), the initialization factor node (green
square) and the virtual factor node (blue square).
4
The DC-BP algorithm is presented in Algorithm 1, and it
is an instance of a loopy GBP applied over a linear model
defined by linear measurement functions h(x). It is well
known that, if loopy GBP applied over a linear model
converges, it will converge to a fixed point representing a
solution of an equivalent WLS problem (12) [35]. Unlike
means, the variances of GBP messages need not converge to
1 We note that each variable node in the initialization step of the algorithm
has the corresponding local factor node attached. Local factor nodes only send,
but do not receive, the message to/from incident variable nodes. Direct and
virtual factor nodes repeatedly transmit the same message to the corresponding
variable node throughout BP iterations.
f θ1
1 MP12
2
fθ2
fP12
f θ3
fP23
3
MP23
θ1
θ2
bus 2
fP4
bus 1
θ3
bus 3
Mθ 3
4
f θ4
MP 4
(a)
θ4
bus 4
(b)
Fig. 5.
Transformation of the bus/branch model and measurement
configuration (subfigure a) into the corresponding factor graph with different
types of factor nodes for the DC model (subfigure b)
correct values. In the following, it is sufficient to investigate
the conditions for the DC-BP convergence as in case of
convergence, the obtained mean values are the WLS solution.
D. Convergence of DC-BP Algorithm
In this part, we present convergence analysis of DC-BP
algorithm with synchronous scheduling. In the following, it
will be useful to consider a subgraph of the factor graph that
contains the set of variable nodes X = {x1 , . . . , xN }, the set
of indirect factor nodes Find = {f1 , . . . , fm } ⊂ F, and a set
of edges B ⊆ X × Find connecting them. The number of
edges in this subgraph is b = |B|. Within the subgraph, we
will consider a factor node fi ∈ Find connected to its
neighboring set of variable nodes Xi = {xq , . . . , xQ } ⊂ X
by a set of edges Bi = {bqi , . . . , bQ
i } ⊂ B, where di = |Xi | is
the degree of fi . Next, we provide results on convergence of
both variances and means of DC-BP messages, respectively.
Convergence of the Variances: From equations (14b) and
(19b), we note that the evolution of the variances is
independent of mean values of messages and measurements.
Let vs ∈ Rb denote a vector of variance values of messages
from indirect factor nodes Find to variable nodes X . Note
that this vector can be decomposed as:
(τ )
(τ ) T
vs(τ ) = [vs,1 , . . . vs,m
] ,
(22)
where the i-th element vs,i ∈ Rdi is equal to:
(τ )
(τ )
(τ )
vs,i = [vfi →xq , . . . , vfi →xQ ].
(23)
Substituting (14b) in (19b), the evolution of variances vs is
equivalent to the following iterative equation:
(24)
vs(τ ) = f vs(τ −1) .
More precisely, using simple matrix algebra, one can obtain
the evolution of the variances vs in the following matrix form:
h
i
e −1 ΠC
e · D(A) −1 + Σa C
e −1 i
vs(τ ) = C
(25)
where
e = CCT
C
A=
T
ΓΣ−1
s Γ
(26a)
+ L.
(26b)
7
(τ −1)
Note that in (25), the dependance on vs
is hidden in matrix
A, or more precisely, in matrix Σs . Next, we briefly describe
both the matrices and matrix-operators involved in (25).
The operator D(A) ≡ diag(A11 , . . . , Abb ), where Aii is
the i-th diagonal entry of the matrix A. The unit vector i is
of dimension b and is equal to i = [1, . . . , 1]T . The diagonal
(τ −1)
matrix Σs is obtained as Σs = diag vs
∈ Rb×b .
The matrix C = diag C1 , . . . , Cm ∈ Rb×b contains
diagonal entries of the Jacobian non-zero elements, where
i-th element Ci = [Cxq, . . . , CxQ ] ∈ Rdi . The matrix
Σa = diag Σa,1 , · · · Σa,m ∈ Rb×b contains indirect factor
node variances, with the i-th entry Σ
[vi , . . . , vi ] ∈ Rdi .
a,i =b×b
The matrix L = diag L1 , · · · Lm ∈ R
contains inverse
variances from singly-connected factor nodes to a variable
node, if such nodes exist, where the i-th element
Li = lxq , · · · lxQ ∈ Rdi . For example, lxq equals:
1
, if xq is incident to fd,q
(27)
lxq = vfd,q →xq
0,
otherwise.
The matrix Π = diag Π1 , . . . Πm ∈ Fb×b
2 , F2 = {0, 1},
is a block-diagonal matrix in which the i-th element is a block
matrix Πi = 1i − Ii ∈ Fd2i ×di , where the matrix 1i is di × di
block matrix of ones, and Ii is di × di identity matrix. The
matrix Γ ∈ Fb×b
is of the following block structure:
2
01,1 Γ1,2 . . . Γ1,m
Γ2,1 02,2 . . . Γ2,m
Γ= .
(28)
,
..
..
..
.
.
Γm,1 Γm,2 . . . 0m,m
d ×d
where 0i,i is a block matrix di ×di of zeros, and Γi,j ∈ F2i j
with the (i, j)-th entry:
(
1, if both bqi and bqj are incident to xq
Γi,j (i, j) =
(29)
0, otherwise.
Note that the following holds: Γj,i =
ΓT
i,j .
Theorem 1. The variances vs from indirect factor nodes to
variable nodes always converge to a unique fixed point
(τ )
(τ =0)
limτ →∞ vs = v̂s for any initial point vs
> 0.
Proof. The theorem can be proved by showing that f vs
satisfies the conditions of the so-called standard function [36],
following similar steps as in the proof of Lemma 1 in [37].
(31d)
Note that the vector of means zs ∈ Rb can be decomposed as:
(τ )
)
) T
z(τ
= [zs,1 , . . . , z(τ
s
s,m ] ,
(32)
where the i-th element zs,i ∈ Rdi is equal to:
(τ )
(τ )
(τ )
(33)
zs,i = [zfi →xk , . . . , zfi →xK ].
T
The vector za = za,1 , · · · za,m ∈ Rb contains means of
indirect factor nodes, where za,i = [zi , . . . , zi ] ∈ Rdi . The
diagonal matrix Σ̂s ∈ Rb×b is obtained as Σ̂s =
(τ )
limτ →∞ Σs . The vector zb = zb,1 , · · · , zb,m ∈ Rb
contains means from direct and virtual factor nodes to a
variable node, if such
nodes exist, where the i-th element
zb,i = zxk , · · · zxK ∈ Rdi . For example, the element zxk of
zb,i is equal to:
(
zfd,k →xk , if xk is incident to fd,k
zxk =
(34)
0,
otherwise.
Theorem 2. The means zs from indirect factor nodes to
variable nodes converge to a unique fixed point
(τ )
limτ →∞ zs = ẑs :
−1
(35)
e
ẑs = I + Ω
z,
(τ =0)
for any initial point zs
ρ(Ω) < 1.
if and only if the spectral radius
Proof. The proof steps follow the proof of Theorem 5.2, [36].
To summarize, the convergence of the DC-BP algorithm
depends on the spectral radius of the matrix:
−1
T
Ω = C−1 ΠC · D(ΓΣ̂−1
· ΓΣ̂−1
. (36)
s Γ + L)
s
If the spectral radius ρ(Ω) < 1, the DC-BP algorithm will
converge and the resulting vector of mean values will be equal
to the solution of the MAP estimator.
E. Convergence of DC-BP with Randomized Damping
(31a)
(31b)
)
(τ −1)
)
zd = z(τ
+ α2 z(τ
q + α1 zr
r ,
(30)
where
−1
e
z = C−1 za − D · D(Â)
· Lzb
−1
−1
Ω = D · D(Â)
· ΓΣ̂s
(31c)
D = C−1 ΠC.
In this section, we propose an improved DC-BP algorithm
that applies synchronous scheduling with randomized
damping. Several previous works reported that damping the
BP messages improves the convergence of BP [37], [38].
Here, we propose a different randomized damping approach,
where each mean value message from indirect factor node to
a variable node is damped independently with probability p,
otherwise, the message is calculated as in the standard
DC-BP algorithm. The damped message is evaluated as a
linear combination of the message from the previous and the
current iteration, with weights α1 and 1 − α1 , respectively.
In Section VI, we demonstrate that the DC-BP with
randomized damping dramatically improves convergence as
compared to the standard DC-BP.
In the proposed damping, the equation (30) is redefined as:
Convergence of the Means: Equations (14a) and (19a)
show that the evolution of the mean values depends on the
variance values. Due to Theorem 1, it is possible to simplify
evaluation of mean values zs from indirect factor nodes Find
to variable nodes X by using the fixed-point values of v̂s .
The evolution of means zs becomes a set of linear equations:
)
−1)
z(τ
=e
z − Ωz(τ
,
s
s
T
 = ΓΣ̂−1
s Γ +L
(τ )
(37)
8
where 0 < α1 < 1 is the weighting coefficient, and α2 =
(τ )
(τ )
1 − α1 . In the above expression, zq and zr are obtained
as:
)
−1)
z(τ
z − QΩz(τ
q = Qe
s
)
z(τ
r
= Re
z−
−1)
RΩz(τ
,
s
(38a)
(38b)
where diagonal matrices Q ∈ Fb×b
and R ∈ Fb×b
are
2
2
defined as Q = diag(1 − q1 , ..., 1 − qb ), qi ∼ Ber(p), and
R = diag(q1 , ..., qb ), respectively, where Ber(p) ∈ {0, 1} is
a Bernoulli random variable with probability p independently
sampled for each mean value message.
Substituting (38a) and (38b) in (37), we obtain:
−1)
(τ )
zd = Q + α2 R e
z − QΩ + α2 RΩ − α1 R z(τ
.
s
(39)
(τ −1)
(τ −1)
Note that zr
= Rzs
. In a more compact form,
equation (39) can be written as follows:
(τ )
−1)
zd = z̄ − Ω̄z(τ
,
s
(40)
where
z̄ = Q + α2 R e
z
(41a)
Ω̄ = QΩ + α2 RΩ − α1 R.
(41b)
Theorem 3. The means zd from indirect factor nodes to
variable nodes converge to a unique fixed point
(τ )
(τ =0)
ẑd = limτ →∞ zd for any initial point zd
if and only if
the spectral radius ρ(Ω̄) < 1. For the resulting fixed point, it
holds that ẑd = ẑs .
Proof. To prove theorem it is sufficient to show that equation
(40) converges to the fixed point defined in (35). We can write:
zr
(τ −1)
= Re
z−
−2)
RΩz(τ
.
s
(42)
Substituting (38a), (38b) and (42) in (37):
(τ )
z
zd = Q + α2 R + α1 R e
−1)
− QΩ + α2 RΩ z(τ
− α1 RΩzs(τ −2) . (43)
s
(τ )
The fixed point ẑd = limτ →∞ zd
is equal to:
−1
ẑd = I + QΩ + α2 RΩ + α1 RΩ
· Q + α2 R + α1 R e
z. (44)
From definitions of Q, R and α2 , we have QΩ + α2 RΩ +
α1 RΩ = Ω and Q + α2 R + α1 R = I, thus (44) becomes:
−1
(45)
e
ẑd = I + Ω
z.
This concludes the proof.
V. BP-BASED D ISTRIBUTED G AUSS -N EWTON M ETHOD
FOR AC SE
The rest of the paper focuses on the more challenging AC
model, where our goal is to derive the BP-based AC SE
algorithm. However, due to non-linearity of measurement
functions, the closed-form expressions for certain classes of
BP messages cannot be obtained. Using approximations, we
derive the AC-BP algorithm as an (approximate) BP solution
for the AC SE problem. Due to space constraints, we refer
the reader to Appendix C, for a detailed derivation of the
AC-BP algorithm. Unfortunately, due to approximations, the
AC-BP algorithm does not match the performance of the
centralized AC SE based on Gauss-Newton method (as we
show in Section VI).
As the main contribution of this paper, we adopt different
methodology to derive efficient BP-based AC SE method.
Namely, instead of using the AC-BP method, we present the
solution where BP is applied sequentially over the AC
model, akin to what is done by the Gauss-Newton method.
The resulting GN-BP algorithm represents a BP counterpart
of the Gauss-Newton method, achieving exactly the same
accuracy as the Gauss Newton method.
A. Gauss-Newton Method as a Sequential MAP Problem
Consider the Gauss-Newton method (11) where, at each
iteration step ν, the algorithm returns a new estimate of x
denoted as x(ν) . Note that, after a given iteration, an
estimate x(ν) is a vector of known (constant) values. If the
Jacobian matrix J(x(ν) ) has a full column rank, the equation
(11a) represents the linear WLS solution of the minimization
problem [39]:
min ||W1/2 [r(x(ν) ) − J(x(ν) )∆x(ν) ]||22 .
∆x(ν)
(46)
Hence, at each iteration ν, the Gauss-Newton method produces
WLS solution of the following system of linear equations:
r(x(ν) ) = g(∆x(ν) ) + u,
(47)
where g(∆x(ν) ) = J(x(ν) )∆x(ν) comprises linear functions,
while u is the vector of measurement errors. The equation
(11a) is the weighted normal equation for the minimization
problem defined in (46), or alternatively, equation (11a) is a
WLS solution of (47). Consequently, the probability density
function associated with the i-th measurement (i.e., the i-th
residual component ri ) at any iteration step ν:
N (ri (x(ν) )|∆x(ν) , vi )
1
exp
=√
2πvi
(
)
[ri (x(ν) ) − gi (∆x(ν) )]2
. (48)
2vi
The MAP solution of (9) can be redefined as an iterative
optimization problem where, instead of solving (11), we solve
the MAP (sub)problem:
∆x̂(ν) = arg max L r(x(ν) )|∆x(ν)
∆x(ν)
= arg max
∆x(ν)
k
Y
N ri (x(ν) )|∆x(ν) , vi
i=1
(ν)
x(ν+1) = x(ν) + ∆x̂
.
(49a)
(49b)
Next, we show that the solution of the above MAP
sub-problem (49a) over increment variables ∆x(ν) can be
efficiently obtained using the BP algorithm applied over the
underlying factor graph.
9
Note that, in general, the resulting BP algorithm represents
an instance of a loopy GBP over a linear model defined by
linear functions g(∆x(ν) ). Thus if it converges, it provides a
solution equal to the linear WLS solution ∆x(ν) of (11a).
The BP solution of ∆x̂(ν) in each iteration ν (outer iteration
loop) is obtained via iterative BP algorithm (inner iteration
loops). Every inner BP iteration τ = {0, 1, 2, . . . η(ν)}
outputs ∆x̂(ν,τ ) , where η(ν) is the number of inner BP
iterations within the outer iteration ν.
Marginal inference: It can be shown that the marginal of
the state variable ∆xs , according to (4), is represented by the
Gaussian function:
p(∆xs ) ∝ N (∆x̂s |∆xs , v∆xs ),
(55)
with mean ∆x̂s which represents the estimated value of the
state variable increment ∆xs and variance v∆xs :
!
X rfc →∆xs
v∆xs
(56a)
∆x̂s =
vfc →∆xs
fc ∈Fs
1
B. The Factor Graph Construction
From the factorization of the likelihood expression (49a),
one easily obtains the factor graph corresponding to the
GN-BP method as follows. The increments ∆x of state
variables x determine the set of variable nodes
X = {(∆θ1 , ∆V1 ), . . . , (∆θN , ∆VN )} and each likelihood
function N (ri (xν )|∆xν , vi ) represents the local function
ψ(Xi ). Since the residual equals ri (xν ) = zi − hi (xν ), in
general, the set of factor nodes F = {f1 , . . . , fk } is defined
by the set of measurements M. The factor node fi connects
to the variable node ∆xs ∈ {∆θs , ∆Vs } iff the increment of
the state variable ∆xs is an argument of the corresponding
function gi (∆x), i.e., if the state variable xs ∈ {θs , Vs } is an
argument of the measurement function hi (x).
C. Derivation of BP Messages
Message from a variable node to a factor node: It is easy
to show that the message µ∆xs →fi (∆xs ) is proportional to:
µ∆xs →fi (∆xs ) ∝ N (r∆xs →fi |∆xs , v∆xs →fi ),
(50)
with mean r∆xs →fi and variance v∆xs →fi obtained as:
!
X rfa →∆xs
v∆xs →fi
(51a)
r∆xs →fi =
vfa →∆xs
fa ∈Fs \fi
1
v∆xs →fi
1
X
=
fa ∈Fs \fi
vfa →∆xs
.
(51b)
Message from a factor node to a variable node: Similarly,
the message µfi →∆xs (∆xs ) is proportional to:
µfi →∆xs (∆xs ) ∝ N (rfi →∆xs |∆xs , vfi →∆xs ),
(52)
with mean rfi →∆xs and variance vfi →∆xs obtained as:
rfi →∆xs =
vfi →∆xs =
1
C∆xs
1
2
C∆x
s
!
ri −
X
C∆xb · r∆xb →fi
(53a)
∆xb ∈Xi \∆xs
!
vi +
X
2
C∆x
b
· v∆xb →fi .
(53b)
∆xb ∈Xi \∆xs
The coefficients C∆xp , ∆xp ∈ Xi , are Jacobian elements
of the measurement function associated with the factor node
fi (see Appendix A and D for details):
C∆xp =
∂h(xs , xl , . . . , xL )
.
∂xp
(54)
v∆xs
=
1
X
fc ∈Fs
vfc →∆xs
.
(56b)
To summarize, the MAP sub-problem defined in (49a) can
be efficiently solved using (51), (53) and (56).
D. Iterative GN-BP Algorithm
To introduce the GN-BP algorithm, we use the same
classification of factor nodes and the same message-passing
schedule as in Section IV-C for the DC-BP algorithm. The
slack factor node is specified as a node where the voltage
angle has a given value, therefore, the residual of the
corresponding state variable is equal to zero, and its variance
tends to zero. Residuals of initialization and virtual factor
nodes approach zero, while their variances tend to infinity. It
is important to note that local factor nodes send messages
represented by a triplet: mean (of the residual), variance and
the state variable value (see Appendix D, for details).
The GN-BP algorithm is presented in Algorithm 2. After
the initialization, the outer loop starts by computing residuals
for direct and indirect factor nodes, as well as the Jacobian
elements, and passes them to the inner iteration loop. The
inner iteration loop represents the main algorithm routine
which includes BP-based message inference described in the
previous subsection. The output of the inner iteration loop is
the estimate of the state variable increments. Finally, the
outer loop updates the set of state variables. The outer loop
iterations are repeated until the stopping criteria is met.
E. Convergence of the GN-BP Algorithm
The GN-BP in the inner iteration loop solves the system
of linear equations (47). By directly applying the results from
Section IV-D of the (linear) DC-BP convergence analysis, the
convergence of the GN-BP with synchronous scheduling in
each outer iteration loop ν depends on the spectral radius of
the matrix:
Ω(x(ν) ) = C(x(ν) )−1 ΠC(x(ν) )
−1
T
· D(ΓΣ̂−1
· ΓΣ̂−1
, (57)
s Γ + L)
s
where C(x(ν) ) contains diagonal entries of the Jacobian
J(x(ν) ) non-zero elements in every outer iteration ν.
According to Theorem 2, the GN-BP with synchronous
scheduling converges to a unique fixed point if
ρ(Ω(x(ν) )) < 1, ν = {0, 1, 2, . . . }. We note that this
10
Algorithm 2 The GN-BP
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
Although similar results can be obtained for the GN-BP
algorithm (see Section V-E), for brevity, we present the
convergence results only for the DC-BP case.
For evaluation of accuracy, we focus on the proposed BPbased AC SE methods and compare them to the Gauss-Newton
method (the centralized SE algorithm). More precisely, we
first consider the IEEE 14 bus test case and provide a detailed
performance comparison between the AC-BP and the GN-BP
algorithms. Then, due to favorable performance, we provide
an in-depth evaluation of the GN-BP algorithm using the IEEE
14 bus test case, providing the results for IEEE 30 and IEEE
118 bus test cases in Appendix F.
procedure I NITIALIZATION ν = 0
for Each xs ∈ X do
(0)
Initialize value of xs , typically as a ”flat start”
end for
end procedure
procedure O UTER ITERATION LOOP ν = 0, 1, 2, . . . ; τ = 0
while stopping criterion for the outer loop is not met do
for Each fs ∈ Fdir do
(ν)
(ν)
compute residual rs = zs − xs
end for
for Each fs ∈ Floc do
(ν)
(ν)
send µfs →∆xs , xs to incident ∆xs ∈ X
end for
for Each ∆xs ∈ X do
(ν)(τ =0)
(ν)
(ν)
send µ∆xs →f = µfs →∆xs , xs to incident fi ∈ Fid
i
end for
for Each fi ∈ Find do
(ν)
compute residual ri = zi − hi (x(ν) )
(ν)
compute Jacobian elements Ci,∆xp ; ∆xp ∈ Xi
end for
procedure I NNER I TERATION LOOP τ = 1, 2, . . .
while stopping criterion for the inner loop is not met do
for Each fi ∈ Find do
(τ )
Compute µf →∆xs using (53)*
i
end for
for Each ∆xs ∈ X do
(τ )
Compute µ∆xs →f using (51)
i
end for
end while
end procedure
for Each ∆xs ∈ X do
(ν)
compute ∆x̂s using (56)
(ν)
(ν)
(ν+1)
update xs
= xs + ∆x̂s
end for
end while
end procedure
*Incoming messages are obtained in previous iteration τ − 1
A. Simulation Setup
convergence condition is required to hold in every outer
iteration ν.
Similarly, the convergence of the GN-BP with randomized
damping in every outer iteration loop ν is governed by the
spectral radius of the matrix:
Ω̄(x(ν) ) = QΩ(x(ν) ) + α2 RΩ(x(ν) ) − α1 R.
In the simulated model, we start with a given IEEE bus
test case and apply the power flow analysis to generate the
exact solution. In case of the DC model, we apply the DC
power flow analysis to calculate voltage angles and active
powers, while for the AC model, we apply the AC power
flow analysis to generate currents, voltages and powers
across the network. Further, we corrupt the exact solution by
the additive white Gaussian noise of variance v and we
observe the set of measurements.
In the DC model, the measurements contain active power
flows and power injections, and bus voltage angles, while in
the AC model, they contain active and reactive power flows,
active and reactive power injections, bus voltage magnitudes
and bus voltage angles. The set of measurements is selected
in such a way that the system is observable. More precisely,
for each scenario, we generate 1000 random measurement
configurations with the number of measurements equal either
to double or triple the size of the number of state variables
(i.e., we consider the redundancy to be equal 2 or 3).
To evaluate the performance of the BP-based SE algorithms,
we convert each of the above randomly generated IEEE bus
test cases with a given measurement configuration into the
corresponding factor graph and we run the DC-BP, AC-BP or
GN-BP algorithms over the factor graph. Note that, in order
to initialize the AC-BP, the GN-BP, and the Gauss-Newton
method, we use the ”flat start” assumption2 (Vi = 1, θi = 0,
i = 1, . . . , N ).
B. Convergence Analysis of DC-BP Algorithm
(58)
According to Theorem 3, in every outer iteration ν, the
algorithm will converge to a unique fixed point if
ρ(Ω̄(x(ν) )) < 1, ν = {0, 1, 2, . . . }, and the resulting fixed
point is equal to the fixed point obtained by the GN-BP with
synchronous scheduling.
VI. N UMERICAL R ESULTS
In this section, using numerical simulations, we analyze the
convergence and the accuracy of the BP-based SE methods.
For convergence evaluation, we focus on the DC-BP
algorithm whose convergence in analyzed in Sections IV-D
and IV-E. We verify these results via numerical simulations.
As detailed in Sec. IV-D and IV-E, the DC-BP with
synchronous scheduling with or without randomized
damping will converge if ρ(Ω) < 1 and ρ(Ω̄) < 1,
respectively. This condition is verified in our simulations,
thus we present the convergence performance by comparing
spectral radii of matrices Ω and Ω̄.
Figure 6 shows spectral radii for different redundancies for
IEEE 14 and IEEE 118 bus test case. For each scenario, the
randomized damping case behaves superior in terms of the
2 Usually, IEEE bus test cases include in-phase transformers with zero series
resistances. Consequently, the ”flat start” will cause undefined messages due to
the corresponding Jacobian elements being equal to zero. This problem can be
alleviated by initializing bus voltage angles with a small random perturbation
at transformers’ buses.
11
1.0
Synchronous
Scheduling
0.9
Randomized
0.8
Damping
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4
Spectral Radius ρ
Empirical CDF
Empirical CDF
spectral radius. As an interesting and somewhat extreme
case, for the IEEE 118 bus test case, the DC-BP algorithm
with synchronous scheduling could not converge at all, while
with randomized damping3 , we recorded convergence with
probability above 0.9. As expected, the algorithm with
randomized damping performs better for larger redundancy.
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.8
Synchronous
Scheduling
Randomized
Damping
0.9
1.0
1.1
Spectral Radius ρ
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
(b)
0.8950
Empirical CDF
Empirical CDF
(a)
0.8945
0.8940
0.8935
1.0
Synchronous
Scheduling
Randomized
Damping
1.0
1.2
1.1 1.2 1.3 1.4
Spectral Radius ρ
1.5
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.9935
0.9930
0.9925
0.9920
1.0
Note that WRSS is the value of the objective function of the
optimization problem (10) we are solving, thus it is suitable
metric for the SE accuracy. Finally, we normalize the obtained
WRSS by WRSSWLS of the centralized SE obtained using the
Gauss-Newton method after 12 iterations (which we adopt as
a normalization constant). This way, we compare the accuracy
of BP-based algorithms to the one of the centralized SE.
Besides WRSS, we use the mean absolute difference (MAD)
between the state variables in two consecutive iterations:
n
1X
|∆xi |.
MAD =
(60)
n i=1
The MAD value represents average component-wise shift of
the state estimate over the iterations, thus it may be used to
quantify the rate of convergence.
The IEEE 14 bus test case with fixed measurement
configuration containing 61 measurement devices, as shown
in Fig. 7, is used to compare the performance of the
BP-based SE algorithms. For each value of noise variance
v = {v1 , v2 , v3 , v4 } = {10−10 , 10−8 , 10−6 , 10−4 } p.u.,
using Monte Carlo approach, we generate 1000 random sets
of measurement values and feed them to the proposed
BP-based SE algorithms.
13
Synchronous
Scheduling
Randomized
Damping
1.0
1.1 1.2 1.3 1.4
Spectral Radius ρ
(c)
14
1.5
12
11
(d)
10
Fig. 6. The spectral radius of matrices Ω for synchronous scheduling and
Ω̄ for randomized damping for redundancy equal 2 for IEEE 14 (subfigure
a) and IEEE 118 (subfigure c) bus test case and for redundancy equal 3 for
IEEE 14 (subfigure b) and IEEE 118 (subfigure d) bus test case.
In scenarios when the standard DC-BP algorithm did not
converge (i.e., ρ(Ω) > 1), the algorithm with randomized
damping had smaller spectral radius, i.e., ρ(Ω̄) < ρ(Ω), and
moreover, the spectral radius was usually smaller than one
(ρ(Ω̄) < 1). On the other hand, when DC-BP converged
(ρ(Ω) < 1), then in most cases the randomized damping
method also converged, albeit with slightly higher spectral
radius (ρ(Ω) < ρ(Ω̄) < 1). These results demonstrate the
benefit of randomized damping for improved DC-BP
stability and convergence4 .
C. Performance Evaluation of AC-BP and GN-BP
In the following, compare the performance of the AC-BP
and GN-BP algorithms via extensive numerical study (the ACBP is presented in detail in Appendix B and C). We use the
weighted residual sum of squares (WRSS) as a metric:
WRSS =
k
X
[zi − hi (x)]2
i=1
σi2
.
(59)
3 Note that randomized damping parameters, which are optimized via
numerical experiments in Sec. VI-D, are set to p = 0.6 and α1 = 0.5.
4 Another trick for improving the convergence behavior is to connect virtual
factor nodes to all variable nodes that do not already have direct factor nodes.
6
9
1
7
5
4
8
2
3
Active and Reactive Power Flow
Injection Active and Reactive Power
Voltage Magnitude and Angle
Fig. 7. The IEEE 14 bus test case with given measurement configuration.
For both algorithms, we apply randomized damping due to
improved convergence, and initially, we set the randomized
damping parameters to p = 0.6 and α1 = 0.5. The GN-BP is
configured so that, for every outer iteration ν, the number of
inner iterations grows exponentially as η(ν) = ν e , where e is
the inner iteration exponent which we set to e = 4 [40]5 .
In order to present average performance results, we use box
plots in which we: i) set the minimum and the maximum value
to 0% and 90% of the obtained simulation results, ii) the box
is defined by the first and the third quartile, iii) the middle line
is the median value, and iv) the results above the maximum
value are represented as outliers.
Fig. 8 shows the weighted residual sum of squares of the
AC-BP WRSSτBP over the iterations τ , normalized by
5 Note that we perform detailed performance analysis as a function of
damping parameters in [40], Appendix E, and as a function of the number of
inner iterations in Section VI-D.
WRSSBP/WRSSWLS
WRSSτBP/WRSSWLS
1.25
1.20
1.15
1.10
1.05
WRSSBP/WRSSWLS
12
1.00003
1.00002
1.00001
1.00000
100
300
500
700
1.0003
1.0002
1.0001
1.0000
AC - BP
1.00
1.0005
1.0004
GN - BP
AC - BP
(a)
900 1100 1300 1500 1700 1900 2100 2275
Number of iterations τ
GN - BP
(b)
1.4
1.3
1.002
1.3
1.2
1.1
1.0
AC - BP
GN - BP
500
700
900 1100 1300 1500 1700 1900 2100 2275
Number of iterations τ
Fig. 8. The AC-BP normalized WRSS (i.e., WRSSτBP /WRSSWLS ) for the
low noise level v1 (subfigure a) and the high noise level v4 (subfigure b).
WRSSWLS (i.e., WRSSτBP /WRSSWLS ). We observe that the
AC-BP converges for both the low and the high noise level,
however, for the high noise level, the solution of the AC-BP
algorithm does not correspond to the solution of the
centralized SE. This is expected, since as the noise variance
increases, the accuracy of Gaussian approximation of the BP
messages is decreasing, which affects the accuracy of the
AC-BP solution.
Fig. 9 shows WRSSνBP /WRSSWLS of the GN-BP over the
outer iterations ν. As shown, the GN-BP algorithm converges
to the solution of the centralized SE for both the low and the
high noise level.
Number of inner iterations τ
979 2275 4676 8722 15333 25333 39974 60710
WRSSνBP/WRSSWLS
Number of inner iterations τ
979 2275 4676 8722 15333 25333 39974 60710
2.0
1.5
6
7
8
9 10 11 12
Number of outer iterations ν
Fig. 10. The normalized WRSS for the AC-BP and GN-BP after 2275
iteration for variances v1 (subfigure a), v2 (subfigure b), v3 (subfigure c)
and v4 (subfigure d).
addition, in terms of implementation, the AC-BP suffers two
drawbacks as compared to the GN-BP: i) the AC-BP
messages have considerably more complex form, and ii) the
AC-BP requires prior knowledge (e.g., historical data).
Consequently, in the remaining part of this section, we focus
on the performance of the GN-BP.
D. Design of Inner Iteration Schemes
In order to reduce the number of non-converging
simulations due to an insufficient number of inner iterations,
in this section we test the GN-BP convergence performance
with different inner iteration schemes. The damping
parameters, investigated in detail in Appendix E, are set to
p = 0.4 and α1 = 0.3. As a reference point, the MAD values
of the centralized SE for variances v1 and v4 are shown in
Fig. 11.
1.004
1.003
1.002
1.001
1.000
1.0
GN - BP
(d)
5
6
7
8
9 10 11 12
Number of outer iterations ν
(b)
(a)
WRSSν
Fig. 9. The GN-BP normalized WRSS (i.e.,
BP /WRSSWLS ) for the
low noise level v1 (subfigure a) and the high noise level v4 (subfigure b).
To compare BP-based algorithms, we consider WRSSBP
of the BP-based algorithms after the same total number of
iterations. More precisely, the AC-BP algorithm is terminated
after τ = 2275 iterations, while the GN-BP is terminated
after ν = 5 outer iterations, which corresponds exactly to the
total of 2275 inner iterations. Fig. 10 demonstrates that, for
the low noise level, the AC-BP converges faster then the
GN-BP. However, as the noise level increases, the GN-BP
converges faster than the AC-BP.
To summarize, the GN-BP is able to reach exactly the
same solution as the centralized SE, while the AC-BP
generally fails to achieve it for higher noise levels. In
10−2
10−2
10−4
10−4
10−6
10−6
MAD
300
(b)
WRSSνBP/WRSSWLS
1.004
1.4
(c)
1.1
100
5
1.006
AC - BP
1.2
1.0
2.5
1.008
1.000
MAD
WRSSτBP/WRSSWLS
1.5
WRSSBP/WRSSWLS
WRSSBP/WRSSWLS
(a)
10−8
10−10
10−8
10−10
10−12
10−12
10−14
10−14
10−16
10−16
1 2 3 4 5 6 7 8 9 10 11 12
Number of iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of iterations ν
(a)
(b)
Fig. 11. The MAD values of the centralized SE for the low noise level v1
(subfigure a) and high noise level v4 (subfigure b).
For comparison, the MAD values of the GN-BP with the
exponential inner iteration scheme (η(ν) = ν e , e = 4) is
shown in Fig. 12. Note that the GN-BP exhibits comparable
convergence performance to the centralized SE algorithm.
Note also that it is difficult to directly compare the two, due
to a large difference in computational loads of a single
(outer) iteration. For example, the complexity of a single
iteration remains constant but significant (due to matrix
13
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
1
10−2
10−2
10−4
10−4
10−6
10−6
10−10
10−4
10−6
10−6
10−8
10−10
10−12
10−12
10−14
10−16
10−16
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(a)
(b)
10−10
10−12
10−12
10−14
10−16
10−16
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(a)
(b)
Fig. 12. The MAD values of the GN-BP with the exponential inner iteration
scheme for the low noise level v1 (subfigure a) and high noise level v4
(subfigure b).
In the following, we define an alternative accuracy-based
inner iteration scheme. Namely, instead of prescribing the
number of inner iterations in advance, we let the number of
inner iterations evolve until a certain accuracy-based
criterion is met. More precisely, the algorithm in the inner
iteration loop is running until the following convergence
criterion is reached:
(ν,τ −1
|rf →∆x − rf →∆x )| < (ν),
(61)
where rf →∆x represents the vector of mean-value messages
from factor nodes to variable nodes, and (ν) is the
threshold at iteration ν. Note that, due the fact that the BP is
a distributed algorithm, imposing this condition in real
systems would require additional communication overhead.
Fig. 13 shows MAD values and the total number of inner
iterations for the threshold (ν) = [10−3 , 10−6 , . . . , 10−15 ,
10−16 , . . . , 10−16 ], where the number of outer iterations is
ν = 1, . . . , 12. Comparing the results with the centralized SE
(Fig. 11), we note that the GN-BP has almost identical
convergence rate. We note that the proposed accuracy-based
inner iteration scheme dramatically reduces the number of
non-converging simulations.
To summarize, we have demonstrated that by using
different inner iteration schemes, one can influence and
control the trade-off between the GN-BP convergence rate
and the total number of inner iterations. For additional
results on scalability and complexity of GN-BP, we refer the
reader to Appendix F.
VII. C ONCLUSIONS
In this paper, we presented an in-depth study of the
application of the BP algorithm to the SE problem in power
systems. We provided detailed derivation, convergence and
performance analysis of BP-based SE algorithms for both
DC and AC model. The main contribution of our study is
the GN-BP algorithm, which is shown to represent a
BP-based implementation of the iterative Gauss-Newton
method. In a thorough numerical study, we have shown how
10−8
10−10
10−14
10−8
10−14
(ν,τ )
10−2
10−4
200000
200000
100000
60710
39974
25333
15333
8722
4676
100000
60710
39974
25333
15333
8722
4676
Number of inner iterations τ
10−8
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
Number of inner iterations τ
MAD
MAD
1
Number of inner iterations τ
10−2
MAD
Number of inner iterations τ
MAD
inversion) over iterations for the centralized SE algorithm,
while it gradually increases for the GN-BP starting from an
extremely low complexity at initial outer iterations.
2275
979
354
98
2275
979
354
98
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(c)
(d)
Fig. 13. The MAD values and the number of inner iterations for the low noise
level v1 (subfigure a and subfigure c) and the high noise level v4 (subfigure
b and subfigure d) of the GN-BP with threshold (ν) = [10−3 , 10−6 , . . . ,
10−15 , 10−16 , . . . , 10−16 ].
to fine-tune the GN-BP implementation and balance between
the accuracy and the convergence rate. In our future work,
we will investigate extensions to asynchronous BP in
dynamic and real-time state estimation, including suitable
models for communication constraints.
APPENDIX A
M EASUREMENT F UNCTIONS AND JACOBIAN E LEMENTS
In the following, we present measurement functions and
corresponding Jacobian elements for both, the AC and DC SE
model. Note that, for simplicity, we assume that the model
does not contain phase-shifting transformers.
AC SE Model
The active and reactive power flow at the branch (i, j) ∈ E
that connects buses i and j:
hPij (·) = Vi2 (gij + gsi ) − Vi Vj (gij cos θij + bij sin θij )
(62a)
hQij (·) = −Vi2 (bij + bsi ) − Vi Vj (gij sin θij − bij cos θij ),(62b)
where Vi and Vj are bus voltage magnitudes, while θij =
θi −θj is the bus voltage angle difference between bus voltage
angles at buses i and j. The parameters in above equations
include the conductance gij and susceptance bij of the branch,
as well as the conductance gsi and susceptance bsi of the
branch shunt element connected at the bus i.
14
The Jacobian expressions corresponding to hPij (·) are:
∂hPij (·)
= Vi Vj (gij sinθij − bij cosθij )
∂θi
∂hPij (·)
= −Vi Vj (gij sinθij − bij cosθij )
∂θj
∂hPij (·)
= −Vj (gij cosθij + bij sinθij ) + 2Vi (gij + gsi )
∂Vi
∂hPij (·)
= −Vi (gij cosθij + bij sinθij ).
∂Vj
The Jacobian expressions corresponding to hQij (·) are as
follows:
∂hQij (·)
= −Vi Vj (gij cosθij + bij sinθij )
∂θi
∂hQij (·)
= Vi Vj (gij cosθij + bij sinθij )
∂θj
∂hQij (·)
= −Vj (gij sinθij − bij cosθij ) − 2Vi (bij + bsi )
∂Vi
∂hQij (·)
= −Vi (gij sinθij − bij cosθij ).
∂Vj
Jacobian expressions corresponding to hPi (·) are:
X
∂hPi (·)
= Vi
Vj (−Gij sinθij + Bij cosθij )
∂θi
j∈Hi \i
∂hPi (·)
= Vi Vj (Gij sinθij − Bij cosθij )
∂θj
X
∂hPi (·)
=
Vj (Gij cosθij + Bij sinθij ) + 2Vi Gii
∂Vi
j∈Hi \i
∂hPi (·)
= Vi (Gij cosθij + Bij sinθij ),
∂Vj
where Hi \ i is the set of buses adjacent to the bus i. The
Jacobian expressions corresponding to hQi (·) are:
X
∂hQi (·)
= Vi
Vj (Gij cosθij + Bij sinθij )
∂θi
j∈Hi \i
∂hQi (·)
= Vi Vj (−Gij cosθij − Bij sinθij )
∂θj
X
∂hQi (·)
=
Vj (Gij sniθij − Bij cosθij ) − 2Vi Bii
∂Vi
j∈Hi \i
∂hQi (·)
= Vi (Gij sinθij − Bij cosθij ).
∂Vj
The current magnitude at the branch (i, j) ∈ E that
connects buses i and j:
hIij (·) = [Ac Vi2 + Bc Vj2
− 2Vi Vj (Cc cos θij − Dc sin θij )]1/2 . (63)
Ac = (gij + gsi )2 + (bij + bsi )2 ;
2
Bc = gij
+ b2ij
Cc = gij (gij + gsi ) + bij (bij + bsi ); Dc = gij bsi − bij gsi .
The Jacobian expressions corresponding to current magnitude
measurement function hIij (·) are:
∂hIij (·) Vi Vj (Dc cos θij + Cc sin θij )
=
∂θi
Iij
∂hIij (·)
Vi Vj (Dc cos θij + Cc sin θij )
=−
∂θj
Iij
∂hIij (·) Vj (Dc sin θij − Cc cos θij ) + Ac Vi
=
∂Vi
Iij
∂hIij (·) Vi (Dc sin θij − Cc cos θij ) + Bc Vj
=
.
∂Vj
Iij
The voltage magnitude and angle on the bus i ∈ V:
hVi (·) = Vi ; hθi (·) = θi .
(65)
The Jacobian expressions corresponding to measurements of
voltage magnitude and voltage angle are as follows:
∂hVi (·)
= 1;
∂Vi
∂hθi (·)
= 1;
∂θi
∂hVi (·)
=0
∂Vj
∂hθi (·)
= 0.
∂θj
DC SE Model
The active power flow at the branch (i, j) ∈ E that connects
buses i and j:
hPij (·) = −bij (θi − θj ).
(66)
The Jacobian expressions corresponding to hPij (·) are as
follows:
∂hPij (·)
∂hPij (·)
= −bij ;
= bij .
∂θi
∂θj
The active and reactive power injection into the bus i ∈ V:
The active power injection into the bus i ∈ V is:
X
hPi (·) = −
bij (θi − θj ),
(67)
j∈Hi \i
hPi (·) = Vi
X
Vj (Gij cos θij + Bij sin θij )
(64a)
Vj (Gij sin θij − Bij cos θij ),
(64b)
j∈Hi
hQi (·) = Vi
X
j∈Hi
where Hi is the set of buses adjacent to the bus i, including
the bus i. The parameters Gij and Bij are conductance and
susceptance of the complex bus matrix [1, Sec. 2.3]. The
where Hi \ i is the set of buses adjacent to the bus i. The
Jacobian expressions corresponding to hPi (·) are as follows:
X
∂hPi (·)
=−
bij ;
∂θi
j∈Hi \i
X
∂hPi (·)
=
bij .
∂θj
j∈Hi \i
The voltage angle on the bus i ∈ V:
hθi (·) = θi .
(68)
15
The Jacobian expressions corresponding to hθi (·) are as
follows:
∂hθi (·)
= 1;
∂θi
∂hθi (·)
= 0.
∂θj
APPENDIX B
BP- BASED A LGORITHM FOR THE AC SE
The AC model is characterized by the set of state
variables x ≡ (θ, V), while measurement functions are
defined in Appendix A, equations (62)-(65).
The factor graph construction: According to (9), in the
AC scenario, the set of state variables θ and V determines the
set of variable nodes X = {(θ1 , V1 ), . . . , (θN , VN )}, while
the set of factor nodes F = {f1 , . . . , fk } is defined by the set
of measurements M. A factor node fi connects to a variable
node xs ∈ X if and only if the state variable xs is an argument
of the corresponding measurement function hi (x).
Example 3 (Constructing factor graph). In this example,
using a simple 3-bus model presented in Fig. 14a, we
demonstrate the conversion from a bus/branch model with a
given measurement configuration into the corresponding
factor graph for the AC model. The variable nodes represent
bus 1
θ1
2
1 MP12
bus 2
fP12
V1
θ2
V2
fP3
3
MP 3
(a)
θ3
V3
bus 3
(b)
Fig. 14.
Transformation of the bus/branch model and measurement
configuration (subfigure a) into the corresponding factor graph for the DC
model (subfigure b) and the AC model (subfigure c)
state variables X = {θ1 , V1 , θ2 , V2 , θ3 , V3 }. Factor nodes
are defined by corresponding measurements, where in our
example, measurements MP12 and MP3 are mapped into
factor nodes F = {fP12 , fP3 }.
4
Message from a variable node to a factor node: Due to
the fact that variable node output messages do not depend on
measurement functions (see equation (2)), relations (13) and
(14) hold for the AC-BP.
Message from a factor node to a variable node: Due to
non-linear measurement functions hi (·), the integral in (3)
for the AC-BP cannot be evaluated in closed form.
Consequently, the message from a factor node to a variable
node will not be Gaussian. In the following, as an
approximation, we assume that for the AC-BP, the message
µfi →xs (xs ) also has the Gaussian form (15). According to
DC-BP we provide arguments that lead us to approximations
used to derive messages for the AC-BP.
Mean value evaluation: The expression for the mean of
the DC-BP zfi →xs is exact and equals (19a). Although the
expression (19a) is obtained by directly evaluating (3) for the
linear DC model, we note that it has a useful interpretation
via conditional expectation. For that purpose, let us define a
vector xb = Xi \ xs , and let zxb →fi denote a vector of mean
values of messages from variable nodes Xi \ xs to the factor
node
fi .
Then,
the
conditional
expectation
E[hi (xs , xb )|xb = zxb →fi ] can be calculated as:
E[hi (xs , xb )|xb = zxb →fi ]
= Cxs E[xs |xb = zxb →fi ] +
X
Cxb zxb →fi = zi . (70)
xb ∈Xi \xs
From the BP perspective, the conditional expected value
E[xs |xb = zxb →fi ] represents the mean zfi →xs . Hence, it is
possible to define the conditional expectation of non-linear
measurement function hi (·):
E[hi (xs , xb )|xb = zxb →fi ] = zi .
(71)
Due different forms of non-linear measurement functions
hi (·), see equations (62)-(64), the equation (71) will produce
different forms of conditional expectation E[xs |zxb →fi ] ≡
zfi →xs :
aE[x2s |xb
aE[xs |zxb →fi ] + b = 0
(72a)
= zxb →fi ] + bE[xs |xb = zxb →fi ] + c = 0
(72b)
aE[sin2 xs |xb = zxb →fi ] + bE[sin xs |xb = zxb →fi ] + c = 0,
(72c)
where a, b and c are coefficients derived from non-linear
measurement functions (see Appendix C for details).
Due to quadratic form of (72b) and (72c), we may obtain
two possible values for the mean value zfi →xs . Thus in order
to unambiguously define zfi →xs , we assume that certain a
e V),
e
e ≡ (θ,
priori knowledge of state variables, denoted as x
is available (e.g., historical data). Given the prior data, we
evaluate the mean value as:
(1)
, if ∆ > 0 and d1 < d2
z
fi →xs
(2)
(73)
zfi →xs = zfi →xs , if ∆ > 0 and d1 > d2
x
es ,
if ∆ < 0
where ∆ is the discriminant of the quadratic polynomial, and
(1)
(2)
d1 = |zfi →xs − x
es |, d2 = |zfi →xs − x
es |, (see Appendix C for
details).
The variance evaluation: The expression for the variance
of the DC-BP vfi →xs is equal (19b). Let us provide another
interpretation of the variance vfi →xs . For this purpose, we
observe the factor graph presented in Fig. 15. Consider the
set of messages µxb →fi = N (zxb →fi |xb , vxb →fi ) arriving to
the factor node fi from any variable node neighbour
xb ∈ Xi . Informally, we note that this message carries a
“belief” about itself that the variable node xb sends to the
factor node fi , representing collective evidence the rest of
the factor graph provides about the variable node xb . Let us
represent this belief by an equivalent factor node attached to
each variable node. Thus for a set of variable nodes Xi , we
introduce a set of factor nodes Feq = {fs , fl , . . . , fL },
16
It can be shown, using (78), that the variance vfi →xs is
governed by (19b) where the coefficients Cxp , xp ∈ Xi are
defined by Jacobian elements (see Appendix A and C for
details):
∂hi (·)
.
C xp =
(79)
∂xp xs =zfi →xs
xl
fl
µ xl → f i ( x l )
x
fi µxs →fi (xs ) s
.
.
.
.
.
.
fL
fs
µ xL → f i ( x L )
xb =zxb →fi
xL
Fig. 15. Factor graph which provides interpretation of the variance σf2
i →xs
.
where for each xb ∈ Xi , the corresponding factor node
fb
∈
Feq is singly-connected to xb and by
N (zxb →fi |xb , vxb →fi ). Note that, from the perspective of
SE, this factor node can observed as a measurement defined
by the value zxb →fi , variance vxb →fi , and measurement
function hb (xb ) = xb .
Let us now solve the system illustrated in Fig. 15 using
the WLS method. It is easy to show that the corresponding
Jacobian matrix6 H and weighted matrix W have the
following form:
Cxs Cxl . . . CxL
1
0
...
0
0
1
.
.
.
0
H=
(74)
..
..
..
.
.
.
0
0
...
1
W = diag(1/vi , 1/vxs →fi , 1/vxl →fi , . . . , 1/vxL →fi ). (75)
A variance-covariance matrix of WLS method is defined as:
V(xi ) = (HT WH)−1
var(xs ) cov(xs , xl ) . . . cov(xs , xL )
cov(xl , xs ) var(xl ) . . . cov(xl , xL )
=
..
.
..
.
.
..
.
cov(xL , xs ) cov(xL , xl ) . . .
(76)
var(xL )
According to (76), and using (74) and (75), the variance
var(xs ) is:
1
var(xs )
=
"
1
vxs →fi
+
1
Cx2s
vi +
X
Cx2b vxb →fi
#−1
.
(77)
xb ∈Xi \xs
Consider the second term on the right-hand side of (77). Recall
that it represents the inverse of the variance vfi →xs of the
message from the factor node fi to the variable node xs , as
defined by (19b). Therefore, we have demonstrated that by
applying WLS on the factor graph in Fig. 15, one can obtain
the expression for the variance of the message from the factor
node fi to the variable node xs .
For the AC SE that deals with non-linear measurement
functions, it is possible to define a linear approximation of
the variance-covariance matrix at a given point xi using the
Gauss-Newton method (11a):
V(xi ) = [J(xi )T WJ(xi )]−1 .
(78)
6 Note that the measurement function of the factor node f is given by (18),
i
while for all other factor nodes fb ∈ Feq , it is equal to hb (xb ) = xb .
Note that the coefficients above are evaluated at the point xi =
(xs , xb ), where the values in xi represent the mean-values of
the corresponding messages.
To summarize, the message evaluation for the AC-BP is
governed by (72) and (19b), where coefficients are obtained
using (79).
Marginal inference: The marginal of the state variable xs
is governed by (21).
Iterative AC-BP algorithms Here, the indirect factor
nodes Find ⊂ F include measurements of power flows,
power injections and currents. The direct factor nodes
Fdir ⊂ F include measurements of voltage magnitudes and
angles.
The AC-BP algorithms are presented in Algorithm 3. Note
that, the initialization step for the DC-BP and AC-BP is
different. This is due to the fact that the variance of the
message from a factor node to a variable node for the
AC-BP depends not only on the mean values of incoming
messages, but also on the mean value of the message whose
variance is being calculated.
Algorithm 3 The AC-BP
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
procedure I NITIALIZATION τ = 0
for Each fs ∈ Floc do
(0)
send µfs →xs to incidence xs ∈ X
end for
for Each xs ∈ X do
(0)
(0)
send µxs →f = µfs →xs , to incidence fi ∈ Fid
i
end for
for Each fi ∈ Find do
(0)
(0)
send µf →xs = µxs →f to incidence xs ∈ X
i
i
end for
end procedure
procedure I TERATION LOOP τ = 1, 2, . . .
while stopping criterion is not met do
for Each fi ∈ Find do
(τ )
Compute µf →xs using (72)*, (19b)*
i
end for
for Each xs ∈ X do
(τ )
Compute µxs →f using (14)
i
end for
end while
end procedure
procedure O UTPUT
for Each xs ∈ X do
Compute x̂s , vxs using (21)
end for
end procedure
*Incomming messages are obtained in previous iteration τ − 1
17
The mean zfi →Vi follows from the quadratic equation, where
we selected a solution using (73).
APPENDIX C
T HE AC-BP A LGORITHM : D ETAILED M ESSAGE
D ERIVATION
Here we present an example of evaluation of the message
from a factor node to a variable node for the AC-BP algorithm.
We consider a simple model containing buses i and j, with
the active power flow measurement Mi ≡ MPij at the branch
(i, j). The mean zi , variance vi and the measurement function
hi (θi , Vi , θj , Vj ) defined as (62a) is associated with the active
power flow measurement Mi . The corresponding factor graph
is shown in Fig. 16.
Further, all incoming messages from variable nodes to the
factor node fi have Gaussian form. Therefore, these
messages, denoted as µθi →fi (θi ), µVi →fi (Vi ), µθj →fi (θj )
and µVj →fi (Vj ), are represented by their mean-variance pair
(zθi →fi , vθi →fi ), (zVi →fi , vVi →fi ), (zθj →fi , vθj →fi ) and
(zVj →fi , vVj →fi ), respectively (Fig. 16a - Fig. 16d).
θi
µ θ i →fi ( θ i )
µ θ j →fi ( θ j )
fi
Vi
µ fi → V i ( Vi )
θj
θi
Vj
Vi
µ fi →θ i ( θ i )
θj
µ θ j →fi ( θ j )
Vi
µ V j → fi ( V j )
(c)
θj
θi
Vj
Vi
σf2i →Vi =
Cθi =
CVi =
∂hi (Vi , xb )
∂θi
∂hi (Vi , xb )
∂Vi
Vi =zfi →Vi
xb =zxb →fi
CVj =
Vi =zfi →Vi
xb =zxb →fi
∂hi (Vi , xb )
∂θj
∂hi (Vi , xb )
∂Vj
Vi =zfi →Vi
xb =zxb →fi
Vi =zfi →Vi
xb =zxb →fi
aE[Vj |xb = zxb →fi ] + b = 0,
where: xb = (θi , Vi , θj ) and zxb →fi = (zθi →fi , zVi →fi ,
zθj →fi ), with coefficients:
µ fi → θ j ( θ j )
fi
a = zi − zV2i →fi (gij + gsi )
θj
b = zVi →fi (gij cos zθij →fi + bij sin zθij →fi ).
Vj
µ V i → fi ( Vi )
Cθj =
• The message µfi →Vj (Fig. 16b): The mean zfi →Vj is defined
according to (72a) as:
Vj
µ θ i →fi ( θ i )
1
(vi + Cθ2i vθi →fi + Cθ2j vθj →fi + CV2j vVj →fi ),
CV2i
where coefficients are defined according to Jacobian elements
of the measurement function hi (·):
µ fi → V j ( V j )
(b)
fi
µ V i → fi ( Vi )
fi
µ V i → fi ( Vi )
µ V j → fi ( V j )
(a)
θi
µ θ j →fi ( θ j )
µ θ i →fi ( θ i )
The variance vfi →Vi is determined using (19b) as:
µ V j → fi ( V j )
(d)
Due the fact that the conditional expected value E[Vj |xb =
zxb →fi ] represents the mean zfi →Vj , we obtain:
Fig. 16. Messages from from factor node fi to variable nodes: Vi (subfigure
a), Vj (subfigure b), θi (subfigure c) and θj (subfigure d).
According to assumption (see Section Appendix B), the
messages from the factor node fi to variable nodes have
Gaussian form: µfi →Vi (Vi ), µfi →Vj (Vj ), µfi →θi (θi ) and
µfi →θj (θj ) with their mean-variance pair (zfi →Vi , vfi →Vi ),
(zfi →Vj , vfi →Vj ), (zfi →θi , vfi →θi ) and (zfi →θj , vfi →θj ). In
the following, we consider calculation of each of these
messages.
• The message µfi →Vi (Fig. 16a): Let us first consider the
mean zfi →Vi . The equation (19a) for the active power flow
measurement boils down to (72b):
azfi →Vj + b = 0.
The variance vfi →Vj is determined using (19b) as:
vfi →Vj =
1
(vi + Cθ2i vθi →fi + CV2i vVi →fi + Cθ2j vθj →fi ),
CV2j
where coefficient are defined according to Jacobian elements
of the measurement function hi (·).
The messages µfi →θi and µfi →θj (Fig. 16c and Fig. 16d):
Means zfi →θi and zfi →θj are defined according to (72c):
•
aE[Vi2 |xb = zxb →fi ] + bE[Vi |xb = zxb →fi ] + c = 0,
aE[sin2 xs |xb = zxb →fi ] + bE[sin xs |xb = zxb →fi ] + c = 0,
where: xb = (θi , θj , Vj ) and zxb →fi = (zθi →fi , zθj →fi ,
zVj →fi ), with coefficients:
where: xb = (Vi , θj , Vj ) and zxb →fi = (zVi →fi , zθj →fi ,
zVj →fi ) for the message µfi →θi , xb = (θi , Vi , Vj ) and
zxb →fi = (zθi →fi , zVi →fi , zVj →fi ) for the message µfi →θj ,
and xs ∈ {θi , θj }. Due the fact that the all variables and
messages preserve Gaussian distribution, the conditional
expectations of sine functions are equal to E[sin2 xs |xb =
zxb →fi ] = sin2 zfi →xs and E[sin xs |xb = zxb →fi ] =
sin zfi →xs , which allows us to compute the mean:
a = gij + gsi
b = −zVj →fi (gij cos zθij →fi + bij sin zθij →fi )
c = −zi ,
where zθij →fi is determined as zθi →fi −zθj →fi . Due the
fact that the conditional expected value E[Vi |xb = zxb →fi ]
represents the mean zfi →Vi , we can write:
a(zf2i →Vi + vfi →Vi ) + bzfi →Vi + c = 0.
a sin2 zfi →xs + b sin zfi →xs + c = 0.
To simplify expressions, we introduce coefficients a = A2 +
18
B 2 , b = −2BC and c = −A2 + C 2 :
A = gij cos zθj →fi − bij sin zθj →fi ,
xs ≡ θ i
A = gij cos zθi →fi + bij sin zθi →fi ,
xs ≡ θj
B = gij sin zθj →fi + bij cos zθj →fi ,
xs ≡ θ i
B = gij sin zθi →fi − bij cos zθi →fi ,
xs ≡ θi
C=
zV2i →fi (gij
+ gsi ) − zi
,
zVi →fi zVj →fi
xs ∈ {θi , θj }
The variance vfi →θi is determined using (19b) as:
vfi →θi =
1
(vi + CV2i vVi →fi + Cθ2j vθj →fi + CV2j vVj →fi ),
Cθ2i
where coefficients are defined, as above, by calculating
Jacobian elements of the measurement function hi (·).
The variance vfi →θj is determined according to (19b) as:
vfi →θj =
1
(vi + Cθ2i vθi →fi + CV2i vVi →fi + CV2j vVj →fi ),
Cθ2j
where coefficient follow Jacobian elements of the
measurement function hPi (·).
Using the same methodology, it is possible to define
corresponding equations for means and variances for every
type of measurement functions.
APPENDIX D
T HE GN-BP A LGORITHM : T OY E XAMPLE
An illustrative example presented in Fig. 5a will be used to
provide a step-by-step presentation of the proposed algorithm.
Input data for SE from measurement devices are
Gaussian-type functions represented by means and variances:
{zV1 , zθ2 , zθ3 , zP12 , zP3 } and {vV1 , vθ2 , vθ3 , vP12 , vP3 }.
The corresponding factor graph is given in Fig. 14, where
indirect factor nodes (red squares) are frP12 and frP3 , while
direct factor nodes (orange squares) are frV1 , frθ2 and frθ3 .
The slack factor node frθ1 (yellow square) corresponds to
the slack or reference bus. The initialization (green square)
and virtual (blue square) factor nodes are frV2 and frV3 ,
respectively.
Each variable node in the initialization step of the
algorithm has the corresponding local factor node. Local
factor nodes only send, but do not receive, the messages to
the incident variable nodes. Direct and virtual factor nodes
always repeat the same message to their corresponding
variable nodes through iterations.
f r θ1
f r θ2
Algorithm Initialization
1) The AC SE in electric power systems assumes ”flat start”
or a priori given values of state variables:
x(ν=0) = [θ1 θ2 θ3 V1 V2 V3 ](ν=0) .
2) The residual of the slack factor node is set to rθ1 = 0 with
variance vθ1 → 0.
3) The value of initialization factor nodes and virtual factor
nodes are set to rV2 → 0 and rV3 → 0, with variances
vV2 → ∞ and vV3 → ∞.
Iterate - Outer Loop: ν= 0, 1, 2, . . .; τ = 0
4) Each direct factor node computes residual, e.g.:
(ν)
(ν)
rV1 = zV1 − V1 .
5) Local factor nodes send messages represented by a triplet
(residual, variance, state variable), to incident variable
nodes, e.g.:
(ν)
(ν)
µfr →∆θ1 := rθ1 , vθ1 , θ1
θ1
(ν)
(ν)
(ν)
µfr →∆V1 := rV1 , vV1 , V1 .
V1
6) Variable nodes forward the incoming messages received
from local factor nodes along remaining edges, e.g.:
(ν)
(ν,τ )
(ν)
(ν)
= µfr →∆θ1 := rθ1 , vθ1 , θ1
µ∆θ1 →fr
P12
θ1
(ν,τ )
r∆θ1 →fr
P
:=
P3
= µfr
θ1
→∆θ1
frP12
∆V1
∆V2
(ν,τ )
P3
(ν,τ )
P3
.
7) Indirect factor nodes compute residuals, e.g.:
(ν)
(ν)
(ν)
(ν)
(ν)
rP12 = zP12 − hP12 (θ1 , θ2 , V1 , V2 ).
8) Indirect factor nodes compute appropriate Jacobian
elements associated with state variables, e.g.:
(ν)
∂hP12 (·)
(ν)
(ν)
(ν) (ν)
= V1 V2 (g12 sinθ12 − b12 cosθ12 )
∂θ1
∂hP12 (·)
(ν)
(ν)
(ν)
=
= −V1 (g12 cosθ12 + b12 sinθ12 ).
∂V2
CP12 ,∆θ1 =
(ν)
CP12 ,∆V2
Iterate - Inner Loop: τ = 1, 2, . . . , η(ν)
9) Indirect factor nodes send messages as pairs along incident
edges according to (53), e.g.:
(τ )
(τ )
(τ )
µfP →∆θ2 := rfP →∆θ2 , vfP →∆θ2
(τ )
rfr
f rV2
P12
→∆θ2
12
=
1
(ν)
CP12 ,∆θ2
12
h
(ν)
(ν)
(ν,τ −1)
rP12 − CP12 ,∆θ1 · r∆θ1 →fr
(ν,τ −1)
(ν)
∆ θ 3 ∆ V3
(ν)
:= rθ1 , vθ1 , θ1
(ν,τ )
−CP12 ,∆V1 · r∆V1 →fr
f r θ3
, θ1
∆θ2
f rP 3
f rV 1
P12
:= r∆θ1 →fr , v∆θ1 →fr , θ1
12
∆θ 1
, v∆θ1 →fr
(ν)
(ν)
(ν,τ )
µ∆θ1 →fr
(ν,τ )
(ν,τ )
12
f rV3
Fig. 17. Factor graph of the bus/branch model with given measurement
configuration shown in Fig. 5a.
(τ )
vfr
P12
→∆θ2
(ν)
=
1
(ν)
(CP12 ,∆θ2 )2
(ν)
P12
i
(ν,τ −1)
− CP12 ,∆V2 · r∆V2 →fr
P12
h
(ν)
(ν,τ −1)
vP12 + (CP12 ,∆θ1 )2 · v∆θ1 →fr
(ν,τ −1)
+(CP12 ,∆V1 )2 · v∆V1 →fr
P12
(ν)
P12
P12
i
.
(ν,τ −1)
+ (CP12 ,∆V2 )2 · v∆V2 →fr
P12
19
=
1
1
+ (τ )
vθ2 v
fr →∆θ2
P3
12
(ν)
(ν,τ )
r∆θ2 →fr
P12
P12
rθ2
+
vθ 2
=
!
(τ )
rfr →∆θ2
P3
(τ )
v∆θ2 →fr .
(τ )
P12
vfr →∆θ2
P
3
Iterate - Outer Loop: ν= 0, 1, 2, . . .; τ = η(ν)
11) Variable nodes compute marginals according to (21), e.g.:
(ν)
833
500
300
100
50
30
15
7
(ν)
∆θ̂2
P12
=
(ν)
rθ2
vθ 2
+
P12
(τ )
→∆θ2
(τ )
vfr
P12
→∆θ2
+
rfr
P3
→∆θ2
(τ )
vfr
P3
(ν+1)
(ν)
50
30
p = 0. 2
p = 0. 4
p = 0. 5
p = 0. 8
p = 0. 6
p = 1. 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Weighting coefficient α1
(b)
!
(ν)
v̂∆θ2 .
→∆θ2
12) Variable nodes update the state variables, e.g.:
θ2
100
Fig. 18. The number of non-converging simulations (out of 1000 simulations
in total) for the low noise level v1 (subfigure a) and the high noise level v4
(subfigure b).
P3
(τ )
rfr
p = 0. 4
p = 0. 6
p = 1.0
300
(a)
1
1
1
=
+ (τ )
+ (τ )
vθ 2 v
vfr →∆θ2
fr
→∆θ2
(ν)
= θ2 + ∆θ̂2 .
13) Repeat steps 4-13 until convergence.
Note that, after step 8, initialization factor nodes are
removed from the factor graph. Also in each iteration, virtual
factor nodes repeat the same message as in the initial step
and messages from a virtual factor node to a variable node
should not be included in calculation of the marginals.
APPENDIX E
S ELECTION OF R ANDOMIZED DAMPING PARAMETERS
In this appendix, we analyze the GN-BP convergence
performance in order to select appropriate values for
randomized damping parameters p and α1 introduced in
equation (37). We generate 1000 random measurement
configurations with the redundancy equals to 3. The set of
measurement values are corrupted by additive noise of both
the low noise level v1 and the high noise level v4 . The
number of inner iterations is the same as in Section VI-B.
We record the number of non-converging simulations and
evaluate the MAD value, both of which will depend on the
convergence parameters p and α1 . To recall, the probability
p defines a fraction of a factor node to variable node
messages from the current iteration that are combined with
the corresponding messages from the previous iteration. The
weighting coefficient α1 defines the ratio that determines
how messages from the current and the previous iteration are
combined. For example, p = 0.2 specifies that 20% of the
messages from the current iteration will be combined with
their values in the previous iteration, while 80% of messages
are keeping the values calculated in the current iteration.
Furthermore, if α1 = 0.1, then for the 20% of messages, the
new value is obtained as a linear combination of the values
calculated in the current and the previous iteration with
Fig. 18 shows the number of non-converging simulations
for low noise level v1 and high noise level v4 as a function
of p and α1 . In general, for the selection of p and α1 for
which only a small fraction of messages are combined with
their values in the previous iteration, and that is the case for
p close to zero or α1 close to one, we observe a large
number of non-converging simulations. This clearly
demonstrates the necessity of using (37) to “slow down” the
BP progress, thus increasing the algorithm stability and
providing improved convergence.
For a fixed value of α1 , we investigate the convergence
rate of GN-BP for different values of p. We expect that, for
any selected α1 , the GN-BP will converge faster for smaller
values of p, as lower p leads to a reduced “slow down”
effect. For example, Fig. 19 shows median MAD values for
α1 = 0.5 over the range of p values (where the median is
calculated over convergent simulation outcomes only). Figure
clearly shows that lower p leads to faster convergence.
However, one needs to be careful with selection of p in
order to avoid the combinations of p and α1 that lead to
large number of non-converging outcomes.
Number of inner iterations τ
1
Number of inner iterations τ
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
10−1
1
10−5
10−7
10−9
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
10−1
p = 0.2
p = 0.5
p = 0.8
p = 1.0
10−3
MAD
(ν)
v̂∆θ2
p = 0. 2
p = 0. 5
p = 0. 8
500
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Weighting coefficient α1
(ν)
p(∆θ2 ) ∝ N (∆θ̂2 |∆θ2 , v̂θ2 )
1
841
Number of non - converging
simulations
1
(ν,τ )
v∆θ2 →fr
P
P12
p = 0.2
p = 0.5
p = 0.8
p = 1.0
10−3
10−5
MAD
P12
coefficients 0.1 and 0.9, respectively. In the following, we
numerically investigate the randomized damping parameter
pairs that lead to acceptable trade-off between the number of
non-converging simulations and the rate of convergence.
Number of non - converging
simulations
10) Variable nodes send messages as pairs along incident edges
to indirect factor nodes according to (14), e.g.:
(ν,τ )
(ν,τ )
(ν,τ )
µ∆θ2 →fr
:= r∆θ2 →fr , v∆θ2 →fr
10−7
10−9
10−11
10−11
10−13
10−13
10−15
10−15
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(a)
(b)
Fig. 19. The median MAD values for α1 = 0.5 over the range of p values
for the low noise level v1 (subfigure a) and the high noise level v4 (subfigure
b).
20
Number of inner iterations τ
1
Number of inner iterations τ
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
1
10−2
10−2
10−4
10−4
10−6
10−6
MAD
MAD
To select convergence parameters, we observed the median
MAD values and the number of non-converging simulations.
Table I lists the p and α1 pairs that demonstrate good
performance of both metrics of interest. From the table, we
can select e.g., p = 0.4 and α1 = 0.3, to achieve good
balance between the number of non-converging simulations
and the convergence rate.
10−8
10−10
10−12
C ONVERGENCE PERFORMANCES
10−14
10−14
0.4
0.4
0.4
0.5
0.5
0.5
0.5
0.6
0.8
0.8
1.0
1.0
α1
0.1
0.2
0.3
0.2
0.3
0.4
0.5
0.5
0.5
0.6
0.6
0.7
Median MAD values per iteration (i.e. 7,8,9)
Variance v1
Variance v2
7
8
9
7
8
9
−8
−9
−11
−5
[10 ] [10 ] [10 ] [10 ] [10−6 ] [10−8 ]
5.547
5.218
4.522
5.929
5.351
4.928
4.264
4.762
5.850
4.887
5.590
4.593
5.123
4.206
3.271
5.340
4.543
3.754
2.835
4.034
5.179
3.823
5.231
3.638
9.942
9.427
6.697
13.782
10.061
7.316
5.309
7.354
14.215
8.595
13.882
7.097
5.402
4.848
4.663
5.634
5.160
4.826
3.971
4.528
5.503
4.644
5.456
4.500
4.776
3.851
3.134
5.177
4.847
3.605
2.887
3.179
5.140
3.948
4.963
3.613
10.411
8.775
6.593
14.042
9.514
8.060
5.096
6.902
14.454
8.800
13.086
7.084
Number of
non-converging
simulations
v1
v2
8
8
9
7
10
11
10
9
9
10
7
8
33
35
31
31
33
30
35
33
32
30
34
32
There are two main causes of non-converging behavior.
The main cause is that the number of inner iterations in each
outer iteration follows the exponential rule (τ = ν q ). In
general, the GN-BP performs very well with such an
exponential increase of inner iterations, however, it suffers
from insufficient number of inner iterations in the first few
outer iterations. Our experiments show that increasing the
number of inner iteration only in the first few outer iterations
can dramatically reduce the number of non-converging
simulations (see Section VI-D). The second cause is a
combination of “weakly” connected variable nodes and
measurements, especially if those nodes represent the buses
incident to transformers. In that case, to prevent the
non-converging behavior, it is sufficient to connect virtual
factor nodes to such variable nodes. Therefore, with some
adjustments, one can eliminate the non-converging cases and
make the GN-BP algorithm fully reliable.
APPENDIX F
S CALABILITY AND C OMPLEXITY OF GN-BP A LGORITHM
Scalability of the GN-BP algorithm: To demonstrate
scalability of the GN-BP algorithm, we provide MAD values
for the IEEE 30 and IEEE 118 bus test case. As before, we
generate 1000 random measurement configurations with the
redundancy equal to 3. The set of measurement values are
corrupted by additive noise of both the low noise level v1
and the high noise level v4 . The convergence parameters are
set to p = 0.4 and α1 = 0.3 and we use the exponential
inner iteration scheme. From Fig. 20 and Fig. 21, we note
that the GN-BP converges, however, the rate of convergence
slightly decreases as the size of the system increases. As
described in the previous subsection, one can use different
inner iteration schemes to fine-tune the rate of convergence.
Complexity of the GN-BP algorithm: Let us first
consider the complexity of the centralized SE based on the
10−16
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(a)
(b)
Fig. 20. The MAD values of the GN-BP for the IEEE 30 bus test case with
the exponential inner iteration scheme (η(ν) = ν e ) for the low noise level
v1 (subfigure a) and high noise level v4 (subfigure b).
Number of inner iterations τ
1
Number of inner iterations τ
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
1
10−2
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
10−2
10−4
10−4
10−6
MAD
p
10−10
TABLE I
MAD
Convergence
parameters
10−8
10−12
10−16
3 3 4 0
4 9 75 76 22 33 33 97 71
17 98 35 97 22 46 87 15 25 39 60
10−8
10−10
10−6
10−8
10−12
10−10
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
1 2 3 4 5 6 7 8 9 10 11 12
Number of outer iterations ν
(a)
(b)
Fig. 21. The MAD values of the GN-BP for the IEEE 118 bus test case with
the exponential inner iteration scheme (η(ν) = ν e ) for the low noise level
v1 (subfigure a) and high noise level v4 (subfigure b).
Gauss-Newton method. For n state variables and k
measurements, each iteration of the centralized SE involves a
constant number of matrix multiplications and a matrix
inversion whose complexity is of the order of O(kn2 ) and
O(n3 ), respectively. Due to the fact that k is proportional to
n for observable systems, and that the centralized SE takes
constant number of iterations to converge, the overall
complexity of the centralized SE scales as O(n3 ). However,
this can be reduced by employing matrix inversion
techniques that exploit the sparsity of involved matrices. The
resulting sparsity-aware centralized SE methods operate with
complexity that scales as O(n2 ) [41], [42].
The complexity of BP depends on the sparsity of the
underlying factor graph, as the computational effort per
iteration is proportional to the number of edges in the factor
graph. For each of the k measurements, the degree (the
number of incident edges) of the corresponding factor node
is limited by a (typically small) constant. Indeed, for any
type of measurements, the corresponding measurement
function depends only on a few state variables corresponding
to the buses in the local neighbourhood of the bus/branch
where the measurement is taken. Thus as n and k grow
large, the number of edges in the factor graph scales as
21
O(n), which corresponds to the computation complexity of
BP per iteration. The scaling of the number of BP iterations
ρ as n grows large is a challenging problem. Based on
discussion in [43] for full matrices, the number of iterations
is likely to scale with condition number of the underlying
matrix, which for well-conditioned matrices may scale as
low as O(1). However, we leave the more detailed analysis
on the scaling of the number of GN-BP iterations for our
future work.
To summarize, BP approach builds upon the factor graph
structure that directly exploits the underlying system sparsity,
thus achieving minimal complexity of O(n) per iteration,
while the scaling of the number of iterations needs further
study. In contrast to the optimized centralized methods
whose complexity scales as O(n2 ), the BP method can be
flexibly distributed by arbitrarily segmenting the underlying
factor graph into disjoint areas. In the extreme case of the
fully-distributed BP algorithm (that we focus on in this
paper), each factor graph node operates locally and
independently. Thus, the SE problem is distributed across
O(n) nodes, and if implemented to run in parallel, can be
O(n) times faster than the centralized solution. In addition,
for fully-distributed BP, none of the nodes need to store the
system-level matrices (whose storage-size typically scales as
O(n2 )), and storing only constant-size set of local
parameters is sufficient.
[2] F. F. Wu, K. Moslehi, and A. Bose, “Power system control centers: Past,
present, and future,” Proc. of the IEEE, vol. 93, pp. 1890–1908, Nov 2005.
[3] A. Monticelli, “Electric power system state estimation,” Proceedings of the
IEEE, vol. 88, no. 2, pp. 262–282, 2000.
[4] F. C. Schweppe and D. B. Rom, “Power system static-state estimation, part ii:
Approximate model,” IEEE Transactions on Power Apparatus and Systems,
vol. PAS-89, no. 1, pp. 125–130, Jan 1970.
[5] Y.-F. Huang, S. Werner, J. Huang, N. Kashyap, and V. Gupta, “State estimation
in electric power grids: Meeting new challenges presented by the requirements
of the future grid,” IEEE Signal Processing Magazine, vol. 29, no. 5, 2012.
[6] G. N. Korres, “A distributed multiarea state estimation,” IEEE Transactions on
Power Systems, vol. 26, no. 1, pp. 73–84, Feb 2011.
[7] W. Jiang, V. Vittal, and G. T. Heydt, “Diakoptic state estimation using phasor
measurement units,” IEEE Transactions on Power Systems, vol. 23, no. 4, pp.
1580–1589, Nov 2008.
[8] L. Zhao and A. Abur, “Multi area state estimation using synchronized phasor
measurements,” IEEE Transactions on Power Systems, vol. 20, no. 2, 2005.
[9] A. Minot, Y. Lu, and N. Li, “A distributed gauss-newton method for power
system state estimation,” in 2016 IEEE Power and Energy Society General
Meeting (PESGM), July 2016, pp. 1–1.
[10] D. Marelli, B. Ninness, and M. Fu, “Distributed weighted least-squares
estimation for power networks,” IFAC-PapersOnLine, vol. 48, no. 28, 2015.
[11] X. Tai, Z. Lin, M. Fu, and Y. Sun, “A new distributed state estimation technique
for power networks,” in 2013 American Control Conference, June 2013.
[12] R. Ebrahimian and R. Baldick, “State estimation distributed processing [for
power systems],” IEEE Transactions on Power Systems, vol. 15, no. 4, pp.
1240–1246, 2000.
[13] A. J. Conejo, S. de la Torre, and M. Canas, “An optimization approach to
multiarea state estimation,” IEEE Transactions on Power Systems, vol. 22,
no. 1, pp. 213–221, 2007.
[14] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed
optimization and statistical learning via the alternating direction method of
multipliers,” Foundations and Trends R in Machine Learning, vol. 3, no. 1,
pp. 1–122, 2011.
[15] H. Zhu and G. Giannakis, “Power system nonlinear state estimation using
distributed semidefinite programming,” IEEE Journal on Selected Topics in
Signal Processing, vol. 8, no. 6, pp. 1039–1050, 12 2014.
[16] J. Matamoros, A. Tsitsimelis, M. Gregori, and C. Antn-Haro, “Multiarea
state estimation with legacy and synchronized measurements,” in 2016 IEEE
International Conference on Communications, May 2016, pp. 1–6.
R EFERENCES
[1] A. Abur and A. Expósito, Power System State Estimation: Theory and
Implementation, ser. Power Engineering. Taylor & Francis, 2004.
[17] V. Kekatos and G. B. Giannakis, “Distributed robust power system state
estimation,” IEEE Transactions on Power Systems, vol. 28, no. 2, 2013.
[18] X. Li and A. Scaglione, “Robust decentralized state estimation and tracking
for power systems via network gossiping,” IEEE Journal on Selected Areas in
Communications, vol. 31, no. 7, pp. 1184–1194, 2013.
[19] L. Xie, D. H. Choi, S. Kar, and H. V. Poor, “Fully distributed state estimation
for wide-area monitoring systems,” IEEE Transactions on Smart Grid, vol. 3,
no. 3, pp. 1154–1169, Sept 2012.
[20] A. Gómez-Expósito, A. de la Villa Jaén, C. Gómez-Quiles, P. Rousseaux, and
T. Van Cutsem, “A taxonomy of multi-area state estimation methods,” Electric
Power Systems Research, vol. 81, no. 4, pp. 1060–1069, 2011.
[21] J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. San Francisco, USA: Morgan Kaufmann Publishers Inc., 1988.
[22] C. M. Bishop, Pattern Recognition and Machine Learning. Springer, 2006.
[23] Y. Weiss and W. T. Freeman, “On the optimality of solutions of the maxproduct belief-propagation algorithm in arbitrary graphs,” IEEE Transactions
on Information Theory, vol. 47, no. 2, pp. 736–744, 2001.
[24] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles and
Techniques. MIT Press, 2009.
[25] Y. Hu, A. Kuh, T. Yang, and A. Kavcic, “A belief propagation based
power distribution system state estimator,” IEEE Computational Intelligence
Magazine, vol. 6, no. 3, pp. 36–46, 2011.
[26] Y. Weng, R. Negi, and M. Ilic, “Graphical model for state estimation in electric
power systems,” in Smart Grid Communications, 2013 IEEE International
Conference on, Oct 2013, pp. 103–108.
[27] T. Sui, D. E. Marelli, and M. Fu, “Convergence analysis of gaussian belief
propagation for distributed state estimation,” in Decision and Control (CDC),
2015 IEEE 54th Annual Conference on. IEEE, 2015.
[28] M. Cosovic and D. Vukobratovic, “Fast real-time DC state estimation in
electric power systems using belief propagation,” 2017. [Online]. Available:
http://arxiv.org/abs/1705.01376
[29] F. R. Kschischang, B. J. Frey, and H. A. Loeliger, “Factor graphs and the sumproduct algorithm,” IEEE Transactions on Information Theory, vol. 47, no. 2,
pp. 498–519, Feb 2001.
[30] H. A. Loeliger, J. Dauwels, J. Hu, S. Korl, L. Ping, and F. R. Kschischang,
“The factor graph approach to model-based signal processing,” Proc. of the
IEEE, vol. 95, no. 6, pp. 1295–1322, 2007.
[31] A. Monticelli, State Estimation in Electric Power Systems: A Generalized
Approach, ser. Kluwer international series in engineering and computer
science. Springer US, 1999.
[32] D. Barber, Bayesian Reasoning and Machine Learning.
Cambridge
University Press, 2012.
[33] A. Wood and B. Wollenberg, Power Generation, Operation, and Control, ser.
A Wiley-Interscience publication. Wiley, 1996.
[34] G. Elidan, I. McGraw, and D. Koller, “Residual Belief Propagation: Informed
Scheduling for Asynchronous Message Passing,” ArXiv e-prints, Jun. 2012.
[35] Y. Weiss and W. T. Freeman, “Correctness of belief propagation in gaussian
graphical models of arbitrary topology,” in Advances in neural information
processing systems, 2000, pp. 673–679.
[36] B. L. Ng, J. Evans, and S. Hanly, “Distributed downlink beamforming in
cellular networks,” in 2007 IEEE International Symposium on Information
Theory, June 2007, pp. 6–10.
[37] C. Fan, X. Yuan, and Y. J. Zhang, “Scalable uplink signal detection in c-rans
via randomized gaussian message passing,” IEEE Transactions on Wireless
Communications, vol. PP, no. 99, pp. 1–1, 2017.
[38] M. Pretti, “A message-passing algorithm with damping,” Journal of Statistical
Mechanics: Theory and Experiment, vol. 2005, no. 11, 2005.
[39] P. Hansen, V. Pereyra, and G. Scherer, “Least squares data fitting with
applications.” Johns Hopkins University Press, 2012, ch. 9, p. 166.
[40] M. Cosovic and D. Vukobratovic, “Distributed gauss-newton method for ac
state estimation: A belief propagation approach,” in 2016 IEEE International
Conference on Smart Grid Communications (SmartGridComm), Nov 2016,
pp. 643–649.
[41] M. K. Enns, W. F. Tinney, and F. L. Alvarado, “Sparse matrix inverse factors
[power systems],” IEEE Transactions on Power Systems, vol. 5, no. 2, 1990.
[42] F. L. Alvarado, “Computational complexity in power systems,” IEEE
Transactions on Power Apparatus and Systems, vol. 95, no. 4, July 1976.
[43] D. Bickson, “Gaussian Belief Propagation: Theory and Aplication,” ArXiv eprints, Nov. 2008.
| 7cs.IT
|
Logical Methods in Computer Science
Vol. 13(4:9)2017, pp. 1–76
https://lmcs.episciences.org/
Submitted
Published
Jun. 01, 2015
Nov. 15, 2017
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
CLOVIS EBERHART a , TOM HIRSCHOWITZ b , AND THOMAS SEILLER c
a,b
c
Univ. Savoie Mont Blanc, CNRS, LAMA, F-73000 Chambéry, France
Department of Computer Science, University of Copenhagen, Copenhagen, Denmark and CNRS,
UMR 7030, Laboratoire d’Informatique de Paris Nord Université Paris 13, Sorbonne Paris Cité,
F-93430 Villetaneuse, France
Abstract. Following previous work on CCS, we propose a compositional model for the
π-calculus in which processes are interpreted as sheaves on certain simple sites. Such
sheaves are a concurrent form of innocent strategies, in the sense of Hyland-Ong/Nickau
game semantics. We define an analogue of fair testing equivalence in the model and show
that our interpretation is intensionally fully abstract for it. That is, the interpretation
preserves and reflects fair testing equivalence; and furthermore, any innocent strategy is
fair testing equivalent to the interpretation of some process. The central part of our work
is the construction of our sites, relying on a combinatorial presentation of π-calculus traces
in the spirit of string diagrams.
Contents
1. Introduction
1.1. Causal models and beyond
1.2. Traces and naive concurrent strategies
1.3. Innocence as a sheaf condition
1.4. Main result
1.5. Contributions
1.6. Related work
1.7. Plan
2. Notation and preliminaries
2.1. Basic notation and labelled transition systems
2.2. A π-calculus
2.3. Fair testing equivalence
2.4. Playgrounds
2
2
3
4
5
7
7
8
9
9
10
11
14
Key words and phrases: Programming languages; categorical semantics;presheaf semantics; game semantics;
concurrency; process algebra.
The authors acknowledge support from French ANR projet blancs PiCoq ANR 2010 BLAN 0305 01 and
Récré ANR-11-BS02-0010.
l
LOGICAL METHODS
IN COMPUTER SCIENCE
©
DOI:10.23638/LMCS-13(4:9)2017
CC
C. Eberhart, T. Hirschowitz, and T. Seiller
Creative Commons
2
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
3. A pseudo double category of traces
3.1. String diagrams
3.2. From string diagrams to actions
3.3. From actions to traces
3.4. The main double category
4. Codomain is a fibration
4.1. A factorisation system
4.2. A first ‘fibred’ double category
4.3. Restriction of seeds
4.4. Opliftings
4.5. Restriction of actions
4.6. Restriction of traces
5. A playground for π
5.1. A candidate playground
5.2. Correctness criterion
5.3. A playground
6. A sheaf model
6.1. Strategies and behaviours
6.2. Decomposing behaviours
6.3. Interpretation of π
6.4. Semantic fair testing
7. Intensional full abstraction
7.1. A first graph with testing for behaviours
7.2. A further graph with testing for behaviours
7.3. Intensional full abstraction
7.4. Generalisation
8. Conclusion and future work
References
Appendix A. Proof of Lemma 7.37
17
18
20
22
23
25
25
26
27
29
32
34
37
38
40
45
47
48
50
52
53
54
54
60
62
64
65
65
70
1. Introduction
1.1. Causal models and beyond. Operational semantics of programming languages standardly model the execution of programs as paths in a certain labelled transition system (lts).
Under this interpretation, different possible interleavings of parallel actions yield different
paths. Verification on ltss thus incurs a well-known state explosion problem. Similarly,
causality between various actions, visible in the syntax, is lost in the lts, thus making, e.g.,
error diagnostics difficult [32].
Causal models, originally designed for Petri nets [63] and Milner’s CCS [72], intend to
remedy both problems, but have yet to be applied to full-scale programming languages. They
have recently been extended in two different directions: (1) by Crafa et al. [18] to Milner’s
π-calculus, and (2) by Melliès [57] to Girard’s linear logic. The former extension accounts for
the subtle interaction of channel creation with synchronisation in π, a significant technical
achievement, 30 years after the first causal semantics for CCS. The latter is the first causal
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
3
model for functional languages (inspired by Hyland-Ong’s and Nickau’s games models for
PCF [62, 42]). An important challenge is now the search for a causal model of full-fledged
languages with both concurrent and functional features. Winskel and collaborators are
currently working in this direction, using extensions of Melliès’s approach [67, 73, 14].
In previous work [40, 38, 39], we have proposed a causal model for CCS based on a
different approach. We here push this approach further by applying it to the π-calculus.
1.2. Traces and naive concurrent strategies. In standard causal models, execution
traces essentially consist of partially ordered sets of atomic ‘events’. Our approach relies on
a new notion of trace, which we now briefly sketch. There is first a (straightforward) notion
of position, which is essentially a finite hypergraph whose nodes are thought of as agents,
and whose hyperedges between nodes x1 , . . . , xn are thought of as communication channels
shared by x1 , . . . , xn . There is then a notion of atomic action from one position to another.
The collection of atomic actions is thought of as a ‘rule of the game’. Examples of atomic
actions are: an agent creates a new, private communication channel; an agent forks into
two new agents connected to the same channels; an agent sends some channel a over some
channel b to some other agent. We finally have a notion of trace which allows several atomic
actions to occur, in a way that only retains some minimal causality information between
them. We here mean, e.g., information such as: ‘such agent outputs on such channel only
after having created such other channel’.
The main purpose of our notion of trace is to interpret π-calculus processes as some
kind of strategies over them. Most naively, a strategy on some position X is a prefix-closed
set of ‘accepted’ traces from X. But what should prefix mean in our setting? Well, we may
view traces with initial position X and final position Y as morphisms Y
X. Sequential
composition of traces, denoted by ●, yields an analogue of prefix ordering, defined by t ≤ t ● w.
Strategies as prefix-closed sets of traces however fail to suit our needs on three counts. First,
such naive strategies may not be stable under isomorphism of traces; second, they are bound
to model coarse behavioural equivalences, at least as coarse as may testing equivalence (a.k.a.
trace equivalence); and third, they permit undesirable interaction between players. Let us
examine these issues in more detail.
The first, easy one is that there is an obvious notion of isomorphism between traces,
under which strategies should be closed. The second problem is more serious: until now,
these too naive strategies are not concurrent enough to adequately model CCS or the
π-calculus.
Example 1.1 (Milner’s coffee machines). Consider the CCS processes P = (a.b + a.c) and
Q = a.(b + c). The process P has two ways of inputting on a and then, depending on
the chosen way, inputs on either b or c. The process Q inputs on a and then has both
possibilities of inputting on b or c. They hence exhibit significantly different interactive
behaviour. Both processes, however, accept exactly the same traces (in the standard sense),
namely {, a, ab, ac}, where denotes the empty trace.
Thus, taking strategies to be prefix-closed sets of traces would prevent us from directly
modelling any reasonably fine behavioural equivalence on processes. Inspired by presheaf
models [45], we remedy both problems at once by passing from prefix-closed sets of traces to
presheaves (of finite sets) on traces. Indeed, in the simple case where traces on X form a mere
poset T(X) by prefix ordering, a prefix-closed set of traces is nothing but a contravariant
functor from T(X) to the ordinal 2, viewed as a category. The latter has two objects 0 and
4
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
1 and just one non-trivial morphism 0 → 1. The idea is that a functor S∶ T(X)op → 2 maps
any trace to 1 when it is accepted, and to 0 otherwise. Furthermore, if t ≤ t′ , i.e., t is a prefix
of t′ , then we have a morphism t → t′ which should be mapped by S to some morphism
S(t′ ) → S(t). If t′ is accepted then S(t′ ) = 1, so this has to be a morphism 1 → S(t).
Because there are no morphisms 1 → 0, this entails S(t) = 1, hence prefix-closedness of the
corresponding strategy. Now in the case where traces form a proper category T(X), whose
morphisms encompass both prefix ordering and isomorphism between traces, considering
functors T(X)op → 2 retains prefix-closedness and solves our first problem: for any t ≅ t′ ,
functoriality imposes S(t) ≅ S(t′ ). Our second problem is then solved by replacing such
functors with presheaves, i.e., functors T(X)op → Set.
Example 1.2. In Example 1.1, the two ways that P has to accept inputting on a may
be reflected by mapping the trace a to some two-element set. More precisely, P may be
modelled by the presheaf S defined on the left and pictured on the right:
⋆
a
a
● S empty otherwise,
● S() = {⋆},
x
x′
● S( ↪ a) = {x ↦ ⋆, x′ ↦ ⋆},
● S(a) = {x, x′ },
c
b
● S(a ↪ ab) = {y ↦ x},
● S(ab) = {y},
′
y
y
.
● S(a ↪ ac) = {y ′ ↦ x′ },
● S(ac) = {y ′ },
Presheaves thus may ‘accept a trace in several ways’: the trace a is here accepted in two
ways, x and x′ . The process Q is of course modelled by equating x and x′ .
As it turns out, we actually only need finitely many ways of accepting each trace. Thus,
we arrive at a first sensible notion of strategy given by presheaves of finite sets, i.e., functors
T(X)op → set, where set denotes the category with as objects all finite subsets of N, with all
maps between them. We call them (naive) strategies in the sequel. (Please note that set is
equivalent to the category of all finite sets.)
Notation 1.3. For any C, let ̃
C denote the category of presheaves of finite sets over C.
1.3. Innocence as a sheaf condition. The third problem evoked above is that functors
T(X)op → set allow some undesirable behaviours. Intuitively, in π just as in CCS, agents
should not have any control over the routing of messages.
Example 1.4. Consider a position X with three agents x, y, and z sharing a communication
channel a, and a strategy S accepting (1) the trace where x outputs on a, (2) the trace
where y inputs on a, and (3) the trace where z inputs on a. Then, both synchronisations
should be accepted by S. However, one easily constructs a naive strategy in which one is
refused (see Example 6.2).
In order to rectify this deficiency, we enrich strategies with ‘local’ information. The idea
is that a strategy should not only accept or refuse traces on the whole position X, but also
traces on all possible subpositions of X. Moreover, this local information should fit together
coherently.
Example 1.5. Consider the position X of Example 1.4. Any strategy on X should now in
particular include independent strategies for each of the three agents x, y, and z. Coherence
means that in order for a trace to be accepted, it should be enough for it to be ‘locally
accepted’, i.e., at every stage in the trace, each agent should approve what they see of the
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
5
next action. E.g., if the next action is a synchronisation x y with x outputting and y
inputting on some channel a, then all that is required for the synchronisation to be accepted
is that x accepts to output and y accepts to input. Consequently, if some other agent z also
accepts to input on a at this stage, then the synchronisation x z is also accepted.
We call this putative coherence condition innocence by analogy with Hyland and Ong’s
notion [42]. In order to formalise it, we first extend our category of traces T(X) on X with
new objects representing traces on subpositions of X. We also add new morphisms, which
are about ‘locality’. Indeed, standardly, plays form a poset for the prefix ordering, but here
we want to enrich this, e.g., by embedding traces on subpositions of X into traces on X.
Example 1.6. Consider a position X with two agents x1 and x2 . There is a trace t on X
in which both agents fork. Consider now the subposition Y of X consisting solely of x1 and
the trace t′ on Y in which x1 merely forks. There is a morphism t′ → t in our new category.
This extended category, TX , yields an intermediate notion of strategy, given by functors
Top
X → set. Among the new objects, we have in particular traces on just one agent of X
obtained by sequentially composing atomic actions whose final position again consists of
one agent. We call this particular kind of trace a view. Views are the most ‘local’ kind of
objects in TX . They form a full subcategory VX of TX .
Example 1.7. If X merely consists of an agent x linked to n communication channels,
consider the atomic action given by x forking into two new agents, say x1 and x2 . This
action, viewed as an object of TX has three subobjects which are views: (1) the ‘identity’
view, in which nothing happens, (2) πnl , which represents the left-hand branch (to x1 ), (3)
and πnr , which represents the right-hand branch (to x2 ).
The inclusion VX ↪ TX induces a simple Grothendieck topology [54] on TX , which
amounts to decreeing that any trace is covered by its views. We finally call any S∶ Top
X → set
innocent precisely when it is a sheaf for this Grothendieck topology. In particular, giving
an innocent presheaf on TX is equivalent (up to isomorphism) to separately giving an
innocent presheaf for each agent of X, which rules out the undesirable behaviour described
in Example 1.4.
Sheaves on TX form a category SX , which is small thanks to our use of set instead of Set.
They furthermore map back to naive strategies, i.e., presheaves on T(X), by forgetting the
local information. Finally, because the considered topology is particularly simple, sheaves
̃
are equivalent to presheaves on views, i.e., SX ≃ V
X (recalling Notation 1.3). In summary,
we have three categories of strategies: naive strategies are presheaves on the ‘global’ category
of traces T(X), innocent strategies SX are sheaves on the extended category of traces TX ,
and so-called behaviours BX are presheaves on the category of views VX . The last two are
̃
equivalent, and we furthermore have an adjunction T(X)
SX .
We use both sides of the equivalence: behaviours directly lead to our compositional
interpretation ⟦−⟧∶ Pi → S of π-calculus processes, and innocent strategies are used below as
the basis for our semantic definition of fair testing equivalence.
1.4. Main result. What should we do in order to demonstrate adequacy of our model? By
definition, causal models expose some intensional information. Hence, equality is generally
much finer than any reasonable behavioural equivalence, so we should not base our main
result on it. On the other hand, causal models are supposed to be ‘compositional’, i.e., to
6
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
come equipped with an interpretation of syntactic operations in the model. A natural thing
to do is thus to choose some behavioural equivalence defined from syntactic operations,
use compositionality to transpose it to the model, and prove that the two coincide. More
precisely, the considered equivalence induces by quotienting two ‘extensional collapses’, one
syntactic and the other semantic, and we want to prove that the translation induces a
bijection between both extensional collapses. Following [3], we call this intensional full
abstraction for the considered equivalence. In fact, all behavioural equivalences end up
relying on some notion of observation, which we will also need to transpose to the model.
We here focus on so-called testing equivalences [20, 61, 9, 66], which are defined in two
stages. First, one chooses a ‘mode of interaction’. That is, one defines what the relevant
tests are for a given process and specifies how the two should interact. Typically, tests for
P are other processes T with the same free communication channels as P , and interaction
is just parallel composition P ∣ T . This part will be easy to transpose to the model by
compositionality. The second stage amounts to choosing when P ∣ T is successful. E.g., in
♡
may testing equivalence P ∣ T is successful just when there exists a transition (P ∣ T ) Ô
⇒ P′
(that is, a ♡ transition, possibly surrounded by silent transitions), where ♡ is some ‘tick’
action fixed in advance. In must testing equivalence, success is when all maximal (possibly
infinite) transition sequences contain at least one ♡ transition. In fair testing equivalence,
♡
one requires that all silent sequences (P ∣ T ) Ô
⇒ P ′ extend to some sequence P ′ Ô
⇒ P ′′ Ð
→ P ′′′
ending with a ♡ transition. These ideas transpose to the model by observing whether a given
trace contains a ♡ action. In this paper, we focus on fair testing equivalence, i.e., we prove
(Theorem 6.25) that our model is intensionally fully-abstract for fair testing equivalence. We
finally show (Section 7.4) that the result generalises to a wide range of testing equivalences,
obtained by varying the notion of success.
In order to fix intuitions, let us quickly motivate must and fair testing, using barbed
congruence [69] as a standard starting point. Barbed congruence equates processes P and Q,
roughly, when for all contexts C, C[P ] and C[Q] are weakly bisimilar w.r.t. reduction (i.e.,
only τ -actions are allowed), and furthermore they have the same interaction capabilities at
all stages. Barbed congruence is sometimes perceived as too discriminating w.r.t. guarded
choice. Consider, e.g., the following CCS processes.
P1
=
●
τ
τ
●
●
τ
a
●
●
τ
b
●
●
P2
=
●
τ
τ
●
●
τ
b
●
●
τ
a
●
●
Both processes may disable both actions a and b, the only difference being that P1 disables a
before disabling b. Barbed congruence distinguishes P1 from P2 (consider the trivial context
C = ◻), which some view as a deficiency.
Another possibility would be must testing equivalence [20]. Recall that P must pass a
test process R iff all maximal executions of P ∣ R perform, at some point, the ‘tick’ action ♡.
Then, P and Q are must testing equivalent iff they must pass the same tests. Must testing
equivalence does equate P1 and P2 above, but is sometimes perceived as too discriminating
w.r.t. divergence. E.g., consider Q1 = !τ ∣ a and Q2 = a. Perhaps surprisingly, Q1 and Q2 are
not must testing equivalent. Indeed, Q2 must pass the test a.♡, but Q1 does not, due to an
infinite, silent reduction sequence.
Fair testing equivalence was originally introduced (for CCS-like calculi) to rectify both
the deficiency of barbed congruence w.r.t. choice and that of must testing equivalence w.r.t.
divergence. The idea is that two processes are equivalent when they should pass the same
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
7
tests. A process P should pass the test T iff their parallel composition P ∣ T never loses the
ability of performing the special ‘tick’ action ♡, after any ♡-free reduction sequence. Fair
testing equivalence thus equates P1 and P2 above, as well as Q1 and Q2 . Cacciagrano et
al. [13] provide an excellent survey of fair testing for π.
Example 1.8 ([13]). The π-calculus features a well-known encoding of internal choice
using channel creation and parallel composition. Mixing this with replication leads to
intriguing examples of fair testing. Consider the following subtly different encodings of
!(b ⊕ c), where ⊕ denotes internal choice and ! denotes replication: let R1 =!νa.(a ∣ a.b ∣ a.c)
and R2 = νa.!(a ∣ a.b ∣ a.c). These are clearly fair testing equivalent. However, each encoding
has an execution that always makes, say, the left choice, and Cacciagrano et al. argue that
for R1 this execution is fair, as the involved channel is different each time. They use similar
examples to argue that fair testing is in fact too coarse, and instead propose alternative
notions (which lie beyond the scope of this paper).
1.5. Contributions. Since this paper follows the same approach as previous work on
CCS [40, 38, 39], we should explain in which sense extending the approach to π is more
than an easy application.
A first contribution comes from the fact that, in order to even define composition in our
category of traces, we need to show that traces form the total category of a fibration [43]
over positions. In previous work, this was done in an ad hoc way. We here introduce a more
satisfactory approach based on factorisation systems [52, 29].
A second significant contribution is prompted by the interplay between synchronisation
and private channels in π, which is notoriously subtle to handle. And indeed, our proof
method for CCS fails miserably on π. One reason for this, we think, is that our notion of
trace for π, though simple and natural, is not ‘modular’ enough, in the sense that a trace
contains strictly more information than the collection of all ‘local’ information accessible to
agents (i.e., of all of its views, in the above sense). Thus, adapting our proof technique from
CCS would have required us to define a much more complex but more modular notion of
trace. Instead, we here take a somewhat rougher route.
Finally, our proof now applies not only to fair testing equivalence, but also to a whole
class of testing equivalences.
1.6. Related work. Beyond the obviously closely related, already mentioned work of
Winskel et al., we should mention other causal models for π [11, 59, 26, 16, 7, 21, 15, 10, 18, 12],
as well as interleaving models [28, 27, 70, 15, 60, 36] and the early approach [44] based on
Girard’s Geometry of Interaction. All of these models are based on some lts for π. Instead,
ours is rather based on reduction rules. The subtleties usually showing up in ltss, related
to mixing synchronisation and private channels, do resurface in our proof of intensional full
abstraction, but not in the definition of our model. Indeed, it merely goes by describing the
‘rule of the game’ in π, and applying the general framework of playgrounds [39].
Another general framework relating operational and denotational descriptions of programs is Kleene coalgebra [6], which is mainly designed for automata theory. Playgrounds
may be viewed as adapting ideas from Kleene coalgebra to the process-algebraic setting.
We should also mention Laird’s games model of (a fragment of) π [47], which accounts
for trace (a.k.a. may testing) equivalence. Standard game models view strategies as sets
of traces (with well-formedness conditions), so, as we have seen, lend themselves better to
8
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
modelling trace equivalence. In a non-deterministic, yet not concurrent setting, Harmer and
McCusker [35] resort to an explicit action for divergence, which allows them to recover a
finer behavioural equivalence. We feel that the presheaf-based approach is more general.
Furthermore, recent work by Tsukada and Ong [71] adapts and extends some ideas
of [40, 38] to nondeterministic, simply-typed λ-calculus. In particular, they show that
innocent strategies as sheaves are compatible with the hiding operation of standard game
semantics. Eberhart and Hirschowitz further establish [22] a formal link between Tsukada
and Ong’s notion of innocence and ours: they construct a model of nondeterministic, simplytyped λ-calculus in our style, and then a morphism of Grothendieck sites, which entails that
both models are equivalent.
Let us moreover mention less closely related work: Girard’s ludics [31], Melliès’s
reworking of game semantics [56, 57], the part of it rediscovered by Levy [51] with a different
presentation, Melliès’s game semantics in string diagrams [58], Harmer et al.’s categorical
combinatorics of innocence [35], McCusker et al.’s graphical foundation for schedules [55],
and Winskel’s strategies as profunctors [73]. Finally, Hildebrandt’s work [37] also uses
sheaves, though as a tool to correctly handle infinite behaviour, as opposed to their use here
to force reactions of agents to depend only on their views.
1.7. Plan. In previous work [39], we have defined an algebraic notion called playground,
which provides a sufficient framework for sheaf-based innocence to make sense. Namely, it
organises positions, atomic actions, and traces into a pseudo double category [24, 25, 33, 34,
49, 30] with additional structure. Any playground D automatically gives rise, among other
things, to
● categories of innocent strategies SX on each position X, organised into a pseudo double
functor from Dop to small categories;
● a simple, yet complete syntax for innocent strategies, together with an lts SD for them
over an alphabet built from atomic actions.
After introducing some notation, the considered variant of π-calculus, fair testing
equivalence (Section 2), and recalling the notion of playground, we construct a playground D
for the π-calculus in Sections 3 to 5. This is a lot of work, and not all aspects of playgrounds
are used in defining the model and proving the main result. The reason we devote so
much energy to it is that playgrounds provide a really helpful setting, in fact a calculus, for
reasoning about positions, traces, views and the various notions of strategies. The underlying
pseudo double category is constructed in Section 3. The main playground axiom, asserting
that a certain functor is a Grothendieck fibration, is established in Section 4. Finally, the
remaining axioms are proved in Section 5.
We then continue in Section 6 by applying results from [39] to define our sheaf model
and semantic fair testing equivalence, as well as our translation ⟦−⟧ of π. We then state
the main result (Theorem 6.25). In Section 7, after introducing the basic notion of definite
residual, we reduce our main theorem to an analogous statement about an lts S for strategies
(derived from SD ). The advantage of the latter statement is that it lies entirely in the realm
of ltss. We then define a further, more syntactic lts M which we prove equivalent to S,
thus further reducing the main result to an analogous statement about M. We finally prove
the latter, which entails the main result.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
9
2. Notation and preliminaries
We start in this section with a few reminders. In Section 2.1, after fixing some basic
notation, we recall ltss. We use a slightly more general, graph-based notion than the
standard, relation-based one. In Section 2.2, we introduce the considered π-calculus, which
is mostly standard except that (1) we use a presentation in the style of Berry and Boudol’s
chemical abstract machine [5], and (2) we consider infinite terms, thus sparing us the need
for recursion or replication constructs. We then go on and recall fair testing equivalence
for π in Section 2.3. In fact, because we will also need to define fair testing for other ltss,
we introduce a general framework in which it makes sense, called graphs with testing. We
further provide sufficient conditions for a relation between the vertices of two graphs with
testing to preserve and reflect fair testing equivalence (Lemma 2.22 and Corollary 2.25).
Finally, in Section 2.4, we recall and briefly explain the definition of playgrounds.
2.1. Basic notation and labelled transition systems. First of all, we adopt the notation
of [39, Section 2], with the slight modification that set now denotes the category with finite
subsets of N as objects, and all maps as morphisms. (This category is equivalent to what we
used in [39], but slightly easier to work with for our purposes.) For all n ∈ N, we often abuse
̂ the category of presheaves
notation and let n denote the finite set {1, . . . , n}. We denote by C
̃
on C, and by C the category of presheaves of finite sets, i.e., of contravariant functors to set.
For any category C, let Cf denote the full subcategory of finitely presentable objects [4], or
finite objects for short. In the only case where we’ll use this, C will be a presheaf category
̂ will be
[Cop , set] and furthermore due to the special form of C, finite presentability of F ∈ C
equivalent to the category of elements of F being finite, and further equivalent to the set of
elements of F being finite, i.e., ∑c∈ob(C) F (c) is finite.
To recall some bare minimum: we often confuse objects C of a category C with the
̂ Gph denotes the category of reflexive graphs,
corresponding representable presheaves yC ∈ C.
and all our graphs are reflexive so we often omit mentioning it. We think of morphisms
p∶ G → A in Gph as ltss over the alphabet A, except that for reasons specific to playgrounds
our convention is that a transition from x to y is represented as an edge x ← y. Using graphs
as alphabets generalises the standard approach based on sets of labels: indeed, in order
to model any set of labels, take for A the graph with one vertex and one endo-edge for
each label. The extra generality is useful, e.g., to add some typing information on labels.
Finally, using graphs as alphabets provides us with standard tools for transporting ltss
across morphisms (by pullback, resp. post-composition).
For any graph G, G⋆ denotes the graph with the same vertices and all paths between
them; on the other hand, fc(G) denotes the free category on G, i.e., the category with the
same vertices and identity-free paths between them. Both (−)⋆ and fc extend to functors,
i.e., act on morphisms. We often silently coerce fc(G) into a reflexive graph, and denote by
̃
− the obvious morphism G⋆ → fc(G).
For any graph p∶ G → A over A, x, y ∈ ob(G), and edge a∶ p(y) → p(x) in A, we denote
a
by x ←
Ð y the existence of an edge e∶ y → x in G such that p(e) = a. When a = id , we just
write x ← y. We denote strong bisimilarity over A by ∼A .
For any graph p∶ G → A over A, x, y ∈ ob(G), and path ρ∶ p(y) → p(x) in A⋆ , we denote
ρ
⋆ (r) = ρ
̃. When ρ is the empty
by x ⇐
Ô y the existence of a path r∶ y → x in G⋆ such that p̃
path we just write x ⇐
Ô y. We denote weak bisimilarity over A by ≈A .
10
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
2.2. A π-calculus. We now present our variant of π, which features a chemical abstract
machine presentation and infinite terms. Also, we keep track of the channels known to the
considered process, i.e., we work with a ‘natural deduction’ presentation of terms.
Processes are infinite terms coinductively generated by the grammar
γ ⊢g P1
γ⊢P
γ ⊢g Pn
...
γ⊢Q
γ ⊢P ∣Q
γ ⊢ ∑ Pi
i∈n
γ, a ⊢ P
γ⊢P
γ⊢P
γ ⊢g νa.P
γ ⊢g ♡.P
γ ⊢g τ.P
a∈γ
γ, b ⊢ P
a, b ∈ γ
γ ⊢g a(b).P
γ⊢P
γ ⊢g ā⟨b⟩.P
,
where
● we have two judgements, ⊢ for processes and ⊢g for guarded processes;
● γ ranges over finite sets of natural numbers, and
● γ, a is defined iff a ∉ γ and then denotes γ ⊎ {a}.
Notation 2.1. Let Pi be the set of all such (non-guarded) processes. Let Pi γ denote the
set of processes γ ⊢ P .
As usual, a is bound in νa.P and b is bound in a(b).P . In the following, processes are
considered equivalent up to renaming of bound channels. Capture-avoiding substitution
extends the assignment γ ↦ Pi γ to a functor set → Set mapping σ∶ γ → γ ′ to P ↦ P [σ].
Let us now describe the dynamics of our π-calculus. They are slightly unusual, in
that they are presented in the style of the chemical abstract machine. In particular, there
are silent transitions for ‘heating’ both parallel composition and name creation. A further
slight peculiarity, which we adopt for its convenience in the chemical abstract machine
presentation, is that name creation is a guard. E.g., we have some processes of the form
(νa.P ) + b(x).Q. This is hardly significant. E.g., the previous process is strongly bisimilar to
(τ.νa.P ) + b(x).Q in more standard settings, and our results are about equivalences coarser
than weak bisimilarity anyway.
Notation 2.2. For any γ ⊢g P , γ ⊢ Q of the form ∑i∈n Qi , and injection h∶ n ↪ n + 1, we
denote by P +h Q the sum ∑j∈n+1 Pj , where Ph(i) = Qi for all i ∈ n and Pk = P , for k the
unique element of (n + 1) ∖ Im(h).
Definition 2.3. Let −⊙ denote the finite multiset monad on sets.
Definition 2.4. A configuration is an element of Conf = ∑γ∈Pf (N) Pi ⊙
γ.
Notation 2.5. Configurations (γ, S) will be denoted by ⟨γ∥S⟩, and we will use list syntax
[P1 , . . . , Pn ] for multisets, sometimes dropping brackets, e.g., as in ⟨γ∥P1 , . . . , Pn ⟩. We
sometimes resort to a hopefully clear ‘multiset comprehension’ notation [P ∣ ϕ(P )]. We use
∪ for multiset union and x ∶∶ M = [x] ∪ M .
Just as Pi , Conf extends to a functor set → Set by capture-avoiding substitution.
Let us now extend Conf to an lts over the alphabet {♡, τ }. This means that we need
to construct a graph morphism Conf → Σ, where Σ denotes
♡
●
τ ,
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
11
τ being the chosen identity edge.
This is done in Figure 1, omitting identity edges. There, we let R and R′ range over
processes of the form ∑i∈n Pi . The last rule makes sense because each transition as in the
premise implicitly comes with an inclusion h∶ γ1 ↪ γ2 , and the second occurrence of S is
implicitly S[h].
τ
τ
⟨γ∥P ∣ Q⟩ ←
Ð ⟨γ∥P, Q⟩
τ
⟨γ∥τ.P +h R⟩ ←
Ð ⟨γ∥P ⟩
⟨γ∥νa.P +h R⟩ ←
Ð ⟨γ, a∥P ⟩
⟨γ∥♡.P +h R⟩ ←
Ð ⟨γ∥P ⟩
♡
τ
⟨γ∥a(b).P +h R, ā⟨c⟩.Q +h′ R′ ⟩ ←
Ð ⟨γ∥P [b ↦ c], Q⟩
α
⟨γ1 ∥S1 ⟩ ←
Ð ⟨γ2 ∥S2 ⟩
α
⟨γ1 ∥S ∪ S1 ⟩ ←
Ð ⟨γ2 ∥S ∪ S2 ⟩
(α ∈ {τ, ♡})
Figure 1: Reduction rules for Conf
2.3. Fair testing equivalence. Let us now define fair testing equivalence for π, together
with our general framework of graphs with testing. Graphs with testing are essentially De
Nicola and Hennessy’s original framework [20], adapted to our graph-based presentation of
ltss. We derive a few results about general graphs with testing, notably sufficient conditions
for a relation between two graphs with testing to preserve and reflect fair testing equivalence.
Remark 2.6. In [39], an abstract framework was defined for studying fair testing equivalence
and its relationship with weak bisimilarity. We won’t use this here, and instead work in a
simpler setting.
We first cover π-calculus, and then generalise. The starting point is that we need to
be able to test processes against other processes, and more generally configurations against
configurations. Because configurations carry their sets of free channels, it makes sense to
consider a partial parallel composition operation:
Definition 2.7. For any ⟨γ∥S⟩, ⟨γ ′ ∥S ′ ⟩ ∈ Conf, let ⟨γ∥S⟩@⟨γ ′ ∥S ′ ⟩ denote ⟨γ∥S ∪ S ′ ⟩ if
γ = γ ′ and be undefined otherwise. Let furthermore γ = ⟨γ∥⟩.
Lemma 2.8. The domain of @, i.e., the set of pairs (C, C ′ ) such that C@C ′ , is an
equivalence relation.
Let us denote by ¨Conf this equivalence relation.
Here is the standard definition of fair testing equivalence:
Definition 2.9. Let Conf denote the set of configurations C such that for all C ⇐
Ô C ′,
♡
there exists C ′ ⇐
Ô C ′′ .
Any two configurations C and C ′ are fair testing equivalent iff C ¨Conf C ′ and for all
D ¨Conf C, (C@D) ∈ Conf iff (C ′ @D) ∈ Conf .
Let us now abstract away from this definition. For this, it would make sense to start
from a partial, parallel composition map. However, in the model, the corresponding map
will involve a pushout of positions which is only determined up to isomorphism. We thus
12
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
generalise from partial maps to relations, but we need to impose some conditions. What
matters is the use we will make of parallel composition for testing. Intuitively, we will check
that the parallel composition C@C ′ belongs to some given pole, which is closed under weak
bisimilarity over Σ. This choice is perhaps slightly arbitrary, but it encompasses all known
testing equivalences. Finally, to gain just a little more generality, we will use the fact that
weak bisimilarity over Σ coincides with strong bisimilarity over fc(Σ), and work with the
latter (Definition 7.40). For ltss over fc(Σ) obtained by applying fc to ones over Σ, this is
equivalent to the previous setting. The extra generality is useful for describing ltss which
are not free categories. E.g., in Section 7.1, we introduce the graph with testing C, whose
transitions are traces, which compose in a non-free way. We thus start from the following
notion, and then define fair testing equivalence.
Definition 2.10. A graph with testing is a graph G together with a morphism p∶ G → fc(Σ)
and a relation R∶ (ob(G))2
ob(G) whose domain is an equivalence relation and which is
partially functional up to strong bisimilarity over fc(Σ).
The domain being an equivalence relation more precisely means that the set {(x, y) ∣
∃z.(x, y) R z} forms an equivalence relation over ob(G).
Partial functionality up to strong bisimilarity means that if (x, y) R z and (x, y) R z ′ ,
then z ∼fc(Σ) z ′ .
Notation 2.11. The relation is called the testing relation, and we denote it by ∣G , i.e.,
(x, y) R z is denoted by z ∈ (x ∣G y). Furthermore, its domain is denoted by ¨G . We use ∣
and ¨ when there is no ambiguity. Since ∣ is partially functional up to strong bisimilarity,
for any (x, y) R z, as long as what we say about z is invariant under strong bisimilarity,
then it also holds for any other z ′ such that (x, y) R z ′ . In such cases, we implicitly make
some global choice of z and consider ∣ as partially functional.
Example 2.12. Figure 1 defines a morphism pConf ∶ Conf → Σ. Because, @ is a partial
map, it induces a partially functional relation ob(Conf)2
ob(Conf), whose domain is an
equivalence relation. Because partially functional implies partially functional up to strong
bisimilarity, we have:
Proposition 2.13. The morphism fc(pConf )∶ fc(Conf) → fc(Σ), with @ as testing relation,
forms a graph with testing.
We may now mimick the standard definition of fair testing equivalence in the abstract
setting:
Definition 2.14. For any graph with testing p∶ G → fc(Σ), let G denote the set of objects
♡
x such that for all x ←
Ð y, there exists y ←
Ð z.
Any two objects x and y are fair testing equivalent iff x ¨G y and for all z ¨G x,
(x ∣G z) ∈ G iff (y ∣G z) ∈ G .
Remark 2.15. Because we are working over fc(Σ), if G has the shape fc(G′ ), then single
♡
transitions like x ←
Ð y denote arbitrary paths of silent transitions. And analogously x ←
Ðy
denotes any path with all edges silent except exactly one.
Notation 2.16. We denote fair testing equivalence in G by ∼G
f . Given x, any z such that
x ¨ z is called a test for x, and x passes the test iff (x ∣G z) ∈ G .
Example 2.17. Definition 2.14 instantiates to Definition 2.9.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
13
In fact, the construction of the graph with testing fc(pConf ) from pConf is easily generalised:
ob(G) whose
Lemma 2.18. For any morphism pG ∶ G → Σ and relation R∶ (ob(G))2
domain is an equivalence relation, R equips fc(pG ) with testing structure iff it is partially
functional up to weak bisimilarity over Σ.
Proof. R is a strong bisimulation for fc(G) iff it is a weak bisimulation for G.
Definition 2.19. A graph with testing is free iff it is of the form fc(pG ).
Our fc(pConf ) is thus a free graph with testing.
Let us conclude this section with a sufficient condition for a relation between the
vertices of two graphs with testing to be adequate for fair testing equivalence, and a natural
specialisation to free graphs with testing.
Definition 2.20. A relation R∶ ob(G)
ob(H) between the vertex sets of two graphs with
testing pG ∶ G → fc(Σ) and pH ∶ H → fc(Σ) is fair iff
● x R y and x′ R y ′ implies (x ¨G x′ ) ⇔ (y ¨H y ′ );
● R is total and surjective, i.e.,
– for all x ∈ G, there exists y ∈ H such that x R y, and
– for all y ∈ H, there exists x ∈ G such that x R y;
● x R y implies x ∼fc(Σ) y;
● if x R y, x′ R y ′ , and x ¨G x′ , then there exist u ∈ (x ∣G x′ ) and v ∈ (y ∣H y ′ ) such that
u R v.
ob(H) between the vertex sets of two graphs with
Lemma 2.21. A relation R∶ ob(G)
G
H
testing p ∶ G → fc(Σ) and p ∶ H → fc(Σ) is fair iff its converse R† is, where R† is defined
by (y R† x) ⇔ (x R y).
Proof. Easy.
Lemma 2.22. For any fair relation R∶ ob(G)
′
H ′
(x ∼G
f x ) ⇔ (y ∼f y ).
ob(H), if x R y and x′ R y ′ , then
For proving this lemma, we need:
Lemma 2.23. For all graphs with testing pG ∶ G → fc(Σ) and pH ∶ H → fc(Σ), x ∈ G and
y ∈ H, if x ∼fc(Σ) y, then (x ∈ G ) ⇔ (y ∈ H ).
Proof. Assume x ∈ G and consider any transition y ← y ′ in H. By bisimilarity, x ← x′ ∼fc(Σ)
Ð x′′ , so by bisimilarity, y ′ ←
Ð y ′′ . Thus, y ∈ H , which entails
y ′ . By hypothesis, we find x′ ←
the result by symmetry.
♡
♡
Proof of Lemma 2.22. Consider any such x, y, x′ , and y ′ . By Lemma 2.21, it suffices to show
′
′
one direction of the desired equivalence. So let us assume that x ∼G
f x . Then x ¨G x , hence
also y ¨H y ′ by fairness of R. By symmetry, it again suffices to check one direction of the
desired implication. Consider thus any t ¨H y such that (y ∣ t) ∈ H . By surjectivity of R, we
find s ∈ G such that s R t. By fairness again, we find u ∈ (x ∣ s) and v ∈ (y ∣ t) such that u R v,
so (x ∣ s) ∼fc(Σ) u ∼fc(Σ) v ∼fc(Σ) (y ∣ t), and hence by the previous lemma (x ∣ s) ∈ G . Because
′
′
G
′
H
x ∼G
f x , this entails (x ∣ s) ∈ , hence by a similar argument (y ∣ t) ∈ , as desired.
14
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
ob(H) between the vertex sets of two free graphs
Definition 2.24. A relation R∶ ob(G)
with testing respectively generated by pG ∶ G → Σ and pH ∶ H → Σ is weakly fair iff it satisfies
the conditions of Definition 2.20, except for the third one, which is replaced by: x R y
implies x ≈Σ y.
Corollary 2.25. For any weakly fair relation R∶ ob(G)
fc(G) ′
fc(H) ′
then (x ∼f
x ) ⇔ (y ∼f
y ).
ob(H), if x R y and x′ R y ′ ,
Proof. Because weak bisimilarity over Σ is the same as strong bisimilarity over fc(Σ), being
weakly fair is the same as being fair for the generated graphs with testing.
2.4. Playgrounds. To conclude this preliminary section, let us recall the axioms for playgrounds [39]. Some constructions and results are developped from these axioms in op. cit.
Some of the main ideas are reviewed and reworked in Sections 6 and 7.
Let us start with a brief recap of pseudo double categories. A pseudo double category
D consists of a set ob(D) of objects, shared by a ‘horizontal’ category Dh and a ‘vertical’
bicategory Dv . Since we won’t consider strict double categories, we’ll often omit the word
‘pseudo’. Following Paré [64], Dh , being a mere category, has standard notation (normal
arrows, ○ for composition, id for identities), while the bicategory Dv earns fancier notation
(
for arrows, ● for composition, id ● for identities). D is furthermore equipped with a set
of double cells α, which have vertical, resp. horizontal, domain and codomain, denoted by
domv (α), codv (α), domh (α), and codh (α). The horizontal domain and codomain of a double
cell are vertical morphisms, while the vertical domain and codomain are horizontal morphisms.
E.g., for α in the diagram delow, we have u = domh (α), u′ = codh (α), h = domv (α), and
h′ = codv (α). D is furthermore equipped with operations for composing double cells: ○
composes them along a common vertical morphism, ● composes along horizontal morphisms.
Both vertical compositions (of morphisms and double cells) may only be associative up to
coherent isomorphism. The full axiomatisation is given by Garner [30], and we here only
mention the interchange law, which says that the two ways of parsing the above diagram
coincide: (β ′ ○ β) ● (α′ ○ α) = (β ′ ● α′ ) ○ (β ● α).
h
X
u
α
Y
v
Z
β
Y
Z
k′
′
Y
β′
v′
h′′
X ′′
α′
u′
h′
k
X′
′
k′′
Z
u′′
′′
v ′′
′′
For any (pseudo) double category D, we denote by DH the category with vertical morphisms as
objects and double cells as morphisms, and by DV the bicategory with horizontal morphisms
as objects and double cells as morphisms. Domain and codomain maps form functors
domv , codv ∶ DH → Dh and pseudofunctors domh , codh ∶ DV → Dv . We will refer to domv and
codv simply as dom and cod, reserving subscripts for domh and codh .
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
15
We then need to recall the notion of fibration (see [43]). Consider any functor p∶ E → B.
A morphism r∶ E ′ → E in E is cartesian when, as in
E ′′
t
s
E′
p(E ′′ )
E
r
p(t)
k
p(E ′ )
p(E),
p(r)
for all t∶ E ′′ → E and k∶ p(E ′′ ) → p(E ′ ), if p(r)○k = p(t) then there exists a unique s∶ E ′′ → E ′
such that p(s) = k and r ○ s = t.
Definition 2.26. A functor p∶ E → B is a fibration iff for all E ∈ E, any h∶ B ′ → p(E) has a
cartesian lifting, i.e., a cartesian antecedent by p.
Notation 2.27. We denote by E∣h the domain of the (chosen) cartesian lifting, and call it
the restriction of E along h.
We may now state the definition of playgrounds.
Remark 2.28. The following differs slightly from the original definition, mostly in presentation and terminology, but more significantly because the class B of basic moves was
mistakenly required to be replete in [39].
We provide some intuition right after the definition.
Definition 2.29. In a double category, a cell α is special when its vertical domain and
codomain cod(α) and dom(α) are identities.
Definition 2.30. A playground is a double category D such that cod is a fibration, equipped
with
● a full subcategory I ↪ Dh of objects called individuals,
● a full, replete1 subcategory M ↪ DH , whose objects are called actions, with full subcategories B and F of basic and full actions, with F replete,
● a map ∣ − ∣∶ ob(DH ) → N called the length,
satisfying the following conditions:
(P1) I is discrete. Basic actions have no non-trivial automorphisms in DH . Vertical
identities on individuals have no non-trivial endomorphisms.
(P2) (Individuality) Basic actions have individuals as both domain and codomain.
(P3) (Atomicity) For any cell α∶ v → u in DH , if ∣u∣ = 0 then also ∣v∣ = 0. Up to a special
isomorphism in DH , all plays u of length n > 0 admit decompositions into n actions.
For any u∶ X
Y of length 0, there is an isomorphism id ●X → u in DH , as in
X
X
αu
X
1Replete means closed under isomorphism.
ū
u
Y.
16
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
(P4) (Fibration, continued) Restrictions of actions (resp. full actions) to individuals either
are actions (resp. full actions), or have length 0.
X, and y∶ d → Y with d ∈ I, there exists a cell
(P5) (Views) For any action M ∶ Y
y
d
Y
αy,M
v y,M
y,M
d
M
X,
yM
where v y,M either is a basic action or has length 0, which is unique up to canonical
isomorphism, i.e., for any y ′ ∶ d′ → X, v ′ ∶ d
d′ , and α′ ∶ v ′ → M with dom(α′ ) = y
′
′
′
M
and cod(α ) = y , we have y = y and there exists a unique isomorphism β∶ v → v ′
making the diagram
y
d
αy,M ′
α
v′
yM
v y,M
dy,M
Y
β
d′
M
X
y′
commute.
(P6) (Left decomposition) Any double cell
h
A
h
A
X
α3
w1
α
decomposes as
Y
u
u
l
C
Y
α2
w2
u2
Z
k
w1
u1
w2
B
X
α1
B
Z
k
with α3 an isomorphism, in an essentially unique way.
(P7) (Right decomposition) Any double cell as in the center below, where b is a basic
action and M is an action, decomposes in exactly one of the forms on the left and
right:
A
X
α1
B
w
Y
α2
C
h
A
¢
α
B
C
A
¨
Y
B
M
k
Z
X
α1
u
b
Z
X
Y
α2
C
Z.
(P8) (Finiteness) For any object X, the comma category I/X (taken in Dh ) is finite.
(P9) (Basic vs. full) For all d ∈ I and actions M ∶ X
d, M ′ ∶ X ′
d, and b∶ d′
d with
′
′
M and M full and b basic, if there exist cells M ← b → M then M ≅ M ′ .
Intuitively, the objects of D are configurations, or positions, in the game. The considered
games are multi-party, so it makes sense to consider embeddings of positions: this is intended
to be described by the horizontal category Dh . The vertical category Dv is that of traces,
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
17
or plays: morphisms u∶ Y
X model plays from the initial position X, to the final one,
Y . Finally, cells model embeddings of plays, preserving initial and final position. E.g., this
could model embedding the part of a play involving a particular agent.
Individuals are intended to model agents in a position, with a role similar to that of
representable presheaves among general ones. Typically, the hom-set Dh (d, X), with d ∈ I,
models the set of agent slots of type d in the position X. For example, in the playground we
construct below for π, individuals bear the number of channels that an agent is connected
to, so that a morphism [n] → X amounts to an n-ary agent in X, i.e., an agent connected
to n channels. Accordingly, the object [n] models a position with just one n-ary agent.
Actions model moves in the game, and atomicity (P3) notably says that any play
decomposes into moves. The distinction between basic and full actions has to do with
innocence. The two notions are more or less dual: basic actions are as thin as possible, while
full ones are as wide as possible. Intuitively, an action is full when it cannot embed into a
larger one (unless possibly some agents are added), while it is basic when it cannot embed
any smaller one. As we will see, in our playground for π, we have an action for forking,
which describes how an agent x may fork into two, say x1 and x2 . This action is full, and it
has two basic sub-actions, which respectively model the passage from x to x1 and to x2 . For
another example, we also have an action for inputting on some given channel: it is obviously
basic, but in fact also full, because the only way to embed it into a wider action is to add an
agent and do a synchronisation. In view of this, it should be natural that basic actions have
individuals as their domain and codomain (P2). As alluded to in (P5), views will be defined
as composites of basic actions. Axiom (P5) intuitively enforces existence of one sub-action
for each agent in the final position of any action. Extending this to general plays will yield
an operation analogous to taking the view of a play in standard game semantics. Axiom (P9)
requires that these basic sub-actions may not be shared among different full actions.
Both (P6) and (P7) are decomposition axioms. The former says that a decomposition
of a play reflects essentially uniquely onto any subplay. The latter essentially says that basic
actions are strictly sequential: if any play of the form b ● w with b basic embeds into some
other play, then the image of w should occur after that of b. This is expressed in a slightly
convoluted way by saying that if the latter play decomposes as M ● u, then
● either b maps to M , in which case w should map to u,
● or b maps to U , in which case w should also map to u.
This should make most of the axioms rather intuitive: the others are technical, which
means that they emerged from our attempts to make things work out, but that we are not
yet able to explain them satisfactorily.
3. A pseudo double category of traces
In this section, we introduce our notion of trace, which is based on certain combinatorial
objects, close in spirit to string diagrams. We first define these string diagrams, and then use
them to define traces. Positions are special, hypergraph-like string diagrams whose vertices
represent agents and whose hyperedges represent channels. A perhaps surprising point is
that actions are not just a binary relation between positions, because we not only want to say
when there is an action from one position to another, but also how this action is performed.
This will be implemented by viewing actions from X to Y as cospans Y → M ← X in a
̃ f , whose objects we call higher-dimensional string diagrams for lack of a
certain category C
better term. The idea is that X and Y respectively are the initial and final positions, and
18
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
that M describes how one goes from X to Y . By combining such actions (by pushout), we
get a bicategory Dv of positions and traces. Finally, we recast Dv as the vertical bicategory
of a pseudo double category D.
̃ f will be a category of finite presheaves over a base
3.1. String diagrams. The category C
category, C. Let us motivate the definition of C by recalling that (directed, multi) graphs
may be seen as presheaves over the category with two objects ⋆ and [1], and two non-identity
morphisms s, t∶ ⋆ → [1]. Any such presheaf G represents the graph with vertices in G(⋆) and
edges in G[1], the source and target of any e ∈ G[1] being respectively G(s)(e) and G(t)(e),
or e⋅s and e⋅t for short. A way to visualise how such presheaves represent graphs is to compute
their categories of elements [54]. Recall that the category of elements el G for a presheaf G
over C has as objects pairs (c, x) with c ∈ C and x ∈ G(c), and as morphisms (c, x) → (d, y)
all morphisms f ∶ c → d in C such that y ⋅ f = x. This category admits a canonical functor πG
y
πG
̂ with the Yoneda embedding.
to C, and G is the colimit of the composite el G Ð→ C Ð
→C
s
t
E.g., the category of elements for y[1] is the poset (⋆, s) Ð
→ ([1], id [1] ) ←
Ð (⋆, t), which could
, where dots represent vertices, the triangle represents the edge, and
be pictured as
links materialise the graph of G(s) and G(t), the convention being that t connects to the
apex of the triangle. We thus recover some graphical intuition.
Let us give the formal definition of C for reference. We advise to skip it on a first
reading, as we then attempt to provide some intuition.
Definition 3.1. Let GC be the graph with, for all n, m, with a, b ∈ n and c, d ∈ m:
● vertices ⋆, [n], πnl , πnr , πn , νn , ♡n , τn , ιn,a , on,a,b , and τn,a,m,c,d ;
● edges s1 , ..., sn ∶ ⋆ → [n], plus, ∀v ∈ {πnl , πnr , ♡n , τn , on,a,b }, edges s, t ∶ [n] → v;
t
s
t
s
● edges [n] Ð
→ νn ←
Ð [n + 1] and [n] Ð
→ ιn,a ←
Ð [n + 1];
l
r
ρ
● edges πnl Ð
→ πn ←
Ð πnr and ιn,a Ð
→ τn,a,m,c,d ←
Ð om,c,d .
Let C be the free category on GC , modulo the equations
s ○ si = t ○ si
l○t=r○t
ρ ○ t ○ sa = ○ t ○ sc
ρ ○ s ○ sn+1 = ○ s ○ sd .
The first equation should be understood in C(⋆, v) for all n ∈ N, i ∈ n, and v ∈ ∪a,b∈n {πnl , πnr ,
♡n , τn , ιn,a , on,a,b , νn }. (This is rather elliptic: if v has the shape ιn,a or νn , s ○ si is really
si
s
⋆ Ð→ [n + 1] Ð
→ v.) The second equation should be understood in C([n], πn ) for all n, and
the last two in C(⋆, τn,a,m,c,d ), for all n, m, a ∈ n, and c, d ∈ m.
̃ f . A presheaf X
Our category of string diagrams is the category of finite presheaves C
over C is a kind of higher-dimensional graph whose components are typed by objects of C:
● X(⋆) is the set of vertices, or channels;
● X[n] is the set of agents connected to n channels (which are given by X(si ));
● X(ιn,a ) is the set of input actions by some n-ary agent on its ath channel;
● X(om,c,d ) is the set of output actions by some m-ary agent on its cth channel of its dth
channel;
● X(τn,a,m,c,d ) is the set of synchronisations between some input and some output on a
common channel;
● X(πn ) is the set of forking actions by some n-ary agent;
● and similarly for X(πnl ), X(πnr ), X(νn ), X(♡n ), and X(τn ).
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
19
(⋆, s1 ) (⋆, s2 ) (⋆, s3 )
([3], id [3] )
ls
lss1
l
s
ts1
rs
id π2
r
lss2
ρs
lt = rt
ts3
β
id τ
n,a,m,c,d
ρ
ts2
t
y′
x′
α
ρt
x
y
Figure 2: Categories of elements for [3], π2 , and τ1,1,3,2,3 , with graphical representation
To see these intuitions at work, let us compute a few categories of elements. Let us start
with an easy one, that of [3] ∈ C (recalling that we implicitly identify any c ∈ C with yc ).
An easy computation shows that it is the poset pictured in the top left part of Figure 2. We
think of it as a position with one agent ([3], id [3] ) connected to three channels, and draw it
as in the top right part, where the bullet represents the agent, and circles represent channels.
Definition 3.2. Positions are finite presheaves empty except perhaps on ⋆ and [n]s.
Let us organise positions into a category, by designing a notion of morphism. We may
equip the objects of C with a dimension: ⋆ has dimension 0, any [n] has dimension 1, all of
τn , πnl , πnr , ♡n , ιn,a , on,c,d , νn have dimension 2, πn has dimension 3, τn,i,m,j,k has dimension 4.
Definition 3.3. We accordingly define the dimension of a presheaf X on C to be the lowest
n ∈ N such that for any m ∈ C of dimension strictly greater than n, X(m) = ∅.
A position is thus equivalently a finite presheaf in [Cop , set] of dimension at most 1. An
interface is one of dimension 0.
̂ is 1-injective iff it is injective in all strictly positive dimensions.
Definition 3.4. A map in C
̂ The intuition for a morphism
A morphism of positions is a 1-injective morphism in C.
X → Y between positions is thus that X embeds into Y , possibly identifying some channels.
Definition 3.5. Positions and morphisms between them form a category Dh .
Returning to our explanation of C through categories of elements, let us consider that of
π2 . It is the poset generated by the left-hand graph in the second row of Figure 2 (omitting
base objects for conciseness). We think of it as a binary agent (lt) forking into two agents
(ls and rs), and draw it as on the right. The equation lt = rt ensures that π2l and π2r are
performed by the same agent. The graphical convention is that a black triangle stands
for the presence of id π2 , l, and r. Below, we represent just l as a white triangle with only
a left-hand branch, and symmetrically for r. Furthermore, in all our pictures, time flows
‘upwards’.
Another category of elements, characteristic of the π-calculus, is the one for synchronisation τn,a,m,c,d . The case (n, a, m, c, d) = (1, 1, 3, 2, 3) is the poset generated by the graph at
20
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
♡
♠
[p]
[p]
[m]
[n + 1]
[p]
[p]
[p + 1]
πpl
πpr
om,c,d
ιn,a
♡p
τp
νp
[p]
[p]
[m]
[n]
[p]
[p]
[p]
Figure 3: Pictures and corresponding cospans for πpl , πpr , om,c,d , ιn,a , ♡p , τp , and νp
the bottom left of Figure 2, which we will draw as on the right. The left-hand ternary agent
x outputs its 3rd channel, here β, on its 2nd channel, here α. The right-hand unary agent y
receives the sent channel on its unique channel, here α. Both agents have two occurrences,
one before and one after the action, respectively marked as x/x′ and y/y ′ . Both x and x′
are ternary here, while y is unary and y ′ , having gained knowledge of β, is binary. There are
actually three actions here, in the sense that there are three higher-dimensional elements.
The first is the output action from x to x′ , graphically represented as the middle point of
(intended to evoke the point where β enters channel α). The second is the input
action ρ from y to y ′ , graphically represented as the middle point of
(where β
exits channel α). The third action is the synchronisation itself, which ‘glues’ the other two
together, as represented by the squiggly line.
We leave the computation of other categories of elements as an exercise to the reader.
The remaining string diagrams are depicted in the top row of Figure 3, for p = 2 and
(n, a, m, c, d) = (1, 1, 3, 2, 3). The first two are views, in the game semantical sense, of the
fork action π2 explained above. The next two, om,c,d (for ‘output’) and ιn,a (for ‘input’),
respectively are views for the sender and receiver in a synchronisation action. The τp action
is a silent, dummy action standard in π-calculus. The ♡p action is the special ‘tick’ action
used for defining fair testing equivalence. The last one, νp , is a channel creation action.
3.2. From string diagrams to actions. In the previous section, we have defined our
̃ f , and provided some intuition on its objects. The next step
category of string diagrams as C
is to construct a bicategory whose objects are positions, and whose morphisms represent
traces. We start in this section by defining in which sense higher-dimensional objects of C
represent actions, and continue in the next one by explaining how to compose actions to
form traces. Actions are defined in two stages: seeds, first, give their local form, their global
form being given by embedding into bigger positions.
To start with, until now, our string diagrams contain no information about the ‘flow of
time’, although we mentioned it informally in the previous section. To add this information,
for each string diagram M representing an action, we define its initial and final positions,
s
t
say X and Y , and view the whole action as a cospan Y Ð
→M ←
Ð X. We have taken care, in
drawing our pictures before, of placing initial positions at the bottom, and final positions at
the top. So, e.g., the initial and final positions for the example synchronisation of Figure 2
are as follows.
↝
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
21
They map into (the representable presheaf over) τ1,1,3,2,3 , yielding the cospan
s
t
Y Ð
→ τ1,1,3,2,3 ←
Ð X.
We leave it to the reader to define, based on the above pictures, the expected cospans for
forking and synchronisation
[p] ∣ [p]
[m] c,d ∣ a,n+1 [n + 1]
πp
τn,a,m,c,d
[p]
[m] c ∣ a [n]
plus the remaining ones specified in the bottom row of Figure 3. Initial positions are at the
bottom, and we use:
Notation 3.6. We denote by [m] a1 ,...,ap ∣ c1 ,...,cp [n] the position consisting of an m-ary agent
x and an n-ary agent y, quotiented by the equations x ⋅ sak = y ⋅ sck for all k ∈ p. When both
lists are empty, by convention, m = n and the agents share all channels in order.
Definition 3.7. These cospans are called seeds.
We now define actions from seeds by embedding the latter into bigger positions. E.g.,
we allow a fork action to occur in a position with more than one agent.
̂ is F (⋆) ⋅ ⋆, the F (⋆)-fold coproduct of
Definition 3.8. The interface IF of a presheaf F ∈ C
⋆ with itself, or in other words the position consisting solely of F ’s channels. The interface
s
t
of a seed Y Ð
→M ←
Ð X is IX .
Since channels occurring in the initial position remain in the final one, we have for each
seed a cone from IX to the seed. For any morphism of positions IX → Z, pushing the cone
along IX → Z using the universal property of pushout as in
Y
Y′
M
M′
IX
(3.1)
Z
X′
X
yields a new cospan, say Y ′ → M ′ ← X ′ .
Definition 3.9. Let actions be all such pushouts of seeds.
Intuitively, taking pushouts glues string diagrams together. Let us do a few examples.
[ls,rs]
lt
Example 3.10. The seed [2] ∣ [2] ÐÐÐ→ π2 ←
Ð [2] has as interface the presheaf I[2] = ⋆ + ⋆,
consisting of two channels, say a and b. Consider the position [2] + ⋆ consisting of an agent
y connected to two channels b′ and c, plus an additional channel a′ . Further consider the
map h∶ I[2] → [2] + ⋆ defined by a ↦ a′ and b ↦ b′ . The pushout
x1
I[2]
is
π2
x2
[2] + ⋆
M′
.
a=a′
b=b′
x
y
c
22
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
y ′′
c
x1
y1
x2
a=a′
y2
c
b=b′
x
y′
a
y
a
b
x
y
c
z
Figure 4: Example traces
The meaning of such an action is that x forks while y is passive.
Example 3.11. Because we push along initial channels, the interface of a seed may not
contain all involved channels. E.g., in an input action (not part of any synchronisation), the
received channel cannot be part of the initial position.
3.3. From actions to traces. Having defined actions, we now define their composition to
̃ f ), the bicategory which
yield our bicategory Dv of positions and traces. Consider Cospan(C
has as objects all finite presheaves on C, as morphisms X → Y all cospans X → U ← Y , and
obvious 2-cells. Composition is given by pushout, and hence is not strictly associative.
Notation 3.12. By convention, the initial position is the target of the morphism in
̃ f ). We denote morphisms in Cospan(C
̃ f ) with special arrows Y
Cospan(C
X; composition
●
and identities are denoted with ● and id , which matches the notation of pseudo double
categories (Section 2.4).
̃ f which is isomorphic to some finite, possibly
Definition 3.13. A trace is any cospan in C
̃ f ). Let Dv denote the subbicategory of Cospan(C
̃f )
empty composite of actions in Cospan(C
obtained by restricting to positions, traces, and 1-injective 2-cells.
Thus, arrows X → Y in Dh denote embeddings of X into Y (up to identification of
channels), whereas arrows Y
X in Dv denote traces with X initial and Y final. Intuitively,
composition in Dv glues string diagrams on top of each other, which yields a truly concurrent
notion of trace: the only information retained in a trace about the order of occurrence of
actions is their causal dependencies.
Example 3.14. Composing the action of Example 3.10 with a forking action by y yields
the first string diagram of Figure 4, which shows that the ordering between remote actions
is irrelevant. To illustrate how composition retains causal dependencies between actions,
consider the second string diagram. It is unfolded for readability: one should identify both
framed nodes, resp. both circled ones. In the initial position, there are channels a, b, and c,
and three agents x(a, b), y(b), and z(a, c) (channels known to each agent are in parentheses).
In a first action, x sends a on b, and y receives it. In a second action, z sends c on a, and
the avatar y ′ of y receives it. The second action is enabled by the first, by which y gains
knowledge of a.
Before going on to construct the base double category for our playground, let us observe
the following two basic facts about traces.
s
t
Lemma 3.15. For any trace Y Ð
→U ←
Ð X, s and t are monos.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
23
(EXPANDED VERSION)
Proof. We proceed by induction on the number of actions involved in any decomposition of
U . The base case is trivial. For the induction step, because composition of cospans is by
pushout, the result follows from the induction hypothesis, stability of monos under pushout
and composition, and the fact that the result holds for actions. The latter in turn follows
from the fact that monos are stable under pushout (again!), the pushout lemma, and the
fact that the result holds for seeds, which holds by case inspection.
s
t
Lemma 3.16. For any trace Y Ð
→U ←
Ð X, s⋆ is an isomorphism.
Proof. Similar to the previous proof.
The intuition behind the last lemma is that no channel is forgotten during the play.
3.4. The main double category. At last, we define the base double category D of our
playground for the π-calculus. It is a sub-double category of a double category of cospans
̂
in C.
Consider the double category D0 with
● positions as objects,
● horizontal morphisms X → Y given by all natural transformations h∶ X → Y ,
s
t
̂
● vertical morphisms X
Y given by cospans X Ð
→U ←
Ð Y in C,
● and double cells U → V given by commuting diagrams
X′
k
Y′
sU
U
sV
l
tU
X
(3.2)
V
tV
h
Y.
Definition 3.17. Let D denote the sub-double category of D0 obtained by restricting
● vertical morphisms to traces,
● horizontal morphisms to 1-injective maps,
● double cells to diagrams (3.2) in which k, l, and h are 1-injective.
Proposition 3.18. D indeed forms a sub-double category of D0 , i.e., is closed under all
composition operations.
Proof. The only non-obvious point is that double cells in D are closed under vertical
composition, and in particular that the middle component of the composite is 1-injective.
This follows from Lemma 3.20 and Corollary 3.22 below.
Definition 3.19. Let V0 denote the set of ‘t’-legs (i.e., lower legs) of seeds.
Lemma 3.20. For any morphism (3.2) in D0H , if U and V are traces, then the upper square
is a pullback.
s
t
Proof. For any trace Y Ð
→P ←
Ð X and n ∈ N, Y [n] consists of all elements of P [n] which
are not in the image of (the action of) any map in V0 .
Now, consider any double cell as in (3.2). Because sV is monic, U ×V Y ′ may be chosen
to be just l−1 (Y ′ ) ⊆ U . By Lemma 3.15 and standard cancellation properties of monos,
24
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
the mediating arrow X ′ → U ×V Y ′ is mono. To show that it is epi, we proceed pointwise.
Over ⋆, the result follows from sU and sV being isomorphisms (Lemma 3.16). Over [n],
if x ∈ (U ×V Y ′ )[n] then x ∈ U [n], and l(x) ∈ V [n] is not in the image of (the action of)
any t ∈ V0 . But if there existed y such that y ⋅ t = x, then by naturality we would have
l(y) ⋅ t = l(x), contradicting the latter. Any natural transformation being both epi and mono
is an isomorphism, hence the result.
Lemma 3.21. In Set, consider any cube
I
B
A
I′
C
B′
A′
f
C ′,
with the marked pushouts and pullback, and with all arrows mono except perhaps f . Then, f
is also mono and the front square is also a pullback.
Proof. Any such cube is naturally isomorphic to some cube of the shape
I +R
I
I +Y
injl
I +Y +R
injl +k
I+h
I +X
f =I+h+k
I +X +S
injl
I +X +Z
I + X + Z + S,
the only non-trivial point being that the map I + R → I + X + S has the given shape. But
this is because we know that its pullback along I + X → I + X + S is injl , so the image of R
has to lie in S.
Corollary 3.22. Consider any cube
X
B
A
X′
C
B′
A′
f
C ′,
̂ in which all arrows except perhaps f are 1-injective, and the marked squares are
in C
pushouts, resp. pullbacks. Then f is also 1-injective.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
25
Proof. We proceed pointwise. On any object C of dimension > 0, we obtain a diagram in
sets for which the lemma applies.
4. Codomain is a fibration
In this section, we prove that the double category D of traces constructed in the previous
section satisfies the primary axiom for playgrounds, namely that the codomain functor
DH → Dh is a fibration. We proceed as follows. We first define a (strong) factorisation
̂ (Section 4.1), from which we derive in Section 4.2 an intermediate sub-double
system on C
category D ↪ D1 ↪ D0 . We further show that by the properties of factorisation, the codomain
functor D1H → D1h is a fibration. Finally, we want to show that DH → Dh is a fibration by
proving that traces are stable under relevant cartesian lifting in D1 , i.e., cartesian liftings of
any trace along any morphism in Dh are again in DH . We first check this for seeds, by case
analysis, in Section 4.3. In order to generalise to actions M ∶ Y
X, the basic idea is to
(1) decompose X as a pushout of X0 , where the generating seed M0 takes place, and Z,
which is passive;
(2) decompose the morphism along which we want to restrict, say h∶ X ′ → X, accordingly,
say as h0 ∶ X0′ → X0 and hZ ∶ Z ′ → Z;
(3) restrict M0 along h0 to obtain P0′ ;
(4) recompose a trace P ′ from P0′ and Z ′ ;
(5) check that P ′ admits a cartesian morphism to M .
Step (4) is non-trivial, so we devote Section 4.4 to it. It works as a kind of formal opposite to
restriction, as we essentially lift P0′ along h0 ∶ X0′ → X ′ . We call this an oplifting of P0′ along
h0 , by analogy with lifting in opfibrations. However, opliftings do not enjoy the relevant
universal property (opcartesianness, which is dual to cartesianness). Instead, we find that
opliftings are in fact cartesian! In Section 4.5, we use opliftings to show that actions are
stable under restriction, following the above plan. Finally, we extend the result to arbitrary
traces in Section 4.6.
4.1. A factorisation system. Let us start by defining the (strong) factorisation system
̂ on which the intermediate sub-double category D1 will be based. The idea is
(V, H) on C,
that all three components of cartesian morphisms in DH are in H, while t-legs of vertical
morphisms are in V. The cartesian lifting of any V ∶ Y ′
Y as in (3.2) along any h∶ X → Y
is then given by factoring tV ○ h as l ○ tU with tU ∈ V and l ∈ H to obtain
X′
k
Y′
sU
U
sV
l
tU
X
(4.1)
V
tV
h
Y,
where the upper square is a pullback.
We recall from Definition 3.19 that V0 denotes the set of ‘t’-legs (i.e., lower legs) of
seeds. Following Bousfield’s [8] construction of ‘cofibrantly generated’ factorisation systems,
we define H = V0 to be the class of maps f such that for any t ∈ V0 and commuting square
̂ → , there exists a unique filler h making the following diagram commute:
(u, v)∶ t → f in C
26
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
X
t
Y
u
X′
h
v
f
′
Y .
In this situation, one says that f is right-orthogonal to t, and t is left-orthogonal to f , which
is denoted by t f .
We finally define V = H to consist of all maps which are left-orthogonal to any map in
H. Of course, we have V0 ⊆ V. The following is an application of [8, Theorem 4.1]:
Proposition 4.1. The pair (V, H) forms a factorisation system.
What does that mean? Here is a modern definition [29]:
Definition 4.2. The classes of maps V and H form a factorisation system iff V = H,
V = H, and any arrow factors as h ○ v with h ∈ H and v ∈ V.
̂ admits a
In the case where H = V0 and V = H, Bousfield proves that any map in C
factorisation using a transfinite construction (a so-called small object argument). But here
we will only need factorisations of particular morphisms, which we will actually be able to
calculate by hand. Bousfield’s results include:
Lemma 4.3. V is stable under pushout and composition, contains all isomorphisms, and
enjoys the right cancellation property, i.e., if v ∈ V and f v ∈ V, then f ∈ V.
H is stable under pullback and composition, contains all isomorphisms, and enjoys the
left cancellation property, i.e., if h ∈ H and hf ∈ H, then f ∈ H.
Remark 4.4. Stability under pushout is ambiguous here: we mean that for any pushout
X
v
Z
Y
v′
T,
if v ∈ V, then v ′ ∈ V. Stability under pullback is defined dually.
4.2. A first ‘fibred’ double category. We now make concrete the idea evoked in the
previous section, of using our factorisation system to obtain a codomain fibration. Consider
the sub-double category D1 of D0 obtained by restricting vertical morphisms to cospans
s
t
XÐ
→U ←
Ð Y with t ∈ V. Its vertical morphisms are stable under composition and contain
identities by Lemma 4.3, i.e.:
Lemma 4.5. D1 forms a sub-double category of D0 .
Lemma 4.6. Traces are in D1v , i.e., we have D ⊆ D1 .
Proof. By Lemma 4.3.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
27
The main interest of introducing D1 is:
Lemma 4.7. The codomain functor cod∶ D1H → D1h is a fibration in which a double cell (3.2)
is cartesian iff l ∈ H and the upper square is a pullback.
Proof. Let us show that the lifting candidate computed in (4.1) is cartesian. Indeed, consider
any double cell (4.1), and any morphism from some vertical morphism Z ′ → W ← Z to V
whose bottom component factors through h∶ X → Y . By unique lifting in (V, H), we obtain
a unique dashed arrow making
Z
X
U
Y
W
V
commute. We finally obtain the desired arrow Z ′ → X ′ by universal property of pullback.
Conversely, any cartesian double cell, being isomorphic to such a lifting, satisfies the
conditions.
As a final observation, let us record:
Lemma 4.8. Any morphism in Dh is automatically in H.
Proof. Indeed, consider any h∶ X → Y in Dh . There cannot be any commuting square
u
X0
X
t
h
U
Y
v
with t ∈ V0 , because U is a representable presheaf of dimension > 1 and Y has dimension ≤ 1,
so there cannot be any v∶ U → Y .
4.3. Restriction of seeds. We now show that restrictions of seeds (in the sense of D1 ) are
traces.
h
t
Lemma 4.9. Consider any diagram X ′ Ð
→XÐ
→ M , where t ∈ V0 and h ∈ Dh (X ′ , X). Its
t′
h′
factorisation X ′ Ð
→ M ′ Ð→ M with t′ ∈ V and h′ ∈ H is such that h′ is 1-injective and the
obtained restriction is a trace of length at most 2. If X ′ is an individual, i.e., a position of
shape [n], then it is isomorphic to some seed. If X ′ is an interface, then the restriction is
an equivalence (in Dv ).
Proof. We proceed by case analysis. In each case, one has to check that h′ is 1-injective,
that X ′ individual implies t′ seed, that X ′ interface implies that t′ is an isomorphism, and
that the upper leg of the obtained cospan is as expected: this is routine so we mention it
here once and for all.
Let us first treat the case where M = yc , for c not of the shape τn,i,m,j,k . Then, we
have X = [n] for some n. If idc ∈ Im(th), then X ′ ≅ [n] + I for some interface I (since h is
1-injective). Consider the diagram
28
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
M +I
[id,tk]
M
t+idI
t
[n] + I
[n].
h=[id,k]
The map t + idI is in V by Lemma 4.3, so we just have to prove that [id, tk] is in V0 , which
is a simple verification.
If now idc ∉ Im(h), then X ′ is an interface, and the relevant factorisation is
t○h
X′
M
X
(4.2)
t
id
′
h
[n],
because t ○ h is easily checked to be in V0 .
The case of τn,i,m,j,k is a bit more complicated. Here, t is actually t0 = [ρt, t]. First of
all, if X ′ is an interface, then we obtain a factorisation analogous to (4.2). Consider now
the case where Im(h) contains both agents of [n] i ∣ j [m]. Let x denote the n-ary one and y
denote the m-ary one (in X ′ ). If x ⋅ si = y ⋅ sj , then X ′ = ([n] i ∣ j [m]) + I for some interface
I and the required factorisation is easily seen to be
[id,t0 k]
τn,i,m,j,k + I
τn,i,m,j,k
t0
t0 +id I
([n] i ∣ j [m]) + I
h=[id,k]
[n] i ∣ j [m].
Consider now the case where X ′ still contains both agents but x ⋅ si ≠ y ⋅ sj . Then
X = [n] + [m] + I for some interface I, and the required factorisation is
′
[ρ,,t0 k]
ιn,i + om,j,k + I
τn,i,m,j,k
t0
t+t+id I
[n] + [m] + I
[x,y,k]
[n] i ∣ j [m].
The only non-trivial point here is to show that [ρ, , t0 k] is in V0 , which easily reduces
to showing that there is no commuting square
[n] i ∣ j [m]
u
t0
τn,i,m,j,k
ιn,i + om,j,k + I
[ρ,,t0 k]
v
τn,i,m,j,k ,
which is true because there is no such u.
The cases where X ′ only contains one agent of [n] i ∣ j [m] are similar to the latter case:
if it contains x then the factorisation is through ιn,i , and otherwise it is through om,j,k .
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
29
4.4. Opliftings. We now aim at extending Lemma 4.9 from seeds to actions. Consider an
arbitrary action Y → M ← X, obtained by pushing some seed Y0 → M0 ← X0 along IX0 → Z.
Consider now a morphism h∶ X ′ → X in Dh , along which we wish to restrict M . As explained
at the beginning of Section 4, our strategy is to consider the pullback
X0′
X0
X′
X.
Indeed, the part of X really concerned by M is the image of X0 , so we would like to first
restrict M0 to X0′ using Lemma 4.9, and then extend it to X ′ . The present section is devoted
to constructing the second step, the extension of (M0 )∣X0′ to X ′ , which we call an oplifting
of (M0 )∣X0′ along X0′ → X ′ .
In the general case, we want to compute the oplifting of Y → U ← X along h∶ X → X ′ .
Because h is 1-injective, we can complete the solid part of
IX
Z
⊆
⊆
′
X
X
h
′
into a pushout. Indeed, we take Z(⋆) = X (⋆), Z[n] = X ′ [n] ∖ Im(h[n] ) for all n, and
IX → Z is uniquely determined by h⋆ . In passing, we have:
Lemma 4.10. This pushout is uniquely determined up to canonical isomorphism by h alone.
Then, we observe that, because U is a trace, IX → X → U factors through Y → U .
Definition 4.11. Let the oplifting of U ∶ Y
as in (3.1) by pushing U along IX → Z.
X along h∶ X → X ′ be the cospan obtained
All horizontal maps are 1-injective by construction, and we have:
Lemma 4.12. The obtained cospan Y ′ → U ′ ← X ′ is a trace.
Proof. We start by showing that U ′ is an action if U is. So assume U is obtained by pushing
a seed Y0 → M0 ← X0 along some IX0 → Z0 . Then, Z0 → X is surjective on ⋆ because
IX0 → X0 is and epis are stable under pushout in presheaf categories. Thus, IX → X factors
through Z0 → X. Let Z ′′ denote the pushout
IX
Z0
Z
Z ′′ .
By the pushout lemma, Y ′ → U ′ ← X ′ is isomorphic to the cospan obtained by pushing
Y → U ← X along Z0 → Z ′′ . By the pushout lemma again, it is isomorphic to the cospan
obtained by pushing Y0 → M0 ← X0 along IX0 → Z0 → Z ′′ . Thus, it is indeed an action.
We now prove the general case by induction on ∣U ∣. This is trivial if U is isomorphic to
an identity. If now U is a composite Z
along h to obtain a double cell, say
V
Y
M
X, then we compute the oplifting of M
30
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Y
M
X
h′
αM
h
Y′
M′
′
X,
by pushing M along some morphism IX → Z1 making
IX
X
Z1
h
X′
into a pushout.
The crucial insight is then that by computing the following pushout Z2 and applying its
universal property, we obtain a diagram
IX
IY
Y
(4.3)
Z1
Z2
Y
′
whose exterior is a pushout by construction. So by the pushout lemma the right-hand square
is again a pushout, which by Lemma 4.10 is the unique pushout along which to push V to
compute the oplifting of V along h′ , say
Z
V
Y
h′′
αV
h′
Z′
V′
′
Y .
By induction hypothesis, V ′ is again a trace. But by the pushout lemma again, αV is also
what we obtain by pushing V along the exterior rectangle of (4.3). A bunch of applications
of the pushout lemma finally yields that αM ● αV is the oplifting of M ● V along h, so M ′ ● V ′
is a trace by construction.
Although opliftings have an opcartesian flavour, they are in fact not opcartesian in
general, and moreover opcartesian liftings do not exist in general.
Example 4.13. Consider the oplifting of the seed ι1,1 along [1] ↪ [1] 1 ∣ 1 [1], say ι1,1 1 ∣ 1 [1],
whose final position is [2] 1 ∣ 1 [1]. To see that it is not opcartesian, consider the diagram
[2] 1,2 ∣ 1,1 [1]
[2]
[2] 1 ∣ 1 [1]
τ1,1,1,1,1
ι1,1
ι1,1 1 ∣ 1 [1]
[1] 1 ∣ 1 [1]
[1]
[1] 1 ∣ 1 [1].
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
31
In this case, opcartesianness would mean finding dashed arrows rendering the diagram
commutative. There is indeed a unique arrow ι1,1 1 ∣ 1 [1] → τ1,1,1,1,1 making the corresponding
square commute, but then no arrow [2] 1 ∣ 1 [1] → [2] 1,2 ∣ 1,1 [1] fits. Indeed, there is only one
arrow of the given type, and it does not make the relevant square
[2] 1 ∣ 1 [1]
[2] 1,2 ∣ 1,1 [1]
ι1,1 1 ∣ 1 [1]
τ1,1,1,1,1
commute, because one side (down; right) of the square maps the unary player to ρ ○ t, while
the other (right; down) maps it to ρ ○ s. This shows that ι1,1 1 ∣ 1 [1] is not an opcartesian
lifting of ι1,1 . But in fact, ι1,1 1 ∣ 1 [1] and τ1,1,1,1,1 are two liftings of ι1,1 along [1] ↪ [1] 1 ∣ 1 [1].
Thus, any opcartesian lifting should have a final position mapping both to [2] 1 ∣ 1 [1] and
[2] 1,2 ∣ 1,1 [1], hence containing just one, binary player: no trace can meet this requirement.
However, even though they are not opcartesian, opliftings are in fact cartesian. Let us
now show this, starting with a few preliminary results.
Definition 4.14. Let Agents(X) = ∑n∈N X[n] denote the set of agents of any position X.
Lemma 4.15. For any seed Y → C ← X, the morphism
[x](n,x)
[n] ÐÐÐÐ→ X
∑
(n,x)∈Agents(X)
is epi.
Proof. This is trivial except in dimension 0, where it holds by case inspection.
Corollary 4.16. For all arrows as in
h
X
U
h′
t
C
U
g
f
′
̂ such that f is 1-injective, t ∈ V0 , and f h = gt = f h′ , we have h = h′ .
in C
Proof. We construct the diagram
he
∑(n,x)∈Agents(X) [n]
U
h′ e
e
h
X
t
C
U
h′
g
f
U ′,
and observe that h′ e = he by 1-injectivity of f , hence h = h′ because e is epi by Lemma 4.15.
32
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Lemma 4.17. Opliftings of traces are cartesian.
f
Proof. Consider any oplifting U Ð
→ U ′ of Y → U ← X along, say X → X ′ . By Lemma 4.7,
it is enough to show that the middle arrow U → U ′ is in H and that its upper square is a
pullback. The latter follows from Lemma 3.20. So we just have to show that any square
Z
u
U
v
U′
t
C
with t ∈ V0 admits a unique lifting C → U . By the Yoneda lemma, v amounts to an element of
U ′ (C) of dimension > 1, but by construction U and U ′ have exactly the same such elements.
This yields a candidate lifting, say k, which is unique by 1-injectivity and makes the bottom
triangle commute by construction. The top one finally commutes by Corollary 4.16 with
h = u and h′ = kt.
4.5. Restriction of actions. Let us now extend Lemma 4.9 from seeds to actions, following
the strategy sketched at the beginning of Section 4.
s
t′
t
→M ←
Ð X and h ∈ Dh (X ′ , X), the factorisation X ′ Ð
→
Lemma 4.18. For any action Y Ð
h′
P ′ Ð→ M of th with t′ ∈ V and h′ ∈ H is such that h′ is 1-injective and the obtained restriction
is a trace of length at most 2. If X ′ is an individual then it is either a seed or an equivalence;
if X ′ is an interface then it is an equivalence.
Proof. Consider any action Y → M ← X obtained by pushing the following seed-withinterface along I → Z:
I
Y0
M0
X0 .
By Lemma 4.17, the morphism M0 → M is cartesian.
Consider the pullback of the bottom square below along h∶ X ′ → X to obtain the top
square
I′
Z′
X0′
I
X′
Z
X0
f
X,
which, because presheaf categories are adhesive [46] and I → X0 is mono, is again a pushout.
Furthermore, consider the front face, which is a pullback in Dh . By Lemma 4.9, restricting
M0 along X0′ → X0 yields a trace, say Y0′ → P0′ ← X0′ with a morphism to Y0 → M0 ← X0 in
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
33
DH . Since it is a trace, I ′ ↪ X0′ ↪ M0′ factors through Y0′ ↪ M0′ . Pushing Y0′ → P0′ ← X0′
along I ′ → Z ′ , we obtain a trace Y ′ → P ′ ← X ′ (we may choose X ′ as initial position because
the top square above is a pushout) with a morphism from Y0′ → P0′ ← X0′ in DH , which is
an oplifting, hence cartesian (by Lemma 4.17). Then, by universal property of pushout we
obtain a unique morphism f ∶ P ′ → M making the following cube commute:
I′
Z′
P0′
P′
I
f
Z
M0
M.
By Corollary 3.22, f is 1-injective, which entails that the induced morphism of traces is
in DH .
We now need to show that P ′ → M is cartesian, which by Lemmas 3.20 and 4.7 amounts
to showing that its middle arrow f ∶ P ′ → M is in H. To this end, consider any morphism
̂ → . First of all, because M0 → M is identity
t∶ Z ′′ → C in V0 and morphism (u, v)∶ t → f in C
in dimensions > 1, the morphism v uniquely factors through M0 → M . Furthermore, in
all cases where f0 ∶ P0′ → M0 is identity in dimensions > 1, the Yoneda lemma entails that
C → M0 uniquely factors through P0′ . This yields a diagram
u
l
Z ′′
?
P0′
k
C
f0
M0
g
P′
f
M
v
which commutes except perhaps for the upper part marked ‘?’. But the latter also commutes
by Corollary 4.16 with h = u and h′ = gl. We thus obtain a lifting, which is unique by
1-injectivity of f .
So what happens when is f0 non-identity in dimensions > 1? By inspection of the proof
of Lemma 4.9, except for the easy case where P0′ ≅ X0′ , this is when M0 = τn,i,m,j,k for some
n, m, i, j, k, and P0′ has one of the shapes ιn,i + J, om,j,k + J, or ιn,i + om,j,k + J, for some
interface J. In the first two cases, C ≠ τn,i,m,j,k because there can be no u∶ [n] i ∣ j [m] → P ′
(one agent is missing in P ′ ), so the previous argument applies. In the third case, letting x
and y respectively denote the n- and m-ary initial agents in P0′ and a = x ⋅ si and a′ = y ⋅ sj
the corresponding channels, one easily shows that g(a) ≠ g(a′ ), so again there can be no
u∶ [n] i ∣ j [m] → P ′ . Thus, C ≠ τn,i,m,j,k and the previous argument again applies.
34
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
4.6. Restriction of traces. So far, we have shown that seeds and actions admit cartesian
liftings in DH . We now show that it is also the case for arbitrary traces. We proceed by
induction on the length of the considered trace, which requires a few preliminary results.
Lemma 4.19. In sets, for any commuting diagram
I
A
C
J
B
D
whose exterior rectangle is a pullback, with the marked pushout and monos, the right-hand
square is also a pullback.
Proof. We check the universal property of pullback for A, relative to 1, which is enough in
sets. So consider any commuting square
1
c
C
b
B
D.
First, we observe that there is at most one mediating arrow 1 → A, because A → B is mono.
If b has an antecedent in A, say a, then because C → D is mono, a makes both required
triangles commute and we are done.
Otherwise, by surjectivity of A + J → B, b admits an antecedent in J, i.e., there exists
j
j∶ 1 → B such that b is 1 Ð
→ J → B, then we have a cone to C → D ← J, so we apply the
universal property of pullback to obtain i as in
1
c
i
I
A
C
J
B
D
j
b
i
making everything commute and so 1 Ð
→ I → A suits our needs.
Corollary 4.20. In any presheaf category, in any commuting cube
I
B
A
I′
C
B′
A′
f
C ′,
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
35
with the marked pushouts, pullback, and mono, the front square is also a pullback.
Proof. It suffices to show the result in sets, as all involved properties are pointwise in presheaf
categories. First, as monos are stable under pullback and pushout, I → B, A → C, and
A′ → C ′ are also monos. Furthermore, pushouts along monos are also pullbacks, so the top
and bottom faces are also pullbacks. By the pullback lemma, the rectangle
I
I′
A′
B
B′
C′
is a pullback. The previous lemma thus entails the result.
Lemma 4.21. For any commuting diagram as the solid part of
V′
T
t
U′
h
k
C
f
V
U
with t ∈ V0 , h ∈ H, and f 1-injective, there is a unique lifting k as shown.
Proof. Because h ∈ H, there is a unique map from k ′ ∶ C → V ′ making both triangles commute.
By composing k ′ with V ′ → U ′ , we obtain a lifting k for the desired square. Uniqueness
follows from 1-injectivity of f .
Lemma 4.22. The codomain functor cod∶ DH → Dh is a subfibration of cod∶ D1H → D1h .
s
t
Proof. First of all, it is sufficient to prove that given any trace Y Ð
→ U ←
Ð X and h ∈
Dh (X ′ , X), any cartesian lifting in D1 lies in D, i.e., the obtained vertical morphism is again
a trace and the double cell to U lies in D (i.e., all its components are 1-injective). Indeed,
mediating morphisms computed in D1 are automatically in D by cancellation.
We proceed by induction on U . If U is an equivalence, then the result is obviously true.
Otherwise, U = M ● V for some action M and trace V . Let us call Z the final position of M .
By Lemma 4.18, M admits a lifting P ′ along h, with a final position Z ′ , and Z ′ → Z and
fM ∶ P ′ → M are 1-injective. By induction hypothesis, V admits a lifting V ′ along Z ′ → Z
with a double cell to V in DH . Therefore, considering the composite P ′ ● V ′ , we have a
commuting diagram as in Figure 4.6, where f is obtained by universal property of pushout.
Because pushouts along monos are also pullbacks in presheaf categories, both marked
pushouts are also pullbacks. Furthermore, by Lemma 3.20, Z ′ = P ′ ×M Z and Y ′ = (V ′ ×V Y ),
as shown. Also, by Corollary 4.20, V ′ = (P ′ ● V ′ ) ×M ●V V , as shown. Furthermore, by
Proposition 3.18, f is 1-injective.
By Lemmas 4.7 and 3.20, it suffices to show that f is in H, i.e., that it is right-orthogonal
t
to any T Ð
→ C in V0 . Consider any commuting square
T
u
f
t
C
P′ ● V ′
v
M ● V.
Since M ●V is the coproduct of M and V in dimensions greater than 1 and C is a representable
of dimension greater than 1, we have that v factors either through t2 or s2 .
36
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Y′
Y
V′
s′2
V
M ●V
Z
t′2
t2
s0
fM
P′
s2
t1
f
P′ ● V ′
Z′
s1
fV
M
t0
X′
X,
Figure 5: Diagram for induction step in Lemma 4.22
If v factors through s2 , then by universal property of pullback we find a map u′ ∶ T → V ′
making
u
T
u′
P′ ● V ′
V′
t
C
V
M ●V
s2
v
commute. Then, by Lemma 4.21, we find a unique lifting as desired.
If v factors as t2 v ′ , then by Lemma 4.21, it is sufficient to show that there is a map
u′ ∶ T → P ′ making
u
T
u′
t′2
P′
P′ ● V ′
f
t
C
M
v′
M ●V
t2
v
x
commute. To that end, it is sufficient to show that for every [n] Ð
→ T , there is a map
fx
[n] Ð→ P ′ such that
[n]
x
T
P
(4.4)
u
fx
′
P ●V
′
t′2
commutes. Indeed, if that is the case, then the square
′
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
∑(n,x)∈Agents(T ) [n]
[fx ](n,x)∈Agents(T )
P
(EXPANDED VERSION)
[x](n,x)∈Agents(T )
37
T
u′
u
P ●V′
′
′
t′2
also commutes, and since its bottom map is mono and its top map is epi by Lemma 4.15,
there is a unique lifting u′ ∶ T → P ′ making both triangles commute. The square
T
u′
P′
t
C
v′
M
also commutes because it commutes when composed with t2 , which is mono.
fx
x
So we now need to show that for every [n] Ð
→ T , there is a map [n] Ð→ P ′ making
[t′2 ,s′2 ]
the square (4.4) commute. Because P ′ + V ′ ÐÐÐ→ P ′ ● V ′ is epi and [n] is a representable
ux
presheaf, [n] Ð→ P ′ ● V ′ factors through either t′2 or s′2 .
If it factors through s′2 , say as s′2 x′ , then we have
t2 (v ′ tx) = vtx = f ux = f s′2 x′ = s2 fV x′ ,
so by universal property of the pullback Z there exists a unique x′′ ∶ [n] → Z such that
s0 x′′ = v ′ tx and t1 x′′ = fV x′ . We thus obtain a commuting diagram
[n]
x′′
Z
s0
x
T
t
C
v′
M,
which is impossible because v ′ tx, as one of the agents performing action C in M , cannot
remain in the final position of M .
Thus, ux factors through t′2 and we are done.
5. A playground for π
In the previous section, we have proved the main playground axiom, asserting that the
codomain functor DH → Dh of the double category of traces constructed in Section 3 is a
fibration. We now prove the remaining axioms. In Section 5.1, we equip D with playground
structure and prove that it satisfies all the needed axioms, except both decomposition axioms
((P6) and (P7)) which require a bit more work. In Section 5.2, we establish a correctness
̂ is a trace. We then use this criterion in
criterion detecting when a given cospan in C
Section 5.3 to prove both remaining axioms.
38
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
5.1. A candidate playground. So we start in this section by defining the needed additional
structure on D.
Definition 5.1. We recall from Lemma 4.9 that I, the set of individuals, consists of
representable positions [n]. Let B, the full subcategory of basic actions, span all seeds of
shape τn , πnr , πnl , νn , ♡n , ιn,i , or om,j,k . Full actions (notation F) are all actions obtained
from seeds of shape τn , πn , νn , ♡n , ιn,i , om,j,k , or τn,i,m,j,k . Closed-world actions are all
actions obtained from seeds of shape τn , πn , νn , ♡n , or τn,i,m,j,k . Let W denote the graph
with positions as vertices and closed-world actions between them as edges (the initial position
being the target). Finally, all decompositions of any trace U into actions have the same
length which we denote by ∣U ∣.
Here is a summary of which actions
τn πnl πnr
Basic ✓ ✓ ✓
Full ✓
Closed-world ✓
are basic, full and closed-world:
πn ιn,i om,j,k τn,i,m,j,k νn
✓
✓
✓
✓ ✓
✓
✓
✓
✓
✓
✓
♡n
✓
✓
✓ .
Remark 5.2. The definition matches the explanations following Definition 2.30. Basic
actions are as small as possible, which here means they start from one agent and only retain
one agent in the final position. Full actions are those that retain all possible agents in the
final position. So the only kind of actions which are not basic are πn and τn,i,m,j,k . They
each have one sub-action for each agent in their final positions, namely πnr and πnl for πn ,
and ιn,i and om,j,k for τn,i,m,j,k . All of these sub-actions are basic, but only ιn,i and om,j,k
are full. Finally, a new class of actions appear here, that of closed-world actions. Intuitively,
it consists of those actions that do not involve any interaction with the environment. Or,
in other words, those that cannot be extended, even by adding new agents. E.g., ιn,i may
be completed to τn,i,m,j,k by adding an agent, hence is not closed-world. But τn,i,m,j,k is.
Closed-world actions will be at the basis of our semantic notion of testing equivalence
(Definition 6.24).
Lemma 5.3.
● (P1) I, viewed as a subcategory of Dh , is discrete. Basic actions have no non-trivial automorphisms in DH . Vertical identities on individuals have no non-trivial endomorphisms.
● (P2) (Individuality) Basic actions have individuals as both domain and codomain.
● (P3) (Atomicity) For any cell α∶ U → U ′ , if ∣U ′ ∣ = 0 then also ∣U ∣ = 0. Up to a special
isomorphism in DH , all traces of length n > 0 admit decompositions into n actions. For
any U ∶ X
Y of length 0, there is an isomorphism id ●X → U in DH as in
X
X
X
αU
Ū
U
Y.
● (P4) Restrictions of actions (resp. full actions) to individuals either are actions (resp. full
actions), or have length 0.
Proof. (P1) and (P2) are direct by Yoneda. (P4) is also easy in view of Lemma 4.22 and
its proof. For (P3), any vertical X → U ← Y of length 0, being a trace, is isomorphic to an
≅
≅
identity cospan, say Z → Z ← Z. To construct αU , just take the composite id ●X Ð
→ id ●Z Ð
→ U.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
39
Let us now treat the axiom for views, which is really easy. It actually becomes stronger
because of Remark 2.28, though this does not affect the rest of the construction:
Definition 5.4. Let B0 be the full subcategory of DH having as objects basic actions and
vertical identities between individuals.
Lemma 5.5. (P5) For any action M ∶ Y
unique cell
X and y∶ d → Y in Dh with d ∈ I, there exists a
y
d
v y,M
αy,M
y,M
d
yM
Y
M
X,
with v y,M ∈ B0 .
Proof. The result holds for seeds, by case analysis. E.g., v ls,πn = πnl , v rs,πn = πnr , and so on.
Now, any action M comes with a cell from its generating seed, say β∶ M0 → M . If y is in the
image of M0 , then the required cell αy,M is β ○ αy,M0 . Otherwise, v y,M = id ●d admits a cell
to M , which suits our needs. Uniqueness follows by (P1).
Remark 5.6. The important result that Axiom (P5) entails [39, Proposition 4.24] says that
when we replace M with any trace u, we get a double cell αy,u which is only unique up to
isomorphism. Below, we still define views up to isomorphism, so our modified Axiom (P5)
does not make this any stronger.
We conclude this section with the (straightforward) verification of (P8) and (P9).
Lemma 5.7. (P8) For any X, I/X is finite.
Proof. All positions are finitely presentable presheaves.
Lemma 5.8. (P9) For all d ∈ I and actions M ∶ X
d, M ′ ∶ X ′
d, and b∶ d′
′
′
and M full and b basic, if there exist cells M ← b → M then M ≅ M ′ .
d with M
Proof. By case analysis. E.g., if b = πnl , then M ≅ πn ≅ M ′ .
Remark 5.9. While the verification of (P9) is straightforward in our case, this axiom does
impose strong constraints on playgrounds. Morally, it demands that the basic subactions of
a given full action should be disjoint from those of a different full action. To see why this is
restrictive, let, for all j ∈ n, ιn,i,j denote the quotient of ιn,i by the equation s ○ sn+1 = s ○ sj .
The equation says that the received channel was already known as channel number j. Further
let [n]/{i = j} denote [n] quotiented by si = sj . We could be tempted to decree that the
cospan
t
s
[n + 1]/{n + 1 = j} Ð
→ ιn,i,j ←
Ð [n]
is an action. An example consequence would be that, e.g., the synchronisation on [n] i,l ∣ j,k [m]
where [m] sends k on j, when restricted to the receiver, would give ιn,i,l instead of ιn,i .
But then ιn,i and ιn,i,l would be two non-isomorphic full actions sharing a common basic
subaction, ιn,i .
We now have proved all playground axioms for D, except right and left decomposition.
These require the development of more machinery, which we undertake in the next section.
40
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
5.2. Correctness criterion. In order to prove the remaining playground axioms for D,
we set up a combinatorial characterisation of traces among cospans. Before delving into
technicalities, let us briefly map out our correctness criterion. Given a trace Y ↪ U ↩ X,
we start by forgetting the cospan structure and exploring the properties of U alone.
The main idea is to construct a binary relation over the elements of U , modeling causality.
So, e.g., if an agent x ∈ U [n] forks into x1 and x2 , then we will have causal relations
x1
x2
r
x,
where r denotes the corresponding element in U (πn ). In order for U to admit a sequential
decomposition into actions, the main criterion is that the causality relation should be acyclic.
In addition to this, a few sanity checks are necessary. First of all, because actions
are merely seeds pushed along 1-injective maps from their interfaces, the neighbourhood
of each action x ∈ U (µ) should not be too degenerate. For instance, the corresponding
map ⌜x⌝∶ µ → U should be 1-injective. Moreover, for inputs and channel creations, the new
channel should really be new. This property, which is a bit tedious to define properly, is
called local 1-injectivity.
Furthermore, when we add a new action to some trace, it is played by an agent in
the final position. This entails that no two actions in U may be performed by the same
agent. We call this target-linearity below. Symmetrically, no two actions may share their
‘created’ agents, which we call source-linearity. Linearity is then the conjunction of sourceand target-linearity.
̃ f has an acyclic causal
These conditions are sufficient, in the sense that if any U ∈ C
relation, and is furthermore locally 1-injective and linear, then it is the middle object of a
trace. But in fact, it is then easy to determine the corresponding initial and final positions.
s
t
We design notions of initial and final morphisms, so that Y Ð
→U ←
Ð X is a trace iff U
satisfies the above conditions, t is initial, and s is final.
Let us first define the causal relation. A first step is to restrict attention to the cores of
U , in the following sense, which are intuitively the main elements. E.g., for a forking action
x ∈ U (πn ), keeping track of x is enough, and handling x ⋅ l and x ⋅ r tends to get in the way.
Technically, an input or output is a core iff it is not part of a synchronisation; and a left or
right fork action is a core iff it is not part of a full fork action. Here is a concise definition:
̂ is an element of dimension > 1 which is not
Definition 5.10. A core of a presheaf U ∈ C
the image of any element of higher dimension.
Our definition of the causal relation will rely on the preliminary notions of sources and
targets of a core, and that of channels created by a core. These notions will fix the direction
of the causal relation.
Definition 5.11. For any U and core µ ∈ U (c),
● the sources of µ are the agents x such that x = µ ⋅ f ⋅ s for some f ;
● the targets of µ are the agents y such that y = µ ⋅ f ⋅ t for some f ;
● a channel a ∈ U (⋆) is created by µ iff µ has the shape νn or ιn,i , and a = µ ⋅ s ⋅ sn+1 .
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
41
Example 5.12. In the representable πn , there is one target, l ○ t (or equivalently r ○ t), and
two sources, s1 = l ○ s and s2 = r ○ s. Another example is τn,i,m,j,k , which has two targets,
○ t and ρ ○ t, and two sources. However, ρ ○ s ○ sn+1 is not created by the input element ρ,
because it is not a core.
Definition 5.13. For any U , let its causal graph GU have:
● as vertices, all channels, agents, and cores in U ,
● for all x ∈ U [n] and i ∈ n, an edge x → x ⋅ si ,
● and, for each core µ, an edge from each of its sources and created channels, and one into
each of its targets, as in
source1
source2
created
core
target1
target2 .
Please note that edges a → µ from a channel to an input action exist only if the involved
action is not part of a synchronisation; for otherwise the synchronisation is a core, not the
input.
The obtained graph is actually a binary relation, since there is at most one edge between
any two vertices. It is also a colored graph, in the sense that it comes equipped with a
morphism to the graph L:
∞
1
0,
mapping cores to ∞, agents to 1, and channels to 0. (In particular, there are no edges from
channels to agents or from cores to channels.) For any graph G, equipped with a morphism
l∶ G → L, we call vertices of G channels, agents, or cores, according to their label.
As expected, we have:
Proposition 5.14. For any trace U , GU is acyclic (in the directed sense).
Proof. By induction on any decomposition of U .
Let us now consider local 1-injectivity, linearity, initiality and finality. First, let us
emphasise that for all seeds Y ↪ M ↩ X, M is a representable presheaf, so, e.g., it makes
sense to consider U (M ).
Definition 5.15. A presheaf U is locally 1-injective iff for any seed Y ↪ M ↩ X with
interface I and core µ ∈ U (M ), if two distinct elements of M are identified by the Yoneda
morphism µ∶ M → U , then they are in (the image of) I(⋆).
This is equivalent to requiring that all morphisms yc → U , for all c ∈ C, are 1-injective,
and that for all core inputs and channel creations x of arity n in U , for all i ∈ n, we have
x ⋅ s ⋅ sn+1 ≠ x ⋅ s ⋅ si .
Proposition 5.16. Any trace U is locally 1-injective.
42
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Proof. Choose a decomposition of U into actions; µ corresponds to precisely one such action,
say M ′ , obtained, by definition, from some seed M as a pushout (3.1). By construction
of pushouts in presheaf categories, M ′ is obtained from M by identifying some channels
according to I → Z.
Let us observe that, because local 1-injectivity is only about cores, an input which is
part of a synchronisation may receive an already known channel, even if its n + 1th channel
is not part of its interface — because it is not a core.
After local 1-injectivity, let us consider linearity.
Definition 5.17. Any G ∈ Gph/L is source-linear iff for any cores µ, µ′ , and other vertex
(necessarily an agent or a channel) x, µ ← x → µ′ in G, then µ = µ′ ; G is target-linear iff
for any cores µ, µ′ and agent x, if µ → x ← µ′ in G, then µ = µ′ ; G is linear iff it is both
source-linear and target-linear.
s
t
Proposition 5.18. For any trace Y Ð
→U ←
Ð X, GU is linear.
Proof. Straightforward, by induction on any decomposition of U into actions, observing that
we glue along agents and channels which are initial on one side and final on the other.
The last of our necessary sanity checks is about initiality and finality. The idea here is
that one may read in U alone what both legs of the cospan Y ↪ U ↩ X should be.
Definition 5.19. An agent is initial in U when it is not the source of any action, i.e., for
no action µ ∈ U , x = µ ⋅ s. A channel is initial when it is not created by any core.
An agent x in U is final iff it is not the target of any action, i.e., for no action µ ∈ U ,
x = µ ⋅ t. All channels are final.
Lemma 5.20. An agent is initial in U iff it has no edge to any core in GU .
Lemma 5.21. An agent is final in U iff it has no edge from any core in GU .
Now, here is the expected characterisation:
Theorem 5.22. A monic cospan Y s U t X of finite presheaves is a trace iff
(C1) U is locally 1-injective,
(C2) X contains exactly the initial agents and channels in U ,
(C3) Y contains exactly the final agents and channels in U ,
(C4) and GU is linear and acyclic.
Of course, we have almost proved the ‘only if’ direction, and the rest is easy, so only
the ‘if’ direction remains to prove. The rest of this section is devoted to this. So given a
cospan satisfying the above conditions, we intend to sequentialise it, i.e., decompose it into
actions. We will proceed by induction on the number of cores in U , by picking a core µ which
is maximal according to GU , removing it from U and applying the induction hypothesis
to the rest. However, it may not be obvious how we should remove µ from U . E.g., the
topos-theoretic difference U ∖ µ does not yield the expected result, as it removes all sources
of µ. Instead, we consider the following operation: for any morphism of presheaves f ∶ U → V
and set W , let U − W = ∑c∈C Im(U (c)) ∖ W ⊆ ∑c∈C V (c). This is a slight abuse of notation,
as f is implicit, but it should be easily inferred from context.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
43
Remark 5.23. We observe that U − W is generally just a set, not a presheaf; i.e., its
elements are not necessarily stable under the action of morphisms in C. Consider for
example U = [1] ∣ [1] and let W consist of the first agent and the unique channel. Then
U − W does not contain the unique channel of U , so the action of s1 on the second agent
steps outside U − W .
But there is one useful case where U − W is indeed a subpresheaf of U , as we show below
in Lemma 5.25.
Definition 5.24. For any seed Y ↪ M ↩ X, let the past past(M ) = M − Y of M be the
set of its elements not in the image of Y . For any such M , presheaf U , and core µ ∈ U (M ),
let past(µ) = Im(past(M )) consist of all images of past(M ).
To explain the statement a bit more, by Yoneda, we see µ as a map M → U , so we have
a set-function
past(M ) ↪ ob(el(M )) → ob(el(U ))
(recalling that el denotes the category of elements). We observe that past(µ) is always a set
of agents and actions only, since channels present in X always are in Y too.
Given a core µ ∈ U , the relevant way of removing µ from U will be:
U ) µ = ⋃{V ↪ U ∣ ob(el(V )) ∩ past(µ) = ∅}.
U ) µ is thus the largest subpresheaf of U not containing any element of the past of µ. The
good property of this operation is:
Lemma 5.25. If µ is a maximal core in GU (i.e., there is no path to any further core) and
GU is target-linear, then (U ) µ)(c) = U (c) ∖ past(µ) for all c.
Proof. The direction (U ) µ)(c) ⊆ U (c) ∖ past(µ) is by definition of ). Conversely, it is
enough to show that c ↦ U (c) ∖ past(µ) forms a subpresheaf of U , i.e., that for any f ∶ c → c′
in C, and x ∈ U (c′ )∖past(µ), x⋅f ∉ past(µ). Assume on the contrary that x′ = x⋅f ∈ past(µ).
Then, of course f cannot be the identity, and w.l.o.g. we may assume that x′ is an agent
and x is a core. But then, because x′ ∈ past(µ), there is an edge µ → x′ in GU , and because
x′ = x ⋅ f , there is an edge x → x′ or x′ → x in GU . The former case is impossible by
target-linearity, and the latter case would imply the existence of a path µ → x in GU , which
contradicts the maximality of µ. So x′ ∈ past(µ) is impossible altogether.
Proof of Theorem 5.22. We proceed by induction on the number of actions in U . If it is
zero, then U is a position; by (C2), t is an iso, and by (C3) so is s, hence the cospan is a
trace. For the induction step, we first decompose U into
s2
t2
s1
t1
Y Ð→ U ′ ←Ð Z Ð→ M ′ ←Ð X,
and then show that M ′ is an action and U ′ satisfies the conditions of the theorem.
First, by acyclicity, pick a maximal core µ in GU , i.e., one with no path to any other
core. Let
I0
Y0
M0
X0
44
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
be the seed with interface corresponding to µ, so we have the Yoneda morphism µ∶ M0 → U .
Let U ′ = (U ) µ), and X1 = X − Agents(X0 ). X1 is a subpresheaf of X, since it contains
all channels. The square
I0
X1
X0
X
is a pushout, since it just adds the missing agents to X1 . Define now Z, M ′ , s1 , and t1 by
the pushouts
Y0
Z
s1
M′
M0
I0
X1
X0
U
t1
X
and the induced arrows. We further obtain arrows to U by universal property of pushout.
Let us show that the arrow f ∶ M ′ → U is mono. First, it is obviously mono in dimensions
> 1. It is also mono in dimension 1, because M ′ [n] = X[n] + Y0 [n] for all n and X → U is
mono with image consisting only of initial agents, which are thus dijoint from the image of
Y0 . Finally, for dimension 0, i.e., at ⋆, the pushout defining M ′ is isomorphic to
I0 (⋆) = X0 (⋆)
X1 (⋆) = X(⋆)
injl
injl
M0 (⋆) = X0 (⋆) + I
M ′ (⋆) = X(⋆) + I
where I = M0 (⋆) ∖ X0 (⋆) is the set of channels created by the action. Consider any
a, b ∈ M ′ (⋆) such that a ≠ b. Because X → U is mono, if a, b ∈ X(⋆) then f (a) ≠ f (b). By
local 1-injectivity of U , if a, b ∈ I then f (a) ≠ f (b). Finally, if a ∈ X(⋆) and b ∈ I, then we
have an edge f (b) → µ in GU , whereas f (a) is initial by (C2), so f (a) ≠ f (b). This shows
that M ′ → U is mono, which also entails that Z → U is a mono, because s1 is a pushout of
the mono Y0 → M0 .
By (C1) and Lemma 5.25, U = M ′ ∪ U ′ , i.e., the square
Z
U′
M′
U
is a pushout, so U is indeed a composite as claimed, with Z ↪ M ′ ↩ X an action by
construction. So, it remains to prove that Y ↪ U ′ ↩ Z satisfies the conditions. First, as
a subpresheaf of U (whose inclusion preserves cores), U ′ is locally 1-injective and has a
linear and acyclic causal graph, so satisfies (C1) and (C4). U ′ furthermore satisfies (C2) by
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
45
(EXPANDED VERSION)
construction of Z and source-linearity of GU , and (C3) because removing past(µ) cannot
make any non-final agent final.
Let us conclude this section with a helpful lemma, whose proof relies on Theorem 5.22:
Lemma 5.26. There is at most one cell filling any diagram
k
Y′
Y
u
u′
X
′
h
X
in D.
In order to prove this, let us introduce:
Definition 5.27. For any action x ∈ U , let core(x), the core associated to x, be the unique
core µ ∈ U for which there exists f in C such that µ ⋅ f = x. If x is an agent or a channel,
then by definition core(x) = x.
s′
t′
s
t
Proof of Lemma 5.26. By definition, we have cospans Y ′ Ð
→ u′ ←
Ð X ′ and Y Ð
→u←
Ð X.
′ ′
′
Suppose we are given l, l ∶ u → u making (k, l, h) and (k, l , h) into cells. By naturality, l
and l′ are determined by their images on channels, agents, and cores. We show by induction
on the ordering induced by Gu′ that they have to agree on these. For the base case: they
have to agree with h on initial agents and channels by definition of cells. For the induction
step, we proceed by case analysis on the kind of element to consider. The image of any
source of or channel created by a core µ is uniquely determined by naturality, which leaves
the case of a core µ, of which we assume that there is an agent x such that µ → x in Gu′
and l(x) = l′ (x). The edge µ → x yields a morphism, say t, in C such that µ ⋅ t = x. But
then by naturality we have l(µ) ⋅ t = l(x) = l′ (x) = l′ (µ) ⋅ t. By linearity of Gu we have
core(l(µ)) = core(l′ (µ)). Now let cµ denote the object of C over which µ lies, and let c′
be the one over which core(l(µ)) lies. By inspection of C, there is exactly one morphism
f ∶ cµ → c′ , and so we have l(µ) = core(l(µ)) ⋅ f = core(l′ (µ)) ⋅ f = l(µ), as desired.
5.3. A playground. In this section, we finally prove:
Theorem 5.28. D forms a playground.
Most axioms have been proved in previous sections, and we are left with both decomposition axioms, which are proved in Lemmas 5.29 and 5.30 below, relying on the correctness
criterion of the previous section.
Lemma 5.29. (P7) Any double cell as in the center below, where B is a basic action and
M is an action, decomposes in exactly one of the forms on the left and right:
A
X
α1
C
U
Y
α2
D
h
A
C
D
A
Y
C
M
k
X
α1
V
B
Z
X
α
Z
Y
α2
D
Z.
46
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
A
h
X
U1
f1
W1
C
fm
Y
f
U
U2
f2
B
k
W
M
Z,
Figure 6: Proof of Lemma 5.30
̂ let its neighbourhood consist of
Proof. For any element a over c ∈ C of any presheaf F ∈ C,
all elements in the image of a∶ c → F .
Let b ∈ B and m ∈ M be the unique cores of B and M , respectively. Let Vm be the
neighbourhood of m in M ● V .
If α(b) ∈ Vm , let us show that the whole of U is mapped to V , and we are in the left-hand
case. It is clear for channels. If there exists an element x of U of dimension ≥ 1 mapped to y
in M − V , i.e., M − Y , then we obtain a path x → x′ to an agent x′ of C, in GB●U . Via α,
this yields a path M − Y → Y in GM ●V between elements of dimension ≥ 1, a contradiction.
If now α(b) ∉ Vm , we show similarly that the whole of B ● U is mapped to V , because
the contrary would imply the existence of a path M − Y → V − Y in GM ●V , which also is a
contradiction. Hence, we are in the right-hand case.
Lemma 5.30. (P6) Any double cell
h
A
X
α
U
f
C
k
Y
α2
W2
B
W1
U1
decomposes as
Y
X
α1
α3
W1
U
h
A
W2
U2
Z
B
k
Z
with α3 an isomorphism, in an essentially unique way.
Proof. Let α = (h, f, k). We first treat the case where W2 is an action M , recalling Definition 5.27. We construct U1 and U2 , as depicted in Figure 6. First, let U1 = U ×W W1 , where
W = W2 ● W1 , and let A → U1 denote the induced arrow. By construction, all of A → U1 → U
are monos and, by Lemma 3.20 and the pullback lemma, A = U1 ×W1 X.
Let us show that the projection f1 ∶ U1 → W1 preserves initiality of channels and agents.
We proceed by contrapositive: consider any channel or agent x ∈ U1 . If x′ = f1 (x) is not
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
47
initial in W1 , then we have an edge x′ → µ′ for some core µ′ of W1 . But, since f is a
morphism between traces, it preserves initiality, so x cannot be initial in U , hence we find
x → µ in GU . By source linearity of GW , core(f (µ)) = µ′ , so the action f (µ) ∈ W has
antecedents both in U and W1 . By universal property of pullback, there exists an action
y ∈ U1 , respectively mapped to µ and f (µ), which by naturality and injectivity of U1 → U
entails that x → y in GU1 . Therefore, x is not initial in GU1 , as required.
Let now C ↪ U1 denote the subpresheaf of U1 consisting of initial channels and agents
(a subpresheaf because if x is an initial, n-ary agent, then x ⋅ si is an initial channel for
any i ∈ n, by Theorem 5.22 and Lemma 5.21). Since f1 preserves initiality, C → U1 → W1
factors through Y → W1 , uniquely since the latter is mono, say as fm (see Figure 6). By
Theorem 5.22, A → U1 ← C is a trace and (h, f1 , fm ) defines a morphism to X → W1 ← Y .
Let then U2 ↪ U denote the subpresheaf of U consisting of elements below C in GU , i.e.,
x ∈ U2 ⇔ ∃c ∈ C.c →⋆ GU core(x).
A first observation is that all initial channels and agents of U are in U2 , so that B → U
factors through U2 . Indeed, consider any such initial x. By acyclicity of GU , each initial
element is reachable from some final element, so x is reachable from some final y. But by
source-linearity the corresponding path y →⋆ x goes through C, so we find a path c →⋆ x for
some c ∈ C, as desired.
Now, because U2 → U and M → W are monos, showing that f maps all elements x of
U2 to M will imply that U2 → U → W uniquely factors through M → W . Let us do this by
case analysis:
● If x is not a channel, then f preserves paths from agents to x, so we find some path
f (c) →⋆ GW core(f (x)) with c ∈ C hence f (c) ∈ Y , which implies that f (x) ∈ M (f (x) ∈
W1 − M would contradict initiality of Y in W1 ).
● If x is some channel initial in U , then since f preserves initiality x is mapped to Z hence
to M .
● If finally x is some non-initial channel, then x → µ for some core µ ∈ U . Now µ ∈ U2 , as
witnessed by the path c →⋆ x → µ. But then x = µ ⋅ u for some morphism u of C, so since
by the above f (µ) ∈ M , we have that f (x) = f (µ) ⋅ u is in M too, as desired.
We thus get a diagram as in Figure 6, which commutes because M → W is mono.
By Theorem 5.22, C → U2 ← B is a trace, and U = U2 ● U1 , which shows existence of the
desired decomposition.
For any decomposition as in Figure 6, we have C = U2 ×M Y by Lemma 3.20, so by
Corollary 4.20, we also have U1 = U ×W W1 . Thus, U1 is uniquely determined up to canonical
isomorphism. But by Theorem 5.22, C → U1 is so too, as the subobject of initial agents and
channels. But then U2 precisely consists of elements below C. Indeed, by finiteness of GU2
and (C3) in Theorem 5.22, all of U2 clearly lies below C. Conversely, for any x ∈ U1 − U2 , by
finiteness of GU1 and (C3) in Theorem 5.22, we have a path x →+ c to some c ∈ C, so x cannot
lie below C by (C4). Our decomposition is thus unique up to canonical isomorphism.
6. A sheaf model
In the previous sections, we have constructed a double category D and equipped it with
playground structure. We now instantiate constructions from [39] on D, which lead to the
definition of our sheaf model for π. We first recall various notions of strategy in Section 6.1:
48
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
naive strategies, innocent strategies, and behaviours. Behaviours are further studied in
Section 6.2, where we introduce a kind of calculus for them. Using this calculus, we then
define our intepretation of π in Section 6.3. Finally, in Section 6.4, we state our semantic
definition of fair testing equivalence and our main result.
6.1. Strategies and behaviours. We first recall notions of strategies. As announced in
the introduction, we define a category T(X) combining prefix ordering and isomorphism
of traces: T(X) has traces u∶ Y
X as objects, and as morphisms u → u′ all pairs (w, α)
′
with w∶ Y
Y and α an isomorphism u ● w → u′ in the hom-category Dv (Y ′ , X), as in
Y′
w
≅
α
Y
u
u′
X,
considered up to the smallest equivalence relation identifying (w, α) and (w′ , α ○ (u ● γ)),
for any w′ ∶ Y ′
Y and special γ∶ w′ → w). Thus, u′ is an extension of u by w.
̃
Definition 6.1. Let the category of (naive) strategies on X be T(X).
Strategies do not yield a satisfactory model for π:
Example 6.2. Consider the position X with three agents x, y, z sharing a channel a, and
the following traces on it: in ux,y , x sends a on a, and y receives it; in ux,z , x sends a on a,
and z receives it; in iz , z inputs on a. One may define a strategy S mapping ux,y and iz to a
singleton, and ux,z to ∅. Because ux,y is accepted, x accepts to send a on a; and because iz
is accepted, z accepts to input on a. The problem is that S rejecting ux,z roughly amounts
to x refusing to synchronise with z, or conversely.
We want to rule out this kind of strategy from our model, by adapting the idea of
innocence. We start by extending T(X) with objects representing traces on subpositions of
X. For this, we consider the following category TX . It has as objects pairs (u, h) of a trace
u∶ Z
Y and a morphism h∶ Y → X in Dh . A morphism (u, h) → (u′ , h′ ) consists of a trace
w∶ T
Z and a cell as below left with h′ r = h. Morphisms are considered up to the smallest
equivalence relation identifying (w, α) with (w′ , α ○ (u ● γ)), for any w′ and γ as below right.
s
T
Z′
T′
w
α
Z
u′
u
w′
γ
s
T
Z′
w
α
Z
u′
u
r
Y
h
X
Y′
h′
r
Y
h
X
Y′
h′
Example 6.3. Recalling the right-hand trace of Figure 4 (page 22), say u∶ Y
X, y’s first
action is an input on its unique channel b. This yields a trace ι1,1 ∶ [2]
[1]. Here is an
examlpe morphism (ι1,1 , y) → (u, id X ) in TX :
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
y ′′
[3]
(EXPANDED VERSION)
49
Y
ι2,2
[2]
u
ι1,1
y
[1]
X.
y
X.
We think of it as an occurrence of the trace ι1,1 in u. Thus, morphisms in TX account both
for prefix inclusion and for ‘spatial’ inclusion, i.e., inclusion of a trace into some other trace
on a larger position.
We now define views within TX :
Definition 6.4. A view is a trace isomorphic to some (possibly empty) composite of basic
actions (Definition 5.1). Let VX denote the full subcategory of TX spanning pairs (u, h)
where u is a view.
Intuitively, basic actions follow exactly one agent through an action. An object of VX
consists of a view, say v∶ [n′ ]
[n], plus a morphism h∶ [n] → X in Dh , which by Yoneda is
just an agent of X. So an object of VX is just an agent of X and a view from it.
Definition 6.5. The inclusion jX ∶ VX ↪ TX induces a Grothendieck topology, for which
(wi ,αi )
a sieve ((ui , hi ) ÐÐÐÐ→ (u, h))i∈I of morphisms to some trace u is covering iff it contains
all morphisms from views into u. Let the category SX ↪ T̃
X of innocent strategies be the
category of sheaves of finite sets for this topology. Let the category BX of behaviours over
̃
X be V
X.
As announced in the introduction, letting ranjop denote right Kan extension [53] along
X
op
op
the inclusion jop
X ∶ VX ↪ TX , we have:
Proposition 6.6. The embedding ranjop induces an equivalence of categories BX ≃ SX .
X
Proof. See [39, Lemma 4.34].
We thus obtain the innocent strategy SB associated to a behaviour B ∈ BX by taking
its right Kan extension as in
jop
X
Vop
X
Top
X
kop
X
T(X)op
SB
B
set.
U(SB )
Explicitly, using standard results, we obtain the end
SB (u, h) = ∫
B(v, x)TX ((v,x),(u,h)) ,
(v,x)∈VX
which is a kind of generalised product. In the boolean setting (functors to 2), this end
reduces to the conjunction ⋀{(v,x)∈VX ∣∃α∶(v,x)→(u,h)} B(v, x), demanding precisely that all
views of u are accepted by B. In the general case, the intuition is that a way of accepting u
for SB is a compatible family of ways of accepting the views of u for B. The forgetful functor
U to naive strategies is then given by restricting along the opposite of kX ∶ T(X) ↪ TX as
above. Some local information may be forgotten by U, which is hence neither injective on
objects, nor full, nor faithful. E.g., if two behaviours differ on one agent, but are both empty
on the views of another, then both are mapped to the empty naive strategy.
50
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Example 6.7. Recalling X and S from Example 6.2, let us show that for any B ∈ BX , the
̃ cannot be S. Indeed, if U(SB ) was S, then because S
associated strategy U(SB ) ∈ T(X)
accepts ux,y and iz , B has to accept the following views: (1) iz , (2) ox , in which x sends
a on a (without any matching input), (3) iy , in which y inputs on a, and (4) all identity
views on x, y, and z. But then U(SB ) has to accept both ux,y and ux,z , because B accepts
all views mapping into them.
6.2. Decomposing behaviours. In this section, we study behaviours a bit more in depth,
which yields the calculus announced at the beginning of Section 6. The starting point is that
the assignment X ↦ BX may be equipped with useful structure, describing how a behaviour
B on some given position restricts to any subposition, and also what remains of it after
a given action has been played. Otherwise said, morphisms of D act contravariantly on
behaviours:
● horizontal morphisms h∶ X → X ′ induce functors Bh ∶ BX ′ → BX , and
● vertical morphisms u∶ Y
X induce functors Bu ∶ BX → BY .
Furthermore, any cell as below left induces a natural isomorphism as below right:
k
Y
u
X
α
Y′
Bu
u′
h
X
Bk
BY
′
BX
BY ′
Bu′
≅
Bh
BX ′ ,
which notably says that B ⋅ u′ ⋅ k ≅ B ⋅ h ⋅ u for any behaviour B ∈ BX ′ . This is worked out in
detail and in full generality in [39], and extended to a pseudo double functor Dop → QCat,
where QCat denotes Ehresmann’s double category of quintets over Cat. But let us explain
how both actions look like in the present, concrete case.
The first, horizontal action is really easy: any horizontal morphism k∶ X ′ → X acts on
a given behaviour B ∈ BX by returning the behaviour B ⋅ k such that for all (v, h) ∈ VX ′ ,
(B ⋅ k)(v, h) = B(v, k ○ h).
Proposition 6.8. The functor BX → ∏n,x∶[n]→X B[n] given at (n, x) by horizontal action
of x, i.e., B ↦ B ⋅ x, is an isomorphism.
Proof. We have VX ≅ ∑n,x∶[n]→X V[n] .
Notation 6.9. If (Bx )x∶[n]→X is a family of behaviours indexed by the agents of X, we
accordingly denote its unique antecedent by [Bx ]x∶[n]→X .
Vertical action is a bit harder. Let us start by recalling the following result from [39],
which generalises Lemma 5.5:
Lemma 6.10. For any trace P ∶ Y
essentially unique cell
X and y∶ d → Y in Dh with d ∈ I, there exists an
d
v y,P
y
αy,P
y,P
d
yP
Y
P
X,
with v y,P a view.
Proof. By induction on P using Lemma 5.5. See [39] for the precise meaning of essential
uniqueness in this case.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
51
This result allows us to describe what remains of a behaviour after a trace:
Definition 6.11. Given any behaviour B ∈ BX and trace P ∶ Y
B along P is the behaviour determined (up to isomorphism) by
X, the residual B ⋅ P of
(B ⋅ P ⋅ y)(w, id [ny ] ) = B(v y,P ● w, y P )
for all agents y∶ [ny ] → Y .
So we may consider residuals of behaviours along arbitrary traces. Conversely, any
behaviour is determined by its initial states and residuals. Let us first consider the following
special kind of behaviour:
Definition 6.12. A behaviour B on [n] is definite iff B(id ● , id [n] ) ≅ 1. A behaviour B on
an arbitrary position X is definite iff for all agents x of X, B ⋅ x is definite. Let DX denote
the full subcategory of BX spanning definite behaviours.
This means that B has exactly one initial state.
Definition 6.13. For any n ∈ N and family Bb ∈ B[nb ] indexed by all basic actions b∶ [nb ]
[n], let ⟨Bb ⟩b denote the definite behaviour B determined by B ⋅ b = Bb for all b.
By construction, we have:
Proposition 6.14. For any definite behaviour D ∈ B[n] , D ≅ ⟨D ⋅ b⟩b .
This extends to arbitrary behaviours on individuals using the fact that any behaviour
on an individual is a coproduct of definite behaviours.
Notation 6.15. For any behaviour B = ∑k∈γ Dk on any [n], let ∣B∣ = γ and, for any k ∈ γ,
let the restriction B∣k of B to k be Dk .
Remark 6.16. In other words, k ∈ γ is just an element of B(id ●[n] , id [n] ) and B∣σ is
determined by
B∣σ (v, id [n] ) = {σ ′ ∈ B(v, id [n] ) ∣ σ ′ ⋅ !v = σ},
where !v denotes the unique morphism (id ●[n] , id [n] ) → (v, id [n] ) in V[n] .
We obtain:
Proposition 6.17. For any behaviour B ∈ B[n] ,
B ≅ ∑ ⟨(B∣k ) ⋅ b⟩b .
k∈∣B∣
Putting this together with spatial decomposition, we obtain:
Corollary 6.18. For any behaviour B ∈ BX ,
B ≅ [ ∑ ⟨((B ⋅ x)∣k ) ⋅ b⟩b∶[nb ]
[nx ] ]x∶[nx ]→X .
k∈∣B⋅x∣
The fact that both actions of D yield a pseudo double functor Dop → QCat essentially
boils down to:
Lemma 6.19. For any B ∈ BX and cell
Y′
α
u′
X
k
′
h
Y
u
X,
52
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
we have D ⋅ h ⋅ u′ ≅ D ⋅ u ⋅ k.
The result is stated (and proved) in full generality as [39, Proposition 4.31].
6.3. Interpretation of π. We now define our interpretation of π-calculus configurations.
We start with processes and then cover configurations. Because the notation for behaviours
introduced in the previous section only covers behaviours on representable positions, while
π-calculus syntax is name-based, we bridge the gap by keeping track, along the recursive
definition, of a bijection between the set γ of free channels of the considered process and its
cardinal ∣γ∣. So we define a family of maps in
∏
Pi ×Bij(γ,∣γ∣)
γ
D[∣γ∣]
,
γ∈Pf (N)
where Bij(A, B) denotes the set of all bijections A −∼ B. For any such γ, P , and h in
the domain, we denote the result by ⟦P ⟧h , or ⟦P ⟧γ,h when needed. Letting n = ∣γ∣, it is
coinductively defined by
⟦∑i Pi ⟧h = ⊞i ⟦Pi ⟧h
π l ↦ ⟦P ⟧h
⟦P ∣ Q⟧h = ⟨ nr
⟩
πn ↦ ⟦Q⟧h
⟦νb.P ⟧h = ⟨ νn ↦ ⟦P ⟧h′ ⟩
⟦♡.P ⟧h = ⟨ ♡n ↦ ⟦P ⟧h ⟩
⟦τ.P ⟧h = ⟨ τn ↦ ⟦P ⟧h ⟩
⟦a(b).P ⟧h = ⟨ ιn,h(a) ↦ ⟦P ⟧h′ ⟩
⟦ā⟨b⟩.P ⟧h = ⟨ on,h(a),h(b) ↦ ⟦P ⟧h ⟩
where
● in any list ⟨b1 ↦ B1 , . . . , bm ↦ Bm ⟩, all unmentioned basic actions are meant to be mapped
to the empty behaviour;
● the definite sum ⊞i Di of definite behaviours Di is the definite behaviour determined by
(⊞i Di ) ⋅ b = ∑i (Di ⋅ b), for all basic actions b∶ [n′ ]
[n];
● and h′ ∶ γ, b −∼ n + 1 maps any a ∈ γ to h(a), and b to n + 1.
Example 6.20. Let us briefly illustrate the translation. Consider any h∶ γ → ∣γ∣ and
processes γ, b ⊢ P , γ, b ⊢ Q, and γ ⊢ R, with a ∈ γ. We can form a(b).P + a(b).Q + ā⟨a⟩.R,
which is mapped to
⟨ι∣γ∣,h(a) ↦ (⟦P ⟧h′ + ⟦Q⟧h′ ), oγ,h(a),h(a) ↦ ⟦R⟧h ⟩.
To emphasise the difference, using coproduct of behaviours instead of definite sum in the
translation would yield a behaviour with three distinct initial states, closer to the internal
choice a(b).P ⊕ a(b).Q ⊕ ā⟨a⟩.R than to the original process.
Generalising this to configurations should really be intuitive: we map any ⟨γ∥P1 , . . . , Pn ⟩
to some behaviour on the position X with X(⋆) = γ and n agents of arity ∣γ∣, given for
each agent i ∈ n by ⟦Pi ⟧. In order to fully define such a position, we need to specify maps
fi ∶ ∣γ∣ → γ. We use for all of them the inverse of the canonical bijection hγ defined by:
Definition 6.21. Let hγ ∶ γ −∼ ∣γ∣ map each a ∈ γ to its position in the ordering induced by
the one on natural numbers.
We call the obtained position X(γ, n).
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
53
Definition 6.22. Let ⟦−⟧∶ ob(Conf) → ∑X BX map any configuration C = ⟨γ∥P1 , . . . , Pn ⟩
to the pair (X(γ, n), ⟦C⟧), where ⟦C⟧ is defined through Proposition 6.8 by
⟦C⟧(∣γ∣)(i) = ⟦Pi ⟧hγ ,
for all i ∈ n. We implicitly consider processes P over γ as configurations ⟨γ∥P ⟩, and hence
allow ourselves to write ⟦P ⟧ for ⟦⟨γ∥P ⟩⟧.
6.4. Semantic fair testing. In order to state our main result, it remains to define our
semantic analogue of fair testing equivalence. It rests on two main ingredients: a notion of
closed-world trace, and an analogue of parallel composition in game semantics.
The intuitive purpose of parallel composition is to let behaviours interact. If we partition
the agents of a position X into two teams, we obtain two subpositions X1 ↪ X ↩ X2 , each
agent of X belonging to X1 or X2 according to its team. The crucial fact is that the
category VX of views on X is isomorphic to the coproduct category VX1 + VX2 . Parallel
̃
̃
composition of any B1 ∈ V
X1 and B2 ∈ VX2 is then simply given by copairing [B1 , B2 ]
(following Notation 6.9), as in
Vop
X1
Vop
X
B1
[B1 ,B2 ]
Vop
X2
B2
set.
We now describe closed-world traces, which are then used as a criterion for success
of tests. Closed-world actions were defined (Definition 5.1) as those not involving any
interaction with the environment, i.e., formally, pushouts of a seed of any shape among
νn ,τn ,♡n ,πn , and τn,a,m,c,d . A trace is closed-world when it is a composite of closed-world
i
actions. Let W(X) X T(X) denote the full subcategory of T(X) consisting of closed-world
̃
traces, and let the category of closed-world strategies be W(X).
Notation 6.23. Recalling the discussion below Proposition 6.6, we denote by B ↦ B the
composite functor
op
∆kop
∆iop
ran jX
X
X
̃ ÐÐ→
̃
̃
̃ ÐÐ
V
→ T(X)
W(X),
X ÐÐÐ→ TX Ð
where ∆f denotes restriction along f .
A closed-world trace is successful when it contains a ♡ action, and unsuccessful otherwise.
̃ over a closed-world trace u∶ Z ′
A state σ ∈ S(u) of any S ∈ W(Z)
Z is successful iff u
̃ such that any unsuccessful
is. Define ⊥ Z as the set of closed-world strategies S ∈ W(Z)
closed-world state admits a successful extension, i.e. S ∈ ⊥ Z iff for all unsuccessful u ∈ W(Z)
and σ ∈ S(u), there exists a successful u′ ∈ W(Z), a morphism f ∶ u → u′ , and a state
σ ′ ∈ S(u′ ) such that σ ′ ⋅ f = σ. Finally, in order to compare behaviours for semantic fair
testing equivalence, we specify what a test is for a given behaviour B ∈ BX . A test consists
of a position Y and a behaviour T ∈ BY . Recalling Definition 3.8, we say that the pair
(X, B), with B ∈ BX , should pass the test (Y, T ) iff IX = IY and [B, T ] ∈ ⊥ Z , where Z is
the pushout X +IX Y (X and Y thus form two teams on Z). At last, we define semantic
fair testing equivalence, for any B ∈ BX and B ′ ∈ BX ′ :
Definition 6.24. Let (X, B) ∼f (X ′ , B ′ ) iff they should pass the same tests.
54
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
We may at last state:
Theorem 6.25. The translation ⟦−⟧∶ ob(Conf) → ∑X BX is intensionally fully abstract for
∼f , i.e.,
● For all configurations C1 and C2 , C1 ∼Pi
f C2 iff ⟦C1 ⟧ ∼f ⟦C2 ⟧;
● Furthermore, for all positions X and behaviours B ∈ BX , there exists C ∈ Pi X(⋆) such that
⟦C⟧ ∼f B.
The proof is the subject of the next section.
7. Intensional full abstraction
In the previous section, exploiting the playground structure of D established in Sections 4
and 5, we have defined and studied the notion of behaviour, into which we have translated
π-calculus processes and configurations. We have then defined our semantic analogue of
fair testing equivalence and stated our main result. We now work towards proving it. In
Section 7.1, we define a graph with testing S whose vertices are pairs (X, B) with B a
definite behaviour on the position X (Definition 6.12), such that fair testing equivalence
in S coincides with fair testing equivalence in the model. In order to prove that this is the
case, we introduce an intermediate graph with testing C which is in fact quite intricate. We
are then in a position where our main result is reduced to intensional full abstraction of a
translation between two graphs with testing, Conf and S, for which we may hope to apply
the results of Section 2.3. In fact, in Section 7.2, we further reduce to a translation to a
more syntactic graph with testing, M. In Section 7.3, we prove intensional full abstraction
of the translation to M, from which we deduce Theorem 6.25. Finally, we generalise to a
large class of testing equivalences in Section 7.4.
7.1. A first graph with testing for behaviours.
Definition 7.1. Let S denote the graph with vertices in ∑X DX , where we recall (Definition 6.12) that DX denotes the category of definite behaviours on X, and with non-identity
edges (X, D) ← (Y, D′ ) all closed-world actions M ∶ Y
X such that for all agents y in Y ,
there exists σy ∈ ∣D ⋅ M ⋅ y∣ such that
D′ ⋅ y ≅ (D ⋅ M ⋅ y)∣σy .
M
Moreover, let pS ∶ S → Σ denote the map sending (X, D) ←Ð (X ′ , D′ ) to the ♡ edge in Σ
if M is a tick action and to τ otherwise.
Let us now equip S with testing structure.
Definition 7.2. We define the relation ∣S by (Z, D) ∈ ((X1 , D1 ) ∣S (X2 , D2 )) iff X1 (⋆) =
X2 (⋆) and there is a pushout square
IX1
X2
injr
X1
injl
Z
such that D1 ≅ D ⋅ injl and D2 ≅ D ⋅ injr , or otherwise said D ≅ [D1 , D2 ].
Lemma 2.18 entails:
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
55
Proposition 7.3. The morphism fc(pS )∶ fc(S) → fc(Σ), with ∣S as testing relation, forms a
free graph with testing.
Proof. We exhibit a weak bisimulation relating any two such pushouts (Z, D) and (Z ′ , D′ ).
The relation containing two such pairs as soon as there exists a horizontal isomorphism
h∶ Z → Z ′ such that D ≅ D′ ⋅ h does the job.
The crucial result for proving that fair testing equivalence in S coincides with semantic
fair testing equivalence is:
Lemma 7.4. For any definite behaviour D ∈ DX , we have D ∈ ⊥ X iff (X, D) ∈ S .
Our strategy for proving this is to introduce and study an intermediate graph with
testing C, which is closer to semantic fair testing in that its transitions may comprise several
actions. However, defining C just as S with arbitrary traces instead of actions would be
wrong:
Example 7.5. Consider the trace P = (τ0 ● π0 ) consisting of a nullary agent performing
a τ action and then forking. Consider now any definite behaviour D such that D(τ0 ) =
D(τ0 ● π0l ) = D(τ0 ● π0r ) = 2, which maps both inclusions τ0 ● π0l ↩ τ0 ↪ τ0 ● π0r to the identity.
Then D(P ) = 2: it consists of pairs (σ l , σ r ) in D(τ0 ● π0l ) × D(τ0 ● π0r ) whose restrictions to
D(τ0 ) coincide, which leaves just (1, 1) and (2, 2). Using the decompositions of Section 6.2,
another way to say this is that D = ⟨τ0 ↦ D1 + D2 ⟩, with Di = ⟨π0l ↦ Dil , π0r ↦ Dir ⟩, for
i = 1, 2. In order for C to correspond to the model, assuming D1l ≇ D2l and D1r ≇ D2r ,
there should be two transitions from ([0], D), one to ([0] ∣ [0], [D1l , D1r ]) and the other to
([0] ∣ [0], [D2l , D2r ]). But if we naively generalise Definition 7.1 to arbitrary traces, we obtain
additional, ‘incoherent’ transitions, to ([0] ∣ [0], [D1l , D2r ]) and ([0] ∣ [0], [D2l , D1r ]).
Instead of relying on ∏y∈Agents(Y ) ∣D ⋅ P ⋅ y∣ as in Definition 7.1, we would like to rely on
D(P ). This may be done by constructing a map
ψPD ∶ D(P ) →
∏
∣D ⋅ P ⋅ y∣,
(7.1)
y∶[ny ]→Y
whose image will consist precisely of all ‘coherent’ elements. (From now on, we omit the
superscript D when clear from context.) Intuitively, this map associates to each global state
the corresponding family of local states. Furthermore, ψP is always injective, but Example 7.5
shows that it is not surjective in general. In order to construct ψP , recalling Definition 6.11,
we have by definition: ∏y∶[ny ]→Y ∣D ⋅ P ⋅ y∣ = ∏y∶[ny ]→Y D(v y,P , y P ). Furthermore, we also
have
D(P ) ≅ ∫
D(v, x)TX ((v,x),(P,id X )) ,
(v,x)∈VX
TX ((v,x),(P,id X ))
which is a subset of ∏(v,x)∈VX D(v, x)
. We may thus define:
ψP (σ)(y) = σ(v y,P , y P )(id ●[ny ] , αy,P ).
y,P
P
Here, σ(v y,P , y P ) is in D(v y,P , y P )TX ((v ,y ),(P,id X )) . So by applying it to αy,P viewed as
a morphism in (v y,P , y P ) → (P, id X ) in TX , we obtain an element of D(v y,P , y P ) = ∣D ⋅ P ⋅ y∣,
as desired.
We may now define the intermediate graph with testing C. We first extend the notion
of restriction:
56
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Notation 7.6. We extend Notation 6.15: if B ∈ BX and σ ∈ ∏n,x∶[n]→X ∣B ⋅ x∣, let B∣σ be
defined up to isomorphism by
B∣σ ⋅ x = (B ⋅ x)∣σ(x) .
Definition 7.7. Let C denote the graph with ob(C) = ob(S), and where C((X ′ , D′ ), (X, D))
is the set of closed-world traces W ∶ X ′ X such that there exists a state σ ∈ D(W ) satisfying
(D ⋅ W )∣ψW (σ) ≅ D′ .
Thus, C is a generalisation of S from closed-world actions to closed-world traces. Let us
turn it into a graph with testing.
Definition 7.8. Let DW
v denote the smallest locally full subbicategory of Dv containing all
closed-world traces. The graph morphism W → Σ, where we recall that W denotes the graph
of closed-world actions (Definition 5.1), extends to a pseudo functor pW ∶ DW
v → fc(Σ), which
C
W
essentially counts the number of ticks. Let p ∶ C → Dv denote the obvious projection.
pC
pW
Proposition 7.9. The composite projection C Ð→ DW
v Ð→ fc(Σ), with ∣S as testing relation,
makes C into a graph with testing.
Proof. Just as Lemma 7.3.
As announced, C is an example of a non-free graph with testing. The rest of this
section is devoted proving Lemma 7.4, and reducing Theorem 6.25 to a statement about S.
Lemma 7.4 follows from the fact that both poles are equivalent to C , as we now set out to
prove. We start with a lemma saying that C has essentially the same transitions over two
(specially) isomorphic traces.
Notation 7.10. For any morphism p∶ G → H in Gph, we denote by pA,B ∶ G(A, B) →
H(p(A), p(B)) the component of p at A and B.
Lemma 7.11. For any (X, D), (X ′ , D′ ) ∈ C, if there exists any special isomorphism W1 ≅ W2
′
in DW
v (X , X), we have
C −1
(pC )−1
(X ′ ,D′ ),(X,D) (W1 ) ≅ (p )(X ′ ,D′ ),(X,D) (W2 ).
Proof. The given special isomorphism induces by pseudo double functoriality of B an
isomorphism ϕ∶ D ⋅ W1 −∼ D ⋅ W2 , hence an isomorphism
ϕid X ′ ∶ D(W1 )
∼
−
D(W2 ).
This isomorphism is such that for any σ,
(D ⋅ W1 )∣ψW1 (σ) ≅ (D ⋅ W2 )∣ψW2 (ϕid
Thus, (D ⋅ W1 )∣ψW1 (σ) ≅ D′ iff (D ⋅ W2 )∣ψW2 (ϕid
X′
(σ))
X′
(σ)) .
≅ D′ .
In order to relate C to S, let us now show that transitions in C behave well w.r.t.
composition of traces. First, transitions compose, and second, transitions over any composite
(closed-world) trace W1 ● W2 always decompose into a transition over W1 followed by one
over W2 .
W
W′
Lemma 7.12. For all edges (X, D) ←Ð (X ′ , D′ ) ←ÐÐ (X ′′ , D′′ ) in C, there is an edge
W ●W ′
(X, D) ←ÐÐÐ (X ′′ , D′′ ).
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
57
Proof. Consider σ ∈ D(W ) such that (D ⋅ W )∣ψW (σ) ≅ D′ and σ ′ ∈ D′ (W ′ ) such that
(D′ ⋅ W ′ )∣ψW ′ (σ′ ) ≅ D′′ . We want to construct an edge W ● W ′ ∶ (X ′′ , D′′ ) → (X, D), i.e.,
find σ ′′ ∈ D(W ● W ′ ) such that (D ⋅ (W ● W ′ ))∣ψW ●W ′ (σ′′ ) ≅ D′′ . Now, the isomorphism
′
′
ϕ∶ (D ⋅ W )∣ψW (σ) −∼ D′ yields a state σ1 = ϕ−1
W ′ (σ ) ∈ (D ⋅ W )∣ψW (σ) (W ) such that
((D ⋅ W )∣ψW (σ) ⋅ W ′ )∣ψW ′ (σ1 ) ≅ (D′ ⋅ W ′ )∣ψW ′ (σ′ ) .
(7.2)
Now we have
′′
(D ⋅ W )∣ψW (σ) (W ′ ) ≅ {σ ′′ ∈ D(W ● W ′ ) ∣ σ∣W
= σ},
′′
where σ∣W
denotes restriction of σ ′′ along the prefix inclusion W ↪ W ● W ′ . So the left-hand
side in (7.2) is just (D ⋅ (W ● W ′ ))∣ψW ●W ′ (σ1 ) , which yields the desired transition.
Lemma 7.13. The projection pC ∶ C → DW
v satisfies the following weak Conduché condition:
for all X ′′
W2
X′
W1
W1 ●W2
X, if there is an edge (X ′′ , D′′ ) ÐÐÐÐ→ (X, D) in C, then there
W2
W1
exists D′ ∈ DX ′ and edges (X ′′ , D′′ ) ÐÐ→ (X ′ , D′ ) ÐÐ→ (X, D).
Proof. Consider any (X, D) ∈ C and σ ∈ D(W1 ● W2 ) witnessing the given edge. Consider
also the morphism u∶ W1 → (W1 ● W2 ) given by (W2 , id ), and let σ1 = σ ⋅ u ∈ D(W1 ). Let
D1 = (D ⋅ W1 )∣ψW1 (σ1 ) . We have σ ∈ {σ ′ ∈ D(W1 ● W2 ) ∣ σ ′ ⋅ u = σ1 }, hence σ ∈ D1 (W2 ).
Furthermore,
(D1 ⋅ W2 )∣ψW2 (σ) ≅ (D ⋅ (W1 ● W2 ))∣ψW1 ●W2 (σ) ≅ D′′ ,
so we have two edges
(W1 ,σ1 )
(W2 ,σ)
(X, D) ←ÐÐÐÐ (X ′ , D1 ) ←ÐÐÐÐ (X ′′ , D′′ )
as desired.
The previous result generalises by induction to n-ary composites:
′
′′
Notation 7.14. By default, composition in DW
v associates to the right, i.e., W ● W ● W
denotes W ● (W ′ ● W ′′ ).
Corollary 7.15. For any path p, say
X = X0
M1
X1
M2
. . . Xn = X ′ ,
(7.3)
W
in W and edge (X ′ , D′ ) Ð→ (X, D) in C over its right-associated, n-ary composition W =
(M1 ● (. . . ● Mn )), there is a path e = (e1 , . . . , en ) in C⋆ ((X ′ , D′ ), (X, D)) such that
(pC )⋆(X ′ ,D′ ),(X,D) (e) = p.
Proof. By induction on n.
Our next goal is to relate transitions in C to sequences of transitions in S. First of all, C
and S coincide on actions:
Lemma 7.16. If W ∶ X ′
X is a closed-world action (i.e., has length 1), then for all D
′
′
and D both fibres of C((X , D′ ), (X, D)) and S((X ′ , D′ ), (X, D)) over W are equal.
Proof. By [39, Proposition 5.23].
58
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Now, let us show that for any sequence of closed-world actions, sequences of transitions
in S correspond to transitions over the composite in C.
Corollary 7.17. For all closed-world paths as in (7.3), and (X, D), (X ′ , D′ ) ∈ S, we have
((pS )⋆(X ′ ,D′ ),(X,D) )−1 (p) ≠ ∅
iff
(pC(X ′ ,D′ ),(X,D) )−1 (P ) ≠ ∅,
for any special isomorphism P ≅ (M1 ● (. . . ● Mn )).
Proof. Consider any special isomorphism α∶ P
((pS )⋆(X ′ ,D′ ),(X,D) )−1 (p) ≠ ∅
iff ((pC )⋆(X ′ ,D′ ),(X,D) )−1 (p) ≠ ∅
iff (pC(X ′ ,D′ ),(X,D) )−1 (M1 ● (. . . ● Mn ))
iff (pC(X ′ ,D′ ),(X,D) )−1 (P ) ≠ ∅
∼
−
(M1 ● (. . . ● Mn )). We have
(by Lemma 7.16)
≠ ∅ (by Lemma 7.12 and Corollary 7.15)
(by Lemma 7.11).
As a corollary, we get that the identity relation on objects is a strong bisimulation
between fc(S) and C:
w
Ô (X ′ , D′ )
Corollary 7.18. For all w ∈ Σ⋆ (⋆, ⋆) and (X, D), (X ′ , D′ ) ∈ S, we have (X, D) ⇐
̃
w
in S iff (X, D) ←
Ð (X ′ , D′ ) in C.
̃
w
The last statement is slightly subtle, in that (X, D) ←
Ð (X ′ , D′ ) denotes a single edge
̃ in fc(Σ).
in C, lying over the composite w
w
̃ and
Proof. If (X, D) ⇐
Ô (X ′ , D′ ) in S, then there exists p ∈ W⋆ such that fc(pW )(p) = w
there is a path e∶ (X ′ , D′ ) → (X, D) over p in S. Let W ∶ X ′
X denote the composition of
̃
p. By Corollary 7.17, we get an edge (X ′ , D′ ) → (X, D) over W in C. So since pW (W ) = w,
this gives us the expected transition.
̃
w
Conversely, if (X, D) ←
Ð (X ′ , D′ ) in C, then let W ∶ X ′
X denote the corresponding
W
̃ Decomposing W as some path p in
edge in Dv . In particular, we have pW (W ) = w.
(pW )⋆ (p)
W, we obtain by Corollary 7.17 a transition sequence (X, D) ⇐ÔÔÔÔ (X ′ , D′ ) in S. But
W )⋆ (p) = pW (W ) = w,
̃ as desired.
(p̃
As promised, we readily obtain:
Corollary 7.19. We have S = C .
Let us also prove the analogous result with the semantic pole ⊥ .
Lemma 7.20. We have D ∈ ⊥ X iff (X, D) ∈ C .
Proof. Assume D ∈ ⊥ X , and consider any (X, D) ← (X ′ , D′ ). The latter is witnessed
by some unsuccessful, closed-world trace W ∶ X ′
X, state σ ∈ D(W ), and isomorphism
h∶ (D ⋅ W )∣ψW (σ) −∼ D′ .
By hypothesis, σ admits an extension σ ′ ∈ D(W ● W ′ ) for some successful W ′ ∶ X ′′
X ′.
′′
′
Letting D = (D ⋅ (W ● W ))∣ψW ●W ′ (σ′ ) , we have
D′′ ≅ (((D ⋅ W )∣ψW (σ) ) ⋅ W ′ )∣ψW ′ (σ′ ) ≅ (D′ ⋅ W ′ )∣ψ
W ′ (hW ′ (σ
♡n
′ ))
,
and hence (X ′ , D′ ) ←Ð (X ′′ , D′′ ) for some n > 0. This shows that (X, D) ∈ C .
Conversely, assume (X, D) ∈ C and consider any unsuccessful, closed-world trace
W ∶ X′
X and state σ ∈ D(W ). Letting D′ = (D ⋅ W )∣ψW (σ) , we have (X, D) ← (X ′ , D′ ).
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
59
By hypothesis, we find some transition (X ′ , D′ ) ←
Ð (X ′′ , D′′ ), witnessed by some successful
′
′′
′
′′
′
′
W ∶X
X . Hence, D ≅ (D ⋅ W )∣ψW ′ (σ′ ) for a certain σ ′ ∈ D′ (W ′ ). By definition of D′ ,
σ ′ is a state in D(W ● W ′ ) such that σ ′ ⋅ u = σ, where u∶ W → (W ● W ′ ) is (W ′ , id ). This
gives the desired successful extension of σ, which shows that D ∈ ⊥ X .
♡
Combining the last two results, we may now prove that the semantic pole coincides with
that of S:
Proof of Lemma 7.4. We have D ∈ ⊥ X iff (X, D) ∈ C iff (X, D) ∈ S by Corollary 7.19 and
Lemma 7.20.
As expected, this entails preservation and reflection of semantic fair testing equivalence:
Corollary 7.21. For all D ∈ DX and D′ ∈ DX ′ , we have (X, D) ∼f (X ′ , D′ ) iff (X, D) ∼Sf
(X ′ , D′ )
Proof. Let us first show that semantic fair testing equivalence may as well be defined only
with definite tests. Indeed, if IX ≠ IX ′ then the result holds trivially. So assuming IX = IX ′ ,
consider any test B ∈ BY with IX = IY , and, w.l.o.g., [D, B] ∈ ⊥ Z and [D′ , B] ∉ ⊥ Z ′ with
Z = X +IX Y and Z ′ = X ′ +IX ′ Y . Then, letting B = ∑i∈γ Di with each Di definite, there
exists i such that [D, Di ] ∈ ⊥ Z and [D′ , Di ] ∉ ⊥ Z ′ , hence Di also distinguishes D from D′ .
Returning to our main goal, for any definite test T ∈ DY with IY = IX , by Lemma 7.4,
[D, T ] ∈ ⊥ Z iff (Z, [D, T ]) = ((X, D) ∣ (Y, T )) ∈ S , which easily entails the result.
From this we may reduce our main theorem to a result on S:
Corollary 7.22. If the translation ⟦−⟧∶ ob(Conf) → ∑X DX is intensionally fully abstract
for ∼Sf , then Theorem 6.25 holds, i.e., ⟦−⟧ is also intensionally fully abstract for ∼f .
Proof. For all configurations C1 and C2 , by Corollary 7.21, ⟦C1 ⟧ ∼f ⟦C2 ⟧ iff ⟦C1 ⟧ ∼Sf ⟦C2 ⟧,
which holds iff C1 ∼Conf
C2 by hypothesis.
f
It remains to prove that surjectivity up to ∼f reduces to surjectivity up to ∼Sf . For this,
let us first show that any behaviour is ∼f -equivalent to some definite one. Indeed, consider
any B ∈ BX . Letting B ⋅ x = ∑i∈nx Dix for all agents x in X, B is fair testing equivalent to
the definite behaviour D such that
D ⋅ x = ⟨τnx ↦ ∑ Dix ⟩,
i∈nx
except if B ⋅ x = ∅ for some x. But in the latter case, B is fair testing equivalent to the
definite behaviour on one nullary player with the same interface which merely ticks.
Thus, we may restrict attention to definite behaviours. So consider any definite D ∈
DX . By hypothesis, there exists C such that ⟦C⟧ ∼Sf (X, D), hence ⟦C⟧ ∼f (X, D) by
Corollary 7.21, which concludes the proof.
60
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
7.2. A further graph with testing for behaviours. In the previous section, we have
characterised semantic fair testing equivalence using the graph with testing S, and reduced
intensional full abstraction of ⟦−⟧ w.r.t. ∼f to intensional full abstraction w.r.t. ∼Sf . We now
define a further graph with testing, M, which will help us bridge the gap between S and
π-calculus configurations. Indeed, we define a surjective morphism m∶ S → M over Σ, and we
then prove that m is intensionally fully abstract (Proposition 7.30), from which we deduce
that our main result follows from intensional full abstractness of the composite translation
T = m ○ ⟦−⟧ (Lemma 7.32).
Recall from Definition 2.3 that −⊙ denotes the finite multiset monad on sets.
Definition 7.23. Let the set M0 of mixed behaviours be
n ⊙
∑ ( ∑ D[n] × γ ) .
γ∈Pf (N) n∈N
The graph M over Σ is inductively defined by the rules in Figure 7, where S[γ1 ⊆ γ2 ] denotes
pointwise composition of the substitution component with the inclusion h∶ γ1 ⊆ γ2 , i.e., each
D[σ] is replaced by D[h ○ σ].
Let pM ∶ M → Σ denote the projection.
i ∈ ∣D ⋅ πnl ∣
j ∈ ∣D ⋅ πnr ∣
i ∈ ∣D ⋅ τn ∣
id
id
⟨γ∥D[σ]⟩ ←Ð ⟨γ∥(D ⋅ πnl )∣i [σ], (D ⋅ πnr )∣j [σ]⟩
⟨γ∥D[σ]⟩ ←Ð ⟨γ∥(D ⋅ τn )∣i [σ]⟩
i ∈ ∣D ⋅ νn ∣
i ∈ ∣D ⋅ ♡n ∣
⟨γ∥D[σ]⟩ ←
Ð ⟨γ∥(D ⋅ ♡n )∣i [σ]⟩
♡
i ∈ ∣D1 ⋅ ιn1 ,a1 ∣
a∉γ
σ+⌜a⌝
id
⟨γ∥D[σ]⟩ ←Ð ⟨γ, a∥(D ⋅ νn )∣i [n + 1 ÐÐÐ→ γ, a]⟩
j ∈ ∣D2 ⋅ on2 ,a2 ,b2 ∣
σ1 (a1 ) = σ2 (a2 )
[σ1 ,⌜σ2 (b2 )⌝]
id
⟨γ∥D1 [σ1 ], D2 [σ2 ]⟩ ←Ð ⟨γ∥(D1 ⋅ ιn1 ,a1 )∣i [n1 + 1 ÐÐÐÐÐÐÐ→ γ], (D2 ⋅ on2 ,a2 ,b2 )∣j [σ2 ]⟩
α
⟨γ1 ∥S1 ⟩ ←
Ð ⟨γ2 ∥S2 ⟩
α
⟨γ1 ∥S ∪ S1 ⟩ ←
Ð ⟨γ2 ∥S[γ1 ⊆ γ2 ] ∪ S2 ⟩
id
⟨γ∥S⟩ ←Ð ⟨γ∥S⟩
Figure 7: Transitions in M
Notation 7.24. Similarly to the notation for configurations, we denote
(γ, [(n1 , D1 , σ1 ), . . . , (np , Dp , σp )])
by
⟨γ∥D1 [σ1 ], . . . , Dp [σp ]⟩.
For the testing structure of M, we mimick Definition 2.7 and put:
Definition 7.25. For any ⟨γ∥S⟩, ⟨γ ′ ∥S ′ ⟩ ∈ M, let ⟨γ∥S⟩@⟨γ ′ ∥S ′ ⟩ denote ⟨γ∥S ∪ S ′ ⟩ if γ = γ ′
and be undefined otherwise. Let furthermore γ = ⟨γ∥⟩.
By Lemma 2.18, we have:
Proposition 7.26. The morphism fc(pM )∶ fc(M) → fc(Σ), with the graph of @ as testing
relation, forms a free graph with testing.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
61
Let us now reduce Theorem 6.25 to a statement about M. In order to do this, we will
use Lemma 2.22 and hence need to exhibit a fair relation ob(S)
M0 . We use the graph of
the following map:
Definition 7.27. Let m∶ ob(S) → M0 map any (X, D) to the mixed behaviour
⟨X(⋆)∥[(D ⋅ x)[σx ] ∣ (n, x) ∈ Agents(X)]⟩,
[⌜x⋅si ⌝]i∈n
where σx is the map n ÐÐÐÐÐ→ X(⋆).
We now need to show that m yields a fair relation. Most points are direct, like totality
or the fact that (X, D) ¨ (X ′ , D′ ) iff m(X, D) ¨ m(X ′ , D′ ). Furthermore, we have:
Proposition 7.28. The map m∶ ob(S) → M0 is surjective. Let a denote any section of m.
The point about bisimilarity is trickier:
Proposition 7.29. We have (X, D) ∼Σ m(X, D) for all (X, D) ∈ ob(S).
Proof. It is enough to prove that (the graph of) m is a strong bisimilarity up to strong
bisimilarity. For this, let us record that clearly for any isomorphism h∶ X → Y of positions
and D ∈ DY , we have (Y, D) ∼Σ (X, D ⋅ h) in S. Let us call I (for isomorphism) the relation
given by all pairs ((Y, D), (X, D ⋅ h)), so that we have I ⊆ ∼Σ . By case analysis, we can
α
show that m is a strong bisimulation up to I, i.e., for all (X, D) ←
Ð (Y, D′ ), there exists
α
(Y, D′ ) I (Z, D′′ ) such that m(X, D) ←
Ð m(Z, D′′ ), as below left. And more tightly, for all
α
α
m(X, D) ←
Ð M ′ , there exists (X, D) ←
Ð (X ′ , D′ ) such that m(X ′ , D′ ) = M ′ , as below right.
(X, D)
m
m(X, D)
α
(Y, D ) I (Z, D )
′′
m
m
α
α
′
(X, D)
m(Z, D )
′′
(Y, D )
′
m
m(X, D)
α
′
M
The right-hand diagram is a tedious yet straightforward case analysis. The left-hand one is
α
also a tedious case analysis, whose main point is that for all transitions (X, D) ←
Ð (Y, D′ ),
some renaming of elements of X may take place, which cannot happen in any transition
from m(X, D). So in each case we need to find the Z and corresponding D′′ which avoids
such renaming. In fact, this goes by indentifying the right transition from m(X, D) and
showing that the obtained M ′ is of the form m(Z, D′′ ) for (Y, D′ ) I (Z, D′′ ). Let us do one
♡
case. If (X, D) ←
Ð (Y, D′ ), then we have an iso h∶ X −∼ Y , and there is some agent (n0 , x0 )
in X and i such that D′ ⋅ h(x0 ) = (D ⋅ x0 ⋅ ♡n0 )∣i and furthermore for all (n, x) ≠ (n0 , x0 ) in
Agents(X), we have D′ ⋅ h(x) = D ⋅ x. So, letting y = h(x) for all such x, we indeed have
M = ⟨γ∥(D ⋅x0 )[σx0 ]∶∶[(D ⋅x)[σx ] ∣ x ≠ x0 ]⟩
♡
⟨γ∥(D′ ⋅y0 )[σx0 ]∶∶[(D ⋅x)[σx ] ∣ x ≠ x0 ]⟩ = M ′
with M = m(X, D), and furthermore letting D′′ be determined by D′′ ⋅ x = D′ ⋅ h(x) = D ⋅ x
for x ≠ x0 and D′′ ⋅ x = D′ ⋅ h(x0 ), we have D′′ = D′ ⋅ h and hence (Y, D′ ) I (X, D′′ ) with
m(X, D′′ ) = M ′ , as desired.
62
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
We are now ready to show:
Proposition 7.30. The map m∶ ob(S) → ob(M) is intensionally fully abstract for fair
testing equivalence.
Proof. As announced, for preservation and reflection of fair testing equivalence we apply
Lemma 2.22: we have established all necessary hypotheses, except the last one, which follows
by choosing a pushout with the same set of channels as its summands.
For surjectivity up to ∼M
f , consider any ⟨γ∥S⟩ ∈ M0 . Because a is a section of m, we have
m(a(⟨γ∥S⟩)) = ⟨γ∥S⟩, hence m(a(⟨γ∥S⟩)) ∼M
f ⟨γ∥S⟩, thus providing the desired antecedent.
Let us now reduce Theorem 6.25 to its analogue about M.
Definition 7.31. Let T∶ ob(Conf) → ob(M) denote the composite
⟦−⟧
m
ob(Conf) ÐÐ→ ob(S) Ð
→ ob(M).
Concretely, we have
−1
T⟨γ∥P1 , . . . , Pn ⟩ = ⟨γ∥⟦P1 ⟧hγ [h−1
γ ], . . . , ⟦Pn ⟧hγ [hγ ]⟩.
Lemma 7.32. The translation ⟦−⟧ from Definition 6.22 is intensionally fully abstract for
∼Sf if T is for ∼M
f .
Proof. Assuming T is intensionally fully abstract, then for all configurations C and C ′ , we
′
have that C ∼f C ′ iff T(C) ∼M
f T(C ). But by Proposition 7.30, we have
′
′
T(C) = m⟦C⟧ ∼M
f m⟦C ⟧ = T(C )
iff
⟦C⟧ ∼Sf ⟦C ′ ⟧,
hence C ∼f C ′ iff ⟦C⟧ ∼f S⟦C ′ ⟧.
Finally, for any (X, D) ∈ S, by intensional full abstractness of T, we find a configuration
C such that T(C) ∼M
f m(X, D), hence by Proposition 7.30 again ⟦C⟧ ∼f (X, D).
7.3. Intensional full abstraction. We at last prove our main result, by proving intensional
full abstractness of T. Our strategy is to define a relation &∶ Conf
M over Σ which
● relates any configuration to its image under T, and
● is surjective, i.e., relates any mixed behaviour to some configuration.
We will then show that this relation is a weak bisimulation over Σ, which will entail the
result.
Let us start with the second point, and define a map Z ∈ ob(M) → ob(Conf), which
associates a configuration to each mixed behaviour. We first coinductively define ζ for
definite behaviours on representable positions by:
⎛
⎜+
⎜
⎜+
ζ(n ⊢D D) = ⎜
⎜+
⎜
⎜
⎜+
⎝+
∑i∈∣D⋅πnl ∣,j∈∣D⋅πnr ∣ τ.(ζ(n ⊢D (D ⋅ πnl )∣i ) ∣ ζ(n ⊢D (D ⋅ πnr )∣j ))
∑i∈∣D⋅τn ∣ τ.ζ(n ⊢D (D ⋅ τn )∣i )
∑i∈∣D⋅♡n ∣ ♡.ζ(n ⊢D (D ⋅ ♡n )∣i )
∑i∈∣D⋅νn ∣ ν(n + 1).ζ(n + 1 ⊢D (D ⋅ νn )∣i )
∑a∈n,i∈∣D⋅ιn,a ∣ a(n + 1).ζ(n + 1 ⊢D (D ⋅ ιn,a )∣i )
∑a,b∈n,i∈∣D⋅on,a,b ∣ ā⟨b⟩.ζ(n ⊢D (D ⋅ on,a,b )∣i )
where n ⊢D D means D ∈ D[n] .
⎞
⎟
⎟
⎟
⎟,
⎟
⎟
⎟
⎟
⎠
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
63
Except perhaps for the first term of the sum, this should be rather natural: each definite
behaviour on a representable position corresponds to a guarded sum, with one term for each
state over each basic action – the translation is direct. The twist in the first term is due to
the fact that forking is not a guard in π, so P ∣ Q cannot appear in any guarded sum. But
τ.(P ∣ Q) can, and it is clearly weakly bisimilar to P ∣ Q, so this is precisely what ζ does.
Let us now extend ζ to arbitrary configurations:
Definition 7.33. Let Z∶ ob(M) → ob(Conf) be defined by
Z⟨γ∥D1 [σ1 ], . . . , Dn [σn ]⟩ = ⟨γ∥ζ(D1 )[σ1 ], . . . , ζ(Dn )[σn ]⟩.
Remark 7.34. Let us emphasise that brackets on the right denote proper substitutions,
while on the left Di [σi ] is just syntactic sugar for (ni , Di , σi ) by Notation 7.24.
In order for Z to return an antecedent up to ∼M
f , we immediately observe that for any
l
r
D and i ∈ ∣D ⋅ πn ∣ and j ∈ ∣D ⋅ πn ∣,
● on the one hand D has a silent transition to ∂i,j D, the behaviour on [n] ∣ [n] such that
(∂i,j D) ⋅ x1 = (D ⋅ πnl )∣i and (∂i,j D) ⋅ x2 = (D ⋅ πnr )∣j (where x1 and x2 denote the two agents
of [n] ∣ [n]);
● on the other hand ζ(D) has a silent transition to
′
ζi,j
(D) = (ζ((D ⋅ πnl )∣i ) ∣ ζ((D ⋅ πnr )∣j )),
which then has a further silent transition to the two-process configuration consisting of
ζ((D ⋅ πnl )∣i ) and ζ((D ⋅ πnr )∣j ).
id
′
Thus, when we try to relate D and ζ(D), the transition ζ(D) ←Ð ζi,j
(D) has to be matched
id
by the former transition D ←Ð ∂i,j D. So our relation & should somehow include pairs
′
(ζi,j
(D), ∂i,j D).
Definition 7.35. Let the relation &∶ Conf
in Figure 8.
M over Σ be defined inductively by the rules
(In the last rule, γ is understood in Conf on the left, and in M on the right.)
γ′ ⊢ P
h∶ γ ′
∼
−
n
n ⊢D D
σ∶ n → γ
P [σ ○ h] & ⟦P ⟧h [σ]
n ⊢D D1 , D2
σ∶ n → γ
2
1
ζ(D)[σ] & D[σ]
C&M
(ζ(D ) ∣ ζ(D ))[σ] & D [σ], D [σ]
1
2
σ∶ n → γ
D&N
C@D & M @N
γ & γ
⋅
Figure 8: The relation &
Lemma 7.36. We have C & T(C) for all C and Z(M ) & M for all M , and so & is total
and surjective.
Proof. By construction.
64
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Lemma 7.37. The relation & is a weak bisimulation over Σ.
Proof. First, we observe that & may equivalently be defined by first letting &0 be generated
by all rules except the last two, and then adding the rule
C1 &0 M1
...
Cn &0 Mn
C1 @ . . . @Cn & M1 @ . . . @Mn
(n ∈ N)⋅
The advantage of this presentation is that proofs of C & M all have depth at most 1.
The rest is then a tedious case analysis, which we defer to Appendix A. In summary, we
easily observe that the ‘forwards’ clause of weak bisimulation is satisfied, the only subtlety
being that heating (ζ(D1 ) ∣ ζ(D2 ))[σ] should be matched by the identity edge on the
corresponding behaviour. Furthermore, the ‘backwards’ clause is also easily satisfied, the
only subtlety being that if the considered behaviour performs a transition involving one or
several [D1 [σ], D2 [σ]]’s, related on the left to (ζ(D1 ) ∣ ζ(D2 ))[σ]’s, then all of the latter
first have to heat to [ζ(D1 )[σ], ζ(D2 )[σ]], and only then perform the matching transition.
This easily entails:
Lemma 7.38. For all C1 , C2 , M1 , M2 , if C1 & M1 and C2 & M2 , then C1 ∼f C2 iff M1 ∼f M2 .
Proof. The relation & is weakly fair, the only difficult points being proved by Lemmas 7.36
and 7.37 above. We thus conclude by Corollary 2.25.
Theorem 7.39. The map T∶ ob(Conf) → ob(M) is intensionally fully abstract.
Proof. Lemma 7.38 directly entails preservation and reflection of fair testing equivalence.
M
Regarding surjectivity up to ∼M
f , for any M ∈ M0 , we have M ∼f T(Z((M ))). Indeed, for
′
any M , we have
M @M ′ ∈
iff Z(M )@Z(M ′ ) ∈
(because Z(M )@Z(M ′ ) & M @M ′ and by Lemma 7.37)
iff T(Z(M ))@M ′ ∈
(because Z(M )@Z(M ′ ) & T(Z(M ))@M ′ and by Lemma 7.37),
as desired.
We are now able to prove our main result:
Proof of Theorem 6.25. By Theorem 7.39, T is intensionally fully abstract for ∼M
f , so by
S
Lemma 7.32 ⟦−⟧ is intensionally fully abstract for ∼f . We thus conclude by Corollary 7.22.
7.4. Generalisation. We now show that our main results generalise beyond fair testing
equivalence. Indeed, let us put:
Definition 7.40. A pole is a property of states over fc(Σ) which is stable under strong
bisimilarity.
There is a slight size issue in this definition, as it quantifies over elements of all graphs
over fc(Σ). The reader may understand this using whatever fix they prefer, e.g., using a
universe or some modal logic.
Example 7.41. Consider any x ∈ G over fc(Σ). We have
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
65
Ð x′′ ;
● x is in the pole for fair testing equivalence iff for all x ← x′ there exists x′ ←
♡ ′
● x is in the pole for may testing equivalence iff there exists x ←
Ðx.
Must testing equivalence is less easy to capture, for reasons explained in [41]. Here is an
exotic, yet perhaps relevant pole: x is in it iff for all finite, not-necessarily silent transition
♡
sequences x ←
Ð⋆ x′ , there exists x′ ←
Ð x′′ . In other words, x never loses the ability to tick.
The induced equivalence is clearly at least as fine as fair testing equivalence, but we leave
open the question of whether or not it is strictly finer.
♡
Definition 7.42. For any such pole , let ∼ denote the testing equivalence induced by
replacing G by in the definition of fair testing equivalence (Definition 2.14).
Semantic testing equivalence may then be taken to be testing equivalence on C (Definition 7.7), and we get the exact analogue of Theorem 6.25 (without changing the model in
any way).
8. Conclusion and future work
We have described our playground for π and the induced sheaf model, which we have proved
intensionally fully abstract for a wide range of testing equivalences.
Regarding future work: our proof that traces form a playground uses a new technique
based on factorisation systems. Since submission of this paper, we have designed [22, 23] a
general setting where this technique applies, and used it to bridge the gap between our notion
of plays based on string diagrams and the standard one based on justified sequences [71].
We also consider applying our notion of trace to error diagnostics [32] or efficient machine
representation of reversible π-calculus processes [19]. Longer-term directions include applying
the approach to more complex calculi, e.g., calculi with passivation [50] or functional calculi,
and eventually consider some full-fledged functional language with concurrency primitives.
Finally, deriving the complex notion of trace evoked in Section 1.5 from the one exposed
here is akin to deriving ltss from reduction rules [48, 65]. Since the issue still seems easier
on traces than on a full operational semantics specification, this might be worthwile to
investigate further. In the same vein, the emphasis we put on traces suggests that we
might be able to deduce properties of type systems (soundness, progress, etc) or compilers
(correctness) from corresponding properties on traces.
References
[1] LICS 1996. Proc. 11th Symposium on Logic in Computer Science, 1996. IEEE.
[2] LICS 2015. Proc. 30th Symposium on Logic in Computer Science, 2015. IEEE.
[3] Samson Abramsky, Radha Jagadeesan, and Pasquale Malacaria. Full abstraction for
PCF. Information and Computation, 163(2):409–470, 2000. doi:10.1006/inco.2000.2930.
[4] Jir̆ı́ Adámek and Jir̆ı́ Rosicky. Locally Presentable and Accessible Categories. Cambridge
University Press, 1994. doi:10.1017/CBO9780511600579.
[5] Gérard Berry and Gérard Boudol. The chemical abstract machine. In Proc. 17th
International Symposium on Principles of Programming Languages, pages 81–94, 1990.
doi:10.1145/96709.96717.
66
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
[6] Filippo Bonchi, Marcello M. Bonsangue, Jan J. M. M. Rutten, and Alexandra Silva.
Deriving syntax and axioms for quantitative regular behaviours. In Proc. 21st International Conference on Concurrency Theory, volume 5710 of LNCS, pages 146–162.
Springer, 2009. doi:10.1007/978-3-642-04081-8 11.
[7] Michele Boreale and Davide Sangiorgi. A fully abstract semantics for causality in the
π-calculus. Acta Informatica, 35(5):353–400, 1998.
[8] A. K. Bousfield. Constructions of factorization systems in categories. Journal of Pure
and Applied Algebra, 9(2-3):287–329, 1977.
[9] Ed Brinksma, Arend Rensink, and Walter Vogler. Fair testing. In CONCUR 1995 [17],
pages 313–327. ISBN 3-540-60218-6. doi:10.1007/3-540-60218-6 23.
[10] Roberto Bruni, Hernán C. Melgratti, and Ugo Montanari. Event structure semantics for nominal calculi. In Proc. 18th International Conference on Concurrency
Theory, volume 4137 of LNCS, pages 295–309. Springer, 2006. ISBN 3-540-37376-4.
doi:10.1007/11817949 20.
[11] Nadia Busi and Roberto Gorrieri. A Petri net semantics for π-calculus. In CONCUR
1995 [17], pages 145–159. ISBN 3-540-60218-6. doi:10.1007/3-540-60218-6 11.
[12] Nadia Busi and Roberto Gorrieri. Distributed semantics for the π-calculus based on
Petri nets with inhibitor arcs. Journal of Logic and Algebraic Programming, 78(3):
138–162, 2009. doi:10.1016/j.jlap.2008.08.002.
[13] Diletta Cacciagrano, Flavio Corradini, and Catuscia Palamidessi. Explicit fairness in
testing semantics. Logical Methods in Computer Science, 5(2), 2009. doi:10.2168/LMCS5(2:15)2009.
[14] Simon Castellan, Pierre Clairambault, and Glynn Winskel. The parallel intensionally
fully abstract games model of pcf. In LICS 2015 [2].
[15] Gian Luca Cattani and Peter Sewell. Models for name-passing processes: Interleaving
and causal. In Proc. 15th Symposium on Logic in Computer Science, pages 322–333.
IEEE, 2000. ISBN 0-7695-0725-5. doi:10.1109/LICS.2000.855781.
[16] Gian Luca Cattani, Ian Stark, and Glynn Winskel. Presheaf models for the π-calculus.
In Category Theory and Computer Science, volume 1290 of LNCS, pages 106–126.
Springer, 1997. ISBN 3-540-63455-X. doi:10.1007/BFb0026984.
[17] CONCUR 1995. Proc. 6th International Conference on Concurrency Theory, volume
962 of LNCS, 1995. Springer. ISBN 3-540-60218-6.
[18] Silvia Crafa, Daniele Varacca, and Nobuko Yoshida. Event structure semantics of
parallel extrusion in the pi-calculus. In Proc. 15th Foundations of Software Science and
Computational Structures, volume 7213 of LNCS, pages 225–239. Springer, 2012. ISBN
978-3-642-28728-2. doi:10.1007/978-3-642-28729-9 15.
[19] Ioana Cristescu, Jean Krivine, and Daniele Varacca. A compositional semantics for the
reversible π-calculus. In Proc. 28th Symposium on Logic in Computer Science, pages
388–397. IEEE, 2013. ISBN 978-1-4799-0413-6. doi:10.1109/LICS.2013.45.
[20] Rocco De Nicola and Matthew Hennessy. Testing equivalences for processes. Theoretical
Computer Science, 34:83–133, 1984. doi:10.1016/0304-3975(84)90113-0.
[21] Pierpaolo Degano and Corrado Priami. Non-interleaving semantics for mobile processes. Theoretical Computer Science, 216(1-2):237–270, 1999. doi:10.1016/S03043975(99)80003-6.
[22] Clovis Eberhart and Tom Hirschowitz. Justified sequences in string diagrams: a
comparison between two approaches to concurrent game semantics. Preprint, 2016.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
67
[23] Clovis Eberhart and Tom Hirschowitz. Justified sequences in string diagrams: a
comparison between two approaches to concurrent game semantics. In Proc. 7th
International Conference on Algebra and Coalgebra in Computer Science, LIPIcs. Schloss
Dagstuhl - Leibniz-Zentrum fuer Informatik, 2017. To appear.
[24] Charles Ehresmann. Catégories structurées. Annales scientifiques de l’Ecole Normale
Supérieure, 80(4):349–426, 1963.
[25] Charles Ehresmann. Catégories et structures. Dunod, 1965.
[26] Joost Engelfriet. A multiset semantics for the pi-calculus with replication. Theoretical
Computer Science, 153(1&2):65–94, 1996. doi:10.1016/0304-3975(95)00118-2.
[27] Marcelo P. Fiore and Daniele Turi. Semantics of name and value passing. In Proc.
16th Symposium on Logic in Computer Science, pages 93–104. IEEE, 2001. ISBN
0-7695-1281-X. doi:10.1109/LICS.2001.932486.
A fully-abstract
[28] Marcelo P. Fiore, Eugenio Moggi, and Davide Sangiorgi.
model for the pi-calculus (extended abstract). In LICS 1996 [1], pages 43–54.
doi:10.1109/LICS.1996.561302.
[29] P. J. Freyd and G. M. Kelly. Categories of continuous functors, I. Journal of Pure and
Applied Algebra, 2:169–191, 1972. doi:10.1016/0022-4049(72)90001-1.
[30] Richard H. G. Garner. Polycategories. PhD thesis, University of Cambridge, 2006.
[31] Jean-Yves Girard. Locus solum: From the rules of logic to the logic of rules. Mathematical
Structures in Computer Science, 11(3):301–506, 2001. doi:10.1007/3-540-44802-0 3.
[32] Gregor Gösler, Daniel Le Métayer, and Jean-Baptiste Raclet. Causality analysis in
contract violation. In Proc. of Runtime Verification, volume 6418 of LNCS, pages
270–284. Springer, 2010. ISBN 978-3-642-16611-2. doi:10.1007/978-3-642-16612-9 21.
[33] Marco Grandis and Robert Paré. Limits in double categories. Cahiers de Topologie et
Géométrie Différentielle Catégoriques, 40(3):162–220, 1999.
[34] Marco Grandis and Robert Paré. Adjoints for double categories. Cahiers de Topologie
et Géométrie Différentielle Catégoriques, 45(3):193–240, 2004.
[35] Russell Harmer, Martin Hyland, and Paul-André Melliès. Categorical combinatorics for
innocent strategies. In Proc. 22nd Symposium on Logic in Computer Science, pages
379–388. IEEE, 2007. doi:10.1109/LICS.2007.14.
[36] Matthew Hennessy. A fully abstract denotational semantics for the π-calculus. Theoretical Computer Science, 278(1-2):53–89, 2002. doi:10.1016/S0304-3975(00)00331-5.
[37] Thomas T. Hildebrandt. Towards categorical models for fairness: fully abstract presheaf
semantics of SCCS with finite delay. Theoretical Computer Science, 294(1/2):151–181,
2003. doi:10.1016/S0304-3975(01)00247-X.
[38] Tom Hirschowitz. Full abstraction for fair testing in CCS. In Proc. 5th International
Conference on Algebra and Coalgebra in Computer Science, volume 8089 of LNCS, pages
175–190. Springer, 2013. doi:10.1007/978-3-642-40206-7 14.
[39] Tom Hirschowitz. Full abstraction for fair testing in CCS (expanded version). Logical
Methods in Computer Science, 10(4), 2014. doi:10.2168/LMCS-10(4:2)2014.
[40] Tom Hirschowitz and Damien Pous. Innocent strategies as presheaves and interactive
equivalences for CCS. In Proc. of Interaction and Concurrency Experience, pages 2–24.
Electronic Proceedings in Theoretical Computer Science, 2011. doi:10.4204/EPTCS.59.2.
[41] Tom Hirschowitz and Damien Pous. Innocent strategies as presheaves and interactive
equivalences for CCS. Scientific Annals of Computer Science, 22(1):147–199, 2012.
doi:10.7561/SACS.2012.1.147. Selected papers from ICE ’11.
68
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
[42] J. M. E. Hyland and C.-H. L. Ong. On full abstraction for PCF: I, II, and III. Information
and Computation, 163(2):285–408, 2000. doi:10.1006/inco.2000.2917.
[43] Bart Jacobs. Categorical Logic and Type Theory. Number 141 in Studies in Logic and
the Foundations of Mathematics. North Holland, Amsterdam, 1999.
[44] Lalita Jategaonkar Jagadeesan and Radha Jagadeesan. Causality and true concurrency:
A data-flow analysis of the pi-calculus (extended abstract). In Proc. of Algebraic
Methodology and Software Technology, volume 936 of LNCS, pages 277–291. Springer,
1995. doi:10.1007/3-540-60043-4 59.
[45] André Joyal, Mogens Nielsen, and Glynn Winskel. Bisimulation and open maps.
In Proc. 8th Symposium on Logic in Computer Science, pages 418–427. IEEE, 1993.
doi:10.1109/LICS.1993.287566.
[46] Stephen Lack and Pawel Sobociński. Adhesive categories. In Proc. 7th Foundations of
Software Science and Computational Structures, volume 2987 of LNCS, pages 273–288.
Springer, 2004. doi:10.1007/978-3-540-24727-2 20.
[47] James Laird. Game semantics for higher-order concurrency. In Proc. 26th Foundations
of Software Technology and Theoretical Computer Science, volume 4337 of LNCS, pages
417–428. Springer, 2006. doi:10.1007/11944836 38.
[48] James J. Leifer and Robin Milner. Deriving bisimulation congruences for reactive systems.
In Proc. 11th International Conference on Concurrency Theory, volume 1877 of LNCS,
pages 243–258. Springer, 2000. ISBN 3-540-67897-2. doi:10.1007/3-540-44618-4 19.
[49] Tom Leinster. Higher Operads, Higher Categories, volume 298 of London Mathematical
Society Lecture Notes. Cambridge University Press, Cambridge, 2004.
[50] Sergueı̈ Lenglet, Alan Schmitt, and Jean-Bernard Stefani. Normal bisimulations in calculi
with passivation. In Proc. 12th Foundations of Software Science and Computational
Structures, volume 5504 of LNCS, pages 257–271. Springer, 2009. doi:10.1007/978-3642-00596-1 19.
[51] Paul Blain Levy. Morphisms between plays. GaLoP, 2013.
[52] Saunders Mac Lane. Duality for groups. Bulletin of the American Mathematical Society,
56, 1950.
[53] Saunders Mac Lane. Categories for the Working Mathematician. Number 5 in Graduate
Texts in Mathematics. Springer, 2nd edition, 1998.
[54] Saunders Mac Lane and Ieke Moerdijk. Sheaves in Geometry and Logic: A First
Introduction to Topos Theory. Universitext. Springer, 1992.
[55] Guy McCusker, John Power, and Cai Wingfield. A graphical foundation for
schedules. Electronic Notes in Theoretical Computer Science, 286:273–289, 2012.
doi:10.1016/j.entcs.2012.08.018.
[56] Paul-André Melliès. Asynchronous games 2: the true concurrency of innocence. In Proc.
15th International Conference on Concurrency Theory, volume 3170 of LNCS, pages
448–465. Springer, 2004. doi:10.1007/978-3-540-28644-8 29.
[57] Paul-André Melliès. Asynchronous games 4: A fully complete model of propositional
linear logic. In Proc. 20th Symposium on Logic in Computer Science, pages 386–395.
IEEE, 2005. ISBN 0-7695-2266-1. doi:10.1109/LICS.2005.6.
[58] Paul-André Melliès. Game semantics in string diagrams. In Proc. 27th Symposium on
Logic in Computer Science, pages 481–490. IEEE, 2012. doi:10.1109/LICS.2012.58.
[59] Ugo Montanari and Marco Pistore. Concurrent semantics for the π-calculus. Electronic Notes in Theoretical Computer Science, 1:411–429, 1995. doi:10.1016/S15710661(04)00024-6.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
69
[60] Ugo Montanari and Marco Pistore. Structured coalgebras and minimal HDautomata for the π-calculus. Theoretical Computer Science, 340(3):539–576, 2005.
doi:10.1016/j.tcs.2005.03.014.
[61] V. Natarajan and Rance Cleaveland. Divergence and fair testing. In Proc. 22nd
International Colloquium on Automata, Languages and Programming, volume 944 of
LNCS, pages 648–659. Springer, 1995. ISBN 3-540-60084-1. doi:10.1007/3-540-600841 112.
[62] Hanno Nickau. Hereditarily sequential functionals. In Proc. Logical Foundations of
Computer Science, volume 813 of LNCS, pages 253–264. Springer, 1994. doi:10.1007/3540-58140-5 25.
[63] Mogens Nielsen, Gordon Plotkin, and Glynn Winskel. Event structures and domains,
part 1. Theoretical Computer Science, 13:65–108, 1981.
[64] Robert Paré. Yoneda theory for double categories. Theory and Applications of Categories,
25(17):436–489, 2011.
[65] Julian Rathke and Pawel Sobociński. Deconstructing behavioural theories of mobility.
In IFIP TCS, volume 273 of IFIP, pages 507–520. Springer, 2008. doi:10.1007/978-0387-09680-3 34.
[66] Arend Rensink and Walter Vogler. Fair testing. Information and Computation, 205(2):
125–198, 2007. doi:10.1016/j.ic.2006.06.002.
[67] Silvain Rideau and Glynn Winskel. Concurrent strategies. In Proc. 26th Symposium on
Logic in Computer Science, pages 409–418. IEEE, 2011. doi:10.1109/LICS.2011.13.
[68] Davide Sangiorgi and Jan Rutten, editors. Advanced Topics in Bisimulation and Coinduction. Number 52 in Cambridge Tracts in Theoretical Computer Science. Cambridge
University Press, 2011.
[69] Davide Sangiorgi and David Walker. The π-calculus - a theory of mobile processes.
Cambridge University Press, 2001.
[70] Ian Stark. A fully abstract domain model for the π-calculus. In LICS 1996 [1], pages
36–42. doi:10.1109/LICS.1996.561301.
[71] Takeshi Tsukada and C.-H. Luke Ong. Nondeterminism in game semantics via sheaves.
In LICS 2015 [2].
[72] Glynn Winskel. Event structure semantics for CCS and related languages. In Proc. 9th
International Colloquium on Automata, Languages and Programming, volume 140 of
LNCS, pages 561–576. Springer, 1982. ISBN 3-540-11576-5. doi:10.1007/BFb0012800.
[73] Glynn Winskel. Strategies as profunctors. In Proc. 16th Foundations of Software Science
and Computational Structures, volume 7794 of LNCS, pages 418–433. Springer, 2013.
doi:10.1007/978-3-642-37075-5 27.
70
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
Appendix A. Proof of Lemma 7.37
One wants to check two properties:
a
a
→A M with C ′ & M ′ ;
→A C with C & M , there exists M ′ Ð
(LH) for all transitions C ′ Ð
a
a
(RH) for all transitions M A ←
Ð M ′ with C & M , there exists C A ⇐
Ô C ′ with C ′ & M ′ .
The attentive reader will have noticed that (LH) imposes y to answer with a single transition.
This means we actually prove that & is an expansion [68, Chapter 6]. Any expansion being
in particular a weak bisimulation, this suffices.
Notation A.1. We sometimes cast processes P (resp. pairs D[σ]) over γ into configurations
⟨γ∥P ⟩ (resp. mixed behaviours ⟨γ∥D[σ]⟩). We proceed similarly for multisets of processes.
We start by proving (LH) for all cases, before proving that (RH) holds as well.
Synchro, (LH). We begin by the case of a synchronisation, i.e., when one has a transition
id
C = ⟨γ∥a(b).P +k1 R1 , ā⟨c⟩.Q +k2 R2 ⟩@C0 ←Ð ⟨γ∥P [b ↦ c], Q⟩@C0 = C ′ .
id
We want to show that there exists a transition M ←Ð M ′ with C ′ & M ′ .
We write P1 = a(b).P +k1 R1 and P2 = ā⟨c⟩.Q +k2 R2 . Neither of them are of the form
(⋅ ∣ ⋅) so they can only be related to mixed behaviours using the first two rules. Therefore,
four sub-cases should be considered, as detailed in Figure 9. If we are in case i1 for P1 and i2
for P2 , then we have two mixed behaviours D1 [σ1 ] and D2 [σ2 ] such that ni ⊢D Di , σi ∶ ni → γ,
and Pi & Di [σi ] for i = 1, 2, plus M = ⟨γ∥D1 [σ1 ], D2 [σ2 ]⟩@M0 with C0 & M0 .
● Case 1 for both P1 and P2 . We have M = ⟦P1′ ⟧h1 [σ1 ]@⟦P2′ ⟧h2 [σ2 ]@M0 , and there is a
transition
id
M ←Ð ⟦P ′ ⟧h′1 [σ1′ ]@⟦Q′ ⟧h2 [σ2 ]@M0 ,
[σ1 ,c]
h1 +!
where h′1 is γ1′ , b′ ÐÐ→ n1 + 1 and σ1′ is n1 + 1 ÐÐÐ→ γ.
h1 +!
σ1 +⌈b⌉
b↦c
Since σ1′ ○ h′1 equals γ1′ ÐÐ→ n1 + 1 ÐÐÐ→ (γ, b) ÐÐ→ γ, we have that P ′ [σ1′ ○ h′1 ] =
P ′ [(σ1 + ⌈b⌉) ○ (h1 +!)][b ↦ c] = P [b ↦ c], and therefore P [b ↦ c] & ⟦P ′ ⟧h′1 [σ1′ ]. Moreover,
it is clear that Q = Q′ [σ2 ○ h2 ] & ⟦Q′ ⟧h2 [σ2 ], and finally [P [b ↦ c], Q]@C0 & M ′ .
● Case 1 for P1 , Case 2 for P2 . We have a transition
id
M ←Ð ⟦P ′ ⟧h′1 [σ1′ ]@(D2 ⋅ on2 ,a2 ,c2 )∣j [σ2 ]@M0 = M ′ ,
h1 +!
[σ1 ,c]
where h′1 is γ1′ , b′ ÐÐ→ n1 + 1 and σ1′ is n1 + 1 ÐÐÐ→ γ.
As is the previous case, one can check that P [b ↦ c] & ⟦P ′ ⟧h′1 [σ1′ ]. Moreover, Q &
(D2 ⋅ on2 ,a2 ,c2 )∣j [σ2 ] and therefore [P [b ↦ c], Q]@C0 & M ′ .
● Case 2 for P1 , Case 1 for P2 . We have a transition
id
[σ1 ,⌈c⌉]
M ←Ð (D1 ⋅ ιn1 ,a1 )∣i [n1 + 1 ÐÐÐÐ→ γ]@⟦Q′ ⟧h′2 [σ2′ ]@M0 = M ′ .
As in the first case, we have Q = Q′ [σ2 ○ h2 ] & ⟦Q′ ⟧h2 [σ2 ]. Furthermore, since
[σ1 ,⌈c⌉]
ζ((D1 ⋅ ιn1 ,a1 )∣i )[n1 + 1 ÐÐÐÐ→ γ]
σ1 +⌈b⌉
= ζ((D1 ⋅ ιn1 ,a1 )∣i )[n1 + 1 ÐÐÐ→ γ, b][b ↦ c]
= P [b ↦ c],
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
P1
(EXPANDED VERSION)
71
Case 1
Case 2
′
′
′ ′
′
′
There exist γ1 ⊢ P1 = a1 (b ).P +k1 R1 There exist n1 ⊢D D1 , a1 ∈ n1 , and i ∈
and h1 ∶ γ1′ −∼ n1 such that
∣D1 ⋅ ιn1 ,a1 ∣ such that
P1 = P1′ [σ1 ○ h1 ]
σ1 (a1 ) = a
σ1 (h1 (a′1 )) = a
σ1 +⌜b⌝
σ1 +⌜b⌝
h1 +!
P = ζ((D1 ⋅ ιn1 ,a1 )∣i )[n1 + 1 ÐÐÐ→ γ, b].
P = P ′ [γ1′ , b′ ÐÐ→ n1 + 1 ÐÐÐ→ γ, b]
D1 = ⟦P1′ ⟧h1
P1 = ζ(D1 )[σ1 ]
R1 = R1′ [σ1 ○ h1 ].
P2 There exist γ2′ ⊢ P2′ = a¯′2 ⟨c′ ⟩.Q′ +k2 R2′ There exist n2 ⊢D D2 , a2 , c2 ∈ n2 , and
j ∈ ∣D2 ⋅ on2 ,a2 ,c2 ∣ such that
and h2 ∶ γ2′ −∼ n2 such that
σ2 (a2 ) = a
σ2 (c2 ) = c
P2 = P ′ [σ2 ○ h2 ]
2
P2 = ζ(D2 )[σ2 ]
σ2 +⌜b⌝
h2 +1
Q = Q′ [γ2′ , b ÐÐ→ n2 + 1 ÐÐÐ→ γ, b]
σ2 (h2 (a′2 )) = a
D2 = ⟦P2′ ⟧h2
Q = ζ((D2 ⋅ on2 ,a2 ,c2 )∣j )[σ2 ].
σ2 (h2 (c′2 )) = c
R2 = R2′ [σ2 ○ h2 ].
Figure 9: Synchro, (LH) cases
we have that [P [b ↦ c], Q]@C0 & M ′ .
● Case 2 for both P1 and P2 . We have a transition
[σ1 ,⌈c⌉]
id
M ←Ð (D1 ⋅ ιn1 ,a1 )∣i [n1 + 1 ÐÐÐÐ→ γ]@(D2 ⋅ on2 ,a2 ,c2 )∣j [σ2 ]@M0 = M ′ .
As in the previous cases, one can show that [P [b ↦ c], Q]@C0 & M ′ .
Heating, (LH). We now consider the case of heating, i.e., when one has a transition
τ
C = (P ∣ Q)@C0 ←
Ð [P, Q]@C0 = C ′ .
τ
We want to show that there exists a transition M ←
Ð M ′ with [P, Q]@C0 & M ′ .
We now have to consider a few cases, depending on which rule is applied for P ∣ Q in
the proof of (P ∣ Q)@C0 & M . Notice that P ∣ Q cannot be of the form ζ(D)[σ]. We are
therefore left with two cases, depending on whether the first or third rule is applied.
● If the first rule is applied, we find M0 , γ ′ ⊢ P ′ ∣ Q′ , h∶ γ ′ −∼ n, and σ∶ n → γ such that
P ∣ Q = (P ′ ∣ Q′ )[σ ○ h] and M = ⟦P ′ ∣ Q′ ⟧h [σ]@M0 , with C0 & M0 . Letting D = ⟦P ′ ∣ Q′ ⟧h ,
we notice that
D = ⟨πnl ↦ ⟦P ′ ⟧h , πnr ↦ ⟦Q′ ⟧h ⟩.
Thus, there is a transition
τ
M ←
Ð M ′ = [⟦P ′ ⟧h [σ], ⟦Q′ ⟧h [σ]]@M0
with [P, Q]@C0 & M ′ , as desired.
● If the third rule is applied, we find M0 , n ⊢D D1 , D2 and σ∶ n → γ such that
P ∣ Q = (ζ(D1 ) ∣ ζ(D2 ))[σ]
72
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
and M = [D1 [σ], D2 [σ]]@M0 . We notice that [P, Q] = [ζ(D1 )[σ], ζ(D2 )[σ]], so [P, Q] &
τ
[D1 [σ], D2 [σ]], hence [P, Q]@C0 & M . The identity transition M ←
Ð M thus fits our
needs.
Nu, (LH). We now consider the case of a ν rule, i.e., when one has a transition
τ
C = ⟨γ∥νa.P +k R⟩@C0 ←
Ð ⟨γ, a∥P ⟩@C0 [γ ⊂ γ, a] = C ′ .
τ
We want to show that there exists a transition M ←
Ð M ′ with ⟨γ, a∥P ⟩@C0 [γ ⊂ γ, a] & M ′ .
We notice that νa.P +k R cannot be obtained from the third rule. We thus consider two
cases corresponding to the first and second rules.
● If the first rule is applied, there exist M0 , γ ′ ⊢ νa.P ′ +k R′ , h∶ γ ′ −∼ n, and σ∶ n → γ such
that M = ⟦νa.P ′ +k R′ ⟧h [σ]@M0 and
νa.P +k R = (νa.P ′ +k R′ )[σ ○ h].
h+1
We write D = ⟦νa.P ′ +k R′ ⟧h as ⟨νn ↦ ⟦P ′ ⟧h′ ⟩ ⊞ ⟦R′ ⟧h , where h′ is γ ′ , a ÐÐ→ n + 1. Thus,
there is a transition
σ+⌜a⌝
τ
M ←
Ð M ′ = ⟨γ, a∥⟦P ′ ⟧h′ [n + 1 ÐÐÐ→ γ, a]⟩@M0 [γ ⊂ γ, a]
and, modulo the fact that C & M implies C[σ] & M [σ], we have P @C0 [γ ⊂ γ, a] &
h′
σ+1
⟦P ′ ⟧h′ [σ + ⌜a⌝]@M0 [γ ⊂ γ, a] since P = P ′ [γ ′ , a Ð→ n + 1 ÐÐ→ γ, a], as desired.
● If the second rule is applied, there exist M0 , n ⊢D D, and σ∶ n → γ such that M =
ζ(D)[σ]@M0 and
νa.P +k R = ζ(D)[σ].
σ+⌈a⌉
Thus, there exists i ∈ ∣D ⋅ νn ∣ such that ζ((D ⋅ νn )∣i )[n + 1 ÐÐÐ→ γ, a] = P . There is thus a
transition
τ
M ←
Ð M ′ = ⟨γ, a∥(D ⋅ νn )∣i [σ + ⌜a⌝]⟩@M0 [γ ⊂ γ, a]
with P @C0 [γ ⊂ γ, a] & M ′ , as desired.
Tick and Tau, (LH). We now consider the cases ♡ and τ , i.e., when one has a transition
ξ
C = (ξ.P +k R)@C0 ←
Ð P @C0 ,
ξ
where ξ ∈ {♡, τ }. We want to show that there exists a transition M ←
Ð M ′ with P @C0 & M ′ .
Once again, the third rule could not have been applied, and we are left with two cases
corresponding to the first and second rules.
● If the first rule is applied, then we find M0 , γ ′ ⊢ ξ.P ′ +k R′ , h∶ γ ′ −∼ n, and σ∶ n → γ such
that M = ⟦ξ.P ′ +k R′ ⟧h [σ]@M0 , C0 & M0 , and
ξ.P +k R = (ξ.P ′ +k R′ )[σ ○ h].
We write D = ⟦ξ.P ′ +k R′ ⟧h as ⟨ξn ↦ ⟦P ′ ⟧h ⟩ ⊞ ⟦R′ ⟧h . Thus, there is a transition
ξ
M = D[σ]@M0 ←
Ð ⟦P ′ ⟧h [σ]@M0 = M ′ ,
with P = P ′ [σ ○ h] and thus P @C0 & M ′ , as desired.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
73
● If the second rule is applied, then we find M0 , n ⊢D D, and σ∶ n → γ such that C0 & M0 ,
M = D[σ]@M0 , and
ξ.P +k R = ζ(D)[σ].
Then, there exists i ∈ ∣D ⋅ ξn ∣ such that ζ((D ⋅ ξn )∣i )[σ] = P . Hence, there is a transition
ξ
M ←
Ð (D ⋅ ξn )∣i [σ]@M0 = M ′
with P @C0 & M ′ , as desired.
We have thus proved that (LH) holds. We now proceed with the case analysis for (RH).
Synchro, (RH). We start, once again, with the case of the synchronisation, i.e., when M
has the shape ⟨γ∥D1 [σ1 ], D2 [σ2 ]⟩@M0 and we consider a silent transition to
[σ1 ,⌜σ2 (b2 )⌝]
M ′ = ⟨γ∥(D1 ⋅ ιn1 ,a1 )∣i [n1 + 1 ÐÐÐÐÐÐÐ→ γ], (D2 ⋅ on2 ,a2 ,b2 )∣j [σ2 ]⟩@M0 ,
where ni ⊢D Di and σi ∶ ni → γ for i = 1, 2, σ1 (a1 ) = a = σ2 (a2 ), and σ2 (c2 ) = c. We want to
id
show that there exists a transition C ←Ð C ′ with C ′ & M ′ . There are exactly 10 cases here.
Firstly, we have the case where the third rule is applied to D1 [σ1 ], D2 [σ2 ]. Otherwise, one
could have used each of the three rules for &0 for each of D1 and D2 , yielding nine cases.
We start with the first case, and then treat the nine others.
● If the third rule is applied on D1 [σ1 ], D2 [σ2 ] (hence n1 = n2 = n and σ1 = σ2 = σ), then we
find C0 such that C0 & M0 and C = (ζ(D1 ) ∣ ζ(D2 ))[σ]@C0 . Then, we have a transition
id
(ζ(D1 ) ∣ ζ(D2 ))[σ]@C0 ←Ð [ζ(D1 )[σ1 ], ζ(D2 )[σ2 ]]@C0 .
Since the latter configuration is again related to M , this reduces to the case where the
second rule is applied for both D1 [σ1 ] and D2 [σ2 ].
● If the third rule is applied for D1 [σ1 ] and any of the three rules is applied for D2 [σ2 ],
then we find P , M1 , M2 , C0 , and n1 ⊢D D3 such that M1 has length 1 if the third rule is
also used for D2 and 0 otherwise, M0 = D3 [σ3 ]@M1 @M2 , P & D2 [σ2 ]@M1 , C0 & M2 , and
C = (ζ(D1 ) ∣ ζ(D3 ))[σ1 ]@P @C0 .
Thus, there is a transition
id
(ζ(D1 ) ∣ ζ(D3 ))[σ1 ]@P @C0 ←Ð [ζ(D1 )[σ1 ], ζ(D3 )[σ1 ]]@P @C0
which reduces this case to the one where the second rule is applied to D1 .
● If the third rule is applied for D2 [σ2 ] and any of the first two rules is applied for D1 , then
we find M1 , P , C0 , and n2 ⊢D D3 such that M0 = D3 [σ3 ]@M1 , P & D1 [σ1 ], C0 & M1 , and
C = (ζ(D2 ) ∣ ζ(D3 ))[σ2 ]@P @C0 . Again, we are reduced to the case where the second rule
is applied for D2 , using the transition
id
(ζ(D2 ) ∣ ζ(D3 ))[σ2 ]@P @C0 ←Ð [ζ(D2 )[σ2 ], ζ(D3 )[σ2 ]]@P @C0 .
In the remaining cases, we have C = P10 @P20 @C0 with Pi0 & Di [σi ] for i = 1, 2, and
C0 & M0 . We thus have four cases, described in Figure 10 just as we did in Figure 9.
● If the second rule is applied for both D1 and D2 , then there is a transition
id
C ←Ð C ′ = ζ((D1 ⋅ ιn1 ,a1 )∣i )[[σ1 , ⌈c⌉]]@ζ((D2 ⋅ on2 ,a2 ,c2 )∣j )[σ2 ]@M0
with C ′ & M ′ as desired.
74
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
D1
Case 1
Case 2
′
′
′
There exist γ1 ⊢ P1 = a1 (b).P1 +k1 R1 P10 = ζ(D1 )[σ1 ].
and h1 ∶ γ1′ −∼ n1 such that
P10 = P1 [σ1 ○ h1 ]
D1 = ⟦P1 ⟧h1
h1 (a′1 ) = a1
i ∈ ∣D1 ⋅ ιn1 ,a1 ∣
(D1 ⋅ ιn1 ,a1 )∣i = ⟦P1′ ⟧h′1
h1 +1
P2
where h′1 is γ1′ , b ÐÐ→ n1 + 1.
There exist γ2′ ⊢ P2 = a¯′2 ⟨c′ ⟩.P2′ +k2 R2 P20 = ζ(D2 )[σ2 ].
and h2 ∶ γ2′ −∼ n2 such that
h2 (a′2 ) = a2
P20 = P2 [σ2 ○ h2 ]
h2 (c′2 ) = c2
D2 = ⟦P2 ⟧h2
j ∈ ∣D2 ⋅ on2 ,a2 ,c2 ∣
(D2 ⋅ on2 ,a2 ,c2 )∣j = ⟦P2′ ⟧h2 .
Figure 10: Synchro, (RH) cases
● If we apply the second rule for D1 and the first rule for D2 , then there is a transition
id
C ←Ð C ′ = ζ((D1 ⋅ ιn1 ,a1 )∣i )[[σ1 , ⌈c⌉]]@P2′ [σ2 ○ h2 ]@C0
with C ′ & M ′ as desired.
● If the first rule is applied for both D1 and D2 , then we have a synchronisation between
P1 [σ1 ○ h1 ] and P2 [σ2 ○ h2 ]. In order to determine the result of this synchronisation we
need to choose a representative for P1 [σ1 ○ h1 ], i.e., pick a channel for what b becomes
after substitution. A reasonable choice here is γ + 1, we choose
P1 [σ1 ○ h1 ] = a(γ + 1).(P1′ [(σ1 + 1) ○ h′1 ]) +k1 R1 [σ1 ○ h1 ].
We thus have a transition to
C ′ = P1′ [(σ1 + 1) ○ h′1 ][γ + 1 ↦ c]@P2′ [σ2 ○ h2 ]@C0 .
But the diagram
n1 + 1
σ1 +1
γ+1
[γ+1↦c]
[σ1 ,⌜c⌝]
γ
commutes, so C ′ = P1′ [[σ1 , ⌜c⌝] ○ h′1 ]@P2′ [σ2 ○ h2 ]@C0 . Finally, we also know:
M ′ = ((D1 ⋅ ιn1 ,a1 )∣i )[[σ1 , ⌜c⌝]]@(D2 ⋅ on2 ,a2 ,c2 )∣j [σ2 ]@M0
= ⟦P1′ ⟧h′1 [[σ1 , ⌜c⌝]]@⟦P2′ ⟧h2 [σ2 ]@M0 ,
which entails C ′ & M ′ as desired.
AN INTENSIONALLY FULLY-ABSTRACT SHEAF MODEL FOR π
(EXPANDED VERSION)
75
● If we apply the first rule for D1 and the second rule for D2 , then, choosing the same
representative as before for P1 [σ1 ○ h1 ], there is a transition
id
C ←Ð P1′ [(σ1 + 1) ○ h′1 ][γ + 1 ↦ c]@ζ((D2 ⋅ on2 ,a2 ,c2 )∣j )[σ2 ]@C0 = C ′
satisfying C ′ & M ′ (for the same reason as in the last case).
Fork, (RH). We consider the case of a forking action, i.e., M = ⟨γ∥D[σ]⟩@M0 with n ⊢D D
and σ∶ n → γ, and we have a transition
id
⟨γ∥D[σ]⟩@M0 ←Ð ⟨γ∥(D ⋅ πnl )∣i [σ], (D ⋅ πnr )∣j [σ]⟩@M0
id
for some i and j. We want to show that there exists a transition C ←Ð C ′ with C ′ & M ′ .
We proceed by case analysis on the rule applied for D[σ] in the proof of C & M .
● If the first rule is applied, then we find C0 & M0 , γ ′ ⊢ P = P1 ∣ P2 , and h∶ γ ′ −∼ n such that
D = ⟦P ⟧h , C = P [σ ○ h]@C0 , (D ⋅ πnl )∣i = ⟦P1 ⟧h , and (D ⋅ πnr )∣j = ⟦P2 ⟧h .
Thus, there is a transition
id
P [σ ○ h]@C0 ←Ð P1 [σ ○ h]@P2 [σ ○ h]@C0 = C ′
with C ′ & M ′ as desired.
● If the second rule is applied, then we find C0 & M0 such that C = ζ(D)[σ]@C0 . Thus,
ζ(D)[σ] has the shape
τ.(ζ((D ⋅ πnl )∣i ∣ ζ((D ⋅ πnr )∣j )))[σ] +k R
(A.1)
so we have
ζ(D)[σ]@C0
id
←Ð (ζ((D ⋅ πnl )∣i ) ∣ ζ((D ⋅ πnr )∣j ))[σ]@C0 = C ′
with C ′ & M ′ (using the third rule) as desired.
● If the third rule is applied, then we find n ⊢D D′ , M1 , and C0 such that C0 & M1 ,
C = (ζ(D) ∣ ζ(D′ ))[σ]@C0 , and M0 = D′ [σ]@M1 . But then, as in the previous case,
ζ(D)[σ] has the shape (A.1) and we have transitions:
(ζ(D) ∣ ζ(D′ ))[σ]@C0
id
←Ð
ζ(D)[σ]@ζ(D′ )[σ]@C0
←Ð
=
ζ((D ⋅ πnl )∣i )[σ]@ζ((D ⋅ πnr )∣j )[σ]@ζ(D′ )[σ]@C0
C′
id 2
with C ′ & M ′ as desired.
Nu, (RH). We consider the case of a ν rule, i.e., one has a transition
σ+⌈a⌉
id
C = ⟨γ∥D[σ]⟩@M0 ←Ð ⟨γ∥(D ⋅ νn )∣i [n + 1 ÐÐÐ→ γ, a]⟩@M0 [γ ⊂ γ, a]
id
with n ⊢D D and σ∶ n → γ. We want to show that there exists a transition C ←Ð C ′ with
C ′ & M ′ . We again proceed by case analysis on the rule applied for D[σ].
● If the first rule is applied, then we find C0 & M0 , γ ′ ⊢ P = νa.P ′ +k R, and h∶ γ ′ −∼ n such
that C = P @C0 , D = ⟦P ⟧h and (D ⋅ νn )∣i = ⟦P ′ ⟧h+1 . There are thus transitions
id
h+1
σ+⌈a⌉
C = P [σ ○ h]@C0 ←Ð P ′ [γ ′ , a ÐÐ→ n + 1 ÐÐÐ→ γ, a]@C0 [γ ⊂ γ, a] = C ′
76
C. EBERHART, T. HIRSCHOWITZ, AND T. SEILLER
with C ′ & M ′ as desired.
● If the second rule is applied, then we find C0 & M0 such that C = ζ(D)[σ]@C0 , and there
is a transition
σ+⌈a⌉
id
ζ(D)[σ]@C0 ←Ð ζ((D ⋅ νn )∣i )[n + 1 ÐÐÐ→ γ, a]@C0 [γ ⊂ γ, a] = C ′
with C ′ & M ′ as desired.
● If the third rule is applied, then we find M1 , C0 , and n ⊢D D′ such that M0 = D′ [σ]@M1 ,
C0 & M1 , and C = (ζ(D) ∣ ζ(D′ ))[σ]@C0 . But then we have
(ζ(D) ∣ ζ(D′ ))[σ]@C0
id
←Ð ζ(D)[σ]@ζ(D′ )[σ]@C0
σ+⌈a⌉
id
←Ð ζ((D ⋅ νn )∣i )[n + 1 ÐÐÐ→ γ, a]@ζ(D′ )[σ][γ ⊂ γ, a]@C0 [γ ⊂ γ, a] = C ′
and C ′ & M ′ as desired.
Tick and Tau, (RH). We now consider the cases ♡ and τ , i.e., when one has a transition
ξ
⟨γ∥D[σ]⟩@M0 ←
Ð ⟨γ∥(D ⋅ ξn )∣i [σ]⟩@M0
where ξ ∈ {♡, τ }, with n ⊢D D and σ∶ n → γ. We want to show that there exists a transition
ξ
C ←
Ð C ′ with C ′ & M ′ . Again, we proceed by case analysis on the rule applied for D[σ].
● If the first rule is applied, we find C0 & M0 , γ ′ ⊢ P = ξ.P ′ +k R, and h∶ γ ′ −∼ n such that
C = P [σ ○ h]@C0 , D = ⟦P ⟧h [σ], and (D ⋅ ξn )∣i = ⟦P ′ ⟧h . But then we have
ξ
P [σ ○ h]@C0 ←
Ð P ′ [σ ○ h]@C0 & ⟦P ′ ⟧h [σ]@M0 ,
as expected.
● If the second rule is applied, then C = ζ(D)[σ]@C0 for some C0 & M0 and we have
ξ
ζ(D)[σ]@C0 ←
Ð ζ((D ⋅ ξn )∣i )[σ]@C0 & (D ⋅ ξn )∣i [σ]@M0 ,
as desired.
● Finally, if the third rule is applied, then we find C0 , M1 , and n ⊢D D′ such that C =
(ζ(D) ∣ ζ(D′ ))[σ]@C0 , M0 = D′ [σ]@M1 , and C0 & M1 . But then, we have
id
ξ
(ζ(D) ∣ ζ(D′ ))[σ]@C0 ←Ð ζ(D)[σ]@ζ(D′ )[σ]@C0 ←
Ð ζ((D ⋅ ξn )∣i )[σ]@ζ(D′ )[σ]@C0 = C ′
with C ′ & M ′ as desired.
| 6cs.PL
|
The Imprecisions of Precision Measures in Process Mining
Niek Tax∗, Xixi Lu, Natalia Sidorova, Dirk Fahland, Wil M.P. van der Aalst
Eindhoven University of Technology, P.O. Box 513, Eindhoven, The Netherlands
arXiv:1705.03303v2 [cs.DB] 16 May 2017
Abstract
In process mining, precision measures are used to quantify how much a process model overapproximates the behavior
seen in an event log. Although several measures have been proposed throughout the years, no research has been done
to validate whether these measures achieve the intended aim of quantifying over-approximation in a consistent way for
all models and logs. This paper fills this gap by postulating a number of axioms for quantifying precision consistently
for any log and any model. Further, we show through counter-examples that none of the existing measures consistently
quantifies precision.
Keywords: Process mining, Formal languages and automata, Petri nets, Design of algorithms
1. Introduction
Process mining [1] is a fast growing discipline that
is focused on the analysis of events logged during the
execution of a business process. Events contain information on what was done, by whom, for whom, where,
when, etc. Such event data are often readily available
from information systems such as ERP, CRM, or BPM
systems. Process discovery, which plays a prominent
role in process mining, is the task of automatically
generating a process model that accurately describes
a business process based on such event data. Many
process discovery techniques have been developed over
the last decade (e.g. [2, 3, 4, 5]), producing process
models in various forms, such as Petri nets [6], process
trees [7], YAWL models [8], and BPMN models [9].
The process model that is pursued by process discovery techniques ideally allows for all the behavior that
was observed in the event log (called fitness), while at
the same time it should not be too general by allowing
for much more behavior than what was seen in the event
log (called precision).
A range of measures have been proposed for quantifying precision [10, 11, 12, 13, 14]. However, to the best
of our knowledge, there is currently no work on verifying whether precision measures actually quantify what
∗ Corresponding
author
Email addresses: [email protected] (Niek Tax), [email protected]
(Xixi Lu), [email protected] (Natalia Sidorova),
[email protected] (Dirk Fahland), [email protected]
(Wil M.P. van der Aalst)
Preprint submitted to Elsevier
they are supposed to measure in a consistent manner.
Conceptually, the precision of a process model in the
context of an event log should be high when the model
allows for few traces not seen in the log, and it should be
low when it allows for many traces not seen in the log.
In this paper we propose a set of axioms that formulate
desired properties of precision measures and systematically validate whether these axioms hold for existing
precision measures.
In Section 2 we introduce basic notation and definitions. In Section 3 we formulate axioms for precision
measures. We then continue with Section 4, where we
describe existing precision measures in more detail and
validate the axioms for these measures. In Section 5 we
describe two contexts in which we are not able to define axioms for precision. In Section 6 we conclude this
paper and state several directions for future work.
2. Preliminaries
In this section we introduce concepts used in later
sections of this paper.
X = {a1 , a2 , . . . , an } denotes a finite set. P(X) denotes
the power set of X, i.e., the set of all possible subsets of
X. X ∗ denotes the set of all sequences over a set X and
σ = ha1 , a2 , . . . , an i denotes a sequence of length n, with
hi the empty sequence. X\Y denotes the set of elements
that are in set X but not in set Y, e.g., {a, b, c}\{a, c}={b}.
A multiset (or bag) over X is a function B : X→N which
we write as [aw1 1 , aw2 2 , . . . , awn n ], where for 1≤i≤n we have
May 18, 2017
ai ∈X and wi ∈N+ . The set of all bags over X is denoted
B(X).
In the context of process mining, we assume the set of
all process activities Σ to be given. Event logs consist
of sequences of events where each event represents a
process activity.
A partial function f ∈X9Y with domain dom( f ) can
be lifted to sequences over X using the following recursive definition: (1) f (hi) = hi; (2) for any σ∈X ∗ and
x ∈ X:
(
f (σ)
if x<dom( f ),
f (σ · hxi) =
f (σ) · h f (x)i if x∈dom( f ).
Definition 1 (Event, Trace, and Event Log). An
event e in an event log is the occurrence of an activity
e∈Σ. We call a sequence of events σ∈Σ∗ a trace. An
event log L∈B(Σ∗ ) is a finite multiset of traces.
Defining an initial and final markings allows to define
the language accepted by a Petri net as a set of finite
sequences of activities.
L=[ha, b, ci2 , hb, a, ci3 ] is an example event log over
process activities Σ={a, b, c}, consisting of 2 occurrences of trace ha, b, ci and three occurrences of trace
hb, a, ci.
Most precision measures have been implemented for
Petri nets, a process modeling formalism frequently
used in the context of process mining. A Petri net is
a directed bipartite graph consisting of places (depicted
as circles) and transitions (depicted as rectangles), connected by arcs. A transition describes an activity, while
places represent the enabling conditions of transitions.
Labels of transitions indicate the type of activity that
they represent. Unlabeled transitions (τ-transitions) represent invisible transitions (depicted as gray rectangles),
which are only used for routing purposes and are not
recorded in the event log.
Definition 3 (Accepting Petri Net). An
accepting
Petri net is a triplet APN=(N, m0 , MF), where N is
a labeled Petri net, m0 ∈B(P) is its initial marking,
and MF⊆B(P) is its set of possible final markings. A
sequence σ∈Σ∗ is a trace of an accepting Petri net APN
γ
if there exists a firing sequence m0 −→m f such that
m f ∈MF, γ∈T ∗ and `(γ)=σ.
The language L(APN) is the set of all its traces, i.e.,
γ
L(APN)={l(γ)|γ∈T ∗ ∧∃m f ∈MF m0 −→m f }, which can be
of infinite size when APN contains loops. Even though
we define language for accepting Petri nets, in theory
L(M) can be defined for any process model M with formal semantics. We denote the universe of process models as M. For each M∈M, L(M) is defined.
For an event log L, L̃={σ∈Σ∗ |L(σ)>0} is the trace
set of L. For example, for log L=[ha, b, ci2 , hb, a, ci3 ],
L̃={ha, b, cihb, a, ci}. For an event log L and a model M
we say that L is fitting on model M if L̃⊆L(M). Precision is related to the behavior that is allowed by a model
M that was not observed in the event log L, i.e., L(M)\L̃.
Definition 2 (Labeled Petri net). A labeled Petri net
N = hP, T, F, `i is a tuple where P is a finite set of
places, T is a finite set of transitions such that P∩T =∅,
F⊆(P×T )∪(T ×P) is a set of directed arcs, called
the flow relation, and `:T 9Σ is a partial labeling
function that assigns a label to a transition, or leaves it
unlabeled (the τ-transitions).
3. Axioms for Precision Metrics
The properties that are desired for precision measures
are not clearly defined in existing work, although they
are often discussed informally. Van der Aalst et al. [15],
describe the precision dimension as “Precision: measure determining whether the model prohibits behavior
very different from the behavior seen in the event log.
A model with low precision is underfitting.”. Vanden
Broucke et al. [13] describe precision as “precision (or:
appropriateness), i.e., the model’s ability to disallow unwanted behavior;”. Mũnoz-Gama and Carmona [12] describe it as “Precision: refers to overly general models,
preferring models with minimal behavior to represent as
closely as possible to the log.”. Buijs et al. [16] describe
precision as “... precision quantifies the fraction of the
behavior allowed by the model which is not seen in the
event log.”.
We write •n and n• for the input and output nodes of
n ∈ P ∪ T (according to F). A state of a Petri net is defined by its marking m∈B(P) being a multiset of places.
A marking is graphically denoted by putting m(p) tokens on each place p∈P. State changes occur through
transition firings. A transition t is enabled (can fire) in
a given marking m if each input place p∈•t contains
at least one token. Once t fires, one token is removed
from each input place p∈•t and one token is added
to each output place p0 ∈t•, leading to a new marking
m0 =m− •t + t•.
A firing of a transition t leading from marking m
t
to marking m0 is denoted as step m−→m0 . Steps are
lifted to sequences of firing enabled transitions, written
γ
m−→m0 and γ∈T ∗ is a firing sequence..
2
L̃
L(M1 )
L(M2 )
L̃1
=⇒ prec(L, M1 )≥prec(L, M2 )
L̃2
L(M)
=⇒ prec(L2 , M)≥prec(L1 , M)
(A2)
(A5)
Figure 1: Two of the five axioms for precision measures visualized with Euler diagrams.
A4 was stated before in an informal manner by Rozinat
and van der Aalst [11], who stated that precision should
be independent of structural properties of the model.
Adding fitting traces to a fitting log can only increase
the precision of a given model with respect to the log.
We consider precision to be a function prec(L, M)
which quantifies which part of the language of model
M is seen in event log L. Below we formalize the
desired properties of function prec through axioms to
consistently hold for any kind of model and any kind
of log.. Note that in the examples that we will show in
this paper all models M will be Petri nets, however the
formulated axioms are more general and apply to any
process model M∈M. Figure 1 visualizes two axioms
using Euler diagrams.
The first axiom states that precision is deterministic,
i.e., given a log and model always the same result is
returned.
Axiom A5. For model M and logs L1 and L2 ,
L̃1 ⊆L̃2 ⊆L(M) =⇒ prec(L2 , M)≥prec(L1 , M)
From A5 it follows as a corollary that precision is
maximal when the log contains all the traces allowed
by the model, and minimal when it contains no traces
allowed by the model.
In the coming sections we will validate whether these
axioms hold for several precision measures. Some articles that introduce precision measures explicitly mention that the measure is intended to be used only with
a certain subclass of Petri nets. An example of such a
subclass of Petri nets are bounded Petri nets, which have
the restriction that all places most have a finite number
of tokens in all reachable markings. When an article
that introduces a precision measure states an explicit assumption on the subclass of Petri nets, then we only validate the axioms on this subclass of Petri nets. When no
explicit assumption on a subclass of Petri nets is stated,
we assume that the precision measure is intended for
Petri nets in general.
Axiom A1. A precision measure is a function prec :
B(Σ∗ ) × M → R, i.e., it is deterministic.
Existing precision measures normalize R to a [0, 1]interval.
The second axiom formulates the conceptual description of precision more formally: if a process model M2
allows for more behavior not seen in a log L than another model M1 does, then M2 should have a lower precision than M1 regarding L.
Axiom A2. For models M1 and M2 and a log L,
L̃⊆L(M1 )⊆L(M2 ) =⇒ prec(L, M1 )≥prec(L, M2 )
Note that A2 does allow L̃⊆L(M1 )⊂L(M2 ) with
prec(L, M1 )=prec(L, M2 ).
Ideally, since L(M1 ) is
smaller than L(M2 ) we would like to see a higher precision for M1 , but this requirement might be too strict.
However, for a process model M with L̃⊆L(M), we
would like the precision of M on L to be higher than
the precision of M on any flower model (i.e., a model
that allows for all behavior over its activities) on log L.
4. Precision Metrics
In this section we give an overview of the precision
measures that have been developed in the process mining field, and validate the axioms for precision measures
introduced in Section 3 for each of those measures.
4.1. Soundness
Axiom A3. For models M1 and M2 and a log L,
Greco et al. [10] were the first to propose a precision
L(M1 )⊂P(Σ∗ )∧L(M2 )=P(Σ∗ ) =⇒ prec(L, M1 )>prec(L, M2 ) measure, defining it as the number of unique executions
of the process that were seen in the event log divided by
The precision of a log on two language equivalent
the number of unique paths through the process model.
models should be equal, i.e., precision should not deThis measure is not usable in practice, because it is zero
pend on the model structure.
when the process model allows for an infinite number
Axiom A4. For models M1 and M2 and a log L,
of paths through the model. Any process model havL(M1 )=L(M2 ) =⇒ prec(L, M1 )=prec(L, M2 )
ing a loop has a precision of 0. More recent precision
3
measures are capable of calculating the precision of a
model for an event log even when the models allows for
infinite behavior.
4.2. Behavioral Appropriateness
c
b
d
Figure 2: Model M.
Rozinat and Van der Aalst [11] proposed the simple
behavioral appropriateness precision measure, which
looks at the average number of enabled transitions during replay. The authors observed themselves that simple
behavioral appropriateness is dependent on the structure
of the model, and not solely dependent on the behavior
that it allows, therefore A4 does not hold for this measure. Furthermore, for a process model that contains
silent transitions or duplicate labels it is possible that
a given trace can be replayed on this model in multiple ways, where the average number of enabled transitions can depend on the chosen replay path through the
model. This replay path through the model is chosen arbitrarily from the possible ways in which the trace can
be replayed. This shows that A1 does not hold for simple behavioral appropriateness, as it is not deterministic.
In the same paper, Rozinat and van der Aalst [11]
propose advanced behavioral appropriateness, which
is independent of the model structure. Advanced behavioral appropriateness calculates the sets S F ⊆Σ×Σ of
pairs of activities that sometimes, but not always, follow
each other. Likewise set S P ⊆Σ×Σ is calculated as the
set of activities that sometimes, but not always, precede
each other. S FL and S PL denote the sometimes-follows
and sometimes-precedes relations on the log, and S FM
and S PM denotes the sometimes-follows and sometimesprecedes relations according to the model. However,
to calculate S FM and S PM , exhaustive exploration of the
state space of the model is required, prohibiting the
application of this measure for large models or highly
concurrent models, where the state-space explosion
problem arises. Advanced behavioral appropriateness
|S PL ∩S PM |
|S FL ∩S FM |
+ 2·|S
precision is defined as a0b =( 2·|S
M|
M | ). Because
F
a
End place, and (iii) every node must be on some path
from Start to End. The soundness property additionally
require that each transition can be potentially executed,
and that the process can always terminate properly, i.e.,
finish with only one token in the End place.
Consider model M of Figure 2, which belongs to
the class of sound WF-nets, and any log L such that
L̃⊆L(M). The loop in model M causes S FM and S PM
to contain all pairs of activities of Σ. Therefore, |S FM |
0
and |S PM | are identical to the sometimes relations |S FM |
0
and |S PM | of any model M 0 with L(M 0 )=P(Σ∗ ), leading
to prec(L, M)=prec(L, M 0 ). As L(M)⊂P(Σ∗ ), this is in
conflict with A3.
4.3. Escaping Edges Precision
Escaping Edges Precision (ETC) [12] calculates precision by constructing a prefix automaton, which consists of one state per unique prefix of the event log.
Figure 3b shows an example prefix automaton for an
event log L = [ha, ci, ha, di]. For each state in the prefix automaton it is then determined which activities are
allowed as next activities by the process model. Activities that are allowed as next activities for some prefix
but that are never observed in the event log after this
prefix are referred to as escaping edges.
In later work [18, 19], alignments [20] are used to
calculate the prefix automaton on the aligned event log
instead of the original event log, making the precision
measure robust to non-fitting traces, i.e., traces that
are not in the language of the model. For a trace σ
from a log L that is fitting on an accepting Petri net
APN, alignments [20] give a sequence of transition
γ
firings γ∈T ∗ such that m0 −→m f with m0 the initial
marking and m f a final marking of APN and `(γ)=σ.
Note that for a given trace σ and model, multiple
possible alignments can exist. For non-fitting traces,
alignments search for a firing sequence γ∈T ∗ such that
`(γ) is as close as possible to σ. Adriansyah et al. [18]
describe two versions of the alignment-based escaping
edges precision: one-align ETC, which calculates
the precision based on one optimal alignment of log
and model, and all-align ETC, which calculates the
precision based on all optimal alignments between log
and model. In practice, it is often computationally
infeasible to calculate all optimal alignments. A later
P
S FM and S PM are obtained through exhaustive exploration
of the state space of the model, it is easy to see that they
depend only on the behavior of the model and not on its
structure, therefore A4 holds. A problem with advanced
behavioral appropriateness occurs for deterministic
models, where |S PM |=|S FM |=0, leading to undefined precision. This shows that advanced behavioral appropriateness is a partial function, which is in conflict with A1.
Rozinat and van der Aalst [11] state that simple behavioral appropriateness and advanced behavioral appropriateness assume the Petri net to be in the class of
sound workflow (WF) nets [17]. A WF-net requires the
Petri net to have (i) a single Start place, (ii) a single
4
a
c
2
b
3
d 1
2
a
b
c 1
d
b
(b)
(a)
a
a
c
b
a
c
d 1
d
d
d
d
d
b a
b
a
b
a
b
a
b
a
c
3 1 1 1 1 1 1 1 1 1 1 1
a
c
c
c
b
c
1
(b)
(a)
b
(c)
c
Figure 3: (a) Model M, and the alignment automata on
Model M for (b) log L1 =[ha, ci, ha, di], and for (c) log
L2 =[ha, ci, ha, di, ha, b, a, b, a, b, a, b, a, b, a, ci].
Red arcs correspond to escaping edges.
a
(c)
Figure 4: (a) The flower-model, which allows for all behavior over its
set of activities, (b) an alternative representation of the flower-model,
and (c) a constrained model which always starts with activity a.
precision measure, representative-align ETC [21],
calculates the escaping edges based on a sample of
optimal alignments, and can therefore be seen as
a trade-off between the computational efficiency of
one-align ETC and the reliability of all-align ETC. The
papers on ETC precision and its variants do not state an
assumption on a subclass of Petri nets. ETC, one-align
ETC, all-align, and representative-align ETC precision
are all implemented in the package ETConformance1
as part of the process mining framework ProM [22].
The one optimal alignment that is used by one-align
ETC is chosen arbitrarily from the set of optimal
alignments of a log on a model. However, different
optimal alignments result in different prefix automata,
which can potentially lead to different precision values.
This shows that A1 does not hold for one-align ETC.
Consider log L1 =[ha, ci, ha, di], log L2 =[ha, ci, ha, di,
ha, ci, ha, b, a, b, a, b, a, b, a, b, a, ci] and model M be the
Petri net of Figure 3a. Note that L̃1 ⊂L̃2 . The alignment
automata generated for the calculation of prec(L1 , M)
and the calculation of prec(L2 , M) are shown in Figure
3b and Figure 3c. The circles represent the states of the
automaton, and the arrows the transitions. The numbers in the states represent the weights of the states for
the precision calculation, i.e., the number of times that
states are visited in the alignment of log L on model
M [18]. In an alternative definition of one-align ETC
[19] the states are weighted by the number of times that
events occurred while being in this state according to
the alignment of L on M, instead of the number of times
that this state was reached according to this alignment.
Figure 3b shows that the initial state was visited twice,
activity a occurred twice at the start in log L1 , resulting
in a state from which activities b, d, and c were enabled.
From this state, activities c and d were seen once, and
activity b was never seen, thus it is an escaping edge.
Escaping edges precision is then the (weighted) average
ratio of non-escaping edges from all outgoing edges,
where states are weighted by the number of times that
they are visited. Counting the weighted number of nonescaping edges in the numerator and the weighted total
number of edges in the denominator in our example, we
6
find prec(L1 , M)= 2×1+2×2+1×0+1×0
2×1+2×3+1×0+1×0 = 8 =0.75. One-align
ETC results in the following precision values for M on
L1 and L2 : prec(L1 , M)=0.75 and prec(L2 , M)=0.7143.
This shows that A5 does not hold for one-align ETC.
By comparing the automata of Figures 3b and 3c it becomes clear that the single trace that is in L2 but not
in L1 brings the model to many states with three escaping edges, reducing precision. The prefix automata and
the precision calculations for M on logs L1 and L2 were
performed manually following the procedure from the
paper and validation using the ETConformance plugin
in ProM.
Now consider log L = [ha, b, ci], and the three
Petri nets M1 , M2 , M3 in Figures 4a, 4b, and 4c
respectively. Note that M1 and M2 are language
equivalent, as L(M1 )=L(M2 )={a, b, c}∗ . M3 is more
behaviorally constrained than M1 and M2 , since all its
traces start with activity a. The one-align precision
of M1 , M2 , M3 on L are: prec(L, M1 )=0.3333,
prec(L, M2 )=0.5238,
and
prec(L, M3 )=0.4444.
L(M3 )⊂L(M1 ), but prec(L, M3 )>prec(L, M1 ), implying
that A2 does not hold for one-align ETC. Furthermore,
L(M1 )=L(M2 ), but prec(L, M1 ),prec(L, M2 ), implying
that A4 does not hold for one-align ETC.
Analyzing the ETConformance plugin in ProM we
1 https://svn.win.tue.nl/trac/prom/browser/Packages/ETConformance
5
found that the prefix automaton generated for one-align
precision for calculation of prec(L, M1 ) results in 6
states, belonging to 3 firings of observable transitions
and 2 firings of τ-transitions. In 3 of the 6 states, which
correspond to M1 being in the initial marking, there
are 4 possible next activities according to the model, of
which only one is observed for that prefix. Furthermore,
it shows that the alignment automaton generated for L
and M1 consists of 6 states, the automaton for L and M2
consists of 12 states, and the automaton for L and M3
consists of 5 states. This shows that the silent (τ) transitions in M2 generate additional states in the alignment
automaton, leading to a higher precision value.
Computing the precision of M1 or of M2 on L did
not finish with all-align ETC and representative-align
ETC after 8 hours of computation time. The long computation time of all-align ETC and representative-align
ETC on models where many optimal alignments exist
is a known issue which hinders the application of those
measures in practice.
c
a
d
f
g
b
e
Figure 5: A process model M1 and its behaviorally restricted variant
M2 which includes the dotted places and arcs.
of events w that directly precede e, then calculating all
subsequences of events in the log that exactly match w,
and finally negative events are identified by calculating
which events have never occurred in the log directly
after any subsequence matching w. This procedure is
repeated for different windows sizes, and the resulting
negative event are weighted by window size.
To induce the events that could not have happened
after e.g. trace prefix σ0 = ha, c, c, d, e, c, d, e, ei, the
method to induce weighted negative events described in
[13] searches for subsequences of events in the log that
are identical to the latest k events of σ0 in the event log.
All the activities that have never succeeded such subsequences are considered to be negative events, furthermore, the confidence of these negative events is based
on the length k of those matching subsequences.
Negative event based precision measures, with the
different methods for negative event induction, are implemented in the ProM package NEConformance2 . In
this paper we evaluate the precision measure that uses
weighted negative events [13], which is the most recent approach to induce negative events and the recommended approach for measuring precision [13]. No
assumption on a subclass of Petri nets is stated in the
papers on negative event precision.
Consider models M1 and M2 of Figure 5 respectively
excluding and including the arcs and places indicated in
dotted lines. L(M2 )⊂L(M1 ), since M2 contains a long
term dependency between activities a and f and between activities b and g, which M1 does not have. Consider an event log L which consists of 10 traces from
M2 , leading to L being fitting on both M1 and M2 . We
found the negative event precision of M1 and M2 on the
same L to be non-deterministic, resulting in slightly different values every time that it is calculated. This shows
that A1 does not hold for negative event precision.
Because negative event precision is nondeterministic, we calculated the precision of M1
4.4. Negative Event Precision
Goedertier et al. [2] proposed a method to induce
negative events, i.e., sets of events that were prevented
from taking place. Negative events are induced for each
position in the event log, i.e., for each event e in each
trace of the log a set of events is induced that could
not have taken place instead of event e. De Weerdt et
al. [23] proposed a precision measure based on negative events, behavioral precision (pB ), which is closely
linked to how precision is defined in the area of data
mining. Negative event precision regards a process
model as a binary classifier that determines whether a
certain event can take place given a certain prefix, and
then evaluates the precision of this classifier in data mining terms taking the induced negative events as ground
truth. For a given trace prefix, true positive (TP) events
are defined as events that are possible according to both
the process model (i.e., a transition labeled with this
event is enabled) and log (i.e., this event is not a negative event). False positive events (FP) are negative
events induced for a given prefix that were possible according to the model. Behavioral precision is defined
TP
, which is in accordance to the definition
as pB = T P+FP
of precision in the data mining field. In later work [24]
induction of artificial negative events has been refined
based on frequent temporal patterns which are mined
from the event log. Finally, weighted artificial events,
where negative events are weighted according to their
confidence, are proposed in [13].
Weighted behavioral precision induces negative
events for an event e in the log by taking a window
2 http://processmining.be/neconformance/
6
and M2 on L both 20 times. The highest precision
found in 20 repetitions for M1 is 0.4876, while the
lowest precision found for M2 is 0.4545, showing that
the non-determinism has the effect that A2 does not
hold for negative event precision. We found an average
value of 0.4744 with a standard deviation of 0.0090 for
the precision of M1 on L and an average value of 0.4640
with a standard deviation of 0.0072 for the precision
of M2 on L. This shows that also in terms of average
precision value A2 does not hold.
To test whether the difference in mean precision between M1 and M2 is due to chance alone we formulate
a null hypothesis:
H0 : The average negative event precision of M2 on L
is higher than or equal to the average negative event
precision of M1 on L.
Testing this null hypothesis with a one-tailed Welch
t-test [25] we found a p-value of 0.0001801, indicating
that we can reject the null hypothesis with significance
level 0.01. This shows that, with statistical significance,
the precision of M1 on L is higher than the precision of
M2 on L, which is in disagreement with A2.
To see why A2 does not hold for negative event precision, consider the negative event inducing procure being
applied to trace prefix σ0 = ha, c, c, d, e, c, d, e, ei from
log L. Petri net M2 generates many different traces because of the parallel length-one-loops on activities c, d
and e, which allows for any sequence of any length over
these activities. Therefore, the matching subsequences
of σ0 in the log generated from M2 are the subsequences
that by chance ended in the same behavior over c, d,
and e. Because the sequences of c, d and e events can
be long and diverse, activity a and b are unlikely to
be present in the matching subsequences, which makes
it unlikely that the procedure can induce the negative
event g for σ0 . Because the negative events that reflect
the constraint that M2 introduces compared to M1 cannot be induced from the log, negative event precision is
not able to recognize that M2 is more precise than M1 .
a
b
b
(a)
b
a
b
a
(b)
Figure 6: (a) A model with a length-one-loop, (b) the same model
with the loop unrolled up to two executions.
behavior that was allowed both in log and model automaton. It then iterates over the states of the model automaton, and calculates over the share of transitions of
this state that is also possible in the corresponding state
in the conjunction automaton. It defines precision as the
average of this share over the model automaton states.
PCC precision is implemented in the ProM package
ProjectedRecallAndPrecision3 . PCC precision assumes
the Petri net to be of the class of bounded Petri nets, i.e.,
the Petri net for which precision is calculated must have
a finite number of tokens in every place for all reachable
markings.
Consider log L = [ha, bi] and Petri nets M1 and M2 of
Figures 6a and 6b respectively. M1 starts with a lengthone-loop on activity a, followed by activity b. M2 unrolls the length-one-loop on activity a of M1 to at most
two executions, thereby limiting the behavior as it only
allows at most two executions of activity a. It is easy to
see that M1 and M2 both belong to the class of bounded
Petri nets, as in both models each place can have at most
one token. For this log and these models, PCC precision
results in prec(L, M1 )=0.6, and prec(L, M2 )=0.5. However, since L(M2 )⊂L(M1 ), A2 states that the precision
of M2 for fitting log L should be higher or equal to its
precision of M1 . This shows that A2 does not hold for
PCC precision.
This drop in precision is an effect of the additional
states that are created in the model DFA as an effect of
unrolling the length-one-loop. The model DFA created
from Petri net M2 (Figure 6b) for example contains a
state s that is reached after firing ha, ai. This state however is never reached based on event log L, which only
contains a trace ha, bi, which has the effect that none
of the enabled transitions from state s were observed in
the log, bringing down the precision. In the DFA generated from Petri net M1 (Figure 6a), this state s is merged
with the state that one reaches after observing a single a
4.5. Projected Conformance Checking
Projected Conformance Checking (PCC) precision
was developed by Leemans et al. [14] as a computationally efficient precision measure that scales to event
logs with billions of events. PCC precision projects both
event log and model on all subsets of activities of size
k, and generates minimal deterministic finite automata
(DFA) for the behavior over these subsets of activities
in the log (i.e., log automaton) and for the behavior over
these events allowed by the model (i.e., model automaton). Based on the log automaton and model automaton
it then builds a conjunction automaton which allows the
3 https://svn.win.tue.nl/trac/prom/browser/
Packages/ProjectedRecallAndPrecision/
7
a
b
L(M1 )
L̃
L(M2 )
L̃
L(M1 )
L(M2 )
c
(a)
(b)
Figure 7: A flower model over activities a, b and c.
Figure 8: Two situations in which the desired properties of precision
measures are unclear: (a) two models on which log L fits, with both
models allowing for behavior that is not allowed by the other model,
and (b) two models and a log on which the models do not fit.
Table 1: Overview of the precision axioms and whether they hold for
each precision measure.
Metric
Simple behavioral appropriateness
Advanced behavioral appropriateness
One-align ETC
Negative Event Precision
PCC precision
A1
7
7
7
7
A2
A3
A4
A5
7
7
3
7
7
7
7
7
5. Contexts With Unclear Requirements for Precision Metrics
The axioms introduced in Section 3 can be regarded
as necessary conditions for precision measures, but they
leave precision unspecified in some contexts. Figure
8a shows a situation in which L̃⊆L(M1 ), L̃⊆L(M2 ), but
L(M1 )\L(M2 ),∅ and L(M2 )\L(M1 ),∅. In this setting,
both M1 and M2 allow for (a possibly infinite amount of)
different behavior that was not seen in L. Precision measures deal with this situation by quantifying the amount
of behavior of M1 and M2 . However, there are no obvious formal properties telling how the precision of M1
and M2 on L should relate.
Furthermore, all axioms define desired properties of
precision measures when the event log L fits the behavior of the model M, i.e., L̃ ⊆ L(M). In practice, process
discovery techniques will return process models with
fitness below 1, i.e., there exists σ∈L : σ<L(M). The
discovery algorithm may deliberately abstract from infrequent behavior. In this paper we do not formulate axioms for precision measures in the context of event logs
that do not fit the process model, since we feel that there
is not enough agreement in the process mining community on how a precision measure should behave in this
context. Figure 8b shows an Euler diagram of a log L
and two models M1 and M2 such that L(M1 )⊂L(M2 ) and
L̃*L(M1 ), which is a non-fitting equivalent of A2. A2
prescribes prec(L, M1 )≥prec(L, M2 ), however, when the
log does not fit the models, the behavior that fits M2
but not M1 , (L̃\L(M1 ))∩L(M2 ), makes it unclear how
the precision of M1 and M2 should relate. Furthermore,
even when (L̃\L(M1 ))∩L(M2 )=∅, it can be the case that
the behavior in L that does not fit the models is behaviorally similar to behavior of M2 .
7
event, as future behavior allowed by the model does not
depend on the number of a-events seen.
Consider
Petri
net
M
of
Figure
7,
and
event
logs
L1 =[hb, a, ci, ha, a, ci],
and
L2 =[hb, a, ci, ha, a, ci, ha, b, b, b, b, b, b, b, b, b, b, b, b, b,
b, bi, hb, a, a, a, a, a, a, a, a, a, a, a, a, a, a, ai]. The single place of M is bounded to one token, therefore M
belongs to the class of bounded Petri nets. It is easy to
see that L̃1 ⊂L̃2 , since the first two traces of log L2 form
log L1 . PCC precision results in prec(L1 , M)=0.3125
and prec(L2 , M)=0.2727, violating A5. The two traces
of L2 that are not in L1 are very long traces to the traces
that are in L1 , leading to additional states in the log
automaton and the conjunction automaton. The additional states of the conjunction automaton have a low
precision of 41 , since for each state the model allows for
four options (firing activity a, b, c, or stopping), while
only one is seen in the log. Therefore, if we would
expand trace hb, a, a, a, a, a, a, a, a, a, a, a, a, a, a, ai
with more events of activity a, then prec(L2 , M) would
approach 41 .
4.6. Overview of Precision Metric Properties
We formulated five axioms that describe desirable
properties for precision measures. Table 1 gives an
overview of that axioms that we showed that do hold
(3) and that do not hold (7) for each precision measure. We found that none of the existing precision measures fulfills all five axioms. Empty cells in the table are
currently unknown, and no formal proof nor a counter
example has been found that proves or disproves the axiom for the respective precision measure.
6. Conclusions & Future Work
In this paper provides a set of minimal requirements
for precision measures through axioms. We validated
these axioms for existing measures. Surprisingly, we
8
discovered that none of the existing precision measures
fulfills all formulated requirements.
In future work, we would like fill the empty cells
of Table 1 and get a complete overview of the axioms
that hold for each precision measure. Furthermore, we
would like to use the insights learned from evaluating
the axioms on the measures to either repair one of the
existing measures or come up with a completely new
measure that fulfills all five axioms.
[14] S. J. J. Leemans, D. Fahland, W. M. P. van der Aalst, Scalable
process discovery and conformance checking, Software & Systems Modeling (2016) 1–33.
[15] W. M. P. van der Aalst, A. Adriansyah, A. K. A. De Medeiros,
F. Arcieri, T. Baier, T. Blickle, J. C. Bose, P. van den Brand,
R. Brandtjen, J. Buijs, et al., Process mining manifesto, in:
International Conference on Business Process Management,
Springer, 2011, pp. 169–194.
[16] J. C. A. M. Buijs, B. F. van Dongen, W. M. P. van der Aalst,
Quality dimensions in process discovery: The importance of
fitness, precision, generalization and simplicity, International
Journal of Cooperative Information Systems 23 (01) (2014)
1440001.
[17] W. M. P. van der Aalst, Verification of workflow nets, in: International Conference on Application and Theory of Petri Nets,
Springer, 1997, pp. 407–426.
[18] A. Adriansyah, J. Munoz-Gama, J. Carmona, B. F. van Dongen,
W. M. P. van der Aalst, Alignment based precision checking,
in: International Conference on Business Process Management,
Springer, 2012, pp. 137–149.
[19] W. M. P. Van der Aalst, A. Adriansyah, B. F. van Dongen,
Replaying history on process models for conformance checking and performance analysis, Wiley Interdisciplinary Reviews:
Data Mining and Knowledge Discovery 2 (2) (2012) 182–192.
[20] A. Adriansyah, B. F. van Dongen, W. M. P. van der Aalst, Conformance checking using cost-based fitness analysis, in: Proceedings of the 15th IEEE International Enterprise Distributed
Object Computing Conference, IEEE, 2011, pp. 55–64.
[21] A. Adriansyah, J. Munoz-Gama, J. Carmona, B. F. van Dongen,
W. M. P. van der Aalst, Measuring precision of modeled behavior, Information Systems and e-Business Management 13 (1)
(2015) 37–67.
[22] B. F. van Dongen, A. K. A. de Medeiros, H. M. W. Verbeek, A. J.
M. M. Weijters, W. M. P. van der Aalst, The ProM framework:
A new era in process mining tool support, in: International Conference on Application and Theory of Petri Nets, Springer, 2005,
pp. 444–454.
[23] J. De Weerdt, M. De Backer, J. Vanthienen, B. Baesens, A robust F-measure for evaluating discovered process models, in:
Proceedings of the IEEE Symposium on Computational Intelligence and Data Mining, IEEE, 2011, pp. 148–155.
[24] S. K. L. M. vanden Broucke, J. De Weerdt, B. Baesens, J. Vanthienen, Improved artificial negative event generation to enhance
process event logs, in: International Conference on Advanced
Information Systems Engineering, Springer, 2012, pp. 254–269.
[25] B. L. Welch, The generalization of student’s’ problem when
several different population variances are involved, Biometrika
34 (1/2) (1947) 28–35.
[26] N. Tax, Validation of precision measures - event logs and
process models, Eindhoven University of Technology. Dataset,
http://dx.doi.org/10.4121/uuid:991753f7-a2404ba6-a8a8-67174a08c51b (2017).
Reproducibility. The event logs and process models
that are used as part of a counterexample for a combination of an axiom and a precision measure can be found
at [26].
References
[1] W. M. P. van der Aalst, Process mining: Data science in action,
Springer, 2016.
[2] S. Goedertier, D. Martens, J. Vanthienen, B. Baesens, Robust
process discovery with artificial negative events, Journal of Machine Learning Research 10 (2009) 1305–1340.
[3] S. J. J. Leemans, D. Fahland, W. M. P. van der Aalst, Discovering block-structured process models from event logs-a constructive approach, in: International Conference on Applications and
Theory of Petri Nets and Concurrency, Springer Berlin Heidelberg, 2013, pp. 311–329.
[4] R. Conforti, M. Dumas, L. Garcı́a-Bañuelos, M. La Rosa,
BPMN miner: automated discovery of BPMN process models
with hierarchical structure, Information Systems 56 (2016) 284–
303.
[5] S. J. van Zelst, B. F. van Dongen, W. M. P. van der Aalst, Avoiding over-fitting in ILP-based process discovery, in: International
Conference on Business Process Management, Springer International Publishing, 2015, pp. 163–171.
[6] T. Murata, Petri nets: Properties, analysis and applications, Proceedings of the IEEE 77 (4) (1989) 541–580.
[7] J. C. A. M. Buijs, B. F. van Dongen, W. M. P. van der Aalst,
A genetic algorithm for discovering process trees, in: Proceedings of the 2012 IEEE Congress on Evolutionary Computation,
IEEE, 2012, pp. 1–8.
[8] W. M. P. van der Aalst, A. H. M. ter Hofstede, YAWL: yet another workflow language, Information systems 30 (4) (2005)
245–275.
[9] Object Management Group, Notation (BPMN) version 2.0,
OMG Specification.
[10] G. Greco, A. Guzzo, L. Pontieri, D. Sacca, Discovering expressive process models by clustering log traces, IEEE Transactions
on Knowledge and Data Engineering 18 (8) (2006) 1010–1027.
[11] A. Rozinat, W. M. P. van der Aalst, Conformance checking of
processes based on monitoring real behavior, Information Systems 33 (1) (2008) 64–95.
[12] J. Muñoz-Gama, J. Carmona, A fresh look at precision in process conformance, in: International Conference on Business
Process Management, Springer, 2010, pp. 211–226.
[13] S. K. L. M. vanden Broucke, J. De Weerdt, J. Vanthienen,
B. Baesens, Determining process model precision and generalization with weighted artificial negative events, IEEE Transactions on Knowledge and Data Engineering 26 (8) (2014) 1877–
1889.
9
| 2cs.AI
|
Detection of parallel steps in programs with arrays
R. Nuriyev
([email protected])
The problem of detecting of information and logically independent (DILD) steps in programs is a
key for equivalent program transformations. Here we are considering the problem of
independence of loop iterations – the concentration of massive data processing and hence the
most challenge construction for parallelizing. We introduced a separated form of loops when
loop‟s body is a sequence of procedures each of them are used array‟s elements selected in a
previous procedure. We prove that any loop may be algorithmically represented in this form and
number of such procedures is invariant. We show that for this form of loop the steps connections
are determined with some integer equations and hence the independence problem is
algorithmically unsolvable if index expressions are more complex than cubical. . We suggest a
modification of index semantics that made connection equations trivial and loops iterations can be
executed in parallel.
We are considering not only algorithmic fullness of the programming language but also
its data fullness when selection functions of elements of arrays are algorithmically full class too.
These two features are independent. We suggest a modification of index semantics that made
connection equations trivial and loops iterations can be executed in parallel. This modified
language is a full in both senses.
We consider a DILD problem for programming languages with arrays and focusing on
the syntax constructions determining the DILDs.
Transformations of programs, and parallelizing in particular, are based on fact that
execution order for given start values can be changed without changing result. For program with
single variables for fragment
x=f(u) // step 1
y=g(x) // step 2
we may say that step 2 informational depends on step 1.
But for the case of indexed variable
x[i] =f(u) // step 3
y=g(x[j]) // step 4
the information dependency takes a place only if values i and j are the same.
Let‟s consider two program fragments, structured and unstructured in Dijkstra‟s terms,
when each loop has one entrance and one exit points as in figure 1:
Figure 1
For the first fragment to identify an execution step for operator q we may use 2dementional vector (p, k): p is a number of work out iterations of the upper loop and k is a
number of lower loop iterations in the iteration p of upper loop. Step q3,7 means that we are
talking about such execution of the operator q when the lower loop made 7 iterations in a third
iteration of the upper loop.
To identify a step for the second fragment we have to use variable length vector q(i1,i2,…)
meaning that we are talking about point of execution when upper loop executes i1 times, then
lower executes i2 times, then upper executes i3 times and then inner executes i4 times and so on.
Good news is that any program can be algorithmically structured (Dijkstra‟s goto
elimination problem) so that only nested loops may be used.
In section 3 we will extend the canonization result to recursion programs.
So we will consider from now and up only structured programs: programs in which
each repetition steps are formed only with nested loops.
We will extend this idea by introducing forced syntax principle approach to modify of
the language that syntactically implicit properties we need to solve problem in hands are
encapsulated in a (new if necessary) syntactically explicit constructions.
Applying this principle to the DILD problem with indexed variables we will introduce a
representation of a loop body as a chain of special subprograms (last one is called kernel, others
are called controllers) which are responsible for calculation values of indexes for variables of
next levels controllers or kernel but not for themselves. Such organized loops are called
separated forms of loops. Other word, in separated loops the two loop body‟s functions of
selecting data from some sets or arrays and processing that data are separated. For programs with
separated loops we have enough syntactical objects to be able to define a condition if some set of
iterations can be run in parallel.
A separated loop for processing connected list has two level controllers. First level
controller selects node and the second one selects pointer to the next element.
In this paper we'll prove that for FORTRAN like languages (or containing FORTRAN)
the class of algorithmically parallelizing programs can't be remarkably bigger than the class with
linear index expression. So to create language with algorithmically parallelizing programs we
have to change some fundamental constructions of the languages and our goal is to point to these
constructions.
1. Basic definitions
We will consider schemas of programs in which functions and predicates are symbols
and they get interpretation (implementations) together with start data. But also we will allow
some of them to have a fixed interpretation.
Let‟s A, X, F be finite alphabets without common elements and A*, X*, F* are a sets of
strings in an appropriate alphabets.
Def. 1. 1. We call a memory M a countable set of cells - two types of elements: x and
a[w1, .., wm], where x belong to the set X* and is called single cell, a belongs to A* and is
called array of cells, w1,…, wn belong to X* and is called indexes of the cell.
A cell may get (be assigned to) a value from set and keep it until next changing by a
program operator. Cells without value are called empty. Each memory element has no more than
one value and the value of element x will be denoted <x>, so <x>
.
To keep this notation more close to programming languages let‟s suppose that one array
name can‟t have different dimensions. Pairs of sets X, A and A, do not have common
elements.
Def. 1.2. Variables of a program schemas are elements of X* and expressions
a[K1(z1),…, Kn(zn)], where a A* is called array name, K1, …, Kn F are called index functions,
z1, …, zn are program variables from X*. Variables of a first type are called simple and of a
second type – indexed variables.
For X={x,y}, A={a,b}, expressions x, y, xx, xyx are a simple variables, expression
ba[xyx, yx] is an example of indexed variables, expression aa[ba[x]] is not a variable –
expression in [ ] parenthesis has to contain only simple variables.
Def. 1.3. Schema of program is a 5D vector (Opers, F, VarX, VarA, subP) where
VarX is a set of single variables, VarA is a set of array variables, F is a set of function an
predicate symbols, subP is a set of (sub) schemas, also called procedures, Opers is a finite set
of program instructions – strings of one of the forms:
a. l1: x0=g(x1, …, xn) then l2; // assignment operator
b. l1: if p(x1,…,xn) then l2 else l3;// conditional operator
c. l1: do P1 while p(x1,…,xn) then l2;// loop with body P1 and iteration condition
p(x1,…,xn),
d. l1: do P1 then l2; // call sub schema or procedure P1.
Text after “//” is a comment, not a part of instructions.
Here l1, l2, l3 are called labels, l1 is called an input label, l2 and l3 are called output
labels. P1 is called a sub schema and set of it labels does not have common labels with upper
level program or any other procedures, p1 is called a repetition predicate,
, x0, x1, …, xn
VarX VarA.
Output labels which are not input label are called final, and only one final label is
allowed. Collection of labels of program or procedure P will be denote L(P).
One separate label l0 is called start label and its operator – start operator.
We assume that each of sets Opers, F, VarX, VarA, subP includes such sets for
procedures too.
Schema is called determined if its instructions have different input labels.
Interpretation of the schema is a map of some finite set of memory elements to elements
from (they called start value for the program), function symbols interpreted as maps [ n
],
n
predicate symbols interpreted as maps [
{true, false}].
Program variable x from X has a value <x> of the element of memory x, variable
a[K1(x1, …, xn)…,Kn(x1, …, xn)] has a value a[IK1(<x1>,.., <xn>), …, IKn(<x1>,…, <xn>)], IKj
is an interpretation of Ki. Value of the variable is empty (not defined) if some of its index
function has no value or one of memory element used as argument for index function is empty.
The execution of operators for a given interpretation consists of three parts calculating some function or sub schema from the operator, changing the memory elements and
marking some operator as next to execute.
More accurate, for operator of type:
a. calculate function Ig (interpretation of g) with <x1>, …, <xn> arguments and put the
result to cell x0, mark next to execute operator with input label l2;
b. calculate Ip(<x1>,…,<xn>), if it is true - mark next to execute operator with label l2, if
it is false - mark next to execute operator with label l3;
c. calculate sub schema P1 and then if Ip(<x1>, …, <xn>) is true repeat this operator
again, if it is false then mark next to execute operator with label l2;
d. calculate sub schema P1 and then mark next to execute operator with label l2;
Operator execution is defined in following cases
a. all variables x1, …, xn are defined and function Ig(<x1>, …,<xn>) is defined as well;
b. all variables x1,…, xn are defined and predicate Ip(<x1>,…,<xn>) is also defined;
c. any iterations of P1 are finished with final label and Ip(<x1>,…,<xn>) after each
iteration has a value true and becomes false after finite number of steps;
d. sub schema P1 stops in his final label.
Execution of schema for a given interpretation is a chain of executions of marked
operators starting with operator with start label l0.
Execution is considered ended successfully if after finite number of steps some of
marked label is final label.
Execution ended without result if some supposed to be calculated predicate or function
does not have value or one of its arguments does not have value.
Remark 1.1. It is possible that schema will run forever without reaching final label. So
schemas calculate partially defined function.
For shortness we will omit letter I in interpretation of functions and predicates if it is
clear from context what it has to be - symbol or its interpretation.
3. Loops in a separated form
Canonic forms of the studying objects always have a theoretical and practical value
because they allow classify objects, to work with smaller diversity and to use smaller description
of the objects. We saw it for step enumeration of structured program. Also it nicely comes with
level of abstraction for program design.
We hope that studying a canonic form for information processes in programs will help to
simplify design, develop better debugging tools and increase the code reusability.
In this section we will show that any loop body can be represent as a sequence of sub
procedures determine values of array indexes for next procedures, but not for itself or upper level
controllers. Last sub procedure called loop‟s kernel, others are called controllers of
corresponding levels. So we separate the loop‟s body execution in two functions: hierarchical
selecting data parts and parts processing them.
Such implementation allows to reduce a debugging of complex processes of data
selection in arbitrary arrays to the chain of debugging a simple blocks without arrays. It comes
from functions of controllers. If each upper level controller works properly, then data stream for
this block works correctly and hence we need to check if this block is correct.
In C++ STL classes and generic collections in C# algorithms represent loops with fixed
controllers for getting next elements from lists, maps, stacks and so on.
For theoretical studies this result is important because it gives syntactical objects for
describing information connections in loops and, as it will be proved next, to divide the loops
with different numbers of controllers or their connection graphs to unequal classes.
Def. 3.1. Let S(P) be a set of sub procedures of P, SS(P) be a transitive closure of this
relations. Then sub procedure P called recursive if P SS(P).
Def .3.2. Schema P is called loop structured (for short L-schema) if for P and for each
of its sub schemas Pi:
a. a relation “input label is less than output labels” is a partial order, called “structural”,
b. Pi is not a recursive procedure.
So there is no recursion call, no spaghetti paths and no loops created by labels.
Remark 3.1. In L-schemas any repetition process can be generated only with nested
loops.
Remark 3.2. L-schemas are structured in Dijkstra‟s terms: each loop has one start point
and one exit point (we consider one final label only).
Def. 3.3. Two schemas with the same set of non interpreted symbols of functions and
predicates are called equal if for each interpretation one schema finished successfully if and only
if another schema finished successfully too and their state of memory is the same.
We can be proved the following
Theorem 3.1. There exists an algorithm to transform any schema to an equivalent Lschema.
Full proof is mostly technical, too bulky and obvious. Instructions that breach the label
orders can be encapsulated in loops. Recursions can be replaced with pair loops and additional
interpreted stack operations. Duplications of sub procedures can be eliminated by renaming and
copying.
Def. 3.4. For a given instruction with input label m denote Ind(m), Arg(m), Val(m) sets
of its index variables, set of its argument variables, set of its output variables. More accurately,
for instruction m of following type
a)
Ind(m) is a set of indexes of array variables x0, x1, …, xn; Arg(m) = {x1,…,
xn} Ind(m); Val(m)={x0},
b)
Ind(m) is a set of indexes of array variables from {x1,…,xn};
Arg(m)={x1,…,xn} Ind(m);Val(m)= ,
c)
Ind(m)= Ind(k|k L(P)); Arg(m)= Arg(k|k L(P)) Ind(m);
Val(m)= Val(k|k L(P));
d)
Ind(m)= Ind(k|k L(P) JI, where J1 is set of indexes of array variables of
predicate p; Arg(m)= Arg(k|k L(P)) Ind(m); Val(m)= Val(k|k L(P)),
For program or sub procedure P a set Ind(P)= Ind(k|k P), Arg(P)= Arg(k|k P) Ind(P),
Val(P)= Val(k|k P).
Def. 3.5. Loop C with non empty set of index variables is called separated if its body P
is a set of instructions
m0: do P1 then m1;
…..
mn: do P2 then mn+1,
where i j [(Ind(Pi) Val(Pi+j)= ) &(Ind(Pi) Val(Pi-1)
)], i, j N.
Here P1,…Pk-1 are called controllers of levels 1,…,k-1, the last Pk is called a kernel of
the loop.
Def. 3.6. The separated loop called strictly separated if
Pi(i<k) Val(Pk) (Arg(Pi) Ind(Pi))= .
Other words, a kernel of a strictly separated loop does not change any variable used by
any controller.
Def. 3.7. Schema is called simple if it consists only of instructions of type a) and d).
Def. 3.8. Two schemas called t-equally if they are equal for each interpretation where
functions and predicates are define for each arguments (also called totally define).
Def .3.9. Schema is called forward oriented if each loop has a body in which index of
variable for any interpretation can be changed only before using. Syntactically, for each branch
with array variables there is no operator changing its indexes with bigger label.
The following auxiliary statement can be proven.
Lemma 3.1. There exists an algorithm of transformation of any L-schema to the t-equal
forward oriented schema.
Idea of proof. Let‟s have two instructions S1 and S2 in one branch of P. S1 is executed
early (input label of S1 is less than for S2) and uses index variable E and S2 is executed later and
changes E in one of the branches. Then we‟ll add new variable newE and add ahead of S1
instruction
newE = E;
and replace index E to newE in S1.
Modified branch looks like next
….
newE=E;
x1=f(x[.., newE,..]…);
….
E=g(…);
…
It clearly equals to old branch for any interpretation of f and g and the branch now satisfies for
forward orientation condition. By repeating such modification for any branch and loop body we
will end up with forward oriented schema.
Now we are ready for the main result of this section.
Theorem 3.2. There exists an algorithm to transform any loop C with arrays to t-equally
separated loops with some n controllers and there is no any equally separated loop with different
number of controllers.
Proof of the first part of theorem. Let‟s B be a body of the loop C. According to lemma
2.1 we may assume that B has a “structural” partial order “<” on L(B). Let‟s for each branch
collect such of instructions k with bigger input label than start instruction in order “<” and built a
set of left side variables Vs = Val(k) until we meet instruction with indexes from Vs or get a
final one. Let‟s call this instruction “limited”, and continue with other branches. Process stops
when all branches are visited and each ended with limited (red) or the final instruction. Visited
set of instructions constitutes a first level controller.
To finish building the first controller let‟s add interpreted instructions with a new variable
vLeb1 which will keep the output label of last executed instructions. It may looks like the
following. Let mr is a final or limited output label. Then we‟ll replace mr with new label mAux
and add instruction
mAux: vLab1=‟mr‟.
We also add the next interpreted instructions to the rest set of instructions after removing
organizer‟s instructions:
maux0: if (vLeb1==m1) then m1 else maux1 // start instruction of next procedure
maux1: if (vLeb1==m2) then m2 else maux2
maux2: ..,
here mauxi are new auxiliary labels.
These additions guarantee that after execution of the first controller the calculations will
continue in an original order.
Clear, that if the rest set of instructions has instruction which changes indexes of others
we may repeat the above process: mark as limited those instructions that have indexes from
Ind(k|k L(P)) – Val(P0) and separate next level controller Pi. If there are no such instructions
then the loop has only i controllers and the set of rest instructions is a kernel.
Second part of the statement (about number of controllers) can be proven using special
interpretations which we borrowed from Gödel‟s model theory. He developed it to study logical
model (and called it model on constants). We will use this technique several times latter in this
study.
Suppose we have some formal system with signature consisting of symbols of functions
from F and predicates from P. Then the set of values for standard interpretations is a set of
terms T, built with elements of F and variable expressions. Formally T can be defined by
induction:
1. Simple variables X, used in schema (finite set) are elements of T;
2. If a[r(x1,.., xn1), ..] is an array variable in schema, then expression (string of symbols)
„a[r1(x1,…, xn1), ..]‟ is in T for any x1,…, xn from T.
3. term f(t1,…,tn) for each t1,..tn from T and f F also belongs to T.
Standard interpretation of a functional symbol f with n arguments is a map terms
t1,…, tn to the term (chain of symbols) „f(t1,…, tn)‟ if it is in T and is not define if it is not in T. In
other word, the value of a function is a path of its calculation.
Standard interpretation of a predicate symbol is determined with finite set D, called a
diagram, of expressions „p(t1,…,tn)‟ or „¬ p( t1,…,tn)‟ (only one of these two expressions is
allowed to be in D ), where t1, …,tn are from T.
Now we may prove the second part of the Theorem 2.2 that number of controllers is
invariant for equivalent transformations.
Let‟s consider loop C for standard interpretation. By construction, each controller
changes at least ones the indexes used in next organizer. So the value of indexes for standard
interpretation have to be a word as „a[…, t,…]‟where term t is a value of previous controller.
Next level controller has to have a branch which changes indexes for next after it level and has to
contain this expression. Otherwise if each path (for total schema it is also a branch) of such
controller does not have this word in expression t for index, then all paths must be included in
the previous levels.
Therefore execution of n controllers has to have value with n-1 deepness of parenthesis
[]. Hence two loops with different number of controllers can‟t be equal.
Proof is finished.
This technique may be used for more detailed classification of the loops. For example,
from the proof it also follows that loops with different dependency graph between controllers
also can‟t be equal.
4. Immediate information dependency between iterations
Each loop iteration has to depend from previous immediate iteration, otherwise iteration
will be just a repetition of the previous iteration (memory used by iteration doesn‟t change) and
hence, loop will run infinitely long. So iterations of whole loop body can‟t be executed in
parallel, but parts of it can be. Iterations of level 1 controller have to have connections with body
iterations; otherwise it can be run only one time.
For body with an indexed variable the result created on one iteration n0 can be used on
another iteration n1 (n0<n1), not immediate next to n0. To determine that the value of a[K0(i|n0)]
created on iteration n0 used with indexed variable a[K1(j|n1)] on iteration n1 we have to solve an
equation
K0(i|n0) =K1(j|n1)
for n0 and n1. The expression i|n0 means value of i on iteration n0.
For this equation we have to identify the system execution steps of the nested loops.
We‟ll use the following notations for nested loops and its elements. If Ci1 is a loop with
body Bi1, then loops that belong to it will be denoted as Ci1,i2. In general Ci1,…,in will denote loops
belonging to the body of the loop Ci1,…,in-1. So depth of the nested loop is reflected in the index
dimension of its name. The next diagram illustrates this notation:
Figure 4.1.
For simplicity we suppose that in a schema all instructions are different and each loop has
a counter of iterations starting at 0 when loop instruction is initialized. Then for loop Ci1,…,in a
vector of counters m=m1,…,mn of the inner loops Ci1,…,ip for p<n +1 will be unique for the step
of execution of any instruction q from its body and we will use the notation qm.
4.1. Connection equation
Immediate connection between steps q1 m1 and q2 m2 (let q1m1<q2m2) takes a place when
there is a simple or indexed variable value of it is created on the step q1 m1 and used on the step
q2 m2, or the result of q1m1 is overwritten with q2 m2 and it is the nearest step with such properties
(there are no any operator q3 and iteration m3 (m3<m2) in between have a property like q2).
For case of simple variable the nearest m2 is the next iteration after vector m1. In case of
indexed variables a[K1(i)] Val(q1) and a[K2(j)] Arg(q2), immediate connection means that
K1(i|m1)=K2(j|m2) and there is no any instruction q3m3 , where q1m1<q3m3<q2m2, with such a
variable a[K3(p)] Val(q3) that K3(p|m3) = K1(j|m1) for m1<m3<m2; i|m1 means value for i
calculated by some controller on iteration m1.
So to detect information connections we have to have solution for following equation:
K1(i|m1)=K2(j|m2).
We will call it a connection equation.
It is a natural number equation which is a superposition of index expressions and
functions of controllers on both sides. We can solve this problem for the system of linear
equations but not for polynomials higher than 3 degrees.
It means that class of programs for solving system of linear equations, matrix operations
and differential equations can be parallelized automatically [1,2]. But this way is a dead end.
Even for a case when connection equation is a polynomial, solving algorithm does not exist: we
have a Diofant‟s equation problem.
Remark 4.1. We can show that the equation is the only problem – if we have solution for
connection equation, no more algorithmically unsolvable problems are left for parallelizing of
free program schemas.
This is technical result and we are not going to show it here.
4.2. Avoiding the connection equation
Let modify programming language (and call it language of programs with
predecessors) by changing only semantics of index variables: instead of asking connection
equation we may ask its solution. Now a variable with index expression like a[g(i)] may be used
only in right side of assignment operator and it means that on current iteration must be used
value for a that was created on previous iteration g(i) and hence g(i)<i. Other word, index
expression reflects a point in iteration space, not a point in an array.
There is no needs to show an index expression in any left side, it always the current
iteration (when this operator was executed).
Example 4.1. Let‟s have a 4-point deferential approximation:
fk(i,j)=1/4(fk(i-1, j)+fk(i, j-1)+fk-1( i+1, j)+f k-1(i, j+1)).
The equivalent program schema with predecessors is next system of loops
for (k=1;k<N+1;k++)
for (i=1;i<N+1;i++)
for (j=1;j<N+1;j++)
f = 1/4(f[k, i-1, j]+f[k, i, j-1]+f[k-1, i+1,j]+f[k-1, i, j+1]);
In reality last line has to be more complex because of start data. Start data or elements of
input arrays might be introduced as negative number indexes. But for simplicity we are not
showing calculation along the coordinate plains.
The natural way to execute a program with predecessors is to repeatedly run in parallel
all iterations of loop‟s body fragments which have data. Obviously more efficient execution is to
look out only iterations that just got data.
For example above controllers are simple and just increase indexes j, i, k. Their values do
not depend on kernel (it a last line of code) and controller iterations can be executed before any
kernel iterations. So we will get a list of 3D vectors (1,1,1), (1,1,2),…, (1,1,N),…, (1,2,1),…,
(1.2,N), …, (1,N,N),…, (N,1,1),…, (N,N,N). For each of them we check if there is data for the
arguments. In practice we have to check only points changed on a previous step.
At the beginning it is the point (1,1,1). Then data will be ready for 3 points (1,2,1),
(1,1,2), (2,1,1).
Let P be a plain having these tree points, and Norm is its normal vector. Then at the next
step data will be ready for iteration lying on next nearest plain with the same normal vector and
including whole number coordinates. One can see that each next available for execution set of
iterations is belonged to next parallel plain with integer points.
Each plain iterations on can be calculated in parallel. Really, for any integer point on the
plain, all points used for this iteration are in an area strictly below this plain. It means that these
points were already calculated and any point of plain can‟t be argument for another point of the
same plain and hence points on one plain can be calculated in parallel.
So this 4-point approximation is highly parallel and for this 3D task parallel execution
time is linear function of the dimension size.
Next figure 4.2 illustrates this case.
Figure 4.2.
Remarkable feature of this language is that it has a trivial algorithmically solvable
problem of finding nearest iteration in which result of current iteration will be used. The index
expression of right side variable is a solution of a connection equation. Still both languages are
algorithmically full. It is easy to prove that there exists a fixed interpretation of functional and
predicate symbols so that each algorithmic function can be calculated with a program from these
classes.
There is no contradiction to Rise‟s theorem about insolvability of mass problems. The
class of our programs is not an exact subclass – it is a full class.
For programs we have to consider two dimensions of fullness: algorithmically when any
partial recursion function has a program for it calculation and data fullness when data selection
functions are an algorithmically full class, any data structure can be represented.
Loops with simple controllers are very suited for parallelizing. They produced well
organized data streams that can be efficiently implemented in parallel systems with separated
memories. Questions are how useful such programs, how to organize parallel execution and how
complex is it. The only comprehensive but expensive answer for that questions is a building of a
real hardware and software system and applications of a wide enough area that executed in the
system. But as it is shown in [3] there are many problems as well.
4.3. Cone of dependency and parallelism
Program with predecessors has a simple geometric interpretation. Let‟s consider
hierarchy of loops with body B of the deepest loop. Then for each iteration i1,…,in we may
calculate all immediate predecessors which values are used on this step. Repeating this for each
predecessor, for predecessors of predecessors and so on we will get a set of iteration which we
call a cone of dependency. It is clear that to get result for the current iteration we need results of
any iteration from this cone.
Figure 4.2 represents the cone for iteration (k,i,j) for the 4-point example above:
Figure 4.2.
Any dependency cones for any point of any plain of iterations executable in parallel do
not include each others, so they are independent from each other. This plain is a tangential to the
cone at point (k, i, j) and the cone lies behind one side of it.
The relationship between sets of independent and sets of iterations executable in parallel
is not simple.
Next example shows that even if any set of iterations are lying on any lines, parallel to
some given one, are independent the parallel execution for such sets does not exist.
Example 4.2.
For (z=1; z<N; z++)
{
For (x=1; x<N; x++)
{
For (y=1; y<N; y++)
{
for (p=1; ((x+p<n)&(y+p<n)); p++)
{
m0: if (x>y) then m1 else m2
m1:v=f1(u[p,x+p], u[x+p,p],v) then m2
m2:v=f2(u[p,y+p], u[y,y+p],v) then m3
m3: u[x,y]=f3(u[x,y], u[x-1,y-1],v) then ms
}
}
}
}
The cone of dependency for these nested loops is shown in figure the 4.4.
Here any set of iterations parallel to L (bold line) consists of independent iterations. But
the two lines cannot be executed in parallel because dependency cone of one of them will contain
points of another one. Thus any parallel to L line is a set of independent iterations, but only one
line iterations can be executed in parallel.
Figure 4.4.
It can be proved that any forward loops can be transform to loops in separated form with
the same technique as for ordinary program schemas. Number of its controllers and their
connection topologies are an invariant for transformation. Therefore loops with different
characteristics like these can‟t be equivalent.
Important feature of the forward loops is its data selection fullness. The constructors
may represent any functions. Hence any data structure might be represented and processed in parallel.
Conclusion. We‟ve shown that problem of loop iterations independency contains the
problem of solving connection equations. The last problem might be avoided by changing
semantic of index expressions. The getting class is algorithmically full and has full data
selections.
References
1. R. Scott, T. Clark, B. Baghery, Scientific parallel computing, Princeton University
Press, 2005.
2. S.L. Graham, M. Sniz, C.A. Patterson, The Future of Supercomputing: An Interim
Report, Washington, DC, National Academies Press. 2005.
3. H. Zima, K. Kennedy, C. Koelbel, The rise and fall of High Performance Fortran: an
historical object lessons, Proceedings of the third ACM SIGPLAN conference on
History of programming languages, 2007, San Diego, California June 09 - 10, 2007
q
q
Figure 1.1. Structured and unstructured loops
C1
C2
C1,2
q3
C1,1
C1,2,1
q1
q2
q4
C3
Figure 4.1. Loop naming
k
j
(k,i-1,j)
(k-1,i,j+1)
3 (k,i,j-1)
(k-1,i,j-1)
2
2
3
i
Figure 4.2. Hyper plains for parallel execution
k
(k,i,j)
i
j
Figure 4.3. A cone for point (k,i,j)
z
L
y
(x0, y0, z0)
x
Figure 4.4. Cone of dependence for system from an example 2.4.2.
| 6cs.PL
|
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC
INITIAL SYSTEMS
arXiv:1608.06670v1 [math.AC] 23 Aug 2016
SARAH MAYES-TANG
Abstract. Several authors investigating the asymptotic behaviour of the Betti diagrams of the graded system {I k } have shown that the shape of the nonzero entries in the
diagrams stabilizes when I is a homogeneous ideal with generators of the same degree.
In this paper, we study the Betti diagrams of graded systems of ideals built by taking the
initial ideals or generic initial ideals of powers, and discuss the stabilization of additional
collections of Betti diagrams. Our main result shows that when I has generators of the
same degree, the entries in the Betti diagrams of the reverse lexicographic generic initial
system {ginτ (I k )} are given asymptotically by polynomials and that the shape of the
diagrams stabilizes.
1. Betti Tables of Graded Systems of Ideals: Examples
Cutkosky, Herzog, and Trung proved that when I is an ideal in S = K[x1 , . . . , xn ], the
regularity of I k is a linear function for k large enough ([CHT99]). Following this lead, later
work examined the asymptotic behaviour of finer invariants of the set of powers of an ideal,
{I k }, such as the the Betti numbers and graded Betti numbers βi,j (I) = dimK Tori (K, I)j
([LV04], [Sin07], [BCH13], [Whi14]). The goal of this paper is to examine the behaviour
of the graded Betti numbers of other families of ideals. In particular, we will consider
graded systems of ideals {Jk } satisfying Ji · Jj ⊆ Ji+j for all i and j.
One type of end behaviour of graded Betti numbers of ideals in a graded system occurs
when the shape of the nonzero entries in the Betti diagrams stabilizes in the following
sense (see [Whi14]).
Definition 1.1. Consider a graded system of ideals {Jk }. We say that the shape of
β(Jk ) stabilizes if there exist integers k0 and r such that for all i, j, and k ≥ k0 ,
βi,kr+j (Jk ) 6= 0 if and only if βi,k0 r+j (Jk0 ) 6= 0.
The smallest k0 satisfying this definition is called the stabilization index of {Jk }.
The first known families with this property are obtained by taking powers of ideals with
homogeneous generators of the same degrees; such ideals are said to be equigenerated.
Theorem 1.2 ([LV04], [Sin07], [Whi14]). Let I be an ideal in a polynomial ring S =
K[x1 , . . . , xn ] that is equigenerated in degree r. Then the shape of β(I k ) stabilizes.
Example 1.3. Let I = (x2 y + z 3 , xyz, yz 2 ). Macaulay2 ([GS]) gives the following Betti
diagrams of I k for small k. Note that βi,i+j (I) is displayed in the ith column and the jth
row of the Betti diagram.
2010 Mathematics Subject Classification. 13D02, 13P10.
Key words and phrases. Betti numbers, generic initial ideals, graded systems of ideals, asymptotic behaviour.
1
2
SARAH MAYES-TANG
β(I) : 0 1
3
3 1
4
− 1
β(I 4 ) : 0
12
15
13
−
1
16
1
2
3
−
β(I 2 ) : 0 1
6
6 4
7
− 1
β(I 5 ) : 0
15
21
16
−
β(I 3 ) : 0 1
9
10 9
10
− 1
1 2
25 6
1 −
β(I 6 ) : 0
18
28
19
−
2
1
−
1
36
1
2
10
.
−
For all k ≥ 3, β(I k ) has the following shape
β(I k ) : 0 1 2
3k
a b c
3k + 1 − d −
and we see from the examples that the stabilization index of {I k } is 3. Looking more
closely at the graded Betti numbers appearing in the tables, we see that the entries are
given by the following polynomials in k:
k 2 3k
k2 3
+
+ 1, b = k 2 , c =
− k + 1, d = 1.
2
2
2
2
Definition 1.4. Consider an indexed collection of ideals {Jk }. We say that the entries of
β(Jk ) are given by polynomials for k >> 0 if there exists a number r such that for all
i, j there is a number k0 (i, j) and a polynomial Pi,j (k) satisfying
a=
βi,rk+i+j (Jk ) = Pi,j (k)
for all k > k0 (i, j)
As suggested by Example 1.3, when I is equigenerated β(I k ) has entries given by polynomials.
Theorem 1.5 (Cor. 2.2 of [Sin07], Prop. 6.3.6 of [LV04]). Let I be an ideal in a polynomial
ring S = K[x1 , . . . , xn ] that is equigenerated in degree r. Then β(I k ) has entries given
by polynomials for k >> 0. Further, the polynomial functions corresponding to nonzero
entries have positive leading coefficients and are of degree less than the analytic spread of
I, `(I).
At first glance, it may seem that if β(Jk ) has entries given by polynomials for k >> 0
then the shape of the β(Jk ) must stabilize. However, notice that as k gets large, the
number of nonzero entries in β(Jk ) may keep increasing, and the collection {k0 (i, j)} may
be unbounded. The following example illustrates this possibility
Example 1.6. Consider the complete intersection ideal I = (xa , y b ) ⊆ K[x, y]. When
1 < a b,
β0,ak+l(b−a) (I k ) = 1
for all k ≥ l
β1,ak+l(b−a)+a (I k ) = 1
for all k ≥ l
and all other Betti numbers are 0 (see [GVT05]). Thus, the Betti numbers of β(I k ) are
given by polynomials but the shape of the diagram does not stabilize because the number
of nonzero entries in β(I k ) is unbounded.
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC INITIAL SYSTEMS
3
The following lemma connects Definitions and 1.1 and 1.4; its proof is immediate from
the definitions.
Lemma 1.7. Let {Jk } be an indexed collection of ideals. Suppose that the entries of β(Jk )
are given by polynomials in k and that there exists a k0 such that the number of nonzero
entries of β(Jk ) is bounded by a constant for all k > k0 . Then the shape of β(Jk ) stabilizes.
Question 1.8. Let {Jk } be a graded system of ideals. Does stabilization of the shape of
β(Jk ) imply that the entries of β(Jk ) are given by polynomials? If not, what conditions
are necessary?
We now consider the asymptotic behaviour of the graded Betti numbers of other graded
systems of ideals. Our main stabilization result concerns graded systems built by taking
the generic initial ideals of powers of an equigenerated ideal.
Theorem 1.9. Let I be an equigenerated ideal in S = K[x1 , . . . , xn ], where K a field of
characteristic 0. Consider the graded reverse lexicographic generic initial system, {gin(I k )}.
a) The shape of β(ginrevlex (I k )) stabilizes.
b) The entries of β(ginrevlex (I k )) are given by polynomials for k >> 0.
When no monomial order is specified, gin(I) represents the generic initial ideal under
the reverse lexicographic order and lexgin(I) represents the generic initial ideal under the
lexicographic order.
Example 1.10. Let I = (x2 y + z 3 , xyz, yz 2 ). Under the reverse lexicographic order, the
Betti diagrams of gin(I k ) for small values of k are as follows.
β(gin(I)) : 0 1
3
3 2
4
1 1
β(gin(I 4 )) : 0
12
15
13
1
1
17
1
2
3
−
β(gin(I 2 )) : 0 1
6
6 5
7
1 1
β(gin(I 2 )) : 0
15
21
16
1
1
26
1
β(gin(I 3 )) : 0
9
10
10
1
2
6
−
1
10
1
2
1
−
β(gin(I 6 )) : 0
18
28
19
1
1
37
1
2
10
−
We see that β(gin(I k )) stabilizes to the following shape
β(gin(I k )) : 0
3k
a
3k + 1
e
1 2
b c
d −
and its stabilization index is 3. The non-zero polynomials guaranteed by part (b) of
Theorem 1.9 are:
a=
k 2 3k
+
+ 1,
2
2
b = k 2 + 1,
c=
k2 3
− k + 1,
2
2
d = e = 1.
Under the lexicographic order, the number of nonzero rows in the Betti diagram of
lexgin(I k ) increases as k increases. For example,
4
β(lexgin(I)) : 0
3
3
2
4
1
5
6
1
SARAH MAYES-TANG
1
3
3
2
1
2
1
1
1
−
β(lexgin(I 2 )) : 0 1
6
6 8
7
4 7
3 5
8
2 4
9
2 3
10
1 2
11
12
1 1
2
3
3
2
2
1
1
−
β(lexgin(I 3 )) : 0
9
10
6
10
4
11
4
12
3
13
3
14
2
15
2
16
1
17
18
1
1
15
10
8
7
6
5
4
3
2
1
We can show that reg(lexgin(I k )) = 6k while the minimal degree generators of lexgin(I k )
are of degree 3k. Therefore, the Betti diagram of lexgin(I k ) has nonzero entries in 3k + 1
rows, and shape of the Betti diagrams does not stabilize. However, it appears that for
any pair (i, j), βi,3k+i+j (I k ) is a polynomial for k >> 0, implying a negative answer to
Question 1.8 for reverse lexicographic generic initial systems. For example,
β1,3k+1+1 (lexgin(I k )) = 4k − 2
for k ≥ 2.
While the shape Betti diagrams of the lexicographic generic initial system {lexgin(I k )}
do not stabilize in the sense of Definition 1.1, there appears to be additional uniformity beyond its entries being given by polynomials. For example, rows 8 through 12 in
β(lexgin(I 2 )) are identical to rows 14 through 18 of β(lexgin(I 3 )). Similar patterns hold for
higher members in the graded system. Further, while the I above shows that the number
of nonzero entries of β(lexgin(I k )) is not always bounded, there are circumstances under
which it does, even when lexgin(I k ) 6= gin(I k ) (for example, consider I = (x2 , xy, z 2 )).
Observations like those lead to the following questions.
Question 1.11. Are the entries of β(lexgin(I k )) given by polynomials for all k >> 0?
Under what conditions does is the number of nonzero entries in β(lexgin(I k )) bounded?
What other types of uniformity do these Betti diagrams exhibit as k → ∞?
The following example demonstrates that generic coordinates are necessary for Theorem
1.9.
Example 1.12. Let I = (x2 y + z 3 , xyz, yz 2 ) and consider the graded system obtained
by taking the lexicographic or reverse lexicographic initial ideals of I k , {in(I k )}. The
regularity of in(I k ) is increasing at a faster rate than the degree of the minimal generators
of I k . In particular, there are k + 1 rows in β(in(I k )) with nonzero entries, so the shape
of β(in(I k )) does not stabilize.
We might hope that the entries in β(in(I k )) are given by polynomials for k >> 0.
However, if we look more carefully at its entries we see that this does not hold. Consider,
for example, β2,3k+2 (in(I k )):
3 4 5 6 7 8 9 10
k
β2,3k+2 (in(I k )) 2 4 8 12 18 24 32 40
2
6
4
4
3
3
2
2
1
1
−
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC INITIAL SYSTEMS
5
For all k ≥ 3,
(
0.5k 2 − k + 0.5
β2,3k+2 (in(I )) =
0.5k 2 − k
k
for odd k
for even k
This function cannot be written as a single polynomial for all k >> 0.
As in the lexicographic generic initial system, there appears to be some asymptotic
uniformity in the Betti diagrams of the initial system, although it is more complicated
than in generic coordinates. Similar statements may be made for the Betti diagrams of
other graded systems of ideals, including {I k } when I is not equigenerated and the system
{I (k) } composed of the symbolic powers of an ideal. This leads to the following question:
Question 1.13. In what ways may the Betti diagrams of a graded system of ideals {Jk }
display uniformity as k → ∞? How can we describe more complicated types of stabilization? In addition to the systems {I k } and {gin(I k )} arising from equigenerated ideals I,
are there other natural graded systems of ideals for which the shape of β(Jk ) stabilizes in
the sense of Definition 1.1?
Bagheri, Chardin, and Ha address similar questions in [BCH13] where they describe
the asymptotic behaviour of the graded Betti numbers of multi-graded systems of ideals.
Section 2 of this paper provides background on the resolutions of generic initial ideals
and Section 3 contains the proof of Theorem 1.9. Additional corollaries of Theorem 1.9
are included in Section 3.3, including a description of the Boij-Söderberg decompositions
of β(gin(I k )).
2. Background: Graded Betti Numbers of Generic Initial Ideals
In this section we record facts related to the graded Betti numbers of generic initial
ideals. Throughout, S = K[x1 , . . . , xn ] is a polynomial ring with the standard grading
over a field of characteristic zero and I is a homogeneous ideal in S. We use ginτ (I) to
denote the generic initial ideal under the term order τ and gin(I) to denote the generic
initial ideal under the reverse lexicographic order.
To describe the graded minimal free resolution of gin(I k ), for any monomial m, we will
denote
max(m) = max{i : xi divides m}.
Definition 2.1 (See Construction 28.6 in [Pee10]). Let M be a Borel-fixed ideal minimally
generated by monomials m1 , . . . , m` . For each mi and for each sequence of natural numbers
1 ≤ j1 < · · · < jp < max(mi ), let S(mi ; xj1 · · · xjp ) denote a free S-module with one
generator in degree deg(mi ) + p. The basis of the Eliahou-Kervaire resolution of M
is given by
B = {S(mi ; j1 , . . . , jp ) : 1 ≤ j1 < · · · < jp < max(mi ), 1 ≤ i ≤ r}
where S(mi xj1 · · · xjp ) is in homological degree p..
Recalling that ginτ (I) is Borel-fixed ([Gal74]), we obtain the following result.
Theorem 2.2 ([EK90]). For any homogeneous ideal I ⊆ S, the Eliahou-Kervaire resolution produces a minimal free resolution of ginτ (I).
6
SARAH MAYES-TANG
Corollary 2.3. Let M be a Borel-fixed ideal minimally generated by monomials m1 , . . . , m` .
Then the graded Betti numbers of M are given by
X
max(mi ) − 1
βp,p+q (M ) =
p
deg(mi )=q,1≤i≤`
The following definition and theorem say that the Betti diagrams of an ideal and any
of its generic initial ideals are related by cancelling numbers along diagonals.
Definition 2.4 (see [Pee10]). We say that a sequence of numbers {qi,j } is obtained from
a sequence of numbers {pi,j } by a consecutive cancellation if there are indices s and r
such that
qs,r = ps,r − 1, qs+1,r = ps+1,r − 1
qi,j = pi,j for all other values of i, j
Theorem 2.5 ([Gre98]). Let I be a homogeneous ideal in S. For any term order τ , the
graded Betti numbers βi,j (I) may be obtained from the graded Betti numbers βi,j (ginτ (I))
by a sequence of consecutive cancellations.
Example 2.6. Let I = (x2 y + z 3 , xyz, yz 2 ). The Betti diagram of gin(I 3 ) is
9
10
0
10
1
1
10
1
2
1
.
−
Theorem 2.5 says that β(I 3 ) will be obtained by making consecutive cancellations. There
are two possible consecutive cancellations in this case: cancelling β0,10 (gin(I 3 )) with
β1,10 (gin(I 3 )) and cancelling β1,11 (gin(I 3 )) with β2,11 (gin(I 3 )). Using Macaulay2, we find
that only the first consecutive cancellation is needed and β(I 3 ) is given by
9
10
0
10
−
1 2
9 1
.
1 −
This example demonstrates that not all possible consecutive cancellations need to occur
when we pass from the Betti digram of gin(I) to the Betti diagram of I. In general, there
are no techniques for determining which consecutive cancellations will occur.
Definition 2.7. A Betti number βi,i+j (I) 6= 0 is called extremal if βk,k+l (I) = 0 for all
pairs (k, l) 6= (i, j) with k ≥ i and l ≥ j. An extremal Betti number is the upper left-hand
corner of a block of zeros within a Betti diagram.
Notice that if I is an equigenerated ideal, the extremal Betti numbers of I k occur in
the same relative positions for all k >> 0 by Theorem 1.2. Our interest in extremal
Betti numbers comes as a result of the following theorem, which says that the reverse
lexicographic generic initial ideal preserves extremal Betti numbers.
Theorem 2.8 (Theorem 4.3.17 of [BCP99]). Let I be a homogeneous ideal of S. Then
for any i, j ⊂ N,
(a) βi,i+j (I) is extremal if and only if βi,i+j (gin(I)) is extremal;
(b) if βi,i+j (I) is extremal, then βi,i+j (I) = βi,i+j (gin(I)).
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC INITIAL SYSTEMS
7
Corollary 2.9 ([BS87]). Under the reverse lexicographic order,
reg(I) = reg(gin(I)).
Another advantage of the reverse lexicographic order is that hyperplane sections
P interact
nicely with generic initial ideals under this order. Consider a linear form h = nj=1 hj xj ∈
S1 such that hn 6= 0. The homomorphism φ : S = K[x1 , . . . , xn ] → Ŝ = K[x1 , . . . , xn−1 ]
1 P
defined by φ(xj ) = xj for j 6= n and φ(xn ) = − ( j6=n hj xj ) defines an isomorphism
hn
between S/(h) and Ŝ that preserves the degree of polynomials not in the kernel.
Definition 2.10. Given an ideal I in S, the h-hyperplane section of I is the ideal φ(I)
in Ŝ. It is denoted Ih .
Theorem 2.11 ([BS87], [Gre98]). Let I be a homogeneous ideal in S and let h ∈ S1 be a
generic linear form. Then
gin(Ih ) = gin(I)|xn →0 .
3. Stabilization of β(gin(I k ))
The goal of this section is to prove Theorem 1.9 describing the stabilization of β(gin(I k )).
Throughout, I is a homogeneous ideal of S = K[x1 , . . . , xn ] with the standard grading.
We denote the generic initial ideal of I with respect to the reverse lexicographic order by
gin(I).
3.1. Part (a) of Theorem 1.9 follows from part (b). The following lemma claims
that for all k >> 0 the number of nonzero entries in β(gin(I k )) is bounded. Note that it
only holds for the reverse lexicographic order (see Example 1.10).
Lemma 3.1. Suppose that I is an ideal equigenerated in degree r, and let c and k0 be the
constants such that reg(I k ) = rk + c for k > k0 , guaranteed by Theorem 1.2 . Then for all
k > k0 , the nonzero entries of β(gin(I k )) are confined to the following window.
0
rk
∗
rk + 1 ∗
.
..
..
.
1
∗
∗
..
.
···
···
···
..
.
∗
∗
···
rk + c
n−1
∗
∗
..
.
∗
Proof. Fix k > k0 . Since all elements of gin(I k ) are of degree at least rk, the least row
of β(gin(I k )) containing nonzero entries is the rkth row. By Corollary 2.9, reg(I k ) =
reg(gin(I k )) = rk + c so the (rk + c)th row of β(gin(I k )) is the greatest containing nonzero
entries. Finally, since projdim(gin(I k )) ≤ n − 1, the (n − 1)st column is the greatest
column containing nonzero entries.
Since Lemma 3.1 says that only a finite number of βi,i+rk+j (gin(I k )) may be nonzero,
part (a) of Theorem 1.9 follows from part (b) of the theorem and Lemma 1.7. Therefore,
we will focus on the proof of part (b).
8
SARAH MAYES-TANG
3.2. Proof of Part (b) of Theorem 1.9.
Proof. We will proceed by induction on the number of variables n in S.
Base Case: Result holds for 2 variables.
Suppose that I ⊆ K[x, y] is equigenerated in degree r. Then by Theorem 1.5, βi,rk+i+j (I k )
is given by a polynomial Pi,j (k) for all k >> 0. By consecutive cancellation (Theorem
2.5), there exist integers ci,j (k) such that βi,rk+i+j (gin(I k )) = Pi,i+j (k) + ci,i+j (k).
Note that projdim(gin(I k )) and projdim(I k ) are at most 2 for all k >> 0, so ci,i+j (k) =
Pi,i+j (k) = 0 whenever i ≥ 2 .
0
β(gin(I k ))
rk
P0,0 (k) + c0,0 (k)
P (k) + c (k)
rk + 1
0,1
0,1
P (k) + c (k)
rk + 2
0,2
0,2
..
.
1
P1,0 (k) + c1,0 (k)
P1,1 (k) + c1,1 (k)
P1,2 (k) + c1,2 (k)
..
.
We will now proceed by induction on the row index j to show that ci,j (k) is a polynomial
in k for all k >> 0. This will imply that the entries in β(gin(I k )) are given by polynomials
for all k >> 0.
According to Lemma 3.1, βi,rk (gin(I k )) = 0 for all i > 0. By consecutive cancellation,
Theorem 2.5, this implies that c0,0 (k) = 0. The Eliahou-Kervaire resolution implies that,
for all k >> 0, P0,0 (k) gives the number of minimal generators of gin(I k ) of degree rk,
while P1,0 (k) + c1,0 (k) gives the number of minimal generators of gin(I k ) of degree rk
divisible by y. We know that xrk the only generator of gin(I k ) of degree rk not divisible
by y, so P1,0 (k) + c1,0 (k) = P0,0 (k) − 1. Therefore c1,0 (k) = P0,0 (k) − 1 − P1,0 (k) is a
polynomial in k for k >> 0.
Suppose that we know ci,j (k) is a polynomial in k for i = 0, 1 and all j < J. First notice
that both P0,J (k) + c0,J (k) and P1,J (k) + c1,J (k) represent the number of generators of
gin(I k ) of degree rk + J since y divides all generators of degree greater than rk. Thus,
P0,J (k) + c0,J (k) = P1,J (k) + c1,J (k) and c1,J (k) is a polynomial for all k >> 0 if c0,J (k) is.
However, by consecutive cancellation, c0,J (k) = c1,J−1 (k) is polynomial by our inductive
assumption. Therefore, Theorem 1.9 holds when I is an equigenerated ideal in K[x, y].
Inductive Step: Hyperplane Restriction
Now suppose that Theorem 1.9 holds for all equigenerated ideals in K[x1 , . . . , xn−1 ].
We will show that this implies the theorem for equigenerated ideals in K[x1 , . . . , xn ].
The key element of this argument is hyperplane restriction, Theorem 2.11. Let I be an
equigenerated ideal in K[x1 , . . . , xn ] and let Ih be a generic hyperplane section of I. Ih is
a homogeneous ideal of K[x1 , . . . , xn−1 ] equigenerated in degree r and
gin((Ih )k ) = gin((I k )h ) = gin(I k )|xn →0 .
Since our inductive assumption applies to {Ih } ⊆ K[x1 , . . . , xn−1 ], there are polynomials
fi,j (k) such that βi,rk+i+j (gin((Ih )k )) = βi,rk+i+j (gin((I k )h )) = fi,j (k) for i, j, and k >>
0. From Theorem 1.5, we know that there are polynomials Pi,j (k) such that βi,rk+i+j (I k ) =
Pi,j (k) for all k >> 0 and by consecutive cancellation βi,rk+i+j (gin(I k )) = Pi,j (k) + ci,j (k)
for some numbers ci,j (k). To complete the inductive step it is sufficient to show that the
ci,j (k) are polynomials in k for all k >> 0.
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC INITIAL SYSTEMS
0
β(gin(I k )h ) :
rk
f0,0 (k)
f0,1 (k)
rk + 1
..
..
.
.
1
f1,0 (k)
f1,1 (k)
..
.
f0,ρ0 (k)
f1,ρ0 (k)
rk + ρ0
β(gin(I k )) :
0
rk
P0,0 (k) + c0,0 (k)
P (k) + c (k)
rk + 1
0,1
0,1
..
.
..
.
rk + ρ
P0,ρ (k) + c0,ρ (k)
···
···
···
n−2
fn−2,0 (k)
fn−2,1 (k)
..
.
fn−2,ρ0 (k)
···
···
1
P1,0 (k) + c1,0 (k)
P1,1 (k) + c1,1 (k)
..
.
9
···
···
···
···
P1,ρ (k) + c1,ρ (k) · · ·
n−1
Pn−1,0 (k) + cn−1,0 (k)
Pn−1,1 (k) + cn−1,1 (k)
..
.
Pn−1,ρ (k) + cn−1,ρ (k)
By Theorem 2.11, for any δ the set of degree δ minimal generators of gin(I k ) not
divisible by xn is the same as the set of degree δ minimal generators of gin((I k )h ). We
will use this fact along with the Eliahou-Kervaire resolution to determine how the fi,j s
and Pi,j s are related, row-by-row.
For a monomial ideal M , let mi,δ (M ) denote the number of minimal generators of M
of degree δ divisible by xi and no larger-indexed variable. β0,rk (M ) gives the number
of generators of M of degree rk. Since βi,rk (gin(I k )) = 0 for all i > 0, c0,0 (k) = 0 by
consecutive cancellation. Then
mn,rk (gin(I k )) = β0,rk (gin(I k )) − β0,rk (gin((I k )h ))
= P0,0 (k) − f0,0 (k).
We know that P0,0 (k) and f0,0 (k) are eventually polynomial in k, so mn,rk (gin(I k )) is a
polynomial in k for all k >> 0.
Similarly, in homological degree I, Corollary 2.3 implies that
n−1
n
X q − 1
X
q−1
k
mq,rk (gin((I k )h ))
mq,rk (gin(I )) −
[PI,0 (k) + cI,0 (k)] − fI,0 (k) =
I
I
q=I+1
q=I+1
n−1
=
mn,rk (gin(I k ))
I
Since we know that the PI,0 (k), fI,0 (k), and mn,rk (gin(I k )) are polynomial for all k >> 0,
cI,0 (k) must be polynomial for k >> 0 as well. Therefore, we know that the cancellation
numbers ci,0 (k) in the first nonzero row of β(gin(I k )) are polynomial for all k >> 0.
Now assume that we know that the cancellation numbers ci,j (k) are polynomial for all
k >> 0 whenever j < J. By consecutive cancellation,
c0,J (k) = c1,J−1 (k) − c2,J−2 (k) + · · · + (−1)n cn−1,J−(n−1) (k).
By our assumption, this implies that c0,J (k) is also polynomial for all k >> 0. As before,
[P0,J (k) + c0,J (k)] − f0,J (k) = mn,rk+J (gin(I k ))
so mn,rk+J (gin(I k )) is polynomial for all k >> 0 as well.
10
SARAH MAYES-TANG
In column I and row rk + J,
n
n−1
X
X q − 1
q−1
k
[PI,J (k) + cI,J (k)] − fI,J (k) =
mq,rk+J (gin(I )) −
mq,rk+J (gin((I k )h ))
I
I
q=I+1
q=I+1
n−1
=
mn,rk+J (gin(I k ))
I
Since mn,rk+J (gin(I k )) is polynomial for all k >> 0, we see that cI,J (k) must also be
polynomial in k for all k >> 0. Therefore, all cancellation numbers in row rk + J are
polynomial for k >> 0. We conclude by induction on the row index that that all entries
in the table β(gin(I k )) are given by polynomials in k for k >> 0 when I ⊆ K[x1 , . . . , xn ]
is equigenerated.
Therefore, by induction on the number of variables, we conclude that the theorem holds
for all equigenerated ideals I in any number of variables.
3.3. Consequences of the Proof. By Theorem 1.5, the polynomials Pi,j (k) that give
the entries of β(I k ) are of degree less than `(I), the analytic spread of I. If we follow the
degrees of the polynomials in the above proof, we see that the same can be said for the
functions giving the cancellation numbers.
Corollary 3.2. Let I be an ideal equigenerated in degree r. Then the degree of each
polynomial Qi,j (k) = βi,i+j+rk (gin(I k )) guaranteed by Theorem 1.9 is less than `(I).
Similarly, we can follow the stabilization indices in the proof to obtain the following
result.
Corollary 3.3. Let I be an equigenerated ideal with generators of degree r. Then the
stabilization index of {gin(I k )} is at most the stabilization index of {I k }.
We may also consider the structure of the Boij-Söderberg decompositions of β(gin(I k )).
The proof of the main result from [MT15] holds for any family of ideals {Jk } such that
the shape of β(Jk ) stabilizes and whose entries are given by polynomials in k. Therefore,
Theorem 1.9 gives the following result.
Corollary 3.4. Let I be an equigenerated ideal with generators of degree r. Then there
are integers m and K such that for all k > K, the positive Boij-Söderberg decomposition
of β(gin(I k )) is of the form
β(gin(I k )) = w1 (k)π1 (k) + · · · + wm (k)πm (k)
where each wi (k) is a polynomial in k with rational coefficients and π1 (k) < π2 (k) < · · · <
πm (k) is a translated family of chains.
References
[AHH00] Annetta Aramova, Jürgen Herzog, and Takayuki Hibi, Ideals with stable Betti numbers, Advances
in Mathematics 152 (2000), 72–77.
[BCH13] Amir Bagheri, Marc Chardin, and Huy Tai Ha, The eventual shape of Betti tables of powers of
ideals, Mathematical Research Letters 20 (2013), no. 6, 1033–1046.
[BCP99] Dave Bayer, Hara Charalambous, and Sorin Popescu, Extremal Betti numbers and applications
to monomial ideals, Journal of Algebra 221 (1999), no. 2, 497–512.
ASYMPTOTIC STABILIZATION OF BETTI DIAGRAMS OF GENERIC INITIAL SYSTEMS
[BS87]
11
David Bayer and Michael Stillman, A criterion for detecting m-regularity, Inventiones mathematicae 87 (1987), no. 1, 1–11.
[CHT99] S Dale Cutkosky, Jürgen Herzog, and Ngô Viêt Trung, Asymptotic behaviour of the castelnuovomumford regularity, Compositio Mathematica 118 (1999), no. 3, 243–261.
[EFW11] David Eisenbud, Gunnar Fløystad, and Jerzy Weyman, The existence of equivariant pure free
resolutions, Annales de l’Institut Fourier, vol. 61, 2011, pp. 905–926.
[EK90] Shalom Eliahou and Michel Kervaire, Minimal resolutions of some monomial ideals, Journal of
Algebra 129 (1990), no. 1, 1–25.
[Gal74] A. Galligo, A propos du théorem de préparation de Weierstrass, Lecture Notes in Mathematics
409 (1974), 543–579.
[Gre98] Mark Green, Generic initial ideals, Six lectures on commutative algebra, Springer, 1998, pp. 119–
186.
[GS]
Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in algebraic geometry, Available at http://www.math.uiuc.edu/Macaulay2/.
[GVT05] Elena Guardo and Adam Van Tuyl, Powers of complete intersections: graded betti numbers and
applications, Illinois Journal of Mathematics 49 (2005), no. 1, 265–279.
[Kod93] Vijay Kodiyalam, Homological invariants of powers of an ideal, Proceedings of the American
Mathematical Society 118 (1993), no. 3, 757–764.
[LV04]
Olga Lavila-Vidal, On the diagonals of a Rees algebra, arXiv:math/0407041 [math.AC].
[MT15] Sarah Mayes-Tang, Stabilization of Boij-Söderberg decompositions of ideal powers,
arXiv:1509.08544 [math.AC].
[Pee10] Irena Peeva, Graded syzygies, vol. 14, Springer Science & Business Media, 2010.
[Sin07]
Pooja Singla, Convex-geometric, homological and combinatorial properties of graded ideals, Ph.D.
thesis, Universität Duisburg-Essen, Fakultät für Mathematik, 2007.
[Whi14] Gwyneth Whieldon, Stabilization of Betti tables, Journal of Commutative Algebra 6 (2014),
no. 1, 113–126.
Quest University Canada, 3200 University Blvd, Squamish, British Columbia, V8B 0N8
| 0math.AC
|
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
arXiv:1603.03450v1 [stat.ME] 9 Mar 2016
Multisensor–Multitarget Bearing–Only Sensor
Registration
Ehsan Taghavi, R. Tharmarasa
and T. Kirubarajan
McMaster University
Hamilton, Ontario, Canada
email: {taghave, tharman, kiruba}@mcmaster.ca
Mike McDonald
Defence Research and Development Canada
Ottawa, Ontario, Canada
email: [email protected]
Abstract
Bearing–only estimation is one of the fundamental and challenging problems in target tracking. As in the case of radar tracking,
the presence of offset or position biases can exacerbate the challenges in bearing–only estimation. Modeling various sensor biases
is not a trivial task and not much has been done in the literature specifically for bearing–only tracking. This paper addresses the
modeling of offset biases in bearing–only sensors and the ensuing multitarget tracking with bias compensation. Bias estimation
is handled at the fusion node to which individual sensors report their local tracks in the form of associated measurement reports
(AMR) or angle-only tracks. The modeling is based on a multisensor approach that can effectively handle a time–varying number
of targets in the surveillance region. The proposed algorithm leads to a maximum likelihood bias estimator. The corresponding
Cramér–Rao Lower Bound to quantify the theoretical accuracy that can be achieved by the proposed method or any other algorithm
is also derived. Finally, simulation results on different distributed tracking scenarios are presented to demonstrate the capabilities
of the proposed approach. In order to show that the proposed method can work even with false alarms and missed detections,
simulation results on a centralized tracking scenario where the local sensors send all their measurements (not AMRs or local
tracks) are also presented.
Index Terms
Bias estimation, bearing–only tracking, target motion analysis, triangulation, filtering, multisensor–multitarget
I. I NTRODUCTION
Multisensor–multitarget bearing–only tracking is a challenging problem with many applications [4], [5], [13], [27]. Some
applications of bearing–only tracking are in maritime surveillance using sonobuoys, underwater target tracking using sonar
and passive ground target tracking using Electronic Support Measures (ESM) or Infra–red Search and Track (IRST) sensors.
In such applications, it is of interest to find the target position as well as any biases that may affect estimation performance.
From the early works in [1], [2], [36] to the more recent works in [29] and the references therein, the focus has been only
on tracking the targets based on measurements from a bearing–only sensor. However, due to the limitations of single sensor
bearing–only tracking, i.e., due to the need for own–ship maneuvers for the observability of state parameters [28], the issue of
biases in passive single sensor tracking has not been addressed in the literature. The main focus of this paper is multisensor
bearing–only tracking in the presence of biases. In multisensor bearing-only tracking, observability is no longer a major issue.
However, in the case of port–starboard ambiguity, the problem of observability was discussed in detail in [10]. Besides, the
presence of sensor biases that are often unaccounted for can degrade the estimation results significantly. Most of the works on
bias estimation have been about radar tracking (see [20], [42], [43] and the references therein) or using other measurements
besides bearing information [6], [7]. For example, when the elevation information is available, one can estimate the offset
biases as in [7], [14].
With the objective of providing a combined bias estimation and target tracking algorithm that is both theoretically sound
and practical, the problem of multisensor bearing-only multitarget tracking is considered in this paper. Having more than one
passive sensor in the surveillance region ensures the observability of the state parameters, i.e., position and velocity of the
target, without the need for maneuvers [33], [35], [44]. One of the issues that can complicate bearing–only tracking is the bias
in the sensors. For example, in maritime surveillance using sonobuoys, which are usually dropped from an aircraft or thrown
from a ship close to an area of interest, the exact locations of the sonobuoys are not known. This leads to position biases [25].
This is also an isuue in modern systems such as autonomous underwater vehicles (AUV) [9]. In addition, the impact with the
water surface and the waves can result in systemic offset biases [6]. In wide area surveillance using airborne IRST sensors,
uncertain platform motion can contribute to biases as well. Offset bias can be modeled as an additive constant term affecting
the measurement equation and the sensor position uncertainty can be modeled using a random walk [31].
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
Negligible biases can be treated as residual errors. This residual error can be used in the form of additional uncertainty
in the measurements later in the filtering step. However, if offset biases are larger than the noise standard deviation of the
bearing–only measurements, a mirror of the target’s bearing is sent to fusion center instead of its actual value, which will
result in totally erroneous estimates. Further, these erroneous measurements can worsen state estimation results when fused
with measurements from similarly biased sensors. In order to benefit from the information available from multiple bearing–only
sensors with offset bias, one needs to model, estimate and then correct the biases. This is precisely the motivation for this
paper.
The focus of this paper is offset bias estimation. In order to model offset biases, one can transform the measurement
space of sensor data to Cartesian coordinates followed by covariance matrix transformation. This transformation will make it
possible to find an exact model for the biases that can be used in bias estimation and correction. However, the full position
information is not available in a single bearing–only measurement. The process of measurement transformation is done by
paring measurements from different sensors in the surveillance region. That is, this transformation is done through triangulation
[21]. With the new pseudo–measurement in Cartesian coordinates, position and velocity of the targets can be estimated over
time as new measurements are generated from paired sensors at subsequent times [3], [19], [24], [35]. Assuming that these
state estimates also carry the effects of offset biases, it is possible to find such biases, if any, and correct them. In the case
of bistatic passive sensors, one can use the methods in [37], [44]. However, previous work on multisensor bearing–only bias
estimation is still limited. The method presented here gives a comprehensive analysis of offset bias modeling in multisensor
passive bearing–only sensors.
The proposed method gives an exact model for bearing–only biases in Cartesian coordinates. In addition, the formulation
of an appropriate likelihood function enables the use of maximum likelihood estimators to find the biases. Also, the proposed
model is robust against large sensor noise standard deviation. Finally, as shown through simulation results, large bearing biases
can be estimated accurately, which leads to correspondingly accurate target state estimation results.
The goal of this paper is to present a step–by–step approach for designing a multisensor–multitarget tracking system based
on biased bearing–only measurements and give a practical solution to the problem of bias estimation. In Section II, a detailed
model of the multisensor bearing–only estimation problem with bias is given. Section III is devoted to modeling the offset biases
in Cartesian coordinates. In Section IV, a practical solution for bias estimation is proposed. Section V presents the derivation of
Cramér–Rao lower bounds. Simulation results are shown in Section VI along with discussions on different scenarios. Finally,
Section VII ends the paper with conclusions.
II. B EARING –O NLY E STIMATION P ROBLEM
Bearing–only sensors with operating ranges of hundreds of meters to a few kilometers are one of the most crucial sensors in
maritime or ground surveillance applications. These sensors can actively or passively detect the directions of arrival of signals
emitted by the targets of interest. While underwater surveillance is the common application of bearing–only tracking, it is also
used in surface and air target tracking. For example, ESM and IRST sensors also use bearing–only sensors for tracking. As
shown in Figure 1, bearing–only sensors can be on the own–ship or deployed separately in the surveillance region. Moreover,
they can operate under different environmental conditions as shown in Figure 1 [8], [26].
Sonar sensor
𝜌2
𝜌1
spherical spreading loss
cylindrical spreading
Fig. 1: Bearing–only sensors and signal propagation types [26]
The bearing–only measurement from passive sensors is written as
zs (k)
=
θs (k) + bs + ws (k),
(1)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
where zs (k) is the direction of arrival at sensor s, θs is the true bearing of the target, bs is the constant bias in the measurements
of sensor s and ws is an additive zero–mean white Gaussian noise with variance σθ2 . It is assumed that there are S bearing–only
sensors in the surveillance region at positions (xs , ys ) for s = 1, 2, ..., S, and, they record targets’ bearings at time instants
k = 1, ..., K. Note that there is no index to denote target ID, but wherever such clarification is needed, it will be included.
In this paper, bias estimation is handled at the fusion node to which individual sensors report their local tracks in the form
of associated measurement reports (AMR) [16], [46] or angle-only tracks. That is, a distributed tracking system is considered
as in the case of [31], [32], [42]. However, the difference is that in these earlier works local full-state tracks were sent to the
fusion node for bias estimation whereas AMRs or bearing-only tracks are sent in the present case where the local sensors do
not have full observability due to bearing-only measurements. As in [31], [32], [42], working with tracks or AMRs obviates
the need at the fusion center to address false alarms and missed detections that are inevitable at the local sensors although
ghost tracks may be present among local tracks. However, in order to show that the proposed method can work even with false
alarms and missed detections in a centralized tracking system, in one experiment in Section VI we assume that the local sensors
send all their measurements (including false alarms and missed detections instead of AMRs or local bearing-only tracks) and
evaluate the performance of the proposed method. In [37] and [44], where bias estimation at measurement level (rather than at
track level as in [31], [32], [42]) is considered, false alarms and missed detections are not addressed at all. In [30], a joint data
association and bias estimation method was proposed for linear measurement models, which is not applicable for bearing-only
systems. A general case of multistatic passive radar system with false alarms and missed detections was considered in [17]
and [15], but, the bias problem was not addressed.
The goal of bearing–only tracking is to find the bias in each sensor and estimate each target’s position as accurately as
possible based on the model given in (1), either as decoupled parameters or jointly [11], [45]. Due to the computational burdens
of joint tracking and parameter estimation methods [22], a decoupled bias and state estimation is presented in this paper.
It is assumed that each target is following the Discretized Continuous White Noise Acceleration (DCWNA) or the nearly
constant velocity (NCV) model [5]. As a result, a target’s state vector in 2D Cartesian coordinate is given by
T
x(k) vx (k) y(k) vy (k)
x (k) =
(2)
with (x(k), y(k)) being the position and (vx (k), vy (k)) being the velocity. The motion model can further be defined as
x (k + 1)
=
F (k)x (k) + ν(k)
(3)
where the state transition matrix is
1
0
F (k) ,
0
0
T
1
0
0
0
0
1
0
0
0
T
1
and the covariance matrix of ν(k) which is an additive zero–mean white Gaussian noise vector is
Qx 0
Q =
0 Qy
(4)
(5)
in which Qx and Qy are defined as
Qx
=
Qy
=
1 3
3T
1 2
2T
1 3
3T
1 2
2T
1 2
2T
T
1 2
2T
T
q̄x
(6)
q̄y
(7)
where q̄x and q̄y are noise intensities with dimension m2 /s3 [4].
In order to transform the measurements from polar to Cartesian coordinates, it is initially assumed that S is even and that
the sensors are paired into S2 one–to–one pairs. Note that this constraint is relaxed in Section IV. For a pair of sensors with a
single target in the common field of view, the best estimate for the location of the target, independent of its previous location,
can be obtained through triangulation [14], [21], [38]. The triangulated estimates of the target position at time k using sensor
pair (i, j), ignoring measurement noise, are given by
x̂ij (k)
=
yj − yi + xi tan (θi (k) + bi )
xj tan (θj (k) + bj )
−
tan (θi (k) + bi ) − tan (θj (k) + bj ) tan (θi (k) + bi ) − tan (θj (k) + bj )
(8)
ŷij (k)
=
yj tan (θi (k) + bi ) − yi tan (θj (k) + bj ) (xi − xj ) tan (θi (k) + bi ) tan (θj (k) + bj )
+
tan (θi (k) + bi ) − tan (θj (k) + bj )
tan (θi (k) + bi ) − tan (θj (k) + bj )
(9)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
where x̂ij (k) and ŷij (k) are the X and Y Cartesian estimates, respectively.
T
In addition, by defining the stacked covariance matrix in the bearing–only coordinate for the stacked measurement θi (k) θj (k)
as
"
#
σθ2i (k)
0
Rij (k) =
(10)
0
σθ2j (k)
T
one can calculate the covariance matrix of the transformed vector xij (k) yij (k)
as
−1
−1
xy
(k) =
J(k)T (Rij (k)) J(k)
Rij
(11)
T
xij (k) yij (k)
and
T
Ji (k) Jj (k)
J(k) =
where J(k) is the Jacobian matrix with respect to
(12)
Further, the elements of J(k) can be written as
#
"
(xij (k) − xi )
− (yij (k) − yi )
Ji (k) =
2
2 ,
2
2
(xij (k) − xi ) + (yij (k) − yi )
(xij (k) − xi ) + (yij (k) − yi )
(13)
and
"
Jj (k)
=
− (yij (k) − yj )
2
(xij (k) − xj ) + (yij (k) − yj )
2
,
#
(xij (k) − xj )
2
2
(xij (k) − xj ) + (yij (k) − yj )
(14)
If no bias estimation is needed, these pseudo–measurements in Cartesian coordinates can be used in a Kalman filter with their
associated covariance matrices to recursively estimate the target’s position [12].
III. B IAS M ODELING IN C ARTESIAN C OORDINATES
As discussed in Section II, bearing–only biases are in the form of the additive constants. Although additive constant biases
have already been dealt with in the case of radar measurements [42] and [43], it is not possible to generate similar pseudo–
measurements with bearing–only data directly. One way to formulate pseudo–measurements with biases is to model in Cartesian
coordinates. In this section, a step–by–step procedure to model the biases and to separate them from original track positions
in Cartesian coordinates is given. In Section IV, the pseudo–measurement generation is discussed in detail.
In Section II, the process of mapping from bearing–only measurements to Cartesian was given. In order to model the biases,
one can start with separating the bias terms in (8) and (9) from the original track position in Cartesian coordinates. This
separation of bias terms provides the necessary information to create a pseudo–measurement that properly addresses the biases
as in the case of radar measurements. Once the pseudo–measurements are generated, it is possible to estimate the biases and
remove them. The process of finding the bias terms that contribute to (8) and (9) starts with expanding the tan (·) function as
tan (α + β)
=
tan (α) + tan (β)
1 − tan (α) tan (β)
(15)
Applying (15) to (8) and (9), and separating the terms related to the bias from those related to the target position will give a
new set of equations to define the position of the target in Cartesian coordinates. To make the parameter separation easier to
follow, the common terms are defined and named first. For the common terms in x̂ij (k) and ŷij (k), one can define
D
=
tan (θi (k)) − tan (θj (k))
(16)
B
=
1 + tan (θi (k)) tan (θj (k))
(17)
Further, define the following for the terms in x̂ij (k):
Dx
= xi tan (θi (k)) − xj tan (θj (k)) + (yj − yi )
(18)
Bxi
= − (yj − yi ) tan (θi (k)) + xj tan (θi (k)) tan (θj (k)) + xi
(19)
Bxj
Bxij
= − (yj − yi ) tan (θj (k)) − xi tan (θi (k)) tan (θj (k)) + xj
(20)
= xj tan (θi (k)) − xi tan (θj (k)) + (yj − yi ) tan (θi (k)) tan (θj (k))
(21)
Similarly, for the terms in ŷij (k), define
Dy
= yj tan (θi (k)) − yi tan (θj (k)) + (xj − xi ) tan (θi (k)) tan (θj (k))
Byi
=
(xj − xi ) tan (θj (k)) + yi tan (θi (k)) tan (θj (k)) + yj
(23)
Byj
Byij
=
(xj − xi ) tan (θi (k)) − yj tan (θi (k)) tan (θj (k)) − yi
(24)
= yi tan (θi (k)) − yj tan (θj (k)) + (xj − xi )
(22)
(25)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
With these factorizations, bias terms can be separated from the target state values in Cartesian coordinates. It can be seen that
T
the vector x̂ij (k) ŷij (k)
can be written as1
u
xij (k) + βx (θi (k), θj (k), bi , bj )
x̂ij (k)
=
u
yij
(k) + βy (θi (k), θj (k), bi , bj )
ŷij (k)
(26)
where
xuij (k)
=
u
yij
(k)
=
Dx
D
Dy
D
(27)
(28)
and the bias terms can be written in as
βx (θi (k), θj (k), bi , bj )
βy (θi (k), θj (k), bi , bj )
=
=
Bxi tan (bi ) + Bxj tan (bj ) + Bxij tan (bi ) tan (bj )
D + D tan (bi ) + B tan (bj ) + D tan (bi ) tan (bj )
Dx B tan (bi ) − Dx B tan (bj ) + Dx D tan (bi ) tan (bj )
− 2
D + BD tan (bi ) − BD tan (bj ) + D2 tan (bi ) tan (bj )
Byi tan (bi ) + Byj tan (bj ) + Byij tan (bi ) tan (bj )
D + D tan (bi ) + B tan (bj ) + D tan (bi ) tan (bj )
Dy B tan (bi ) − Dy B tan (bj ) + Dy D tan (bi ) tan (bj )
− 2
D + BD tan (bi ) − BD tan (bj ) + D2 tan (bi ) tan (bj )
(29)
(30)
Note that in the above formulations it is assumed that the biases and the true bearings are available. In practice, only the noisy
or estimated values are available. Assuming small values for the bias and noise terms, one can use the above formulation
without significant loss in accuracy. Similar assumptions has been made in the previous works on bias estimation [31], [32],
[42]. A technical discussion on the range of bias and noise values for which the above formulation is valid is given in Section
VI.
IV. B EARING –O NLY T RACKING AND R EGISTRATION
Bearing–only sensor registration is a challenging problem in target tracking that has been addressed in [34], [37], [41].
In order to find the biases and correct the measurements, one should first look into the observability of the bias variables.
Note that in (26), provided that the target is not on the line that connects the two sensors used in the triangulation or in the
vicinity of one of the sensors, the state parameters are observable [4]. In addition, if there are two pairs of sensors tracking
the same target, the biases become observable as it is shown in IV-A. To estimate the biases decoupled from the state vector, a
pseudo–measurement that can address the bias vector directly must be defined. In this section, a new formulation is proposed
to create a pseudo–measurement that can be used for bias estimation with bearing–only data. The key requirement of this
method in order to ensure observability of all parameters is to have at least two sensor pairs in the surveillance region. In the
following, two practical scenarios that can be expanded to a more general formulation to handle varying number of sensors
and targets are discussed in detail.
A. Pseudo–measurement of bearing–only measurements
To handle practical bearing–only scenarios, two different cases are analyzed here. In each case, a separate pseudo–measurement
model is proposed along with its associated covariance matrix. The main idea is to use two different position approximations
to create a pseudo–measurement as discussed below.
1) Four–sensor (or any even number of sensors) case: Assuming that there are two pairs of sensors in the surveillance
region, a vector of nonlinear pseudo–measurements can be defined by subtracting the target positions based on the pairs (i, j)
and (m, n) as
x̂ij (k) − x̂mn (k)
zb (k) =
+ w(k)
(31)
ŷij (k) − ŷmn (k)
where w(k) is the additive zero–mean white Gaussian noise associated with the pseudo–measurement and its covariance matrix
is defined as
Rw (k)
1 The
superscript “u” indicates that the parameter is unbiased.
=
xy
xy
Rij
(k) + Rij
(k)
(32)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
Using the fact that in the absence of bias and noise terms, measurements from any two sensors point to the same target location
regardless of the sensor locations, the pseudo–measurement zb (k) can be written as
βx (bi , bj ) − βx (bm , bn )
zb (k) =
+ w(k)
(33)
βy (bi , bj ) − βx (bm , bn )
for two uncorrelated pairs of sensors. This can be applied to any even number of sensors.
2) Three–sensor (or any odd number of sensors) case: In this case, one must create two position approximations from
triangulation to be able to create a pseudo–measurement for biases. Since there are only three sensors, the possible pairs are
(i, j) and (i, m). Then, the pseudo–measurement can be approximated by
x̂ij (k) − x̂im (k)
zb (k) =
+ w(k)
(34)
ŷij (k) − ŷim (k)
where w(k) is approximately additive zero–mean white Gaussian noise associated with the pseudo–measurement and its
covariance matrix is defined as
xy
xy
= Rij
(k) + Rim
(k)
Rw (k)
(35)
Because of the correlation between the two tracks from three sensors in Cartesian coordinates, the noise is not white anymore
and this formulation is only an approximation.
As in the case for four sensors, the pseudo–measurement zb (k) can be written as
βx (bi , bj ) − βx (bi , bm )
zb (k) =
+ w(k)
(36)
βy (bi , bj ) − βx (bi , bm )
Note that for simplicity, the arguments θi (k), θj (k), θm (k) and θn (k) have been dropped from (33) and (36). With an arbitrary
odd number of sensors, one sensor need to be paired with two other, resulting in in a similar approximation. As for the more
general case of n arbitrary sensors, methods similar to [42] can be adopted to handle the situation.
B. Batch maximum–likelihood estimator
To apply a batch estimator for bias estimation, one needs to form a likelihood function. Following (35) and (33), and
assuming that the noise is white, zero–mean and Gaussian, the likelihood function of the bias parameters given two pairs of
sensors is
1
1
T
exp
−
z
(k)
−
h(b)
p (zb (k) | b) =
b
1
−
2
2π |Rw (k)| 2
−1
(Rw (k)) [zb (k) − h(b)]
(37)
where the nonlinear function h(b) of the bias vector is given by
βx (bi , bj ) − βx (bm , bn )
h(b) =
βy (bi , bj ) − βx (bm , bn )
(38)
Assuming independence over time, one can write the likelihood function over k = 1, ..., K as
p (Zb | b)
=
K
Y
p (zb (k) | b)
(39)
k=1
where
Zb
{zb (1), zb (2), . . . , zb (K)}
=
(40)
Finally, the vector b̂ that maximizes the likelihood function can be written as
b̂
=
arg max p (Zb | b)
b
(41)
The above assumes that there is only one target, but it can be extended to the multitarget case using the stacked parameter
estimator.
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
V. C RAM ÉR –R AO L OWER BOUND FOR B EARING ONLY BIAS ESTIMATION
This section is devoted to the calculation of the Cramér–Rao Lower Bound (CRLB) on the estimation accuracy of bias
parameters by using the pseudo–measurements introduced in Subsection IV-A. Note that based on (38), the measurement
equation for target t at time k is
t
βx (bi , bj ) − βxt (bm , bn )
t
hb (k) =
+ wt (k)
(42)
βyt (bi , bj ) − βxt (bm , bn )
Assuming that there are N targets in the surveillance region and that the bias parameters are constant over time k = 1, ...K,
the stacked measurement vector can be written as
Z
= h(b) + w
(43)
where
zb1 (1)T
=
h1b (1)T
=
Z =
h(b)
w
w1 (1)T
···
zbN (1)T
···
T
hN
b (1)
···
···
zb1 (K)T
···
h1b (K)T
wN (1)T
Further, the covariance matrix of the noise vector w is
1
R (1) · · ·
R = diag
RN (1)
···
···
···
···
w1 (K)T
···
R1 (K) · · ·
zbN (K)T
T
T
wN (K)T
RN (K)
(44)
T
T
hN
b (K)
(45)
(46)
(47)
The covariance matrix of an unbiased estimator b̂ is bounded from below by [4]
T
b̂ − b b̂ − b
E
≥ J−1
(48)
where J is the Fisher Information Matrix (FIM) given by
n
o
T
J = E [∇b log p(Y | b)] [∇b log p(Y | b)]
|b=b̃
n
o
T
= E [∇b λ] [∇b λ]
|b=b̃
(49)
in which b̃ is the true value of the bias vector b, p(Z | b) is the likelihood function of b, λ = − ln p(Y | b) and ∇ is
gradient operator. Based on the model for the stacked measurement vector in (43), one can define the Jacobian matrix of htb (k)
evaluated at the true value b [39] as
T
H̃bt (k) = 5b htb (k)T
(50)
Then, defining
H̃b
=
H̃b1 (1)T
···
H̃bN (1)T
···
H̃b1 (K)T
···
H̃bN (K)T
T
(51)
one can write
J
=
n
o
−1
H̃b
E H̃T
bR
(52)
VI. S IMULATION R ESULTS
To evaluate how the proposed method performs in practical scenarios, a series of simulations is presented in this section.
The implementation details on bias estimation, filtering and fusion are also discussed. Simulation results on different scenarios
are given with discussions on the advantages and disadvantages of the proposed bias estimator.
A. Motion models and measurement generation
A tracking scenario with four bearing–only sensors and sixteen targets is considered as shown in Figure 2. It is assumed
that all sensors are synchronized and that the bias ranges are between −0.05 rad and 0.05 rad. The standard deviation of
measurement noise is σθ = 0.0261 rad = 1.5◦ for target bearing measurements, which is higher than what was previously
assumed in the literature [6]. The true motion of the targets is modeled using the DCWNA or the NCV model [4] with
qx = qy = 0.001m2 /s3 . In the local trackers, DCWNA and Continuous Wiener Process Acceleration (CWPA) are used with
qx = qy = 0.1m2 /s3 to create different scenarios for the simulation. In all scenarios, the sampling time is T = 1s. To validate
the proposed bias model and to quantify its performance at different bias values, the biases are set to both positive and negative
values in different ranges as follows:
T
0.04 rad −0.02 rad 0.03 rad −0.02 rad
btest1 =
(53)
T
−0.04 rad −0.02 rad −0.03 rad −0.02 rad
btest2 =
(54)
T
0.04 rad 0.02 rad 0.03 rad 0.02 rad
btest3 =
(55)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
10000
Target locations
Sensor locations
8000
y (m)
6000
4000
2000
0
−2000
0
0.5
1
x (m)
1.5
2
x 10
4
Fig. 2: Initial locations of the targets and sensors
TABLE I: Parameter settings for the Genetic Algorithm
Parameter
Lower bound
Upper bound
Number of generations
Tolerance value
TABLE II: Comparison of
TABLE III: Comparison of
Value
−0.05 rad
0.05 rad
100
1 × 10−15
p
CRLB {[b]i } and RMSE of GA output for bias estimation of btest1 in a distributed system
Bias parameter
q
CRLB [b]i (rad)
RMSE (rad) of the GA bias estimate
bi
bj
bm
bn
0.071 × 10−3
0.071 × 10−3
0.537 × 10−3
0.462 × 10−3
3.32 × 10−3
3.28 × 10−3
3.92 × 10−3
1.45 × 10−3
p
CRLB {[b]i } and RMSE of the GA output for bias estimation of btest2 in a distributed system
Bias parameter
q
CRLB [b]i (rad)
bi
bj
bm
bn
10−3
0.0904 ×
0.0941 × 10−3
0.4243 × 10−3
0.3787 × 10−3
RMSE (rad) of the GA bias estimate
1.72 × 10−3
1.81 × 10−3
2.95 × 10−3
1.73 × 10−3
B. Bias estimation and the Genetic Algorithm
In this paper, the Genetic Algorithm (GA) [18] is used to solve the optimization problem in (41). The Genetic Algorithm is an
efficient optimization algorithm for highly nonlinear objective functions [40] that is widely used in different applications [18].
Note that although the GA is a batch ML estimator, the length of the window can be varied depending on user criteria to meet
the real–time requirements. The parameters used in the simulations are shown in Table I. The algorithms were implemented
TM
on a computer with Intel R Core i7-3720Qm 2.60GHz processor and 8GB RAM.
C. Bias estimation: Four–sensor distributed problem
In this scenario, all four sensors defined earlier are used to implement the GA. Four out of sixteen targets are used for
performance evaluation. AMRs or local bearing–only tracks are collected over 100 time steps and the GA is applied to the
whole data in batch mode. The GA is run with the settings in Table I and the final results are gathered after the termination of
the GA. Then, the estimated bias vector is used over 100 Monte Carlo runs to calculate the Root Mean Square Error (RMSE)
for
p comparison. As the benchmark, the CRLB is also calculated based on the derivations in Section V. The RMSE values and
CRLB {[b]i } of the ML estimates with the three different
p sets of bias parameters are shown in Tables II, III and IV.
Although there is a difference between the RMSE and CRLB {[b]i }, the RMSE results are nearly an order of magnitude
smaller than the bias values, which indicates that any correction made based on the estimated biases will result in better
position estimates. Note that the CRLB in bearing–only tracking problems can be overly optimistic and may even approach
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
TABLE IV: Comparison of
p
CRLB {[b]i } and RMSE of GA output for bias estimation of btest3 in a distributed system
Bias parameter
q
CRLB [b]i (rad)
RMSE (rad) of the GA bias estimate
10−3
0.0973 ×
0.0954 × 10−3
0.5347 × 10−3
0.4880 × 10−3
bi
bj
bm
bn
2.73 × 10−3
3.75 × 10−3
2.43 × 10−3
2.66 × 10−3
zero (i.e., perfect estimates) in a network of bearing–only sensors [23]. Thus, the difference between the theoretical CRLB
and the empirical RMSE is not of major concern.
To show how much the proposed bias estimation method can help in correcting the target tracks, another simulation is
conducted. In this simulation, it is assumed that the tracker has access to the final estimated bias vector (output of the GA)
and then a Kalman filter is run with the bias estimates in hand. To use the estimated bias parameters, one should, first, correct
the bearing–only measurements with the estimated values. The correction must be done both in the measurement vector and
its associated covariance matrix. Since the estimated biases do not have the covariance information, a scaled version of the the
calculated CRLB of the bias parameters is used instead. The scaling factor can be determined through experiments. Then, the
tracker can be run with these corrected measurements to find the position and velocity estimates of all targets in the surveillance
region. The position RMSE of the original tracks before correction, the RMSE of the corrected estimates and the Cramér–Rao
lower bounds are shown in Figures 3 and 4.
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 3: Position RMSE with distributed tracking for corrected and original tracks of target 2 (set 1)
As shown in Figures 3 and 4, for case 1, Figures 5 and 6 for case 2 and Figures 7 and 8 for case 3, the position error
is reduced significantly in terms of the position RMSE. This demonstrates the effectiveness of the bias model proposed and
the ML estimation algorithm, i.e., the GA, that is applied to the data. Although, according to the results, the correction factor
varies based on the sensor–target orientation, the corrected track and its associated covariance do follow the bias–free values
accurately, which demonstrates the capability of the proposed algorithm in estimating the biases. To process a batch of data
with K = 100, the computational time is 44.1s in MATLAB.
D. Bias estimation: Three–sensor distributed problem
To show the accuracy of bias estimation when there are only three sensors in the surveillance region, the same Genetic
Algorithm is used to solve the ML problem in (41). The primary issue with three sensors is that one of them, here sensor
i, is used to pair with both sensors j and m. This leads to correlation between the two tracks over time. Assuming that the
correlation is negligible, the same method is applied to estimate the biases. Further, for the case of three sensors, the bias
values are set to
T
0.04 rad 0.02 rad 0.03 rad
btest =
(56)
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 4: Position RMSE with distributed tracking for corrected and original tracks of target 3 (set 1)
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 5: Position RMSE with distributed tracking for corrected and original tracks of target 2 (set 2)
to be able to distinguish between the original and corrected tracks. The constraints on the lower and upper bounds of the
optimization algorithm are set to −0.05 rad and 0.05 rad, respectively. The tolerance value for the GA is also set to 1 × 10−15 .
Figures 9 and 10 show the result for position estimates in Cartesian coordinates.
The corrected tracks are sill better in terms of RMSE, which means that the estimated biases are accurate enough in spite
of the correlation.
E. Real–time window–based Genetic Algorithm
It is important for the proposed method to be able to work in real–time. For this purpose, the GA can be set to run, in each
iteration, for a specific window size or duration. The settings of the GA for real time scenarios are given in Table V.
The final estimates and the population matrix in one window can be used as the initial conditions for the next window.
Thus, the estimates of the biases can be used to correct the measurements at the end of processing each window of data. The
results of the simulations using this approach for the case of four sensors (test 1) are given in Figures 11 and 12.
As shown in Figures 11 and 12, the GA is still able to find the biases with a smaller window size and fewer generations.
Note that updating the biases with smaller window sizes enables the use of methods similar to [42], [43] for arbitrary number
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 6: Position RMSE with distributed tracking for corrected and original tracks of target 3 (set 2)
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 7: Position RMSE with distributed tracking for corrected and original tracks of target 2 (set 3)
TABLE V: Parameter settings for real time genetic algorithm
Parameter name
Lower bound
Upper bound
Number of generations
Tolerance value
Window size
Value
−0.05 rad
0.05 rad
50
1 × 10−15
10
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 8: Position RMSE with distributed tracking for corrected and original tracks of target 3 (set 3)
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 9: Position RMSE of corrected and original tracks for the three–sensor distributed tracking case (Target 2)
of sensors in the surveillance region. With this setting, biases can be updated every 9.25 s in MATLAB. This simple example
shows that even when the processing time is a crucial parameter in the design, one can still handle bias estimation in real–time
using the proposed method.
F. Bias estimation with false measurements: Four–sensor centralized problem
To demonstrate how the proposed method performs in the presence of false alarms and missed associations in a centralized
fusion framework where local sensors send all their measurements instead of AMRs or bearing–only tracks, a simulation is
presented in this section. The probability mass function of the number of false alarms or clutter points in surveillance volume
V as a function of their spatial density λ is defined as
m
(λV )
(57)
m!
where m is the number of false alarms [5]. With bearing–only measurements, the volume is V = 2π (rad). It is assumed
that the average number of false alarms per unit volume in a scan, i.e., λ, is 0.5. Also, it is assumed that PD = 0.7 for each
sensor. Note that in this centralized case, the local sensors send all measurements (rather than AMRs or local bearing–only
µFA (m)
= e−λV
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 10: Position RMSE of corrected and original tracks for the three–sensor distributed tracking case (Target 3)
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 11: Position RMSE for corrected and original tracks for the four–sensor distributed tracking case and window size of 10
(Target 2)
tracks) to the fusion node. Because the proposed method is a batch estimator and uses measurements from both sensor pairs to
create a pseudo–measurement for bias estimation, false tracks are often removed prior to generating the pseudo–measurement
vector, which then is sent to the bias estimator. Typically, false tracks do not exist for more than a few time steps as they
are dependent on all four sensors creating false alarms at the same time steps, in the same region, and for a reasonably long
interval of time.
To show the accuracy of bias estimation
p in a centralized system, the same Genetic Algorithm is used to solve the ML
problem in (41). The RMSE values and CRLB {[b]i } of the ML estimates with the bias parameters as defined in (53) are
shown in Table VI. Note that the CRLB values are optimistic because they do not factor in the false alarms or the missed
detections and that the ML estimator does not factor in the false alarms or the missed detections explicitly. A comprehensive
centralized bias estimator is under development. The focus of this paper is the decentralized one.
VII. C ONCLUSIONS
In this paper, a new mathematical model for bearing–only bias estimation in distributed tracking systems was proposed.
This model was based on triangulation using the associated measurement reports or local bearing–only tracks from different
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
3
Position RMSE (m)
10
RM SEOriginal
RM SECorrected
RM
p SEBiasf ree
√ ΣOriginal
pΣCorrected
ΣBiasf ree
2
10
1
10
20
40
60
Time Step k
80
100
Fig. 12: Position RMSE for corrected and original tracks for the four–sensor distributed tracking case and window size of 10
(Target 3)
TABLE VI: Comparison of CRLB and GA output for bias estimation of btest1 with measurement origin uncertainty in a
centralized system (λ = 0.5 and PD = 0.7)
Bias parameter
CRLB
GA bias estimate
bi
bj
bm
bn
0.2123 × 10−3
0.2123 × 10−3
0.7674 × 10−3
0.5779 × 10−3
7.419 × 10−3
12.85 × 10−3
6.143 × 10−3
4.352 × 10−3
sensor pairs. It was shown that the proposed bias model has the advantage of being practical in scenarios with multiple
sensors. In particular, the proposed algorithm is effective when the sensor noise level and bias values are high. In addition,
previously proposed algorithms were dependent on target–sensor maneuvers and/or limited to certain noise levels. The new bias
model can handle any type of target–sensor motion and it is effective against 5◦ of offset bias in each sensor and uncertainty
levels up to 1.5◦ of noise standard deviation, which is higher than what was assumed in the literature previously. Also, the
proposed method can handle false alarms and missed detections in a centralized architecture. That is, the proposed algorithm
is practical in scenarios with realistic sensor parameter values. Finally, a batch ML estimator was proposed to solve the bias
estimation problem along with simulation results. A comprehensive centralized bias estimation algorithm with data association
for bearing–only sensors is in progress.
R EFERENCES
[1] V. J. Aidala, “Kalman filter behavior in bearings–only tracking applications”, IEEE Transactions on Aerospace and Electronic Systems, vol. 1, pp.
29–39, 1979.
[2] V. J. Aidala and S. E. Hammel, “Utilization of modified polar coordinates for bearings–only tracking”, IEEE Transactions on Automatic Control, vol.
28, no. 3, pp. 283–294, 1983.
[3] M. S. Arulampalam, B. Ristic, N. Gordon, and T. Mansell, “Bearings–only tracking of manoeuvring targets using particle filters”, EURASIP Journal
on Applied Signal Processing, pp. 2351–2365, 2004.
[4] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navigation: Theory, Algorithms and Software, Wiley, NY,
2001.
[5] Y. Bar-Shalom, P. K. Willett, and X. Tian, Tracking and Data Fusion: A Handbook of Algorithms, YBS Publishing, Storrs, CT, 2011.
[6] D. Belfadel, R. W. Osborne, and Y. Bar-Shalom, “Bias estimation for optical sensor measurements with targets of opportunity”, 16th International
Conference on Information Fusion, pp. 1805–1812, Istanbul, Turkey, July 2013.
[7] D. Belfadel, R. W. Osborne, and Y. Bar-Shalom, “A minimalist approach to bias estimation for passive sensor measurements with targets of opportunity”,
Proc. SPIE Conference on Signal and Data Processing of Small Targets, Vol. 8857, 2013.
[8] D. A. Blank and A. E. Bock, Introduction to Naval Engineering, Naval Institute Press, Annapolis, MD, U.S.A, 2005.
[9] P. Braca, R. Goldhahn, G. Ferri, and K. LePage, “Distributed information fusion in multistatic sensor networks for underwater surveillance”, IEEE
Sensors Journal, 2015.
[10] P. Braca, P. Willett, K. Lepage, S. Marano, and V. Matta, “Bayesian tracking in underwater wireless sensor networks with port-starboard ambiguity”,
IEEE Transactions on Signal Processing, vol. 62, no. 7, pp. 1864–1878, April 2014.
[11] M. Bugallo, T. Lu, and P. Djuric, “Bearings–only tracking with biased measurements”, 2nd IEEE International Workshop on Computational Advances
in Multi-Sensor Adaptive Processing, pp. 265–268, 2007.
[12] D. E. Catlin, Estimation, Control, and the Discrete Kalman Filter, vol. 71, Springer Science & Business Media, 2012.
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
[13] Y. T. Chan and S. W. Rudnicki, “Bearings–only and Doppler–bearing tracking using instrumental variables”, IEEE Transactions on Aerospace and
Electronic Systems, vol. 28, no. 4, pp. 1076–1083, 1992.
[14] H. Chen and F. Lian, “Bias estimation for multiple passive sensors”, International Conference on Measurement, Information and Control (MIC), vol.
2, pp. 1081–1084, May 2012.
[15] S. Choi, D. F. Crouse, P. Willett, and S. Zhou, “Approaches to Cartesian data association passive radar tracking in a DAB/DVB network”, IEEE
Transactions on Aerospace and Electronic Systems, vol. 50, no. 1, pp. 649–663, 2014.
[16] R. L. Cooperman, “Tactical ballistic missile tracking using the interacting multiple model algorithm”, Fifth International Conference on Information
Fusion, vol. 2, pp. 824–831, Annapolis, Maryland, USA, 2002.
[17] M. Daun, U. Nickel, and W. Koch, “Tracking in multistatic passive radar systems using DAB/DVB-T illumination”, Signal Processing, vol. 92, no. 6,
pp. 1365–1386, 2012.
[18] L. Davis, Handbook of Genetic Algorithms, vol. 115, Van Nostrand Reinhold, New York, 1991.
[19] A. Farina, “Target tracking with bearings–only measurements”, Signal Processing, vol. 78, no. 1, pp. 61–78, 1999.
[20] S. Fortunati, A. Farina, F. Gini, A. Graziano, M. S. Greco, and S. Giompapa, “Least squares estimation and Cramér–Rao type lower bounds for relative
sensor registration process”, IEEE Transactions on Signal Processing, vol. 59, no. 3, pp. 1075–1087, 2011.
[21] G. Hendeby, R. Karlsson, F. Gustafsson, and N. Gordon, “Recursive triangulation using bearings–only sensors”, IEE Seminar on Target Tracking:
Algorithms and Applications, pp. 3–10, Birmingham, England, March 2006.
[22] M. A. Hopkins and H. F. Vanlandingham, “Pseudo-linear identification: Optimal joint parameter and state estimation of linear stochastic MIMO systems”,
IEEE American Control Conference, pp. 1301–1306, 1988.
[23] P. R. Horridge and M. L. Hernandez, “Performance bounds for angle–only filtering with application to sensor network management”, 6th International
Conference on Information Fusion, pp. 695–703, Cairns, Queensland Australia, July 2003.
[24] R. A. Iltis and K. L. Anderson, “A consistent estimation criterion for multisensor bearings–only tracking”, IEEE Transactions on Aerospace and
Electronic Systems, vol. 32, no. 1, pp. 108–120, 1996.
[25] K. Johansson, K. Jöred, and P. Svensson, “Submarine tracking using multi–sensor fusion and reactive planning for the positioning of passive sonobuoys”,
Hydroakustik, vol. 97, 1997.
[26] N. Z. Kolev, Sonar Systems, InTech, Croatia, 2011.
[27] T. R. Kronhamn, “Bearings–only target motion analysis based on a multihypothesis Kalman filter and adaptive own–ship motion control”, IEE
Proceedings–Radar, Sonar and Navigation, vol. 145, no. 4, pp. 247–252, 1998.
[28] B. La Scala and M. Morelande, “An analysis of the single sensor bearings–only tracking problem”, 11th International Conference on Information
Fusion, pp. 1–6, Cologne, Germany, June 2008.
[29] P. H. Leong, S. Arulampalam, T. A. Lamahewa, and T. D. Abhayapala, “A Gaussian–sum based cubature Kalman filter for bearings–only tracking”,
IEEE Transactions on Aerospace and Electronic Systems, vol. 49, no. 2, pp. 1161–1176, 2013.
[30] Z. Li, S. Chen, H. Leung, and E. Bosse, “Joint data association, registration, and fusion using EM–KF”, IEEE Transactions on Aerospace and Electronic
Systems, vol. 46, no. 2, pp. 496–507, 2010.
[31] X. Lin, Y. Bar-Shalom, and T. Kirubarajan, “Exact multisensor dynamic bias estimation with local tracks”, IEEE Transactions on Aerospace and
Electronic Systems, vol. 1, no. 40, pp. 576–590, 2004.
[32] X. Lin, Y. Bar-Shalom, and T. Kirubarajan, “Multisensor multitarget bias estimation for general asynchronous sensors”, IEEE Transactions on Aerospace
and Electronic Systems, vol. 41, no. 3, pp. 899–921, 2005.
[33] M. Mallick and T. Kirubarajan, “Multi-sensor single target bearing–only tracking in clutter”, NASA STI/Recon Technical Report N, vol. 3, pp. 15779,
2001.
[34] D. W. McMichael and N. N. Okello, “Maximum likelihood registration of dissimilar sensors”, The First Australian Data Fusion Symposium, pp. 31–34,
November 1996.
[35] D. Mušicki, “Bearings–only multi-sensor maneuvering target tracking”, Systems & Control Letters, vol. 57, no. 3, pp. 216–221, 2008.
[36] S. C. Nardone, A. G. Lindgren, and K. F. Gong, “Fundamental properties and performance of conventional bearings–only target motion analysis”, IEEE
Transactions on Automatic Control, vol. 29, pp. 775–787, 1984.
[37] N. Okello and B. Ristic, “Maximum likelihood registration for multiple dissimilar sensors”, IEEE Transactions on Aerospace and Electronic Systems,
vol. 39, no. 3, pp. 1074–1083, July 2003.
[38] Y. Qi, Z. Jing, and S. Hu, “Modified maximum likelihood registration based on information fusion”, Chinese Optics Letters, vol. 5, no. 11, pp. 639–641,
2007.
[39] B. Ristic, S. Arulampalam, and N. Gordon, Beyond the Kalman Filter: Particle Filters for Tracking Applications, vol. 685, Artech House, Boston, MA,
2004.
[40] K. C. Sharman and G. D. McClurkin, “Genetic algorithms for maximum likelihood parameter estimation”, International Conference on Acoustics,
Speech, and Signal Processing, pp. 2716–2719, Glasgow, Scotland, May 1989.
[41] Q. Song and Y. He, “A real–time registration algorithm for passive sensors with TOA and angle biases”, 3rd International Congress on Image and
Signal Processing, vol. 9, pp. 4170–4173, October 2010.
[42] E. Taghavi, R. Tharmarasa, T. Kirubarajan, and Y. Bar-Shalom, “A practical bias estimation algorithm for multisensor–multitarget tracking”, IEEE
Transaction on Aerospace and Electronic Systems, Accepted for publication, Ocotber 2015.
[43] E. Taghavi, R. Tharmarasa, T. Kirubarajan, and Y. Bar-Shalom, “Bias estimation for practical distributed multiradar-multitarget tracking systems”, 15th
International Conference on Information Fusion, pp. 1304–1311, Istanbul, Turkey, July 2013.
[44] B. Xu and Z. Wang, “Biased bearings–only parameter estimation for bistatic system”, Journal of Electronics (China), vol. 24, no. 3, pp. 326–331, 2007.
[45] B. Xu, Z. Wu, and Z. Wang, “On the Cramér–Rao lower bound for biased bearings–only maneuvering target tracking”, Signal Processing, vol. 87, no.
12, pp. 3175–3189, 2007.
[46] J. Yosinski, N. Coult, and R. Paffenroth, “Network–centric angle only tracking”, SPIE Optical Engineering+ Applications, pp. 74450O–74450O,
International Society for Optics and Photonics, 2009.
Accepted for publication in IEEE Transactions on Aerospace and Electronics Systems, March 2016
Ehsan Taghavi received the M.Sc. degree in communication engineering in 2012 from Chalmers University of Technology, Gothenburg, Sweden, where he worked on particle filter smoother. He is currently pursuing the Ph.D. degree in computational science and
engineering at McMaster University, Hamilton, Canada. His research interests include estimation theory, scientific computing, signal
processing, parameter estimation, mathematical modeling and algorithm design.
Ratnasingham Tharmarasa was born in Sri Lanka in 1975. He received the B.Sc.Eng. degree in electronic and telecommunication
engineering from University of Moratuwa, Sri Lanka in 2001, and the M.A.Sc and Ph.D. degrees in electrical engineering from
McMaster University, Canada in 2003 and 2007, respectively.
From 2001 to 2002 he was an instructor in electronic and telecommunication engineering at the University of Moratuwa, Sri Lanka.
During 2002-2007 he was a graduate student/research assistant in ECE department at the McMaster University, Canada. Currently
he is working as a Research Associate in the Electrical and Computer Engineering Department at McMaster University, Canada. His
research interests include target tracking, information fusion and sensor resource management.
Thiagalingam Kirubarajan (S’95M’98SM’03) was born in Sri Lanka in 1969. He received the B.A. and M.A. degrees in electrical
and information engineering from Cambridge University, England, in 1991 and 1993, and the M.S. and Ph.D. degrees in electrical
engineering from the University of Connecticut, Storrs, in 1995 and 1998, respectively.
Currently, he is a professor in the Electrical and Computer Engineering Department at McMaster University, Hamilton, Ontario.
He is also serving as an Adjunct Assistant Professor and the Associate Director of the Estimation and Signal Processing Research
Laboratory at the University of Connecticut. His research interests are in estimation, target tracking, multisource information fusion,
sensor resource management, signal detection and fault diagnosis. His research activities at McMaster University and at the University
of Connecticut are supported by U.S. Missile Defense Agency, U.S. Office of Naval Research, NASA, Qualtech Systems, Inc.,
Raytheon Canada Ltd. and Defense Research Development Canada, Ottawa. In September 2001, Dr. Kirubarajan served in a DARPA
expert panel on unattended surveillance, homeland defense and counterterrorism. He has also served as a consultant in these areas
to a number of companies, including Motorola Corporation, Northrop-Grumman Corporation, Pacific-Sierra Research Corporation,
Lockhead Martin Corporation, Qualtech Systems, Inc., Orincon Corporation and BAE systems. He has worked on the development of a number of engineering
software programs, including BEARDAT for target localization from bearing and frequency measurements in clutter, FUSEDAT for fusion of multisensor data
for tracking. He has also worked with Qualtech Systems, Inc., to develop an advanced fault diagnosis engine.
Dr. Kirubarajan has published about 100 articles in areas of his research interests, in addition to one book on estimation, tracking and navigation and two
edited volumes. He is a recipient of Ontario Premiers Research Excellence Award (2002).
Michael McDonald received a B.Sc (Hons) degree in Applied Geophysics from Queens University in Kingston, Canada in 1986 and
a M.Sc. degree in Electrical Engineering in 1990, also from Queen’s University. He received a Ph.D in Physics from the University
of Western Ontario in London, Canada in 1997. He was employed at ComDev in Cambridge, Canada from 1989 through 1992 in
their space science and satellite communications departments and held a post-doctoral position in the Physics department of SUNY
at Stony Brooke from 1996 through 1998 before commencing his current position as Defence Scientist in the Radar Systems section
of Defence Research and Development Canada, Ottawa, Canada. His current research interests include the application of STAP
processing and nonlinear filtering to the detection of small maritime and land targets as well as the development and implementation
of passive radar systems.
| 3cs.SY
|
Dense 3D Regression for Hand Pose Estimation
Chengde Wan1 , Thomas Probst1 , Luc Van Gool1,3 , and Angela Yao2
arXiv:1711.08996v1 [cs.CV] 24 Nov 2017
1
ETH Zürich 2 University of Bonn 3 KU Leuven
Abstract
The use of deep neural networks has become the norm for
hand pose estimation[43, 25, 51, 10].
We present a simple and effective method for 3D hand
pose estimation from a single depth frame. As opposed to
previous state-of-the-art methods based on holistic 3D regression, our method works on dense pixel-wise estimation.
This is achieved by careful design choices in pose parameterization, which leverages both 2D and 3D properties of
depth map. Specifically, we decompose the pose parameters
into a set of per-pixel estimations, i.e., 2D heat maps, 3D
heat maps and unit 3D directional vector fields. The 2D/3D
joint heat maps and 3D joint offsets are estimated via multitask network cascades, which is trained end-to-end. The
pixel-wise estimations can be directly translated into a vote
casting scheme. A variant of mean shift is then used to aggregate local votes while enforcing consensus between the
the estimated 3D pose and the pixel-wise 2D and 3D estimations by design. Our method is efficient and highly accurate.
On MSRA and NYU hand dataset, our method outperforms
all previous state-of-the-art approaches by a large margin.
On the ICVL hand dataset, our method achieves similar accuracy compared to the nearly saturated result obtained
by [5] and outperforms various other proposed methods.
Code is available online1 .
In standard hand pose estimation pipelines, depth maps
are almost always treated as images. This is especially
true for deep learning-based approaches, which heavily rely
on the machinery of (2D) convolutional neural networks
(CNNs). One line of work for 3D hand poses estimation is
holistic regression, that is aiming to directly map the depth
images to 3D pose parameters such as joint angles or 3D coordinates. It bypasses having to solve for intermediate representations such as 2D coordinates and is able to capture
global constraints and correlations among different joints.
However, regressing from highly disparate domains such as
image and pose is a very challenging learning task. Furthermore, holistic regression cannot generalize to combinations of local evidence such as different individual finger
poses and suffers from translational variance and sensitivity
to hand bounding box locations.
CNNs have been successfully applied to 2D body pose
estimation[23, 44, 47]; in particular, fully convolutional networks (FCNs) can perform pixel-wise joint detection very
accurately [23, 47]. This is formulated as a pixel-wise classification of each pixel being the location of a joint. As
such, a second line of work in pose estimation tries to create analogous networks for detecting joints in 2D. Through
pixel-wise classification, joint detection can exploit local
patterns more explicitly than holistic regression, helping the
network to learn better feature maps. The 2D detections
and 3D regression can then be combined with a multi-task
setup [32, 30, 17, 42], either by feeding the 2D detection
heat map as an input to a 3D regression network, or by sharing the feature maps between detection and holistic regression. However, there is no guarantee that the regressed 3D
joints, if they were to be projected back to 2D, will be in
consensus with the original 2D detection heat-map. Moreover, by design, the aforementioned drawbacks of holistic
regression are still not eliminated with this line of work.
Other works in 2D detection apply inverse kinematics and
use a model-based optimization; however, the severe selfocclusion of the hand creates ambiguities which are diffi-
1. Introduction
Vision-based hand pose estimation has made significant progress in recent years. The increased performance
can be attributed to two dominating trends: depth imaging and deep learning. First of all, hand pose estimation
techniques have shifted almost entirely to using only depth
inputs[33, 31, 27, 52] since commodity depth sensors such
as the MS Kinect and Intel Realsense have become widely
available. As a 2.5D source of information, depth significantly resolves much of the ambiguities present in monocular RGB input. Secondly, deep learning has fundamentally
transformed the way that vision problems are being solved.
1 https://github.com/melonwan/denseReg
1
cult to resolve and as such, suffers from accuracy problems
which are otherwise not present in body pose estimation.
Despite all the drawbacks of working in 2D, we do not
want to directly solve a discrete volumetric detection problem with a 3D CNN. This becomes very parameter-heavy
and as a result, severely limits the working resolution[11,
29, 21]. Moreover, as input depth maps are only 2.5D, 3D
CNNs struggle to resolve the ambiguities caused by the selfocclusion common in hand poses.
At the core of the problem is the mismatch between 2.5D
depth data and traditional CNNs, be it in 2D or 3D. By
treating depth maps as a 2D image, we can leverage the
advances of CNNs, but we still under-utilize the information present. Yet we also want to avoid converting depth
information to a volumetric representation due to the computational overhead and the associated ambiguities. To that
end, we propose a combined pixel-wise detection and dense
regression method for hand pose estimation. Our proposed
method enjoys the benefits of 2D FCN-based detection such
as translational-invariance and generalization to different
finger gesture combinations. At the same time, dense regression allows us to make 3D estimates and benefit from
the merits of holistic 3D regression, such as accounting
for correlations and skeleton constraints, without having to
work in the discrete volumetric domain.
We make two careful design choices in parameterization to stabilize our training and improve regression robustness. First, we work with offsets instead of absolute
joint positions, i.e. we regress each pixel to a 3D offset
of each joint. Joint offsets have been used in previous
works[33, 37, 40, 46] and offer invariance towards translation. It also allows us to keep the original spatial resolution
in spite of pooling operations in the CNN. Secondly, we reparameterize the 3D offsets as a heat map and directional
unit vector, leading naturally to a joint detection and regression problem to solve for the two respectively. This form of
parameterization leverages both the 2D and the 3D geometric properties of a 2.5D depth map. For a given depth map,
we use a 2D CNN to capture local surface patterns but also
treat the depth map as a set of 3D points to arrive at a final
pose estimate in 3D.
To do so, we first extend the 2D detection heat map into
3D, i.e., value of the heat map is inversely proportional to
the 3D distance of corresponding point on the depth map to
a specific joint. In addition, we predict unit vector fields,
where each vector field corresponds to the direction from
the point on the depth map to a certain finger joint. Finally,
we also detect the joints in 2D, in the form of a projected
heat map. We aggregate all of the estimates together with
the mean shift algorithm into a global estimate with consensus between the 2D and 3D estimates.
The proposed method is highly accurate and out performs all previous state-of-arts on three publicly available
datasets, i.e., NYU[43], ICVL[39] and MSRA[37]. We also
compare our method against several baselines that combine
holistic regression with 2D joint detection. In these experiments we observe that, unlike in the case of full body pose
estimation, those combination strategies can hardly improve
holistic regression and are less accurate than our proposed
method by a large margin. We attribute this to the depth
ambiguity caused by self occlusion, towards which our proposed method is much more robust.
Our contribution can be summarized as follows:
• we formulate 3D hand pose estimation as a dense regression through a pose re-parameterization that can
leverage both 2D surface geometric and 3D coordinate
properties;
• we provide a non-parametric post-processing method
aggregating pixel-wise estimates to 3D joint coordinates; this post-processing explicitly handles the holistic estimation and ensures consensus between the 2D
and 3D estimates;
• we implement several baselines to investigate fusion
strategies for holistic regression and 2D joint detection
in a multi-task setup; such an analysis has never carried out before for hand pose estimation and provides
valuable insights to the field.
2. Related Works
Coupling 2D joint detection with 3D estimation 3D
pose estimation based on 2D observations has a long history
in computer vision. Early works [35, 50, 34] are mainly
based on low level visual cues, e.g., silhouette or optical
flow, and use generative models to resolve the depth ambiguity. More recent works have shifted towoards mid- and
high-level features, e.g. 2D joint detection heat maps or representations from CNNs, due to the availability of highly
accurate 2D joint detectors [23, 47]. One line of work [32,
30, 17, 42] formulates 3D pose estimation as a regression
problem and couples 2D joint detection and 3D regression
in a multi-task setup. Others [2, 22, 4, 19, 53, 43, 51, 10]
treat 3D estimation as an model-based optimization on top
of the 2D joint detections.
Our approach is similar to [32, 30, 17, 42] in that both
2D and 3D estimations are performed in a multi-task setup.
However, rather than using a holistic 3D regression, we perform pixel-wise 3D estimation. This type of fusion scheme
is translation invariant and can better generalize to different combinations of finger gestures. Like many others, we
also use a post-processing, but ours is much simpler with
negligible effort when compared to the computationally expensive energy minimization of [2, 22, 19, 43, 51, 10], nearest neighbour search [4], to employing an additional neural
network [19, 53].
2
Pose Parameterization Skeleton models don’t necessarily need to be parameterized with 3D joint coordinates.
Many works have modelled pose parameters in other spaces
to better exploit the skeleton structure. For example, [25,
45] learn a latent space to model the correlation among
different joints, while [49, 54, 36] parameterize pose hierarchically, i.e., location of child joint is dependent on its
parent joint along the skeleton tree, to leverage dependencies in the skeleton. [22] models skeleton as distance matrix
among different joints and [28, 3] formulate pose parameters as heat maps together with offset vector fields to handle
multiple instances 2D detection. Ours is inspired by [28, 3]
whereas we work on 3D estimation.
forms of supervision, with all the estimated results being
fed into the next stage to boost pose accuracy. We adopt
mean shift (Sec 3.3) to aggregate the pixel-wise regression
estimates, while enforcing the 2D projections of the final
estimated 3D joints to be in consensus with the pixel-wise
2D joint detections.
3.1. Pose Parameterization
Instead of directly regressing 3D joint coordinates from
the depth map, like most other regression-based methods [25, 26, 11, 45, 24, 12, 5, 6], we want to estimate an offset vector between depth points and hand joints. This makes
the estimate translation-invariant and also generalizes better
to different combinations of finger poses. However, directly
regressing the 3D offset vector field is non-ideal. First of
all, the regression for points that are far from a given hand
joint will result in offset vectors with large norms that dominate the training loss. Furthermore, far away hand joints are
beyond the scope of the receptive field of the convolutional
filters anyway. As such, we decompose the 3D offset vector
into two components – a 3D heat map S, estimated via detection, and a directional unit vector, V , via regression, as
follows:
(
θ − kp − pj k2
kp − pj k2 ≤ θ,
Sj (p) =
(1)
0
otherwise;
( p−p
j
kp − pj k2 ≤ θ,
(2)
Vj (p) = kp−pj k2
0
otherwise.
Hand Pose Estimation We limit our discussion to deep
learning-based methods and refer the reader to [38] for a detailed review of other model-based and random forest-based
methods. Deep learning-based methods fall into two camps:
two-stage approaches[43, 10, 51] with 2D joint detection
followed model-based optimization versus single-stage approaches [25, 26, 11, 45, 24, 12, 5, 6, 9] of holistic pose regression. The current best-performing methods [5, 12, 24]
are all single stage, most likely due to the effective exploitation of joint correlations. Our method takes the advantages
from both camps and well exploits the 2D and 3D properties
of depth maps.
Offset Regression and Hough Voting Several previous
works have successfully employed offset regression for localization and pose estimation tasks [33, 37, 40, 46, 16].
Due to their local nature, they offer invariance to translation
and their compatibility for bottom-up estimation. However,
these methods typically rely on hand-crafted features, with
the exception of works on 2D localization [20, 48]. In this
work, we extend this idea by learning dense 3D offset regression end-to-end.
where p ∈ R3 and pj ∈ R3 are the 3D coordinates of a
point from the depth map and of joint j respectively. θ defines the radius of a 3D ball centered at the joint position
that establishes a candidate region(see Fig. 1) from which
we consider support. The 3D heat map Sj (p) can be regarded a direct extension of the 2D heat map.
In addition, we estimate the joint’s 2D projection as a
heatmap R,
(
τ − kΠ(p) − Π(pj )k2
kΠ(p) − Π(pj )k2 ≤ τ
Rj (p) =
,
0
otherwise
(3)
where Π(·) denotes the 2D perspective projection function
and τ is the radius of the candidate disk. Even though Eq. 1
and 2 are sufficient to recover the 3D joint location, the
over-complete estimation with the 2D projection adds robustness to the local estimate. The 2D projection can be
combined with the 3D joint estimate with non-parametric
methods, which we elaborate in Section 3.3.
3. Method
We leverage both the 2D and 3D properties of a depth
map to formulate hand pose estimation as a pixel-wise regression problem. From a 2D perspective, we treat the
depth map as a 2D surface embedded in 3D and use a convolutional neural network(CNN) composed of 2D convolutional layers to capture surface local geometric patterns.
From a 3D perspective, the depth map can also be regarded
as a set of 3D points. It is for this set of points that we want
to estimate offsets to the hand joints. More specifically, we
use a CNN to estimate a dense vector field of offsets for
each joint of hand. We re-parameterize the joint offset as
a 3D heat map and a directional unit vector and solve for
the two via detection and regression respectively (Sec 3.1).
The resulting network is fully convolutional and compatible
with current joint detection network architectures (Sec 3.2).
Several networks can be stacked together as intermediate
3.2. Network architecture
The architecture of the detection and regression network
is shown in Fig 1. We use the hourglass network[23] as the
3
Algorithm 1 Mean-shift estimation of one joint
backbone because it is highly efficient, though any other
joint detection network architecture, e.g. [47, 28] could potentially be used. The 2D and 3D joint heat maps and the
unit vector fields are estimated by network cascades in a
learning multi-task manner. Specifically, for J joints, the
network first outputs 2D and 3D joint heat maps with two
separate sliding pixel-wise fully-connected layers on top of
the output feature map of the hourglass module. Since the
unit vector field Vj is correlated with the heat map estimates, we concatenate the heat maps together with the hourglass output feature map to determine the unit vector field.
To handle the discontinuity of 3D heat map and unit vector
field regressions at surface edges, the initial depth map is
also provided via concatenation to the input of the 3D heat
map network. Similar to [8, 13], we multiply the binarized
depth map as a mask with feature map and concatenate it
with the initial feature map. This serves as the input for our
unit vector field regression component.
Following the paradigm of [23], we stack together several modules with identical architectures to increase the
learning power. Estimates from previous modules are used
as inputs to the subsequent ones, while intermediate supervision is applied at the end of each module. Specifically, we
define a L2 loss over the J joints from T stacks as follows:
L=
=
T
X
t=1
(t)
(t)
(t)
t=1 j=1
(t)
Input:
D ∈ Rh×w×3 ∈ R
. input point cloud coordinates
outputs from neural network:
R ∈ Rh×w×1
. 2D heat map, see Eq. 3
S ∈ Rh×w×1
. 3D heat map, see Eq. 1
V ∈ Rh×w×3
. 3D offset unit vector field, see Eq. 2
P = D + θ(1 − S) V
. recover the joint coordinate
I = topK(S) ∈ N K×2
. select top K values’ indices
P = P (I) ⊂ R3
. fetch estimated 3D joint coordinates
P2d = {Π(p)|∀p ∈ P} ⊂ R2
. 2D projection
W = R(P2d ) ⊂ R
. fetch corresponding 2D heat map
values as weights
6: p = init(W, P) ∈ R3
. Initialization
7: for n in NPdo
2
1:
2:
3:
4:
5:
8:
(t)
(t)
kRj − Rj∗ k2 + kSj − Sj∗ k2 + kVj
Ppi ,wi ∈(P,W )
K(pi −p)wi pi
pi ,wi ∈{P,W } K(pi −p)wi
−
. K(x) = e
kxk
2σ 2
i.e. K = 5 and the kernel width σ as 40mm based on ablative
analysis.
− Vj∗ k2 ,
where
represent the respective ground-truth 2D
heat maps, 3D heat maps and vector offsets of joint j and
Rjt , Sjt , Vjt are corresponding estimates from tth stack.
Rj∗ , Sj∗ , Vj∗
3.3. Inference
During inference, we aggregate all of the pixel-wise estimated evidences into holistic 3D joint coordinates with the
mean shift algorithm. By design, this process explicitly ensures consensus between the joint detections in 2D and 3D.
Since each joint is estimated with the same mean shift process, we omit the joint index j in this section for simplicity.
As shown in Alg. 1, the N nearest points to the joint are selected based on the estimated 3D distance. We only select
K because points with larger estimated 3D distances tend
to amplify the estimation error of offset direction and thus
degrade the recovered 3D joint position estimation.
In addition, we provide a more efficient “unweighted”
approximation to Algorithm 1 without the 2D projection
(step 4) and replace the weights with (1 + R) S 2 . Table
1 shows that both strategies have nearly identical results. In
practice, we choose 5 nearest points as input to mean shift,
2
p←
9: end for
10: Output: p
(4)
LR + LS + LV
T X
J
X
predefined constants:
θ
. 3D distance threshold between point from D to joint
K
. number of points selected as input to mean shift
σ
. kernel width of mean shift kernel function
N
. number of mean shift iterations
3.4. Implementation Details
The network is implemented with Tensorflow[1] and optimized using the Adam [15] with the initial learning rate
set to 0.001 and the exponential decay rate of the momentum β1 = 0.5. Following [11, 24], we randomly rotate the
input depth map and change the aspect ratio for data augmentation. The batch size is set as 40 and we use batch
re-normalization to accelerate training, which works better
on small training mini-batches compared to batch normalization [14]. During testing, we use two network stacks and
have an average run time of 36ṁs per image (27.8ḞPS) on
a single NVIDIA Titan X GPU card.
4. Experiments
We conduct experiments on 3 publicly available datasets,
i.e. NYU[43], MSRA[37] and ICVL[39]. We choose the
NYU dataset to conduct ablation experiments and compare
against the baseline methods since it has a wider coverage
of hand poses as opposed to the other two.
We quantitatively evaluate our method with two metrics:
mean joint error (in mm) averaged over all joints and all
frames, and percentage of frames in which all joints are below a certain threshold [41]. Qualitative results of the estimated hand poses are shown in Fig. 5 and Fig. 8.
denotes element-wise multiplication
4
L(2)
L(1)
2D heat map
Hourglass
Module
7×7 3×3 2×2 3×3 3×3
C
R
P
R
R
Element-wise
Element
wise add
3×3
R
1×1
C
Concatenate
3D heat map
stride=2
Element-wise mul unit vector field
Figure 1. Network architecture. The abbreviations C, P, R stands for convolution layer, pooling and residual module respectively. We
choose 128*128 as size of input depth map and 32*32 as the input and output resolution of hourglass module[23] with 128 feature channels
in each layer. In this paper, we use 2 stacks due to real-time performance constraint. The network estimate 2D,3D heat maps and unit
vector field for each joint, we only show the pinky tip point here. Figure is best viewed in colour.
4.1. Baseline methods
we design an alternative method using the identical network architecture as detection+coordinate regression(see
Fig. 2 (b)) except for the output layer. Instead of regressing (x, y, z) as per baseline 2, we regress only the z coordinate, and refer to this as baseline 3 (detection+depth
regression). This output in the z axis is then combined with
the 2D detection results which are used directly as the coordinates for x, y plane.
Surprisingly, detection+depth regression outperforms
detection+coordinate regression both in terms of the average joint error and the percentage of frames below the error
threshold from 20 to 50 mm (see Fig. 3). This suggests that
2D detection provides a more accurate estimate than coordinate regression. We conclude that it should be beneficial
to explicitly enforce some form of consensus between the
3D estimates and the 2D detections. While the accuracy of
this baseline is still lower than our proposed approach by a
large margin (see Fig. 3), it shows that treating depth maps
as 2D images and using CNNs for holistic depth regression
is not enough to resolve the depth ambiguity in 3D hand
pose estimation.
Besides fusing the 2D detection heat maps as input for
coordinate regression, a second line of work [43, 10, 51]
conducts model-based tracking based on inverse kinematics
to recover the 3D pose. We compare against previous stateof-the-art methods [43, 10, 51] based on such a strategy and
out-perform all of them (see Section 4.3). This validates
the effectiveness of our proposed method in handling depth
ambiguities arising from the severe self occlusions in the
hand.
In this section we analyze whether regression of 2D joint
detections helps 3D regression and how different strategies
to fuse 2D joint detections and 3D regressions impact the
final pose accuracy. In addition, we also show the influence
of choosing different 3D offset parameterizations.
Does 2D joint detection help with 3D regression? First,
we would like to find out if 2D joint detection actually is
helpful 3D holistic regression. To that end, we design two
baseline methods: directly regressing 3D joint coordinates
versus coupling 2D joint detection and 3D regression in a
multi-task setup. Specifically, for baseline 1 (coordinate
regression), the regression network follows the architecture
from Fig. 2(a) which takes a depth map as input and directly outputs 3D joint coordinates. For baseline 2 (detection+coordinate regression), we adopt a similar regression
network architecture (see Fig. 2(b)) but add an hourglass
module[23]. We feed the depth map, the feature map from
the hourglass module, and the 2D joint detection heat map
all concatenated together as input into the brown module in
Fig. 2,and train for regression. Furthermore, to ensure a fair
comparison to our proposed method, we also stack two of
such networks together for baseline 2.
As is shown in Fig. 3, there is only a minor improvement
of 0.16mm in terms of the average joint error from direct
coordinate regression to detection+coordinate regression.
Furthermore, both methods perform similarly when the error threshold is larger than 25mm. We conclude that while
2D detection may help in learning a better feature map, coupling 2D detection together with 3D regression does not
solve the inherent problems of 3D regression, e.g., translation variance and inability to generalize through combining
local evidence.
Impact of offset re-parameterization We implement a
network which directly regresses the 3D offset without reparameterization into the 3D heatmap and directional unitvector as baseline 4 (mask loss)). As is shown in Fig. 2 (c),
the offset regression architecture follows exactly the same
structure as the offset unit direction regression in Fig. 1.
Impact of fusion strategies To further explore better
strategies for fusion of 2D detection and 3D regression,
5
Percentage of frames with all joints error within D
L
(a)
L(1)
100%
90%
80%
70%
60%
50%
40%
ours(10.21mm)
coordinate regression(15.04mm)
detection+coordinate regression(14.88mm)
detection+depth regression(13.91mm)
mask loss(12.56mm)
30%
20%
10%
0%
0
10
20
30
40
50
60
70
80
Max Allowed distance to GT D(mm)
Figure 3. Comparison with baselines. We compare our approach
to four baseline methods (Sec. 4.1) on the NYU dataset[43]. Number in the parenthesis of the legend indicates the average 3D error
of the corresponding method.
(b)
L(1)
Given the insights drawn from these baseline experiments, we attribute the high accuracy achieved by our
method to the reparameterization and imposing the loss to
all points for vector field regression. Decomposing the 3D
offsets into the joint 3D heat map and offset direction and
regressing the two in a cascaded way is easier to learn than
directly regressing the offsets. Secondly, setting the offset
vector to zero for outlier points instead of excluding them
from the loss makes the estimation more robust to errors in
regression during testing.
(c)
Hourglass
Module
1× 1
3× 3
Element-wise
El
i add
dd
C
R
Concatenate
7× 7 3× 3 2× 2 3× 3 3× 3
C
R
P
R
Element-wise mul
R
4.2. Exploration studies
stride=2
3 × 3 3 × 3 2 × 2 3 × 3 3 × 3 1024 1024
R
R
P
R
R
FC
FC
N
FC
We first experiment on the number of stacked networks
and the hyperparameters of mean-shift, i.e., the number of
selected candidate points as input to the mean shift and the
kernel width. As indicated in Table 1, we find that the proposed method is quite robust to the mean shift hyperparameters. On the other hand, the number of network stacks is
critical to the estimation accuracy. We test only up to 2
stacks to maintain real-time performance; however, as already shown in [23, 47], adding more stacks could improve
the accuracy.
In addition, as shown in the last two rows of Table 1, the
un-weighted mean shift approximation has a similar accuracy as the weighted version, with only 0.09mm difference
with respect to the mean joint error. As such, we choose 2
stacks and 5 candidate points as input to mean shift, kernel
width σ = 40mm and weighted mean shift as described in
Alg. 1 in the following experiments.
(d)
Figure 2. Baseline network architectures. (a) Direct 3D coordinate regression from depth map (baseline 1); (b) Network regresses 3D joint coordinates (baseline 2) or z-axis coupled 2D
joint detection (baseline 3) together with the 2D joint detection
heatmaps; (c) Regressing 3D offset vector field by masking the
loss with the 3D distance to joint (baseline 4); (d) Detailed architecture configurations. The abbreviations C, P, R, FC stands for
convolutional layer, pooling, residual module, and fully connected
layer respectively. For (b) and (c), we experiment with a stack of
2 in the same way as the proposed method for fair comparison.
We use the 2D detection scores to select candidate points
as inputs to the mean shift. In this baseline, we apply a 3D
distance threshold to the loss function of the offset, as was
done in [28], and effectively masks the regression so that we
only regress a joint’s neighbour points. Ideally, this baseline should be conducted without masking, but the training
failed completely, with the loss oscillating back and forth
without decreasing. As is shown in Fig. 3, pixel-wise dense
estimation out-performs holistic regression method and validates the benefits of regressing point-wise 3D offsets.
4.3. Comparison to state-of-the-art
NYU Dataset The NYU hand dataset [43] contains over
72K training and 8K testing frames. Its wide coverage of
hand poses and noisy input depths make this dataset quite
challenging. Since the hand region is not cropped out, we
use an hourglass joint detector [23] with one stack to locate
6
# Stacks T
Network parameters
1
11.20
Method
Xu et al. [49] (Lie-X)
Wan et al. [45] (Crossing Nets)
Oberweger et al.[24] (DeepPrior++)
Guo et al.[12] (REN)
Chen et al.[5] (Pose Guided)
Ours
2
10.21
Mean-shift parameters
1
5
10
30
# Candidates K
10.6
10.21
10.21
10.96
10mm 40mm 80mm 100mm
Kernel width σ
10.35
10.21
10.21
10.21
weighted
unweighted
Weights
10.21
10.26
Average 3D error
14.5mm
15.5mm
12.3mm
12.7mm
11.8mm
10.2mm
Table 2. Comparison with state-of-the-art on NYU. We report
average 3D error on the NYU[43] dataset.
Table 1. Impact of hyperparameters. We report the mean 3D
error (in mm) averaged over all joints and all frames on NYU
dataset[43]. We choose 2 stacks and 5 nearest points as input to
mean shift, kernel width σ = 40mm and weighted mean shift as
described in Alg. 1 as the default parameters.
Percentage of frames with all joints error within D
(a)
100%
90%
80%
70%
60%
50%
Ge et. al.(CVPR 2017)
Xu et. al.(IJCV 2017)
Wan et. al.(CVPR 2017)
Oberweger et. al.(ICCVW 2017)
Guo et. al.(ICIP 2017)
Chen et. al.(arXiv 2017)
ours
40%
30%
20%
10%
0%
0
10
20
30
40
50
60
70
(b)
Figure 5. Qualitative results. Hand pose estimation results on
NYU dataset[43]. (a) Successful samples with largest joint error
below 20mm; (b) Failed samples (top row) and the corresponding
ground-truth(bottom row).
80
Max Allowed distance to GT D(mm)
Figure 4. Comparison with state-of-the-art on NYU [43]. We
plot the percentage of frames in which all joints are below a threshold.
MSRA Dataset The MSRA hand dataset[37] contains
76.5K images from 9 subjects with 17 hand gestures. Following the protocol of [37], we use a leave-one-subjectout training / testing split and average the results over
the 9 subjects. We compare our methods with state-of-art
methods[11, 37, 46, 45, 24, 12, 5]. Specifically, [46, 37] are
based on the hierarchical regression forest. Similar to our
approach, [46, 37] regress 3D offsets and aggregate local estimations with the mean-shift algorithm. [11, 45, 24, 12, 5]
are CNN based 3D holistic regression methods and outperforms other existing methods[10, 18].
Again, our method outperforms all state-of-the-art by
a large margin both in terms of percentage of successful
frames (see Fig. 6) and average joint error (see Tab. 3). As
is shown in Fig. 6, over 81% and 91% of frames have joint
errors below 20mm and 30mm. This is a huge improvement
over the most accurate existing results from [5], which has
only 60% and 81% respectively. The qualitative results is
shown in Fig. 8(b).
the joints and take the median of estimated x and y coordinates over all joints respectively as the center point for
cropping out the hand region. We only use view 1 for both
training and testing and evaluate on a subset of 14 joints as
in [43] to make a fair comparison.
We compare our method to the most recently proposed
methods [49, 45, 24, 11, 12, 5]. All are 3D regression-based
methods with sophisticated network architectures and surpass earlier works [43, 51, 40, 7, 25, 26, 54] by a large margin. As is shown in Fig. 4 and Tab. 2, our method outperforms all these state-of-the-art methods with a large margin for both metrics. Specifically, according to Fig. 4, our
method significantly increases the percentage of successfully estimated frames by 8% (from 50% to 58%) on the
error threshold of 20mm and by 9.2% (from 70% to 79.2%)
on 30mm when compared to most accurate methods published to date ( [5] and [24, 12, 5] respectively). We also
show qualitative results on Fig. 5. The main reasons for
the failure cases are severe self occlusions and noise in the
depth map.
ICVL Dataset The ICVL hand dataset[39] has 22K
frames for training and 1.5k for testing. An additional 160k
augmented frames with in-plane rotations are provided by
7
Percentage of frames with all joints error within D
Method
Wan et al.[45] (Crossing Nets)
Wan et al. [46] (Surface Normal)
Sun et al. [37] (Cascaded Regression)
Oberweger et al.[24] (DeepPrior++)
Guo et al.[12] (REN)
Chen et al. [5] (Pose Guided)
Ours
100%
90%
80%
70%
60%
Ge et. al.(CVPR 2017)
Sun et. al.(CVPR 2015)
Wan et. al.(ECCV 2016)
Wan et. al.(CVPR 2017)
Oberweger et. al.(ICCVW 2017)
Guo et. al.(ICIP 2017)
Chen et. al.(arXiv 2017)
ours
50%
40%
30%
20%
10%
0%
0
10
20
30
40
50
60
70
Average 3D error
10.2mm
8.2mm
9.9mm
8.1mm
7.5mm
6.8mm
7.3mm
Table 4. Comparison with state-of-art on ICVL[39] dataset.
80
Max Allowed distance to GT D(mm)
Figure 6. Comparison with state-of-the-art on MSRA[37]. We plot
the percentage of frames in which all joints are below a threshold.
Method
Ge et al. [11] (3D CNN)
Wan et al.[45] (Crossing Nets)
Oberweger et al.[24] (DeepPrior++)
Guo et al.[12] (REN)
Chen et al. [5] (Pose Guided)
Ours
Average 3D error
9.5mm
12.2mm
9.5mm
9.8mm
8.6mm
7.2mm
(a)
Percentage of frames with all joints error within D
Table 3. Comparison with state-of-art on MSRA [37]. We plot
the percentage of frames in which all joints are below a threshold.
100%
90%
80%
(b)
70%
Figure 8. Qualitative results. Hand pose estimation results from
(a) ICVL[39], (b) MSRA[37].
60%
50%
40%
Tang et. al.(ICCV 2015)
Wan et. al.(CVPR 2017)
Oberweger et. al.(ICCVW 2017)
Guo et. al.(ICIP 2017)
Chen et. al.(arXiv 2017)
ours
30%
20%
10%
0%
0
10
20
30
40
50
60
70
5. Conclusion and discussion
We propose a highly accurate method for 3D hand pose
estimation from single depth map inputs. Given a depth
camera frame, we decompose 3D pose parameters into a
set of 2D/3D joint heat maps and 3D unit vector fields of
offset directions. This reparameterization allows us to consider both the 2D and 3D properties of the depth map and
makes it easy to leverage fully convolutional networks. We
aggregate local estimations by a non-parametric mean shift
variant, which explicitly enforces the estimated 3D joint coordinates to be in accordance with the 2D and 3D local estimations. Our method provides a better fusion scheme between 2D detection and 3D regression than previous stateof-the-art and the various baselines. As future work, we
plan to further extend our method for 3D pose estimation
from RGB inputs as well as for hands grasping objects.
80
Max Allowed distance to GT D(mm)
Figure 7. Comparison with state-of-the-art on ICVL [39]. We
plot the percentage of frames in which all joints are below a threshold.
the dataset but we do not use them as we perform data augmentation on the fly during training as described in Sec. 3.4.
The variance in pose is much smaller in ICVL compared to
the NYU and MSRA datasets. We compare our method
against [37, 46, 40, 45, 24, 12, 5]. [37, 46, 40] are based on
hierarchical regression forest and others [45, 24, 12, 5] on
3D holistic regression.
As is shown in Fig. 7, our method achieves similar accuracy as [5] and outperforms the rest. Our method has an
average 3D error on par with [5] and better than the others.
We consider the differences between our method and [5]
as being less significant given the result is nearly saturated.
The qualitative results can be seen in Fig. 8.
References
[1] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen,
C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, et al.
Tensorflow: Large-scale machine learning on heterogeneous
8
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21] G. Moon, J. Y. Chang, Y. Suh, and K. M. Lee. Holistic planimetric prediction to local volumetric prediction for 3d human
pose estimation. arXiv preprint arXiv:1706.04758, 2017. 2
[22] F. Moreno-Noguer. 3D human pose estimation from a single
image via distance matrix regression. In CVPR, 2017. 2, 3
[23] A. Newell, K. Yang, and J. Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 1, 2, 3,
4, 5, 6
[24] M. Oberweger and V. Lepetit. Deepprior++: Improving fast
and accurate 3D hand pose estimation. In ICCVW, 2017. 3,
4, 7, 8
[25] M. Oberweger, P. Wohlhart, and V. Lepetit. Hands deep
in deep learning for hand pose estimation. arXiv preprint
arXiv:1502.06807, 2015. 1, 3, 7
[26] M. Oberweger, P. Wohlhart, and V. Lepetit. Training a feedback loop for hand pose estimation. In ICCV, 2015. 3, 7
[27] I. Oikonomidis, N. Kyriazis, and A. A. Argyros. Efficient
model-based 3D tracking of hand articulations using kinect.
In BMVC, 2011. 1
[28] G. Papandreou, T. Zhu, N. Kanazawa, A. Toshev, J. Tompson, C. Bregler, and K. Murphy. Towards accurate multiperson pose estimation in the wild. In CVPR, 2017. 3, 4,
6
[29] G. Pavlakos, X. Zhou, K. G. Derpanis, and K. Daniilidis.
Coarse-to-fine volumetric prediction for single-image 3D human pose. In CVPR, 2017. 2
[30] A.-I. Popa, M. Zanfir, and C. Sminchisescu. Deep multitask
architecture for integrated 2D and 3D human sensing. In
CVPR, 2017. 1, 2
[31] C. Qian, Q. Chen, S. Xiao, W. Yichen, T. Xiaoou, and S. Jian.
Realtime and robust hand tracking from depth. In CVPR,
2014. 1
[32] G. Rogez, P. Weinzaepfel, and C. Schmid. LCR-Net:
Localization-classification-regression for human pose. In
CVPR, 2017. 1, 2
[33] J. Shotton, R. Girshick, A. Fitzgibbon, T. Sharp, M. Cook,
M. Finocchio, R. Moore, P. Kohli, A. Criminisi, A. Kipman,
and A. Blake. Efficient human pose estimation from single
depth images. TPAMI, 2013. 1, 2, 3
[34] H. Sidenbladh, M. J. Black, and D. J. Fleet. Stochastic tracking of 3D human figures using 2D image motion. In ECCV,
2000. 2
[35] L. Sigal, A. O. Balan, and M. J. Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. IJCV,
87(1):4–27, 2010. 2
[36] X. Sun, J. Shang, S. Liang, and Y. Wei. Compositional human pose regression. In ICCV, 2017. 3
[37] X. Sun, Y. Wei, S. Liang, X. Tang, and J. Sun. Cascaded
hand pose regression. In CVPR, 2015. 2, 3, 4, 7, 8
[38] J. S. Supancic, G. Rogez, Y. Yang, J. Shotton, and D. Ramanan. Depth-Based hand pose estimation: Data, methods,
and challenges. In ICCV, 2015. 3
[39] D. Tang, H. J. Chang, A. Tejani, and T.-K. Kim. Latent regression forest: Structured estimation of 3D articulated hand
posture. In CVPR, 2014. 2, 4, 7, 8
distributed systems. arXiv preprint arXiv:1603.04467, 2016.
4
F. Bogo, A. Kanazawa, C. Lassner, P. Gehler, J. Romero,
and M. J. Black. Keep it SMPL: Automatic estimation of 3D
human pose and shape from a single image. In ECCV, 2016.
2
Z. Cao, T. Simon, S.-E. Wei, and Y. Sheikh. Realtime
multi-person 2D pose estimation using part affinity fields. In
CVPR, 2017. 3
C.-H. Chen and D. Ramanan. 3D Human Pose Estimation =
2D Pose Estimation + Matching. In CVPR, 2017. 2
X. Chen, G. Wang, H. Guo, and C. Zhang. Pose guided
structured region ensemble network for cascaded hand pose
estimation. arXiv preprint arXiv:1708.03416, 2017. 1, 3, 7,
8
C. Choi, S. Kim, and K. Ramani. Learning hand articulations
by hallucinating heat distribution. In ICCV, 2017. 3
C. Choi, A. Sinha, J. H. Choi, S. Jang, and K. Ramani. A
collaborative filtering approach to Real-Time hand pose estimation: Supplementary material. In ICCV, 2015. 7
J. Dai, K. He, and J. Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016. 4
E. Dibra, T. Wolf, C. Oztireli, and M. Gross. How to Refine
3D Hand Pose Estimation from Unlabelled Depth Data? In
3DV, 2017. 3
L. Ge, H. Liang, J. Yuan, and D. Thalmann. Robust 3D
hand pose estimation in single depth images: from singleview cnn to multi-view cnns. In CVPR, 2016. 1, 2, 3, 5,
7
L. Ge, H. Liang, J. Yuan, and D. Thalmann. 3D convolutional neural networks for efficient and robust hand pose estimation from single depth images. In CVPR, 2017. 2, 3, 4,
7, 8
H. Guo, G. Wang, X. Chen, C. Zhang, F. Qiao, and H. Yang.
Region ensemble network: Improving convolutional network for hand pose estimation. In ICIP, 2017. 3, 7, 8
K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask RCNN. In ICCV, 2017. 4
S. Ioffe. Batch renormalization: Towards reducing minibatch
dependence in batch-normalized models. arXiv preprint
arXiv:1702.03275, 2017. 4
D. Kingma and J. Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 4
I. Lifshitz, E. Fetaya, and S. Ullman. Human pose estimation
using deep consensus voting. In ECCV, 2016. 3
M. Lin, L. Lin, X. Liang, K. Wang, and H. Cheng. Recurrent
3D pose sequence machines. In CVPR, 2017. 1, 2
M. Madadi, S. Escalera, A. Carruesco, C. Andujar, X. Baró,
and J. Gonzàlez. Occlusion aware hand pose recovery from
sequences of depth images. In FG, 2017. 7
J. Martinez, R. Hossain, J. Romero, and J. J. Little. A simple yet effective baseline for 3D human pose estimation. In
ICCV, 2017. 2
F. Milletari, S.-A. Ahmadi, C. Kroll, A. Plate, V. Rozanski,
J. Maiostre, J. Levin, O. Dietrich, B. Ertl-Wagner, K. Bötzel,
et al. Hough-cnn: Deep learning for segmentation of deep
brain regions in mri and ultrasound. CVIU, 2017. 3
9
[40] D. Tang, J. Taylor, P. Kohli, C. Keskin, T.-K. Kim, and
J. Shotton. Opening the black box: Hierarchical sampling
optimization for estimating human hand pose. In ICCV,
2015. 2, 3, 7, 8
[41] J. Taylor, J. Shotton, T. Sharp, and A. Fitzgibbon. The vitruvian manifold: Inferring dense correspondences for one-shot
human pose estimation. In CVPR, 2012. 4
[42] D. Tome, C. Russell, and L. Agapito. Lifting from the deep:
Convolutional 3D pose estimation from a single image. In
CVPR, 2017. 1, 2
[43] J. Tompson, M. Stein, Y. Lecun, and K. Perlin. Real-time
continuous pose recovery of human hands using convolutional networks. ACM Transactions on Graphics (TOG),
2014. 1, 2, 3, 4, 5, 6, 7
[44] A. Toshev and C. Szegedy. DeepPose: Human pose estimation via deep neural networks. In CVPR, 2014. 1
[45] C. Wan, T. Probst, L. Van Gool, and A. Yao. Crossing nets:
Combining GANs and VAEs with a shared latent space for
hand pose estimation. In CVPR, 2017. 3, 7, 8
[46] C. Wan, A. Yao, and L. Van Gool. Hand pose estimation
from local surface normals. In ECCV, 2016. 2, 3, 7, 8
[47] S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In CVPR, 2016. 1, 2, 4, 6
[48] Y. Xie, X. Kong, F. Xing, F. Liu, H. Su, and L. Yang. Deep
voting: A robust approach toward nucleus localization in
microscopy images. In International Conference on Medical Image Computing and Computer-Assisted Intervention,
2015. 3
[49] C. Xu, L. N. Govindarajan, Y. Zhang, and L. Cheng. Lie-X:
Depth image based articulated object pose estimation, tracking, and action recognition on lie groups. IJCV, 2017. 3,
7
[50] A. Yao, J. Gall, L. V. Gool, and R. Urtasun. Learning probabilistic non-linear latent variable models for tracking complex activities. In NIPS, 2011. 2
[51] Q. Ye, S. Yuan, and T.-K. Kim. Spatial attention deep net
with partial PSO for hierarchical hybrid hand pose estimation. In ECCV, 2016. 1, 2, 3, 5, 7
[52] S. Yuan, Q. Ye, B. Stenger, S. Jain, and T.-K. Kim. Bighand2. 2m benchmark: Hand pose dataset and state of the
art analysis. In CVPR, 2017. 1
[53] X. Zhou, Q. Huang, X. Sun, X. Xue, and Y. Wei. Towards
3D human pose estimation in the wild: A weakly-supervised
approach. In ICCV, 2017. 2
[54] X. Zhou, Q. Wan, W. Zhang, X. Xue, and Y. Wei. Modelbased deep hand pose estimation. In IJCAI, 2015. 3, 7
10
| 1cs.CV
|
A note on faithful coupling of Markov chains
Debojyoti Dey1 , Pranjal Dutta2 , and Somenath Biswas1
arXiv:1710.10026v1 [cs.DS] 27 Oct 2017
1
Department of Computer Science and Engineering, Indian Institute of Technology Kanpur,
Kanpur-208016.
2
Chennai Mathematical Institute, Chennai-603103.
October 2017
Abstract
One often needs to turn a coupling (Xi , Yi )i≥0 of a Markov chain into a sticky
coupling where once XT = YT at some T , then from then on, at each subsequent
time step T ′ ≥ T , we shall have XT ′ = YT ′ . However, not all of what are considered
couplings in literature, even Markovian couplings, can be turned into sticky couplings,
as proved by Rosenthal [Ro97] through a counter example. Rosenthal then proposed a
strengthening of the Markovian coupling notion, termed as faithful coupling, from which
a sticky coupling can indeed be obtained. We identify the reason why a sticky coupling
could not be obtained in the counter example of Rosenthal, which motivates us to define
a type of coupling which can obviously be turned into a sticky coupling. We show then
that the new type of coupling that we define, and the faithful coupling as defined by
Rosenthal [Ro97], are actually identical. Our note may be seen as a demonstration of
the naturalness of the notion of faithful coupling.
1
Introduction
First, we recall certain basic definitions and fix our notations. For a random variable X,
we denote its distribution as D(X). Let M be a finite Markov chain on the state space
Ω, and with the transition matrix P . Following [No97], we abbreviate the Markov chain
(Xi )i≥0 , with D(X0 ) = λ and D(Xi+1 ) = D(Xi )P for i ≥ 0, as Markov(λ, P ). A coupling
of the Markov chain M is a process (Xi , Yi )i≥0 , with each Xi , Yi taking values in Ω and
evolving on a common probability space, where (Xi )i≥0 satisfies the property that, for i ≥ 0,
D(Xi+1 ) = D(Xi )P and similarly, (Yi )i≥0 satisfies, for i ≥ 0, D(Yi+1 ) = D(Yi )P .
Remark A: The definition of Markov chain coupling, as given in [LPW09], says that both (Xi )i≥0
and (Yi )i≥0 are Markov chains with transition matrix P . The reason we prefer the above weaker
definition is because in literature we find processes considered as couplings that do not satisfy
1
the stronger notion of [LPW09], but do satisfy our above definition. A Markov chain (Xi )i≥0
with transition matrix P , by definition, satisfies the property (π) that for i ≥ 0, conditioned on
Xi = l, Xi+1 has the distribution P (l, ·) and is indepedent of X0 , . . . , Xi−1 . As we shall see in the
next section, there is a Markovian coupling (Xi , Yi )i≥0 , where (Xi )i≥0 does not satisfy the property
(π), although Xi ’s do evolve (under certain conditions) as per the transition matrix P .
Let D(X0 ) be µ and D(Y0 ) be ν. In the above, since each pair, Xi and Yi , i ≥ 0, is defined
on the same probability space, the two random variables, Xi , Yi , for each i ≥ 0, are coupled,
and so we have, from the above definition of Markov chain coupling, and Proposition 4.7 of
[LPW09]:
kµP i − νP i kTV = kD(Xi ) − D(Yi )kTV ≤ Pr(Xi 6= Yi )
(1)
(Here, k · kTV denotes the total variational distance between two distributions.)
Let us suppose that a coupling of M has the ‘now-equals-forever’ property ([Ro97]), namely,
if Xn = Yn for some n, then for all j ≥ n, Xj = Yj . Then, if T is defined as the random
time
def
T = inf{i|Xi = Yi }
(2)
we get from the relation (1)
|kD(Xi ) − D(Yi )kTV ≤ Pr(Xi 6= Yi ) = Pr(T > i)
(3)
as the two events (Xi 6= Yi ) and (T > i) imply each other, making use of the ‘now-equalsforever’propery.
As such, the applicability of the inequality (3) is limited as most couplings of Markov chains
will not have the ‘now-equals-forever’ property. The way that is often resorted to to tackle
this problem is to define a new process (Zi )i≥0 as follows: ([Ro97])
Xi if i ≤ T
Zi =
(4)
Yi
otherwise
where T is as in (2) above. If it so happens that (Zi )i≥0 = Markov(µ, P ) then, since
(Zi , Yi )i≥0 is a coupling of the Markov chain satisfying ‘now-equals-forever’ property, we
have
kµP i − νP i kTV = kD(Zi ) − D(Yi )kTV ≤ Pr(T > i)
The construction of (Zi ) has been evocatively termed as sticking and the resultant new
coupling as a sticky coupling [HM17].
Definition 1 (Markovian coupling of a Markov chain) A coupling (Xi , Yi )i≥0 of a Markov
chain M on state space Ω, and with its transition matrix as P , where D(X0 ) = µ and
def
D(Y0 ) = ν, for given µ and ν, is said to be a Markovian coupling if (Wi )i≥0 , with Wi =
(Xi , Yi ) is itself a Markov chain on state space Ω × Ω, with a specified joint distribution
of D(X0 ) and D(Y0 ) as the distribution D(W0 ). In other words, there is a (Ω × Ω, Ω × Ω)
2
transition matrix, say Q, such that for some distribution θ0 on Ω × Ω, satisfying that its
def
marginals are µ = D(X0 ) and ν = D(Y0 ), we will have, for i ≥ 1, with θi = θi−1 Q, the two
marginals of θi will be µP i , that is, D(Xi ) and νP i , that is D(Yi ).
Remark B: We note that the defintion above is specific to a joint distribution of (X0 , Y0 ). The
reason for the specificity is that in the example given by Rosenthal ([Ro97]) of a Markovian coupling
for which sticking fails, as we note in the next section, the ability to evolve the two copies as per
the transition matrix P crucially depends on the initial coupling of X0 and Y0 .
As noted in [HM17], it has been stated at times that sticking will work for all Markovian
couplings. However, this is not correct, Rosenthal [Ro97] provides a counter example. He
then provides a stronger version of Markovian couplings, termed as faithful couplings, for
which it is proved that sticking will provably result in a coupling of M where (Zi )i≥0 , as
defined in (4) above, will indeed be Markov(µ, P ). We discuss Rosenthal’s counterexample
in Section 2 and identify what we consider to be the reason why the sticking operation fails
there. This reason then motivates us in defining a stronger version of Markovian coupling
(Section 3) for which it is easy to see that the sticking will indeed work. We then prove that
the new notion of coupling that we define is actually equivalent to the notion of faithful
coupling.
2
Rosenthal’s counterexample
In his example, Rosenthal [Ro97] considers the Markov chain M with state space Ω = {0, 1},
and with the transition matrix P defined as:
0
1
P = 0 1/2 1/2
1 1/2 1/2
(Xi ), (Yi ), i ≥ 0 are defined as follows:
1. Each of D(X0 ), D(Y0 ) is the uniform distribution on the state space {0, 1},
def
2. Each D(Yi ), i > 0 is also the uniform distribution, whereas for i ≥ 0, Xi+1 = Xi ⊕ Yi ,
⊕ being the exclusive-or operation.
Thus, the joint evolution of the two copies (Xi , Yi ) of the chain, on the state space Ω × Ω,
that is, {(0, 0), (0, 1), (1, 0), (1, 1)} is governed by the following transition matrix Q:
(0, 0)
Q = (0, 1)
(1, 0)
(1, 1)
(0, 0) (0, 1) (1, 0) (1, 1)
1/2
1/2
0
0
0
0
1/2
1/2
0
0
1/2
1/2
1/2
1/2
0
0
3
Convention: The distribtion of a random variable X on Ω = {0, 1} is specified as the
vector
[Pr(X = 0) Pr(X = 1)]
and similarly, the distribution of a random variable W on Ω × Ω as the vector
[Pr(W = (0, 0))
Pr(W = (0, 1))
Pr(W = (1, 0))
Pr(W = (1, 1))]
def
We note that for any distribution σ = [p 1 − p], Ω = {0, 1}, σP will be the uniform
distribution. We also note that for the trivial coupling θ, of [1/2 1/2] and [1/2 1/2],
which is [1/4 1/4 1/4 1/4], θQ will again be θ. Therefore, with the joint distribution θ
of D(X0 ) and D(Y0 ), Q indeed defines a Markovian coupling of the Markov chain M .
Remark C: We also note that there are other joint distributions of the same D(X0 ) and D(Y0 ),
def
for which Q will not be a Markovian coupling: consider θ′ = [3/8 1/8 1/8 3/8), θ′ Q =
[3/8 3/8 1/8 1/8]. The marginals of the latter distributions are [3/4 1/4] and [1/2 1/2]
respectively. Thus, if we use the joint distribution θ′ of the same two initial distributions, while Q
evolves Yi ’s (understandably) as per P , it does not evolve Xi ’s as per P . Hence with the initial joint
distribution θ′ , Q fails to be a Markovian coupling of M . This is why we felt that it is necessary
to make explicit the initial joint distribution in the definition of Markovian coupling of a Markov
chain.
Rosenthal proves that although Q is a Markovian coupling of M with θ as the initial joint
distribution, the result of the sticking operation will fail to evolve M correctly: with T
defined as in (2) and Zi ’s as defined in (4),
Pr(Z0 = 1, Z1 = 0)
= Pr(T = 0, Y0 = 1, Y1 = 0, X0 = 1) + Pr(T > 0, X0 = 1, Y0 = 0, X1 = 0)
= 1/8 + 0
= 1/8
However, had Zi ’s been evolving as per P with the same initial distribution as the uniform
distribution, the probability above would have been 1/4, and not 1/8. Hence, sticking fails.
We see that sticking failed with T = 0. As it happens, sticking will fail here for every value
of T (except, of course, for T = ∞). The reason is provided by the Proposition below. We
need the following definition:
Definition 2 For a state space Ω with s an element in it, δs denotes the distribution on Ω
in which the probability of s is 1, and (therefore,) all the other states have probability 0.
Proposition 3 Let M be a Markov chain with Ω as its state space and P as its transition
matrix, and let Q be a transition matrix for the state space Ω × Ω that defines a Markovian
4
coupling (Wi = (Xi , Yi ))i≥0 of M where D(X0 ) = µ and D(Y0 ) = ν, where the initial
joint distribution used is θ0 . This Markovian coupling of M can be turned into a sticky
coupling using the sticking operation if the following condition holds: for every s ∈ Ω,
defining η0 as the unique joint distribution on Ω × Ω with δs and δs as its two marginals,
and further defining ηi+1 as ηi Q, for each i ≥ 0, we have that for each ηi , i ≥ 0, will be a
joint distribution that has δs P i and δs P i as its marginals. On the other hand, if Q does not
satisfy the condition, then the Q-defined Markovian coupling, in general, cannot be turned
into a sticky coupling.
Proof Sketch.
The condition ensures that both (Xi )i≥0 and (Yi )i≥0 of the Markovian
coupling of M satisfy the strong Markov property (Theorem 1.4.2, [No97]) with respect to
the stopping time T , as defined in (2). Therefore, conditioned on T = m and XT = YT = s,
each of (XT +n )n≥0 and (YT +n )n≥0 will be Markov(δs , P ) and the former will be independent
of X0 , X1 , . . . , XT and the latter will be independent of Y0 , Y1 , . . . , YT . Thus, (XT +n )n≥0
and (YT +n )n≥0 can replace each other. Because of this, sticking is guaranteed to work as
(Xi )i≥0 and (X0 , X1 , . . . , XT , YT +1 , . . .) will be indistinguishable.
We now show that the counter example of Rosenthal does not satify the condition as stated
in the statement of the proposition. As we have seen that the Markovian coupling there
cannot be turned into a sticky coupling, the counter example proves the second part of
the Proposition. Let us suppose that T = m and XT = YT = 0. Conditioned on these
events, D(XT ) = D(YT ) = [1 0]. The only joint distribution with these two distributions
as marginals is [1 0 0 0]. The joint distribution of XT +1 and YT +1 will be given by
[1 0 0 0]Q, which is [1/2 1/2 0 0]. This gives D(XT +1 ) as [1 0], which is not
[1/2 1/2] = D(XT )P . Thus, Q fails to ensure that, conditioned as above, D(XT +1 ) = δ0 P .
On the other hand, D(YT +1 ) = δ0 P , and so (XT +n )n≥0 and (YT +n )n≥0 are different and
cannot replace each other, as required for the sticking operation to work.
3
Strong Markovian coupling
We saw in Section 2 that sticking fails for Q because it does not evolve (Xi )i≥0 as per P
for certain joint distributions. Therefore, the following definition of a type of coupling of
Markov chains immediately suggests itself to correct the defect:
Definition 4 (Strong Markovian coupling) A coupling of a finite Markov chain M
with state space Ω and transition matrix P is a strong Markovian coupling if there is a
transition matrix Q from Ω × Ω to Ω × Ω such that for every pair of distributions µ and ν,
each on Ω, and for every joint distribution θ of µ and ν, θQ will be a joint distribution of
µP and νP .
From this definition it follows that:
5
Claim 5 Let Q be a strong Markovian coupling of a finite Markov chain M with state space
Ω and transition matrix P , and for two random variables X0 and Y0 on Ω let D(X0 ), D(Y0 )
be µ and ν respectively, and let θ0 be any joint distribution of µ and ν. If we define θi+1
as θi Q, for i ≥ 0, and Xi+1 and Yi+1 as two random variables whose distributions are
respectively the two marginals of θi+1 then
1. (Xi )i≥0 will be Markov(µ, P ) and (Yi )i≥0 will be Markov(ν, P ). Consequently,
2. D(Xi ) = µP i and D(Yi ) = νP i , for i ≥ 0,
3. Both (Xi )i≥0 and (Yi )i≥0 will satisfy the strong Markov property, and therefore, the
coupling Q of M can be turned into a valid sticky coupling through sticking.
We see next that the strong Markovian coupling is actually equivalent to the familiar notion
of faithful coupling which is defined as:
Definition 6 (Faithful coupling [Ro97]) A Markov coupling of a finite Markov chain
M , with Ω as its state space and P as its transition matrix, is a faithful coupling given by
a Markov chain (Wi )i≥0 = (Xi , Yi )i≥0 , on state space Ω × Ω, with transition matrix Q, if Q
satisfies, for all i, j, i′ , j ′ ∈ Ω, the following
X
Q((i, j), (i′ , j ′ )) = P (i, i′ ), and
j ′ ∈Ω
X
Q((i, j), (i′ , j ′ )) = P (j, j ′ )
i′ ∈Ω
Equivalently, for all t ≥ 0 and all i, j, i′ , j ′ ∈ Ω
Pr(Xt+1 = i′ |Xt = i, Yt = j) = P (i, i′ ), and
Pr(Yt+1 = j ′ |Xt = i, Yt = j) = P (j, j ′ )
Remark D: This kind of coupling has been termed in [LPW09], as well as in [MU05], as Markovian
coupling; however, we follow the terminology of [HM17] for reasons given therein. It is to be noted
that most coupling constructions of Markov chains turn out to be faithful couplings, as the various
coupling examples in [LPW09] demonstrate.
Proposition 7 A Markov chain coupling is faithful if and only if it is a strong Markovian
coupling.
The two lemmas below prove the two directions of the above proposition.
6
Lemma 8 Every faithful coupling of a Markov chain M is a strong Markovian coupling.
Proof. Let M be a Markov chain with state space Ω and transition matrix P . Let the
transition matrix Q, giving transition probabilities from Ω × Ω to Ω × Ω, define a faithful
coupling of the Markov chain M . Let µ and ν be two distributions on Ω, and let θ be a
joint distribution of µ and ν. We need to prove that µP and νP are the two marginals of
θQ. In particular, we need to show:
X
For all x ∈ Ω,
(θQ)(x, y) = (µP )(x)
(5)
y∈Ω
and,
For all y ∈ Ω,
X
(θQ)(x, y) = (νP )(y)
(6)
x∈Ω
We prove (5) as follows:
X
(θQ)(x, y)
y∈Ω
=
X
X
θ(u, v)Q((u, v), (x, y))
y∈Ω (u,v)∈Ω×Ω
=
X
θ(u, v)
X
θ(u, v)P (u, x), because Q defines a faithful coupling
Q((u, v), (x, y))
y∈Ω
(u,v)∈Ω×Ω
=
X
(u,v)∈Ω×Ω
=
X
P (u, x)
X
P (u, x)µ(u), θ being the joint ditribution of µ and ν
u∈Ω
=
X
θ(u, v)
v∈Ω
u∈Ω
= (µP )(x)
In a similar manner we can prove (6).
For the other direction, we prove
Lemma 9 Every strong Markovian coupling of a Markov chain M is a faithful coupling of
M.
Proof. Let M be a Markov chain with state space Ω and transition matrix P . Let the
transition matrix Q, giving transition probabilities from Ω × Ω to Ω × Ω, define a strong
Markovian coupling of the Markov of M . For any (u, v) ∈ Ω × Ω, consider the probabilty
7
distribution δ(u,v) . This distribution is the joint distribution δu and δv , both defined on Ω.
(The definition of δx as in Definition 2.) δu P is the uth row of P , namely, P (u, ·). Similarly,
δv P will be the vth row of P , namely, P (v, ·), and δ(u,v) Q will be the (u, v)th row of Q, that
is, Q((u, v), ·). As Q defines a strong Markovian coupling of M , we have that Q((u, v), ·)
will be the joint distribution of P (u, ·) and P (v, ·). From this condition, we get for any
x, u, v ∈ Ω:
X
Q((u, v), (x, y)) = P (u, x)
(7)
y∈Ω
Similarly, for any y, u, v ∈ Ω, we get
X
Q((u, v), (x, y)) = P (v, y)
(8)
x∈Ω
As (7) and (8) are precisely the conditions to be met by Q to be a faithful coupling, we
conclude that Q defines a faithful coupling of M .
4
Concluding remarks
We have seen that it is sufficient for a Markovian coupling to have the faithfulness property
in order to be turned to a sticky coupling. Is the faithfulness property also a necessary
property? It may be that there is a Markovian coupling of a Markov chain M which evolves
two copies of the chain correctly for a pair of initial distributions µ, ν of M , using a joint
distribution of the pair, and satisfies the condition in the statement of Proposition 3, but it
either does not work for some other joint distribution of the same pair µ and ν, or, does not
satisfy the condition for some other pair of initial distributions of M . Such a Markovian
coupling can be turned into a sticky coupling for the µ, ν pair, but will not be strongly
Markovian, and hence will not be a faithful coupling of M . One feels that even if such
an example exists, it is unlikely to be a natural example. Faithfulness appears to be the
only natural strengthening of the Markovian coupling notion that ensures that the sticking
operation will work.
References
[HM17] Hirscher, Timo and Anders Martinsson, Segregating Markov chains, manuscript,
arXiv:1510.036661v2 [math.PR], 2017.
[MU05] Mitzenmacher, Michael and Eli Upfal, Probability and Computing: Randomized
Algorithms and Probabilistic Analysis, Cambridge University Press, 2005.
[No97] Norris, J.R., Markov Chains, Cambridge University Press, 1997. Paperbackk edition
1998, 15th printing 2009.
8
[LPW09] Levin, David A., Yuval Peres and Elizabeth L. Wilmer, Markov Chains and Mixing Times, American Mathematical Society, 2009.
[Ro97] Rosenthal, Jeffrey S., Faithful Couplings of Markov Chains: Now Equals Forever,
Advances in Applied Mathematics, Vol. 18, pp. 372–381, 1997.
9
| 8cs.DS
|
arXiv:1705.10696v2 [math.ST] 26 Sep 2017
Localized Gaussian width of M -convex
hulls with applications to Lasso and
convex aggregation
Pierre C. Bellec
Rutgers University, Department of Statistics and Biostatistics
Abstract: Upper and lower bounds are derived for the Gaussian mean
width of the intersection of a convex hull of M points with an Euclidean
ball of a given radius. The upper bound holds for any collection of extreme point bounded in Euclidean norm. The upper bound and the
lower bound match up to a multiplicative constant whenever the extreme points satisfy a one sided Restricted Isometry Property.
This bound is then applied to study the Lasso estimator in fixeddesign regression, the Empirical Risk Minimizer in the anisotropic persistence problem, and the convex aggregation problem in density estimation.
1. Introduction
Let T be a subset of Rn . The Gaussian width of T is defined as
ℓ(T ) := E sup uT g,
u∈T
where g = (g1 , ..., gn )T and g1 , ..., gn are i.i.d. standard normal random variables. For any vector u ∈ Rn , denote by |u|2 its Euclidean norm and define
the Euclidean balls
B2 = {u ∈ Rn : |u|2 ≤ 1},
sB2 = {su ∈ Rn , u ∈ B2 } for all s ≥ 0.
We will also use the notation S n−1 = {u ∈ Rn : |u|2 = 1}. The localized
Gaussian width of T with radius s > 0 is the quantity ℓ(T ∩ sB2 ). For any
P
u ∈ Rp , define the ℓp norm by |u|p = ( ni=1 |ui |p )1/p for any p ≥ 1, and let
|u|0 be the number of nonzero coefficients of u.
This paper studies the localized Gaussian width
ℓ(sB2 ∩ T ),
where T is the convex hull of M points in Rn .
1
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
2
If T = B1 = {u ∈ Rn : |u|1 ≤ 1}, then matching upper and lower bounds
are available for the localized Gaussian width:
q
√
ℓ(sB2 ∩ B1 ) ≍ log (en(s2 ∧ 1)) ∧ s n ,
(1)
cf. [14] and [21, Section 4.1]. In the above display, a ≍ b means that a ≤ Cb
and b ≤ Ca for some large enough numerical constant C ≥ 1.
The first goal of this paper is to generalize this bound to any T that is
the convex hull of M ≥ 1 points in Rn .
Contributions. Section 2 is devoted to the generalization of (1) and provides sharp bounds on the localized Gaussian width of the convex hull of
M points in Rn , see Propositions 1 and 2 below. Sections 3 to 5 provide
statistical applications of the results of Section 2. Section 3 studies the Lasso
estimator and the convex aggregation problem in fixed-design regression. In
Section 4, we show that Empirical Risk Minimization achieves the minimax
rate for the persistence problem in the anisotropic setting. Finally, Section 5
provides results for bounded empirical processes and for the convex aggregation problem in density estimation.
2. Localized Gaussian width of a M -convex hull
The first contribution of the present paper is the following upper bound on
localized Gaussian width of the convex hull of M points in Rn .
Proposition 1. Let n ≥ 1 and M ≥ 2. Let T be the convex hull of M
points in Rn and assume that T ⊂ B2 . Let g be a centered Gaussian random
variable with covariance matrix In×n . Then for all s > 0,
q
ℓ(T ∩ sB2 ) ≤ 4 log+
(4eM (s2
√
∧ 1)) ∧ s n ∧ M
(2)
where log+ (a) = max(1, log a).
Proposition 1 is proved in the next two subsections. Inequality
√
ℓ(T ∩ sB2 ) ≤ s n ∧ M
(3)
√
is a direct consequence of the Cauchy-Schwarz inequality and E|P g|2 ≤ d
where P ∈ Rn×n is the orthogonal projection onto the linear span of T and
d ≤ (n ∧ M ) is the rank of P . The novelty of (2) is inequality
q
ℓ(T ∩ sB2 ) ≤ 4 log+ (4eM (s2 ∧ 1)).
(4)
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
3
Inequality (4) was known for the ℓ1 -ball T = {u ∈ Rn : |u|1 ≤ 1} [14], but
to our knowledge (4) is new for general M -convex hulls. If T is the ℓ1 -ball,
then the bound (2) is sharp up to numerical constants [14], [21, Section 4.1].
The above result does not assume any type of Restricted Isometry Property (RIP). The following proposition shows that (4) is essentially sharp
provided that the vertices of T satisfies a one-sided RIP of order 2/s2 .
Proposition 2. Let n ≥ 1 and M ≥ 2. Let g be a centered Gaussian
random variable with covariance matrix In×n . Let s ∈ (0, 1] and assume for
simplicity that m = 1/s2 is a positive integer such that m ≤ M/5. Let T be
the convex hull of the 2M points {±µ1 , ..., ±µM } where µ1 , ..., µM ∈ S n−1 .
Assume that for some real number κ ∈ (0, 1) we have
for all θ ∈ RM such that |θ|0 ≤ 2m,
κ|θ|2 ≤ |µθ |2
where µθ =
PM
j=1 θj µj .
Then
s
√
M s2
ℓ(T ∩ sB2 ) ≥ ( 2/4)κ log
.
5
(5)
The proof of Proposition 2 is given in Appendix A.
2.1. A refinement of Maurey’s argument
This subsection provides the main tool to derive the upper bound (4). Define
the simplex in RM by
n
ΛM = θ ∈ RM ,
M
X
j=1
θj = 1,
o
∀j = 1 . . . M, θj ≥ 0 .
(6)
Let m ≥ 1 be an integer, and let
Q(θ) = θ T Σθ,
where Σ = (Σjj ′ )j,j ′ =1,...,M is a positive semi-definite matrix of size M . Let
θ̄ ∈ ΛM be a deterministic vector such that Q(θ̄) is small. Maurey’s argument [27] has been used extensively to prove the existence of a sparse vector
θ̃ ∈ ΛM such that Q(θ̃) is of the same order as that of Q(θ̄). Maurey’s
argument uses the probabilistic method to prove the existence of such θ̃. A
sketch of this argument is as follows.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
4
Define the discrete set ΛM
m as
ΛM
m
:=
(
m
1 X
uk , u1 , ..., um ∈ {e1 , ..., eM } ,
m k=1
)
(7)
where (e1 , ..., eM ) is the canonical basis in RM . The discrete set ΛM
m is a
subset of the simplex ΛM that contains only m-sparse vectors.
Let (e1 , ..., eM ) be the canonical basis in RM . Let Θ1 , ..., Θm be i.i.d.
random variables valued in {e1 , ..., eM } with distribution
P (Θk = ej ) = θ̄j
for all k = 1, ..., m.
(8)
Next, consider the random variable
θ̂ =
m
1 X
Θk .
m k=1
(9)
The random variable θ̂ is valued in ΛM
m and is such that EΘ [θ̂] = θ̄, where
EΘ denotes the expectation with respect to θ̂. Then a bias-variance decomposition yields
EΘ [θ̂] ≤ Q(θ̄) + R2 /m,
Q(θ) ≤
where R > 0 is a constant such that maxj=1,...,M Σjj ≤ R2 . As minθ∈ΛM
m
M
EΘ [θ̂], this yields the existence of θ̃ ∈ Λm such that
Q(θ̃) ≤ Q(θ̄) + R2 /m.
If m is chosen large enough, the two terms Q(θ̄) and R2 /m are of the same
order and we have established the existence of an m-sparse vector θ̃ so that
Q(θ̃) is not much substantially larger than Q(θ̄).
For our purpose, we need to refine this argument by controlling the deviation of the random variable Q(θ̂). This is done in Lemma 3 below.
M → [0, +∞) be a
Lemma 3. Let m ≥ 1 and define ΛM
m by (7). Let F : R
convex function. For all θ ∈ RM , let
Q(θ) = θ T Σθ,
where Σ = (Σjj ′ )j,j ′ =1,...,M is a positive semi-definite matrix of size M . Assume that the diagonal elements of Σ satisfy Σjj ≤ R2 for all j = 1, ..., M .
Then for all t > 0,
sup
θ∈ΛM :Q(θ)≤t2
F (θ) ≤
Z
1
+∞
"
max
2
2
θ∈ΛM
m : Q(θ)≤x(t +R /m)
#
F (θ)
dx
.
x2
(10)
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
5
In the next sections, it will be useful to bound from above the quantity
F (θ) maximized over ΛM subject to the constraint Q(θ) ≤ t2 . An interpretation of (10) is as follows. Consider the two optimization problems
maximize F (θ)
maximize F (θ)
for θ ∈ ΛM
for θ ∈ ΛM
m
subject to Q(θ) ≤ t2 ,
subject to Q(θ) ≤ Y (t2 + R2 /m),
for some Y ≥ 1. Equation (10) says that the optimal value of the first
optimization problem is smaller than the optimal value of the second optimization problem averaged over the distribution of Y given by the density
y 7→ 1/y 2 on [1, +∞). The second optimization problem above is over the
2
2
discrete set ΛM
m with the relaxed constraint Q(θ) ≤ Y (t + R /m), hence we
have relaxed the constraint in exchange for discreteness. The discreteness of
the set ΛM
m will be used in the next subsection for the proof of Proposition 1.
Proof of Lemma 3. The set {θ ∈ ΛM : Q(θ) ≤ t2 } is compact. The function
F is convex with domain RM and thus continuous. Hence the supremum in
the left hand side of (10) is achieved at some θ̄ ∈ ΛM such that Q(θ̄) ≤ t2 .
Let Θ1 , ..., Θm , θ̂ be the random variable defined in (8) and (9) above. Denote
by EΘ the expectation with respect to Θ1 , ..., Θm . By definition, θ̂ ∈ ΛM
m
and EΘ θ̂ = θ̄. Let E = EΘ [Q(θ̂)]. A bias-variance decomposition and the
independence of Θ1 , ..., Θm yield
E := EΘ [Q(θ̂)] = Q(θ̄) + EΘ (θ̂ − θ̄)T Σ(θ̂ − θ̄),
1
= Q(θ̄) + EΘ [(Θ1 − θ̄)T Σ(Θ1 − θ̄)].
m
Another bias-variance decomposition yields
EΘ (Θ1 − θ̄)T Σ(Θ1 − θ̄) = EΘ [Q(Θ1 )] − Q(θ̄) ≤ EΘ Q(Θ1 ) ≤ R2 ,
where we used that Q(·) ≥ 0 and that Θ1 ΣΘ1 ≤ R2 almost surely. Thus
E = EΘ [Q(θ̂)] ≤ Q(θ̄) + R2 /m ≤ t2 + R2 /m.
(11)
Define the random variable X = Q(θ̂)/E, which is nonnegative and satisfifes ER Θ [X] = 1. By Markov inequality, it holds that PΘ (X > t) ≤
1/t = 1+∞ (1/x2 )dx. Define the random variable Y by the density function x → 1/x2 on [1, +∞). Then we have PΘ (X > t) ≤ P(Y > t) for any
t > 0, so by stochastic dominance, there exists a rich enough probability
space Ω and random variables X̃ and Ỹ defined on Ω such that X̃ and X
have the same distribution, Ỹ and Y have the same distribution, and X̃ ≤ Ỹ
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
6
almost surely on Ω (see for instance Theorem 7.1 in [12]). Denote by EΩ the
expectation sign on the probability space Ω.
By definition of θ̄ and θ̂, using Jensen’s inequality, Fubini’s Theorem and
the fact that θ̂ ∈ ΛM
m we have
sup
θ∈ΛM :Q(θ)≤t2
F (θ) = F (θ̄) = F (EΘ [θ̂]) ≤ EΘ [F (θ̂)] ≤ EΘ [g(Q(θ̂)/E)]
where g(·) is the nondecreasing function g(x) = maxθ∈ΛM
F (θ). The
m :Q(θ)≤xE
right hand side of the previous display is equal to to EΘ [g(X)]. Next, we use
the random variables X̃ and Ỹ as follows:
EΘ [g(X)] = EΩ [g(X̃)] ≤ EΩ [g(Ỹ )] =
Z
1
+∞
g(x)
dx.
x2
Combining the previous display and (11) completes the proof.
2.2. Proof of (4)
We are now ready to prove Proposition 1. The main ingredients are Lemma 3
and the following upper bound on the cardinal of ΛM
m
log |ΛM
m|
!
M +m−1
2M
= log
≤ log
m
m
!
≤ m log
2eM
m
.
(12)
Proof of (4). If s2 < 1/M then by (3) we have ℓ(T ∩ sB2 ) ≤ 1, hence (4)
holds. Thus it is enough to focus on the case s2 ≥ 1/M .
Let r = min(s, 1) and set m = ⌊1/r 2 ⌋, which satisfies 1 ≤ m ≤ M . As T
is the convex hull of M points, let µ1 , ..., µM ∈ Rn be such that
T = convex hull of {µ1 , ..., µM } = {µθ , θ ∈ ΛM },
M
where µθ = M
j=1 θj µj for θ ∈ Λ .
Let Q(θ) = |µθ |22 for all θ ∈ RM . This is a polynomial of order 2, of
the form Q(θ) = θ T Σθ, where Σ is the Gram matrix with Σjk = µTk µj for
all j, k = 1, ..., M . As we assume that T ⊂ B2 , the diagonal elements of Σ
satisfy Σjj ≤ 1. For all θ ∈ RM , let F (θ) = g T µθ . Applying Lemma 3 with
the above notation, R = 1, m = ⌊1/r 2 ⌋ and t = r, we obtain
P
E
sup
θ∈ΛM :Q(θ)≤r 2
T
g µθ ≤ E
Z
1
+∞
"
max
2
θ∈ΛM
m : Q(θ)≤x(r +1/m)
#
F (θ)
dx
.
x2
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
7
By definition of m, r 2 ≤ 1/m so that x(r 2 + 1/m) ≤ 2x/m. Using Fubini
Theorem and a bound on the expectation of the maximum of |ΛM
m | centered
Gaussian random variables with variances bounded from above by 2x/m,
we obtain that the right hand side of the previous display is bounded from
above by
Z
1
+∞
1
x2
s
q
4x log |ΛM
m|
dx ≤ log(2eM/m)
m
Z
+∞
1
2
x3/2
dx.
where we used the bound
(12). To complete the proof of (4), notice that we
R
2
dx = 4.
have 1/m ≤ 2r 2 and 1+∞ x3/2
3. Statistical applications in fixed-design regression
Numerous works have established a close relationship between localized
Gaussian widths and the performance of statistical and compressed sensing procedures. Some of these works are reviewed below.
• In a regression problem with random design where the design and the
target are subgaussian, Lecué and Mendelson [21] established that two
quantities govern the performance of empirical risk minimizer over a
convex class F. These two quantities are defined using the Gaussian
width of the class F intersected with an L2 ball [21, Definition 1.3],
• If p, p′ > 1 are such that p′ ≤ p ≤ +∞ and log(2n)/(log(2en) ≤ p′ .
Gordon et al. [14] provide precise estimates of ℓ(Bp ∩ sBp′ ) where
Bp ⊂ Rn is the unit Lp ball and sBp′ is the Lp′ ball of radius s > 0.
These estimates are then used to solve the approximate reconstruction
problem where one wants to recover an unknown high dimensional
vector from a few random measurements [14, Section 7].
• Plan et al. [28] shows that in the semiparametric single index model,
if the signal is known to belong to some star-shaped set T ⊂ Rn , then
the Gaussian width of T and its localized version characterize the gain
obtained by using the additional information that the signal belongs
to T , cf. Theorem 1.3 in [28].
• Finally, Chatterjee [9] exhibits connection between localized Gaussian
widths and shape-constrained estimation.
These results are reminiscent of the isomorphic method [17, 3, 2], where localized expected supremum of empirical processes are used to obtain upper
bounds on the performance of Empirical Risk Minimization (ERM) procedures. These results show that Gaussian width estimates are important to
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
8
understand the statistical properties of estimators in many statistical contexts.
In Proposition 1, we established an upper bound on the Gaussian width of
M -convex hulls. We now provide some statistical applications of this result
in regression with fixed-design. We will use the following Theorem from [7].
Theorem 4 ([7]). Let K be a closed convex subset of Rn and ξ ∼ N (0, σ 2 In×n ).
Let f 0 ∈ Rn be an unknown vector and let y = f 0 + ξ. Denote by f ∗0 the
projection of f 0 onto K. Assume that for some t∗ > 0,
t2
1
ξ T (u − f ∗0 ) ≤ ∗ .
E
sup
n
2
u∈K: 1 |f ∗ −u|2 ≤t2∗
n
0
(13)
2
Then for any x > 0, with probability greater than 1 − e−x , the Least Squares
estimator fˆ = argminf ∈K |y − f |22 satisfies
1
4σ 2 x
1 ˆ
|f − f 0 |22 ≤ |f ∗0 − f 0 |22 + 2t2∗ +
.
n
n
n
Hence, to prove an oracle inequality of the form (14), it is enough to prove
the existence of a quantity t∗ such that (13) holds. If the convex set K in
the above theorem is the convex hull of M points, then a quantity t∗ is given
by the following proposition.
Proposition 5. Let σ 2 > 0, R > 0, n ≥ 1 and M ≥ 2. Let µ1 , ..., µM ∈ Rn
such that n1 |µj |22 ≤ R2 for all j = 1, ..., M . For all θ ∈ ΛM , let µθ =
P
a centered Gaussian random variable with covarij=1,...,M θj µj . Let g be √
ance matrix σ 2 In×n . If R n ≤ M σ then the quantity
t2∗ = 31σR
v
u
σ
u log eM
√
t
R n
n
satisfies
1
E
sup
n θ∈ΛM : 1 |µ
n
2
2
θ |2 ≤t∗
g T µθ ≤
t2∗
,
2
(14)
provided that t∗ ≤ R.
Proof. Inequality
q
1
√ E
(σg)T µθ ≤ 4σR log (4eM min(1, r 2 /R2 )).
sup
n θ∈ΛM : 1 |µθ |2 ≤r2
n
2
is a reformulation of Proposition 1 using the notation of Proposition 5. Thus,
in order to prove (14), it is enough to establish that for γ = 31 we have
q
√
4eM σγ log(eM σ/(R n))
eM σ
γ2
√
√
log
≤
(∗) := 64 log
.
(15)
R n
4
R n
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
9
√
As 1 ≤ log(eM σ/(R n)) and log t ≤ t for all t > 0, the left hand side of the
previous display satisfies
√
eM σ
1
√
(∗) ≤ 64 log
+ log(4γ) + log log(eM σ/(R n)) ,
R n
2
eM σ
√
.
≤ 64(3/2 + log(4γ)) log
R n
Thus (15) holds if 64(3/2 + log(4γ)) ≤ γ 2 /4, which is the case if the absolute
constant is γ = 31.
Inequality (14) establishes the existence of a quantity t∗ such that
t2
1
E
sup
g T µθ ≤ ∗ ,
n µ∈T : 1 |µ|2 ≤t2∗
2
n
(16)
2
where T is the convex hull of µ1 , ..., µM . Consequences of (16) and Theorem 4 are given in the next subsections.
We now introduce two statistical frameworks where the localized Gaussian
width of an M -convex hull has applications: the Lasso estimator in highdimensional statistics and the convex aggregation problem.
3.1. Convex aggregation
Let f 0 ∈ Rn be an unknown regression vector and let y = f 0 + ξ be
an observed random vector, where ξ satisfies E[ξ] = 0. Let M ≥ 2 and
let f 1 , ..., f M be deterministic vectors in Rn . The set {f 1 , ..., f M } will be
referred to as the dictionary. For any θ = (θ1 , ..., θM )T ∈ RM , let f θ =
PM
M is given, the goal of the aggregation problem
j=1 θj f j . If a set Θ ⊂ R
induced by Θ is to find an estimator fˆ constructed with y and the dictionary
such that
1
1 ˆ
2
2
|f − f 0 |2 ≤ inf
|f θ − f 0 |2 + δn,M,Θ ,
(17)
θ∈Θ n
n
either in expectation or with high probability, where δn,M,Θ is a small quantity. Inequality (17) is called a sharp oracle inequality, where "sharp" means
that in the right hand side of (17), the multiplicative constant of the term
inf θ∈Θ n1 |f θ − f 0 |22 is 1. Similar notations will be defined for regression with
random design and density estimation. Define the simplex in RM by (6).
The following aggregation problems were introduced in [26, 34].
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
10
• Model Selection type aggregation with Θ = {e1 , ..., eM }, i.e., Θ is the
canonical basis of RM . The goal is to construct an estimator whose
risk is as close as possible to the best function in the dictionary. Such
results can be found in [34, 22, 1] for random design regression, in
[23, 10, 5, 11] for fixed design regression, and in [16, 6] for density
estimation.
• Convex aggregation with Θ = ΛM , i.e., Θ is the simplex in RM . The
goal is to construct an estimator whose risk is as close as possible to the
best convex combination of the dictionary functions. See [34, 20, 19, 33]
for results of this type in the regression framework and [29] for such
results in density estimation.
• Linear aggregation with Θ = RM . The goal is to construct an estimator whose risk is as close as possible to the best linear combination of
the dictionary functions, cf. [34, 33] for such results in regression and
[29] for such results in density estimation.
One may also define the Sparse or Sparse Convex aggregation problems:
construct an estimator whose risk is as close as possible to the best sparse
combination of the dictionary functions. Such results can be found in [31, 30,
33] for fixed design regression and in [24] for regression with random design.
These problems are out of the scope of the present paper.
A goal of the present paper is to provide a unified argument that shows
that empirical risk minimization is optimal for the convex aggregation problem in density estimation, regression with fixed design and regression with
random design.
Theorem 6. Let f 0 ∈ Rn , let ξ ∼ N (0, σ 2 In×n ) and define y = f 0 + ξ. Let
P
T
M
f 1 , ..., f M ∈ Rn and let f θ = M
j=1 θj f j for all θ = (θ1 , ..., θM ) ∈ R .
Let
θ̂ ∈ argmin |f θ − y|22 .
θ∈ΛM
Then for all x > 0, with probability greater than 1 − exp(−x),
1
4σ 2 x
1
|f θ̂ − f 0 |22 ≤ min |f θ − f 0 |22 + 2t2∗ +
,
n
n
θ∈ΛM n
√
√
σ/(R n))
4σ2 M 31σR log(eM
2
√
where t∗ = min
and R2 = 14 maxj=1,...,M n1 |f j |22 .
n ,
n
Proof of Theorem 6. Let V be the linear span of f 1 , ..., f M and let P ∈
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
11
Rn×n be the orthogonal projector onto V . If t2∗ = 4σ 2 M/n, then
1
ξT v =
E
sup
n v∈V : 1 |v|2 ≤t2∗
n
2
s
t2∗
E|P ξ|2 ≤
n
s
t2∗ q
E|P ξ|22 =
n
s
t2∗ σ 2 M
= t2∗ /2.
n
Let K be the convex hull of f 1 , ..., f M . Let f ∗0 be the convex projection of
f 0 onto K. We apply Proposition 5 to K − f ∗0 which is a convex hull of M
points, and for all v ∈ K, n1 |v|22 ≤ R2 . By (21) and (14), the quantity t∗
satisfies (13). Applying Theorem 4 completes the proof.
3.2. Lasso
We consider the following regression model. Let x1 , ..., xM ∈ Rn and assume
that n1 |xj |22 ≤ 1 for all j = 1, ..., M . We will refer to x1 , ..., xM as the
covariates. Let X be the matrix of dimension n×M with columns x1 , ..., xM .
We observe
y = f 0 + ξ,
ξ ∼ N (0, σ 2 In×n ).
(18)
where f 0 ∈ Rn is an unknown mean. The goal is to estimate f 0 using the
design matrix X.
Let R > 0 be a tuning parameter and define the constrained Lasso estimator [32] by
β̂ ∈ argmin |y − Xβ|22 .
(19)
β∈R M :|β|1 ≤R
Our goal will be to study the performance of the estimator (19) with respect
to the prediction loss
1
|f − Xβ̂|22 .
n 0
Let x1 , ..., xM ∈ Rn and assume that n1 |xj |22 ≤ 1 for all j = 1, ..., M . Let
X be the matrix of dimension n × M with columns x1 , ..., xM .
Theorem 7. Let R > 0 be a tuning parameter and consider the regression
model (18). Define the Lasso estimator β̂ by (19). Then for all x > 0, with
probability greater than 1 − exp(−x),
1
4σ 2 x
1
|Xβ̂ − f 0 |22 ≤
min
|Xβ − f 0 |22 + 2t2∗ +
,
n
n
β∈R M :|β|1 ≤R n
where t2∗ = min
√
4σ2 rank(X) 62σR
,
n
(20)
√
log(2eM σ/(R n))
√
.
n
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
12
Proof of Theorem 7. Let V be the linear span of x1 , ..., xM and let P ∈
Rn×n be the orthogonal projector onto V . If t2∗ = 4σ 2 rank(X)/n, then
1
ξT v =
E
sup
n v∈V : 1 |v|2 ≤t2∗
n
2
s
t2∗
E|P ξ|2 ≤
n
s
t2∗ q
E|P ξ|22 =
n
s
t2∗ σ 2 rank(X)
= t2∗ /2.
n
(21)
Let K be the convex hull of {±Rx1 , ..., ±RxM }, so that K = {Xβ : β ∈
RM : |β|1 ≤ R}. Let f ∗0 be the convex projection of f 0 onto K. We apply
Proposition 5 to K − f ∗0 which is a convex hull of 2M points of empirical
norm less or equal to R2 . By (21) and (14), the quantity t∗ satisfies (13).
Applying Theorem 4 completes the proof.
The lower bound [30, Theorem 5.4 and (5.25)] states that there exists an
√
absolute constant C0 > 0 such that the following holds. If log(1+eM/ n) ≤
√
C0 n, then there exists a design matrix X such that for all estimator fˆ,
σ 2 rank(X)
1
1
EXβ |Xβ−fˆ|22 ≥
min
n
C
M
0
β∈R :|β|1 ≤R
sup
n
, σR
v
u
u log 1 +
t
n
eM
√σ
R n
,
where for all f 0 ∈ Rn , Ef 0 denotes the expectation with respect to the
distribution of y ∼ N (f 0 , σ 2 In×n ). Thus, Theorem 7 shows that the Least
Squares estimator over the set {Xβ, β ∈ RM : |β|1 ≤ R} is minimax optimal. In particular, the right hand side of inequality (20) cannot be improved.
4. The anisotropic persistence problem in regression with
random design
Consider n iid observations (Yi , Xi )i=1,...,n where (Yi )i=1,...,n are real valued
and the (Xi )i=1,..,n are design random variables in RM with E[Xi XiT ] = Σ
for some covariance matrix Σ ∈ RM ×M . We consider the learning problem
over the function class
n
fβ :
fβ (x) = xT β for some β ∈ RM with |β|1 ≤ R
o
for a given constant R > 0. We consider the Emprical Risk Minimizer defined
by
β̂ =
argmin
β∈R M :|β|
1 ≤R
n
X
i=1
(Yi − β T Xi )2
(22)
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
13
This problem is sometimes referred to as the persistence problem or the
persistence framework [15, 4]. The prediction risk of fβ̂ is given by
R(fβ̂ ) = E[(fβ̂ (X) − Y ) | (Xi , Yi )i=1,...,n],
where (X, Y ) is a new observation distributed as (X1 , Y1 ) and independent
from the data (Xi , Yi )i=1,...,n . Define also the oracle β ∗ by
β∗ =
argmin
R(β)
(23)
β∈R M :|β|1 ≤R
and define σ > 0 by
σ = kY − X T β ∗ kψ2 ,
(24)
√
where the subgaussian norm k·kψ2 is defined by kZkψ2 = supp≥1 E[|Z|p ]1/p / p
for any random variable Z (see Section 5.2.3 in [35] for equivalent definitions
of the ψ2 norm).
To analyse the above learning problem, we use the machinery developed
by Lecué and Mendelson [21] to study learning problems over subgaussian
classes. Consider the two quantities
(
rn (γ) = inf r > 0 : E
(
sn (γ) = inf s > 0 : E
sup
β:|β|1 ≤2R, E[(GT β)2 ]≤s2
T
√
)
β G ≤ γr n ,
sup
β: |β|1 ≤2R, E[(GT β)2 ]≤s2
β T G ≤ γs
2√
(25)
)
n/σ ,
(26)
where G ∼ N (0, Σ). In the present setting, Theorem A from Lecué and
Mendelson [21] reads as follows.
Theorem 8 (Theorem A in Lecué and Mendelson [21]). There exist absolute
constants c1 , c2 , c4 > 0 such that the following holds. Let R > 0. Consider
iid observations (Xi , Yi ) with E[Xi XiT ] = Σ. Assume that the design random
vectors Xi are subgaussian with respect to the covariance matrix Σ in the
sense that kXiT τ kψ2 ≤ 10|Σ1/2 τ |2 for any τ ∈ Rp . Define β ∗ by (23) and σ
by (24). Assume that the diagonal elements of Σ are no larger than 1. Then,
there exists absolute constants c0 , c1 , c2 , c3 > 0 such that the estimator β̂
defined in (22) satisfies
R(fβ̂ ) ≤ R(fβ∗ ) + max(s2n (c1 ), rn2 (c2 )),
with probability at least 1 − 6 exp(−c4 n min(c2 , sn (c1 ))).
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
14
In the isotropic case (Σ = IM ), [25] proves that
rn2 (γ)
≤
(
c3 R2
n
0
log( c3nM ) if n ≤ c4 M
if n > c4 M,
(27)
for some constants c3 , c4 > 0 that only depends on γ, while
s2n (γ) ≤
q
5 Rσ
c√
log( c√5 M σ )
n
c5 σ 2 M
n
nR
if n ≤ c6 σ 2 M 2 /R2
if n > c6 σ 2 M 2 /R2 ,
(28)
for some constants c5 , c6 > 0 that only depend on γ.
Using Proposition 1 and Equation (9) above lets us extend these bounds
to the anisotropic case where Σ is not proportional to the identity matrix.
Proposition 9. Let R > 0, let G ∼ N (0, Σ) and assume that the diagonal
elements of Σ are no larger than 1. For any γ > 0, define rn (γ) and sn (γ)
by (26) and (25). Then for any γ > 0, there exists constants c3 , c4 , c5 , c6 > 0
that depend only on γ such that (28) and (27) hold.
The proof of Proposition 9 will be given at the end of this subsection.
The primary improvement of Proposition 1 over previous results is that this
result is agnostic to the underlying covariance structure. This lets us handle
the anisotropic case with Σ 6= IM in the above proposition.
Proposition 9 combined with Theorem 8 lets us obtained the minimax rate
of estimation for the persistence problem in the anisotropic case. Although
the minimax rate was previously obtained in the isotropic case, we are not
aware of a previous result that yields this rate for general covariance matrices
Σ 6= IM .
Proof of Proposition 9. In this proof, c > 0 is an absolute constant whose
value may change from line to line. Let γ > 0. We first bound rn (γ) from
above. Let r > 0 and define
Tr (R) = {β ∈ Rp : |β|1 ≤ 2R, β T Σβ ≤ r 2 }.
The random variable X ∼ N (0, Σ) has the same distribution as Σ1/2 g where
g ∼ N (0, IM ). Thus, the expectation inside the infimum in (25) is equal to
E sup β T Σ1/2 g.
(29)
β∈Tr (R)
To bound rn (γ) from above, it is enough to find some r > 0 such that (29)
√
is bounded from above by γr n.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
15
By the Cauchy-Schwarz
inequality, the right hand side is bounded from
√
√
above by r M , which is smaller than γr n for all small enough s > 0
provided that n > c4 M for some constant c4 that only depends on γ.
We now bound rn (γ) from above in the regime n ≤ c4 M . Let u1 , ..., uM be
the columns of Σ and let T̃ be the convex hull of the 2M points {±u1 , ..., ±uM }.
Using the fact that Tr (R) = 2RTr/(2R) (1) ⊂ 2R(T̃ ∩ (r/(2R))B2 ), the right
hand side of the previous display is bounded from above by
q
r 2
2R ℓ(T̃ ∩ (r/R))B2 ) ≤ 8R log+ (4eM ( 2R
) ),
(30)
where we used Proposition 1 for
p the last inequality. By simple algebra, one
can show that if r = c3 (γ) √Rn log(c3 (γ)M/n) for some large enough constant c3 (γ) that only depends on γ, then the right hand side of (30) is
√
bounded from above by γr n.
We now bound sn (γ) from above. Let s > 0. By definition of sn (γ), to
prove that sn (γ) ≤ s, it is enough to show that
σEξ
sup β T Σ1/2 g
β∈Ts (R)
√
is smaller than γs2 n. We use Proposition 1 to show that the right hand
side of the previous display is bounded from above by
q
√
s 2
) ) .
cσ min s M , R log+ (4eM ( 2R
By simple algebra very similar to that of the proof of Proposition 5, we obtain
that if s2 equals the right hand side of (28) for large enough c5 = c5 (γ) and
c6 = c6 (γ), then the right hand side of the previous display is bounded from
√
above by γs2 n. This completes the proof of (28).
5. Bounded empirical processes and density estimation
We now prove a result similar to Proposition 1 for bounded empirical processes indexed by the convex hull of M points. This will be useful to study
the convex aggregation problem for density estimation. Throughout the paper, ε1 , ..., εn are i.i.d. Rademacher random variables that are independent
of all other random variables.
Proposition 10. There exists an absolute constant c > 0 such that the following holds. Let M ≥ 2, n ≥ 1 be integers and let b, R, L > 0 be real numbers.
Let Q(θ) = θ T Σθ for some semi-positive matrix Σ. Let Z1 , ..., Zn be i.i.d.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
16
random variables valued in some measurable set Z. Let h1 , ..., hM : Z → R
P
T
M
be measurable functions. Let hθ = M
j=1 θj hj for all θ = (θ1 , ..., θM ) ∈ R .
Assume that almost surely
Q(ej ) = Σjj ≤ R2 ,
|hj (Z1 )| ≤ b,
E[h2θ (Z1 )] ≤ LQ(θ),
(31)
√
for all j = 1, ..., M and all θ ∈ ΛM . Then for all r > 0 such that R/ M ≤
r ≤ R we have
E
"
sup
θ∈ΛM : Q(θ)≤r 2
where F (θ) =
√
#
F (θ) ≤ c max LR
1 Pn
i=1 εi hθ (Zi )|
n|
s
log(eM r 2 /R2 )
n
,
bR2 log(eM r 2 /R2 )
r2n
(32)
for all θ ∈
,
RM .
Proof of Proposition 10. Let m = ⌊R2 /r 2 ⌋ ≥ 1. The function F is convex
since it can be written as the maximum of two linear functions. Applying
Lemma 3 with the above notation and t = r yields
E
sup
θ∈ΛM :Q(θ)≤r 2
F (θ) ≤ E
Z
+∞
M (x)
1
dx
=
x2
Z
+∞
1
E [M (x)]
dx
.
x2
(33)
where the second inequality is a consequence of Fubini’s Theorem and for
all x ≥ 1,
M (x) =
max
F (θ).
2
2
θ∈ΛM
m : Q(θ)≤x(r +R /m)
Using (31) and the Rademacher complexity bound for finite classes given in
[18, Theorem 3.5], we obtain that for all x ≥ 1,
s
E[M (x)] ≤ c′ max
M
Lx(r 2 + R2 /m) log |ΛM
m | b log |Λm |
,
,
n
n
(34)
M
where c′ > 0 is a numerical constant and |ΛM
m | is the cardinal of the set Λm .
M
By definition of m we have r 2 ≤ R2 /m. The cardinal |ΛM
m | of the set Λm is
bounded from above by the right hand side of (12).
Combining
inequality
R +∞ dx
R +∞ dx
(33), inequality (34), the fact that the integrals 1
are
x2 and 1
x3/2
finite, we obtain
E
sup
θ∈ΛM :Q(θ)≤r 2
√
F (θ) ≤ c′′ max LR
s
log(eM/m) bm log(eM/m)
,
n
n
for some absolute constant c′′ > 0. By definition of m, we have R2 /(2r 2 ) ≤
m ≤ R2 /r 2 . A monotonicity argument completes the proof.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
17
Next, we show that Proposition 10 can be used to derive a condition
similar to (14) for bounded empirical processes. To bound from above the
performance of ERM procedures in density estimation, Theorem 13 in the
appendix requires the existence of a quantity r∗ > 0 such that
E
"
sup
θ∈ΛM : Q(θ)≤r∗2
#
F (θ) ≤
r∗2
,
16
(35)
where F is the function defined in Proposition 10 above.
To obtain such quantity r∗ > 0 under√the assumptions of Proposition 10,
we proceed as follows. Let K = max(b, L) and assume that
√
M K > R n.
q
√
Define r 2 = CKR log(eM K/(R n)) where C ≥ 1 is a numerical constant
that will be chosen later. We now bound from above the right hand side of
(32). We have
√
√
log(eM r 2 /R2 ) ≤ log(C) + log(eM K/(R n)) + (1/2) log log(eM K/(R n))
√
≤ (log(C) + 3/2) log(eM K/(R n)),
where for the last inequality we used that log log(u) ≤ log u for all u > 1 and
√
√
that log(C) ≤ log(C) log(eM K/(R n)), since C ≥ 1 and M K/(R n) ≥ 1.
Thus, the right hand side of (32) is bounded from above by
c max
p
log(C) + 3/2 log(C) + 3/2
,
C
C2
!
r2.
It is clear that the above quantity is bounded from above by r 2 /16 if the
numerical constant C is large enough. Thus we have proved that as long as
√
M K > R n, inequality (35) holds for
s
√
log(eM K/(R n))
r∗2 = CRK
,
n
where C ≥ 1 is a numerical constant.
ERM and convex aggregation in density estimation
The minimax optimal rate for the convex aggregation problem is known to
be of order
v
u
u log eM
√
M t
n
,
φC
M (n) := min
n
n
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
18
for regression with fixed design [30] and regression with random design [34]
√
√
√
if the integers M and n satisfy eM σ ≤ R n exp( n) or equivalently
φC
M (n) ≤ 1. The arguments for the convex aggregation lower bound from
[34] can be readily applied to density estimation, showing that the rate
φC
M (n) is a lower bound on the optimal rate of convex aggregation for density
estimation.
We now use the results of the previous sections to show that ERM is
optimal for the convex aggregation problem in regression with fixed design,
regression with random design and density estimation.
Theorem 11. There exists an absolute constant c > 0 such that the following holds. Let (Z, µ) be a measurable space with measure µ. Let p0 be
an unknown density with respect to the measure µ. Let Z1 , ..., Zn be i.i.d.
random variables valued in Z with density p0 . Let p1 , ..., pM ∈ L2 (µ) and let
P
T
M
pθ = M
j=1 θj pj for all θ = (θ1 , ..., θM ) ∈ R . Let
θ̂ ∈ argmin
θ∈ΛM
Z
p2θ dµ
n
2X
pθ (Zi ) .
−
n i=1
!
Then for all x > 0, with probability greater than 1 − exp(−x),
Z
(pθ̂ −p0 )2 dµ ≤ min
where R2 =
θ∈ΛM
1
4
Z
v
√
u
eM √b∞
u
p t log( R n ) 88b∞ x
b∞ M
+
b∞
(pθ −p0 )2 dµ+c max
,
R
,
n
n
3n
maxj=1,...,M
R
p2j dµ and b∞ = maxj=0,1,...,M kpj kL∞ (µ) .
√
If M b∞ ≤
Proof. It is a direct application of Theorem 13 in the appendix.
√
√
√
R n, a fixed point t∗ is given
by Lemma 14. If M b∞ > R n, we use
R ∗
Proposition 10 with Q(θ) = (p0 − pθ )2 , L = b∞ and b = b∞ . The bound
(35) yields the existence of a fixed point t∗ in this regime.
References
[1] Jean-Yves Audibert and Alexandre B. Tsybakov. Fast learning rates
for plug-in classifiers. Ann. Statist., 35(2):608–633, 04 2007. . URL
http://dx.doi.org/10.1214/009053606000001217.
[2] Peter L Bartlett and Shahar Mendelson. Empirical minimization. Probability Theory and Related Fields, 135(3):311–334, 2006.
[3] Peter L. Bartlett, Olivier Bousquet, and Shahar Mendelson. Local
rademacher complexities. Ann. Statist., 33(4):1497–1537, 08 2005. .
URL http://dx.doi.org/10.1214/009053605000000282.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
19
[4] Peter L Bartlett, Shahar Mendelson, and Joseph Neeman.
L1regularized linear regression: persistence and oracle inequalities. Probability theory and related fields, 154(1-2):193–224, 2012.
[5] Pierre C. Bellec.
Optimal bounds for aggregation of affine
estimators.
Annals of Statistics, to appear, 2017.
URL
https://arxiv.org/pdf/1410.0346v4.pdf.
[6] Pierre C. Bellec.
Optimal exponential bounds for aggregation
of density estimators. Bernoulli, 23(1):219–248, 2017. . URL
http://dx.doi.org/10.3150/15-BEJ742.
[7] Pierre C. Bellec. Sharp oracle inequalities for least squares estimators in
shape restricted regression. Annals of Statistics, to appear, 2017. URL
https://arxiv.org/pdf/1510.08029.pdf.
[8] Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Concentration
inequalities: A nonasymptotic theory of independence. Oxford University Press, 2013.
[9] Sourav Chatterjee. A new perspective on least squares under convex constraint. Ann. Statist., 42(6):2340–2381, 12 2014. . URL
http://dx.doi.org/10.1214/14-AOS1254.
[10] D. Dai, P. Rigollet, and T. Zhang. Deviation optimal learning using
greedy Q-aggregation. The Annals of Statistics, 40(3):1878–1905, 2012.
[11] Arnak S. Dalalyan and Alexandre B. Tsybakov. Aggregation by
exponential weighting and sharp oracle inequalities. In Learning
theory, volume 4539 of Lecture Notes in Comput. Sci., pages 97–
111. Springer, Berlin, 2007. ISBN 978-3-540-72925-9.
. URL
http://dx.doi.org/10.1007/978-3-540-72927-3_9.
[12] Frank den Hollander. Probability theory: The coupling method.
[13] Christophe Giraud. Introduction to high-dimensional statistics, volume
138. CRC Press, 2014.
[14] Yehoram Gordon, Alexander E Litvak, Shahar Mendelson, and Alain
Pajor. Gaussian averages of interpolated bodies and applications to
approximate reconstruction. Journal of Approximation Theory, 149(1):
59–73, 2007.
[15] Eitan Greenshtein, Ya’Acov Ritov, et al.
Persistence in highdimensional linear predictor selection and the virtue of overparametrization. Bernoulli, 10(6):971–988, 2004.
[16] Anatoli Juditsky, Philippe Rigollet, and Alexandre B. Tsybakov. Learning by mirror averaging. The Annals of Statistics, 36(5):2183–2206,
2008.
[17] Vladimir Koltchinskii. Local rademacher complexities and oracle inequalities in risk minimization. Ann. Statist., 34(6):2593–2656, 12 2006.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
20
. URL http://dx.doi.org/10.1214/009053606000001019.
[18] Vladimir Koltchinskii. Oracle Inequalities in Empirical Risk Minimization and Sparse Recovery Problems: Ecole d’Eté de Probabilités de SaintFlour XXXVIII-2008, volume 2033. Springer, 2011.
[19] Guillaume Lecué. Empirical risk minimization is optimal for the convex
aggregation problem. Bernoulli, 19(5B):2153–2166, 11 2013. . URL
http://dx.doi.org/10.3150/12-BEJ447.
[20] Guillaume Lecué and Shahar Mendelson. Aggregation via empirical
risk minimization. Probability theory and related fields, 145(3-4):591–
613, 2009.
[21] Guillaume Lecué and Shahar Mendelson. Learning subgaussian classes:
Upper and minimax bounds. arXiv preprint arXiv:1305.4825, 2013.
[22] Guillaume Lecué and Philippe Rigollet. Optimal learning with Qaggregation. Ann. Statist., 42(1):211–224, 2014.
[23] Gilbert Leung and Andrew R. Barron. Information theory and mixing
least-squares regressions. Information Theory, IEEE Transactions on,
52(8):3396–3410, 2006.
[24] K Lounici. Generalized mirror averaging and D-convex aggregation.
Mathematical methods of statistics, 16(3):246–259, 2007.
[25] Shahar Mendelson. Learning without concentration. arXiv preprint
arXiv:1401.0304, 2014.
[26] Arkadi Nemirovski. Topics in non-parametric statistics. In Lectures
on probability theory and statistics (Saint-Flour, 1998), volume 1738 of
Lecture Notes in Mathematics. Springer, Berlin, 2000.
[27] Gilles Pisier. Remarques sur un résultat non publié de b. maurey. Séminaire Analyse fonctionnelle, pages 1–12, 1980.
[28] Yaniv Plan, Roman Vershynin, and Elena Yudovina. High-dimensional
estimation with geometric constraints. arXiv preprint arXiv:1404.3749,
2014.
[29] Ph. Rigollet and A. B. Tsybakov. Linear and convex aggregation of
density estimators. Math. Methods Statist., 16(3):260–280, 2007. . URL
http://dx.doi.org/10.3103/S1066530707030052.
[30] Philippe Rigollet and Alexandre Tsybakov. Exponential screening and
optimal rates of sparse estimation. The Annals of Statistics, 39(2):731–
771, 2011.
[31] Philippe Rigollet and Alexandre B. Tsybakov. Sparse estimation by
exponential weighting. Statist. Sci., 27(4):558–575, 11 2012. URL
http://dx.doi.org/10.1214/12-STS393.
[32] Robert Tibshirani. Regression shrinkage and selection via the lasso.
Journal of the Royal Statistical Society. Series B (Methodological),
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
21
pages 267–288, 1996.
[33] A.B. Tsybakov.
Aggregation and minimax optimality in highdimensional estimation. In Proceedings of the International Congress
of Mathematicians, Seoul, 2014. To appear.
[34] Alexandre B. Tsybakov. Optimal rates of aggregation. In Learning
Theory and Kernel Machines, pages 303–313. Springer, 2003.
[35] Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. arXiv preprint arXiv:1011.3027, 2010.
Appendix A: Proof of the lower bound (5)
Proof of Proposition 2. By the Varshamov-Gilbert extraction lemma [13,
Lemma 2.5], there exist a subset Ω of {0, 1}M such that
|ω|0 = m,
|ω − ω ′ |0 > m,
log |Ω| ≥ (m/2) log(M/(5m))
for any distinct ω, ω ′ ∈ Ω.
For each ω ∈ Ω, we define s(ω) ∈ {−1, 0, 1}M , a signed version of ω,
as follows. Let ε1 , ..., εM be M iid Rademacher random variables. Then we
have
E[|
M
X
j=1
ωj εj µj |22 ] =
M
X
j=1
ωj |µj |22 = m.
Hence, there exists some s(ω) ∈ {−1, 0, 1}M with |s(ω)j | = ωj for all j =
1, ..., M such that |µs(ω) |22 ≤ m.
Define TΩ = {s2 µs(ω) , ω ∈ Ω}. Since s2 = 1/m, each element of TΩ is of
the form (1/m)(±µj1 ± ... ± µjm ) where µj1 , ..., µjm are m distinct elements
of {µ1 , ..., µM }, hence by convexity of T we have TΩ ⊂ T . By definition of
s(ω), it holds that TΩ ⊂ sB2 , and thus TΩ ⊂ T ∩ sB2 . For any two distinct
u, v ∈ TΩ ,
|u − v|22 ≥ κ2 s4 sup |s(ω) − s(ω ′ )|22 > κ2 s4 m = κ2 s2 ,
ω,ω ′
where the supremum is taken over any two distinct elements of Ω. By Sudakov’s inequality (see for instance [8, Theorem 13.4]) we have
√
p
√ q
ℓ(T ∩ sB2 ) ≥ ℓ(TΩ ) ≥ (1/2)κs log Ω ≥ 1/(2 2)κs m log(M/5m).
Since 1/m = s2 , the right hand side of the previous display is equal to the
right hand side of (5) and the proof is complete.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
22
Appendix B: Local Rademacher complexities and density
estimation
In the last decade emerged a vast literature on local Rademacher complexities to study the performance of empirical risk minimizers (ERM) for general
learning problems, cf. [3, 2, 17] and the references therein. The following result is given in [3, Theorem 2.1]. Let ε1 , ..., εn be independent Rademacher
random variables, that are independent from all other random variables considered in the paper.
Theorem 12 (Bartlett et al. [3]). Let Z1 , ..., Zn be i.i.d. random variables
valued in some measurable space Z. Let H : Z → [−b∞ , b∞ ] be a class of
measurable functions. Assume that there is some v > 0 such that E[h(Z1 )2 ] ≤
v for all h ∈ H. Then for all x > 0, with probability greater than 1−exp(−x),
n
1X
εi h(Zi ) +
sup (P − Pn )h ≤ 4E sup
h∈H
h∈H n i=1
#
"
r
2vx 8b∞ x
+
.
n
3n
Theorem 12 is a straightforward consequence of Talagrand inequality. We
now explain how Theorem 12 can be used to derive sharp oracle inequalities
in density estimation.
Theorem 13. Let (Z, µ) be a measurable space with measure µ. Let p0 be an
unknown density with respect to the measure µ. Let Z1 , ..., Zn be i.i.d. random
variables valued in Z with density p0 . Let P
be a convex subset of RL2 (µ).
R
∗
Assume that there exists p0 ∈ P such that (p0 − p∗0 )2 dµ = inf p∈P (p0 −
p)2 dµ. Assume that for some t∗ > 0,
E
p∈P:
R sup∗
(p−p0 )2 dµ≤t2∗
n
1X
n
i=1
εi (p − p∗0 )(Zi ) ≤
t2∗
.
16
(36)
Assume that there exists an estimator p̂ such that almost surely,
p̂ ∈ argmin
p∈P
Z
n
1X
2p(Zi ) .
p dµ −
n i=1
!
2
Then for all x > 0, with probability greater than 1 − exp(−x),
Z
2
(p̂ − p0 ) dµ ≤ min
p∈P
Z
2
(p − p0 ) dµ + 2 max
t2∗ ,
4(kp0 kL∞ (µ) + 8b∞ /3)x
n
!
,
where b∞ = supp∈P kpkL∞ (µ) .
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
23
Proof of Theorem 13. By optimality of p̂ we have
Z
2
(p̂ − p0 ) dµ ≤
Z
(p∗0 − p0 )2 dµ + 2Ξp̂ .
where for all p ∈ P, Ξp is the random variable
Ξp = (P − Pn )(p − p∗0 ) −
1
2
Z
(p∗0 − p)2 dµ.
Let ρ = max t2∗ , 4(kp0 kL∞ (µ) + 8b∞ /3)x/n and define
H = h = p∗0 − p for some p ∈ P such that
Z
h2 dµ ≤ ρ .
The class H is convex, 0 ∈ H and t2∗ ≤ ρ so that h ∈ H implies
any linear form L,
1
1
sup
L(h), ≤ 2
R
ρ h∈H: h2 dµ≤ρ
t∗
!
t2∗
sup
L
h
R
ρ
h∈H: h2 dµ≤ρ
≤
1
t2∗
h∈H:
t2∗
ρh
Rsup
∈ H. For
L(h)
h2 dµ≤t2∗
so that by taking expectations, (36) holds if t2∗ is replaced by ρ.
For any h ∈ H, E[h(Z1 )2 ] ≤ kp0 kL∞ (µ) ρ and h is valued in [−2b∞ , 2b∞ ]
µ-almost surely. We apply Theorem 12 to the class H. This yields that with
probability greater than 1 − e−x , if p ∈ P is such that p80 − p ∈ H, then
s
2ρkp0 kL∞ (µ) x 16b∞ x
ρ
(P − Pn )(p∗0 − p) ≤ +
+
,
4
n
n
x
ρ
≤ + 2 kp0 kL∞ (µ) + 8b∞ /3
≤ ρ.
2
n
−x
On the same event of probability greater than
qR 1 − e , if p ∈ P is such that
R ∗
√
(p∗0 − p)2 dµ which belongs to
(p0 − p)2 dµ > ρ, consider h = ρ(p∗0 − p)/
H. We have (P − Pn )h ≤ ρ, which can be rewritten
(P
− Pn )(p∗0
− p) ≤
√
ρ
sZ
(p∗0
−
p)2 dµ
≤ ρ/2 +
Z
(p∗0 − p)dµ/2,
so that Ξp ≤ ρ/2 ≤ ρ. In summary, we have proved that on an event of
probability greater than 1 − e−x , supp∈P Ξp ≤ ρ. In particular, this holds for
p = p̂ which completes the proof.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
/
24
Appendix C: A fixed point t∗ for finite dimensional classes
Lemma 14. Consider the notations of Theorem 13 and assume that the
linear span of P is finite dimensional of dimension d. Then (36) is satisfied
for t2∗ = 256kp0 kL∞ (µ) d/n.
Proof. Let e1 , ..., ed be anRorthonormal basis of the linear span of P, for the
scalar product hp1 , p2 i = p1 p2 dµ. Then
E
p∈P:
R sup∗
(p−p0 )2 dµ≤t2∗
d
n
n
X
1X
1X
ej (Xi )
εi (p − p∗0 )(Zi ) ≤ E
sup
εi
n i=1
θ∈R d : |θ|2 ≤t2∗ n i=1
j=1
2
v
u
d
uX
u
≤ t∗ t
j=1
≤
!2
n
1X
εi ej (Xi )
n i=1
q
t∗ kp0 kL∞ (µ) d
t2
√
= ∗,
n
16
where we have used the Cauchy-Schwarz inequality, Jensen’ inequality, and
that Eej (X)2 ≤ kp0 kL∞ (µ) for all j = 1, ..., d.
imsart-generic ver. 2014/07/30 file: deviations.tex date: September 28, 2017
| 10math.ST
|
Physeter catodon localization by sparse coding
arXiv:1306.3058v1 [cs.LG] 13 Jun 2013
Sébastien PARIS
DYNI team, LSIS CNRS UMR 7296, Aix-Marseille University
[email protected]
Yann DOH
DYNI team, LSIS CNRS UMR 7296, Université Sud Toulon-Var
[email protected]
Hervé GLOTIN
DYNI team, LSIS CNRS UMR 7296, Université Sud Toulon-Var
[email protected]
Xanadu HALKIAS
DYNI team, LSIS CNRS UMR 7296, Université Sud Toulon-Var
[email protected]
Joseph RAZIK
DYNI team, LSIS CNRS UMR 7296, Université Sud Toulon-Var
[email protected]
Abstract
This paper presents a spermwhale’ localization architecture using jointly a bag-offeatures (BoF) approach and machine learning framework. BoF methods are known, especially in computer vision, to produce from
a collection of local features a global representation invariant to principal signal transformations. Our idea is to regress supervisely from these local features two rough estimates of the distance and azimuth thanks
to some datasets where both acoustic events
and ground-truth position are now available. Furthermore, these estimates can feed
a particle filter system in order to obtain a
precise spermwhale’ position even in monohydrophone configuration. Anti-collision system and whale watching are considered applications of this work.
1. Introduction
Most of efficient cetacean localisation systems are
based on the Time Delay Of Arrival (TDOA) estimation from detected1 animal’s click/whistles signals
1
As click/whistles detector, matching filter is often prefered
Proceedings of the 30 th International Conference on Machine Learning, Atlanta, Georgia, USA, 2013. JMLR:
W&CP volume 28. Copyright 2013 by the author(s).
(Nosal & Frazer, 2006; Bénard & Glotin, 2009). Longbase hydrophones’array is involving several fixed, efficient but expensive hydrophones (Giraudet & Glotin,
2006) while short-base version is requiring a precise array’s self-localization to deliver accurate results. Recently (see (Glotin et al., 2011)), based on Leroy’s attenuation model versus frequencies (Leroy, 1965), a
range estimator have been proposed. This approach
is working on the detected most powerful pulse inside the click signal and is delivering a rough range’
estimate robust to head orientation variation of the
animal. Our purpose is to use i) these hydrophone’
array measurements recorded in diversified sea conditions and ii) the associated ground-truth trajectories of
spermwhale (obtained by precise TDAO and/or Dtag
systems) to regress both position and azimuth of the
animal from a third-party hydrophone2 (typically onboard, standalone and cheap model).
We claim, as in computer-vision field, that BoF approach can be successfully applied to extract a global
and invariant representation of click’s signals. Basically, the pipeline of BoF approach is composed of
three parts: i) a local features extractor, ii) a local feature encoder (given a dictionary pre-trained on
data) and iii) a pooler aggregating local representations into a more robust global one. Several choice for
encoding local patches have been developed in recent
years: from hard-assignment to the closest dictionary
basis (trained for example by Kmeans algorithm) to
2
We assume that the velocity vector is colinear with the
head’s angle.
Physeter catodon localization by sparse coding
a sparse local patch reconstruction (involving for example Orthognal Maching Pursuit (OMP) or LASSO
algorithms).
2. Global feature extraction by spare
coding
2.1. Local patch extraction
Let’s denote by C , {C j }, j = 1, . . . , H the collection
of detected clicks associated with the j th hydrophone
of the array composed by H hydrophones. Each matrix C j is defined by C j , {cji }, i = 1, . . . , N j where
cji ∈ Rn is the ith click of the j th hydrophone. For
our Bahamas2 dataset (Giraudet & Glotin, 2006), we
choose typically n = 2000 samples surrounding the
detected click. The total number of available clicks is
H
P
equal to N =
Nj.
i=1
As local features, we extract simply some local signal patches of p ≤ n samples (typically p = 128)
and denoted by z ji,l ∈ Rp . Furthermore all z ji,l are
`2 normalized. For each cji , a total of L local patches
n
e samZ ji , {z ji,l }, l = 1, . . . , L equally spaced of d L
ples are retrieved (see Fig. 1). All local patches associated with the j th hydrophone is denoted by Z j ,
{Z ji }, i = 1, . . . , N j while Z , {Z j } is denoting all
the local patches matrix for all hydrophones. A final
post-processing consists in uncorrelate local features
by PCA training and projection with p0 ≤ p dimensions.
2.2. Local feature encoding by sparse coding
In order to obtain a global robust representation of
c ⊂ C, each associated local patch z ⊂ Z are first
linearly encoded via the vector α ∈ Rk such as z ≈
Dα where D , [d1 , . . . , dk ] ∈ Rp×k is a pre-trained
dictionary matrix whose column vectors respect the
constraint dTj dj = 1. In a first attempt to solve this
linear problem, α can be the solution of the Ordinary
Least Square (OLS) problem:
1
2
lOLS (α|z; D) , min
kz − Dαk2 .
(1)
α∈Rk 2
OLS formulation can be extended to include regularization term avoiding overfitting. We obtain the ridge
regression (RID) formulation:
1
kz − Dαk22 + βkαk22 . (2)
lRID (α|z; D) , min
α∈Rk 2
This problem have an analytic solution α = (D T D +
βI k )−1 D T z. Thanks to semi-positivity of D T D +
βI k , we can use a cholesky factor on this matrix to
solve efficiently this linear system. In order to decrease
reconstruction error and to have a sparse solution, this
problem can be reformuled as a constrained Quadratic
Problem (QP):
1
lSC (α|z; D) , min kz − Dαk22 s.t. kαk1 = 1. (3)
α∈Rk 2
To solve this problem, we can use a QP solver involving
high combinatorial computation to find the solution.
Under RIP assumptions (Tibshirani, 1994), a greedy
approach can be used efficiently to solve and eq. 3 and
this latter can be rewritten as:
1
lSC (α|z; D) , min kz − Dαk22 + λkαk1 ,
k
α∈R 2
(4)
where λ is a regularization parameter which controls
the level of sparsity. This problem is also known as
basis pursuit (Chen et al., 1998) or the Lasso (Tibshirani, 1994). To solve this problem, we can use the
popular Least angle regression (LARS) algorithm.
2.3. Pooling local codes
The objective of pooling (Boureau et al.; Feng et al.)
is to transform the joint feature representation into
a new, more usable one that preserves important information while discarding irrelevant detail. For each
click signal, we usually compute L codes denoted V ,
{αi }, i = 1, . . . , L. Let define v j ∈ RL , j = 1, . . . , k as
the j th row vector of V . It is essential to use feature
pooling to map the response vector v j into a statistic value f (v j ) from some spatial pooling operation f .
We use v j , the response vector, to summarize the joint
distribution of the j th compounds of local features over
the region of interest (ROI). We will consider the `µ norm pooling and defined by:
fn (v; µ) =
L
X
! µ1
|vm |µ
s.t. µ 6= 0.
(5)
m=1
The parameter µ determines the selection policy for
locations. When µ = 1, `µ -norm pooling is equivalent to sum-pooling and aggregates the responses over
the entire region uniformly. When µ increases, `µ norm pooling approaches max-pooling. We can note
the value of µ tunes the pooling operation to transit
from sum-pooling to max-pooling.
2.4. Pooling codes over a temporal pyramid
In computer vision, Spatial Pyramid Matching (SPM)
is a technic (introduced by (Lazebnik et al.)) which
improves classification accuracy by performing a more
Physeter catodon localization by sparse coding
0.25
0.2
20
0.15
40
0.1
60
0.05
80
0
100
−0.05
120
−0.1
0
500
1000
1500
2000
100
200
300
400
500
600
700
800
900
1000
Figure 1. Left: Example of detected click with n = 2000. Right: extracted local features with p = 128, L = 1000 (one
local feature per column).
robust local analysis. We will adopt the same strategy
in order to pool sparse codes over a temporal pyramid
(TP) dividing each click signal into ROI of different
sizes and locations. Our TP is defined by the matrix
Λ of size (P × 3) (Paris et al.):
risk RM :
RM (V , D) ,
M
1 X1
kz i − Dαi k22 + λkαi k1
M i=1 2
(7)
s.t. dTj dj = 1.
Λ = [a, b, Ω],
(6)
where a, b, Ω are 3 (P × 1) vectors representing subdivision ratio, overlapping ratio and weights respectively. P designs the number of layers in the pyramid.
Each row of Λ represents a temporal layer of the pyramid, i.e. indicates how do divide the entire signal into
sub-regions possibly overlapping. For the ith layer, the
i
click signal is divided into Di = b 1−a
bi +1c ROIs where
ai , bi are the ith elements of vector a, b respectively.
P
P
For the entiere TP, we obtain a total of D =
Di
i=1
ROIs. Each click signal c (n × 1) is divided into temporal ROI Ri,j , i = 1, . . . , P , j = 1, . . . , Di of size
(bai .nc × 1). All ROIs of the ith layer have the same
weight Ωi . For the ith layer,
by bbi .nc
ROIs are shifted
1 1 1
samples. A TP with Λ = 1 1
is designing a
1
2
4
2-layers pyramid with D = 1 + 4 ROIs, the entiere signal for the first layer and 4 half-windows of n2 samples
with 25% of overlapping for the second layer. At the
end of pooling stage over Λ, the global feature x ∈ Rd ,
d = D.k is defined by the weighted concatenation (by
factor Ωi ) of L pooled codes associated with c.
2.5. Dictionary learning
To encode each local features by sparse coding (see
eq. 4), a dictionary D is trained offline with an important collection of M ≤ N.L local features as input. One would minimize the regularized empirical
Unfortunatly, this problem is not jointly convex but
can be optimized by alternating method:
RM (V |D̂) ,
M
1 X1
kz i − D̂αi k22 + λkαi k1 ,
M i=1 2
(8)
which can be solved in parallel by LASSO/LARS and
then:
M
1 X1
kz i − D α̂i k22 s.t. dTj dj = 1.
M i=1 2
(9)
Eq. 9 have an analytic solution involving a large matrix (k × k) inversion and a large memory occupation
for storing the matrix V (k × M ). Since M is potentially very large (up to 1 million), an online method to
update dictionary learning is prefered (Mairal et al.).
Figure 2 depicts 3 dictionary basis vectors learned via
sparse coding. As depicted, some elements reprensents
more impulsive responses while some more harmonic
responses.
RM (D|V̂ ) ,
3. Range and azimuth logistic
regression from global features
After the pooling stage, we extracted unsupervisly N
global features X , {xi } ∈ Rd×N . We propose to
regress via logistic regression both range r and azimuth az (in x − y plan, when animal reach surface
to breath) from the animal trajectory groundtruth denoted y. For the current train/test splitsets of the
Physeter catodon localization by sparse coding
4
x 10
0.1
−1
0
−1.1
−0.1
H
−1.2
−0.2
0
20
40
60
80
100
120
8
−1.3
0.1
−1.4
m
0
7
−1.5
H
−0.1
−1.6
9
H
−0.2
0
20
40
60
80
100
120
−1.7
0.2
H10
−1.8
0
−1.9
−0.2
−0.4
−2
0.7
0
20
40
60
80
100
H11
0.8
0.9
1
1.1
m
120
Figure 2. Example of trained dictionary basis with sparse
coding.
S
S
data, such as X = X train X test , y = y train y test
and N = Ntrain + Ntest , ∀ {xi , yi } ∈ X train × y train ,
we minimize:
(
)
NX
train
1 T
−yi w T
x
i
θ
b θ = arg min
w
log(1 + e
w w +C
) ,
wθ 2 θ θ
i=1
(10)
where yi denotes ri and azi for θ = r and θ = az respectively. Eq. 10 can be efficiently solved for example
with Liblinear software (Fan et al., 2008). In the test
part, range and azimuth for any xi ∈ X test are reconb Tr xi and by a
b Taz xi
tructed linearly by rbi = w
cz i = w
respectively.
1.2
1.3
1.4
1.5
4
x 10
Figure 3. The 2D trajectory (in xy plan) of the single
sperm whale observed during 25 min and corresponding
hydrophones positions.
l
l
l
represent the ground truth,
where yi,j
, ybi,j
and Ntest
the estimate and the number of test samples for the lth
hydrophone respectively. The global ARMSE is then
H
P
1
calculated by ARM SE = H
ARM SE(l).
l=1
4.2. `µ -norm pooling case study
For prilimary results, we investigate the influence of
the µ parameter during the pooling stage. We fix the
number of dictionary basis to k = 128 and the temporal pyramid equal to Λ1 = [1, 1, 1], i.e. we pool sparse
codes on whole the temporal click signal. A value of
4. Experimental results
ARMSE on range
4.1. bahamas2 dataset
1900
To extract local features, we chose n = 2000, p = 128
and L = 1000 (tuned by model selection). For both
the dictionary learning and the local features encoding, we chose λ = 0.2 and fixed 15 iterations to train
dictionary on a subset of M = 400.000 local features drawn uniformaly. We performed K = 10 crossvalidation where training sets reprensented 70% of the
total of extracted global features, the rest for the testing sets. Logistic regression parameter C is tuned by
model selection. We compute the average root mean
square error (ARMSE) of range/azimuth
estimates per
s
hydrophone: ARM SE(l) =
1
K
K
P
l
NP
test
i=1
j=1
l −y
l )2
(yi,j
bi,j
Λ1
1800
1700
ARMSE (m)
This dataset (Giraudet & Glotin, 2006) contains a total of N = 6134 detected clicks for H = 5 different
hydrophones (named H 7 , H 8 , H 9 , H 10 and H 11 and
with N 7 = 1205, N 8 = 1238, N 9 = 1241, N 10 = 1261
and N 11 = 1189 respectively).
1600
1500
1400
1300
1200
1100
0.25
0.5
1
1.5
2
2.5
µ
3
4
6
8
10
20
Figure 4. ARM SE vs. µ for range estimation.
µ = {3, 4} seems to be a good choice for this pooling
procedure. For µ ≥ 20, results are similar to those obtained by max-pooling. For azimuth, we observe also
the same range of µ values.
Physeter catodon localization by sparse coding
4.3. Range and azimuth regression results
Here, we fixed the value of µ = 3 and we varied the
number of dictionary basis k from 128 to 4096 elements. We also investigated the influence of the temporal pyramid and we give results fortwo particu
lary choices: Λ1 = [1, 1, 1] and Λ2 =
1
1
1
3
1
3
1
1
.
For Λ2 , the sparse are first pooled over all the signal
then pooled over 3 non-overlapping windows for a total of 1 + 3 = 4 ROIs. In order to compare results
of our presented method, we also give results for an
hand-craft feature (Glotin et al., 2011) specialized for
spermwhales and based on the spectrum of the most
energetic pulse dtected inside the click. This specialized feature, denoted Spectrum feature, is a 128 points
vector.
ARMSE on range
Λ1
1200
Λ2
Spectrum Feat
ARMSE (m)
1100
1000
900
temporal pyramid for pooling permits also to improve
slightly results.
5. Conclusions and perspectives
We introduced in the paper, for spermwhale localization, a BoF approach via sparse coding delivering
rough estimates of range and azimuth of the animal,
specificaly towarded for mono-hydrophone configuration. Our proposed method works directly on the
click signal without any prior pulses detection/analysis
while being robust to signal transformation issue by
the propagation. Coupled with non-linear filtering
such as particle filtering (Arulampalam et al., 2002),
accurate animal position estimation could be perform
even in mono-hydrophone configuration. Applications
for anti-collision system and whale whatching are targeted with this work.
As perspective, we plan to investigate other local features such as spectral features, MFCC (Davis & Mermelstein, 1980; Rabiner & Juang, 1993), Scattering
transform features (Andén & Mallat). These latter
can be considered as a hand-craft first layer of a deep
learning architecture with 2 layers.
800
References
700
600
128
256
512
1024
k
2048
3072
4096
Figure 5. ARM SE vs. k for range estimation with µ = 3.
Andén, Joakim and Mallat, Stéphane. Multiscale scattering for audio classification. In ISMIR, 11.
Arulampalam, M. Sanjeev, Maskell, Simon, and Gordon, Neil. A tutorial on particle filters for online
nonlinear/non-gaussian bayesian tracking. IEEE
Trans. SP, 50:174–188, 2002.
ARMSE on azimuth
Bénard, Frédéric and Glotin, Hervé. Whales localization using a large array : performance relative to cramer-rao bounds and confidence regions.
In e-Business and Telecommunications, pp. 294–
306. Springer - Verlag, Berlin Heidelberg, september
2009.
120
115
110
ARMSE (deg)
105
100
95
90
Boureau, Y-Lan, Ponce, Jean, and Lecun, Yann.
A theoretical analysis of feature pooling in visual
recognition. In ICML’ 10.
85
Λ1
80
Λ2
75
Spectrum Feat
128
256
512
1024
k
2048
3072
4096
Figure 6. ARM SE vs. k for azimuth estimation with µ =
3.
For both range and azimuth estimate, from k = 2048,
our method outperforms results of the Spectrum feature and particulary for azimuth estimate. Using a
Chen, Scott Shaobing, Donoho, David L., Michael, and
Saunders, A. Atomic decomposition by basis pursuit. SIAM Journal on Scientific Computing, 20:
33–61, 1998.
Davis, S. and Mermelstein, P. Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Trans.
ASSP, 28:357–366, 1980.
Physeter catodon localization by sparse coding
Fan, Rong-En, Chang, Kai-Wei, Hsieh, Cho-Jui,
Wang, Xiang-Rui, and Lin, Chih-Jen. LIBLINEAR:
A library for large linear classification. JMLR, 2008.
Feng, Jiashi, Ni, Bingbing, Tian, Qi, and Yan,
Shuicheng. Geometric lp -norm feature pooling for
image classification. In CVPR ’11.
Giraudet, Pascale and Glotin, Hervé. Real-time 3d
tracking of whales by precise and echo-robust tdoas
of clicks extracted from 5 bottom-mounted hydrophones records of the autec. Applied Acoustics,
67:1106–1117, 2006.
Glotin, H., Doh, Y., Abeille, R., and Monnin, A. Physeter distance estimation using sub-band leroy transmission loss model. In 5th Internationnal Workshop
on Detection, Classification, Localization and Density Estimation of Marine Mammals using Passive
Acoustics, 2011.
Lazebnik, Svetlana, Schmid, Cordelia, and Ponce,
Jean. Beyond bags of features: Spatial pyramid
matching for recognizing natural scene categories.
In CVPR ’06, pp. 2169–2178.
Leroy, C. Sound attenuation between 200 and 10000
cps mesured along single paths. Technical Report 43, Saclant ASW Research Center, 1965.
Mairal, Julien, Bach, Francis, Ponce, Jean, and Sapiro,
Guillermo. Online dictionary learning for sparse
coding. In ICML ’09.
Nosal, E.-M. and Frazer, L. Track of a sperm whale
from delays between direct and surface-reflected
clicks. Applied Acoustics, 67:1187–1201, 2006.
Paris, Sébastien, Halkias, Xanadu, and Glotin, Hervé.
Efficient bag of scenes analysis for image categorization. In ICPRAM’ 13.
Rabiner, L. and Juang, B.H. Fundamentals of Speech
Recognition. Prentice Hall PTR, 1993.
Tibshirani, Robert. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical
Society, Series B, 58:267–288, 1994.
| 5cs.CE
|
An In-place Framework for Exact and Approximate
Shortest Unique Substring Queries ⋆
Wing-Kai Hon1 , Sharma V. Thankachan2 , and Bojian Xu3 ⋆⋆
1
Department of CS, National Tsing Hua University, Taiwan
School of CSE, Georgia Institute of Technology, USA
3
Department of CS, Eastern Washington University, USA
[email protected], [email protected], [email protected]
arXiv:1512.00378v1 [cs.DS] 1 Dec 2015
2
Abstract. We revisit the exact shortest unique substring (SUS) finding problem, and propose its approximate version where mismatches are allowed, due to its applications in subfields such as computational biology. We design
a generic in-place framework that fits to solve both the exact and approximate k-mismatch SUS finding, using the
minimum 2n memory words plus n bytes space, where n is the input string size. By using the in-place framework,
we can find the exact and approximate k-mismatch SUS for every string position using a total of O(n) and O(n2 )
time, respectively, regardless of the value of k. Our framework does not involve any compressed or succinct data
structures and thus is practical and easy to implement.
Keywords: string pattern matching, shortest unique substring, in-place algorithms
1 Introduction
We consider a string S[1..n], where each character S[i] is drawn from an alphabet Σ = {1, 2, . . . , σ}. We
say the character S[i] occupies the string position i. A substring S[i..j] of S represents S[i]S[i + 1] . . . S[j]
if 1 ≤ i ≤ j ≤ n, and is an empty string if i > j. We call i the start position and j the ending position of
S[i..j]. We say the substring S[i..j] covers the kth position of S, if i ≤ k ≤ j. String S[i′ ..j ′ ] is a proper
substring of another string S[i..j] if i ≤ i′ ≤ j ′ ≤ j and j ′ − i′ < j − i. The length of a non-empty
substring S[i..j], denoted as |S[i..j]|, is j − i + 1. We define the length of an empty string as zero.
The Hamming distance of two non-empty strings A and B of equal length, denoted as H(A, B), is
defined as the number of string positions where the characters differ. A substring S[i..j] is k-mismatch
unique, for some k ≥ 0, if there does not exist another substring S[i′ ..j ′ ], such that i′ 6= i, j − i = j ′ − i′ ,
and H(S[i..j], S[i′ ..j ′ ]) ≤ k. A substring is a k-mismatch repeat if it is not k-mismatch unique.
Definition 1 (k-mismatch SUS). For a particular string position p in S and an integer k, 0 ≤ k ≤ n − 1,
the k-mismatch shortest unique substring (SUS) covering position p, denoted as SUS kp , is a k-mismatch
unique substring S[i..j], such that (1) i ≤ p ≤ j, and (2) there does not exist another k-mismatch unique
substring S[i′ ..j ′ ], such that i′ ≤ p ≤ j ′ and j ′ − i′ < j − i.
We call 0-mismatch SUS as exact SUS, and the case k > 0 as approximate SUS.
For any k and p, SUS kp must exist, because at least the string S can be SUS kp , if none of its proper
substrings is SUS kp . On the other hand, there might be multiple choices for SUS kp . For example, if S =
abcbb, SUS 02 can be either S[1, 2] = ab or S[2, 3] = bc, and SUS 12 can be either S[1..3] = abc or
⋆
⋆⋆
Authors are listed in alphabetical order. A preliminary version of this paper appears in Proceedings of the 26th International
Symposium on Algorithms and Computation (ISAAC), Nagoya, Japan, 2015.
Corresponding author. Phone: +1 (509) 359-2817. Fax: +1 (509) 359-2215.
S[2..4] = bcb. Note that in Definition 1, we require k < n, because finding SUS np is trivial: SUS np ≡ S for
any string position p.
Problem (k-mismatch SUS finding). Given the string S, the value of k ≥ 0, and two empty integer arrays
A and B, we want to work in the place of S, A, and B, such that, in the end of computation: (1) S does
not
of the rightmost4 SUS ki , i.e.,
change. (2) Each k(A[i], B[i]) pair saves the start and ending positions
2
S A[i]..B[i] = SUS i , using a total of O(n) time for k = 0 and O(n ) time for any k ≥ 1.
1.1
Prior work and our contribution
Exact SUS finding was proposed and studied recently by Pei et al. [7], due to its application in locating
snippets in document search, event analysis, and bioinformatics, such as finding the distinctness between
closely related organisms [3], polymerase chain reaction (PCR) primer design in molecular biology, genome
mapability [2], and next-generation short reads sequencing [1]. The algorithm in [7] can find all exact SUS
in O(n2 ) time using a suffix tree of O(n) space. Following their proposal, there has been a sequence of
improvements [8,5] for exact SUS finding, reducing the time cost from O(n2 ) to O(n) and alleviating the
underlying data structure from suffix tree to suffix array of O(n) space. Hu et al. [4] proposed an RMQ
(range minimum query) technique based indexing structure, which can be constructed in O(n) time and
space, such that any future exact SUS covering any interval of string positions can be answered in O(1)
time. In this work, we make the following contributions:
– We revisit the exact SUS finding problem and also propose its approximate version where mismatches
are allowed, which significantly increases the difficulty as well as the usage of SUS finding in subfields
such as bioinformatics, where approximate string matching is unavoidable due to genetic mutation and
errors in biological experiments.
– We propose a generic in-place algorithmic framework that fits to solve both the exact and approximate
k-mismatch SUS finding, using 2n words plus n bytes space. It is worth mentioning that 2n words plus
n bytes is the minimum memory space needed to save those n calculated SUSes: (1) It needs 2 words
to save each SUS by saving its start and ending positions (or one endpoint and its length) and there are
n SUSes. (2) It needs another n bytes to save the original string S in order to output the actual content
of any SUS of interest from queries. Note that all prior work [7,8,5,4] use O(n) space but there is big
leading constant hidden within the big-oh notation (see the experimental study in [5]).
– After the suffix array is constructed, all the computation in our solution happens in the place of two
integer arrays, using non-trivial techniques. It is worth noting that our solution does not involve any
compressed or succinct data structures, making our solution practical and easy to implement. Our preliminary experimental study shows that our solution for exact SUS finding is even faster than the fastest
one among [7,8,5]5 , in addition to a lot more space saving than them, enabling our solution to handle
larger data sets. Due to page limit, we will deliver the details of our experimental study in the journal
version of this paper.
2 Preparation
A prefix of S is a substring S[1..i], 1 ≤ i ≤ n. A proper prefix S[1..i] is a prefix of S where i < n. A
suffix of S is a substring S[i..n], denoted as Si , 1 ≤ i ≤ n. Si is a proper suffix of S, if i > 1.
4
5
Since any SUS may have multiple choices, it is our arbitrary decision to resolve the ties by picking the rightmost choice.
However, our solution can also be easily modified to find the leftmost choice.
Note that the work of [4] studies a different problem and its computation is of the query-answer model, and thus is not comparable
with [7,8,5] and ours.
2
For two strings A and B, we write A = B (and say A is equal to B), if |A| = |B| and H(A, B) = 0.
We say A is lexicographically smaller than B, denoted as A < B, if (1) A is a proper prefix of B, or (2)
A[1] < B[1], or (3) there exists an integer k > 1 such that A[i] = B[i] for all 1 ≤ i ≤ k−1 but A[k] < B[k].
The suffix array SA[1..n] of the string S is a permutation of {1, 2, . . . , n}, such that for any i and j,
1 ≤ i < j ≤ n, we have S[SA[i]..n] < S[SA[j]..n]. That is, SA[i] is the start position of the ith smallest
suffix in the lexicographic order. The rank array RA[1..n] is the inverse of the suffix array, i.e., RA[i] = j
iff SA[j] = i. The k-mismatch longest common prefix (LCP) between two strings A and B, k ≥ 0,
denoted as LCP k (A, B), is the LCP of A and B within Hamming distance k. For example, if A = abc and
B = acb, then: LCP 0 (A, B) is A[1] = B[1] = a and | LCP 0 (A, B)| = 1; LCP 1 (A, B) is A[1..2] = ab
and B[1, 2] = ac and | LCP 1 (A, B)| = 2.
Definition 2 (k-mismatch LSUS). For a particular string position p in S and an integer k, 0 ≤ k ≤ n − 1,
the k-mismatch left-bounded shortest unique substring (LSUS) starting at position p, denoted as LSUS kp ,
is a k-mismatch unique substring S[p..j], such that either p = j or any proper prefix of S[p..j] is not
k-mismatch unique.
We call 0-mismatch LSUS as exact LSUS, and the case k > 0 as approximate LSUS.
Observe that for any k, LSUS k1 = SUS k1 always exists, because at least the whole string S can be
LSUS k1 . However, for any k ≥ 0 and p ≥ 2, LSUS kp may not exist. For example, if S = dabcabc, none of
LSUS 0i and LSUS 1j exists, for all i ≥ 5, j ≥ 4. It follows that some string positions may not be covered by
any k-mismatch LSUS. For example, for the same string S = dabcabc, positions 6 and 7 are not covered
by any exact or 1-mismatch LSUS. On the other hand, if any LSUS kp does exist, there must be only one
choice for LSUS kp , because LSUS kp has its start position fixed on p and need to be as short as possible. Note
that in Definition 2, we require k < n, because finding LSUS np is trivial as LSUS n1 ≡ S and LSUS np does
not exist for all p > 1.
Definition 3 (k-mismatch SLS). For a particular string position p in S and an integer k, 0 ≤ k ≤ n − 1,
we use SLS kp to denote the shortest k-mismatch LSUS covering position p.
We call 0-mismatch SLS as exact SLS, and the case k > 0 as approximate SLS.
SLS kp may not exist, since position p may not be covered by any k-mismatch LSUS at all. For example,
if S = dabcabc, then none of SLS 0p and SLS 1p exists, for all p ≥ 6. On the other hand, if SLS kp exists, there
might be multiple choices for SLS kp . For example, if S = abcbac, SLS 02 can be either LSUS 01 = S[1..2]
or LSUS 02 = S[2..3], and SLS 13 can be any one of LSUS 11 = S[1..3], LSUS 12 = S[2..4], and LSUS 13 =
S[3..5]. Note that in Definition 3, we require k < n, because finding SLS np is trivial as SLS np ≡ S for all p.
Lemma 1. For any k and p: (1) LSUS k1 always exists. (2) If LSUS kp exists, then LSUS ki exists, for all
i ≤ p. (3) If LSUS kp does not exist, then none of LSUS ki exists, for all i ≥ p.
Proof. (1) LSUS k1 must exist, because at least the string S can be LSUS k1 if every proper prefix of S is a
k-mismatch repeat. (2) If LSUS kp exists, say LSUS kp = S[p..q], q ≥ p, then LSUS ki exists for every i ≤ p,
because at least S[i..q] is k-mismatch unique. (3) It is true, because otherwise we get a contradiction to the
second statement in the lemma.
⊔
⊓
Lemma 2. For any k and p, | LSUS kp | ≥ | LSUS kp−1 | − 1, if LSUS kp exists.
Proof. Suppose the k-mismatch substring LSUS kp = S[p..q], for some q ≥ p. Then, S[p − 1..q] is also
k-mismatch unique. It follows immediately that, | LSUS kp−1 | ≤ |S[p − 1..q]| = 1 + | LSUS kp |.
⊔
⊓
3
Lemma 3. For any k and p, SUS kp is either SLS kp or S[i..p], for some i, i + | LSUS ki | − 1 < p. That is,
SUS kp is either the shortest k-mismatch LSUS that covers position p, or a right extension (through position
p) of a k-mismatch LSUS.
Proof. We know SUS kp must exist, because at least the string S can be SUS kp . Let’s say SUS kp = S[i..j],
i ≤ p ≤ j. If S[i..j] is neither LSUS ki nor a right extension of LSUS ki , it means S[i..j] is a proper prefix
of LSUS ki and thus is a k-mismatch repeat, which is a contradiction to the fact that S[i..j] = SUS kp is
k-mismatch unique. Therefore, SUS kp = S[i..j] is either LSUS ki , or a right extension of LSUS ki (clearly,
j ≡ p in this case). Further, if SUS kp = S[i..j] = LSUS ki , it is obvious that LSUS ki must be the shortest
k-mismatch LSUS covering position p, i.e., SUS kp = SLS kp .
⊔
⊓
For example, let S = dabcabc, then: (1) SUS 03 can be either S[3..5] = LSUS 03 , or S[1..3], which is a
right extension of LSUS 01 = S[1]. (2) SUS 05 = S[4..5] = LSUS 04 . (3) SUS 06 = S[4..6], which is a right
extension of LSUS 04 = S[4..5]. (4) SUS 14 = S[3..5] = LSUS 13 . (5) SUS 16 = S[3..6], which is a right
extension of LSUS 13 .
The next lemma further says that if SUS kp is an extension of an k-mismatch LSUS, SUS kp can be quickly
obtained from SUS kp−1 .
Lemma 4. For any k and p, if SUS kp = S[i..p] and i + | LSUS ki | − 1 < p, i.e., SUS kp is a right extension
(through position p) of LSUS ki , then the following must be true: (1) p > 2; (2) the rightmost character of
SUS kp−1 is S[p − 1]; (3) SUS kp = SUS kp−1 S[p], the substring SUS kp−1 appended by the character S[p].
Proof. If SUS kp is a right extension (through position p) of a k-mismatch LSUS, it is certain that p > 1,
because SUS k1 ≡ LSUS k1 , which always exists (Lemma 1).
Because SUS kp is a right extension (through position p) of a k-mismatch LSUS, we have SUS kp =
S[i..p] for some i < p, and LSUS ki = S[i..j] for some j < p. We also know S[i..p − 1] is k-mismatch
unique, because the k-mismatch unique substring S[i..j] is a prefix of S[i..p − 1]. Note that any substring
starting from a position before i and covering position p − 1 is longer than the k-mismatch unique substring
S[i..p − 1], so SUS kp−1 must be starting from a position between i and p − 1, inclusive. Next, we show
SUS kp−1 actually must start at position i.
The fact that SUS kp = S[i..p] implies | LSUS kt | ≥ | SUS kp | = p − i + 1 for every t = i + 1, i + 2, . . . , p;
otherwise, rather than S[i..p], any one of these LSUS kt whose size is smaller than p − i + 1 would be a better
choice for SUS kp . That means, any k-mismatch unique substring starting from t = i + 1, i + 2, . . . , p − 1
has a length at least p − i + 1. However, |S[i..p − 1]| = p − i < p − i + 1 and S[i..p − 1] is k-mismatch
unique already and covers position p − 1 as well, so S[i..p − 1] is the only choice for SUS kp−1 . This also
means SUS kp is indeed the substring SUS kp−1 appended by the character S[p].
⊔
⊓
3 The High-Level Picture
In this section, we present an overview of our in-place framework for finding both the exact and approximate
SUS. The framework is composed of three stages, where all computation happens in the place of three arrays,
S, A, and B, each of size n. Arrays A and B are of integers, whereas array S always saves the input string.
The following table summarizes the roles of A and B at different stages by showing their content at the end
of each stage.
4
Stages
1
2
3
A[i]
B[i]
Used as temporary workspace during stage 1, but Ending position of LSUS ki , if LSUS ki exthe content is useless for stages 2 and 3.
ists; otherwise, NIL.
The largest j, such that LSUS kj is an SLS ki , if Ending position of LSUS ki , if LSUS ki exists; otherwise, NIL.
SLS ki exists; otherwise, NIL.
Start position of the rightmost SUS ki
Ending position of the rightmost SUS ki
Stage 1 (Section 4). We take the array S that saves the input string as input to compute LSUS ki for all
i, in the place of A and B. At the end of the stage, each B[i] saves the ending position of LSUS ki , if
LSUS ki exists. Since each existing LSUS ki has its start position fixed at i, at the end of stage 1, each existing
LSUS ki = S i..B[i] . For those non-existing k-mismatch LSUSes, we assign NIL to the corresponding
B array elements. The time cost of this stage is O(n) for exact LSUS finding (k = 0), and is O(n2 ) for
approximate LSUS finding, for any k ≥ 1.
Stage 2 (Section 5). Given the array B (i.e., the k-mismatch LSUS array of S) from stage 1, we compute
the rightmost SLS ki , the rightmost shortest LSUS covering position i, for all i, in the place of A and B. At
the end of stage 2, each A[i] saves the largest j, such that LSUS kj is an SLS ki , i.e., the rightmost SLS ki =
S A[i]..B[A[i]] , if SLS ki exists; otherwise, we assign A[i] = NIL. Array B does not change during stage 2.
The time cost of this stage is O(n), for any k ≥ 0.
Stage 3 (Section 6). Given A and B from stage 2, we compute SUS ki , for all i, in the place of A and B.
At the end of
stage 3, each (A[i], B[i]) pair saves the start and ending positions of the rightmost SUS ki , i.e.,
SUS ki = S A[i]..B[i] . The time cost of this stage is O(n), for any k ≥ 0.
Algorithms 1, 2, 3, and 4 in the appendix give the pseudocode of the in-place procedures that we will
describe in Sections 4.1, 4.2, 5, and 6, respectively.
4 Finding k-mismatch LSUS
The goal of this section is that, given the input string S and two integer arrays A and B, we want to work in
the place of A and B, such that B[i] saves the ending position of LSUS ki for all existing LSUS ki ; otherwise,
B[i] is assigned NIL. We take different approaches in finding the exact LSUS (k = 0) and approximate
LSUS (k ≥ 1).
4.1
Finding exact LSUS (k = 0)
Lemma 5 (Lemma 7.1 in [6]). Given the string S of size n, drawn from an alphabet of size σ, we can
construct the suffix array SA of S in O(n) time, using n + σ words plus n bytes, where the space of n bytes
saves S, the space of n words saves SA, and the extra space of σ words is used as the workspace for the run
of the SA construction algorithm.
Given the input string S, we first use the O(n)-time suffix array construction algorithm from [6] to create
the SA of S, where the array A is used to save the SA and the array B is used as the workspace. Note that
σ ≤ n is always true, because otherwise we will prune from the alphabet those characters that do not appear
in the string. After SA (saved in A) is constructed, we can easily spend another O(n) time to create the rank
array RA of S (saved in B): RA[SA[i]] ← i (i.e., B[A[i]] ← i), for all i. Next, we use and work in the place
of A (i.e., SA) and B (i.e., RA) to compute the ending position of each existing LSUS 0i and save the result
in B[i], using another O(n) time.
5
Definition 4.
LCP 0 S[i..n], S [SA [RA[i] − 1] ..n] , if RA[i] > 1
xi =
0,
otherwise
LCP 0 S[i..n], S [SA [RA[i] + 1] ..n] , if RA[i] < n
yi =
0,
otherwise
That is, xi (yi , resp.) is the length of the longest common prefix of S[i..n] and its lexicographically preceding
(succeeding, resp.) suffix, if the preceding (succeeding, resp.) suffix exists.
Fact 1 For every string position i, 1 ≤ i ≤ n:
S [i..i + max{xi , yi }] , if i + max{xi , yi } ≤ n
0
LSUS i =
not existing,
otherwise.
First, observe that in the sequence of xi ’s, if xi > 0, then xi+1 ≥ xi − 1 must be true, because at least
S[SA[RA[i] − 1] + 1..n] can be the lexicographically preceding suffix of S[i + 1..n], and they share the
leading xi − 1 characters. That means, when we compute xi+1 , we can skip over the comparisons of the
first xi − 1 pair of characters between S[i + 1..n] and its lexicographically preceding suffix. It follows that,
given the SA and RA of S and using the above observation, we can compute the sequence of xi ’s in O(n)
time. Using the similar observation, we can compute the sequence of yi ’s in O(n) time, provided that S and
its SA and RA are given.
Second, since we can compute the sequences of xi ’s and yi ’s in parallel (i.e., compute the sequence of
(xi , yi ) pairs), we can use Fact 1 to compute the sequence of LSUS 0i in O(n) time. Further, since RA[i] is
used only for retrieving the lexicographically preceding and succeeding suffixes of S[i..n] when we compute
the pair (xi , yi ), we can save each computed LSUS 0i (indeed, i+max{xi , yi }, the ending position of LSUS 0i )
in the place of RA[i] (i.e., B[i]). In the case i + max{xi , yi } > n, meaning LSUS 0i does not exist,we will
assign NIL to RA[j] (i.e., B[j]) for all j ≥ i (Lemma 1). The overall time cost for computing the sequence
of LSUS 0i is thus O(n), yielding the following lemma.
Lemma 6. Given the character array S of size n that saves the input string, and the integer arrays A and
B, each of size n, we can work in the place of S, A, and B, using O(n) time, such that at the end of the
computation, S does not change, B[i] saves the ending of position of LSUS 0i , if LSUS 0i exists (otherwise,
B[i] = NIL).
4.2
Finding approximate LSUS (k ≥ 1)
Definition 5. For a particular string position p in S and an integer k, 0 ≤ k ≤ n − 1, the k-mismatch
left-bounded longest repeat (LLR) starting at position p, denoted as LLR kp , is a k-mismatch repeat S[p..j],
such that either j = n or S[p..j + 1] is k-mismatch unique.
Fact 2 (1) If | LLR kp | < n − p + 1, i.e., the ending position of LLR kp is less than n, then LSUS kp =
S p..p + | LLR kp | , the substring of LLR kp appended by the character following LLR kp . (2) Otherwise,
LSUS kp does not exist.
Our high-level strategy for finding LSUS ki for all i is as follows. We first find LLR ki for all i. Then we
use Fact 2 to find each LSUS ki from LLR ki : If LLR ki does not end on position n, we will extend it for one
more character on its right side and make the extension to be LSUS ki ; otherwise, LLR ki does not exist. Next,
we explain how to find LLR ki , for all i.
6
Clearly, | LLR ki | = max{| LCP k (Si , Sj )|, j 6= i}, for all i. The way we calculate | LLR ki | for all i is
simply to let every pair of two distinct suffixes to be compared with each other. In order to do so, we work
over n − 1 phases, named as P1 through Pn−1 . On a particular phase Pδ , we compare suffixes Si and Si−δ
for all i = n, n − 1, . . . , δ + 1. Obviously, over these n − 1 phases, every pair of distinct suffixes have
been compared with each other exactly once. Over these n − 1 phases, we simply record in B[i], which is
initialized to be 0, the length of the longest k-mismatch LCP that each suffix Si has seen when compared
with any other suffixes. Next, we explain the details of a particular phase Pδ .
On a particular phase Pδ , 1 ≤ δ ≤ n−1, we compare suffixes Si and Si−δ for all i = n, n−1, . . . , δ +1.
When we compare Si and Si−δ , we save in A[1..k + 1], which is initialized to be empty at the beginning
of each phase, the leftmost mismatched k + 1 positions in Si . We will see later how to update A[1..k + 1]
efficiently over the progress of a particular phase and use it to update the B array.
We treat A[1..k + 1] as a circular array, i.e., i − 1 = k + 1 when i = 1, and i + 1 = 1 when i = k + 1. Let
size, which is initialized to be 0 at the beginning of each phase, denote the number of mismatched positions
being saved in A[1..k + 1] so far in Pδ . We can describe the work of phase Pδ , inductively, as follows.
1. We compare Sn and Sn−δ by only comparing S[n] and S[n − δ], since Sn = S[n].
(a) If S[n] 6= S[n − δ]: Save n in any position in A[1..k + 1]; size ← 1.
(b) B[n] ← max{B[n], 1}; B[n − δ] ← max{B[n − δ], 1}.
2. Suppose we have finished the comparison between the suffixes Si+1 and Si+1−δ , for some i, δ + 1 ≤
i ≤ n − 1. The leftmost k + 1 mismatched positions (if existing) between them have been saved in the
circular array A[1..k + 1]. Let A[cursor] be the element that is saving the first mismatched position (if
existing) between the two suffixes.
3. Next, we compare the suffixes Si and Si−δ by only comparing S[i] and S[i − δ], since Si+1 and Si+1−δ
have been compared. Remind that cursor − 1 below is in its cyclic manner.
(a) If S[i] 6= S[i − δ]: cursor ← cursor − 1; Save i in A[cursor] and overwrite the old content in
A[cursor] if there is; size ← min{size + 1, k + 1}.
(b) If size < k + 1: B[i] ← max{B[i], n − i + 1}; B[i − δ] ← max{B[i − δ], n − i + 1}.
(c) Else: B[i] ← max{B[i], A[cursor − 1] − i}; B[i − δ] ← max{B[i − δ], A[cursor − 1] − i}. Note
that A[cursor − 1] is saving the (k + 1)th mismatched position between Si and Si−δ .
After the computation of all LLR ki is finished, using the above n−1 phases, each B[i] is saving | LLR ki |.
Next, we can use Fact 2 to convert each LLR ki to LSUS ki by simply checking each B[i]: If i + B[i] − 1 < n,
i.e., LLR ki does not end on position n, then we assigne B[i] = i + B[i], the ending position of LSUS ki ;
otherwise, we assign B[i] = NIL, meaning LSUS ki does not exist.
The computation of all LLR ki takes n−1 phases and each phase clearly has no more than n comparisons,
giving a total of O(n2 ) time cost. The procedure of converting each LLR ki to LSUS ki spends another O(n)
time. Altogether, we get an O(n2 )-time in-place procedure for finding approximate LSUS, for any k ≥ 1.
Lemma 7. Given the character array S of size n that saves the input string, the integer arrays A and B,
each of size n, and the value of integer k ≥ 1, we can work in the place of S, A, and B, using O(n2 ) time,
such that at the end of the computation, S does not change, B[i] saves the ending of position of LSUS ki , i.e.,
LSUS ki = S[i..B[i]], if LSUS ki exists; otherwise, B[i] = NIL.
7
5 Finding k-mismatch SLS
Now we are given the array B, where each B[i] saves the ending position of LSUS ki if LSUS ki exists
and NIL otherwise. In this section, we want to work in the place of A and B, such that in the end of
computation: A[i] saves j, such that LSUS kj is the rightmost SLS ki , if such j exists; otherwise, A[i] = NIL.
That means, in the end of this section, the rightmost SLS ki = S A[i]..B[A[i]] , if SLS ki exists; otherwise,
A[i] = B[i] = NIL.
Recall that some k-mismatch LSUS may not exist and some positions may not be covered by any kmismatch LSUS (see the examples after Definition 2). Further, due to Lemmas 1 and 2, we know such
positions that are not covered by any k-mismatch LSUS must comprise a continuous chunk on the right end
of string S.
Definition 6. Let LSUS kr , 1 ≤ r ≤ n, be the rightmost existing k-mismatch LSUS of the input string S. Let
z, 1 ≤ z ≤ n, be the rightmost string position that is covered by any k-mismatch LSUS of the string S.
Again, due to Lemmas 1 and 2, it is trivial to find the values of r and z in O(n) time: scan array B (i.e.
LSUS array) from right to left, and stop when seeing the first non-NIL B array element, which is exactly
B[r], then z = B[r]. If z < n, we can then simply set A[i] = NIL for all i > z. Recall that B[i] = NIL
already for all i > r from stage 1. In the rest of this section, we only need to work with the two subarrays
A[1..z] and B[1..z], wanting to make A[i] to be the start position of the rightmost SLS ki , for all i ≤ z.
Let B[1..z] and an integer r, 1 ≤ r ≤ z, be the input, where (1) B[1..r] is of monotonically nondecreasing integers (Lemma 2), with i ≤ B[i], (2) B[r + 1..z] are all NIL, if r < z, and (3) B[r] = z.
We can use each B[i], i ≤ r, as a compact representation of the interval Ii = (i, B[i]). Let I = { Ii |
i ∈ [1..r] }, and ℓi = |B[i] − i + 1| be the length of Ii . Let A[1..z] be an output array such that A[j] = i,
where Ii is the rightmost shortest interval in I that covers j.
To illustrate the ideas and concepts that we will present in the rest of this section, let us use the following
as a running example, where r = 9, z = 15, and n = 17 (we add (0, B[0]) = (0, 0) as a sentinel).
i
B[i]
ℓi
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
0 3 4 7 8 10 10 10 11 15 − − − − − − − −
0 3 3 5 5 6 5 4 4 7 − − − − − − − −
pred[i] − − − 2 2 4 2 2 2 8 − − − − − − − −
ti
− 1 2 5 5 9 6 7 8 12 − − − − − − − −
max t−1
− 1 2 −− 4 6 7 8 5 − − 9 − − − − −
i
A[i]
− 1 2 2 2 4 6 7 8 8 8 8 9 9 9 9 − −
Definition 7. For an interval
Ii , wedefine the effective covering region with respect to the previous intervals
I<i = { Ik | k < i } to be ti , B[i] where
n
o
ti = max i, max {B[k] + 1 | Ik is shorter than Ii , k < i } .
We call ti the starting point of the effective covering region of Ii .
The effective covering region of Ii is exactly those regions that would set Ii as the answer, provided that all
the intervals I<i before Ii are present, and all the intervals I>i = { Ik | k > i } are absent.
−1
6
We next define t−1
i as a list , such that j ∈ ti if and only if tj = i. Observe that since ti ≥ i by definition,
any value j in t−1
i must have j ≤ i, and the effective region of Ij must cover i.
6
In actual run, t−1
saves the largest number in that list, as we will see more clearly later.
i
8
Lemma 8. For i = 1, 2, . . . , z:
A[i] = max
i
[
−1
t−1
k = max { A[i − 1], max ti }.
k=1
Si
Proof. Let j = max k=1 t−1
k . This means that for the effective region of any Ih , with h > j, none of
them covers i. Next, observe that Ij must cover i; otherwise, for all the intervals Ih with h < j, we have
B[h] ≤ B[j] < i, so that none of them can cover i, and thus a contradiction occurs. Finally, we show that
for those h < j, Ih can be pruned by Ij , thus implying that A[i] = j is a correct answer.
Consider all those h with h < j:
1. If Ih is longer than Ij , Ih can be pruned away directly.
2. Else, if Ih and Ij have equal length, Ih can be pruned away also, regardless of its coverage on i, since
we pick the rightmost shortest
interval that covers i.
Si
−1
3. Else, h must appear in k=1 tk . By the definition of tj , we have B[h] < tj ≤ i; thus, Ih does not cover
i, and can be pruned away.
Thus, the first equality in the lemma follows, while the second equality in the lemma is trivial once we have
the first equality.
⊔
⊓
Lemma 9. Suppose that all ti , 1 ≤ i ≤ r, can be generated incrementally in O(n) time. Then, we can
obtain all max t−1
i , 1 ≤ i ≤ z, in O(n) time.
Proof. We examine each ti , i = 1, 2, . . . , r, and write i at entry ti = j of the t−1 array; if such an entry
contains a value i′ already, we simply overwrite i′ with the latter i.
⊔
⊓
Indeed, we may scan ti from right to left, i.e., i = r, r − 1, . . . , 1, and update max t−1
as we proceed.
i
Firstly, if ti > i, we set t−1
=
undefined.
Else,
let
j
=
t
(whose
value
is
at
least
i),
and
we
check if t−1
i
i
j is
−1
defined: If not, simply set tj = i; otherwise, no update is needed.
The advantage of the ‘right-to-left’ approach is that we can construct t−1
i in-place, by re-using the memory
space of ti . To see why it is so, by the time we need to update a certain entry j = ti at step i, the information
tj has been used (and will never be used), so that we can safely overwrite the original entry, storing tj , to
store t−1
j instead. This gives the following corollary.
Corollary 1. Suppose that all ti ’s are generated, and are stored in a certain array A[1..z]. Then, we can
obtain max ti−1 for all i’s, in-place, by storing the results in the same array A[1..z]; the time cost is O(n).
Our goal is to make our algorithm in-place. Suppose that we can have in-place incremental generation of ti .
Then, by the above lemma, we may store max t−1
i temporarily at A[i]; afterwards, by the second equality
of Lemma 8, we can compute the correct output A by a simple scan of A from left to right.
Thus, to make the whole process in-place, it remains to show how ti can be computed in O(n) time, inplace. For this, we define pred[i] to be the largest j (if it exists) such that j < i and length of Ij is shorter than
Ii . It is easy to check that if pred[i] = j is defined, then ti = max { B[j] + 1, i } (and ti = i otherwise).7
Moreover, pred[i] for all i’s can be computed incrementally, with a way analogous to the construction of
the failure function in KMP algorithm: we check pred[i − 1], pred[pred[i − 1]], pred[pred[pred[i − 1]],
7
For each j ′ < j, if Ij ′ covers i, Ij would also cover i; in such a case, B[j] + 1 ≥ B[j ′ ] + 1. For each j ′ ∈ [pred[i], i − 1], Ij ′
is longer than Ii .
9
and so on, until we obtain j in the process such that Ij is shorter than Ii , and set pred[i] := j.8 If such j
does not exist, we set pred[i] = NIL. The running time is bounded by O(n).
This gives the following O(n)-time in-place algorithm (where B is read-only):
1. Compute pred[i], i = 1, 2, . . . , r, and store this in A[i]. Note that this step requires the length information of the intervals of Ii , which can be obtained in O(1) time, on the fly, from B[i] .
2. Scan A[1..r] (i.e., pred) incrementally, and obtain ti from the above discussion. The value of ti is stored
in A[i]. Note that this step requires the access to the original B.
3. Scan A[1..r] (i.e., ti ) from right to left, and obtain max t−1
i decrementally (stored in A[i]) by Corollary 1.
4. Scan A[1..z] (i.e., max t−1
i ) incrementally (i = 1, 2, . . . , z), and obtain the desired A[i] by the second
equality in Lemma 8.
Lemma 10. Given the integer array A and B, each of size n, where each B[i] saves the ending position
of LSUS ki , if LSUS ki exists and NIL otherwise, we can work in the place of array A and B, using O(n)
time, such that, in the end of computation, array B does not change, and A[i] saves j, where LSUS kj is the
rightmost SLS ki , if such j exists; otherwise, A[i] = NIL. That is, SLS ki = S A[i]..B[A[i]] , if SLS ki exists;
otherwise, A[i] = B[i] = NIL.
6 Finding k-mismatch SUS
Now we have array A, where A[i] = j, such that LSUS kj is the rightmost SLS ki , if position i is covered
by any k-mismatch LSUS; otherwise, A[i] = NIL. Note that A[i] = j is recording the start position of
the rightmost SLS ki already, because LSUS kj starts on position j. We also have array B, where B[i] =
i + | LSUS ki | − 1, the ending position of LSUS ki , if LSUS ki exists; otherwise, B[i] = NIL.
Step I. We want to transform A and B, such that each (A[i], B[i]) pair saves the start and ending positions of
SLS ki , if SLS ki exists; otherwise, we set (A[i], B[i]) = (NIL, NIL). Since each A[i] is already recording the
start position of SLS ki already, as we have explained at the beginning of this section, we only need to make
changes to array B. We first set B[i] = NIL for all i > z (Definition 6). Then, we scan array B from right
to left, starting from position z through 1, and set each B[i] = B[A[i]], the ending position of the rightmost
SLS ki . Because the leftmost position that any existing LSUS ki can cover is position i, we know A[i] ≤ i
and we no longer need B[i] (i.e., the information of LSUS ki ) after SLS i is computed. Therefore, it is safe to
record SLS ki by overwriting B[i] by B[A[i]] (i.e., the ending position of SLS ki ), in this right-to-left scan.
Step II. We use arrays A and B to calculate SUS ki for each i and save the result in the place of A and B,
i.e., each (A[i], B[i]) pair saves the start and ending position of SUS ki . Because of Lemma 3 and 4, we can
use arrays A and B to compute each SUS ki inductively, as follows:
1. SUS k1 = LSUS k1 = SLS k1 = S A[1]..B[1] .
2. For i = 2, 3, . . . , n, we compute SUS ki :
k
k
be SUS ki−1 appended by
(a) If (A[i], B[i]) = (NIL, NIL), meaning
SLS i does not exist,
we set SUS i to
k
k
the character S[i], i.e., SUS i = S A[i − 1]..B[i − 1] + 1 , and save SUS i by setting (A[i], B[i]) =
(A[i − 1], B[i − 1] + 1);
8
Intuitively, pred defines the shortcuts so that we can skip some intervals in I<i to compute ti .
10
(b) Else, if SUS ki−1 ends at position i − 1 and SUS ki−1 S[i] = S A[i − 1]..B[i − 1] + 1 is shorter than
SLS ki = S A[i]..B[i] , we set (A[i], B[i]) = (A[i − 1], B[i − 1] + 1);
(c) Else, SUS ki = SLS ki and thus we leave A[i] and B[i] unchanged.
Lemma 11. Given arrays A and B:
– A[i] = j, such that LSUS kj is the rightmost SLS ki , if SLS ki exists; otherwise, A[i] = NIL;
– B[i] = i + | LSUS ki | − 1, the ending position of LSUS ki , if LSUS ki exists; otherwise, B[i] = NIL.
we can work in the place of A and B, using O(n) time, such that,
in the end
of computation, each (A[i], B[i])
saves the start and ending positions of SUS ki , i.e., SUS ki = S A[i]..B[i] , i = 1, 2, . . . , n.
By concatenating the claims in Lemmas 6, 7, 10, and 11, we get the final result.
Theorem 1. Given the array S of size n that saves the input string, two empty integer arrays A and B,
each of size n, and the value of integer k ≥ 0, we can work in the place of arrays S, A, and B, using a
total of O(n) time for k = 0 and O(n2 ) time for any k ≥ 1, such that in the end of computation, S does
k
not change, each (A[i],
B[i]) pair represents the start and ending positions of the rightmost SUS i , i.e.,
k
SUS i = S A[i]..B[i] .
7 Conclusion
In this paper, we revisited the exact SUS finding problem, and proposed its approximate version where
mismatches are allowed, and thus significantly extended the usage of SUS finding in subfields such as
computational biology. We designed a generic in-place algorithmic framework that uses the minimum 2n
words plus n bytes space and can fit to find both exact and approximate k-mismatch SUS, with O(n) and
O(n2 ) time complexities, respectively, regardless of the value of any k ≥ 1. An urgent future work will
be researching for a faster (and still practical) in-place algorithm for finding approximate LSUS to replace
the current algorithm discussed in Section 4.2. Such new algorithm will lead to an overall faster in-place
solution for approximate SUS finding.
References
1. Adaş, B., Bayraktar, E., Faro, S., Moustafa, I., Külekci, M.O.: Nucleotide sequence alignment and compression via shortest
unique substring. In: Proceedings of the 3rd International Work-Conference on Bioinformatics and Biomedical Engineering
(IWBBIO). pp. 363–374 (2015)
2. Derrien, T., Estell, J., Marco Sola, S., Knowles, D.G., Raineri, E., Guig, R., Ribeca, P.: Fast computation and applications of
genome mappability. PLoS ONE 7(1), e30377 (01 2012)
3. Haubold, B., Pierstorff, N., Möller, F., Wiehe, T.: Genome comparison without alignment using shortest unique substrings. BMC
Bioinformatics 6, 123 (2005)
4. Hu, X., Pei, J., Tao, Y.: Shortest unique queries on strings. In: Proceedings of the 21st International Symposium on String
Processing and Information Retrieval (SPIRE). pp. 161–172 (2014)
5. İleri, A.M., Külekci, M.O., Xu, B.: A simple yet time-optimal and linear-space algorithm for shortest unique substring queries.
Theoretical Computer Science 562(0), 621 – 633 (2015), (also in CPM2014.)
6. Nong, G.: Practical linear-time O(1)-workspace suffix sorting for constant alphabets. ACM Transactions on Information Systems (TOIS) 31(3), 15:1–15:15 (Aug 2013)
7. Pei, J., Wu, W.C.H., Yeh, M.Y.: On shortest unique substring queries. In: Proceedings of IEEE International Conference on Data
Engineering (ICDE). pp. 937–948 (2013)
8. Tsuruta, K., Inenaga, S., Bannai, H., Takeda, M.: Shortest unique substrings queries in optimal time. In: Proceedings of International Conference on Current Trends in Theory and Practice of Computer Science (SOFSEM). pp. 503–513 (2014)
11
Appendix
Algorithm 1: Finding exact LSUS
Input: String S and integer arrays A and B, each of size n.
Output: S does not change. B[i] = ending position of LSUS 0i , if LSUS 0i exists; otherwise, B[i] = NIL.
1
2
Create the SA of S using the suffix array construction algorithm from [6], where array A is used to save the resulting SA
and B is used as the workspace for the run of the algorithm.
/* Create the RA of S and save the result in the array B.
*/
for i = 1 . . . n do RA[SA[i]] ← i ;
// i.e., SA = A, RA = B, and B[A[i]] ← i.
/* From here on, A and SA are the same physical array. B, RA, and LSUS are the
same physical array.
*/
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
x ← 0; y ← 0;
for i = 1, 2, . . . , n do
if RA[i] > 1 then
j ← SA[RA[i] − 1];
/* Calculate the length of the 0-mismatch LCP between S [i..n] and its
lexicographically preceding suffix.
while S [i + x] = S [j + x] do x ← x + 1
else x ← 0;
if RA[i] < n then
j ← SA[RA[i] + 1];
/* Calculate the length of the 0-mismatch LCP between S [i..n] and its
lexicographically succeeding suffix.
while S [i + y] = S [j + y] do y ← y + 1;
else y ← 0;
if i + max{x, y} ≤ n then LSUS [i] ← i + max{x, y};
else
for j = i . . . n do LSUS [j] ← NIL;
Break;
if x > 0 then x ← x − 1;
if y > 0 then y ← y − 1;
12
*/
*/
// ending position of | LSUS i |
// LSUS i does not exist. Early stop.
Algorithm 2: Finding approximate LSUS
Input: String S and integer arrays A and B, each of size n, the value of k ≥ 1.
Output: S does not change. B[i] = ending position of LSUS ki , if LSUS 0i exists; otherwise, B[i] = NIL.
1
for i = 1 . . . n do B[i] ← 0 ;
// Initialization
/* We use A[1 . . . k + 1] as a circular array to save the k + 1 most recently found
mismatched positions.
*/
2
capacity ← k + 1;
// The capacity of the circular array that records at most k + 1
mismatched positions.
3
cursor ← 1;
// The index of the circular array position that is saving the most
recently founded mismatched position. It can be initialized to be any value from
{1, 2, . . . , capacity}.
4
for δ = 1 . . . n − 1 do
// n − 1 phases
size ← 0 ; // The number of recorded mismatched positions in the circular array in
the current phase.
5
6
for i = n down to δ + 1 do
/* Comparing suffixes Si and Si−δ by comparing their leading characters, as
their remaining characters have been compared in previous steps of this
phase.
*/
7
8
9
10
if S[i] 6= S[i − δ] then
cursor ← (cursor − 2 + capacity) mod capacity + 1;
A[cursor] ← i;
size ← min(size + 1, capacity);
13
if size < capacity then
B[i] ← max(B[i], n − i + 1) ;
B[i − δ] ← max(B[i − δ], n − i + 1);
14
else
11
12
15
// We use 1-based indexing.
// = size − i + 1
B[i] ← max B[i], A (cursor − 1 + k) mod capacity + 1 − i ;
// We use 1-based
indexing.
16
17
18
19
B[i − δ] ← max B[i − δ], A (cursor − 1 + k) mod capacity + 1 − i ;
for i = 1 . . . n do
if B[i] = size − i + 1 then B[i] ← NIL;
else B[i] ← i + B[i];
// LSUS ki does not exist.
// The ending position of LSUS ki .
13
Algorithm 3: Finding SLS (exact or approximate)
Input: Integer arrays A and B, each of size n. Each B[i] saves the ending position of LSUS ki , if LSUS ki exists; NIL,
otherwise.
Output: Array B does not change. Each A[i] = j, such that LSUS kj is the rightmost SLS ki , if SLS ki exists; otherwise,
A[i] = NIL.
1
/* Find the index of the rightmost existing k-mismatch LSUS.
for r = n down to 1 do if B[r] 6= NIL then break;
*/
/* Compute the pred array, using the memory space of A array. pred[i] is the largest
j, such that j < i and | LSUS kj | < |LSU Sik |, if such j exists; otherwise pred[i] = NIL.
If LSUS ki does not exist, pred[i] = NIL also. From here on, pred and A are the same
physical array.
*/
2
if r < n then for i = r + 1 . . . n do pred[i] ← NIL ; // Positions that do not have k-mismatch LSUS
3
pred[1] ← NIL;
for i = 2 . . . r do
ℓi ← B[i] − i + 1 ;
j ← i − 1;
while pred[j] 6= NIL and B[j] − j + 1 >= ℓi do j ← pred[j];
if B[j] − j + 1 < ℓi then pred[i] ← j; else pred[i] ← NIL;
4
5
6
7
8
// | LSUS ki |
/* Compute the t array, using the memory space of A array. t[i] is the start
position of the effective region of LSUS ki , if LSUS ki exists; NIL, otherwise.
From here on, t and A are the same physical array.
*/
9
10
for i = 1 . . . r do
if pred[i] = NIL then t[i] ← i; else t[i] ← max(B[pred[i]] + 1, i);
/* Compute the t−1 array, using the memory space of array A. t−1 [i] is the largest
j, such that the effective region of LSUS kj starts on position i, if such j
exists; otherwise, NIL. From here on, t−1 and A are the same physical array. */
11
12
13
for i = r down to 1 do
if t−1 [t[i]] = NIL then t−1 [t[i]] ← i;
if i < t[i] then t−1 [i] ← NIL;
// Enable us to update this place in the future when
needed.
/* Compute SLS array using the memory space of array A. SLS [i] = j, such that LSUS kj
is the rightmost SLS ki , if SLS ki exists; NIL, otherwise. From here on, SLS and A
are the same physical array.
*/
14
15
SLS [1] ← 1;
for i = 2 . . . B[r] do SLS [i] ← max(SLS [i − 1], t−1 [i]);
14
Algorithm 4: Finding SUS (exact or approximate)
Input: Integer arrays A and B, each of size n. (1) A[i] = j, such that LSUS kj is the rightmost SLS ki , if SLS ki exists;
otherwise, A[i] = NIL. (2) B[i] is the ending position of LSUS ki , if LSUS ki exists; otherwise, B[i] = NIL.
Output: Each (A[i], B[i]) pair represents the start and ending positions of SUS ki .
1
2
3
4
5
6
7
8
9
for i = n down to 1 do
if B[i] 6= NIL then
z ← i + B[i] − 1;
LSUS.
break;
// The rightmost position covered by at least one k-mismatch
if z < n then
for i = z + 1 . . . n do
B[i] ← NIL ;
// Positions not covered by any k-mismatch LSUS.
for i = z down to 1 do
B[i] ← B[A[i]] ;
// The ending position of SLS ki .
/* By this point, S A[i]..B[i] = SLS ki , if SLS ki exists; otherwise A[i] = B[i] = NIL.
Note that SUS k1 = SLS k1 = S A[1]..B[1] , which must be existing and has been
computed.
Next, we compute SUS ki for all i ≥ 2.
10
11
12
13
14
for i = 2 . . . n do
if A[i] = B[i] = NIL then
A[i] ← A[i − 1]; B[i] ← B[i − 1] + 1 ;
*/
// SUS ki = SUS ki−1 S[i]
else if B[i − 1] = i − 1 and B[i − 1] − A[i − 1] + 2 < B[i] − A[i] + 1 then
A[i] ← A[i − 1]; B[i] ← B[i − 1] + 1 ;
/* Otherwise, do nothing. SUS ki = SLS ki .
15
// SUS ki = SUS ki−1 S[i]
*/
| 8cs.DS
|
DEEP LEARNING FOR ACCELERATED ULTRASOUND IMAGING
Yeo Hun Yoon and Jong Chul Ye ∗
arXiv:1710.10006v1 [cs.CV] 27 Oct 2017
Bio Imaging & Signal Processing Lab., Dept. of Bio and Brain Engineering, KAIST
291 Daehak-ro, Yuseong-gu, Daejeon, Republic of Korea
ABSTRACT
In portable, 3-D, or ultra-fast ultrasound (US) imaging systems, there is an increasing demand to reconstruct high quality images from limited number of data. However, the existing solutions require either hardware changes or computationally expansive algorithms. To overcome these limitations,
here we propose a novel deep learning approach that interpolates the missing RF data by utilizing the sparsity of the RF
data in the Fourier domain. Extensive experimental results
from sub-sampled RF data from a real US system confirmed
that the proposed method can effectively reduce the data rate
without sacrificing the image quality.
Index Terms— Deep learning, ultrasound imaging, lowrank Hankel matrix
1. INTRODUCTION
Due to the the excellent temporal resolution with reasonable
image quality and minimal invasiveness, ultrasound imaging
has been adopted as a golden-standard for many disease diagnosis in heart, liver, etc. Accordingly, there have been many
research efforts to extend the US imaging to new applications
such as portable imaging in emergency care, 3-D imaging,
ultra-fast imaging, etc. However, the common technical hurdle in these applications is to reconstruct high resolution images from significantly reduced radio-frequency (RF) data acquired with an US transducer.
While the compressed sensing approach has been investigated to address this issue [1], accurate modeling of wave
propagation is usually required. Recently, Wagner et al [2]
models a scan line profile as finite rate of innovation, and
proposed a specially-designed hardware architecture that enables high resolution scan line reconstruction using sparsity
driven recovery algorithm [2]. Another recent proposal is to
use a low rank Hankel matrix completion approach [3]. In
particular, thanks to the strong correlation between scan lines
and adjacent temporal frames, Jin et al [4] showed that a rank
deficient Hankel structured matrix can be obtained from the
reordered RF measurement data. Accordingly, the missing
∗ This work is supported by National Research Foundation of Korea
(NRF-2016R1A2B3008104).
data can be accurately reconstructed using a low rank Hankel matrix completion algorithm [3]. However, the algorithm
is computationally very expensive, which is not suitable for
routine diagnostic applications.
One of the most important contributions of this paper is
therefore to show that the low-rank interpolations problem of
RF data can be solved using deep convolutional neural network (CNN), which enables very fast run-time reconstruction.
In particular, inspired by recent finding that a CNN can be
interpreted as a deep convolutional framelets obtained from
a decomposition of Hankel matrix [5], we construct a CNN
that performs direct interpolation of RF data. Thus, the final
reconstruction can be done using the standard delay-and-sum
(DAS) beamformer without changing any hardware/software
structures. Compared to an image domain CNN that attempts
to learn acquisition geometry-dependent artifacts, one of the
most important advantages of the proposed RF domain CNN
is its universality. Specifically, although an image domain
deep learning requires many data set from various acquisition
geometry and body parts [6], our CNN can be trained using
a RF data measured by, for example, the linear array transducer for a particular organ, which can be then used for other
type of transducers and/or different body parts. Therefore, the
proposed systems is very practical in real applications.
Extensive experimental results using RF data from a real
US system confirmed that the proposed method has significant potential for accelerated US systems.
2. THEORY
2.1. Sub-sampled dynamic aperture US system
As a proof-of-concept, we consider a sub-sampled dynamic
aperture US system [4]. Specifically, a B-mode ultrasound
scans the body and displays the 2-D image as shown in Fig.1.
In Fig.1, SC, Rx and DAS denote scan line for each ultrasound beam, receivers of the transducer, and delay-and-sum
(DAS) beamformer, respectively. After a focused ultrasound
beam is illuminated along the scan line as shown in Fig.1(a),
ultrasound beam is reflected at some tissue boundaries and
the reflected US data are recorded by receivers as RF data
(see Fig.1(b)). Thus, for each scan line (SC), a Depth-Rx coordinate RF data is obtained, and this is repeated for each scan
(a) Beam emission
(b) RFs record
Fig. 2. Extraction of Rx-SC data for each depth direction.
when the Fourier component of the Rx-SC image is given by
fb, then we can find an annihilating function b
h in the spectral
domain such that their multiplication becomes zero:
fb b
h=0
(1)
This is equivalent to the convolutional relationship in the RxSC domain:
f ~ h = 0,
(c) conventional data acquisitions
Fig. 1. B-mode ultrasound imaging flow.
line to obtain a 3-D cube of RF data in Depth-Rx-SC coordinates. Then, a DAS beamformer uses the 3-D stacked data to
generate one temporal frame as shown in Fig.1(c). Because
of huge dat set, a hardware beamformer is usually incorporated to a transducer and the generated image is transferred
to an image workstation. On the other hand, recent softwarebased US systems directly transfer RF data to the image workstation, where a software-based beamformer reconstructs images. However, for many portable or high spatio/temporal
applications, online reconstruction or huge data transfer may
not be a viable option.
In this paper, we thus consider a randomly sub-sampled
dynamic aperture US system that acquires only partial data
as shown in Fig.3. Because hardware is not changed, this
method can be easily used in any conventional ultrasound system. Also, it can relieve the frame rate limitation caused by
pulse repetition frequency.
2.2. Redundancy of RF data and low-rank Hankel matrix
Recall that B-mode ultrasound measurements are acquired by
point-by-point scanning for each scan line. Since the scan line
only changes incrementally, the acquired Rx data along detectors do not change rapidly for each scan line. This implies
that there exists some level of skewed redundancy in Rx-SC
coordinate data, which can be also observed in Fig. 2.
This type of redundancy can be easily seen as a sparsity
in the Fourier domain as demonstrated in Fig. 2. Therefore,
(2)
which can be represented in a matrix form :
H(f )h = 0,
(3)
where H(f ) is a Hankel structured matrix constructed from
f . This implies that the Hankel matrix constructed from reordered raw data is rank-deficient. Furthermore, its rank is
determined by sparsity level as theoretically proven in [3]. In
fact, Jin et al [4] utilized this idea to interpolate missing RF
data using low-rank Hankel matrix completion. However, the
main limitation of [4] is its computational complexity, which
is prohibitive in real applications.
2.3. Deep Convolutional Framelets
Recently, it was shown that CNN is closely related to Hankel
matrix decomposition [5]. Specifically, for a given Hankel
matrix H(f ) ∈ Rn×d , let Φ and Φ̃ ∈ Rn×m (resp. Ψ and
Ψ̃ ∈ Rd×q ) are frame and its dual frame, respectively, satisfying the frame condition: Φ̃Φ> = I, ΨΨ̃> = I such that it
satisfies the perfect reconstruction condition:
H(f ) = Φ̃Φ> H(f )ΨΨ̃> .
(4)
One of the most important discoveries in [5] is to show that
(4) can be equivalently represented in the signal domain using
so-called deep convolutional framelets expansion, where the
decoder-part convolution is given by
f =
Φ̃C ~ τ (Ψ̃)
(5)
where C is the framelet coefficient matrix obtained from the
encoder part of convolution:
C = Φ> f ~ Ψ .
(6)
Fig. 3. Network architecture for B-mode ultrasound reconstruction
Here, the local filters for the encoder and decoder part convolution are given by
ψ̃1
1 .
d×q
, τ (Ψ̃) := .. ∈ Rdq .
Ψ := ψ 1 · · · ψ q ∈ R
d
ψ̃q
where v represents a flipped version of a vector v. Moreover,
for a given matrix input Z ∈ Rn×p , we have the following
convolutional framelet expansion [5]:
Z
=
(ΦC) ~ τ (Ψ̃)
where the framelet coefficients C is given by
C = Φ> Z ~ Ψ
Here, the encoder and decoder filters are defined by
1
1
ψ1 · · · ψq
.
dp×q
..
..
Ψ :=
.
..
. ∈R
p
ψ1
τ (Ψ̃)
:=
1
ψ̃1
1 .
.
d .
ψ̃q1
···
···
..
.
···
(7)
(8)
(9)
p
ψq
ψ̃1p
.. ∈ Rdq×p
.
(10)
ψ̃qp
The simple convolutional framelet expansion using (5), (7),
(6) and (8) is so powerful that a CNN with the encoderdecoder architecture emerges from them by inserting the pair
(8) and (7) between the pair (6) and (5). Moreover, a deep
CNN training can be interpreted to learn the basis matrix Ψ
for a given basis Φ such that maximal energy compaction can
be achieved. For more detail, see [5].
In brief, due to the redundancy in Rx-SC domain data,
the associated Hankel matrix is low-ranked, which makes the
convolutional framelet coefficients sparse. This implies that
a directly interpolation of Rx-SC domain RF data is feasible
using deep CNN.
3. METHOD
US data in the RF domain were acquired by E-CUBE 12R US
system (Alpinion Co., Korea). Real RF data were acquired by
a linear array transducer (L3-12H) with the center frequency
of 8.48MHz and a convex array transducer (SC1-4H) with the
center frequency of 3.2MHz. The sampling frequencies were
40MHz. We first acquired RF data from nine subjects using a linear array transducer. The size of each Rx-SC planes
is 64×384. 15000 Rx-SC planes of the seven people data
sets are randomly selected for generating training data set,
and 3000 Rx-SC planes of the another person data sets are
randomly selected for generating validation data set. The remaining one person data sets are used as test set. In addition,
we acquired RF data of liver from one subject using a convex
array transducer. This data set was used to verify the universality of the algorithm. The proposed CNN is composed
of convolution layers, batch normalization, ReLU and bypass
connection with concatenation as shown in Fig. 3. Specifically, the network consists of 28 convolution layers composed
of a batch normalization and a ReLU except for the last convolution layer. The first 27 convolution layers use 3×3 kernel,
and the last convolution layer use 1×1 kernel. Four bypass
connections with concatenation exist. As input data, we used
randomly sub-sampled data at the downsampling ratio of 4.
Because our proposed method assumes a dynamic aperture
US system, the RF data are sub-sampled along the receivers.
And the full data of Rx-SC plane was used as the label data.
For network training, the number of epoch was 400. The regularization parameter was 10−4 . The network was trained by
the stochastic gradient descent. The learning rate started from
10−7 and gradually decreased down to 10−9 . The network
was implemented using MatConvNet.
4. EXPERIMENTAL RESULTS
Fig. 4 illustrates B-mode ultrasound images acquired from
linear array transducer. The proposed CNN-based interpola-
Fig. 4. Linear transducer US images from (first row) full RF
data, (second row) x4 downsampled RF data, and (last row)
proposed CNN-based interpolation from x4 sub-sampled RF
data.
Fig. 5. Convex transducer US images from (first row) full RF
data, (second row) x4 downsampled RF data, and (last row)
proposed CNN-based interpolation from x4 sub-sampled RF
data.
important platform for accelerated US imaging.
tion successfully reconstructed the missing RF data from only
25% of RF data such that the beamformer removes the blurring artifacts and generates near artifact-free images. The average of PSNR for the reconstruction B-mode image is around
30.92, which is about 10dB improvement compared to the Bmode images from sub-sampled RF data.
Next, we use the network trained using the RF data from
linear array transducer to interpolate the missing RF data from
a convex array probe. The reconstruction results from a DAS
beamformer are shown in in Fig.5. Because the Rx-SC data of
linear array transducer and convex array transducer are similar, the DAS beamformer provided very accurate reconstruction results without any line artifacts or blurring from CNNbased RF interpolation. It is also remarkable that the accurate reconstruction was obtained for the liver region, which
was never seen by the network trained using the linear array
transducer data. The results confirmed the universality of the
algorithm.
5. CONCLUSIONS
In this paper, we proposed a novel deep learning approach
for accelerated B-mode ultrasound imaging. To exploit the
redundancy in the RF domain, the proposed CNN was applied
to Rx-SC domain. Compared with the methods that need to
change the hardware, the proposed method does not need any
hardware change and can be applied to any B-node ultrasound
system or any transducer. Therefore, this method can be an
6. REFERENCES
[1] Hervé Liebgott, Rémy Prost, and Denis Friboulet, “Prebeamformed RF signal reconstruction in medical ultrasound using compressive sensing,” Ultrasonics, 2012.
[2] Noam Wagner, Yonina C Eldar, and Zvi Friedman,
“Compressed beamforming in ultrasound imaging,”
IEEE Trans. Signal Process., vol. 60, no. 9, pp. 4643–
4657, 2012.
[3] Jong Chul Ye, Jong Min Kim, Kyong Hwan Jin, and
Kiryung Lee, “Compressive sampling using annihilating
filter-based low-rank interpolation,” IEEE Transactions
on Information Theory, vol. 63, no. 2, pp. 777–801, Feb.
2017.
[4] Kyong Hwan Jin, Yo Seob Han, and Jong Chul
Ye, “Compressive dynamic aperture b-mode ultrasound
imaging using annihilating filter-based low-rank interpolation,” in 2016 IEEE 13th International Symposium on
Biomedical Imaging (ISBI). IEEE, 2016, pp. 1009–1012.
[5] Jong Chul Ye and Yo Seob Han, “Deep convolutional
framelets: A general deep learning for inverse problems,”
arXiv preprint arXiv:1707.00372, 2017.
[6] Eunhee Kang, Jaejun Yoo, and Jong Chul Ye, “Wavelet
residual network for low-dose CT via deep convolutional
framelets,” arXiv preprint arXiv:1707.09938, 2017.
| 1cs.CV
|
arXiv:1606.01342v2 [cs.DS] 27 Aug 2016
Recoverable robust spanning tree problem under interval
uncertainty representations
Mikita Hradovich† , Adam Kasperski‡ , Pawel Zieliński†
†
Faculty of Fundamental Problems of Technology,
Wroclaw University of Technology, Wroclaw, Poland
‡
Faculty of Computer Science and Management,
Wroclaw University of Technology, Wroclaw, Poland
{mikita.hradovich,adam.kasperski,pawel.zielinski}@pwr.edu.pl
Abstract
This paper deals with the recoverable robust spanning tree problem under interval
uncertainty representations. A strongly polynomial time, combinatorial algorithm for the
recoverable spanning tree problem is first constructed. This problem generalizes the incremental spanning tree problem, previously discussed in literature. The algorithm built
is then applied to solve the recoverable robust spanning tree problem, under the traditional interval uncertainty representation, in polynomial time. Moreover, the algorithm
allows to obtain several approximation results for the recoverable robust spanning tree
problem under the Bertsimas and Sim interval uncertainty representation and the interval
uncertainty representation with a budget constraint.
Keywords: robust optimization; interval data; recovery; spanning tree
1
Introduction
Let G = (V, E), |V | = n, |E| = m, be an undirected graph and let Φ be the set of all spanning
trees of G. In the minimum spanning tree problem, a cost is specified for each edge, and we
seek a spanning tree in G of the minimum total cost. This problem is well known and can be
solved efficiently by using several polynomial time algorithms (see, e.g., [1, 18]).
In this paper, we first study the recoverable spanning tree problem (Rec ST for short).
Namely, for each edge e ∈ E, we are given a first stage cost Ce and a second stage cost ce
(recovery stage cost). Given a spanning tree X ∈ Φ, let ΦkX be the set of all spanning trees
Y ∈ Φ such that |Y \ X| ≤ k (the recovery set), where k is a fixed integer in [0, n − 1], called
the recovery parameter. Note that ΦkX can be seen as a neighborhood of X containing all
spanning trees which can be obtained from X by exchanging up to k edges. The Rec ST
problem can be stated formally as follows:
!
X
X
Rec ST : min
Ce + min
(1)
ce .
X∈Φ
e∈X
Y ∈ΦkX
e∈Y
We thus seek a first stage spanning tree X ∈ Φ and a second stage spanning tree Y ∈ ΦkX , so
that the total cost of X and Y for Ce and ce , respectively, is minimum. Notice that Rec ST
1
generalizes the following incremental spanning tree problem, investigated in [8]:
X
Inc ST : min
ce ,
Y ∈Φkb
X
(2)
e∈Y
b ∈ Φ is a given spanning tree. So, we wish to find an improved spanning tree Y
where X
b determined by Φk . Several interesting
with the minimum cost, within a neighborhood of X
b
X
practical applications of the incremental network optimization were presented in [8]. It is
worth pointing out that Inc ST can be seen as the Rec ST problem with a fixed first
b whereas in Rec ST both the first and the second stage trees are
stage spanning tree X,
unknown. It has been shown in [8] that Inc ST can be solved in strongly polynomial time
by applying the Lagrangian relaxation technique. On the other hand, no strongly polynomial
time combinatorial algorithm for Rec ST has been known to date. Thus proposing such an
algorithm for this problem is one of the main results of this paper.
The Rec ST problem, beside being an interesting problem per se, has an important connection with a more general problem. Namely, it is an inner problem in the recoverable robust
model with uncertain recovery costs, discussed in [4, 5, 6, 7, 15, 17]. Indeed, the recoverable
spanning tree problem can be generalized by considering its robust version. Suppose that
the second stage costs ce , e ∈ E, are uncertain and let U contain all possible realizations of
the second stage costs, called scenarios. We will denote by cSe the second stage cost of edge
e ∈ E under scenario S ∈ U , where S = (cSe )e∈E is a cost vector. In the recoverable robust
spanning tree problem (Rob Rec ST
P for short), we choose an initial spanning tree X in the
first stage, with the cost equal to e∈X Ce . Then, after scenario S ∈ U reveals, X can be
k
modified by exchanging at most k edges, obtaining
Pa newS spanning tree Y ∈ ΦX . The second
stage cost of Y under scenario S ∈ U is equal to e∈Y ce . Our goal is to find a pair of trees
X andPY such that
P |X \ Y | ≤ k, which minimizes the sum of the first and the second stage
costs e∈X Ce + e∈Y cSe in the worst case. The Rob Rec ST problem is defined formally
as follows:
!
X
X
(3)
cSe .
Ce + max min
Rob Rec ST : min
X∈Φ
e∈X
S∈U Y ∈Φk
X
e∈Y
If Ce = 0 for each e ∈ E and k = 0, then Rob Rec ST is equivalent to the following
min-max spanning tree problem, examined in [2, 13, 12], in which we seek a spanning tree
that minimizes the largest cost over all scenarios:
X
Min-Max ST : min max
cSe .
(4)
X∈Φ S∈U
e∈X
If Ce = 0 for each e ∈ E and k = n − 1, then Rob Rec ST becomes the following
adversarial problem [17] in which an adversary wants to find a scenario which leads to the
greatest increase in the cost of the minimum spanning tree:
X
Adv ST : max min
cSe .
(5)
S∈U Y ∈Φ
e∈Y
We now briefly recall the known complexity results on Rob Rec ST. It turns out that
its computational complexity highly relies on the way of defining the scenario set U . There
are two popular methods of representing U , namely the discrete and interval uncertainty
representations. For the discrete uncertainty representation (see, e.g., [13]), scenario set,
2
denoted by U D , contains K explicitly listed scenarios, i.e. U D = {S1 , S2 , . . . , SK }. In this
case, the Rob Rec ST problem is known to be NP-hard for K = 2 and any constant k [11].
Furthermore, it becomes strongly NP-hard and not at all approximable when both K and k
are a part of the input [11]. It is worthwhile to mention that Min-Max ST is NP hard even
when K = 2 and becomes strongly NP-hard and not approximable within O(log1−ǫ n) for
any ǫ > 0 unless NP ⊆ DTIME(npoly log n ), when K is a part of input [13, 12]. It admits an
FPTAS, when K is a constant [2] and is approximable within O(log2 n), when K is a part of
the input [12]. The Adv ST problem, under scenario set U D , is polynomially solvable, since
it boils down to solving K traditional minimum spanning tree problems.
For the interval uncertainty representation, which is considered in this paper, one assumes
that the second stage cost of each edge e ∈ E is known to belong to the closed interval
[ce , ce + de ], where ce is a nominal cost of e ∈ E and de ≥ 0 is the maximum deviation of the
cost of e from its nominal value. In the traditional case U , denoted by U I , is the Cartesian
product of all these intervals [13], i.e.
U I = {S = (cSe )e∈E : cSe ∈ [ce , ce + de ], e ∈ E}.
(6)
In [4] a polynomial algorithm for the recoverable robust matroid basis problem under
scenario set U I was constructed, provided that the recovery parameter k is constant. In consequence, Rob Rec ST under U I is also polynomially solvable for constant k. Unfortunately,
the algorithm proposed in [4] is exponential in k. Interestingly, the corresponding recoverable
robust version of the shortest path problem (Φ is replaced with the set of all s − t paths in G)
has been proven to be strongly NP-hard and not at all approximable even if k = 2 [5]. It has
been recently shown in [10] that Rob Rec ST under U I is polynomially solvable when k is a
part of the input. In order to prove this result, a technique called the iterative relaxation of
a linear programming formulation, whose framework was described in [14], has been applied
This technique, however, does not imply directly a strongly polynomial algorithm for Rob
Rec ST, since it requires the solution of a linear program.
In [3] a popular and commonly used modification of the scenario set U I has been proposed.
The new scenario set, denoted as U1I (Γ), is a subset of U I such that under each scenario in
U1I (Γ), the costs of at most Γ edges are greater than their nominal values ce , where Γ is
assumed to be a fixed integer in [0, m]. Scenario set U1I (Γ) is formally defined as follows:
X
δe ≤ Γ}.
(7)
U1I (Γ) = {S = (cSe )e∈E : cSe ∈ [ce , ce + δe de ], δe ∈ {0, 1}, e ∈ E,
e∈E
The parameter Γ allows us to model the degree of uncertainty. When Γ = 0, then we get
Rec ST (Rob Rec ST with one scenario S = (ce )e∈E ). On the other hand, when Γ = m,
then we get Rob Rec ST under the traditional interval uncertainty U I . It turns out that
the Adv ST problem under U1I (Γ) is strongly NP-hard (it is equivalent to the problem of
finding Γ most vital edges) [17, 16, 9]. Consequently, the more general Rob Rec ST problem
is also strongly NP-hard. Interestingly, the corresponding Min-Max ST problem with U1I (Γ)
is polynomially solvable [3].
Yet another interesting way of defining scenario set, which allows us to control the amount
of uncertainty, is called the scenario set with a budget constraint (see, e.g,. [17]). This scenario
set, denoted as U2I (Γ), is defined as follows:
X
U2I (Γ) = {S = (cSe )e∈E : cSe = ce + δe , δe ∈ [0, de ], e ∈ E,
δe ≤ Γ},
(8)
e∈E
3
where Γ ≥ 0 is a a fixed parameter that can be seen as a budget of an adversary, and represents
the maximum total increase of the edge costs from their nominal values. Obviously, if Γ is
sufficiently large, then U2I (Γ) reduces to the traditional interval uncertainty representation U I .
The computational complexity of Rob Rec ST for scenario set U2I is still open. We only know
that its special cases, namely Min-Max ST and Adv ST, are polynomially solvable [17].
In this paper we will construct a combinatorial algorithm for Rec ST with strongly
polynomial running time. We will apply this algorithm for solving Rob Rec ST under
scenario set U I in strongly polynomial time. Moreover, we will show how the algorithm for
Rec ST can be used to obtain several approximation results for Rob Rec ST, under scenario
sets U1I (Γ) and U2I (Γ). This paper is organized as follows. Section 2 contains the main result
of this paper – a combinatorial algorithm for Rec ST with strongly polynomial running time.
Section 3 discusses Rob Rec ST under the interval uncertainty representations U I , U1I (Γ),
and U2I (Γ).
2
The recoverable spanning tree problem
In this section we construct a combinatorial algorithm for Rec ST with strongly polynomial
running time. Since |X| = n − 1 for each X ∈ Φ, Rec ST (see (1)) is equivalent to the
following mathematical programming problem:
X
X
ce
Ce +
min
e∈Y
e∈X
(9)
|X ∩ Y | ≥ L,
X, Y ∈ Φ,
s.t.
where L = n − 1 − k. Problem (9) can be expressed as the following MIP model:
X
X
Opt = min
Ce xe +
ce ye
e∈E
s.t.
X
(10)
e∈E
xe = n − 1,
(11)
e∈E
X
xe ≤ |U | − 1,
∀U ⊂ V,
(12)
e∈E(U )
X
ye = n − 1,
(13)
e∈E
X
ye ≤ |U | − 1,
∀U ⊂ V,
(14)
xe − ze ≥ 0,
∀e ∈ E,
(15)
ye − ze ≥ 0,
X
ze ≥ L,
∀e ∈ E,
(16)
e∈E(U )
(17)
e∈E
xe , ye , ze ≥ 0, integer
∀e ∈ E,
(18)
where E(U ) stands for the set of edges that have both endpoints in U ⊆ V . We first apply the
Lagrangian relaxation (see, e.g., [1]) to (10)-(18) by relaxing the cardinality constraint (17)
4
with a nonnegative multiplier θ. We also relax the integrality constraints (18). We thus get
the following linear program (with the corresponding dual variables which will be used later):
X
X
X
φ(θ) = min
Ce xe +
ce ye − θ
ze + θL
(19)
e∈E
s.t.
X
e∈E
−
e∈E
e∈E
xe = n − 1,
X
[µ],
xe ≥ −(|U | − 1),
∀U ⊂ V,
[wU ],
e∈E(U )
X
e∈E
−
ye = n − 1,
X
[ν],
ye ≥ −(|U | − 1),
∀U ⊂ V,
[vU ],
xe − ze ≥ 0,
∀e ∈ E,
[αe ],
ye − ze ≥ 0,
∀e ∈ E,
[βe ],
xe , ye , ze ≥ 0,
∀e ∈ E.
e∈E(U )
For any θ ≥ 0, the Lagrangian function φ(θ) is a lower bound on Opt. It is well-known
that φ(θ) is concave and piecewise linear. By the optimality test (see, e.g., [1]), we obtain the
following theorem:
Theorem 1. Let (xe , ye , ze )e∈E be an optimal solution to (19) for
Psome θ ≥ 0, feasible
to (11)-(18) and satisfying the complementary slackness condition θ( e∈E ze − L) = 0. Then
(xe , ye , ze )e∈E is optimal to (10)-(18).
Let (X, Y ), X, Y ∈ Φ, be a pair of spanning trees of G (a pair for short). This pair
corresponds to a feasible 0 − 1 solution to (19), defined as follows: xe = 1 for e ∈ X,
ye = 1 for e ∈ Y , and ze = 1 for e ∈ X ∩ Y ; the values of the remaining variables are
set to 0. From now on, by a pair (X, Y ) we also mean a feasible solution to (19) defined
as above. Given a pair (X, Y ) with the corresponding solution (xe , ye , ze )e∈E , let us define
the partition (EX , EY , EZ , EW ) of the set of the edges E in the following way: EX = {e ∈
E : xe = 1, ye = 0}, EY = {e ∈ E : ye = 1, xe = 0}, EZ = {e ∈ E : xe = 1, ye = 1}
and EW = {e ∈ E : xe = 0, ye = 0}. Thus equalities: X = EX ∪ EZ , Y = EY ∪ EZ and
EZ = X ∩ Y hold. Our goal is to establish some sufficient optimality conditions for a given
pair (X, Y ) in the problem (19). The dual to (19) has the following form:
X
X
φD (θ) = max −
(|U | − 1)wU + (n − 1)µ −
(|U | − 1)vU + (n − 1)ν + θL
(20)
U ⊂V
s.t. −
X
U ⊂V
wU + µ ≤ Ce − αe ,
∀e ∈ E,
vU + ν ≤ ce − βe ,
∀e ∈ E,
{U ⊂V : e∈E(U )}
−
X
{U ⊂V : e∈E(U )}
αe + βe ≥ θ,
∀e ∈ E,
wU , vU ≥ 0,
U ⊂ V,
αe , βe ≥ 0,
∀e ∈ E.
5
Lemma 1. The dual problem (20) can be rewritten as follows:
φD (θ) =
max
{αe ≥0,βe ≥0 : αe +βe ≥θ, e∈E}
min
X∈Φ
X
(Ce − αe ) + min
Y ∈Φ
e∈X
X
!
(ce − βe )
e∈Y
+ θL.
(21)
Proof. Fix some αe and βe such that αe + βe ≥ θ for each e ∈ E in (20). For
P these constant
values ofPαe and βe , e ∈ E, using the dual to (20), we arrive to minX∈Φ e∈X (Ce − αe ) +
minY ∈Φ e∈Y (ce − βe ) + θL and the lemma follows.
Lemma 1 allows us to establish the following result:
Theorem 2 (Sufficient pair optimality conditions). A pair (X, Y ) is optimal to (19)
for a fixed θ ≥ 0 if there exist αe ≥ 0, βe ≥ 0 such that αe + βe = θ for each e ∈ E and
(i) X is a minimum spanning tree for the costs Ce − αe , Y is a minimum spanning tree for
the costs ce − βe ,
(ii) αe = 0 for each e ∈ EX , βe = 0 for each e ∈ EY .
Proof. By the primal-dual relation, the inequality φD (θ) ≤ φ(θ) holds. Using (21), we obtain
X
X
X
X
X
φD (θ) ≥
(Ce − αe ) +
(ce − βe ) + θL =
Ce +
ce +
(Ce + ce − θ) + θL
e∈X
=
X
Ce +
e∈EX
=
X
e∈X
Ce +
X
e∈EZ
X
e∈Y
Ce +
X
e∈EY
ce +
X
e∈EX
e∈EY
e∈EZ
ce − θ|EZ | + θL
e∈EZ
ce − θ|EZ | + θL = φ(θ).
e∈Y
The Weak Duality Theorem implies the optimality of (X, Y ) in (19) for a fixed θ ≥ 0.
Lemma 2. A pair (X, Y ), which satisfies the sufficient pair optimality conditions for θ = 0,
can be computed in polynomial time.
Proof. Let X be a minimum spanning tree for the costs Ce and Y be a minimum spanning
tree for the costs ce , e ∈ E. Since θ = 0, we set αe = 0, βe = 0 for each e ∈ E. It is clear that
(X, Y ) satisfies the sufficient pair optimality conditions.
Assume that (X, Y ) satisfies the sufficient pair optimality conditions for some θ ≥ 0. If,
for this pair, |EZ | ≥ L and θ(|EZ | − L) = 0, then we are done, because by Theorem 1, the
pair (X, Y ) is optimal to (10)-(18). Suppose that |EZ | < L ((X, Y ) is not feasible to (10)-(18)).
We will now show a polynomial time procedure for finding a new pair (X ′ , Y ′ ), which satisfies
the sufficient pair optimality conditions and |EZ ′ | = |EZ | + 1. This implies a polynomial
time algorithm for the problem (10)-(18), since it is enough to start with a pair satisfying
the sufficient pair optimality conditions for θ = 0 (see Lemma 2) and repeat the procedure at
most L times, i.e. until |EZ ′ | = L.
Given a spanning tree T in G = (V, E) and edge e = {k, l} 6∈ T , let us denote by PT (e)
the unique path in T connecting nodes k and l. It is well known that for any f ∈ PT (e),
T ′ = T ∪ {e} \ {f } is also a spanning tree in G. We will say that T ′ is the result of a move
on T .
6
Consider a pair (X, Y ) that satisfies the sufficient pair optimality conditions for some fixed
θ ≥ 0. Set Ce∗ = Ce − αe and c∗e = ce − βe for every e ∈ E, where αe and βe , e ∈ E, are
the numbers which satisfy the conditions in Theorem 2. Thus, by Theorem 2(i) and the path
optimality conditions (see, e.g., [1]), we get the following conditions which must be satisfied
by (X, Y ):
for every e ∈
/X
Ce∗ ≥ Cf∗
for every f ∈ PX (e),
(22a)
for every e ∈
/Y
c∗e
for every f ∈ PY (e).
(22b)
≥
c∗f
We now build a so-called admissible graph GA = (V A , E A ) in two steps. We first associate
with each edge e ∈ E a node ve and include it to V A , |V A | = |E|. We then add arc (ve , vf )
to E A if e ∈
/ X, f ∈ PX (e) and Ce∗ = Cf∗ . This arc is called an X-arc. We also add arc
(vf , ve ) to E A if e ∈
/ Y , f ∈ PY (e) and c∗e = c∗f . This arc is called an Y -arc. We say that
ve ∈ V A is admissible if e ∈ EY , or ve is reachable from a node vg ∈ V A , such that g ∈ EY ,
by a directed path in GA . In the second step we remove from GA all the nodes which are not
admissible, together with their incident arcs. An example of an admissible graph is shown in
Figure 1. Each node of this admissible graph is reachable from some node vg , g ∈ EY . Note
that the arcs (ve7 , ve6 ) and (ve7 , ve10 ) are not present in GA , because ve7 is not reachable from
any node vg , g ∈ EY . These arcs have been removed from GA in the second step.
Observe that each X-arc (ve , vf ) ∈ E A represents a move on X, namely X ′ = X ∪{e}\{f }
is a spanning tree in G. Similarly, each Y -arc (ve , vf ) ∈ E A represents a move on Y , namely
Y ′ = Y ∪ {f } \ {e} is a spanning tree in G. Notice that the cost, with respect to Ce∗ , of X ′
is the same as X and the cost, with respect to c∗e , of Y ′ is the same as Y . So, the moves
indicated by X-arcs and Y -arcs preserve the optimality of X and Y , respectively. Observe
that e ∈
/ X or e ∈ Y , which implies e ∈
/ EX . Also f ∈ X or f ∈
/ Y , which implies f ∈
/ EY .
Hence, no arc in E A can start in a node corresponding to an edge in EX and no arc in E A
can end in a node corresponding to an edge in EY . Observe also that (ve , vf ) ∈ E A can be
both X-arc and Y -arc only if e ∈ EY and f ∈ EX . Such a case is shown in Figure 1 (see the
arc (ve1 , ve2 )). Since each arc (ve , vf ) ∈ E A represents a move on X or Y , e and f cannot
both belong to EW or EZ .
(a)
(b)
e1 (2; 2)
e6 (2; 3)
(Ce∗ ; c∗e )
ve1 EY
X; Y
X
X
v e2
Y
e10
(2; 2)
e7
e5
e3
(2; 2)
(2; 3)
(2; 3)
v e3 E Z
EX
Y
e2
X
(2; 2)
X
v e8
e4 (1; 5)
e8
v e6
e9
Y
EW
(2; 3)
EX
(5; 0)
v e5 E Y
X
X
ve10
EZ
Figure 1: (a) A pair (X, Y ) such that X = {e2 , e3 , e4 , e6 , e10 } and Y = {e1 , e3 , e5 , e9 , e10 }.
(b) The admissible graph GA for (X, Y ).
We will consider two cases: EX ∩{e ∈ E : ve ∈ V A } =
6 ∅ and EX ∩{e ∈ E : ve ∈ V A } = ∅.
The first case means that there is a directed path from ve , e ∈ EY , to a node vf , f ∈ EX ,
7
in the admissible graph GA and in the second case no such a path exists. We will show that
in the first case it is possible to find a new pair (X ′ , Y ′ ) which satisfies the sufficient pair
optimality conditions and |EZ ′ | = |EZ | + 1. The idea will be to perform a sequence of moves
on X and Y , indicated by the arcs on some suitably chosen path from ve , e ∈ EY , to vf ,
f ∈ EX in the admissible graph GA . Let us formally handle this case.
Lemma 3. If EX ∩ {e ∈ E : ve ∈ V A } =
6 ∅, then there exists a pair (X ′ , Y ′ ) with |EZ ′ | =
|EZ | + 1, which satisfies the sufficient pair optimality conditions for θ.
Proof. We begin by introducing the notion of a cycle graph G(T ) = (V T , AT ), corresponding
to a given spanning tree T of graph G = (V, A). We build G(T ) as follows: we associate with
each edge e ∈ E a node ve and include it to V T , |E| = |V T |; then we add arc (ve , vf ) to AT
if e 6∈ T and f ∈ PT (e). An example is shown in Figure 2.
(a)
(b)
f1
f2
f4
f3
e1
f5
f6
ve2
ve1
e2
vf1
vf6
vf7
vf5
vf2
vf3
vf4
e3
f7
ve3
Figure 2: (a) A graph G with a spanning tree T (the solid lines). (b) The cycle graph G(T ).
Claim 1. Given a spanning tree T of G, let F = {(ve1 , vf1 ), (ve2 , vf2 ), . . . , (veℓ , vfℓ )} be a
subset of arcs of G(T ), where all vei and vfi (resp. ei and fi ), i ∈ [ℓ], are distinct. If
T ′ = T ∪ {e1 , . . . , eℓ } \ {f1 , . . . , fℓ } is not a spanning tree, then G(T ) contains a subgraph
depicted in Figure 3, where {j1 , . . . , jκ } ⊆ [ℓ].
vej1
vfjκ
vfj1
vej2
vfj3
vejκ
vfj2
vej3
Figure 3: A subgraph of G(T ) from Claim 1.
Let us illustrate Claim 1 by using the sample graph in Figure 2. Suppose that F =
{(ve1 , vf5 ), (ve2 , vf2 ), (ve3 , vf3 )}. Then T ′ = T ∪ {e1 , e2 , e3 , } \ {f5 , f2 , f3 } is not a spanning tree
and G(T ) contains the subgraph composed of the following arcs (see Figure 2):
(ve1 , vf2 ), (ve2 , vf2 ), (ve2 , vf3 ), (ve3 , vf3 ), (ve3 , vf5 ), (ve1 , vf5 ).
Proof of Claim 1. We form T ′ by performing a sequence of moves consisting in adding edges ei
and removing edges fi ∈ PT (ei ), i ∈ [ℓ]. Suppose that, at some step, a cycle appears, which
is formed by some edges from {e1 , . . . , eℓ } and the remaining edges of T (not removed from
T ). Such a cycle must appear, since otherwise T ′ would be a spanning tree. Let us relabel
the edges so that {e1 , . . . , es } are on this cycle, i.e. the first s moves consisting in adding
8
ei and removing fi create the cycle, i ∈ [s]. An example of such a situation for s = 4
is shown in Figure 4. The cycle is formed by the edges e1 , . . . , e4 and the paths Pv2 v3 ,
Pv4 v5 and Pv1 v6 in T . Consider the edge e1 = {v1 , v2 }. Because T is a spanning tree,
PT (e1 ) ⊆ Pv2 v3 ∪ PT (e2 ) ∪ PT (e3 ) ∪ Pv4 v5 ∪ PT (e4 ) ∪ Pv1 v6 . Observe that f1 ∈ PT (e1 ) cannot
belong to any of Pv2 v3 , Pv4 v5 and Pv1 v6 . If it would be contained in one of these paths, then
no cycle would be created. Hence, f1 must belong to PT (e2 ) ∪ PT (e3 ) ∪ PT (e4 ). The above
argument is general and, by using it, we can show that for each i ∈ [s], fi ∈ PT (ej ) for some
j ∈ [s] \ {i}.
(a)
(b)
f2
PT (e2 )
Pv 2 v 3
v f1
f1 2 PT (e1 )
f1 2 PT (e2 )
v e2
v f2
f2 2 PT (e2 )
f2 2 PT (e3 )
v e3
v f3
f3 2 PT (e3 )
f3 2 PT (e2 )
v3
v2
e1
e2
v1
e3
Pv 1 v 6
v6 e4
v e1
v5
PT (e3 ) f3
v4
Pv 4 v 5
PT (e4 )
f4
Figure 4: Illustration for the proof of Claim 1. (a) The bold lines represent paths in T
(not necessarily disjoint); f1 ∈ PT (e2 ) ∪ PT (e3 ) ∪ PT (e4 ). (b) The subgraph G′ (T ) with the
corresponding cycle.
We are now ready to build a subgraph depicted in Figure 3. Consider a subgraph G′ (T )
of the cycle graph G(T ) built as follows. The nodes of G′ (T ) are ve1 , . . . , ves , vf1 , . . . , vfs .
Observe that G′ (T ) has exactly 2s nodes, since all the edges e1 , . . . , es , f1 , . . . , fs are distinct
by the assumption of the claim. For each i ∈ [s] we add to G′ (T ) two arcs, namely (vei , vfi ),
fi ∈ PT (ei ) and (vej , vfi ), fi ∈ PT (ej ) for j ∈ [s] \ {i} (see Figure 4). The resulting graph
G′ (T ) is bipartite and has exactly 2s arcs. In consequence G′ (T ) (and thus G(T )) must
contain a cycle which is of the form depicted in Figure 3.
After this preliminary step, we can now return to the main proof. If EX ∩ {e ∈ E : ve ∈
V A} =
6 ∅, then, by the construction of the admissible graph, there exists a directed path in
A
G from a node ve , e ∈ EY , to a node vf , f ∈ EX . Let P be a shortest such a path from
ve to vf , i.e. a path consisting of the fewest number of arcs, called an augmenting path. We
need to consider the following cases:
1. The augmenting path P is of the form:
EY
ve
EX
→ vf
If (ve , vf ) is X-arc, then X ′ = X ∪ {e} \ {f } is an updated spanning tree of G such that
|X ′ ∩ Y | = |EZ | + 1. Furthermore X ′ is a minimum spanning tree for the costs Ce∗ and
the new pair (X ′ , Y ) satisfies the sufficient pair optimality conditions (EX ′ ⊆ EX , so
9
condition (ii) in Theorem 2 is not violated). If (ve , vf ) is Y -arc, then Y ′ = Y ∪ {f } \ {e}
is an updated spanning tree of G such that |X ∩ Y ′ | = |EZ | + 1. Also Y ′ is a minimum
spanning tree for the costs c∗e and the new pair (X, Y ′ ) satisfies the sufficient pair
optimality conditions. An example can be seen in Figure 1. There is a path ve1 → ve2
in the admissible graph. The arc (ve1 , ve2 ) is both X-arc and Y -arc. We can thus choose
one of the two possible moves X ′ = X ∪ {e1 } \ {e2 } or Y ′ = Y ∪ {e2 } \ {e1 }, which
results in (X ′ , Y ) or (Y ′ , X).
2. The augmenting path P is of the form:
EY
EZ
X
EW
Y
(a) ve1 → vf1 → ve2
EZ
X
EW
Y
→ vf 2 → ve 3
EZ
X
EW
Y
Y
→ vf 3 → · · · → ve ℓ
EZ
X
EX
Y
→ vfℓ → veℓ+1
EX
X
→ vf ℓ
(b)
Let X ′ = X ∪ {e1 , . . . , eℓ } \ {f1 , . . . , fℓ }. Let Y ′ = Y ∪ {e2 , . . . , eℓ+1 } \ {f1 , . . . fℓ } for case
(a), and Y ′ = Y ∪ {e2 , . . . , eℓ } \ {f1 , . . . fℓ−1 } for case (b). We now have to show that the
resulting pair (X ′ , Y ′ ) is a pair of spanning trees. Suppose that X ′ is not a spanning tree.
Observe that the X-arcs (ve1 , vf1 ), . . . , (veℓ , vfℓ ) belong to the cycle graph G(X). Thus,
by Claim 1, the cycle graph G(X) must contain a subgraph depicted in Figure 3, where
{j1 , . . . , jκ } ⊆ [ℓ]. An easy verification shows that all edges ei , fi , i ∈ {j1 , . . . , jκ } must
have the same costs with respect to Ce∗ . Indeed, if some costs are different, then there
exists an edge exchange which decreases the cost of X. This contradicts our assumption
that X is a minimum spanning tree with respect to Ce∗ . Finally, there must be an arc
(vei′ , vfi′′ ) in the subgraph such that i′ < i′′ . Since Ce∗i′ = Cf∗ ′′ , the arc (vei′ , vfi′′ ) is
i
present in the admissible graph GA . This leads to a contradiction with our assumption
that P is an augmenting path. Now suppose that Y ′ is not a spanning tree. We consider
only the case (a) since the proof of case (b) is just the same. For a convenience, let us
number the nodes vei on P from i = 0 to ℓ, so that Y ′ = {e1 , . . . , eℓ } \ {f1 , . . . , fℓ }. The
arcs (ve1 , vf1 ), . . . , (veℓ , vfℓ ), which correspond to the Y -arcs (vf1 , ve1 ), . . . , (vfℓ , veℓ ) of
P , belong to the cycle graph G(Y ). Hence, by Claim 1, G(Y ) must contain a subgraph
depicted in Figure 3, where {i1 , . . . , iκ } ⊆ [ℓ]. The rest of the proof is similar to the
proof for X. Namely, the edges ei and fi for i ∈ {i1 , . . . , iκ } must have the same costs
with respect to c∗e . Also, there must exist an arc (vei′ , vfi′′ ) in the subgraph such that
i′ > i′′ . In consequence, the arc (vfi′′ , vei′ ) belongs to the admissible graph, which
contradicts the assumption that P is an augmenting path.
An example of the case (a) is shown in Figure 5. Thus X ′ = X ∪ {e1 , e2 , e3 , e4 } \
{f1 , f2 , f3 , f4 } and Y ′ = Y ∪ {e2 , e3 , e4 , e5 } \ {f1 , f2 , f3 , f4 }. An example of the case
(b) is shown in Figure 6. In this example X ′ is the same as in the previous case and
Y ′ = Y ∪ {e2 , e3 , e4 } \ {f1 , f2 , f3 }.
It is easy to verify that |EZ ′ | = |X ′ ∩ Y ′ | = |EZ | + 1 holds (see also the examples in
Figures 5 and 6). The spanning trees X ′ and Y ′ are optimal for the costs Ce∗ and c∗e ,
respectively. Furthermore, EX ′ ⊆ EX and EY ′ ⊆ EY , so (X ′ , Y ′ ) satisfies the sufficient
pair optimality conditions (the condition (ii) in Theorem 2 is not violated).
10
(2,0)
(2,0)
(2,0)
(Ce∗ ; c∗e )
X
(0,2)
e1
f2 (1,1)
f1 (1,1)
Y
e2 (1,1)
(1,0)
e3 (1,1)
v e1 X
EZ
v f1
f4 (1,1)
(0,1)
e5
e4 (1,1)
(2,0)
(0,2)
EY
f3 (1,1)
EZ
EW
v e2 X
Y
(0,2)
v f2
(0,2)
EW
EZ
v e3 X
Y
v f3 Y
EW
v e4 X
EZ
v f4 Y
EX
v e5
Figure 5: A pair (X, Y ) and the corresponding admissible graph for the case 2a.
(2,0)
(2,0)
(2,0)
(Ce∗ ; c∗e )
X
(0,2)
f1 (1,1)
Y
f3 (1,1)
f2 (1,1)
f4 (1,2)
e2 (1,1)
(1,0)
e3 (1,1)
e4 (1,1)
e1
(0,2)
EY
ve1 X
EZ
EW
Y
v f1
v e2 X
(0,2)
EZ
v f2
Y
(0,2)
EW
EZ
v e3 X
EW
v e4 X
v f3 Y
EX
v f4
Figure 6: A pair (X, Y ) and the corresponding admissible graph for the case 2b.
3. The augmenting path P is of the form
EY
EW
Y
(a) ve1 → vf1
EZ
X
EW
Y
→ ve2 → vf2
EZ
X
EW
Y
→ ve 3 → vf 3
EZ
X
X
EW
Y
EX
X
→ . . . → veℓ → vfℓ → veℓ+1
EX
Y
→ vf ℓ
(b)
Let X ′ = X ∪ {f1 , . . . , fℓ } \ {e2 , . . . , eℓ+1 } for the case (a) and X ′ = X ∪ {f1 , . . . , fℓ−1 } \
{e2 , . . . eℓ } for the case (b). Let Y ′ = Y ∪ {f1 , . . . , fℓ } \ {e1 , . . . eℓ }. The proof that
X ′ and Y ′ are spanning trees follows by the same arguments as for the symmetric
case described in point 2. An example of the case (a) is shown in Figure 7. Thus
X ′ = X ∪ {f1 , f2 , f3 , f4 } \ {e2 , e3 , e4 , e5 } and Y ′ = Y ∪ {f1 , f2 , f3 , f4 } \ {e1 , e2 , e3 , e4 }.
An example for the case (b) is shown in Figure 8. The spanning tree Y ′ is the same
as in the previous case and X ′ = X ∪ {f1 , f2 , f3 } \ {e2 , e3 , e4 }. The equality |EZ ′ | =
|X ′ ∩ Y ′ | = |EZ | + 1 holds. Also, the trees X ′ and Y ′ are optimal for the costs Ce∗ and
c∗e , respectively, EX ′ ⊆ EX , EY ′ ⊆ EY , so (X ′ , Y ′ ) satisfies the sufficient pair optimality
conditions.
11
(0,2)
(0,2)
(0,2)
(Ce∗ ; c∗e )
X
(0,0)
f1 (1,1)
Y
e3 (1,1)
e2 (1,1)
(2,1)
f3 (1,1)
f2 (1,1)
e4
f4 (1,1)
(1,1)
e1
(0,0)
(2,0)
(2,0)
EY
EW
v e1 Y
v f1
(1,0)
e5
EZ
X
v e2
Y
(2,0)
EZ
EW
EW
v e3 Y
v f2 X
v f3 X
EZ
v e4
Y
EW
v f4 X
EX
v e5
Figure 7: A pair (X, Y ) and the corresponding admissible graph for the case 3a.
(0,2)
(0,2)
(0,2)
(Ce∗ ; c∗e )
X
(0,0)
Y
f1 (1,1)
f2 (1,1)
f3 (1,1)
e3
e4
f4 (1,1)
e1
e2 (1,1)
(2,1)
(1,1)
(2,0)
(2,0)
EY
v e1 Y
EW
v f1
X
EZ
v e2 Y
EW
v f2
X
EZ
v e3 Y
(1,1)
(2,0)
EW
v f3
X
EZ
v e4
EX
Y
v f4
Figure 8: A pair (X, Y ) and the corresponding admissible graph for the case 3b.
We now turn to the case EX ∩ {e ∈ E : ve ∈ V A } = ∅. Fix δ > 0 (the precise value of δ
will be specified later) and set:
Ce (δ) = Ce∗ − δ,
Ce (δ) =
Ce∗ ,
ce (δ) = c∗e
ce (δ) =
c∗e
ve ∈ V A ,
−δ
A
ve ∈
/V .
(23a)
(23b)
Lemma 4. There exists a sufficiently small δ > 0 such that the costs Ce (δ) and ce (δ) satisfy
the path optimality conditions for X and Y , respectively, i.e:
for every e ∈
/X
Ce (δ) ≥ Cf (δ)
for every f ∈ PX (e),
(24a)
for every e ∈
/Y
ce (δ) ≥ cf (δ)
for every f ∈ PY (e).
(24b)
Proof. If Ce∗ > Cf∗ (resp. c∗e > c∗f ), e ∈
/ X, f ∈ PX (e) (resp. e ∈
/ Y, f ∈ PY (e)), then there
is δ > 0, such that after setting the new costs (23) the inequality Ce (δ) ≥ Cf (δ) (resp.
ce (δ) ≥ cf (δ)) holds. Hence, one can choose a sufficiently small δ > 0 such that after setting
12
the new costs (23), all the strong inequalities are not violated. Therefore, for such a chosen δ
it remains to show that all originally tight inequalities in (22) are preserved for the new costs.
Consider a tight inequality of the form:
/ X, f ∈ PX (e).
Ce∗ = Cf∗ , e ∈
(25)
On the contrary, suppose that Ce (δ) < Cf (δ). This is only possible when Ce (δ) = Ce∗ − δ and
Cf (δ) = Cf∗ . Hence and from the construction of the new costs, we have vf ∈
/ V A (see (23b))
A
A
A
and ve ∈ V (see (23a)). By (25), we obtain (ve , vf ) ∈ E . Thus vf ∈ V , a contradiction.
Consider a tight inequality of the form:
/ Y, f ∈ PY (e).
c∗e = c∗f , e ∈
(26)
On the contrary, suppose that ce (δ) < cf (δ). This is only possible when ce (δ) = c∗e − δ and
/ V A and vf ∈ V A (see (23)). From (26), it follows that
cf (δ) = c∗f . Thus we deduce that ve ∈
(vf , ve ) ∈ E A and so ve ∈ V A , a contradiction.
We are now ready to give the precise value of δ. We do this by increasing the value of δ
until some inequalities, originally not tight in (22), become tight. Namely, let δ∗ > 0 be the
smallest value of δ for which an inequality originally not tight becomes tight. Obviously, it
/ Y , e ∈ PY (f ).
/ X, f ∈ PX (e) or c∗f − δ∗ = c∗e for f ∈
occurs when Ce∗ − δ∗ = Cf∗ for e ∈
A
A
∗
By (23), ve ∈ V and vf ∈
/ V . Accordingly, if δ = δ , then at least one arc is added to GA .
Observe also that no arc can be removed from GA - the admissibility of the nodes remains
unchanged. It follows from the fact that each tight inequality for ve ∈ V A and vf ∈ V A is
still tight. This leads to the following lemma.
Lemma 5. If EX ∩{e ∈ E : ve ∈ V A } = ∅, then (X, Y ) satisfies the sufficient pair optimality
conditions for each θ ′ ∈ [θ, θ + δ∗ ].
Proof. Set θ ′ = θ + δ, δ ∈ [0, δ∗ ]. Lemma 4 implies that X is optimal for Ce (δ) and Y
is optimal for ce (δ). From (23) and the definition of the costs Ce∗ and c∗e , it follows that
Ce (δ) = Ce − α′e and ce (δ) = ce − βe′ , where α′e = αe + δ and βe′ = βe for each ve ∈ V A ,
α′e = αe and βe′ = βe + δ for each ve ∈
/ V A . Notice that α′e + βe′ = αe + βe + δ = θ + δ = θ ′ for
each e ∈ E. By (23), ce (δ) = ce for each e ∈ EY (recall that e ∈ EY implies ve ∈ V A ), and
thus βe = 0 for each e ∈ EY . Since EX ∩ {e ∈ E : ve ∈ V A } = ∅, Ce (δ) = Ce∗ = Ce holds for
each e ∈ EX , and so αe = 0 for each e ∈ EX . We thus have shown that there exist α′e , βe′ ≥ 0
such that α′e + βe′ = θ ′ for each e ∈ E satisfying the conditions (i) and (ii) in Theorem 2,
which completes the proof.
We now describe a polynomial procedure that, for a given pair (X, Y ) satisfying the
sufficient pair optimality conditions for some θ ≥ 0, finds a new pair of spanning trees (X ′ , Y ′ ),
which also satisfies the sufficient pair optimality conditions with |EZ′ | = |EZ | + 1. We start
by building the admissible graph GA = (V A , E A ) for (X, Y ). If this graph contains an
augmenting path, then by Lemma 3, we are done. Otherwise, we determine δ∗ and modify
the costs by using (23). Lemma 5 shows that (X, Y ) satisfies the sufficient pair optimality
conditions for θ + δ∗ . For δ∗ some new arcs are added to the admissible graph GA (all the
previous arcs must be still present in GA ). Thus GA is updated and we set Ce∗ := Ce (δ∗ ),
c∗e := ce (δ∗ ) for each e ∈ E, and θ := θ + δ∗ . We repeat this until there is an augmenting path
in GA = (V A , E A ). Note that such a path must appear after at most m = |E| iterations,
13
(Ce∗ ; c∗e )
X
Y
e4
(0,9)
(8,3)
e8
(2,5)
e5
e3 (8,3)
e7 (2,3)
(1,2)
(1,1)
e9
e2
(3,2)
(0,6)
e10 (5,9)
e6
(7,2)
e11
e1
(a)
θ=0
EY
v e8 Y
EW
v e7
EY
ve11
EY
v e3
e4
(0,8)
(7,3)
e8
(2,4)
e5
e3 (7,3)
e7 (1,3)
(1,1)
(1,0)
e9
e2
(3,1)
(0,5)
e10 (5,8)
1; θ = 1
e6
(6,2)
e11
e1
(b)
δ∗ =
EY
v e8 Y
EW
EZ
v e7 X v e5
EY
ve11
EY
v e3
e4
(0,7)
(6,3)
e8
(2,3)
e5
e3 (6,3)
e7 (0,3)
(0,1)
(1,-1)
e9
e2
(3,0)
(0,4)
e10 (5,7)
e6
1; θ = 2
(5,2)
e11
e1
(c)
δ∗ =
EY Y
v e8
EW
v e7 X
EY X
ve11
EX
ve10
EY
v e3 Y
EW
v e1
X
EZ
v e5
EX
v e6
Figure 9: Sample computations, X = {e2 , e4 , e5 , e6 , e9 , e10 } and Y = {e2 , e3 , e5 , e8 , e9 , e11 }.
which follows from the fact that at some step a node ve such that e ∈ EX must appear in
GA .
Sample computations are shown in Figure 9. We start with the pair (X, Y ), where
X = {e2 , e4 , e5 , e6 , e9 , e10 } and Y = {e2 , e3 , e5 , e8 , e9 , e11 }, which satisfies the sufficient pair
optimality conditions for θ = 0 (see Figure 9a). Observe that in this case it is enough to check
that X is optimal for the costs Ce∗ = Ce and Y is optimal for the costs c∗e = ce , e ∈ E. For
θ = 0, the admissible graph does not contain any augmenting path. We thus have to modify
the costs Ce∗ and c∗e , according to (23). For δ∗ = 1, a new inequality becomes tight and one
arc is added to the admissible graph (see Figure 9b). The admissible graph still does not have
an augmenting path, so we have to again modify the costs. For δ∗ = 1 some new inequalities become tight and three arcs are added to the admissible graph (see Figure 9c). Now
the admissible graph has two augmenting paths (cases 1 and 3a, see the proof of Lemma 3).
Choosing one of them, and performing the modification described in the proof of Lemma 3
we get a new pair (X ′ , Y ′ ) with |EZ ′ | = |EZ | + 1.
Let us now estimate the running time of the procedure. The admissible graph has at most
m nodes and at most mn arcs. It can be built in O(nm) time. The augmenting path in
the admissible graph can be found in O(nm) time by applying the breath first search. Also
the number of inequalities which must be analyzed to find δ∗ is O(nm). Since we have to
update the cost of each arc of the admissible graph at most m times, until an augmenting
path appears, the required time of the procedure is O(m2 n). We thus get the following result.
14
Theorem 3. The Rec ST problem is solvable in O(Lm2 n) time, where L = n − 1 − k.
3
The recoverable robust spanning tree problem
In this section we are concerned with the Rob Rec ST problem under the interval uncertainty representation, i.e. for the scenario sets U I , U1I (Γ), and U2I (Γ). Using the polynomial
algorithm for Rec ST, constructed in Section 2, we will provide a polynomial algorithm for
Rob Rec ST under U I and some approximation algorithms for a wide class of Rob Rec
ST under U1I (Γ) and U2I (Γ). The idea will be to solve Rec ST for a suitably chosen second
stage costs. Let
X
F (X) =
Ce + max min f (Y, S),
e∈X
S∈U Y ∈Φk
X
e∈X
Y ∈ΦkX
P
where f (Y, S) = e∈Y cSe . It is worth pointing out that under scenario sets U I and U2I (Γ),
the value of F (X), for a given spanning tree X, can be computed in polynomial time [8, 17].
On the other hand, computing F (X) under U1I (Γ) turns out to be strongly NP-hard [17, 9].
Given scenario S = (cSe )e∈E , consider the following Rec ST problem:
!
X
Ce + min f (Y, S) .
(27)
min
X∈Φ
Problem (27) is equivalent to the formulation (1) for S = (ce )e∈E and it is polynomially
solvable, according to the result obtained in Section 2. As in the previous section, we denote
by pair (X, Y ) a solution to (27), where X ∈ Φ and Y ∈ ΦkX . Given S, we call (X, Y ) an
optimal pair under S if (X, Y ) is an optimal solution to (27).
The Rob Rec ST problem with scenario set U I can be rewritten as follows:
!
!
X
X
X
X
Ce + min
cSe = min
(ce + de ) .
(28)
min
Ce + max min
X∈Φ
e∈X
S∈U I Y ∈ΦkX
e∈Y
X∈Φ
e∈X
Y ∈ΦkX
e∈E
Thus (28) is (27) for S = (ce + de )e∈E ∈ U I . Hence and from Theorem 3 we immediately get
the following theorem:
Theorem 4. For scenario set U I , the Rob Rec ST problem is solvable in O((n − 1− k)m2 n)
time.
We now address Rob Rec ST under U1I (Γ) and U2I (Γ). Suppose that ce ≥ α(ce + de )
for each e ∈ E, where α ∈ (0, 1] is a given constant. This inequality means that for each
edge e ∈ E the nominal cost ce is positive and ce + de is at most 1/α greater than ce . It is
reasonable to assume that this condition will be true in many practical applications for not
very large value of 1/α.
Lemma 6. Suppose that ce ≥ α(ce + de ) for each e ∈ E, where α ∈ (0, 1], and let (X̂, Ŷ ) be
an optimal pair under S = (ce )e∈E . Then for the scenario sets U1I (Γ) and U2I (Γ) the inequality
F (X̂) ≤ α1 F (X) holds for any X ∈ Φ.
Proof. We give the proof only for the scenario set U1I (Γ). The proof for U2I (Γ) is the same.
Let X ∈ Φ. The following inequality is satisfied:
X
X
X
F (X) =
Ce + max min f (Y, S) =
Ce + f (Y ∗ , S ∗ ) ≥
Ce + f (Y ∗ , S).
e∈X
S∈U1I (Γ) Y ∈ΦkX
e∈X
15
e∈X
Clearly, (X, Y ∗ ) is a feasible pair to (27) under S. From the definition of (X̂, Ŷ ) we get
X
X
X
X
X
X
F (X) ≥
Ce + f (Ŷ , S) =
Ce +
ce ≥
Ce +
α(ce + de ) =
Ce + αf (Ŷ , S),
e∈X̂
e∈X̂
e∈Ŷ
e∈X̂
e∈Ŷ
e∈X̂
(29)
where S = (ce + de )e∈E . Hence
X
X
F (X) ≥
Ce + α max f (Ŷ , S) ≥
Ce + α max
S∈U1I (Γ)
e∈X̂
≥ α
X
Ce + max
e∈X̂
e∈X̂
X̂
min f (Y, S) = αF (X̂)
S∈U1I (Γ) Y ∈Φk
X̂
and the lemma follows.
min f (Y, S)
S∈U1I (Γ) Y ∈Φk
The condition ce ≥ α(ce + de ), e ∈ E, in Lemma 6, can be weakened and, in consequence,
the set of instances to which the approximation ratio of the algorithm applies can be extended.
Indeed, from inequality (29) it follows
that the
P
P bounds of the uncertainty intervals are only
required to meet the condition e∈Ŷ ce ≥ α e∈Ŷ (ce + de ). This condition can be verified
efficiently, since Ŷ can be computed in polynomial time. P
We now focus on Rob Rec ST for U2I (Γ). Define D = e∈E de and suppose that D > 0
(if D = 0, then the problem is equivalent to Rec ST for the second stage costs ce , e ∈ E).
′
Consider scenario S ′ under which cSe = min{ce + de , ce + Γ dDe } for each e ∈ E. Obviously,
P
P
de
S ′ ∈ U2I (Γ), since
e∈E δe ≤
e∈E Γ D ≤ Γ. The following theorem provides another
approximation result for Rob Rec ST with scenario set U2I (Γ):
Lemma 7. Let (X̂, Ŷ ) be an optimal pair under S ′ . Then the following implications are true
for scenario set U2I (Γ):
(i) If Γ ≥ βD, β ∈ (0, 1], then F (X̂) ≤ β1 F (X) for any X ∈ Φ.
(ii) If Γ ≤ γF (X̂), γ ∈ [0, 1) then F (X̂) ≤
1
1−γ F (X)
for any X ∈ Φ.
Proof. Let X ∈ Φ. Since S ′ ∈ U2I (Γ), we get
X
X
Ce + min f (Y, S ′ ).
F (X) =
Ce + max min f (Y, S) ≥
e∈X
S∈U2I (Γ) Y ∈ΦkX
e∈X
Y ∈ΦkX
(30)
We first prove implication (i). By (30) and the definition of (X̂, Ŷ ), we obtain
X
X
X
de
F (X) ≥
Ce + f (Ŷ , S ′ ) =
Ce +
min{ce + de , ce + Γ }
D
e∈X̂
e∈Ŷ
e∈X̂
X
X
X
X
X
≥
Ce +
min{ce + de , ce + βde } =
Ce +
(ce + βde ) ≥
Ce + βf (Ŷ , S),
e∈X̂
e∈Ŷ
e∈X̂
e∈Ŷ
e∈X̂
where S = (ce + de )e∈E . The rest of the proof is the same as in the proof of Lemma 6. We
now prove implication (ii). By (30) and the definition of (X̂, Ŷ ), we have
X
X
X
F (X) ≥
Ce + f (Ŷ , S ′ ) ≥
Ce + f (Ŷ , S) ≥
Ce + max f (Ŷ , S) − Γ
e∈X̂
≥
X
e∈X̂
e∈X̂
Ce + max
e∈X̂
min f (Y, S) − Γ = F (X̂) − Γ.
S∈U2I (Γ) Y ∈Φk
X̂
16
S∈U2I (Γ)
If Γ ≤ γF (X̂). Then F (X) ≥ F (X̂) − γF (X̂) = (1 − γ)F (X̂) and F (X̂) ≤
1
1−γ F (X).
Note that the value of F (X̂) under U2I (Γ) can be computed in polynomial time [17]. In
consequence, the constants β and γ can be efficiently determined for every particular instance
of the problem. Clearly, we can assume that de ≤ Γ for each e ∈ E, which implies D ≤ mΓ,
1
D for every instance of the problem.
where m = |E|. Hence, we can assume that Γ ≥ m
We thus get from Lemma 7 (implication (i)) that F (X̂) ≤ mF (X) for any X ∈ Φ and the
problem is approximable within m. If α, β and γ are the constants from Lemmas 6 and 7,
then the following theorem summarizes the approximation results:
Theorem 5. Rob Rec ST is approximable within α1 under scenario set U1I (Γ) and it is
1
} under scenario set U2I (Γ).
approximable within min{ β1 , α1 , 1−γ
Observe that Lemma 6 and Lemma 7 hold of any sets Φ and ΦkX (the particular structure
of these sets is not exploited). Hence the approximation algorithms can be applied to any
problem for which the recoverable version (27) is polynomially solvable.
4
Conclusions
In this paper we have studied the recoverable robust spanning tree problem (Rob Rec ST)
under various interval uncertainty representations. The main result is the polynomial time
combinatorial algorithm for the recoverable spanning tree. We have applied this algorithm
for solving Rob Rec ST under the traditional uncertainty representation (see, e.g., [13]) in
polynomial time. Moreover, we have used the algorithm for providing several approximation
results for Rec ST with the scenario set introduced by Bertsimas and Sim [3] and the scenario
set with a budged constraint (see, e.g,. [17]). There is a number of open questions concerning
the considered problem. Perhaps, the most interesting one is to resolve the complexity of
the robust problem under the interval uncertainty representation with budget constraint. It
is possible that this problem may be solved in polynomial time by some extension of the
algorithm constructed in this paper. One can also try to extend the algorithm for the more
general recoverable matroid base problem, which has also been shown to be polynomially
solvable in [10].
Acknowledgements
The second and the third authors were supported by the National Center for Science (Narodowe Centrum Nauki), grant 2013/09/B/ST6/01525.
References
[1] R. K. Ahuja, T. L. Magnanti, and J. B. Orlin. Network Flows: theory, algorithms, and
applications. Prentice Hall, Englewood Cliffs, New Jersey, 1993.
[2] H. Aissi, C. Bazgan, and D. Vanderpooten. Approximation of min-max and min-max
regret versions of some combinatorial optimization problems. European Journal of Operational Research, 179:281–290, 2007.
17
[3] D. Bertsimas and M. Sim. Robust discrete optimization and network flows. Mathematical
Programming, 98:49–71, 2003.
[4] C. Büsing. Recoverable robustness in combinatorial optimization. PhD thesis, Technical
University of Berlin, Berlin, 2011.
[5] C. Büsing. Recoverable robust shortest path problems. Networks, 59:181–189, 2012.
[6] C. Büsing, A. M. C. A. Koster, and M. Kutschka. Recoverable robust knapsacks: the
discrete scenario case. Optimization Letters, 5:379–392, 2011.
[7] A. Chassein and M. Goerigk. On the recoverable robust traveling salesman problem.
Optimization Letters, DOI 10.1007/s11590-015-0949-5, 2015.
[8] O. Şeref, R. K. Ahuja, and J. B. Orlin. Incremental network optimization: theory and
algorithms. Operations Research, 57:586–594, 2009.
[9] N. G. Frederickson and R. Solis-Oba. Increasing the weight of minimum spanning trees.
Journal of Algorithms, 33:244–266, 1999.
[10] M. Hradovich, A. Kasperski, and P. Zieliński. The recoverable robust spanning
tree problem with interval costs is polynomially solvable. Optimization Letters, DOI
10.1007/s11590-016-1057-x, 2016.
[11] A. Kasperski, A. Kurpisz, and P. Zieliński. Recoverable robust combinatorial optimization problems. In Operations Research Proceedings 2012, pages 147–153, 2014.
[12] A. Kasperski and P. Zieliński. On the approximability of robust spanning problems.
Theoretical Computer Science, 412:365–374, 2011.
[13] P. Kouvelis and G. Yu. Robust Discrete Optimization and its Applications. Kluwer
Academic Publishers, 1997.
[14] L. C. Lau, R. Ravi, and M. Singh. Iterative Methods in Combinatorial Optimization.
Cambridge University Press, 2011.
[15] C. Liebchen, M. E. Lübbecke, R. H. Möhring, and S. Stiller. The concept of recoverable
robustness, linear programming recovery, and railway applications. In Robust and Online
Large-Scale Optimization, volume 5868 of Lecture Notes in Computer Science, pages 1–
27. Springer-Verlag, 2009.
[16] K. Lin and M. S. Chern. The most vital edges in the minimum spanning tree problem.
Information Processing Letters, 45:25–31, 1993.
[17] E. Nasrabadi and J. B. Orlin. Robust optimization with incremental recourse. CoRR,
abs/1312.4075, 2013.
[18] C. H. Papadimitriou and K. Steiglitz. Combinatorial optimization: algorithms and complexity. Dover Publications Inc., 1998.
18
| 8cs.DS
|
arXiv:1111.1294v5 [math.AC] 23 Jan 2013
When the associated graded ring of a
semigroup ring is Complete Intersection
M. D’Anna∗
V. Micale
A. Sammartano
Abstract
Let (R, m) be the semigroup ring associated to a numerical semigroup
S. In this paper we study the property of its associated graded ring
grm (R) to be Complete Intersection. In particular, we introduce and
characterize β-rectangular and γ-rectangular Apéry sets, which will be
the fundamental concepts of the paper and will provide, respectively,
a sufficient condition and a characterization for grm (R) to be Complete Intersection. Then we use these notions to give four equivalent
conditions for grm (R) in order to be Complete Intersection.
MSC: 13A30; 13H10.
Introduction
Let (R, m) be a Noetherian local ring with |R/m| = ∞ and let grm (R) =
⊕i≥0 mi /mi+1 be the associated graded ring of R with respect to m. The study
of the properties of grm (R) is a classical subject in local algebra, not only in
the general d-dimensional case, but also under particular hypotheses (that
allow to obtain more precise results). One main problem in this context is to
estimate the depth of grm (R) and to understand when this ring is a CohenMacaulay ring (see, e.g., [17], [18] and [21]). In connection to this problem,
it is natural to investigate if grm (R) is a Buchsbaum ring (see [11], [12]), a
Gorenstein ring or if it is Complete Intersection (see [13]).
In this paper we are interested in the properties of grm (R), when R is a numerical semigroup ring. The study of numerical semigroup rings is motivated
∗
email [email protected]
1
by their connection to singularities of monomial curves and by the possibility
of translating algebraic properties into numerical properties. However, even
in this particular case, many pathologies occur, hence these rings are also a
great source of interesting examples.
In the numerical semigroup case, the Cohen-Macaulayness of grm (R) has
been extensively studied (see, e.g., [10], [15] and [20]); recently, different
authors studied the Buchsbaumness and the Gorensteinness of grm (R) (see
[3], [5], [6] and [19]).
In this paper we investigate when grm (R) is Complete Intersection. About
this problem not much is known (see [1]). When the embedding dimension
of R is small, it is possible to list the generators of the defining ideals of
R and of grm (R) (as it is done in [14] and [20] when the embedding dimension is 2 or 3), but, as soon as the number of generators of m increases, the
computations become too huge. On the other hand, a useful tool to study
the general case (when m is n-generated) is the so called Apéry set of the
semigroup. The properties of the Apéry set reveal much information on the
Cohen-Macaulayness, the Buchsbaumness and the Gorensteinness of grm (R).
In this paper, using the Apéry set, we are able to characterize when grm (R)
is Complete Intersection.
The structure of the paper is the following. In Section 1 we fix the
notation and give some preliminaries about numerical semigroups, semigroup
rings associated to a numerical semigroup and their associated graded ring.
In particular, we qualitatively discuss the form of the elements in the defining
ideals of R, of its quotient R modulo an element of minimal value and of their
associated graded rings (cf. Discussion 1.5).
In Section 2 we define two sets of integers βi and γi , that yield to the
definition and the characterization of two classes of numerical semigroups:
semigroups with β-rectangular and γ-rectangular Apéry set (cf. Definition
2.9, Theorem 2.16 and Theorem 2.22). These notions will provide, respectively, a sufficient condition and a characterization for grm (R) to be Complete
Intersection. These classes are strictly connected and it is useful to study
both of them together. The idea in these definitions is to use the “shape” of
the Apéry set of the semigroup in order to find integers that give information
on the degree and on the nature of the generators of the defining ideals of R,
of its quotient R modulo an element of minimal value and of their associated
graded rings.
In Section 3 we prove the main theorem of the paper (cf. Theorem 3.6),
2
that is the characterization of those numerical semigroup rings whose associated graded ring is Complete Intersection. More precisely, we give four
equivalent conditions for grm (R) in order to be Complete Intersection. To
obtain this result we need to deepen carefully the nature of the elements of
the defining ideals of R, of its quotient R modulo an element of minimal value
and of their associated graded rings, using the integers βi and γi introduced
in the previous section (cf. Discussion 3.2 and Lemma 3.4). Finally, we give
an alternative proof of a sufficient condition for grm (R) being Complete Intersection presented in [2] (cf. Corollary 3.13) and we briefly study the case
of embedding dimension 3 (cf. Theorem 3.14).
The computations made for this paper are performed by using the GAP
system [9] and, in particular, the NumericalSgps package [7].
1
Preliminaries
Let N denote the set of natural numbers, including 0. A numerical semigroup
is a submonoid S of the monoid (N, +) with finite complement in it. Each
numerical semigroup S has a natural partial ordering where, for every s
and t in S, s t if there is an element u ∈ S such that t = s + u. The set
{gi } of the minimal elements in the poset (S \ {0}, ) is called minimal set of
generators for S; indeed all the elements in S are linear combinations, with
coefficients in N, of minimal elements. Note that the set {gi } is finite since for
any s ∈ S, s 6= 0, we have gi 6≡ gj (mod s) if i 6= j. A numerical semigroup
minimally generated by g1 < g2 < . . . < gν is denoted by hg1 , g2 , . . . , gν i; the
condition |N \ S| < ∞ is equivalent to gcd(g1 , . . . , gν ) = 1.
There are several invariants associated to a numerical semigroup S. The
integer m = g1 = min{s ∈ S, s > 0} is called multiplicity, while the minimal
number of generators ν is called embedding dimension; it is well known that
ν ≤ m. Finally, the integer f = max{z ∈ Z, z ∈
/ S} is called Frobenius
number of S.
Following the notation in [2], we denote by Ap(S) = {ω0 , . . . , ωm−1 } the
Apéry set of S with respect to m, that is, the set of the smallest elements
in S in each congruence class modulo m. More precisely, ω0 = 0 and ωi =
min{s ∈ S | s ≡ i (mod m)}. The largest element in the Apéry set is always
f + m.
Very often we will use the following result.
3
Proposition 1.1 ([8], Lemma 6). Let t ∈ Ap(S) and u t, then u ∈ Ap(S).
A numerical semigroup S is called symmetric if f − x ∈
/ S implies that
x ∈ S for every integer x (notice that the converse is true for every numerical
semigroup S).
Proposition 1.2 ([16], Corollary 4.12). S is symmetric if and only if f + m
is the unique maximal element in (Ap(S), ).
An ideal of a semigroup S is a nonempty subset H of S such that
H + S ⊆ H. The ideal M = {s ∈ S | s 6= 0} is called maximal ideal
of S. It is straightforward to see that, if H and L are ideals of S, then
H + L = {h + l| h ∈ H, l ∈ L} and kH(= H + · · · + H, k summands, for
k ≥ 1) are also ideals of S.
Let k be an infinite field; the rings R = k[[tS ]] = k[[tg1 , . . . , tgν ]] and
R = k[tS ]m are called the numerical semigroup rings associated to S. The ring
R is a one-dimensional local domain, with maximal ideal m = (tg1 , . . . , tgν )
and quotient field k((t)) and k(t), respectively. In both cases the associated
graded ring of R with respect to m, grm (R) = ⊕i≥0 mi /mi+1 , is the same.
From now on, we will assume that R = k[[tS ]], but the other case is perfectly
analogous.
Let (A, n) be the local ring of formal power series k[[x1 , . . . , xν ]] and let
ϕ : A −→ R be the map defined by ϕ(xi ) = tgi . Clearly R = A/I and
m = n/I, with I = ker ϕ. Notice that I is a binomial ideal generated by all
the elements of the form
(∗)
xj11 · · · · · xjνν − xh1 1 · · · · · xhν ν ,
with j1 g1 + · · · + jν gν = h1 g1 + · · · + hν gν .
It is well known that this presentation induces a presentation of the corresponding associated graded rings:
ψ : grn (A) −→ grm (R),
where the kernel is the initial ideal of I, i.e. the ideal I ∗ generated by the initial forms of the elements of I; hence grm (R) ∼
= grn (A)/I ∗ ∼
= k[x1 , . . . , xν ]/I ∗
canonically.
Notice that I ∗ is an homogenous ideal generated by all the monomials of
the form xj11 · · · · · xjνν coming from a binomial (∗) for which j1 + · · · + jν <
h1 +· · ·+hν and by all the binomials (∗) such that j1 +· · ·+jν = h1 +· · ·+hν .
4
We denote by µ(·) the minimal number of generators of an ideal. The
ring R is Complete Intersection if µ(I) = ν − 1 and the associated graded
ring grm (R) is Complete Intersection if µ(I ∗ ) = ν − 1. It is well known that,
if grm (R) is Complete Intersection, then also R is Complete Intersection.
Numerical semigroups for which R is Complete Intersection are well known
(and they are called Complete Intersection numerical semigroups; for the definition see, e.g., [16]). We are interested in studying when grm (R) is Complete Intersection. Let R = R/(tm ) and G = grm (R), where m is the maximal
ideal of R.
Remark 1.3. In our hypotheses, it is clear that grm (R) is Complete Intersection if and only if G is Complete Intersection and grm (R) is Cohen-Macaulay
(i.e., as it is proved in [10], (tm )∗ ∈ m/m2 is not a zero-divisor in grm (R)). In
fact, in case grm (R) is Cohen-Macaulay, from the isomorphism R/(tm ) ∼
=R
we get the isomorphism grm (R)/((tm )∗ ) ∼
= G. We notice also that, in general, there is a surjective homomorphism of graded rings grm (R) → G, whose
kernel is the initial ideal of (tm ) in grm (R).
Remark 1.4. We note that R = htωi | ωi ∈ Ap(S)ik , since
ts = 0 in R ⇐⇒ ts ∈ (tm ) ⇐⇒ s − m ∈ S.
We also have G = R as k-vector spaces (but not as rings) since a nonzero
monomial in R is still nonzero in G.
Discussion 1.5. Recalling the isomorphism R ∼
= k[[x1 , x2 , . . . , xν ]]/I, where xi
corresponds to tgi (hence x1 corresponds to tm ), we have the isomorphism R ∼
=
k[[x2 , x3 , . . . , xν ]]/H. More precisely H is the kernel of the homomorphism
defined by xi 7→ tgi and it is generated by all the binomials of the form
(∗∗)
xj22 · · · · · xjνν − xh2 2 · · · · · xhν ν ,
with j2 g2 + · · · + jν gν = h2 g2 + · · · + hν gν ∈ Ap(S), and by all the monomials
of the form
(∗ ∗ ∗)
xj22 · · · · · xjνν ,
where j2 g2 + · · · + jν gν ∈
/ Ap(S).
It follows that G = grm (R) ∼
= k[x2 , x3 , . . . , xν ]/J, where J is the kernel
of the homomorphism defined by xi 7→ tgi (where now tgi is viewed as an
5
element of G) and it is the initial ideal of H. Hence J is a binomial ideal
generated by all the the binomials of the form
(+)
xj22 · · · · · xjνν − xh2 2 · · · · · xhν ν ,
with j2 g2 +· · ·+jν gν = h2 g2 +· · ·+hν gν ∈ Ap(S) and j1 +· · ·+jν = h1 +· · ·+hν ,
and by all the monomials of the form
(++)
xj22 · · · · · xjνν ,
where either j2 g2 + · · · + jν gν ∈
/ Ap(S) or j2 g2 + · · · + jν gν ∈ Ap(S) and there
exist h2 , . . . , hν ∈ N, such that j2 g2 + · · · + jν gν = h2 g2 + · · · + hν gν and
j2 + · · · + jν < h2 + · · · + hν .
In particular, let j = j2 + · · · + jν ; then a binomial of the form (+) is not
necessary as generator of J, if xj22 · · · · · xjνν ∈ (x2 , . . . , xν )j+1. Furthermore,
the monomial xj22 · · · · · xjνν ∈ (x2 , . . . , xν )j \ (x2 , . . . , xν )j+1, such that j2 g2 +
· · · + jν gν ∈ Ap(S), does not belong to J.
Finally, since the Krull dimension of G is 0, we must have µ(J) ≥ ν − 1.
Hence G is Complete Intersection if and only if µ(J) = ν − 1.
From the previous remarks and discussion, it is clear why, to study the
Complete Intersection property for grm (R) it is necessary to study the Apéry
set of S. When G is Complete Intersection, its Hilbert function is completely
determined by the degree of the generators of J and its dimension as k-vector
space is the product of these degrees. Hence we will have to determine
these degrees using numerical conditions; moreover, since monomials in G
correspond bijectively to elements of the Apery set, we will have to determine
the “shape” of Ap(S) corresponding to G Complete Intersection.
2
β- and γ-rectangular Apéry Sets
Within this section we introduce two sets of integers and two corresponding
classes of numerical semigroups, defined via the shape of the Apéry set. The
first class will provide a sufficient condition for G to be complete intersection,
while the second one will give a characterization.
Given a numerical semigroup S = hg1 , g2 , . . . , gν i and s, t ∈ S, we recall
that s t if there exists u ∈ S such that s + u = t. Now we want to define
another partial ordering on S as in [3]. If s ∈ S and M = S \ {0} then
6
there exists a unique h ∈ N such that s ∈ hM \ (h + 1)M; this integer is
defined as the order of s and we will write ord(s) = h. Given s, t ∈ S, we
say that s M t if there exists u ∈ S such that s + u = t (hence s t) and
ord(s) + ord(u) = ord(t). The partial order M is particularly helpful in the
study of the associated graded ring.
The sets of maximal elements of Ap(S) with respect to and M are
denoted with maxAp(S) and maxApM (S), respectively.
Remark 2.1. We note that maxAp(S) ⊆ maxApM (S) and the inclusion can
be strict. For example, let S = h8, 9, 15i. The only maximal element in
Ap(S) with respect to is 45. Anyway maxApM (S) = {30, 45}. Note that
ord(45) = 5 > 3 = ord(30) + ord(15).
A numerical semigroup S is called M-pure if every element in maxApM (S)
has the same order. M-pure symmetric semigroups are characterized in a
similar way to symmetric semigroups:
Proposition 2.2 ([3], Proposition 3.7). A semigroup S is M-pure symmetric
if and only if ω M f + m, for every ω ∈ Ap(S).
Every element s ∈ S can be written, not necessarily in a unique way,
as s = λ1 g1 + · · · + λν gν ; we call this combination of the generators a representation of s. Throughout the paper, we call “representation” both the
expression s = λ1 g1 + · · · + λν gν and the tuple (λ1 , λ2 , . . . , λν ). We say that
an element s ∈ S has a unique representation if it can be written in a unique
way as a linear combination of g1 , g2 , . . . , gν . Notice that, by definition of
Apéry set, an element ω ∈ Ap(S) can have only representations where g1
does not appear.
A representation of an element s ∈ S as s = λ1 g1 + λ2 g2 + · · · + λν gν is
called maximal if λ1 + λ2 + · · · + λν = ord(s). This kind of representations
and, in particular, the number of maximal representations of elements of S
have been studied in [4]. We say that Ap(S) is of unique maximal expression
if every ω ∈ Ap(S) has a unique maximal representation.
Let us define now the following integers, for every i = 2 . . . , ν:
βi = max{h ∈ N | hgi ∈ Ap(S) and ord(hgi ) = h};
γi = max{h ∈ N | hgi ∈ Ap(S), ord(hgi ) = h and
hgi has a unique maximal representation}.
7
Notice that, in the second definition, from ord(hgi ) = h it follows that hgi
must be the unique maximal representation. The following proposition is
straightforward:
Proposition 2.3. For each index i = 2, . . . , ν, we have γi ≤ βi .
Examples 2.4. Let S = h8, 10, 15i; so Ap(S) = {0, 10, 15, 20, 25, 30, 35, 45}.
Since 30, 45 ∈ Ap(S) and 40, 60 ∈
/ Ap(S) then β2 , β3 ≤ 3. We have the
double representation 30 = 3 · 10 = 2 · 15, implying that ord(2 · 15) = 3 > 2;
hence β2 = 3 and β3 = 1. It is easy to check that every element in Ap(S)
has a unique maximal representation, thus γ2 = β2 = 3 and γ3 = β3 = 1.
Let S = h7, 9, 10, 11, 12i; we have Ap(S) = {0, 9, 10, 11, 12, 20, 22}. Analogously to the previous example, we have β2 = 1, β3 = 2, β4 = 2, β5 = 1.
The only double representations in Ap(S) are 20 = 9 + 11 = 10 + 10 and
22 = 10 + 12 = 11 + 11, and they are all maximal. In particular, it follows
γ2 = γ3 = γ4 = γ5 = 1.
In correspondence to the two families of numbers introduced above, we
can define the following two sets:
B =
ν
nX
o
λi gi | 0 ≤ λi ≤ βi ;
Γ =
ν
nX
o
λi g i | 0 ≤ λi ≤ γ i .
i=2
i=2
The sets B, Γ, consist of the elements of S representable via a tuple (λ2 , . . . , λν )
belonging to the hyper-rectangle of Nν−1 whose vertices are respectively given
by βi and γi . By Proposition 2.3, it follows that Γ ⊆ B. Notice also that, as
can be easily seen by the previous examples,
Γ can
Qand
Qν since elements in B
ν
have more than one representation, |B| ≤ i=2 (βi +1) and |Γ| ≤ i=2 (γi +1).
It is natural to ask how the sets B and Γ, are related to the Apéry set.
The inclusion Ap(S) ⊆ B is always true, as we can deduce from the next
lemma:
P
Lemma 2.5. Let ω ∈ Ap(S) and let ω = νi=2 λi gi be a maximal representation. Then λi ≤ βi for each i.
8
Proof. By Lemma 1.1 we have λi gi ∈ Ap(S). If there is a index i such
thatP
βi < λi , then
Pνord(λi gi ) > λi , by definitions of βi . It follows that
ord( νi=2 λi gi ) >
i=2 λi and thus it is not a maximal representation of
ω; contradiction.
Notice also that an integer in B could have more different maximal representations, as can be seen in the same example as above S = h7, 9, 10, 11, 12i.
We want to show that actually the stronger inclusion Ap(S) ⊆ Γ holds;
this is a consequence of the following useful result, which is somehow analogous to Lemma 2.5. In what follows, we denote with lex and grlex respectively
the usual lexicographic order and graded lexicographic order in Nν−1 .
Lemma 2.6. Let ω ∈ Ap(S) and set
ν
ν
o
n
X
X
λi = ord(ω)
R = (λ2 , . . . , λν ) ∈ Nν−1
λi gi = ω and
i=2
i=2
i.e. the set of maximal representations of w. Let (µ2 , . . . , µν ) be the maximum
in R with respect to lex, then we have µi ≤ γi for each i.
Proof. By Lemma 2.5 we have µi ≤ βi . Let us suppose that there exists an
index i such that γi < µi ≤ βi , and take the minimum i with this property.
By definitions of βi , γi , the fact that γi < βi implies a double maximal
representation of (γi +1)gi ; moreover, since γi +1 is the least integer with this
property, the other maximal representation does not involve gi . Explicitly,
we have the relation
X
X
ηj .
ηj gj ,
γi + 1 =
(2.1)
(γi + 1)gi =
j6=i
j6=i
Let us substitute the relation just found in (µ2 , . . . , µν ), i.e. consider the
tuple (µ′2 , . . . , µ′ν ) where
µ′i = µi − γi − 1,
µ′j = µj + ηj
for j 6= i
in particular by (2.1) (µ′2 , . . . , µ′ν ) ∈ R. Now, if there is a index j < i such
that ηj 6= 0, then (µ2 , . . . , µν ) < (µ′2 , . . . , µ′ν ) with respect to lex, yielding a
contradiction to the choice of (µ2 , . . . , µν ). Hence ηj = 0 for each j <
Pi. But
this is a contradiction to (2.1), since gi < gj for i < j and γi + 1 = j>i ηj .
The lemma is proved.
9
Corollary 2.7. Let S be a numerical semigroup, then Ap(S) ⊆ Γ.
Remark 2.8. It is straightforward to check that, under the notation of Lemma
2.6, (µ2 , . . . , µν ) is also the maximum with respect to grlex in the set of (not
necessarily maximal) representations of ω.
We are interested in numerical semigroups for which the inclusions shown
so far turn out to be equalities.
Definition 2.9. Let S be a numerical semigroup:
1. the semigroup S has β-rectangular Apéry set if Ap(S) = B;
2. the semigroup S has γ-rectangular Apéry set if Ap(S) = Γ.
Corollary 2.10. We have the following implication:
Ap(S) is β-rectangular ⇒ Ap(S) is γ-rectangular.
Proof. It follows immediately by the inclusions Ap(S) ⊆ Γ ⊆ B
Remark 2.11. It would be natural to introduce also another set of integers:
αi = max{h ∈ N | hgi ∈ Ap(S)}; it is clear that αi ≥ βi . These integers,
as for the βi ’s and the γi ’s, yield to another class of semigroups (that we
could call semigroups with α-rectangular Apéry set) interesting to discuss
and somehow similar to the two classes just defined. However, this class
would provide a too strong condition with respect to the property of G to be
Complete Intersection, hence, for brevity, we will omit its study.
Examples 2.12. (1) Let S = h8, 10, 15i; we have seen in Examples 2.4 that
Ap(S) = {0, 10, 15, 20, 25, 30, 35, 45} and β2 = 3, β3 = 1. Hence we obtain
that Ap(S) is β-rectangular.
(2) Let S = h8, 10, 11, 12i; we have Ap(S) = {0, 10, 11, 12, 21, 22, 23, 33}.
Since 20, 44, 24 ∈
/ Ap(S) and ord(10) = 1, ord(33) = 3, ord(12) = 1 we have
β2 = 1, β4 = 3, β4 = 1. It is clear that γ2 = γ4 = 1; since 22 = 11 + 11 =
10 + 12 has two maximal representation, we find γ3 = 1. Thus Ap(S) is
γ-rectangular but Ap(S) is not β-rectangular.
(3) Let S = h5, 6, 9i; we have Ap(S) = {0, 6, 9, 12, 18} and the only double
representation is 18 = 3 · 6 = 2 · 9. In this case γ2 = 3, as 18 = 3 · 6 is the
unique maximal representation, while γ3 = 1 because ord(2 · 18) = 3 > 2. It
follows that Ap(S) is not γ-rectangular, since λ2 · 6 + λ3 · 9 ∈
/ Ap(S) as soon
as both λ2 > 0 and λ3 > 0.
10
Our main purpose is to characterize these kinds of numerical semigroups
in terms of some of their invariants, namely the Frobenius number and the
multiplicity. This will allow simpler tests for those properties. We also want
to study the relations between these classes of semigroups and the properties
of unique maximal expression of Ap(S) and of the partial orders and M .
The next lemmas concern maximal representations and are necessary to
establish a characterization of semigroups with β-rectangular Ap(S).
Lemma 2.13. If s M t and t has a unique maximal representation, then s
has a unique maximal representation.
Proof. We have that s + u = t for some u ∈ S and ord(s) + ord(u) =
ord(t). If s had two maximal representations, the same should be true for t;
contradiction.
Lemma 2.14. Let S be a numerical semigroup with β-rectangular Apéry set
and let ω ∈ Ap(S). Then any representation ω = λ2 g2 + · · · + λν gν , with
λi ≤ βi for each i = 2, . . . , ν, is maximal.
Proof. Assume that the representation ω = λ2 g2 + · · · + λν gν is not maximal.
Let ω = µ2 g2 + · · · + µν gν be a maximal representation of ω; since ω ∈
Ap(S), by Lemma 2.5,
have µP
i ≤ βi for every i = 2, . . . , ν. Moreover, by
Pwe
ν
maximality, we have i=2 λi < νi=2 µi , hence there exists an index i such
that λi < µi . On the other hand, since λ2 g2 + · · · + λν gν = µ2 g2 + · · · + µν gν ,
there exists another index j such that λj > µj . Subtracting from both sides
of the previous equality the common summands, we get the equality
X
X
ηi gi
ηi gi =
(2.2)
i∈T2
i∈T1
where T1 , T2 are two non-empty
disjoint
P
P subsets of {2, . . . ν}, 0 ≤ ηi ≤ βi ,
for each i = 2, . . . , ν, and i∈T1 ηi < i∈T2 ηi .
P
Since S has β-rectangular Apéry set, the element t = νi=2 βi gi belongs
to Ap(S). Let us substitute the relation (2.2) in this representation of t, that
is to say consider the tuple (ξ2 , . . . , ξν ) where
ξi = βi − ηi if i ∈ T1 ,
ξi = βi + ηi if i ∈ T2 ,
ξi = βi otherwise.
Pν
Pν
This is another representation of t and i=2 βi < i=2 ξi by (2.2). It follows
that (β2 , . . . , βν ) is not a maximal representation and, therefore, t has a
maximal representation with some coefficient bigger than βi ; contradiction
to Lemma 2.5.
11
Remark 2.15. Combining the last lemma and Lemma 2.5, we have that the
maximality of a representation (λ2 , . . . , λν ) of an element of Ap(S) is equivalent to have λi ≤ βi for each i, under the assumption of β-rectangular Apéry
set.
We are ready to give some characterizations of semigroups with β-rectangular Apéry set.
Theorem 2.16. The following conditions are equivalent:
(i) Ap(S) is β-rectangular;
(ii) Ap(S) has a unique maximal element with respect to M and this element has unique maximal representation;
(iii) S is M-pure, symmetric and Ap(S) is of unique maximal expression;
P
(iv) f + m = νi=2 βi gi ;
Q
(v) m = νi=2 (βi + 1).
Proof. (i) ⇒ (ii) Since Ap(S) is β-rectangular, we immediately get that
P
ν
i=2 βi gi is the unique maximal element in (Ap(S), ) and hence it is f + m.
Moreover, by Lemma 2.14, any element ω ∈ Ap(S) is maximally
P represented
by a tuple (λ2 , . . . , λν ) with λi ≤ βi ; in particular ord(ω) = νi=2 λi and we
can deduce that ω M f + m,
Pfor each ω ∈ Ap(S).
Finally, by Remark 2.15, νi=2 βi gi is the unique maximal representation
of f + m.
(ii) ⇒ (iii) By Proposition 2.2, S is M-pure symmetric; by Lemma 2.13,
every ω ∈ Ap(S) has a unique maximal representation.
(iii) ⇒ (iv) The unique maximal element in (Ap(S), M ) is necessarily
P
f + m. Its unique maximal representation is of the form f + m = νi=2 λi gi
with λi ≤ βi by Lemma 2.5.
Since S is M-pure symmetric βi gi M f + m for each i = 2, . . . , ν.
Hence f + m − β2 g2 ∈ Ap(S) and,
Pν by Lemma 2.13, it has unique maximal
representation f + m − β2 g2 = i=2 εi gi ; moreover εP
i ≤ βi , by maximality
of the representation. It follows that f + m = β2 g2 + νi=2 εi gi is a maximal
representation, hence ε2 = 0 (again by Lemma 2.5). Moreover, since f + m
has a unique maximal representation, λ2 = β2 ; arguing recursively, the thesis
follows.
12
(iv) ⇒ (i) It is a consequence of Proposition 1.1.
(i) ⇒ (v) It follows by m = |Ap(S)|, the fact that Ap(S) is of unique
maximal expression and Remark 2.15.
Q
(v) ⇒ (i) We already noticed that Ap(S) ⊆ B and since m = νi=2 (βi +
1) = |Ap(S)|, we must have an equality.
Example 2.17. We may apply the last theorem to show that the Apéry set of
S = h12, 14, 16, 23i is β-rectangular, without even computing the whole set
Ap(S). We need to determine the βi ’s:
2 · 14
2 · 16
3 · 16
2 · 23
=
=
=
=
12 + 16 ∈ 12 + S
32 ∈ 2M \ 3M and 32 − 12 ∈
/S
4 · 12 ∈ 12 + S
2 · 16 + 14 ∈ 3M
Q
and so β2 = 1, β3 = 2, β4 = 1 and m = 12 = 2 · 3 · 2 = νi=2 (βi + 1).
We have seen (cf. Corollary 2.10) that the following implication holds:
Ap(S) is β-rectangular ⇒ Ap(S) is γ-rectangular.
But a priori we still might have βi > γi when Ap(S) is β-rectangular, for
some i. We show that this is not the case, using the theorem just proved.
Corollary 2.18. Let S be a numerical semigroup. If Ap(S) is β-rectangular,
then βi = γi , for every i = 2, . . . , ν.
Proof. If there is an index i such that γi < βi , then, by definition of βi
and of γi , βi gi is in the Apéry set and it has more maximal representations.
Contradiction to Theorem 2.16, (ii).
We now turn to the study of semigroups with γ-rectangular Apéry set,
starting with a result that is analogous to Lemma 2.14.
Lemma 2.19. Let S be a semigroup with γ-rectangular Apéry set and let
ω ∈ Ap(S). Then any representation ω = λ2 g2 + · · · + λν gν , with λi ≤ γi
(for every i = 2 . . . , ν), is maximal.
13
Proof. Assume, by absurd,
that there exists ω ∈ Ap(S) with a non-maximal
Pν
representation ω = i=2 λi gi , where λi ≤ γi (for every i = 2 . . . , ν). Notice
that, as we
Pνhave already seen for the case Ap(S) = B, if Ap(S) = Γ, then
up to substituting the non-maximal repref + m = i=2 γi gi . In particular,
P
sentation of ω in f + m = νi=2 γi gi , we may assume ω = f + m and hence
λi = γi , for each i.
If we take a maximal representation (µ2 , . . . , µν ) of f + m, then we have
the strict inequality
(2.3)
(γ2 , . . . , γν ) < (µ2 , . . . , µν )
with respect to grlex (since the sums of the respective coefficients are different). In particular (2.3) holds if we choose (µ2 , . . . , µν ) to be the maximum
in the set of all the representations of f +m with respect to grlex. By Remark
2.8 we have
(γ2 , . . . , γν ) < (µ2 , . . . , µν ) ≤ (γ2 , . . . , γν )
with respect to grlex, and thus we reach a contradiction.
Lemma 2.20. Let S be a semigroup with γ-rectangular Apéry set. Then each
ω ∈ Ap(S) has a unique representation of the form ω = λ2 g2 + · · · + λν gν ,
with λi ≤ γi , for every i = 2, . . . ν.
Proof. Assume, by absurd, that there are two distinct representations of
ω ∈ Ap(S)
ν
ν
X
X
ω=
λi g i =
µi gi , with λi , µi ≤ γi .
i=2
i=2
By
PνLemma 2.19 both representations are maximal, and in particular
i=2 µi ; since they are distinct we have, for instance,
Pν
i=2
λi =
(λ2 , . . . , λν ) < (µ2 , . . . , µν )
with respect to lex. By adding the tuple (γ2 − λ2 , . . . , γν − λν ) to both sides
of the last equality we have two maximal representations of f + m:
(γ2 , . . . , γν ) < (µ2 + γ2 − λ2 , . . . , µν + γν − λν )
with respect to lex. We reach an absurd by Lemma 2.6.
Corollary 2.21. If S has γ-rectangular Apéry set, then it is M-pure symmetric.
14
Pν
Proof. Let ω ∈ Ap(S). By Lemma 2.6, ω =
i=2 λi gi , with λi ≤ γi ,
and,
Pν by Lemma 2.19, this representation is maximal; therefore ord(ω) =
i . Since this is valid for an arbitrary ω ∈ Ap(S), we easily have
i=2 λP
ω M νi=2 γi gi = f + m and the thesis follows by Proposition 2.2.
We are ready to characterize semigroups with γ-rectangular Apéry set.
Theorem 2.22. The following conditions are equivalent:
(i) Ap(S) is γ-rectangular;
P
(ii) f + m = νi=2 γi gi ;
Q
(iii) m = νi=2 (γi + 1).
Proof. (i) ⇒ (ii) Clear, as f + m is the biggest element in Ap(S).
(ii) ⇒ (i) It follows by Lemma 1.1.
(i) ⇒ (iii) It follows by Lemma 2.20.
Q
(iii) ⇒ (i) Since Ap(S) ⊆ Γ and |Ap(S)| = m = νi=2 (γi + 1) ≥ |Γ|, the
inclusion must be an equality.
Notice that we do not obtain a full analogous result of Theorem 2.16; more
precisely, we cannot recover conditions (ii) and (iii). The closest analogous
to those conditions is actually expressed by Lemma 2.20. If we look at
the second semigroup in Examples 2.12, S = h8, 10, 11, 12i (with Ap(S) =
{0, 10, 11, 12, 21, 22, 23, 33} and γi = 1, for each i = 2, 3, 4), we notice that
both 22 and 33 have two maximal representations, but only one in Γ.
3
The Main Theorem
In this section, we apply the results contained in the previous section in order
to give a characterization of the numerical semigroup rings whose associated
graded ring is Complete Intersection.
We recall that in the first section we defined R = R/(tm ) and G = grm (R),
where m is the maximal ideal of R.
We also need to introduce two more invariants associated to R. The ideal
Q = (tm ) is a principal reduction of the maximal ideal m, that is a principal
ideal Q ⊆ m such that Qmh = mh+1 for some non-negative integer h. The reduction number is the integer r = rQ (m) = min{h ∈ N, Qmh = mh+1 }, while
15
the index of nilpotency is defined as s = sQ (m) = min{h ∈ N, mh+1 ⊆ Q}.
In the case of numerical semigroup rings we have r = min{h ∈ N, m + hM =
(h + 1)M} and s = max{ord(ωi ) | ωi ∈ Ap(S)}; from the last two equalities
it is easy to see that s ≤ r.
We will also need the following result of Bryant.
Theorem 3.1 ([3],Theorem 3.14). Under the above notation, we have:
(1) S is M-pure symmetric if and only if G is Gorenstein;
(2) if grm (R) is Cohen-Macaulay, then s = r. The converse holds if S is
M-pure;
(3) grm (R) is Gorenstein if and only if S is M-pure symmetric and s = r.
To prove the main result of the paper we have to be more precise about
the generators of the ideal J defined in Discussion 1.5.
Discussion 3.2. Using the terminology introduced in the previous section, in
Discussion 1.5 we have shown that J is generated by all the the binomials of
the form
(+)
xj22 · · · · · xjνν − xh2 2 · · · · · xhν ν ,
where j2 g2 + · · · + jν gν = h2 g2 + · · · + hν gν ∈ Ap(S) are two maximal representations, and by all the monomials of the form
(++)
xj22 · · · · · xjνν ,
/ Ap(S) or j2 g2 + · · · + jν gν ∈ Ap(S) and it is
where either j2 g2 + · · · + jν gν ∈
not a maximal representation.
By definition of βi it follows that xβi i +1 ∈ J. In fact, if ord(βi + 1)gi >
βi + 1, then (tgi )βi +1 ∈ mβi +2 and therefore (tgi )βi +1 = 0 in grm (R) and hence
in G. On the other hand, if (βi + 1)gi ∈
/ Ap(S), then (tgi )βi +1 = 0 in R and
hence in G. Moreover, by definition of βi and by Discussion 1.5 it is clear
that xβi i ∈
/ J for every index i.
On the other hand, by definition of γi we have that, γi < βi if and only if
(γi + 1)gi ∈ Ap(S) is a maximal representation, but it is not unique. Hence
P
P
Q
λ
(γi +1)gi = j6=i λj gj and γi +1 = j6=i λj ; equivalently xiγi +1 − j6=i xj j ∈ J.
Notice that, for some h ≤ βi (hence, also for some h ≤ γi ), it could
P
P
Q
λ
happen that hgi = j6=i λj gj and h > j6=i λj ; in this case, j6=i xj j ∈ J.
16
Hence the smallest pure power of xi appearing in a monomial or a binomial
of J is xiγi +1 .
Finally, it is clear that, if a binomial (+) is in J, then we can cancel all
the common factors: in fact, by Proposition 1.1, if ω ∈ Ap(S) and u ∈ S is
such that u ω, then u ∈ Ap(S).
In the next corollary we summarize the results of Discussions 1.5 and 3.2,
that we will need in the rest of the paper.
Corollary 3.3. For every i = 2, . . . , ν, we have:
(i) xβi i +1 ∈ J and xβi i ∈
/ J;
(ii) γi < βi ⇐⇒ (γi + 1)gi =
Q
λ
xiγi +1 − j6=i xj j ∈ J;
P
j6=i
λj gj and γi + 1 =
P
j6=i
λj
⇐⇒
(iii) the smallest pure power of xi appearing in a monomial or a binomial
of J is xiγi +1 .
Pν
Pν
(iv) xλ2 2 · · · ·· xλν ν ∈
/ J ⇐⇒
j=2 λj gj is a maximal
j=2 λj gj ∈ Ap(S) and
representation.
The next result is a key step in order to prove the main theorem.
Lemma 3.4. The ring G is Complete Intersection if and only if the defining
ideal J is of the following form
Y λ
xj j : i = 2 . . . , ν),
J = (xγi i +1 − ρi
j6=i
where
ρi = 0, if βi = γP
i and 1 otherwise; as soon as ρi = 1, (γi + 1)gi =
P
λ
g
and
γ
+
1
=
i
j6=i j j
j6=i λj .
Proof. We have that G ∼
= k[x2 , x3 , . . . , xν ]/J; hence, if J is of the form
described in the statement, µ(J) = ν − 1 and G is Complete Intersection.
Conversely, assume that G is Complete Intersection, that is µ(J) = ν − 1.
By Corollary 3.3 (i), we know that (xβ2 2 +1 , . . . , xβν ν +1 ) ⊆ J and that xβi i ∈
/J
for every index i.
Hence, if J ) (xβ2 2 +1 , . . . , xβν ν +1 ), for every index i such that xβi i +1 is not a
Q
h
minimal generator, there exists a unique binomial of the form xhi − j6=i xj j ,
which is a minimal generator. By Corollary 3.3 (ii) and (iii), we have that
17
Q
P
λ
γi < βi and the generator is xiγi +1 − j6=i xi j with γi + 1 = j6=i λj , since
there is not any binomial in J involving a pure power of xi with exponent
smaller than γi + 1.
Q
λ
Remarks 3.5. (1) Let J˜ = (xiγi +1 − ρi j6=i xj j : i = 2 . . . , ν) (with the same
notation of the previous lemma); by Corollary 3.3, it is clear that we always
˜
have the inclusion J ⊇ J.
(2) We can assume that in the binomials appearing as generators of J˜ every
exponent λj is less than or equal to γj : choose the biggest (ν − 2)-tuple
(λ2 , . . . λi−1 , λi+1 , . . . λν ) with respect to lex, among all the possible (ν − 2)tuples corresponding to the maximal representations of (γi + 1)gi and then
argue in a similar way as in Lemma 2.6.
(3) Since g2 < g3 < · · · < gν , necessarily ρ2 = ρν = 0, that is xγ22 +1 and xγνν +1
˜
are minimal generators of J.
Theorem 3.6. The following conditions are equivalent:
(i) grm (R) is Complete Intersection;
(ii) Ap(S) is γ-rectangular and grm (R) is Cohen-Macaulay;
(iii) Ap(S) is γ-rectangular and r = s;
(iv) Ap(S) is γ-rectangular and grm (R) is Buchsbaum;
(v) Ap(S) is γ-rectangular and grm (R) is Gorenstein.
Proof. (i) ⇒ (ii) Clearly grm (R) is Cohen-Macaulay. By Remark 1.3 we know
that G is Complete Intersection. Using the previous lemma, we know that J
is generated by ν − 1 forms of degree γi + 1, for every
Qν i = 2, . . . , ν.
∼
Since G = k[x2 , . . . , xν ]/J, we have dimk (G) = i=2 (γi + 1), hence every
monomial xλ2 2 . . . xλν ν , with λi ≤ γi , does not belong to J and their images are
pairwise different in G. By Corollary 3.3 (iv), the corresponding elements
λ2 g2 +· · ·+λν gν of S belong to Ap(S) (and all of these are maximal representations). Hence we obtain Γ = {λ2 g2 + λ3 g3 + · · · + λν gν | λi = 0, . . . , γi , i =
2, . . . , ν} = Ap(S), that is Ap(S) is γ-rectangular.
(ii) ⇒ (i) grm (R) is Cohen-Macaulay and, by [10, Theorem 7], (tm )∗ is
regular. Hence, by Remark 1.3, G is Complete Intersection if and only if G
is Complete Intersection.
18
We know that, as k-vector space, G = htωi | ωi ∈ Ap(S)ik . Moreover, by
˜ hence
Remarks 3.5 (1), it is clear that J ⊇ J,
˜ =
m = |Ap(S)| = dimk (G) ≤ dimk (k[x2 , . . . , xν ]/J)
ν
Y
(γi + 1) .
i=2
Qν
Since Ap(S) is γ-rectangular, by Theorem 2.22, m = i=2 (γi + 1); thus in
˜ that is G is
the above chain we have all equalities and, therefore, J = J,
Complete Intersection.
(ii) ⇔ (iii) By Corollary 2.21, S is M-pure; under this hypothesis G is
Cohen-Macaulay if and only if r = s (Theorem 3.1).
(ii) ⇔ (iv) By [6, Proposition 5.5].
(iv) ⇔ (v) By [6, Corollary 5.6].
Remark 3.7. Using the last theorem, in order to know if grm (R) is Complete Intersection we have to check if Ap(S) is γ-rectangular and if grm (R)
is Cohen-Macaulay (or Buchsbaum, or Gorenstein). In [2, Theorem 2.6]
there is a characterization of the Cohen-Macaulayness of grm (R) that has
been strengthened in [6, Proposition 5.1]; in particular, in case Ap(S) is γrectangular, one has to compute the integers ai and bi defined in these characterizations, only for ωi = f + m, which is the only element in maxApM (S).
Notice also that, in this case, to check the Buchsbaumness of grm (R) it is not
easier than to check the Cohen-Macaulayness (cf. [6, Proposition 3.6]).
Examples 3.8. Let us consider the semigroups (2) and (3) of the Examples
2.12.
(2) S = h8, 10, 11, 12i: here Ap(S) is γ-rectangular (and not β-rectangular)
and grm (R) is Cohen-Macaulay, since r = 3 = ord(33). Hence grm (R) is Complete Intersection. Computing the defining ideals we get: I = (x22 −x1 x4 , x23 −
x2 x4 , x31 − x24 ), I ∗ = (x22 − x1 x4 , x23 − x2 x4 , x24 ) and J = (x22 , x23 − x2 x4 , x24 ).
(3) S = h5, 6, 9i: here Ap(S) is not γ-rectangular and grm (R) is CohenMacaulay (as can be checked using [6, Proposition 5.1]); S is symmetric,
but not M-pure (since 9 6M 18). Therefore, grm (R) is not Complete Intersection (nor Gorenstein). Computing he defining ideals we obtain: I =
(x31 − x2 x3 , x32 − x23 ), I ∗ = (x2 x3 , x23 , x42 − x31 x3 ) and J = (x23 , x2 x3 , x42 ).
Remark 3.9. The proof of equivalence (i) ⇔ (ii) of Theorem 3.6 shows that
G = grm (R) is Complete Intersection if and only if Ap(S) is γ-rectangular.
19
If we want to extend the result to grm (R) the hypothesis of the CohenMacaulayness is indispensable, that is
Ap(S) is γ − rectangular 6⇒ grm (R) is Complete Intersection.
Indeed, let S = h6, 7, 15i. Here Ap(S) is β-rectangular (hence γ-rectangular),
since f = 23, β2 = 2, β3 = 1 and f + m = β2 g2 + β3 g3 . It follows that G
is Complete Intersection and J = (x23 , x32 ). However, in this case, G is not
Complete Intersection, since grm (R) is not Cohen-Macaulay, because s = 3
and r = 6. It is not difficult to compute that I = (x23 − x51 , x32 − x1 x3 ) (S
is symmetric, hence R is Gorenstein that, if the embedding dimension is 3,
implies Complete Intersection) and that I ∗ = (x23 , x1 x3 , x62 ).
Remark 3.10. We note that:
grm (R) is Gorenstein and R is Complete Intersection 6⇒
grm (R) is Complete Intersection.
In fact, let S = h16, 18, 21, 27i; we have Ap(S) = {0, 18, 21, 27, 36, 39, 42, 45,
54, 57, 60, 63, 72, 78, 81, 99}. Let us compute the integer γ3 : 2 · 21 = 42 ∈
Ap(S) and it has a unique representation, while 3 · 21 = 2 · 18 + 27 are two
maximal representations of 63; hence γ3 = 2. The multiplicity of S is m = 16
and γ3 + 1 = 3 does not divide m; therefore, by Theorem 2.22, Ap(S) is not
γ-rectangular (hence grm (R) is not Complete Intersection).
On the other hand, it is not difficult to check that S is symmetric, M-pure
and that ord(99) = 5 = r; hence, by Theorem 3.1 (3), grm (R) is Gorenstein.
Finally, R is Complete Intersection, since I = (x31 − x3 x4 , x32 − x24 , x22 x4 − x33 ).
Remark 3.11. If we substitute the condition “γ-rectangular Apéry set” with
the condition “β-rectangular Apéry set” in the previous Theorem 3.6 we
obtain not only that grm (R) is Complete Intersection, but also that J is
monomial. However this fact does not implies that the defining ideal I ∗
of grm (R) is monomial. For example, let us consider the semigroup of the
Example 2.17: S = h12, 14, 16, 23i; its Apéry set is β-rectangular and r =
s = 4, hence grm (R) is Complete Intersection. In this case the defining ideals
are the following: I = (x22 − x1 x3 , x24 − x2 x23 , x33 − x41 ), I ∗ = (x22 − x1 x3 , x24 , x33 )
and J = (x22 , x33 , x24 ).
As a corollary to Theorem 3.6, we get the following corollary, that can
be also obtained easily by Bézout’s theorem applied to the algebraic variety
defined by G. Given a positive integer x, we call ℓ(x) the length of its unique
factorization i.e. the number of (possibly equal) prime factors of x.
20
Corollary 3.12. Let grm (R) be Complete Intersection, then we have:
(1) ν ≤ ℓ(m) + 1;
(2) if m is a prime number then ν = 2.
Proof. (1) Q
By Theorem 3.6, Ap(S) is γ-rectangular. Thus we have m =
|Ap(S)| = νi=2 (γi + 1) and the thesis follows as γi ≥ 1.
(2) If m is prime, we get ν ≤ ℓ(m) + 1 = 2 by (1); but ν > 1, otherwise
m = 1.
Using our method, we also obtain an alternative proof of a result from
[2]:
Corollary 3.13.
Q Let nν < · · · < n1 be pairwise relatively prime positive
integers, N = νi=1 ni . Let gi = nNi and S = hg1 , g2, . . . , gν i. Then grm (R) is
Complete Intersection.
Proof. As proved in [2, Proposition 3.6], grm (R) is Cohen-Macaulay; we only
need toQprove that Ap(S)
Qνis γ-rectangular. We easily have γi ≤ ni − 1 and
ν
hence
= g1 = m. Since m = |Ap(S)| ≤ |Γ| ≤
i=2 (γi + 1) ≤
i=2 ni Q
Qν
ν
i=2 (γi + 1) holds and, by Theorem 2.22, we
i=2 (γi + 1), the equality m =
get the thesis.
We finish the paper studying the case ν(S) = 3. By Remarks 3.5 (3),
it follows immediately that Ap(S) is γ-rectangular if and only if it is βrectangular. However, we can prove something more.
Theorem 3.14. Let S = hg1 , g2, g3 i be a three-generated semigroup (with
g1 < g2 < g3 ). The following conditions are equivalent:
(i) Ap(S) is β-rectangular;
(ii) Ap(S) is γ-rectangular;
(iii) S is M-pure symmetric.
Proof. By Corollary 2.10 and Corollary 2.21 we only need to show the implication (iii) ⇒ (i). Assume that S is M-pure symmetric, that is ω M f + m
for each ω ∈ Ap(S). If we have two maximal representations of f + m:
f + m = λ2 g 2 + λ3 g 3 = µ 2 g 2 + µ 3 g 3 ,
21
with λ2 + λ3 = µ2 + µ3
then it follows that
(λ2 − µ2 )g2 = (µ3 − λ3 )g3 ,
with λ2 − µ2 = µ3 − λ3
and, since g2 < g3 , λ2 − µ2 = λ3 − µ3 = 0. Thus f + m has a unique maximal
representation and the thesis follows by Theorem 2.16 (iii).
As a consequence of the last result, we get the well known fact that
Goresteinness and Complete Intersection are equivalent in codimension two
(i.e., in our hypotheses, in embedding dimension three).
Corollary 3.15. Let R be a numerical semigroup ring with ν(R) = 3. If
grm (R) is Gorenstein, then it is Complete Intersection (and the ideal J is
monomial).
Proof. Apply Theorems 3.1, 3.14 and 3.6.
Acknowledgments. The authors are grateful to Yi Huang Shen for
some helpful remarks on a preliminary version of the paper that allowed to
define properly the γ-rectangularity and to the referee for several suggestions
and comments that improved the paper.
References
[1] V. Barucci, M. D’Anna, R. Fröberg, On plane algebroid curves, Commutative Ring Theory and Applications, Dekker L.N. in Pure and Applied
Mathematics, Vol. 231, (2003), 37-50.
[2] V. Barucci, R. Fröberg, Associated graded rings of one dimensional analytically irreducible rings, J. Algebra 304 (2006) n.1, 349-358.
[3] L. Bryant, Goto Numbers of a Numerical Semigroup Ring and the
Gorensteiness of Associated Graded Rings, Comm. Alg. 38 n.6 (2010),
2092 - 2128.
[4] L. Bryant, J. Hamblin, L. Jones Maximal Denumerant of a Numerical
Semigroup with Embedding Dimension Less Than Four, to appear in
Journal of Commutative Algebra.
22
[5] M. D’Anna, M. Mezzasalma, V. Micale, On the Buchsbaumness of the
associated graded ring of a one-dimensional local ring, Comm. Alg. 37
n.5 (2009), 1594 - 1603.
[6] M. D’Anna, V. Micale, A. Sammartano, On the Associated Graded Ring
of a Semigroup Ring, Journal of Commutative Algebra 3 n.2 (2011), 147
- 168.
[7] M. Delgado, P. A. Garcia-Sanchez, J. Morais, NumericalSgps - a GAP
package, 0.96 (2005).
[8] R. Fröberg, C. Gottlieb, R. Häggkvist, On Numerical Semigroups, Semigroup Forum, 35 (1987), 63-83.
[9] The GAP Group, GAP - Groups, Algorithms, and Programming, Version 4.4.10 (2007).
[10] A. Garcia, Cohen-Macaulayness of the Associated Graded Ring of a
Semigroup Ring, Comm. Algebra 10 (1982), 393-415.
[11] S. Goto, Buchsbaum rings of maximal embedding dimension, J. Algebra
76 (1982), 383-399.
[12] S. Goto, Noetherian local rings with Buchsbaum associated graded rings,
J. Algebra 86 (1984), 336-384.
[13] W. Heinzer, M. Kim, B. Ulrich, The Gorenstein and complete intersection properties Of associated graded rings, J. Pure Appl. Algebra 201
(2005), 264 - 283.
[14] J. Herzog, Generators and relations of abelian semigroup rings,
Manuscripta Math. 3 (1970), 748-751.
[15] S. Molinelli, D. P. Patil, G. Tamone On the Cohen-Macaulayness of the
Associated Graded Ring of Certain Monomial Curves, Contributions to
Algebra and Geometry 39 (1998), 433-446.
[16] J. C. Rosales, P. A. Garcı́a Sanchez, Numerical Semigroups, Springer
(2009).
[17] L. Robbiano, G. Valla, On the equations defining tangent cones,
Proc. Camb. Phil. Soc. 88 (1980), 281-297.
23
[18] J. D. Sally, On the associated graded ring of a local Cohen-Macaulay
ring, J. Math. Kyoto Univ. 17 n.1 (1977), 19-21.
[19] V. A. Sapko, Associated graded rings of numerical semigroup rings,
Comm. Algebra 29 n.10 (2001), 4759-4773.
[20] Y. H. Shen, Tangent cone of numerical semigroup rings with small embedding dimension, Comm. Algebra 39 n.5 (2011), 1922-1940.
[21] P. Valabrega, G. Valla, Form Rings and Regular Sequences,
Nagoya Math. J. 72 (1978), 93-101.
24
| 0math.AC
|
POWERS OF GENERIC IDEALS AND THE WEAK LEFSCHETZ
PROPERTY FOR POWERS OF SOME MONOMIAL COMPLETE
INTERSECTIONS
arXiv:1612.00411v1 [math.AC] 1 Dec 2016
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
Abstract. Given an ideal I = (f1 , . . . , fr ) in C[x1 , . . . , xn ] generated by forms of
degree d, and an integer k > 1, how large can the ideal I k be, i.e., how small can
the Hilbert function of C[x1 , . . . , xn ]/I k be? If r ≤ n the smallest Hilbert function is
achieved by any complete intersection, but for r > n, the question is in general very
hard to answer. We study the problem for r = n + 1, where the result is known for
k = 1. We also study a closely related problem, the Weak Lefschetz property, for S/I k ,
where I is the ideal generated by the d’th powers of the variables.
1. Introduction
Let I = (f1 , . . . , fr ) be an ideal generated by forms of degree d in C[x1 , . . . , xn ], and
let k ≥ 1. How large can the ideal I k be, i.e., how small can the Hilbert function of
C[x1 , . . . , xn ]/I k be? It is clear that we get the smallest Hilbert series if the fi ’s are
general. If r ≤ n, I is a complete intersection, and the Hilbert series (and even the
graded Betti numbers) for I k are known ([5]). But if r > n not much is known even if
k = 1, and we are not aware of any result for k > 1. For k = 1 the main classes for
which the result is known is when n ≤ 3 ([1],[4]) or when r = n + 1 ([8]). In all known
cases the series for an idealP
with r general generators of degree d ≥ 2 in n variables is
[(1 − td )r /(1 − t)n ], where [ i≥0 ai ti ] means truncate before the first nonpositive term.
A first guess for the Hilbert series of C[x1 , . . . , xn ]/I k , I generated by r generalforms of
general
degree d, could be that the series for k ≫ 1 and r > n is the same as for k+r−1
r−1
forms of degree dk. We will show that this is not always the case for geometric reasons.
We also study a closely related problem, the Weak Lefschetz property (WLP) for
powers of some monomial complete intersections ideals. Recall that a graded algebra
A satisfies the WLP if there exists a linear form L such that the multiplication map
×L : Ai → Ai+1 has maximal rank for all degrees i. See [7] for a survey on the WLP.
For a graded algebra R = ⊕Ri we let R(t) denote the Hilbert series of R.
2. Powers of ideals of general forms
Let f1 , f2 , . . . , fn+1 be general forms of degree d in C[x1 , . . . , xn ]. We will start by
showing that when k = dn−1 , the dimension of C[x1 , . . . , xn ]/(f1 , f2 , . . . , fn+1 )k in degree
dk is one less than expected.
Lemma 1. For general forms f1 , f2 , . . . , fn+1 of degree d in C[x1 , . . . , xn ] there is exactly
one relation of degree dn−1 in C[f1 , . . . , fn+1 ].
The first author was partially supported by the grant VR2013-4545.
1
2
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
Proof. General forms f1 , f2 , . . . , fn+1 give a well-defined map Φ : Pn−1 −→ Pn since they
have no common zeroes. The closure of the image is a variety of dimension n−1, so it is a
hypersurface. The inverse image of a general line gives a reduced complete intersection,
as we can see from producing one such example by choosing the forms to be powers of
general linear forms and by considering a line that gives the inverse image given by the
radical complete intersection ideal (f1 − fn , f2 − fn , . . . , fn−1 − fn ). Thus a general line
meets the hypersurface in dn−1 distinct points and hence the image is not contained in
any hypersurface of lower degree than dn−1 . We conclude that there is a unique equation
of degree dn−1 defining the image of Φ .
In the next two subsections, we will focus on the situation in two and three variables.
We will need the following lemma.
Lemma 2. Let I be an ideal in C[x1 , . . . , xn ] generated by homogeneous elements in
degree d and suppose that xdi ∈ I for i = 1, . . . , n. Suppose that C[x1 , . . . , xn ]/I k is zero
in degree a and that a ≥ d(n − 1). Then C[x1 , . . . , xn ]/I k+1 is zero in degree a + d.
Proof. Let R = C[x1 , . . . , xn ] and write R = R0 ⊕ R1 ⊕ · · · . By the assumption, Ra ⊂ I k ,
so Ra xdi ⊂ I k+1 for i = 1, . . . , n. Let m be a monomial in Ra+d . Since a + d ≥ dn, there
is at least one variable, say xj , whose exponent in m is greater than or equal to d. But
then m ∈ Ra xdj , that is, m ∈ I k+1 .
2.1. The case n = 2, r = 3. Let R2,d,k = C[x, y]/(f1 , f2 , f3 )k , where the fi ’s are general
forms of degree d. If d = 2, then (f1 , f2 , f3 ) = (x, y)2 , so R2,2,k = C[x, y]/(x, y)2k . We
assume in this section that d ≥ 3.
P
d−1 dk
i
Lemma 3. If k ≥ d − 2, then R2,d,k (t) ≥ dk−1
coefficientwise.
i=0 (i + 1)t +
2 t
Proof. Consider
the case
when k = d. By Lemma 1 the dimension of R2,d,k in degree dk
d+2
k−d+2
2
is d + 1 −
− 1 = d−1
2
2 . This implies that in degree k ≥ d, there are
2
k+2
relations among the monomials of degree k in the fi ’s. We have 2 monomials of
degree
k in the fi ’s. Thus the dimension of R2,d,k in degree dk is at least dk + 1 −
k+2
= d−1
− k−d+2
2 .
2
2
Finally, let us consider the case when k = d − 1 and k = d − 2. The dimension of
R2,d,k in degree dk is at least dk + 1 − k+2
2 . If k = d − 1 or k = d − 2, this equals
d−1
2 .
Lemma
4. If k ≥ d − 2, the Hilbert function of C[x, y]/(xd , y d , xd−1 y)k in degree dk is
d−1
2 .
Pdk−1
i + d−1 tdk .
Proof. By Lemma 3, we have C[x, y]/(xd , y d , xd−1 y)k (t) ≥
(i
+
1)t
i=0
2
Thus it is enough to show that the Hilbert function of C[x, y]/(xd , y d , xd−1 y)k in degree
dk is at most d−1
2 .
Let h = min(k, d − 1). We construct h + 1 separate groups indexed by 0, . . . , h, where
we let the b’th group consist of the polynomials xda (xy d−1 )b y dc of degree dk such that
a + b ≤ k. The leading monomials in this group are
xb y dk−b , xb+d y dk−(b+d) , . . . , xb+(k−b)d y dk−(b+(k−b)d) ,
POWERS OF GENERIC IDEALS AND THE WLP FOR POWERS OF SOME MONOMIAL CI’S
3
so there are k − b + 1 leading monomials in group b.
P
In total this gives hb=0 (k − b + 1) = (h + 1)(k + 1) − h+1
elements. If k = d − 2,
2
d−1
then the dimension in degree dk is at most d(d − 2) + 1 − (d − 1)(d − 1) + d−1
2 =
2 .
d
d−1
2
If k = d − 1, then the dimension in degree dk is at most d(d − 1) + 1 − d + 2 = 2 .
Finally, if k ≥ d, the dimension in degree dk is at most kd+1−d(k +1)+ d2 = d−1
2 .
Lemma 5. The Hilbert function of the algebra C[x, y]/(xd , y d , xd−1 y + xy d−1 )k is zero
in degree dk + 1 if k ≥ max(d − 3, 1).
Proof. We will prove by induction that all the monomials of degree dk + 1 are in the
ideal J = (xd , y d , xd−1 y + xy d−1 )k . Assume by induction on ℓ that we have all the
monomials of the form xid+j y (k−i)d+1−j and x(k−i)d+1−j y id+j in J for 0 ≤ i ≤ k − j,
0 ≤ j ≤ ℓ + 1. Denote the ideal generated by these monomials by J (ℓ) . Observe that
J (0) = (x, y)(xd , y d )k ⊆ J by the defintion of J.
We compute for 0 ≤ i ≤ k − ℓ − 1 the following element in J
x · (xd−1 y + xy d−1 )ℓ+1 · xid y (k−i−ℓ−1)d =
ℓ+1
X
ℓ+1
j
j=0
x(i+j)d+ℓ+2−2j y (k−i−j)d−ℓ−1+2j
≡ xid+ℓ+2 y (k−i)d−ℓ−1
(mod J (ℓ) )
since all but the first term are in J (ℓ) . By symmetry we also get that the following is in
J
y · (x
d−1
y + xy
d−1 ℓ+1
)
·x
(k−i−ℓ−1)d id
y
=
ℓ+1
X
ℓ+1
j
j=0
x(k−i−j)d−ℓ−1+2j y (i+j)d+ℓ+2−2j
≡ x(k−i)d−ℓ−1 y id+ℓ+2
(mod J (ℓ) )
since all but the last term are in J (ℓ) . Hence we have proved that J (ℓ+1) ⊆ J for
0 ≤ ℓ ≤ k.
Now we need to prove that J (k) contains all monomials of degree dk+1. Any monomial
of degree dk + 1 can be written as xjd+r y (k−j)d−(r−1) , where 0 ≤ r < d. This monomial
is of the form xid+ℓ+2 y (k−i)d−ℓ−1 in J (k) if r = ℓ + 2 and j ≤ k − ℓ − 1 = k − r + 1, i.e.,
r ≤ k − j + 1 and it is of the form x(k−i)d−ℓ−1 y id+ℓ+2 in J (k) if jd + r = (k − i)d − ℓ − 1,
i.e., j = k − i − 1, r = d − ℓ − 1 and 0 ≤ i ≤ k − ℓ − 1, i.e., 0 ≤ k − j − 1 ≤ r + k − d, which
can be written as r ≥ d − j − 1 ≥ d − k. We will get all monomials in degree dk + 1 if
d − j − 1 ≤ (k − j + 1) + 1, i.e., when d − 3 ≤ k.
Theorem 6. If k ≥ d − 2, then
R2,d,k (t) =
dk−1
X
i=0
d − 1 dk
(i + 1)t +
t .
2
i
The Betti numbers of R2,d,k are β0,0 = 1, β1,dk =
β2,dk+2 = d−1
2 .
2dk−d2 +3d
,
2
β2,dk+1 = (dk − d2 + 3d − 2),
4
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
P
d−1 dk
i
Proof. By Lemma 4 and Lemma 5, R2,d,k (t) ≤ dk−1
i=0 (i + 1)t +
2 t . Together with
Lemma 3 we get the claimed Hilbert series.
P
i
i+j )/(1 − t)2 . We know that the number of
The Hilbert series equals ( dk+2
i=0 (−1) βi,jt
2
generators of (f1 , f2 , f3 )k are k+2
− k−d+2
= 2dk−d2 +3d , and they are all of degree dk,
2
2
2
so β1 = β1,dk = 2dk−d2 +3d . Also β2,dk+2 = d−1
and β2,i = 0 if i > dk + 2 since R2,d,k
2
d−1
has socle of dimension 2 in degree dk, and no socle in higher degree.
Remark 7. We are convinced that also the algebra C[x, y]/(xd , y d , (x + y)d )k , is zero in
degree dk + 1 for any d and k ≥ max(d − 3, 1), but we have not been able to prove this.
See also Conjecture 1.
2.2. The case n = 3, r = 4.
Theorem 8. The Hilbert series of R3,2,k is
2k−1
X i + 2
ti + (3k − 1)t2k .
2
i=0
Proof. A similar calculation as in the proof of Lemma 3 shows that R3,2,k has at least
dimension 2k+2
− k+3
+ k−4+3
= 3k − 1 in degree 2k. Thus we have an inequality.
2
3
3
That there is equality follows from the following lemma.
Lemma 9. The Hilbert series of
C[x, y, z]/(x2 , y 2 , z 2 , (x + y + z)2 )k = C[x, y, z]/(x2 , y 2 , z 2 , xy + xz + yz)k
is
2k−1
X
i=0
i+2 i
t + (3k − 1)t2k .
2
Proof. It is enough to show that the Hilbert series has at most dimension 3k −1 in degree
2k and that the algebra is zero in degree 2k + 1.
We first consider degree 2k. We claim that all monomials of degree 2k except the
3k − 1 monomials
x2k−1 z, x2k−3 z 3 , . . . , xz 2k−1 , y 2k−1 z, y 2k−3 z 3 , . . . , yz 2k−1 , xy 2k−2 z, xy 2k−4 z 3 , . . . , xy 2 z 2k−3
occur as leading monomials in lexicographic ordering. We have (x2 , y 2 , z 2 , (x+y +z)2 ) =
(x2 , y 2 , z 2 , f ), where f = xy + xz + yz. If m = xa y b z c is a monomial of degree 2k, we let
t(m) = (a (mod 2), b (mod 2), c (mod 2)). If t(m) = (0, 0, 0), then m lies in the ideal.
If t(m) = (1, 1, 0), then m = (xy)min(a,b) M , where t(M ) = (0, 0, 0), so m is the leading
monomial of f min(a,b) M . If t(m) = (1, 0, 1) we claim that m is a leading monomial
except when b = 0 or a = 1. If t(m) = (0, 1, 1) we claim that m is a leading monomial
except when a = 0. To prove the claims, it suffices to show that all monomials of the
form x2 yzM , t(M ) = (0, 0, 0) and all monomials of the form x3 y 2 zM , t(M ) = (0, 0, 0)
are leading. Now x2 yz is the leading monomial of f 2 − x2 y 2 , and x3 y 2 z is the leading
monomial of f (f 2 − x2 y 2 − x2 z 2 ).
We now consider degree 2k + 1. A calculation shows that C[x, y, z]/(x2 , y 2 , z 2 , xy +
xz + yz) is zero in degree three and that C[x, y, z]/(x2 , y 2 , z 2 , xy + xz + yz)2 is zero in
POWERS OF GENERIC IDEALS AND THE WLP FOR POWERS OF SOME MONOMIAL CI’S
5
degree five. Since 2k + 1 ≥ 2(n − 1) when k ≥ 2, the remaining cases follow from Lemma
2.
P3k−1 i+2 i
3k
Theorem 10. The Hilbert series of R3,3,k is i=0
when 9 ≤
2 t + (27k − 56)t
k+3
k ≤ 40 and [(1 − t3k )( 3 ) /(1 − t)3 ] when k < 9.
Proof. Consider first the case k ≥ 9. By Lemma 1 and a similar
calculation
as in the proof
3k+2
k+3
k−6
of Lemma 3, we get that R3,3,k has at least dimension 2 − 3 + 3 = 27k − 56
in degree 3k.
Next, consider the case k < 9. If the k+3
generators were generic, the series would
3
k+3
be [(1 − t3k )( 3 ) /(1 − t)3 ], so the series is a lower bound.
To get an upper bound, it is enough to find an example. According to Macaulay2, the
ideal (x3 , y 3 , z 3 , x2 y +11xy 2 −50x2 z +48xyz −29y 2 z −9xz 2 +30yz 2 ) ⊂ Z/101Z[x1 , x2 , x3 ]
has the desired property for k ≤ 40.
P3k−1 i+2 i
3k in general,
Remark 11. We are convinced that R3,3,k (t) = i=0
2 t + (27k − 56)t
cf. Conjecture 1.
2.3. The general case. We believe that some of the results have generalizations to
n−1
any n. By Lemma 1, there is unique relation among the d n +n generators of I k ,
n−1
n +n−1
≥ d n +n − 1. To show that this relation is not trivial, we need to
that is, d n−1
n−1
show that there is room for d n +n generators of degree dn in C[x1 , . . . , xn ]. When
(d, n) = (2, 2), we have five generators but only room for four. When (d, n) = (3, 2), we
have ten generators and room for ten. In all other cases, we have a strict inequality.
n−1
n +n−1
> d n +n .
Lemma 12. Let d, n ≥ 2, (d, n) 6= (2, 2), (3, 2). Then d n−1
Proof. We need to show that
n(dn + n − 1) · · · (dn + 1) > (dn−1 + n) · · · (dn−1 + 1).
We have (dn + n − 1) · · · (dn + 1) > (dn + 1)n−1 and (dn−1 + n) · (dn−1 + n − 1)n−1 >
(dn−1 + n) · · · (dn−1 + 1). Thus we are done if we can show that
n−1
dn + 1
> dn−1 + n.
n·
dn−1 + n − 1
n−1
d·(dn−1 +n−1)+1−dn+d
n
n
dn +1
=
=
>
d(1
−
).
We
have
1
−
Now dn−1
n−1
n−1
n−1
+n−1
d
+n−1
d
d
n−1
n−1 n
n−1 n2
n−1 n3
n−1 na
· · and n−1−a (dn−1 )a −
n−1 − n−1−1 dn−1 + n−1−2 (dn−1 )2 − n−1−3 (dn−1 )3 + ·
na
a
a+1
n−1
n−1
n−1
n
n−a−1
n2
n
n
=
1
−
>
1
−
.
·
n−1
a+1
n−1
a
n−1
n−1
a
n−1
n−1−a (d
n−1−a (d
n−1−(a+1) (d
a+1
)
)
d
)
d
2
n
The inequality 1 − dn−1
≥ 0 ⇔ dn−1 ≥ n2 holds true for d ≥ 3, n ≥ 3 and for
n−1
n
n2
d = 2, n ≥ 7. In these cases, we get the inequality 1 − dn−1
> 1 − dn−1
, so it is
then enough to show that
ndn−1 (1 −
2
n2
) > dn−1 + n.
dn−1
n
We have ndn−1 (1 − dn−1
) > dn−1 + n ⇔ dn−1 >
d ≥ 2, n ≥ 7 and n ≥ 3, d ≥ 3.
n3 +n
n−1 .
This inequality holds true for
6
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
We are left with a few special cases only. In the n = 2 case, we have to check the
2 +1
> d + 2 ⇔ d2 > 3d, which holds when d ≥ 4.
inequality 2 · dd+1
The remaining cases are covered by the conditions 3 ≤ n ≤ 6, d = 2 and then the
n +n−1
n−1
inequality d n−1
> d n +n is checked to be true by hand.
dk+n−1
By Lemma 1 and 12, the dimension of Rn,d,k in degree dk is at least
−
n−1
k+n
k−dn−1 +n
n−1
n−1
+
when k ≥ d
, with equality when k = d
. We conjecture that
n
n
the algebra is zero in higher degrees.
Conjecture 1. If k ≥ dn−1 and (d, n) 6= (2, 2), the Hilbert series of the algebra
C[x1 , . . . , xn ]/(f1 , . . . , fn+1 )k , fi general of degree d, is
dk−1
X i + n − 1
dk + n − 1
k+n
k − dn−1 + n
ti +
−
+
tdk .
n−1
n−1
n
n
i=0
Remark 13. Theorem 6 shows that Conjecture 1 holds in the the case n = 2. Theorem
8 shows that Conjecture 1 holds in the the case n = 3, d = 2, while Theorem 10 shows
that the conjecture holds when n = 3, d = 3, k ≤ 40.
We now look at the case when k < dn−1 . There are obvious relations between the
generators of (f1 , . . . , fn+1 )k of type f1 · f2k = f2 · (f1 f2k−1 ). These are of degree d(k + 1).
Now, if the algebra C[x1 , . . . , xn ]/(f1 , . . . , fn+1 )k is zero in degree d(k+1), these relations
do not show up. This happens if
d+n−1 k+n
Sn,d,k =
−
n−1
n
k+n
k+1+n
d(k + 1) + n − 1
(n + 1) −
−
≥ 0.
n
n
n−1
Conjecture 2. Let k < dn−1 and suppose that for Rd,k,n = C[x1 , . . . , xn ]/(f1 , . . . , fn+1 )k
we have Sn,d,k ≥ 0. Then the Hilbert series of Rd,k,n equals the Hilbert series of
general forms of degree dk.
C[x1 , . . . , xn ]/I, where I is an ideal generated by k+n
n
It is known that C[x1 , . . . , xn ]/(f1 , . . . , fn+1 ) has the same Hilbert series whether the
fi ’s are general forms of degree d, or d’th powers of general linear forms [8]. In the second
case we can assume that (f1 , . . . , fn+1 ) = (xd1 , . . . , xdn , (x1 + · · · + xn )d ). We will end this
section by giving an explicit relation of degree dn in C[xd1 , . . . , xdn , (x1 + · · · + xn )d ].
n−1
we have
Theorem 14. In T = C[x1 , . . . , xn+1 ]/(xd1 , . . . , xdn+1 )d
Y
F =
(x1 + ǫ1 x2 + · · · + ǫn−1 xn + ǫn xn+1 ) = 0,
where the ǫi ’s vary over all dn combinations of a d’th root of unity. In particular,
F/(x1 + x2 + · · · + xn+1 ) is a form of degree dn − 1 which is in the kernel of the map
×(x1 + x2 + · · · + xn+1 ) : T → T .
Proof. Let Gd ⊆ C∗ be the group of d’th roots of unity. The form F is invariant under the
, acting with multiplication on the variables, so it has to be a polynomial
action of Gn+1
d
in xd1 , xd2 , . . . , xdn+1 , which means that F = 0 in T .
POWERS OF GENERIC IDEALS AND THE WLP FOR POWERS OF SOME MONOMIAL CI’S
7
Let us now consider the form F as an element in C[x1 , . . . , xn+1 ]. When n + 1 = 2
and d = 2, F = (x1 + x2 )(x1 − x2 ) = x21 − x22 — the conjugate rule, and for general d,
F = xd1 − (−1)d xd2 .
When n + 1 ≥ 3, the form F is symmetric. To show this, notice first the the form is
invariant under permutation of the variables x2 , x3 , . . . , xn+1 . Thus it is enough to show
that F is invariant with respect to the transposition (12). Let G denote the
Q result iafter
letting (12) act on F . Let ǫ be a primitive
d’th
root
of
unity
and
let
f
=
(x1 + ǫ x2 +
i
Q
Q
ǫ2 x3 +· · ·+ǫn−1 xn +ǫn xn+1 ), so that F = i fi . Let gi = (ǫi1 x1 +x2 +ǫ2 x3 +· · ·+ǫn−1 xn +
Q
ǫn xn+1 ). SinceQn + 1 ≥Q3, we have fi = ǫd−i (x1 + ǫi x2 + ǫ2 x3 + · · · + ǫn−1 xn + ǫn xn+1 ) =
gd−i , so F = i fi = i gi = G.
Example 15. In general, it seems hard to get an explicit description of F . We can get
it for d = 2, n + 1 = 3, 4, 5 and for n + 1 = d = 3. It is the symmetrization of the
following polynomials.
(n + 1, d)
(3, 2)
x4 + x2 y 2
(4, 2)
x8 − 4x6 y 2 + 6x4 y 4 + 4x4 y 2 z 2
(5, 2)
x16 − 8x14 y 2 + 28x12 y 4 + 40x12 y 2 z 2 − 56x10 y 6 − 72x10 y 4 z 2
−176x10 y 2 z 2 u2 + 70x8 y 8 + 40x8 y 6 z 2 + 36x8 y 4 z 4 + 344x8 y 4 z 2 u2
−757x8 y 2 z 2 u2 v 2 + 16x6 y 6 z 4 − 416x6 y 6 z 2 u2 − 272x6 y 4 z 4 u2
+928x6 y 4 z 2 u2 v 2 + 2008x4 y 4 z 4 u4 − 1520x4 y 4 z 4 u2 v 2
(3, 3)
x27 + 36x24 y 3 − 9x21 y 3 z 3 + 684x18 y 9 − 234x18 y 6 z 3 + 3339x18 y 3 z 3 u3
+126x15 y 12 − 711x15 y 9 z 3 + 513x15 y 6 z 6 + 1512x15 y 6 z 3 u3 − 990x12 y 12 z 3
+2961x12 y 9 z 6 − 12222x12 y 6 z 6 u3 + 278371x12 y 6 z 6 u3 − 12171x9 y 9 z 9
−6867x9 y 9 z 6 u3 + 120312x9 y 6 z 6 u6
3. The Weak Lefschetz property
We now turn to the WLP for Tn,d,k := C[x1 , . . . , xn ]/(xd1 , . . . , xdn )k . The algebra T2,d,k
has the WLP, since quotients of C[x1 , x2 ] has the Strong Lefschetz property ([6]), which
implies the WLP. When k = 1, the algebra is of the form C[x1 , . . . , xn ]/(xd1 , . . . , xdn ),
which is a monomial complete intersection, and therefore also Tn,d,1 has the Strong
Lefschetz property ([8]). When d = 1, we have Tn,1,k = C[x1 , . . . , xn ]/(x1 , . . . , xn )k .
Since C[x1 , . . . , xn ] has the WLP, so has C[x1 , . . . , xn ]/(x1 , . . . , xn )k . When n ≥ 3 and
d, k ≥ 2, the situation is more involved.
The WLP for algebras given by monomial ideals have been studied before, for example
in [3] and [2] but not for our kind of monomial ideal.
Since Tn,d,k is a monomial algebra, (C∗ )n acts on Tn,d,k and therefore any general
linear form can be identified with L = x1 + · · · + xn . When taking the quotient by
the ideal (L), we get T̃n,d,k := C[x1 , . . . , xn−1 ]/(xd1 , . . . , xdn−1 , (x1 + · · · + xn−1 )d )k . Thus
Tn,d,k has the WLP if and only if [(1 − t)Tn,d,k (t)] = T̃n,d,k (t).
8
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
From Lemma 12 and Theorem 14 we can immediately get a negative result on the
WLP.
Theorem 16. Suppose that k ≥ dn−2 , n ≥ 3, (n, d) 6= (3, 2). Then Tn,d,k fails the WLP.
Proof. By Lemma 12, T̃n,d,k is non-zero in degree dk. Thus it is enough to show that
the map ×(x1 + x2 + · · · + xn ) : (Tn,d,k )dk−1 → (Tn,d,kQ
)dk is not injective.
n−2
Suppose that k = d
. By Theorem 14, the form ( (x1 + ǫ1 x2 + · · · + ǫn−1 xn ))/(x1 +
x2 + · · · + xn ) of degree dk − 1 is in the kernel of the map. Hence it cannot be injective.
n−2
(F/(x1 + x2 + · · · + xn )) is in the
Suppose instead that k > dn−2 . Then xk−d
1
kernel.
3.1. The case n = 3. We have a conjecture on the WLP for T3,d,k .
Conjecture 3. The algebra C[x, y, z]/(xd , y d , z d )k has the WLP if and only if one of
the following conditions is satisfied.
(1) d ≤ 2,
(2) k ≤ 2,
(3) d > k = 2j + 1 ∈ {3, 7}, d 6= (j + 2)(2ℓ + 1),
(4) d > k = 2j + 1 > 2, k ∈
/ {3, 7}, d ∈
/ {(j + 2)(2ℓ + 1) − 1, (j + 2)(2ℓ + 1), (j +
2)(2ℓ + 1) + 1},
(5) d > k = 2j > 2, d ∈
/ {(j + 1)(2ℓ + 1) + ℓ, (j + 1)(2ℓ + 1) + ℓ + 1}.
We have already proven that the algebra T3,d,k fails the WLP when 3 ≤ d ≤ k and
that T3,d,1 and T3,1,k have the WLP. We will now prove Conjecture 3 in some more cases.
Theorem 17. The algebra T3,2,k has the WLP.
Proof. It is enough to show that T̃3,2,k is zero in degree 2k, which follows from the
observation that (x2 , y 2 , (x + y)2 )k = (x2 , y 2 , xy)k .
In Theorem 18 below we prove the necessity of conditions (3)-(5) of Conjecture 3.
Theorem 18. The algebra T3,d,k fails to have the WLP in the cases
(1) k = 2j + 1 > 2, d = (j + 2)(2ℓ + 1), where ℓ ≥ 1.
(2) k = 2j + 1 > 2, k ∈
/ {3, 7}, d = (j + 2)(2ℓ + 1) ± 1, where ℓ ≥ 1.
(3) k = 2j > 2, d ∈ {(j + 1)(2ℓ + 1) + ℓ, (j + 1)(2ℓ + 1) + ℓ + 1}.
Proof. By [5] we have that the Hilbert series of T3,d,k is given by
dk
2 + 2k)td(k+1) − k+1 td(k+2)
1 − k+2
t
+
(k
2
2
.
T3,d,k (t) =
(1 − t)3
From this, we get that in the range dk ≤ i ≤ d(k + 1) the Hilbert function is given
i+2
i − dk + 2 k + 2
−
2
2
2
with first difference equal to (i + 1) − (i − dk + 1)(k + 1)(k + 2)/2 which is positive when
i + 1 < d(k + 1)(k + 2)/(k + 3) and negative when i + 1 > d(k + 1)(k + 2)/(k + 3). The
turning point is in the interval where this expression for the Hilbert function is valid.
POWERS OF GENERIC IDEALS AND THE WLP FOR POWERS OF SOME MONOMIAL CI’S
9
In order to show that the WLP fails, we will use the representation theory of the symmetric group, S3 . Since the algebra is monomial, it is sufficient to consider multiplication
by the linear form L = x + y + z which is symmetric and hence gives an equivariant map.
In all of the cases we consider, we have that d = (2ℓ + 1)(k + 3)/2 + ǫ, where −1 ≤ ǫ ≤ 1.
We will look at the multiplicity of the trivial and the alternating representations in T3,d,k
in degrees dk + 2ℓ − 1 and dk + 2ℓ. We compute
these
by means of the
multiplicities
i+2
k+2 i−dk+2
characters. In degree i of T3,d,k we have 2 − 2
monomials that are all
2
fixed by the identity permutation. The other two even permutations fixes only powers
of xyz so the character is 1 if i ≡ 0 (mod 3) and k 6≡ 0 (mod 3) and 0 otherwise. The
transposition (12) fixes the monomials of the form (xy)m z n and we have that the value
of the character on the three transpositions is
k+2
i − dk + 2
i+2
−
.
2
2
2
Thus we get that the multiplicity of the trivial representation is
i+2
k+2
i − dk + 2
k + 2 i − dk + 2
i+2
1
−
−
+m+3
6
2
2
2
2
2
2
where the middle term m is −2, 0 or 2 depending on the congruence modulo 3 of the
sum of the other terms. In the same way we get the multiplicity of the alternating
representation as
1
i+2
k+2
i − dk + 2
k + 2 i − dk + 2
i+2
−
−
+m−3
6
2
2
2
2
2
2
When we compute the difference of the Hilbert function between degree dk + 2ℓ − 1 and
dk + 2ℓ we get
dk + 2ℓ + 2
dk + 2ℓ + 1
k+2
2ℓ + 2
2ℓ + 1
−
−
−
2
2
2
2
2
k+2
k(k + 3)
= kǫ.
= dk + 2ℓ + 1 −
(2ℓ + 1) = dk − (2ℓ + 1)
2
2
The difference in the number of monomials fixed by a transposition is
dk + 2ℓ + 1
k+2
2ℓ + 2
2ℓ + 1
dk + 2ℓ + 2
−
−
−
2
2
2
2
2
(
dk
dk − 1
k+2
−j − 1 if k and d are odd,
=
−
−
=
2
2
2
−j
otherwise.
When k is odd and ǫ = ±1, d is odd when j is even and the difference above can be
expressed as −2⌊j/2⌋ − 1. When k is odd and ǫ = 0, d is odd when j is odd and the
same difference can be expressed as −2⌊(j + 1)/2⌋.
First we consider the case k = 2j even and ǫ = ±1/2. We compute the difference in the
multiplicity of the trivial representation when ǫ > 0 and the alternating representation
when ǫ < 0. This difference equals
j+1
1
(±j + m ± 3(−j)) = ∓
6
3
10
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
which has opposite sign to the difference in the Hilbert function when j + 1 ≥ 3. Thus
the multiplication by L = x + y + z cannot have maximal rank by Schur’s lemma.
For the case k = 2j + 1 and ǫ = ±1, we have the corresponding difference
1
j
j
j
±(2j + 1) + m ± 3 −2
−1
=∓
−
.
6
2
2
3
Again, this has different sign than the difference in the Hilbert function when ⌊j/2⌋ >
⌊j/3⌋, i.e., when j = 2 or j ≥ 4.
In the last case k = 2k + 1 and ǫ = 0 it is sufficient to show that the multiplicity
of the trivial representation changes since the Hilbert function has difference zero. The
difference in the multiplicity of the trivial representation is
j+1
j+1
1
0 + m + 3 −2
=−
6
2
2
which is negative for all j ≥ 1.
Remark 19. We can see that the argument of the proof does not work when ǫ = ±3/2
or ǫ = ±2, so in these cases we have the same turning point for the Hilbert function of
the three isotypic components.
Remark 20. By computations in Macaulay2, we have verified the unproven parts of
Conjecture 3 for d, k ≤ 30.
3.2. The general case. As the number of variables increases, the number of non-trivial
pairs (d, k) for which C[x1 , . . . , xn ]/(xd1 , . . . , xdn )k has the WLP seems to decrease.
n
4
5
6
7
8
9
10
11
12, 14, 16
2a + 1
WLP
(2, 2), (2, 3), (3, 2), (4, 2)
(2, 2), . . . , (2, 7), (3, 2), (4, 2), (4, 3), (6, 2)
(2, 2), (2, 3), (3, 2)
(2, 2), (2, 3), (3, 2)
(2, 2), (3, 2)
(2, 2), (2, 3), (3, 2)
(2, 2), (3, 2)
(2, 2), (2, 3)
(2, 2)
(2, 2)
Table 1. The right hand column consists of pairs (d, k) with d, k ≥ 2
for which C[x1 , . . . , xn ]/(xd1 , . . . , xdn )k has the WLP, detected by computations in Macaulay2, except for the last row, which relies on Theorem
22.
In Table 1 we list the non-trivial cases where have been able detect the WLP. Based
upon these observations together with some negative results for the WLP in positive
characteristic, we believe that Table 1 is complete, except for the trivial pairs, when n ≤
10. When n ≥ 11, we guess that the WLP holds at most in the cases (2, 2), (2, 3), (3, 2)
POWERS OF GENERIC IDEALS AND THE WLP FOR POWERS OF SOME MONOMIAL CI’S
11
for n odd, and at most in the cases (2, 2), (3, 2) for n even. We can support our guesses
with two theoretical results.
Theorem 21.
• The algebra
• The algebra
• The algebra
• The algebra T4,2,k has the WLP if and only if k ≤ 3.
T4,3,k has the WLP if and only if k ≤ 2.
T4,4,k has the WLP if and only if k ≤ 2.
T5,2,k has the WLP if and only if k ≤ 7.
Proof. By Theorem 16, the algebra T4,2,k fails the WLP when k ≥ 22 = 4, T4,3,k fails
the WLP when k ≥ 32 = 9, T4,4,k fails the WLP when k ≥ 42 = 16, T5,2,k fails the WLP
when k ≥ 23 = 8,
Hence it is enough to check the WLP over Q for the remaining cases. This has been
done with Macaulay2.
Finally, we can prove that C[x1 , . . . , xn ]/(x21 , . . . , x2n )2 has the WLP when n is odd.
We believe that it is also true for n even, but our method of proof does not apply in
that case.
Theorem 22. For any odd n, the algebra C[x1 , . . . , xn ]/(x21 , . . . , x2n )2 has the WLP.
Proof. Let R = C[x1 , . . . , xn ]/(x21 , . . . , x2n )2 . When d ≥ 2, write Rd = Ad,1 ⊕ · · · ⊕
· · · Ad,n−1 ⊕ Bd , where Ad,i is spanned by all monomials of the form x2i M with M
a squarefree monomial, and where Bd is the set of squarefree monomials of degree d
n
together with all monomials of the form x2n M , with M squarefree. We have |Ad,i | = d−2
n
and |Bd | = nd + d−2
, where |V | denotes the dimension of the vector space V . Notice
that Ri Ad,j ⊂ Ai+d,j .
Multiplication by (x1 + · · · + xn ) on Ad,i agrees with multiplication by (x1 + · · · + xn )
on the basis in degree d − 2 in C[x1 , . . . , xn ]/(x21 , . . . , x2n ). This algebra has the SLP, so
multiplication by (x1 + · · · + xn ) on Ad,i has full rank.
Since (x21 , . . . , x2n )2 ⊆ (x21 , . . . , x2n−1 , x4n ), the dimension of (x1 + · · · + xn )Bd in R
is greater than or equal to the dimension of (x1 + · · · + xn )Bd in the algebra C :=
C[x1 , . . . , xn ]/(x21 , . . . , x2n−1 , x4n ), where we abuse notation and regard Bd as a part of
C. This algebra also has the SLP, so the dimension
of (x1 + · · · + xn )Bd in C equals
n
n
n
n
min(|Bd |, |Bd+1 |) = min d + d−2 , d+1 + d−1 .
If |Bd | ≤ |Bd+1 | and |Ad | ≤ |Ad+1 | we can conclude that multiplication by (x1 + · · · +
xn ) on Rd is injective. If |Bd | ≥ |Bd+1 | and |Ad,i | ≥ |Ad+1,i |, we can conclude that
multiplication by (x1 + · · · + xn ) on Rd is surjective.
Thus we are left with two cases.
In the first case wehave |Bd | < |Bd+1 | and |A
have the
is, we
d+1,i |, that
d,i | >n |A
n
n
n
n
n
n
n
inequalities d + d−2 < d+1 + d−1 and d−2 > d−1 . But d−2 > d−1 implies
n
n
n
n
. Thus the inequalities |Bd | < |Bd+1 |
+ d−1
> d+1
, so nd + d−2
that nd > d+1
and |Ad,i | > |Ad+1,i | cannot be simultaneously satisfied.
n
n
to
< d−1
In the second case we have |Bd | > |Bd+1 | and |Ad,i | < |Ad+1,i |. For d−2
n
n
n
n
hold we need to have d − 1 ≤ ⌊n/2⌋. For d + d−2 > d+1 + d−1 to simultaneously
12
MATS BOIJ, RALF FRÖBERG, AND SAMUEL LUNDQVIST
n
hold, we especially need the inequality nd > d+1
to hold. This inequality is satisfied
if and only if d ≥ ⌊(n + 1)/2⌋.
Now the inequalities d − 1 ≤ ⌊n/2⌋ and d ≥ ⌊(n + 1)/2⌋ together
gives
us that
2k+1
2k+1
d = ⌊n/2⌋ + 1. In the case n = 2k + 1, we have |Bk+1 | = 2k+1
=
+
=
k+1
k+1
k−1
2k+1
2k+1
+ k+2 = |Bk+2 |, so the inequality |Bd | > |Bd+1 | is not fulfilled. This shows
k
that the case |Bd | > |Bd+1 | and |Ad,i | < |Ad+1,i | is empty when n is odd, that is, R has
the WLP when n is odd.
Remark: When n = 2k, we can however not draw any conclusion regarding the rank
of the map ×(x1 + · · · + xn ) from degree k + 1 to k + 2.
References
[1] David J. Anick. Thin algebras of embedding dimension three. J. Algebra, 100(1):235–259, 1986.
[2] David Cook, II and Uwe Nagel. The weak Lefschetz property, monomial ideals, and lozenges. Illinois
J. Math., 55(1):377–395 (2012), 2011.
[3] David Cook, II and Uwe Nagel. The weak Lefschetz property for monomial ideals of small type. J.
Algebra, 462:285–319, 2016.
[4] Ralf Fröberg. An inequality for Hilbert series of graded algebras. Math. Scand., 56(2):117–144, 1985.
[5] Elena Guardo and Adam Van Tuyl. Powers of complete intersections: graded Betti numbers and
applications. Illinois J. Math., 49(1):265–279 (electronic), 2005.
[6] Tadahito Harima, Juan Migliore, Uwe Nagel, and Junzo Watanabe. The weak and strong Lefschetz
properties for Artinian K-algebras. J. Algebra, 262(1):99–126, 2003.
[7] Juan Migliore and Uwe Nagel. Survey article: a tour of the weak and strong Lefschetz properties. J.
Commut. Algebra, 5(3):329–358, 2013.
[8] Richard P. Stanley. Weyl groups, the hard Lefschetz theorem, and the Sperner property. SIAM J.
Algebraic Discrete Methods, 1(2):168–184, 1980.
Department of Mathematics, KTH - Royal Institute of Technology
E-mail address: [email protected]
Department of Mathematics, Stockholm University
E-mail address: [email protected]
Department of Mathematics, Stockholm University
E-mail address: [email protected]
| 0math.AC
|
Performance Analysis of Robust Stable PID Controllers Using
Dominant Pole Placement for SOPTD Process Models
Saptarshi Das1, Kaushik Halder2, and Amitava Gupta2
1) Department of Mathematics, College of Engineering, Mathematics and Physical Sciences, University of
Exeter, Penryn Campus, Penryn TR10 9FE, United Kingdom
2) Department of Power Engineering, Jadavpur University, Salt Lake Campus, LB-8, Sector 3, Kolkata700098, India
Author’s Emails:
[email protected], [email protected] (S. Das*)
[email protected] (K. Halder)
[email protected] (A. Gupta)
Phone number: +44-7448572598
Abstract:
This paper derives new formulations for designing dominant pole placement based proportionalintegral-derivative (PID) controllers to handle second order processes with time delays (SOPTD).
Previously, similar attempts have been made for pole placement in delay-free systems. The presence
of the time delay term manifests itself as a higher order system with variable number of interlaced
poles and zeros upon Pade approximation, which makes it difficult to achieve precise pole placement
control. We here report the analytical expressions to constrain the closed loop dominant and nondominant poles at the desired locations in the complex s-plane, using a third order Pade
approximation for the delay term. However, invariance of the closed loop performance with different
time delay approximation has also been verified using increasing order of Pade, representing a closed
to reality higher order delay dynamics. The choice of the nature of non-dominant poles e.g. all being
complex, real or a combination of them modifies the characteristic equation and influences the
achievable stability regions. The effect of different types of non-dominant poles and the
corresponding stability regions are obtained for nine test-bench processes indicating different levels of
open-loop damping and lag to delay ratio. Next, we investigate which expression yields a wider
stability region in the design parameter space by using Monte Carlo simulations while uniformly
sampling a chosen design parameter space. The accepted data-points from the stabilizing region in the
design parameter space can then be mapped on to the PID controller parameter space, relating these
two sets of parameters. The widest stability region is then used to find out the most robust solution
which are investigated using an unsupervised data clustering algorithm yielding the optimal centroid
location of the arbitrary shaped stability regions. Various time and frequency domain control
performance parameters are investigated next, as well as their deviations with uncertain process
parameters, using thousands of Monte Carlo simulations, around the robust stable solution for each of
the nine test-bench processes. We also report, PID controller tuning rules for the robust stable
solutions using the test-bench processes while also providing computational complexity analysis of
the algorithm and carry out hypothesis testing for the distribution of sampled data-points for different
classes of process dynamics and non-dominant pole types.
Keywords:
dominant pole placement, PID controller tuning, second order plus time delay (SOPTD), control
performance, robust stable controller, stability region, signal/system norms, gain/phase margin
1
Nomenclature:
m : Non-dominance parameter for pole placement,
K : Open loop process DC gain,
L : Open loop process time delay,
T : Open loop process time constant or lag,
ol 1 T : Open loop process natural frequency,
ol : Open loop process damping ratio,
cl : Closed loop process natural frequency,
cl : Closed loop process damping ratio,
K p : Proportional gain,
K i : Integral gain,
K d : Derivative gain,
s : Laplace operator,
2 : System norms,
2 : Signal norms,
Gm : Gain margin,
m : Phase margin,
gc : Gain crossover frequency,
d(t): Disturbance input,
u(t): Control signal,
y(t): output variable,
e(t): control error signal,
J: Performance measure,
S e s : sensitivity function,
T s : complementary sensitivity function,
S d s : disturbance sensitivity function,
S u s : control sensitivity function
1. Introduction
2
The dynamic behaviours of many industrial processes are affected and governed by significant
amount of time delays in the control loops. The time delay is caused by the flow of information,
energy and transport of physical variables, computer processing time etc. (Normey-Rico & Camacho
2007). The introduction of the time delay makes the continuous time closed loop system to have an
infinite order (Åstrӧm & Wittenmark 2011) upon exponential series expansion of the delay term (
e Ls ) which is difficult to handle with a finite term controller (Michiels et al. 2002). To alleviate this
problem, there have been several works to design controllers for such systems e.g. in (Zhong 2006). It
is well known that most of the controllers used in the process industries are of PID type (Ǻstrӧm &
Hägglund 1995) (Zhong 2006) due to its simplicity and ease of implementation, nice disturbance
rejection, tracking performance etc. Amongst many other available approaches, the Internal Model
Control (IMC) based tuning of PID controllers has been quite popular to handle First Order Plus Time
Delay (FOPTD) and Second Order Plus Time Delay (SOPTD) processes, as well as Integral Process
with Dead Time (IPDT) (Shamsuzzoha & Lee 2007; Rivera et al. 1986) because of its good
robustness on uncertain plants. Another approach on the design of smith predictor augmented PID
controller to handle time delay processes have been reported in (Astrom et al. 1994) which yields
improved tracking and load disturbance rejection performances. A modified methodology is proposed
with combined Smith predictor and PID controller in (Matausek & Micic 1996) considering
challenging higher order integrating plants with delays. However the main drawback of this method is
that it cannot handle unstable process with delay (Normey-Rico & Camacho 2007), unless an
additional observer is used (Furukawa & Shimemura 1983). To overcome these problems of
complicated time delay processes, the model predictive control (MPC) has got attention by many
researchers but there are only few results for time delay systems (Ellis & Christofides 2015). Initially
the MPC was developed mainly to control slower processes as it requires large computational burden
for prediction and optimisation-based control. MPC controller design for time delay systems is mostly
an open area and there are only few results like (Holis & Bobál 2015). Another important area is
designing output feedback controller (De la Sen 2005) as well as state feedback controller (Michiels et
al. 2010) for stabilizing time delayed systems which are gradually gaining increased attention. In the
literature some control algorithms are proposed using linear matrix inequalities (LMIs) for time
delayed systems e.g. (Niculescu 1998) to enforce robustness and several closed loop performance
measures. Despite having these results, traditional pole placement remained quite challenging for time
delay systems because of its increased or even infinite order.
In this paper, we propose an analytical formulation for dominant pole placement tuning of PID
controllers to handle SOPTD systems. This is due to the fact that in many process industries, the
dynamical behaviours in a large variety of self-regulating processes can be modelled using the
SOPTD template with the flexibility of showing both sluggish and oscillatory open loop dynamics as
well as different lag to delay ratio or normalized dead time (O’Dwyer 2009). PID controllers are
traditionally tuned by various means like time and frequency domain performance criteria or design
specifications (Cominos & Munro 2002). Amongst many available approaches, the dominant pole
placement method has been quite promising as the designer can specify his demand of closed loop
performance, as the equivalent second order system’s damping ratio, time constant or natural
frequency (Wang et al. 2009). Amongst previous approaches, dominant pole placement based PID
controller design for delay free second order systems have been addressed in (Saha et al. 2012)(Das,
Halder, et al. 2012) whereas its time delay version has been extended in this paper and the method has
been verified on several test-bench SOPTD plants.
A continuous pole placement method (controlling the rightmost root of the closed loop system
and shift it to the left half of the s-plane in quasi-continuous way) has been proposed to design a semiautomated pole placement based state feedback controller for retarded (Michiels et al. 2002) and
neutral type (Michiels & Vyhlidal 2005) delay systems. By this method, the closed loop roots lying in
the extreme right-hand side is shifted to the far possible left-hand side. However, the methodology
3
does not allow direct pole placement for SOPTD system and only monitoring the real part of the
roots. To overcome the above problem another methodology is proposed in (Michiels et al. 2010)
which combines direct pole placement and the minimization of the spectral abscissa for determining
controller parameters in retarded time-delay systems. There are some other approaches on stability
analysis of time delayed single input single output (SISO) systems to derive controller gains by
computing the root locus. Using the characteristic equation which leads to a transcendental equation
in the presence of delays which is also known as the quasi-polynomial, several methods have been
proposed to construct the root locus which creates horizontal asymptotes (Krall 1970; Yeung & Wong
1982; Huang & Li 1967). Other root locus based stabilization methods are also reported to analyse
state space models with input delay (Engelborghs et al. 2001), state delay or both (Suh & Bien 1982)
by using the root locus.
The PID controller tuning by direct pole assignment is found to be a difficult approach for time
delay systems, as the time delay in a process makes the closed loop system to have an infinite order.
Therefore, to handle time delay systems, the direct pole placement in complex s-plane is not
recommended as suggested in the pioneering work on dominant pole placement tuning in (Wang et al.
2009). The methodology in (Wang et al. 2009) also suggested a Nyquist based design for frequency
domain stabilization of the time delay systems using PID controllers. The main hurdle with the pole
placement for delay systems has been the fact that the exponential delay term in Laplace domain (i.e.
e Ls ), manifests itself as very high order transfer function upon approximations using Pade/Routh
methods of a specified order (Silva et al. 2007). Therefore, such a pole placement approach will need
relocation of many open loop poles at a time with a compact finite term (three-term for PID)
controller when the number of interlacing pole-zeroes, arising due to the approximation of the time
delay term are too many to handle for a chosen order of approximation. In our derivations, the order
and approximation method are considered to be fixed. In particular, we apply a third order Pade
approach to approximate the time delay term in the process model. Therefore, it can be considered as
a new area of research to get a clearer picture of the dominant pole placement design for time delay
systems where the task is to handle many poles and zeros of the combined plant and delay with finite
number of controller parameters. To demonstrate the methodology, SOPTD processes with various
delay to lag ratio have been used which shows the strength of the algorithm for even quite challenging
plants e.g. delay dominant systems with low open loop damping which are much harder to control
using standard PID controller tuning methods.
2. Theoretical formulation
In this section, the dominant pole placement based PID controller design has been shown to
handle SOPTD systems. The time delay term has been approximated using a third order Pade’s
approximation instead of considering it as transcendental term in the quasi-polynomial (Silva et al.
2007). The co-efficient matching based pole placement method has been described previously in
(Ǻstrӧm & Hägglund 1995)(Kiong et al. 1999) for the control of delay free systems. We apply here a
similar co-efficient matching method to design dominant pole placement based PID controller gains
using the characteristic equation having third order Pade approximation of the delay term.
Now, let us consider the open loop SOPTD process can be represented by:
G ( s)
K
e Ls ,
s 2 2 ol ol s ol2
(1)
which is to be controlled by the PID controller
C (s) K p
K d s2 K p s Ki
Ki
Kd s
.
s
s
(2)
4
Then the closed loop system with PID controller can be written as:
Gcl s
G s C s
1 G s C s
.
(3)
Again, using a third order Pade approximation for the time delay term ( e
SOPTD process model (1) can be written as:
e Ls
Ls
) of the open loop
L3 s3 12L2 s 2 60 Ls 120
.
L3 s3 12 L2 s 2 60Ls 120
(4)
Therefore, the full expression of the closed loop system transfer function including the open loop
plant and PID controller is given by:
Gcl
K K d s 2 K p s K i L3 s 3 12 L2 s 2 60 Ls 120
s s 2 2 ol ol s ol2 L3 s3 12 L2 s 2 60 Ls 120 K K d s 2 K p s K i L3 s 3 12 L2 s 2 60 Ls 120
.
(5)
It is evident from (5) that the closed loop process has six poles and five zeros which can be
modulated using suitable choice of the three controller gains K p , Ki , Kd . As the closed loop system
in (5) has six poles, the corresponding desired closed-loop characteristic polynomial should also
contain six roots out of which two must meet the equivalent desired closed loop system specifications.
Rest of the four non-dominant poles can be allowed to have different characteristics e.g. all four
complex conjugate poles, all four real poles, two real and two complex conjugate poles which have
been described in the following sub-sections along with the respective derivations for the dominant
pole placement tuning.
2.1.
Deriving unique and alternative expressions for stabilizing PID controller gains and
obtaining the stability regions
Using the coefficient matching of the desired vs. the given characteristic equation of the closed
loop system, we show here that for a chosen set of process parameters K , L, T , ol , the expressions
for integro-differential gains Ki and Kd produce unique values. But there can be many alternative
expressions for the proportional gain Kp which can be derived from the two controller gains Ki and Kd.
The multiple expressions for Kp arise due to an overdetermined system of algebraic equations from
various higher powers of the Laplace variable s. Therefore, for a given set of plant parameters and
design specifications, after obtaining the controller gains Ki , Kd , there can be many possible
solutions for Kp. The most robust expression for selecting Kp can be found out by comparing the
stabilizing regions for all these expressions having the highest volume, compared to the same with the
others. This is determined by drawing uniformly distributed random samples from a chosen design
parameter space with a specified range and then evaluating the stability at each of these sampled
points, as determined by the real part of the closed loop poles being negative.
In order to show this, we use Monte Carlo simulations by sampling the design specifications
m, cl ,cl from chosen intervals for fixed open loop plant parameters ol , ol , L . The effect
of
system’s DC-gain K has not been investigated here, since it simply decreases all the three PID
controller gains in a linear fashion. However, the effect of other parameters may be complicated and
therefore needs to be explored rigorously using Monte Carlo simulations which is adopted next.
5
Let us now assume that for a given set of open loop plant parameters ol , ol , L , one can choose
different design specifications m, cl , cl to derive the PID controller gains. However, due to the
finite term approximation of the time delay, the resulting closed loop system is not necessarily stable
for any randomly chosen design parameter or arbitrary PID controller gains within a specified range.
We substantiate this argument, using a computational approach of randomly sampling the design
specification space within a chosen bound, for fixed open loop SOPTD process parameters which
yields a set of stabilizing PID controller gains K p , Ki , Kd . However, with different design
specifications m, cl , cl , the closed loop system may be stable, but the closed loop performance is
expected to vary widely within this entire stability region. Such thousands of randomly sampled guess
values for the design specifications yielding stable closed loop systems can be quantified as a fraction
of the accepted stable samples to the total number of samples drawn from the entire design
specification space. For a more robust system design, the method producing largest stabilizing volume
in the design parameter space or the equivalent controller parameter space should be selected.
For the set of stabilizing controller gains, we also calculate various closed loop performance
measures like set-point tracking, disturbance rejection, control effort to investigate their trade-offs
(Das & Pan 2014), peak sensitivity and complementary sensitivity functions etc. (Åstrӧm & Hägglund
2004) signifying parametric robustness and high frequency measurement noise rejection performances
respectively. We also demonstrate here the wide applicability of our proposed approach on three
different class of SOPTD processes e.g. lag-dominant ( T L ), delay dominant ( T L ) and balanced
lag-delay ( T L ) system, also each of them with three different damping levels viz. overdamped (
ol 1 ), critically damped ( ol 1 ), and underdamped ( ol 1 ). Therefore, we have nine different
cases to explore, as shown in the subsequent sections. For the Monte Carlo simulations, the ranges for
the desired control performance parameters are chosen as m 1,10 , cl 1,5 ,cl 1,10 , as per the
previous reports like (Panda et al. 2004)(Wang et al. 2009).
2.2.
Pole placement PID controller design with all non-dominant complex conjugate poles
Now, as discussed in (Kiong et al. 1999)(Ǻstrӧm & Hägglund 1995) in order to ensure guaranteed
dominant pole placement with PID controllers, let us consider that the closed loop system (5) has six
complex (conjugate) poles. Amongst these six, two of them are dominant meeting the desired closed
loop design specifications cl , cl . The rest four poles are non-dominant in nature and their
locations can be controlled by selecting the non-dominance pole placement parameter m. One can
easily choose the m in such a way that these poles do not have any significant effect on the closed
loop performance of the process with design specifications cl , cl . Under this assumption, the non-
nd
dominant closed loop complex conjugate poles become s1,2
m cl cl jcl 1 cl2 . Therefore,
the resulting characteristic polynomial with both dominant and non-dominant poles can be written as:
4
( s) s 2 2 cl cl s cl2 s s nd
s s
s 2 2 cl cl s cl2 s s1nd s s2nd
s 2 2 cl cl s cl2
2
nd
1
2
s2nd s1nd s2nd
2
s 2 2 cl cl s cl2 s 2 2m cl cl s m 2 cl2 cl2 m 2cl2 1 cl2
(6)
2
s 2 2 cl cl s cl2 s 2 2m cl cl s m 2cl2 0.
2
In order to find out the controller parameters using the co-efficient matching method which is also
reported in (Kiong et al. 1999)(Ǻstrӧm & Hägglund 1995) for the delay free cases, the closed loop
6
characteristic polynomial (6) can be expanded in terms of the open loop process parameters and the
PID controller gains as:
s 6 L3
s 5 12 L2 2 ol ol L3 KK d L3
s 4 60 L 24 ol ol L2 ol2 L3 12 KK d L2 KK p L3
s 3 120 120 ol ol L 12ol2 L2 60 KK d L 12 KK p L2 KK i L3
(7)
s 240 olol 60 L 120 KK d 60 KK p L 12 KK i L
s1 120ol2 120 KK P 60 KK i L
2
2
ol
2
s 0 120 KK i
0.
Now the characteristic polynomial in terms of the desired closed loop poles in (6) can also be written
as:
s 6 1
s5 4m cl cl 2 cl cl
s 4 2m2cl2 1 2 cl2 8m cl2 cl2 cl2
s3 4m3 cl cl3 4m2 cl cl3 1 2 cl2 4m cl cl3
(8)
s 2 m 4cl4 8m3 cl2 cl4 2m2cl4 1 2 cl2
s1 2m 4 cl cl5 4m3 cl cl5
s 0 m4cl6
0.
For comparing the coefficients of these two characteristic polynomials, after dividing (7) by L3 and
comparing it with (8), the corresponding PID controller gains can be easily obtained as:
m 4cl6 L3
120 K
12
s 5 : K d 2 ol ol 2 cl cl 1 2m K
L
s0 : Ki
s1 : K p 4m3 cl cl5 L3 m 2 m 4cl6 L4 240ol2 240 K
240 ol ol 60o2l 120 KK d 12 KK i
s 2 : K p L2
2
m 4cl4 8m3 cl2 cl4 2m 2cl4 1 2 cl2 60 K
3
3
L
L
L
L
s 3 : K p L3 4m3 cl cl3 4m 2 cl cl3 1 2 cl2 4m cl cl3 120 120 ol ol L 12ol2 L2 60 KK d L KK i L3 12 KL2
12 KK d
60 24 ol ol
s4 : K p 2
ol2
2m 2cl2 1 2 cl2 8m cl2 c2l cl2 K .
L
L
L
(9)
It is observed from (9) that the expressions for the two PID controller gains K i , K d are unique but the
proportional gain K p can have four possible expressions.
2.3.
Pole placement PID controller design with all non-dominant real poles
7
Similar to the above closed-loop characteristic equation in order to find the PID controller gains,
the non-dominant poles can also be considered as real (instead of imaginary) and still be manipulated
by the non-dominance pole placement parameter m while meeting the same dominant closed loop
specification cl , cl . Therefore, for all real non-dominant poles, the characteristic equation
becomes:
( s) s 2 2 cl cl s cl2 s m cl cl
4
s 2 2 cl cl s cl2 s 4 4m cl cl s3 6m2 cl2cl2 s 2 4m3 cl3 cl3 s m4 cl4 cl4 0.
(10)
The desired characteristic polynomial (10) can be expanded and rearranged according to the power of
Laplace variable s, as required for the coefficient matching with that of the open loop system with
PID controller in (7), which yields:
s 6 1
s 5 4m cl cl 2 cl cl
s 4 6m2 cl2 cl2 8m cl2 cl2 cl2
s 3 4m3 cl3 cl3 12m 2 cl3 cl3 4m cl cl3
(11)
s 2 m 4 cl4 cl4 8m3 cl4 cl4 6m 2 cl2 cl4
s1 2m 4 cl5 cl5 4m3 cl3 cl5
s 0 m4 cl4 cl6
0.
Now in a similar way dividing (7) by L3 and matching the coefficients of s with (11) yields the
corresponding PID controller gains for the real non-dominant poles as:
s 0 : K i m 4 cl4 cl6 L3 120 K
12
s 5 : K d 2 ol ol 2 cl cl 1 2m K
L
1
3
4 5 5
3 3 5
s : K p L 2m cl cl 4m cl cl 120ol2 60 KK i L 120 K
240 ol ol 60ol2 120 KK d 12 KK i
s 2 : K p L2
2
m 4 cl4 cl4 8m3 cl4 cl4 6m 2 cl2 c4l 60 K
3
3
L
L
L
L
.
s 3 : K p L3 4m3 cl3 cl3 12m 2 cl3 cl3 4m 2 c2l cl4 120 120 ol ol L 12ol2 L2 60 KK d L KK i L3 12 KL2
12 KK d
60 24 ol ol
s4 : K p 2
ol2
6m 2 cl2 c2l 8m cl2 c2l c2l K .
L
L
L
(12)
Here also, the two PID controller gains K i , K d can be uniquely derived from the open loop and
desired process parameters, but the proportional gain K p can have four different expressions.
2.4.
Pole placement PID controller design with two real and two complex conjugate nondominant poles
The characteristic equation using two real and two complex conjugate non-dominant poles can be
written as:
( s) s 2 2 cl cl s cl2 s 2 2m cl cl s m2cl2 s m cl cl 0 .
2
This yields the coefficients in decreasing order of Laplace variable s as:
8
(13)
s 6 1
s 5 4m cl cl 2 cl cl
s 4 5m 2 cl2 cl2 m 2cl2 8m cl2 cl2 cl4
s 3 2m3 cl3 cl3 2m3 cl cl3 2m 2 cl3 cl3 8m 2 cl3 cl3 2m 2 cl cl3 4m cl cl3
s 2 m 4 cl2 cl4 4m3 cl4 cl4 4m3 cl2 cl4 m 2 cl2 cl4 4m 2 cl2 cl4 m 2cl4
(14)
s1 2m 4 cl3 cl5 2m3 cl3 cl5 2m3 cl cl5
s 0 m 4 cl2 cl6
0.
Again, dividing (7) by L3 and match the co-efficient with (14) as done previously, yields the
corresponding PID controller gains as:
s 0 : K i m 4 cl2 cl6 L3 120 K
12
s 5 : K d 2 ol ol 2 cl cl 1 2m K
L
1
3
4 3 5
3 3 5
s : K p L 2 m cl cl 2m cl cl 2m 3 cl cl5 120ol2 60 KK i L 120 K
240 ol ol 60ol2 120 KK d 12 KK i
s 2 : K p L2
2
m 4 cl2 cl4 4m3 cl4 cl4 4m3 cl2 c4l m 2 cl2 cl4 4m 2 cl2 cl4 m 2cl4 60 K
3
3
L
L
L
L
L3 2m3 cl3 cl3 2m3 cl cl3 2m 2 cl3 cl3 8m 2 cl3 cl3 2m 2 cl cl3 4m cl cl3
12 KL2
s : Kp
120 120 ol ol L 12ol2 L2 60 KK d L KK i L3
3
12 KK d
60 24 ol ol
s4 : K p 2
ol2
5m 2 cl2 cl2 m 2cl2 8m cl2 cl2 cl4 K .
L
L
L
(15)
Similar to the previous cases, K i , K d gains have unique expressions but the gain K p has four
possible values depending on the choice of different coefficients of Laplace variable s.
3. Designing robust stable PID controller and evaluation of different closed loop
performance measures
3.1.
Determining the robust stable solutions using centroid of the stability region
We here use the k-means clustering algorithm to determine the centroid of the stability region in
the PID controller space which can have a complex shape in the 3D parameter space of the controller
gains K p , Ki , K d . The stability regions are determined for nine classes of SOPTD plants and it is
also checked that a single centroid represents the stability regions of a unimodal distribution in the
controller parameter space. Otherwise if a multi-modal distribution is discovered, indicating more
than one possible robust stable solution for the controller gains, the number of centroids (k) in the kmeans clustering algorithm can also be set to the number of modes in the distribution of controller
gains. Previously determination of the stability region centroids in the controller parameter space were
done by looking for only circular (for two gains), spherical (for three gains) or (hyper)-spherical
clusters (more than two controller gains) as shown in (Pan et al. 2011), which has been extended here
to find out centroids of more generalised complicated shaped clusters like ellipsoids or other
structures by assuming that the centroid is likely to lie in the high density region of the clusters where
most of the samples are accepted in the random Monte Carlo sampling.
Although the chosen design parameter space has been explored using 105 uniformly distributed
random samples, certain areas of the design specification or the equivalent controller parameter space
9
has more stable solutions than the others. Therefore, the present random sampling approach helps in
understanding the shape of the stability region using various expressions for deriving the PID
controller gains, while also giving the flexibility to specify the nature of the non-dominant poles (i.e.
all real, all complex or mixed), as explored in the previous sections.
The k-means clustering algorithm starts with a random initial guess for the centroid of the multidimensional data space and iteratively move the centroids based on minimizing the squared Euclidean
distance criteria from all the data points (Rogers & Girolami 2012). However, to ensure that the best
possible estimate of the centroid or the robust stable solution has been discovered in the iterative
process, the k-means clustering has been run 10 times, for each case with random starting points and
the best solution with minimum Euclidean distance is chosen as the final estimate of the centroid,
signifying the most robust stable solution in the PID controller parameter space.
Therefore, for a given process model characterised by the constants K , ol , ol , L , the robust
stable PID controller gains obtained via the dominant pole placement can be obtained using the
following steps:
Step 1: Choose the non-dominant pole types amongst all complex, all real or mixed using the
expressions in (9), (12) or (15) to map the open loop and desired closed loop system parameters on to
the PID controller gains.
Step 2: Obtain the stabilizing PID controller gains using few thousands of uniformly distributed
samples within a chosen range of all design parameters m, cl , cl .
Step 3: Cluster the stabilizing PID gains to get the centroid as the robust stable solution.
Step 4: Evaluate different performance measures with the robust stable PID controller on the nominal
and perturbed process models.
3.2.
Performance measures with the stabilizing PID controller gains
The previous section derives the expressions for obtaining stabilizing PID controller gains by
choosing the nature of the non-dominant poles being all complex, real or mixture of them. However,
in a realistic control system design problem, apart from the stability, the control loop performance is
also of major concern. Therefore, the robust stable solutions might not always show an acceptable
control performance. On contrary the controller setting for optimised performance criteria may not
have a sufficient robustness against plant or controller parameter variation. Therefore, we chose to
design the PID controller settings based on maximum robust stability and then compare the
achievable control performances. However the robustness checking of optimally designed controllers
is also a valid approach as previously studied for single control objective (Das, Pan, et al. 2012; Das et
al. 2011) and multiple conflicting control objectives (Das et al. 2015)(Das & Pan 2014). Also, the
parametric robustness of optimal controllers and optimality of robust design have been previously
discussed in (Pan & Das 2016).
After the robust stable solution is determined using clustering which yields the centroid of the
arbitrary shaped stability region, few well-known performance measures are calculated next to
compare the effect of having different type of non-dominant poles. Amongst these, both time and
frequency domain performance measures are evaluated, however both of them are neither specified
nor can be guaranteed together, under the present design approach, as also shown in (Das et al. 2011).
However for a fair comparison we evaluate different control performances with the PID controller
gains obtained using different non-dominant pole types e.g. gain and phase margin (Gm and Φm) and
gain crossover frequency (ωgc) controlling the overshoot vs. speed of operation (Das et al. 2011), peak
sensitivity and complementary sensitivity (Ms and MT) (Åstrӧm & Hägglund 2004), 2 norms for
10
tracking or command following mode (servo) and disturbance rejection (regulatory) mode (Alcántara
et al. 2013)(Arrieta et al. 2010).
Let, Gol s be the open-loop transfer function comprising of the time-delayed SOPTD system
G s with PID controller C s which can be represented as:
Gol s C s G s .
(16)
The basic PID control loop with different inputs e.g. set-point (r), disturbance (d) and noise (n)
and the measurement points e.g. error (x1), control signal (x2) and noisy process variable (x3) are
shown in Figure 1. To guarantee internal stability and also for evaluating different performance
measures of a feedback control loop the following nine transfer functions in (17) play a major role
(Doyle et al. 2013)
x1 e
1
x u d 1 C
2
1 GC
x3 y n
GC
G
1
G
1 r
C d .
1 n
(17)
These matrix shows the effective transfer functions from different inputs to various measurement
points. Amongst the nine, four transfer functions play the central role to characterise the control
system performance (Doyle et al. 2013)(Das & Pan 2014)(Herreros et al. 2002), i.e. the sensitivity
S e s , complementary sensitivity T s , disturbance sensitivity S d s and control sensitivity S u s
as follows:
Se s
T s
1
Gre s Gdx2 s Gnx3 s ,
1 Gol s
Gol s
1 Gol s
S d s
G s
Grx3 s , S e s T s 1,
1 Gol s
Gdx3 s , S u s
C s
1 Gol s
(18)
Grx2 s .
These four transfer functions can be quantified using various systems norms ( 2 ) for
e
different system inputs (Herreros et al. 2002). It is clear from (18) that the sensitivity function S s
has three different interpretations. In other words, the sensitivity function signifies the effective
transfer function at all the three measurement points in Figure 1, for set-point, disturbance and noise
inputs, as also revealed from the same diagonal elements in (17). Therefore, in a process control
design, it has been considered as one of the fundamentally important criteria which is often
considered to be norm bounded in various tuning rules like MIGO (M-constrained integral gain
optimization) and approximate MIGO (AMIGO) etc. for PI/PID controller design (Åstrӧm &
Hägglund 2004; Hägglund & Åstrӧm 2004)(Hägglund & Åstrӧm 2002).
First we check the performance J d for a step disturbance input d s and calculating the
2 norms of the disturbance sensitivity function which can be represented by:
J 2d d1 s Sd s 2 , d1 s 1 s ,
(19)
J d d 2 s Sd s , d 2 s 1 s .
11
Using the final value theorem for Laplace transform, the functions with finite time domain
integral are quantified and accordingly the set-point/disturbance inputs are selected between
step/impulse as r, d 1 s or r , d 1 respectively, as the input to different sensitivity functions.
Figure 1: Schematic diagram of the PID control loop with different inputs and measurement points.
The controller output or the control signal leads to the actuator limits and frequent oscillatory
u
inputs to the actuator J which can be represented by the 2 norms of the control sensitivity
function. However, for standard PID controller structure without a derivative filter as being used here
for obtaining 3 controller parameter based 3D stability regions, the control sensitivity becomes an
improper transfer function with more zeros than poles which forbids calculating system norms
directly. As an alternative approach, the 2 norms of the control signal can be computed for a
step input to S u s if it is proper, or impulse input to S u s s if the control sensitivity is improper, as
in the present case. Therefore the control sensitivity norms can be quantified as the 2 norm for
step change in the set-point using inverse-Laplace transform ( 1 ) of S u s s as:
J 2u S u t
J u S u t
2
1 d1 s S u s , d1 s 1 s ,
(20)
2
1 d 2 s S u s , d 2 s 1 s .
e
The set point tracking performance J is analysed using the sensitivity function subjected to step
change in the set-point yielding
J 2e d1 s S e s , d1 s 1 s ,
2
J d2 s S
e
e
(21)
s , d2 s 1 s .
n
The noise rejection performance can be quantified as J using the complementary sensitivity function
subjected to impulse input in the set-point yielding
J 2n d1 s T s 2 , d1 s 1,
(22)
J n d 2 s T s , d 2 s 1.
For all the different sensitivity functions, the corresponding 2 system norms and 2 signal
norms are defined as:
12
1
2
G s 2
g t 2
G j
2
d ,
G s
u t
2
dt ,
g t
sup G j ,
(23)
sup g t .
t
The 2 norms here in (19), (21) and (22) represent large and sustained oscillations in the disturbance
response, error signal and process variable, whereas the norms denote the peak gain of the
frequency response for a particular type of input (servo/regulatory). For stable closed loop system, the
corresponding 2 norms of the sensitivity functions with a chosen type of input excitation need to be
finite and should have a low-pass characteristic i.e. as , the Bode magnitude plot should be
drooping in nature.
Next we also compute the gain margin ( Gm ), phase margin ( m ) and the gain cross-over
frequency ( gc ) of the open loop system which signify the robustness, oscillatory nature and speed of
the closed loop response respectively. These three important quantities for a given process model and
set of PID controller gains can be derived as:
Arg Gol j gc Arg C j gc G j gc m
Gol j gc C j gc G j gc 1
(24)
Gol j pc C j pc G j pc 1 Gm
Arg Gol j pc Arg C j pc G j pc .
4. Simulation and results
4.1.
Test-bench SOPTD processes for performance evaluation
For each of the nine classes of test-bench plants under investigation, we also tabulate which pole
configuration and expression for the PID controller gains yield the largest stability region, as explored
from the number of stabilizing solutions obtained from the Monte Carlo simulations on the chosen
design parameter space. The percentage volume can be represented as the ratio of accepted stable
solutions to the total number of uniformly distributed (105) random samples drawn from the chosen
parameter space. In addition, the correlation amongst the stabilizing controller gains or the design
parameters are also investigated. For example, it can be observed that a high cl can only be achieved
with small m i.e. allowing more effect of the non-dominant poles, thus unnecessarily affecting the
control performance. Similarly, an inverse relation is observed between the stabilizing cl and cl . In
principle, a high value of m is desired which keeps the effect of non-dominant poles on the control
performance as minimum. However it will be evident from the simulation examples presented later
that the high m regions are sparse and is also limiting the design specification on the cl and cl .
The nine classes of test-bench processes and their characteristics are as follows:
i)
Lag-dominant (L<T) underdamped (ζol<1) process (Jahanmiri & Fallahi 1997)
G1 (s )
1
e s ,
9s 2.4s 1
(25)
2
with K = 1/9, L = 1, T = 3, ζol = 0.4, L/T = 0.333<1.
ii)
Lag-dominant (L<T) critically-damped (ζol=1) process (Hwang & others 1995)
13
G2 ( s)
1
e0.8 s ,
s 2s 1
(26)
2
with K = 1, L = 0.8, T = 1, ζol = 1, L/T = 0.8<1.
iii)
Lag-dominant (L<T) overdamped (ζol>1) process (Pomerleau et al. 1996)
G3 ( s )
e2 s
,
1 10s 1 4s
(27)
with K = 1/40, L = 2, T = 6.3246, ζol = 1.1068, L/T = 0.3162<1.
iv)
Balanced lag-delay (L≈T) underdamped (ζol<1) process (Hwang & others 1995)
G4 ( s)
0.5
e s ,
s 1.2s 1
(28)
2
with K = 0.5, L = 1, T = 1, ζol = 0.6, L/T = 1.
v)
Balanced lag-delay (L≈T) critically-damped (ζol=1) process (Hägglund & Åstrӧm 2004)
G5 (s)
e s
1 s
2
,
(29)
with K = 1, L = 1, T = 1, ζol = 1, L/T = 1.
vi)
Balanced lag-delay (L≈T) overdamped (ζol>1) process (Panda et al. 2004)
e3s
G6 ( s) 2
,
9s 24s 1
(30)
with K = 1/9, L = 3, T = 3, ζol = 4, L/T = 1.
vii)
Delay dominant (L>T) underdamped (ζol<1) process (Wang & Shao 2000)
e 1.2755 s
G7 ( s )
,
3.2158s 2 3.1614 s 3.0568
with K = 1/3.2158, L = 1.2755, T = 1.0257, ζol = 0.5042, L/T = 1.2435>1.
This is a reduced order model of a highly oscillatory higher order
1
G s 2
e s .
s s 1 s 3
viii)
(31)
process
Delay dominant (L>T) critically-damped (ζol=1) process (Thyagarajan & Yu 2003)
G8 (s)
e10 s
s 1
2
,
(32)
with K = 1, L = 10, T = 1, ζol = 1, L/T = 10>1.
ix)
Delay dominant (L>T) overdamped (ζol>1) process (Bi et al. 2000)
G9 ( s )
e 2 s
,
0.12s 2 1.33s 1.24
(33)
with K = 1/0.12, L = 2, T = 0.3111, ζol = 1.7239, L/T = 6.4288>1. This represents an HVAC system
model between fan speed to the supply air pressure control loop.
14
The next sub-section reports the stability regions of each of these nine classes of test-bench processes,
representing different dynamical behaviour with various lag to delay ratio and open loop oscillation
levels.
4.2.
Stability regions and the robust stable PID controller design for the test-bench SOPTD
processes
Out of the 12 possible expressions (3 non-dominant pole types × 4 different coefficient orders of
Laplace variable to find out Kp) for the stabilizing gains, we now identify the stabilizing data points
within a range of design parameters for the expression where the number of stable solutions is
maximum. The stabilizing data-points in the design parameter space m, cl , cl are next projected on
to the controller parameter space K p , Ki , Kd by an one to one mapping using the expressions in (9),
(12) and (15). These sets of stabilizing controller gains are then fed to a clustering algorithm to find
out the centroids of the stabilizing regions, for each of the non-dominant pole types. Next, the
performances of these robust stable solutions are also compared, using various criteria introduced in
section 3.2.
Figure 2: 3D stability region in the design specification and controller parameter space for the three lag-dominant
processes. (top) design parameter space, (bottom) PID controller parameter space.
For the nine test-bench processes in (25)-(33), the stability regions in the design parameter space
(top panels) and PID controller parameter space (bottom panels) are shown in Figure 2-Figure 4
respectively. Out of the 105 uniformly drawn samples in the chosen design parameter space
m, cl ,cl ,
the number of stabilizing solutions are reported in Table 1 for the nine test-bench
processes and different non-dominant pole types and expressions for proportional gain Kp. It is also
evident from Table 1 that the all complex non-dominant pole type and the first (s1) expression for Kp,
yields a larger stabilizing region given by the percentage (%) volume or the number of stable
solutions, obtained from the randomly drawn samples. This is due to the reason that for all four
complex non-dominant poles, they have wider flexibility by adjusting their real and imaginary parts to
stabilize the closed loop system. However, for the two complex/two real case, lesser number of nondominant poles can explore a larger portion of the negative s-plane. Similarly, a more stringent
criterion is imposed for the all real non-dominant poles case, as they are forced to lie only on the
negative real axis and cannot explore the entire negative s-plane, which gives them lesser degrees of
freedom and hence yields smaller stability regions. For each type of non-dominant poles and for all
15
the test-bench processes in Table 1, the expression with highest percentage volume is highlighted in
italics. Next considering the robust stable solution as the centroid of the three largest stability regions
representing different pole types, we now compare the performances of these robust stable solutions.
Figure 3: 3D stability region in the design specification and controller parameter space for the three balanced lag-delay
processes. (top) design parameter space, (bottom) PID controller parameter space.
Figure 4: 3D stability region in the design specification and controller parameter space for the three delay-dominant
processes. (top) design parameter space, (bottom) PID controller parameter space.
It is observed from Figure 2-Figure 4 that the stability regions are wider for high cl with the
two/four complex non-dominant poles, whereas the number of stabilizing data points are only fewer
with a demand of high closed loop damping in the case of all real poles. This implies that the centroid
of the stabilizing all real non-dominant poles will lie in low closed loop damping region and will yield
a less sluggish response than that with the two/four complex non-dominant poles. Also, in the
corresponding controller parameter space, the all real non-dominant pole case yields a higher value of
Kd.
16
Table 1: Number of stable controller gains for the nine test-bench processes using 105 random Monte Carlo evaluations
Process
type
Open loop
damping
Underdamped
Criticallydamped
Lag
dominant
Overdamped
Underdamped
Criticallydamped
Balanced
lag and
delay
Overdamped
Underdamped
Criticallydamped
Delay
dominant
Overdamped
Nature of non-dominant
pole
all complex
all real
two complex/two real
Alternate stabilizing
expressions for Kp
1
s
s2
s3
s4
503
442
33
1
107
108
32
0
447
39
47
0
Max. no.
of stable
solutions
503
108
447
%
Volume
0.503
0.108
0.447
all complex
all real
two complex/two real
all complex
all real
two complex/two real
1401
630
1336
361
62
351
264
330
156
292
11
25
242
167
140
37
76
37
23
22
4
0
0
0
1401
630
1336
361
76
351
1.401
0.63
1.336
0.361
0.076
0.351
all complex
all real
two complex/two real
all complex
all real
two complex/two real
1028
376
1038
1403
645
1363
807
316
128
295
423
204
142
137
108
237
161
140
22
15
5
34
41
13
1028
376
1038
1403
645
1363
1.028
0.376
1.038
1.403
0.645
1.363
all complex
all real
two complex/two real
all complex
all real
two complex/two real
1356
341
1262
933
327
933
1146
32
189
766
199
122
268
123
142
114
123
81
33
38
13
19
30
6
1356
341
1262
933
327
933
1.356
0.341
1.262
0.933
0.327
0.933
all complex
all real
two complex/two real
all complex
all real
two complex/two real
1007
394
992
5043
2462
4640
365
0
313
12
568
884
1389
384
707
2487
1057
2229
180
162
130
349
335
293
1389
394
992
5043
2462
4640
1.389
0.394
0.992
5.043
2.462
4.64
The stabilizing controller gains having the highest percentage volume amongst the four alternative
expressions for Kp in Table 1 are now clustered for each type of test-bench process and non-dominant
pole types. The centroids of the stability region as identified by 10 independent runs of the k-means
clustering algorithm can be considered as the robust stable PID controller gains and are reported in
Table 2 for each type of processes and non-dominant pole type. The median distance of all the
stabilizing samples from the centroid are also reported in Table 2 indicating compactness of the
clusters which is consistently smallest for the all real non-dominant pole type. The projected stability
regions in the 2D pair-wise controller parameter space are shown in Figure 5-Figure 13 along with the
centroids or the robust stable PID controller gains as the red star, for all the test-bench process types
with different lag to delay ratio and open loop oscillation levels. It is evident in most cases that the Kp
and Ki (left panels in Figure 5-Figure 13) are more correlated compared to the other pairs of PID
controller gains.
We now investigate which cluster has the highest spread in the 3D space of PID controller gains.
We also report the median distance between the stabilizing samples and the cluster centroid or the
robust stable PID controller gain. The complex shapes of the stabilizing cluster of data-points indicate
17
a highly skewed distribution of the stabilizing controller gains as revealed from their histograms in
Figure 14. This skewed distribution is a result of different degree of influence of the design
parameters or the equivalent controller gains on the shape of the stability regions for each type of testbench process. This indicates that a different variation of the PID controller gains is possible along
different direction or alternatively different extent of parametric uncertainty in the plant model can be
allowed using the robust stable solution for the PID controller gains (Silva et al. 2007). Almost in all
the cases of the lag-dominant plants, the all real non-dominant poles have the minimum spread (interquartile range) and hence the most compact cluster as revealed from the top panel of Figure 14.
However for the delay dominant and balanced lag-delay plants the spread become comparable but
creates more outliers indicating distant stabilizing regions far away from the centroid which is also
evident from the projected 2D scatter diagrams in Figure 8-Figure 13.
Table 2: Robust stable gains and median distance of the stabilizing samples from the centroid
Process
Type
Open loop
damping
Median distance
21.3949
all real
two complex/two real
all complex
all real
two complex/two real
all complex
all real
7.8779
21.1823
2.6631
3.3457
2.5681
189.5916
39.5518
3.1883
0.5443
-0.5196
0.6856
-0.2923
0.1429
6.2367
0.7876
0.1902
0.1229
0.6052
0.1630
0.0827
0.7632
7.5194
7.4275
0.5173
1.2991
0.8199
15.0279
17.0616
two complex/two real
all complex
all real
two complex/two real
all complex
all real
161.5911
5.2558
6.0959
5.2462
2.2106
2.5821
1.3567
-1.4974
-0.0842
-1.2407
-0.6477
0.3531
0.1650
0.1001
0.5029
0.1305
0.0743
0.3623
15.6003
0.4146
1.5163
0.7598
0.3124
1.0217
two complex/two real
all complex
all real
two complex/two real
all complex
all real
2.2602
134.3780
108.3602
120.4336
9.2729
10.0374
-0.4108
0.4472
2.4796
1.3153
-2.4113
-1.0152
0.1177
0.1281
0.2550
0.1828
0.1106
0.3795
0.4655
3.2445
7.9074
6.5032
0.1404
1.1933
Criticallydamped
two complex/two real
all complex
all real
two complex/two real
all complex
all real
9.8668
1.2044
0.8523
0.9227
0.6347
0.7960
-2.1395
0.0140
-0.2052
-0.5277
-0.7776
-0.2768
0.1349
0.0096
0.0199
0.0179
0.0962
0.1922
0.0955
1.1905
1.3696
0.3647
0.2456
0.7399
Overdamped
two complex/two real
0.6297
-0.5778
0.0984
0.4143
Underdamped
Criticallydamped
Lag
dominant
Overdamped
Underdamped
Criticallydamped
Balanced
lag and
delay
Overdamped
Underdamped
Delay
dominant
Robust PID gains
Kp
Ki
Kd
-0.3412
0.0826
6.5390
Nature of nondominant pole
all complex
18
Figure 5: Stability regions of PID controller gains to control test-bench plant G1 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
Figure 6: Stability regions of PID controller gains to control test-bench plant G2 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
19
Figure 7: Stability regions of PID controller gains to control test-bench plant G3 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
Figure 8: Stability regions of PID controller gains to control test-bench plant G4 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
20
Figure 9: Stability regions of PID controller gains to control test-bench plant G5 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
Figure 10: Stability regions of PID controller gains to control test-bench plant G6 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
21
Figure 11: Stability regions of PID controller gains to control test-bench plant G7 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
Figure 12: Stability regions of PID controller gains to control test-bench plant G8 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
22
Figure 13: Stability regions of PID controller gains to control test-bench plant G9 using different non-dominant pole types as
(top) all complex poles, (middle) all real poles, (bottom) two real and two complex poles.
Figure 14: Distance of the stabilizing samples from the centroid for the nine test-bench processes, indicating compactness of
the clusters of data points in the respective stability regions.
4.3.
Performance of the robust stable solutions for the test-bench SOPTD processes
In this section, the control performance of the robust stable solutions are quantified for all the nine
test-bench processes using the signal norms ( 2 ) or system norms ( 2 ) of different
sensitivity functions, as introduced in section 3.2. The performances are quantified in terms of the
disturbance rejection (d), control effort (u), measurement noise filtering (n), tracking error (e), gain
margin (Gm), phase margins ( m ) and gain cross-over frequency ( gc ). Many recent literatures
argued that in PID control loops, the disturbance rejection and the control signal affecting the actuator
size are considered to be the two most significant criteria (Doyle et al. 2013). However apart from the
23
sensitivity function based 2/∞-norms, the traditional performance measures like Gm , m , gc are also
compared in Table 3 that may help in understanding the robustness to gain variations, oscillations and
the speed of the closed loop system respectively, for different robust stable solutions corresponding to
the three non-dominant pole types.
Table 3: Control performance with the robust stable PID controller gains for the nine classes of test-bench processes with
fixed process parameters
Process
type
Open loop
damping
Nature of nondominant pole
all complex
Underdamped
all real
two complex/
two real
all complex
Criticallydamped
all real
two complex/
two real
all complex
Lag
dominant
Overdamped
all real
two complex/
two real
all complex
Underdamped
all real
two complex/
two real
all complex
Criticallydamped
all real
two complex/
two real
all complex
Balanced
lag and
delay
Overdamped
all real
two complex/
two real
all complex
Underdamped
all real
two complex/
two real
all complex
Criticallydamped
all real
two complex/
two real
all complex
Delay
dominant
Overdamped
all real
two complex/
two real
Performance measures of the robust stable PID controller with nominal process parameters
J 2d
J d
J 2u
J u
J 2n
J n
J 2e
J e
Gm
m
gc
3.3044
19.9771
82.7783
4.4093
0.7415
1.0458
3.2822
19.1229
2.3691
57.1641
0.0624
0.4736
1.2697
65.8704
9.4701
1.2351
2.2471
1.3674
2.5484
1.7022
28.6972
0.9056
1.4191
6.0633
56.2284
6.6734
0.8891
1.3416
1.6370
5.7103
2.0109
54.0378
0.8813
3.3439
14.0064
78.7187
1.4461
0.5511
2.0063
3.5184
14.4937
1.6132
38.4221
0.1305
0.8524
2.0666
37.1731
2.3733
0.8002
1.0695
1.1890
2.3550
2.4179
62.6255
0.4096
2.3366
8.3256
67.7462
1.2429
0.5160
1.2158
2.5147
8.6536
2.5592
49.7392
0.1481
2.5287
20.5746
76.7873
5.7889
0.4322
1.1142
3.1520
23.3761
2.6707
56.9813
0.0459
0.3742
1.3103
70.3277
13.5142
0.7863
1.7215
1.8449
4.1672
1.8411
37.6897
0.4085
1.2487
7.3045
37.4806
7.4410
0.4795
1.0250
2.0481
8.9502
2.4889
90.0900
0.0897
3.6135
17.8317
177.2366
3.5547
0.8981
4.1980
7.2048
35.4755
1.2433
32.0756
0.0740
0.8637
2.3685
78.2454
2.2434
0.5553
1.0199
1.7981
4.6522
3.0004
59.6834
0.2182
2.5086
10.3044
137.8486
2.7726
0.6468
2.2735
5.0116
20.4856
1.4595
41.4284
0.0803
5.0056
24.4705
84.7916
1.5971
0.6311
2.8661
5.1543
24.9874
1.3805
34.1570
0.0926
1.2026
3.2422
41.9556
1.3970
0.6216
1.0275
1.4781
3.4699
2.7956
62.0941
0.2782
3.0487
11.6793
63.3211
1.2953
0.4548
1.4710
3.2131
12.0332
1.9285
45.5859
0.1198
1.9220
14.1479
94.6679
1.8423
0.2737
1.6764
3.7439
26.1029
7.9973
34.8646
0.0647
0.8470
4.3265
57.1623
9.4263
0.4782
1.1939
2.2895
9.3214
3.0638
59.7635
0.1123
1.2354
7.4122
72.6952
6.0139
0.3896
1.3105
2.7746
14.7869
3.8712
49.1662
0.0813
3.0527
16.3049
271.0282
5.9197
1.0286
5.2373
9.2686
49.4705
1.1931
29.9875
0.0589
0.9421
2.9722
124.1419
3.5254
0.4522
1.1578
2.8602
8.9635
2.1639
53.6541
0.1262
2.2789
10.0616
108.0883
4.6069
0.7963
3.2323
6.9138
30.4903
1.3145
37.2722
0.0618
7.4984
103.7771
64.5575
1.4312
0.7079
1.0000
7.5587
103.7771
2.5635
84.2525
0.0095
6.2557
50.2831
24.6987
1.6705
0.8474
1.0021
6.3342
50.2831
2.0175
64.5762
0.0198
8.9267
71.3828
59.5067
1.3151
0.3884
1.8875
8.9794
71.4827
1.5833
43.7891
0.0209
4.0000
19.1321
116.2718
1.9777
0.8174
2.8856
5.0481
23.9850
1.3746
34.3820
0.0965
1.9469
6.5425
59.9512
4.2854
1.5701
1.1506
2.5436
8.2736
1.9999
51.8288
0.1446
3.1262
12.5264
77.9316
1.5506
0.9072
1.5167
3.9621
15.6585
1.7872
47.1110
0.0865
The corresponding time domain responses for control variable due to step change in set-point (left
top), step change in disturbance input (right top), manipulated variable or the control signal for step
24
change in set-point (left bottom) and control signal for step change in disturbance input (right bottom)
are shown in Figure 15-Figure 23, for the nine test-bench processes. Both J 2d and J d are found to be
the smallest for the robust stable PID controller with the all real non-dominant pole type, as evident
from Table 3 as well as the right top panel of Figure 15-Figure 23.
Figure 15: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G1
Figure 16: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G2
The all complex non-dominant pole case has the worst disturbance rejection performance and the
two-complex/two-real case falls in between these two cases. The tracking performance is also the best
for the all real non-dominant poles as revealed from the values of J 2e and J e in Table 3 and the
responses of the control variables, subjected to a step change in the set-point in the left top panels of
Figure 15-Figure 23. For the all complex non-dominant poles, the step response is sluggish and highly
oscillatory, amongst the three types of non-dominant poles. However the best tracking and
disturbance rejection performance of the all real non-dominant poles comes at the cost of increased
25
actuator size ( J u ) or higher control effort ( J 2u ) or both as reported in Table 3 and the two bottom
panels of Figure 15-Figure 23, for different test-bench processes. The noise rejection performance is
also not always the best using the all real non-dominant poles, as evident from J 2n and J n in Table 3.
The gc in Table 3 is always found to be the highest using the all real non-dominant poles, yielding a
faster time response. But this increased speed of the robust stable solution comes at the cost of
reduced gain and phase margin ( Gm , m ) in some cases. Therefore, as a summary, control systems
where the disturbance rejection, speed of set-point tracking are of utmost importance, the all real nondominant pole based robust stable PID controller can be employed. In addition, the all real nondominant poles also produce the best phase margin particularly for the balanced lag-delay family of
processes, as evident from Table 3.
Figure 17: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G3
Figure 18: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G4
26
Figure 19: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G5
Figure 20: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G6
However some of the above mentioned measures in Table 3, may indicate a similar performance
and hence a performance correlation analysis as shown in Figure 24 may reveal which measures are
inter-related and which are not. A threshold based analysis of the correlation coefficients (R) of the
robust stable performance measures shows that two pairs viz. J 2d , J d and J d , J e are highly
correlated with R>0.9, hence any one of them within the pair would be a sufficiently independent
measure of the closed loop performance. This is inherently different from the trade-off based design
approach for PID control loops, where it is usually considered that the performance measures are
always independent of each other. Therefore, the correlation plots in Figure 24 for the all real nondominant pole case with best achievable overall performance indicate that improvement in one
performance criterion may not always lead to deterioration of the others. In recent years, there have
been many works e.g. (Das & Pan 2014)(Herreros et al. 2002; Pan & Das 2012; Pan & Das
2013)(Hajiloo et al. 2012) which have used multi-objective optimization for controller design by
27
minimizing multiple conflicting objectives together. However, in most of these cases, the performance
correlation analysis was not shown to illustrate whether a chosen set of cost functions or performance
measures are at all independent which is an important criterion to judge before designing such
optimization-based control systems.
Figure 21: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G7
Figure 22: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G8
The performance correlation analysis in Figure 24 is particularly important in controller design
tasks since there are many recent attempts with an aim of optimization based PID controller tuning
using clearly redundant and highly correlated performance measures or cost functions which should
have otherwise gone through such performance correlation analysis first, before applying heuristic
optimization algorithms on them and comparing marginal performance improvement amongst various
competing global optimizers. As prominent examples of such redundant and correlated cost function
based PID controller design can be noted in (Panda et al. 2012) and elsewhere. In particular, some
approaches reported a comparison of different cost functions with different units and also having
28
different ranges which are unjustified e.g. in (Panda et al. 2012). Moreover, for time domain tracking
comparisons, usually integral performance indices like integral of absolute error (IAE), integral of
squared error (ISE), integral of time multiplied absolute error (ITAE) and integral of time multiplied
squared error (ITSE) capture the combined effects of overshoot, steady state error, rise-time, settlingtime and peak-time and do not need to be added separately in the cost function unlike the approaches
reported in many recent works e.g. (dos Santos Coelho 2009)(Zhu et al. 2009)(Ramezanian et al.
2013)(Sahib 2015)(Bendjeghaba 2014)(Zeng et al. 2015). Therefore including these time domain
features either in the cost function along with an integral performance criteria can be considered
redundant and unnecessary as previously adopted in (Zamani et al. 2009)(Aguila-Camacho & DuarteMermoud 2013) or formulating customized cost functions as weighted average of many correlated
cost functions (Das et al. 2011).
Figure 23: Controlled variable (top) and manipulated variable (bottom) due to step change in set-point (left) and
disturbance input (right) for process G9
Figure 24: Correlation analysis of the 11 performance measures for the robust stable solutions of the test-bench processes.
29
In some cases, performance criteria in time vs. frequency domain can be expected to be highly
correlated to each other from physical understanding of the underlying control principles e.g.
overshoot vs. phase margin (Zamani et al. 2009). Another optimization based approaches in (Hasanien
2013) also reported tuning subset of controller parameters leaving the integral gain, optimization on
fitted polynomials sampled on limited number of points in the search space, and even reporting nonzero steady state error in PID control loops which are unrealistic and limiting. However, our approach
is not only an improvement over these literatures, reporting different system/signal norms for the
robust stable solutions using a standardised set of traditional control loop performance measures but
also investigates how much redundant information is embedded within these chosen set of control
performance quantification measures. Various performance measures were also previously used in
(Zamani et al. 2017)(Chen et al. 2014)(Das & Pan 2014)(Pan & Das 2012; Pan & Das 2013), where
explanations for such possible conflicts amongst a chosen set of cost functions can be found from
physical understanding of the controlled, manipulated and other state variables as well as their role in
the cost function formulation e.g. speed (gain crossover frequency) vs. overshoot (phase margin)
trade-off, tracking vs. control effort trade-off, servo (tracking) vs. regulatory (disturbance rejection)
trade-off etc. (Das & Pan 2014)(Alcántara et al. 2013). However, for multiple objective based control
system design, a negatively correlated pattern or conflicting nature amongst the chosen cost functions
is a necessary criterion to balance various hidden aspects of the overall achievable control
performance, apart from finding only robust stable solutions what the present paper proposes.
4.4.
Invariance of the order of Pade approximation on the control performance
Next, we show the effect of changing the order of Pade approximation using the best found robust
controller i.e. using the all real non-dominant pole configuration which is found to have the fastest
set-point tracking as well as disturbance rejection performances. The Pade approximation for the time
delay term e Ls can be represented in a general form (Silva et al. 2007), with user defined order
(NPade) as:
e Ls N r Ls Dr Ls ,
2r k ! Ls k ,
k 0 k ! r k !
r
2r k ! Ls k ,
Dr Ls
k 0 k ! r k !
r
N r Ls
(34)
where r is the order of the Pade approximation i.e. NPade = r.
Using (34), the Pade approximation for different orders of NPade = {3, 5, 7, 9} becomes:
L3 s3 12 L2 s 2 60 L 120
L3 s 3 12 L2 s 2 60 L 120
L5 s5 30 L4 s 4 420 L3 s 3 3360 L2 s 2 15120 Ls 30240
5 5
L s 30 L4 s 4 420 L3 s 3 3360 L2 s 2 15120 Ls 30240
L7 s 7 56 L6 s 6 1512L5 s 5 25200 L4 s 4 277200 L3 s 3 1995840 L2 s 2 8648640 Ls 17297280
7 7
L s 56 L6 s 6 1512 L5 s 5 25200 L4 s 4 277200 L3 s 3 1995840 L2 s 2 8648640 Ls 17297280
L9 s 9 90 L8 s8 3960 L7 s 7 110880 L6 s 6 2162160 L5 s 5 30270240 L4 s 4 302702400 L3 s 3
2075673600 L2 s 2 8821612800 Ls 1.76432256 1010
.
L9 s 9 90 L8 s8 3960 L7 s 7 110880 L6 s 6 2162160 L5 s 5 30270240 L4 s 4 302702400 L3 s 3
2 2
10
2075673600 L s 8821612800 Ls 1.76432256 10
(35)
eN Ls
Pade 3
eN Ls
Pade 5
eN Ls
Pade 7
eN Ls
Pade 9
30
Figure 25: Set-point and disturbance response for the lag-dominant processes with different order of Pade approximation.
Figure 26: Set-point and disturbance response for the balanced lag-delay processes with different order of Pade
approximation.
Now, the time delay term of all the nine types of test-bench processes have been approximated by
(35). Approximation of the delay term using (35) not only increases the overall system order but also
changes the process dynamics due to increased number of root locus branches. Now by increasing the
Pade order as NPade = {3, 5, 7, 9}, the nine test-bench processes are controlled by the respective robust
PID controllers which was originally designed using a third order approximation (NPade = 3) and the
deviation in the performance measures have also been noted. The set-point tracking and disturbance
rejection tasks for step input remain almost unchanged with the robust-stable PID controller even with
higher order of Pade approximation which are shown for the three classes of processes in Figure 25Figure 27 respectively. The lag-dominant and balanced lag-delay plants show almost imperceptible
difference with increasing order of NPade. However, for delay dominant processes, at the initial dead
time phase of a step-input response, the number of peaks and troughs in the initial oscillations might
31
vary depending on the delay approximation order NPade, although the overall dynamical characteristics
do not change significantly due to the dominant pole placement criteria imposed in the controller
design phase.
Figure 27: Set-point and disturbance response for the delay-dominant processes with different order of Pade
approximation.
Figure 28: Closed loop pole-zero maps for dominant pole placement of test-bench processes with increasing order of Pade
approximation NPade = 1 to 10.
Although in the design process the four roots of equation (10) are considered to be real, however
due to the over-determined nature of the problem having 4 alternative solutions for Kp in (12), this
would yield different stability regions, amongst which the largest one is selected as discussed before.
These sets of stabilizing set of design parameters {m, ζcl, ωcl} are then mapped on to the controller
parameter {Kp, Ki, Kd} space where the clustering is carried out to find out the robust stable solution
with the largest stability region. However, using this controller gain does not always guarantee that
32
the characteristic polynomial involving open loop process parameters and three controller gains upon
factorization, would always yield real poles. The assumption of different type of non-dominant poles
are just to get many alternative candidate expressions for controller gains, so that one can choose the
most robust stable solution while also maintaining acceptable closed loop performance.
One of the biggest hurdles of pole placement controller design for time delay systems is any
parameterized approach of coefficient matching of the characteristic polynomial is usually infeasible,
due to having infinite number of roots. Therefore, the approximation method and order for the delay
term need to adapt to the unknown number of closed loop poles. In order to keep the dominant
dynamics intact upon variable order of Pade approximation, the effect of the rest of closed loop poles
can be minimized using the non-dominance criteria, described above. To verify this, we have shown
here the closed loop pole locations with different orders of Pade approximation. The true delay
manifests itself as a very high order transfer function, thus creating many interlaced poles and zeros
which the PID controller design should be able to handle as reported in Figure 28 for NPade varying
from 1 to 10. It is clear from the pole-zero map in Figure 28 that the non-dominant pole is far away
from the rest as expected and the closed loop poles with increasing order of Pade approximation move
almost along a constant damping line such that the closed loop performances remain unaltered which
are also revealed from the set-point tracking and disturbance rejection responses in time domain
(Figure 25-Figure 27). It is also observed from Figure 28 that with the even order of NPade, the nondominant pole becomes real and for odd NPade, the non-dominant poles become complex conjugates,
creating different patterns of closed loop pole locations for even and odd orders of NPade.
4.5.
Performance of the robust stable controller with perturbed plant parameters
In this section we now investigate the performance variations considering the process parameters
to be uncertain around the robust stable solutions which were found as the centroids of the stability
regions using k-means clustering. This was obtained by random sampling of the design parameter
space m, cl , cl and then mapping the accepted stable solutions in the three PID controller
parameter space K p , K i , K d , followed by applying k-means clustering algorithm to find out the
centroid of the stability region in the 3D space of controller gains. These robust stable solutions are
considered as the mid-points of the images showing the iso-performance contours where the midpoint represent the performance with fixed robust stable PID controller gains on the nominal process
model parameters as reported in section 4.1. Moving towards a better performance as often done in
optimization based controller design approaches using single or multiple control objectives may not
necessarily yield a robust stable solution (Pan & Das 2016). Rather here we take a different route and
design a robust stable solution first (as the centroid of the stability region in controller parameter
space) followed by investigating the resulting performance improvement or degradation for uncertain
process parameters using this robust stable solution. The maximum allowable positive/negative
perturbation in all the parameters are determined until at least one out of the 1000 randomly sampled
values yields unbounded performance measure i.e. 2/∞-norms or gain/phase margins, within a chosen
range of process parameters around the robust stable point (designed with the fixed nominal process
model parameters) i.e. L Lmin , Lmax , T T min , T max , ol olmin , olmax .
This is justified from the fact that for linear control systems, the parametric uncertainty of the
process under control can be equivalently expressed as uncertainty in the controller gains (Silva et al.
2007). Therefore, a robust stable solution at the centroid of the controller parameter space is expected
to handle equivalent model parameter uncertainties, around its nominal values. It has been found that
for the lag-dominant processes a maximum of 40-60% perturbations in all three parameters L,T , ol
can be allowed without encountering an unstable solution out of 1000 randomly sampled points. Also,
for balanced lag-delay processes a range of maximum 50-70% perturbation can be allowed. However,
33
this range significantly reduces to 20-30% for delay dominant processes, since stabilizing these
processes are more difficult. The variation with lag (T) or open loop natural frequency (ωol) with other
parameters like L, ol have been found to be smoother compared to the other combinations. The
highly heterogeneous performance contours are found for uncertain parameter pairs of L, ol where
using the robust stable PID controller, the performance contours has many local maxima or minima as
revealed from Figure 29-Figure 37 for the nine test-bench processes. In all the contour plots a reddish
hue represent a higher value and bluish hue represent lower value.
Figure 29: Iso-performance contours as joint distribution of 40% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G1.
Figure 30: Iso-performance contours as joint distribution of 60% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G2.
34
The 2D contour plots in Figure 29-Figure 37 for all the nine types of processes are explored with
parameter variations around their respective robust stable controllers, designed with the nominal
process parameters, using a perturbed range in the open loop parameters L,T , ol . These are helpful
to get an idea about how each process parameter affects the 11 different performance criteria,
mentioned above. This can be justified from almost parallel lines either along the x-axis or y-axis
which is found in several 2D sub-plots of certain process parameter pairs. For lag-dominant processes
it can be observed that the performance criteria Gm , m , gc are more influenced by the delay (L)
compared to the rest two process parameters
T , ol
as shown in Figure 29-Figure 31. The
Gm , m can be maintained to higher value (for better robust stability and
performance) with smaller L,T for the lag-dominant processes. Whereas gain crossover frequency (
gain/phase margins
gc ) can be increased to achieve faster time response with a lower value of damping ( ol ) for the lagdominant processes. Also, for the lag-dominant processes a lower value of delay helps maintaining a
smaller 2/∞-norms.
Figure 31: Iso-performance contours as joint distribution of 60% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G3.
For the balanced lag-delay processes (G4-G6) most of the performance criteria are highly
influenced by the lag (T) compared to the rest two parameters L, ol which are observed from almost
parallel lines along the axes with T in Figure 32-Figure 34. Also, a large phase margin ( m ) can be
maintained for low delay (L). Similar to the balanced lag-delay processes, the delay-dominant process
(G7-G9) the lag parameter (T) has more influence to all the 11 performance criteria as shown in Figure
35-Figure 37, compared to the rest two process parameter L, ol . Moreover, low delay (L) has got
mild effect to achieve a higher phase margin ( m ) and hence less overshoot for the delay-dominant
processes.
35
Figure 32: Iso-performance contours as joint distribution of 60% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G4.
Figure 33: Iso-performance contours as joint distribution of 70% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G5.
36
Figure 34: Iso-performance contours as joint distribution of 50% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G6.
Figure 35: Iso-performance contours as joint distribution of 30% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G7.
37
Figure 36: Iso-performance contours as joint distribution of 20% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G8.
Figure 37: Iso-performance contours as joint distribution of 30% perturbed process parameters {L, T, ζol} with the robust
stable PID controller having all real non-dominant poles for process G9.
4.6.
PID controller tuning rule generation for robust stability
The robust stable PID controller gains as the centroid of the stabilizing cluster of points for each
test-bench processes can now be condensed in the form of polynomial functions or tuning rules
involving the process parameters like delay to time constant ratio (L/T) and open loop damping (ζol),
while following a similar method adopted in (Das et al. 2015) using polynomial regression analysis on
the open loop process model parameters. However in this case the criteria is changed from several
38
optimal performances to the robust stable solutions unlike (O’Dwyer 2009)(Das et al. 2015). The
tuning rule generation for robust stability can be summarized as (36) for mapping the open loop
process parameters in (25)-(33) on to the controller gains in Table 2. Owing to the lesser number of
process models (9 plants) and in order to obtain a consistent estimate, the maximum order for Kp
tuning rule is limited to 2 in both the free parameters L T , ol while for K i , K d tuning rules, a
second order polynomial lag to delay ratio (L/T) and first order polynomial in damping (ζol) are
considered along with the interaction terms between them, as shown in (36).
K
p
, Ki , K d f L T , ol K
K p p00 p10 L T p01 ol p20 L T p11 L T ol p02 ol2
2
Ki , K d p00 p10 L T p01 ol p20 L T
2
(36)
p11 L T ol
Table 4: Model parameter for tuning rule generation and fitting statistics
PID parameters
Kp
Ki
Kd
Order of L/T
2
2
2
Order of ζol
2
1
1
RMSE
1.3263
0.0706
4.4964
R2
0.8762
0.9629
0.6693
Adjusted R2
0.6698
0.9258
0.3386
This final model order selection (amongst many other combinations) to develop the optimum
tuning rule for the PID controller gains satisfying all the nine test-bench plants was carried out using
the adjusted coefficient of determination (Adj R2) criteria as shown in Table 4. This is a preferred
criterion since it penalizes both model complexity and fitting error, rather than using the simple
criteria like root mean squared error (RMSE) and R2 as both of them are prone to overfitting small
datasets. It has been found that the best fitting performance ( Adj R 2 1 ) is found for the integral gain
Ki. The best fitted coefficients of the tuning rule (36) for the three PID controller gains are given in
(37) along with their respective uncertainties.
Kp :
p00 5.4815 2.648, p10 8.2986 2.2708, p01 1.7928 3.0842,
p20 0.5475 0.14441, p11 2.1487 0.81942, p02 0.68766 0.59573;
p00 1.1271 0.10005, p10 0.64476 0.12037, p01 0.25487 0.045225,
Ki :
p20 0.036846 0.0076874, p11 0.19132 0.041707;
Kd :
(37)
p00 17.247 6.3708, p10 17.696 7.6642, p01 2.8946 2.8796,
p20 1.1552 0.48948, p11 4.8674 2.6557.
4.7.
Computational complexity of the algorithm
In the overall design method proposed in this paper, there are mainly two steps where the
computational complexity needs to be considered:
i)
ii)
Random Monte Carlo sampling to get the stability regions
Using the k-means clustering algorithm on the stabilizing data points in the 3D space of
PID controller gains
Theoretical computational complexity analysis for Monte Carlo algorithms may be quite involved
(Troyer & Wiese 2005). We have therefore provided a numerical study of the computational
complexity, similar to the work reported in (Karlsson et al. 2005) for Sequential Monte Carlo
algorithms. A similar numerical simulation based approach has been adopted for the complexity
analysis of the k-means clustering algorithm as shown in (Velmurugan & Santhanam 2010).
39
Figure 38: Scalability of the Monte Carlo sampling with increasing sample number for plant G1 and all three non-dominant
pole types.
Figure 39: Convergence of the percentage volume of the stability region with increasing sample number for plant G1 and all
three non-dominant pole types.
The simulations were run on a 64-bit Windows PC with 16 GB memory and an Intel I5, 3.40 GHz
processor on 4 parallel cores. The run times of the Monte Carlo sampling increases with the number
of samples drawn within the chosen interval of the design parameters which are compared in Figure
38 for three different non-dominant pole types. The corresponding percentage volume of the stability
region converges to their final value as the number samples increases which is shown in Figure 39.
The k-means clustering algorithm also takes slightly longer time when finding the centroid in a larger
volume of datasets which is also shown in Figure 40. It is important to note that here the sampling or
clustering algorithms are employed in a 3D parameter space. Usually for the same dimension and
number of data-points, increased number of compact clusters in the data-set may increase the
computational time of the k-means algorithm (Velmurugan & Santhanam 2010). Similar to the
40
previous cases, here as well 10 runs of the k-means algorithm were carried out and the best results
with minimum distance criteria was selected in order to avoid the effect of randomized searching from
different initial conditions.
Figure 40: Time requirements for the k-means clustering algorithm with increasing sample number for plant G1 and all three
non-dominant pole types.
4.8.
Hypothesis testing using the sampled data for different processes and non-dominant
pole types
As shown in Table 1, there can be 27 possible combinations for different process types based on
lag to delay ratios (lag-dominated, balanced and delay-dominated), damping levels (under-damped,
critically-damped and over-damped) and non-dominant pole types (all complex, all real and mixed
type). It would be interesting to study whether these groups have significant effect on the sampled
data-points which represent the stability regions for each case. Here, the number of factors is 27
(different combinations of plant and non-dominant pole types) and there are 6 response variables (
K p , Ki , K d , cl , cl , m ). The Multivariate Analysis of Variance (MANOVA) compares multivariate
group means with two or more continuous response variables. This helps in avoiding the chance of
inflated type-I error in group assignment compared to carrying out many univariate analysis of
variance (ANOVA) tests on each response variable (Field et al. 2012). The simple ANOVA also
would have ignored any relationship between the response variables which can be seen from the joint
scatter plots reported in the earlier sections amongst the set of PID controller gains and the design
parameters. Before, the MANOVA test was conducted, it is worthwhile to look at the underlying
distributions of the datasets representing the stability regions in the joint parameter space of PID
controller gains and three design variables. The group-wise scatter plots in Figure 41 show that there
are some partial overlap between the 27 groups. However, the hypothesis testing will help us to
answer the question whether the mean of these distributions are different or not.
For the MANOVA test, usually there are some assumptions about the nature of the dataset e.g.
independence, multivariate normality and homogeneity of the covariance matrices. The normality of
the datasets are violated as observed from the scatterplots in Figure 41 and also in the earlier sections
as well as more formally from the Kolmogorv-Smirnov test on the univariate datasets out of the
multivariate one ( K p , Ki , K d , cl , cl , m ) or even a multivariate normality test. The independence
condition is violated if the data comes from the same process which is the case here. Therefore, here
41
we need to use the repeated measure MANOVA. Four different test statistics have been used to
quantify repeated measure MANOVA results i.e. Pillai-Bartlett trace (V), Hotelling-Lawley trace (U),
Wilk’s lambda (Λ), Roy’s largest root statistic (Θ) as given in (38).
V trace Qh Qh Qe
Qe
Qh Qe
1
,
i
1
,
1 i
(38)
U trace Qh Qe1 i ,
max eig Qh Qe1 .
In (38), the hypothesis sum of squares and products matrix ( Qh ) and the residuals sum of squares and
products matrix ( Qe ) are given as:
Qh T T Z 1T
(39)
Qe C T RT R C,
.
where, R Y X B
The other parameters in (38) are defined as:
D
T ABC
(40)
Z A X T X AT ,
1
with ABC D , B is the matrix of coefficients, A is the matrix defining hypothesis based on the
between subject models, C is the matrix defining hypotheses based on the within-subjects model and
D is the matrix containing hypothesized value. More details on the multivariate hypothesis testing can
be found in (Davis 2002).
Figure 41: Group-wise bivariate scatter plots for different process and non-dominant pole types.
It has been suggested in (Field et al. 2012), that even though the conditions like multivariate
normality or homogeneity of the covariance matrices within the groups are violated, some measures
42
can still be useful like the Pillai-Bartlett trace. However the Roy’s largest root statistic may be
affected by the non-normality (especially platykurtic distributions) as well as non-homogeneity of the
covariance matrices as discussed in (Field et al. 2012). The repeated measure MANOVA results are
given in Table 5 which shows that for all the different statistics, the group differences are indeed
significant (p<0.01).
Table 5: Repeated measure MANOVA results for the 27 groups representing different process dynamics (lag to delay ratio
and damping) and non-dominant pole types
Within
Constant
Constant
Constant
Constant
Constant
Constant
Constant
Constant
Between Statistic
Intercept
Pillai
Intercept
Wilks
Intercept Hotelling
Intercept
Roy
group
Pillai
group
Wilks
group
Hotelling
group
Roy
Value
0.76717
0.23283
3.2951
3.2951
0.82663
0.35719
1.3332
0.93328
F Statistic
20238
20238
20238
20238
233.98
270.48
314.92
1102.5
R2
0.76717
0.76717
0.76717
0.76717
0.16533
0.18856
0.2105
0.48274
df1
5
5
5
5
130
130
130
26
df2
30710
30710
30710
30710
153570
151310
153540
30714
p-value
0
0
0
0
0
0
0
0
However, with an assumption of independence, as such each data-point came from different
plants and there being no relation between the data-points which were actually sampled from the same
plant, we can use the standard one-way MANOVA. The one-way MANOVA results for all the 6
observed variables show significant results (p<0.01) for comparing 27 groups with the respective pvalues being p 0, 0, 0, 0, 4.7 10 251 ,8.05 10 38 with degrees of freedom d = 6.
Table 6: Results of univariate Kruskal-Wallis tests on each variable
Variable
Kp
Ki
Kd
ζcl
ωcl
m
χ2 statistic
10132.8
3784.994
3963
3265.832
8316.116
2388.299
p-value
0
0
0
0
0
0
It is common to do a post-hoc analysis after a significant MANOVA test, by carrying out several
univariate ANOVA tests with Bonferroni correction by diving the significance level with the number
of tests conducted. However due the violation of the normality condition, here we carry out the nonparametric version of ANOVA i.e. the Kruskal-Wallis test. This test uses the null hypothesis that each
of the 6 parameters K p , K i , K d , cl , cl , m between different groups come from the same
distribution whereas the alternative hypothesis being that not all the samples come from the same
distribution. The respective -statistics (instead of the F-statistics in standard one way ANOVA) and
p-values have been reported in Table 6 for each of the six controller gain and design parameters,
along with the group-wise box-plots in Figure 42 which clearly shows that their medians are different
as verified from the hypothesis tests. With the Bonferroni correction, these six univariate nonparametric tests have shown significant results (p<0.01) for all the variables amongst
2
K
p
, K i , K d , cl , cl , m , indicating significant group differences amongst the 27 possible
43
combinations. The significantly low p-values show that the tests reject the null hypothesis indicating
a significant group difference.
Figure 42: Univariate box-plots for {Kp, Ki, Kd, ζcl, ωcl, m} as given by the Kruskal-Wallis hypothesis test.
5. Discussions
This paper proposes three types of dominant pole placement based PID controller design methods
for processes with time delays using different criteria for the non-dominant poles i.e. all complex
conjugate, all real and mixture of them. The assumption of different types of non-dominant poles are
used to get many alternative candidate expressions for the PID controller gains, so that the designer
can choose the most robust stable one amongst competing solutions, while also maintaining
acceptable closed loop performance, characterised by 11 different control performance measures
which is followed by a correlation analysis amongst them. Although the present design was primarily
focussed on dominant pole placement technique, depending on the nature of the dominant poles the
closed loop performance may vary drastically. Also, the dominant pole placement parameters
m, cl ,cl
include the designer’s choice of the expected level of non-dominance, closed loop
oscillation and speed of operation respectively. For a range of demanded closed loop parameters
m, cl ,cl , there can be many alternative stabilizing controller gains which can form a
stability
region in the controller parameter space which was found out by randomly sampling the specified
interval of design parameters. The robust stable solution is chosen as the centroid of the stability
region for different constructs of the characteristic equations considering either all real, all complex or
their combinations for the closed loop non-dominant poles. However, as a side effect of such robust
stable design, while satisfying the dominant pole placement criteria, the other control performance
objectives like servo/regulatory, tracking/disturbance rejection, control effort, noise rejection tradeoffs in the form of various system/signal norms as well as traditional frequency domain measures like
gain/phase margin, gain cross-over frequency might take some arbitrary values, rendering
unacceptable or poor overall closed loop performance. It is well known that all these performance
measures cannot be directly controlled in either the pole placement or robust stable design approaches
but may be of prime concern in many industrial process control applications. Therefore our design
approach first finds out the centroid of the stabilizing PID controller gains, thus maximizing
parametric robustness of the process model, since uncertainty in the linear system parameters can be
equivalently represented as uncertainty in the PID controller parameters (Silva et al. 2007). Then with
44
the best robust stable solutions which can be judged either by comparing the highest volume of the
stability region or some other time/frequency domain performance criteria, we investigate rest of the
performance measure as an additional exploration. However, incorporating multiple of these desired
criteria can be better managed under a unified or conflicting control design within a linear or bilinear
matrix inequality (LMI/BMI) approach or by using a multi-objective optimisation which can be
investigated in a future research.
The overdetermined nature of the PID controller design problem having 4 alternative solutions for
Kp in different non-dominant pole-types produce different volume and shape of the stability regions.
Amongst them the largest stability region as indicated by the maximum number of accepted stable
solutions within the same interval of design parameters, is selected. Then k-means clustering
algorithm is used to find out the robust stable PID controller gains within the largest stability region.
It was observed from the previous sections that the robust stable PID controller which has been
designed considering the all real non-dominant pole criteria performs the best compared to the other
non-dominant pole types i.e. all complex conjugate and the mixture of them. Again, the robust stable
PID controller, designed with all real non-dominant poles, is also capable to handle the SOPTD
processes with large parametric uncertainties around their nominal values where this uncertainty has
been considered in the process parameters L, T , ol , instead of the controller gains. Also, simple
tuning rules has been established via polynomial regression analysis to design the PID controller gains
to achieve robust stability using the nine types of test-bench SOPTD processes. Computational
complexity analysis of the two step algorithm has been provided, followed by hypothesis testing for
the distribution of the sampled data-points for different process and non-dominant pole types.
However, in spite of demonstrating the effectiveness of the proposed PID controller design
methodology to handle the stable test-bench SOPTD processes e.g. performance invariance against
order of time delay approximation using Pade method, there are also some limitations of this
approach:
The design has not considered the effect of any strong zero or lead dynamics in the time delay
processes. As the presence of the time delay term in the SOPTD process itself makes the system
to have an infinite order (upon infinite series expansion of the exponential term), the inclusion of
the zero dynamics in the delayed process makes it even more difficult to design dominant pole
placement criteria based PID controller which can be considered as our future scope of research.
Performance contours and stability regions are obtained by random sampling but not through any
deterministic analytical expression.
This design methodology assumes about prior knowledge of the plant model, although in a more
realistic case the plant models can be uncertain either being norm bounded or having varying
parameters within certain intervals. For such a scenario, the robust stable PID controller design
usually should not be based on the dominant pole placement criteria on the nominal system and is
left as future scope of research.
Regarding the main achievement of this paper, we first attempt to not only find the stability
region of an SOPTD process which can be thought of a higher order system under Pade
approximation of the delay term but also, we can maintain a desired equivalent delay-free second
order response, characterised by a user-specified closed loop damping and frequency. In other words,
the effect of undesired poles and zeros due to the Pade approximation of delay term can be
constrained with the non-dominance criteria (controlled by the parameter m) such that even with
arbitrary order of NPade and hence unknown number of poles and zeros in the closed loop system, the
dominant process dynamics can still be maintained near to the desired level which has been
successfully shown in Figure 25-Figure 27. It is a new finding on top of the existing dominant pole
placement PID controller design related literatures mentioned before, since our method is insensitive
45
of the actual number of non-dominant poles/zeros, arising from arbitrary order of Pade approximation
for the delay term and thus both the servo/regulatory modes can be maintained at an acceptable level
for such time-delay systems via a simpler dominant pole placement approach.
6. Conclusion
In order to handle SOPTD processes, dominant pole placement based PID controller tuning has
been considered with three different non-dominant pole types. The k-means clustering technique is
used here to achieve the robust stable PID controller gains inside the stability region in the controller
parameter space, obtained by random Monte Carlo sampling of the chosen pole placement parameter
intervals. Different closed loop performance measures in both time and frequency domain have been
analysed for nine test-bench SOPTD processes with their corresponding robust tuned PID controller
which was originally designed using analytical criteria derived for a third order Pade approximation
(NPade = 3). It has been shown later that the proposed robust-stable PID controllers are able to control
the resulting higher order processes and dynamics of arbitrary number of poles and zeros, due to
expansion of the delay term i.e. increasing the order of the Pade approximation while maintaining a
similar closed loop performance. In future, the proposed methodology can be extended for nonlinear
and uncertain processes. Also, to design a digital control scheme, the proposed methodology can be
further extended in the discrete time version where choice of the sampling time as a fraction of
desired closed loop time constant may play a significant role.
Acknowledgement
KH acknowledges the support from the University Grants Commission (UGC), Govt. of India under
its Basic Scientific Research (BSR) scheme.
References
Aguila-Camacho, N. & Duarte-Mermoud, M.A., 2013. Fractional adaptive control for an automatic
voltage regulator. ISA transactions, 52(6), pp.807–815.
Alcántara, S., Vilanova, R. & Pedret, C., 2013. PID control in terms of robustness/performance and
servo/regulator trade-offs: A unifying approach to balanced autotuning. Journal of Process
Control, 23(4), pp.527–542.
Arrieta, O., Visioli, A. & Vilanova, R., 2010. PID autotuning for weighted servo/regulation control
operation. Journal of Process Control, 20(4), pp.472–480.
Astrom, K.J., Hang, C.C. & Lim, B., 1994. A new Smith predictor for controlling a process with an
integrator and long dead-time. IEEE transactions on Automatic Control, 39(2), pp.343–345.
Åstrӧm, K. & Hägglund, T., 2004. Revisiting the Ziegler-Nichols step response method for PID
control. Journal of process control, 14(6), pp.635–650.
Ǻstrӧm, K.J. & Hägglund, T., 1995. PID controllers: theory, design, and tuning,
Åstrӧm, K.J. & Wittenmark, B., 2011. Computer-Controlled Systems: Theory and Design, Courier
Corporation.
Bendjeghaba, O., 2014. Continuous firefly algorithm for optimal tuning of PID controller in AVR
system. Journal of Electrical Engineering, 65(1), pp.44–49.
Bi, Q. et al., 2000. Advanced controller auto-tuning and its application in HVAC systems. Control
Engineering Practice, 8(6), pp.633–644.
46
Chen, Z. et al., 2014. Design of a fractional order PID controller for hydraulic turbine regulating
system using chaotic non-dominated sorting genetic algorithm II. Energy Conversion and
Management, 84, pp.390–404.
Cominos, P. & Munro, N., 2002. PID controllers: recent tuning methods and design to specification.
IEE Proceedings-Control Theory and Applications, 149(1), pp.46–53.
Das, S., Pan, I., et al., 2012. Improved model reduction and tuning of fractional-order PIλDμ
controllers for analytical rule extraction with genetic programming. ISA Transactions, 51(2),
pp.237–261.
Das, S., Halder, K., et al., 2012. Inverse optimal control formulation for guaranteed dominant pole
placement with PI/PID controllers. In Computer Communication and Informatics (ICCCI), 2012
International Conference on. pp. 1–6.
Das, S. et al., 2011. On the selection of tuning methodology of FOPID controllers for the control of
higher order processes. ISA Transactions, 50(3), pp.376–388.
Das, S. & Pan, I., 2014. On the Mixed H2/H∞ Loop-Shaping Tradeoffs in Fractional-Order Control of
the AVR System. IEEE Transactions on Industrial Informatics, 10(4), pp.1982–1991.
Das, S., Pan, I. & Das, S., 2015. Multi-objective LQR with optimum weight selection to design
FOPID controllers for delayed fractional order processes. ISA transactions, 58, pp.35–49.
Davis, C.S., 2002. Statistical methods for the analysis of repeated measurements, Springer Science &
Business Media.
Doyle, J.C., Francis, B.A. & Tannenbaum, A.R., 2013. Feedback control theory, Courier Corporation.
Ellis, M. & Christofides, P.D., 2015. Economic model predictive control of nonlinear time-delay
systems: Closed-loop stability and delay compensation. AIChE Journal, 61(12), pp.4152–4165.
Engelborghs, K., Dambrine, M. & Roose, D., 2001. Limitations of a class of stabilization methods for
delay systems. IEEE Transactions on Automatic Control, 46(2), pp.336–339.
Field, A., Miles, J. & Field, Z., 2012. Discovering Statistics Using R.
Furukawa, T. & Shimemura, E., 1983. Predictive control for systems with time delay. International
Journal of Control, 37(2), pp.399–412.
Hägglund, T. & Åstrӧm, K.J., 2002. Revisiting The Ziegler-Nichols Tuning Rules For PI Control.
Asian Journal of Control, 4(4), pp.364–380.
Hägglund, T. & Åstrӧm, K.J., 2004. Revisiting the Ziegler-Nichols tuning rules for PI control—part
II: the frequency response method. Asian journal of control, 6(4), pp.469–482.
Hajiloo, A., Nariman-Zadeh, N. & Moeini, A., 2012. Pareto optimal robust design of fractional-order
PID controllers for systems with probabilistic uncertainties. Mechatronics, 22(6), pp.788–801.
Hasanien, H.M., 2013. Design optimization of PID controller in automatic voltage regulator system
using Taguchi combined genetic algorithm method. IEEE Systems Journal, 7(4), pp.825–831.
Herreros, A., Baeyens, E. & Perán, J.R., 2002. Design of PID-type controllers using multiobjective
genetic algorithms. ISA transactions, 41(4), pp.457–472.
47
Holis, R. & Bobál, V., 2015. Model predictive control of time-delay systems with measurable
disturbance compensation. In Process Control (PC), 2015 20th International Conference on. pp.
209–214.
Huang, I.-B. & Li, L.L.-C., 1967. Root locus determination of linear systems with transport lag. IEEE
Transactions on Automatic Control, 12(5), pp.632–634.
Hwang, S.-H. & others, 1995. Closed-loop automatic tuning of single-input single-output systems.
Industrial & engineering chemistry research, 34(7), pp.2406–2417.
Jahanmiri, A. & Fallahi, H., 1997. New methods for process identification and design of feedback
controller. Chemical Engineering Research and Design, 75(5), pp.519–522.
Karlsson, R., Schon, T. & Gustafsson, F., 2005. Complexity analysis of the marginalized particle
filter. IEEE Transactions on Signal Processing, 53(11), pp.4408–4411.
Kiong, T.K., Wang, Q.G. & Chang, C.H., 1999. Advances in PID control, Springer.
Krall, A.M., 1970. The root locus method: a survey. SIAM Review, 12(1), pp.64–72.
Matausek, M. & Micic, A., 1996. A modified Smith predictor for controlling a process with an
integrator and long dead-time. IEEE transactions on automatic control, 41(8), pp.1199–1203.
Michiels, W. et al., 2002. Continuous pole placement for delay equations. Automatica, 38(5), pp.747–
761.
Michiels, W. & Vyhlidal, T., 2005. An eigenvalue based approach for the stabilization of linear timedelay systems of neutral type. Automatica, 41(6), pp.991–998.
Michiels, W., Vyhlidal, T. & Zitek, P., 2010. Control design for time-delay systems based on quasidirect pole placement. Journal of Process Control, 20(3), pp.337–343.
Niculescu, S.-I., 1998. H∞ memoryless control with an α-stability constraint for time-delay systems:
an LMI approach. IEEE Transactions on Automatic Control, 43(5), pp.739–743.
Normey-Rico, J.E. & Camacho, E.F., 2007. Control of dead-time processes, Springer.
O’Dwyer, A., 2009. Handbook of PI and PID controller tuning rules, World Scientific.
Pan, I. et al., 2011. Stabilizing gain selection of networked variable gain controller to maximize
robustness using particle swarm optimization. In Process Automation, Control and Computing
(PACC), 2011 International Conference on. pp. 1–6.
Pan, I. & Das, S., 2012. Chaotic multi-objective optimization based design of fractional order PIλDμ
controller in AVR system. International Journal of Electrical Power & Energy Systems, 43(1),
pp.393–407.
Pan, I. & Das, S., 2016. Fractional Order AGC for Distributed Energy Resources Using Robust
Optimization. IEEE Transactions on Smart Grid, 7(5), pp.2175–2186.
Pan, I. & Das, S., 2013. Frequency domain design of fractional order PID controller for AVR system
using chaotic multi-objective optimization. International Journal of Electrical Power & Energy
Systems, 51, pp.106–118.
48
Panda, R.C., Yu, C.-C. & Huang, H.-P., 2004. PID tuning rules for SOPDT systems: Review and
some new results. ISA transactions, 43(2), pp.283–295.
Panda, S., Sahu, B. & Mohanty, P., 2012. Design and performance analysis of PID controller for an
automatic voltage regulator system using simplified particle swarm optimization. Journal of the
Franklin Institute, 349(8), pp.2609–2625.
Pomerleau, A. et al., 1996. Development and evaluation of an auto-tuning and adaptive PID
controller. Automatica, 32(1), pp.71–82.
Ramezanian, H., Balochian, S. & Zare, A., 2013. Design of optimal fractional-order PID controllers
using particle swarm optimization algorithm for automatic voltage regulator (AVR) system.
Journal of Control, Automation and Electrical Systems, 24(5), pp.601–611.
Rivera, D.E., Morari, M. & Skogestad, S., 1986. Internal model control: PID controller design.
Industrial & engineering chemistry process design and development, 25(1), pp.252–265.
Rogers, S. & Girolami, M., 2012. A first course in machine learning, CRC Press.
Saha, S. et al., 2012. A conformal mapping based fractional order approach for sub-optimal tuning of
PID controllers with guaranteed dominant pole placement. Communications in Nonlinear Science
and Numerical Simulation, 17(9), pp.3628–3642.
Sahib, M.A., 2015. A novel optimal PID plus second order derivative controller for AVR system.
Engineering Science and Technology, an International Journal, 18(2), pp.194–206.
Dos Santos Coelho, L., 2009. Tuning of PID controller for an automatic regulator voltage system
using chaotic optimization approach. Chaos, Solitons & Fractals, 39(4), pp.1504–1514.
De la Sen, M., 2005. On pole-placement controllers for linear time-delay systems with commensurate
point delays. Mathematical Problems in Engineering, 2005(1), pp.123–140.
Shamsuzzoha, M. & Lee, M., 2007. IMC- PID Controller Design for Improved Disturbance Rejection
of Time-Delayed Processes. Industrial & Engineering Chemistry Research, 46(7), pp.2077–2091.
Silva, G.J., Datta, A. & Bhattacharyya, S.P., 2007. PID controllers for time-delay systems, Springer.
Suh, I. & Bien, Z., 1982. A root-locus technique for linear systems with delay. IEEE Transactions on
Automatic Control, 27(1), pp.205–208.
Thyagarajan, T. & Yu, C.-C., 2003. Improved autotuning using the shape factor from relay feedback.
Industrial & engineering chemistry research, 42(20), pp.4425–4440.
Troyer, M. & Wiese, U.-J., 2005. Computational complexity and fundamental limitations to fermionic
quantum Monte Carlo simulations. Physical review letters, 94(17), p.170201.
Velmurugan, T. & Santhanam, T., 2010. Computational complexity between K-means and K-medoids
clustering algorithms for normal and uniform distributions of data points. Journal of computer
science, 6(3), p.363.
Wang, Q.-G. et al., 2009. Guaranteed dominant pole placement with PID controllers. Journal of
Process Control, 19(2), pp.349–352.
49
Wang, Y.-G. & Shao, H.-H., 2000. PID auto-tuner based on sensitivity specification. Chemical
Engineering Research and Design, 78(2), pp.312–316.
Yeung, K. & Wong, W., 1982. Root-locus plot of systems with time delay. Electronics Letters,
18(11), pp.480–481.
Zamani, A.-A., Tavakoli, S. & Etedali, S., 2017. Fractional order PID control design for semi-active
control of smart base-isolated structures: A multi-objective cuckoo search approach. ISA
transactions, 67, pp.222–232.
Zamani, M. et al., 2009. Design of a fractional order PID controller for an AVR using particle swarm
optimization. Control Engineering Practice, 17(12), pp.1380–1387.
Zeng, G.-Q. et al., 2015. Design of fractional order PID controller for automatic regulator voltage
system based on multi-objective extremal optimization. Neurocomputing, 160, pp.173–184.
Zhong, Q.-C., 2006. Robust control of time-delay systems, Springer Science & Business Media.
Zhu, H. et al., 2009. CAS algorithm-based optimum design of PID controller in AVR system. Chaos,
Solitons & Fractals, 42(2), pp.792–800.
50
| 3cs.SY
|
A Solution to the Flowgraphs Case Study using Triple Graph
Grammars and eMoflon
Anthony Anjorin∗
Marius Lauder∗
Technische Universität Darmstadt
Real-Time Systems Lab
Merckstr. 25
64283 Darmstadt, Germany
Technische Universität Darmstadt
Real-Time Systems Lab
Merckstr. 25
64283 Darmstadt, Germany
[email protected]
[email protected]
After 20 years of Triple Graph Grammars (TGGs) and numerous actively maintained implementations, there is now a need for challenging examples and success stories to show that TGGs can be
used for real-world bidirectional model transformations. Our primary goal in recent years has been to
increase the expressiveness of TGGs by providing a set of pragmatic features that allow a controlled
fallback to programmed graph transformations and Java.
Based on the Flowgraphs case study [7] of the Transformation Tool Contest (TTC 2013), we
present (i) attribute constraints used to express complex bidirectional attribute manipulation, (ii)
binding expressions for specifying arbitrary context relationships, and (iii) post-processing methods
as a black box extension for TGG rules. In each case, we discuss the enabled trade-off between
guaranteed formal properties and expressiveness.
Our solution, implemented with eMoflon (www.emoflon.org) our metamodelling and model
transformation tool, is available as a virtual machine hosted on Share [15].
1
Introduction and Motivation
Triple Graph Grammars (TGGs) [13] are a declarative, rule-based, bidirectional model transformation
language and can be used to specify a consistency relation over source and target models, with which
various operational scenarios, such as a forward and backward transformation, can be automatically
supported. As TGGs have been in use for about 20 years and numerous formal results based on algebraic
graph transformation [8] have been established, it is now time to tackle convincingly realistic real-world
case studies with TGGs.
In our opinion, based on practical experience with our TGG implementation in eMoflon [1], what is
necessary is to increase the expressiveness of TGGs by providing a pragmatic set of features that allow a
controlled fallback to a general purpose (transformation) language such as standard Java or programmed
graph transformations via, e.g., Story Driven Modelling (SDM) [4].
Our contribution in this paper is to present a set of advanced TGG features that enable a controlled
integration of SDM and Java code in TGG rules. These features are: (i) attribute constraints for complex bidirectional attribute manipulation in TGG rules, (ii) binding expressions for expressing complex,
possibly recursive context relationships, and (iii) post-processing methods that allow for arbitrary postprocessing after a TGG rule has been applied. Using the case-study Flowgraphs of the TTC 2013 [7],
we present each feature, explaining when it is necessary and discussing the involved trade-off between
expressiveness and formal properties.
∗ Supported by the ‘Excellence Initiative’ of the German Federal and State Governments and the Graduate School of Computational Engineering at TU Darmstadt.
Pieter Van Gorp, Louis M. Rose, Christian Krause (Eds.):
Sixth Transformation Tool Contest (TTC 2013)
EPTCS 135, 2013, pp. 69–74, doi:10.4204/EPTCS.135.8
70
2
A Solution to the Flowgraphs Case Study using TGGs and eMoflon
Solution With Triple Graph Grammars
The Flowgraphs case study of the TTC 2013 is a text-to-model transformation involving the static analysis of a small Java subset. The task is to transform a simple Abstract Syntax Tree (AST) for Java to a
flowgraph, i.e., a model containing explicit information about the control and data flow in the program.
In this context, bidirectionality is a crucial requirement to reflect program manipulations (refactorings,
quick-fixes), applied via model transformations, back in the code. To show that eMoflon supports using TGGs directly with standard parser and unparser technology (including XML), we modify the case
study slightly by using ANTLR [12] and StringTemplate [11] for parsing and unparsing Java code, respectively. Our TGG rules operate directly on the AST produced by ANTLR without a model parser
such as EMFText [5].
A TGG consists of a set of rules describing how source and target models evolve simultaneously.
Each TGG rule consists of elements (objects and links), typed according to the classes and references
in the respective metamodels. The TGG rule MethodRule (Fig. 1) creates the root element of the AST
representing a class containing a single method (on the left hand side), and a Method with an explicit
Exit element in the model (on the right hand side). As can be seen from the rule, the AST metamodel
(not shown explicitly) consists basically of labelled nodes with children and attributes.
MethodRule is referred to as an axiom as it only creates elements (green with a “++” markup) and
does not require any context (precondition). Note how correspondence elements such as nodeToBlock,
depicted visually as hexagons, are created to connect AST elements (classNode) with model elements
(method). These elements are used in other rules as traceability links and conform to a simple correspondence metamodel, connecting source and target metamodels. The triple of source, correspondence,
and target metamodels is referred to as a TGG schema.
Simple attribute manipulation can be expressed as attribute assignments such as txt = “Exit” in the
exit element. Note that this is automatically interpreted as an assertion for the backward transformation,
where the model is parsed and the AST is created.
++
classNode : Node
name := "CLASS"
nodeToBlock :
NodeToBlock
++
parentNode
children ++
++
methodNode : Node
++
++
++
++
exit
++
nodeToExit :
NodeToExit
source
target
++
exit : Exit
target
++
method : Method
txt := "Exit"
txt
node
++ attribute
++
name : Attribute
value
name := "METHOD_NAME"
++
target
++
name := "METHOD"
parentNode
++
++
source
{addSuffix(name.value, "()", method.txt)}
children
statementsNode : Node
source
name := "STATEMENTS"
++
++
nodeToMethod :
NodeToBlock
Figure 1: TGG axiom MethodRule for creating the corresponding tree structure for Methods
statementsNode : Node
source
name == "STATEMENTS"
++
parentNode
children ++
assignmentNode : Node
nodeToBlock :
NodeToBlock
target
++
source
block : Block
++
target
stmts ++
cfPrev
A. Anjorin & M. Lauder
71
More complex attribute manipulation can be specified using bidirectional attribute constraints, such
as the constraint addSuffix(name.value, “()”, method.txt) in MethodRule, which expresses that the
value of the name attribute in the AST must be equal to the txt attribute of the method element in the
model, after adding “()” as a suffix. Attribute constraints such as addSuffix are interpreted appropriately
when compiling the TGG rule for each operational scenario (e.g., forward, backward) and can thus be
expressed in a direction agnostic manner fitting to the rest of the TGG rule.
A second TGG rule AssignmentWithExpRule (Fig. 2) handles, for example, statements of the form
“a = b + 3;” represented in the AST as an assignment tree with lhs = “a”, rhs = “+”, operandL = “b”,
and operandR = “3”. To test the flexibility of the transformation language, the case study requires an
embedded model-to-text transformation to transform the tree structure back to text and store it as the attribute value of the SimpleStmt element in the model. This is accomplished in AssignmentWithExpRule
with a set of attribute constraints, referred to as the attribute Constraint Satisfaction Problem (CSP) of
the TGG rule. How this works is best explained using a concrete example; To handle “a = b + 3;” in
the forward transformation, the set of constraints is sorted and solved as follows:
isAnIdentifier("a") => true
concatWithOperatorSymbol("+", "b", "3", temp1) => temp1 = "b + 3"
concat("=", "a", "b + 3", temp2) => temp2 = "a = b + 3"
addSuffix("a = b + 3", ";", declStm.txt) => declStm.txt = "a = b + 3;"
The constraints isAnIdentifier and concatWithOperatorSymbol are user defined constraints and were
implemented in Java specifically for the Flowgraphs case study. All other constraints are library constraints and are directly available for use in our tool. In this way, the set of constraints can be seamlessly
extended by the user, reused, and combined with other constraints in different TGG rules. The constraint
isAnIdentifier ensures that the rule is not applied for statements such as “int a = b + 3”, where “int a” is
not an identifier and is handled with a different rule for declarations. The second user defined constraint
concatWithOperatorSymbol is an extension of the normal concat constraint and is able to split expressions using a list of supported operators as potential characters for splitting/concatenation as required.
The attribute CSP is sorted and solved differently but analogously for the backward transformation.
Formally, attribute CSPs serve as application conditions for TGG rules, and, as access is restricted
to only attribute values, most formal results for TGGs still hold as we have shown in previous work [2].
Attribute constraints are implemented in Java and thus allow a controlled integration of Java code for
complex attribute manipulation in TGG rules.
Note how the rule requires context elements (black without any markup) that must already exist
and have been created by applying a different rule such as MethodRule. In the model, AssignmentWithExpRule requires (i) the parent block element that is to contain the created assignment statement
declStatement, and (ii) the succeeding element nextFlow in the flowgraph.
Although the created cfPrev and cfNext references are explicit in the FlowGraph metamodel, determining the corresponding next node nextFlowNode in the AST structure is more challenging and
requires a non-trivial recursive search in the AST. In our approach, this search is specified using SDM or
Java, and is integrated in the TGG rule via a virtual link referred to as a binding expression. In AssignmentWithExpRule, the binding expression is the link from assignmentNode to nextFlowNode, which is
visualized as a dashed arrow leading to an object with a bold border. Using binding expressions, complex
context relationships in TGG rules can be indicated by requiring a virtual link that does not really exist
according to the corresponding metamodel. Our tool automatically creates a helper method, which is
invoked when navigating along this link, and can be implemented with SDM or Java.
name := "STATEMENTS"
72
NodeToBlock
++
A Solution to the Flowgraphs Case Study using TGGs and eMoflon
statementsNode : Node
nodeToBlock :
NodeToBlock
source
name == "STATEMENTS"
++
parentNode
children ++
++
++
assignmentNode : Node
source
name := "ASSIGNMENT"
++
parentNode
children
++
++
nodeToSimpleStmt :
NodeToSimpleStmt
target
++
stmts ++
declStatement :
SimpleStmt
cfPrev
rhs : Node
index := 1
index := 0
parentNode
parentNode
++
children
++
operandL : Node
name
block : Block
parentNode
children
++
++
lhs : Node
name
target
index := 0
name
++
++
name
txt
children
operandR : Node
index := 1
++
name
{concatWithOperatorSymbol(rhs.name, operandL.name, operandR.name, temp1)
concat("=", lhs.name, temp1, temp2)
addSuffix(temp2, ";", declStatement.txt)
isAnIdentifier(lhs.name)}
nextFlowNode : source
Node
nodeToFlowInstr :
NodeToFlowInstr
target
nextFlow :
FlowInstr
cfNext
Figure 2: TGG rule AssignmentWithExpRule for creating assignment statements
Formally, binding expressions are equivalent to a separate pre-processing step in which all virtual
links are created explicitly in an instance of an appropriately augmented metamodel. The TGG rules
are then viewed as operating on this augmented metamodel. In practice, however, it is much more
convenient to “find” these links on-the-fly as required. In this manner, binding expressions allow the
controlled integration of SDM or Java code in TGG rules to determine complex context relationships.
As a final feature, every TGG rule can invoke a post-processing method (implemented with SDM
or Java) to perform some final tasks that are difficult or impossible to specify directly in the TGG rule.
In AssignmentWithExpRule, the index of the created assignmentNode has to be set correctly for the
backward transformation, which requires a non-trivial recursive search in the tree structure due to if/else,
loop, and return/break statements.
Post-processing methods are clearly a black-box extension and beyond any formal reasoning. In our
practical experience, however, post-processing can be kept in most cases to a bare minimum by using a
combination of attribute constraints and binding expressions.
A. Anjorin & M. Lauder
3
73
Related Work
Related approaches can be divided into three main groups: (i) Other ways of increasing the expressiveness of TGGs, (ii) Other bidirectional approaches, and (iii) Achieving bidirectionality with unidirectional
transformation languages.
(i) TGGs with a separated pre- and post-processing step: A justified question is if it is not better
to use TGGs with a clearly separated pre-processing and post-processing step instead of our features
for integrating auxiliary methods in the TGG rules. Our approach, however, avoids an extra traversal
through the input and output models as binding expressions can be used as required to determine context
relationships before a TGG rule is applied and a post-processing method to complete the rule after application. In this way, the traversal strategy applied by the TGG control algorithm can be used for on-the-fly
processing as required. Our practical experience is that users tend to regard TGGs as not being worth
the effort if complex pre-processing and post-processing with a separate traversal is necessary. With our
approach, the TGG is more in focus and is used to structure the transformation in an iterative manner.
(ii) Other bidirectional languages and tools: For a detailed survey of bidirectional languages other
than TGGs we refer to [14]. Compared to other bidirectional approaches, TGGs are advantageous
as there exist multiple, actively maintained TGG implementations (interpretative, generative and hybrid approaches) with different strengths and weaknesses. Furthermore, most TGG implementations are
Ecore/EMF/Eclipse compatible and are thus suitable in an MDE context together with other EMF tools
and infrastructure. In a different context, however, other tools are probably more suitable, e.g., for bidirectional XML manipulation [10] and for bidirectional string and tree manipulation [3]. Depending on
the application scenario, it might be more appropriate to derive a backward transformation from a given
forward transformation [6] as opposed to describing the simultaneous build-up of model triples.
(iii) Combination of unidirectional transformation languages: A combination of unidirectional transformation languages as an alternative to a bidirectional language is advantageous for obvious reasons:
Unidirectional transformation languages are typically better established, stable, expressive and have better tool support. We, however, regard transformation languages with explicit support for bidirectionality
as superior as they enable a high-level specification. In the case of TGGs, the specified consistency relation can be used to automatically derive other useful operational transformations to support conformance
testing, consistency checking and link creation for existing source and target models, and incremental
model synchronization [9]. Supporting these scenarios and guaranteeing suitable formal properties [8] is
quite challenging using separate forward and backward transformations.
4
Future Work
In the near future we will establish, compare, and experiment with various ideas for modularity concepts
to improve the maintainability of TGG rules. Efficiency and scalability of transformations with TGGs
are also crucial points that must be improved. We are currently using TGGs for an increasing number of
non-trivial transformations including industrial projects together with Siemens AG and internally (i.e.,
as part of eMoflon) for a bootstrap of a textual syntax for Ecore/SDM/TGGs with TGGs. We aim to
establish a transformation zoo for TGGs consisting of numerous representative examples.
74
A Solution to the Flowgraphs Case Study using TGGs and eMoflon
References
[1] Anthony Anjorin, Marius Lauder, Sven Patzina & Andy Schürr (2011): eMoflon : Leveraging EMF and
Professional CASE Tools. In: Proc. of Informatik 2011, LNI 192, GI, p. 281.
[2] Anthony Anjorin, Gergely Varró & Andy Schürr (2012): Complex Attribute Manipulation in TGGs with
Constraint-Based Programming Techniques. In: Proc. of Bx 2012, ECEASST 49, EASST.
[3] Aaron Bohannon, J.N. Foster, B.C. Pierce, Alexandre Pilkiewicz & Alan Schmitt (2008):
Boomerang: Resourceful Lenses for String Data.
ACM SIGPLAN Notices 43(1), pp. 407–419,
doi:10.1145/1328897.1328487.
[4] Thorsten Fischer, Jörg Niere, Lars Torunski & Albert Zündorf (2000): Story Diagrams: A New Graph Rewrite
Language Based on the Unified Modeling Language and Java. In Hartmut Ehrig, Gregor Engels, HansJörg Kreowski & Grzegorz Rozenberg, editors: Proc. of TAGT 1998, LNCS 1764, Springer, pp. 296–309,
doi:10.1007/978-3-540-46464-8 21.
[5] Florian Heidenreich, Jendrik Johannes, Sven Karol, Mirko Seifert & Christian Wende (2009): Derivation
and Refinement of Textual Syntax for Models. In Richard F. Paige, Alan Hartman & Arend Rensink, editors:
Proc. of ECMDA-FA 2009, LNCS 5562, Springer, pp. 114–129, doi:10.1007/978-3-642-02674-4 9.
[6] Soichiro Hidaka, Zhenjiang Hu, Kazuhiro Inaba, Hiroyuki Kato & Keisuke Nakano (2011): GRoundTram:
An Integrated Framework for Developing Well-Behaved Bidirectional Model Transformations. In Perry
Alexander, Corina Pasareanu & John Hosking, editors: Proceedings of ASE 2011, IEEE, pp. 480–483,
doi:10.1109/ASE.2011.6100104.
[7] Tassilo Horn (2013): The TTC 2013 Flowgraphs Case. In: Sixth Transformation Tool Contest (TTC 2013),
EPTCS this volume.
[8] Felix Klar, Marius Lauder, Alexander Königs & Andy Schürr (2010): Extended Triple Graph Grammars
with Efficient and Compatible Graph Translators. In Andy Schürr, C. Lewerentz, G. Engels, W. Schäfer
& B. Westfechtel, editors: Nagl Festschrift, LNCS 5765, Springer, pp. 141–174, doi:10.1007/978-3-64217322-6 8.
[9] Marius Lauder, Anthony Anjorin, Gergely Varró & Andy Schürr (2012): Efficient Model Synchronization
with Precedence Triple Graph Grammars. In Hartmut Ehrig, Gregor Engels, Hans-Jörg Kreowski & Grzegorz
Rozenberg, editors: Proc. of ICGT 2012, LNCS 7562, Springer, pp. 401–415, doi:10.1007/978-3-642-336546 27.
[10] Christian Nentwich, Licia Capra, Wolfgang Emmerich & Anthony Finkelstein (2002): Xlinkit: A Consistency
Checking and Smart Link Generation Service. ACM Transactions on Internet Technology 2(2), pp. 151–185,
doi:10.1145/514183.514186.
[11] Terence John Parr (2004): Enforcing strict model-view separation in template engines. Proc. of WWW ’04,
p. 224, doi:10.1145/988672.988703.
[12] Terence John Parr (2007): The Definitive ANTLR Reference: Building Domain-Specific Languages. The
Pragmatic Bookshelf.
[13] Andy Schürr (1994): Specification of Graph Translators with Triple Graph Grammars. In G. Tinhofer,
editor: Proc. of WG 1994, Lecture Notes in Computer Science 903, Springer, pp. 151–163, doi:10.1007/3540-59071-4 45.
[14] Perdita Stevens (2008): A Landscape of Bidirectional Model Transformations. In Ralf Lämmel, Joost Visser
& João Saraiva, editors: Proc. of GTTSE 2007, LNCS 5235, Springer, pp. 408–424, doi:10.1007/978-3-54088643-3 10.
[15] Transformation Solution with eMoflon hosted on Share (2013): http: // is. ieis. tue. nl/ staff/
pvgorp/ share/ ?page= ConfigureNewSession&vdi= XP-TUe_ TGG-Comparison_ eMoflon_ TTC13_
eMoflon. vdi .
| 6cs.PL
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.